lang
stringclasses 1
value | s2FieldsOfStudy
listlengths 0
8
| url
stringlengths 78
78
| fieldsOfStudy
listlengths 0
5
| lang_conf
float64 0.8
0.98
| title
stringlengths 4
300
| paperId
stringlengths 40
40
| venue
stringlengths 0
300
| authors
listlengths 0
105
| publicationVenue
dict | abstract
stringlengths 1
10k
⌀ | text
stringlengths 1.94k
184k
| openAccessPdf
dict | year
int64 1.98k
2.03k
⌀ | publicationTypes
listlengths 0
4
| isOpenAccess
bool 2
classes | publicationDate
timestamp[us]date 1978-02-01 00:00:00
2025-04-23 00:00:00
⌀ | references
listlengths 0
958
| total_tokens
int64 509
40k
|
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
en
|
[
{
"category": "Education",
"source": "s2-fos-model"
},
{
"category": "Environmental Science",
"source": "s2-fos-model"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/020122fcd3271980e2e32e33bf7f15599b812a32
|
[] | 0.884879
|
Building a Blockchain-Based Decentralized Crowdfunding Platform for Social and Educational Causes in the Context of Sustainable Development
|
020122fcd3271980e2e32e33bf7f15599b812a32
|
Sustainability
|
[
{
"authorId": "2267990636",
"name": "B. Țigănoaia"
},
{
"authorId": "2267993831",
"name": "George-Madalin Alexandru"
}
] |
{
"alternate_issns": null,
"alternate_names": null,
"alternate_urls": [
"http://mdpi.com/journal/sustainability",
"http://nbn-resolving.de/urn/resolver.pl?urn=urn:nbn:ch:bel-172127"
],
"id": "8775599f-4f9a-45f0-900e-7f4de68e6843",
"issn": "2071-1050",
"name": "Sustainability",
"type": "journal",
"url": "http://www.e-helvetica.nb.admin.ch/directAccess?callnumber=bel-172127"
}
|
Blockchain technology contributes to achieving the Sustainable Development Goals. Education for sustainable development (ESD) is UNESCO’s education sector response to the urgent and dramatic challenges the planet faces. The traditional way of donating money to charitable causes, such as education, has been through centralized methods and organizations that lack transparency, and donors often do not have a clear understanding of how their contributions are being utilized. Blockchain technology, particularly, platforms like Ethereum and Polygon, has the potential to address the issues associated with traditional donation systems. This paper proposes a decentralized web3 application that utilizes blockchain technology to enhance transparency and efficiency in educational donations in the context of sustainable development. The platform leverages decentralized protocols and smart contracts to ensure secure and transparent transactions, enabling donors to track the utilization of their contributions and ensuring their funds reach their intended beneficiaries. This paper discusses the design and implementation of the platform, highlighting its features and potential for transforming the landscape of charitable donations. This software application can be used in education, and a demo plus some scenarios/work cases are presented/analyzed. The main results and contributions open other future research directions for not only authors.
|
## sustainability
_Article_
# Building a Blockchain-Based Decentralized Crowdfunding Platform for Social and Educational Causes in the Context of Sustainable Development
**Bogdan Tiganoaia** **[1,]*** **and George-Madalin Alexandru** **[2]**
1 Entrepreneurship and Management Department, National University of Science and Technology
POLITEHNICA Bucharest, 060042 Bucharest, Romania
2 Computer Science Department, National University of Science and Technology POLITEHNICA Bucharest,
060042 Bucharest, Romania
***** Correspondence: bogdantiganoaia@gmail.com
**Abstract: Blockchain technology contributes to achieving the Sustainable Development Goals. Edu-**
cation for sustainable development (ESD) is UNESCO’s education sector response to the urgent and
dramatic challenges the planet faces. The traditional way of donating money to charitable causes,
such as education, has been through centralized methods and organizations that lack transparency,
and donors often do not have a clear understanding of how their contributions are being utilized.
Blockchain technology, particularly, platforms like Ethereum and Polygon, has the potential to address the issues associated with traditional donation systems. This paper proposes a decentralized
web3 application that utilizes blockchain technology to enhance transparency and efficiency in educational donations in the context of sustainable development. The platform leverages decentralized
protocols and smart contracts to ensure secure and transparent transactions, enabling donors to
track the utilization of their contributions and ensuring their funds reach their intended beneficiaries.
This paper discusses the design and implementation of the platform, highlighting its features and
potential for transforming the landscape of charitable donations. This software application can be
used in education, and a demo plus some scenarios/work cases are presented/analyzed. The main
results and contributions open other future research directions for not only authors.
**Citation: Tiganoaia, B.; Alexandru,**
G.-M. Building a Blockchain-Based
Decentralized Crowdfunding
Platform for Social and Educational
Causes in the Context of Sustainable
Development. Sustainability 2023, 15,
[16205. https://doi.org/10.3390/](https://doi.org/10.3390/su152316205)
[su152316205](https://doi.org/10.3390/su152316205)
Academic Editor: Yang (Jack) Lu
Received: 3 October 2023
Revised: 7 November 2023
Accepted: 15 November 2023
Published: 22 November 2023
**Copyright:** © 2023 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
[Attribution (CC BY) license (https://](https://creativecommons.org/licenses/by/4.0/)
[creativecommons.org/licenses/by/](https://creativecommons.org/licenses/by/4.0/)
4.0/).
**Keywords: education; blockchain; decentralization; web3 platform; Ethereum; smart contracts;**
transparency; sustainable development
**1. Introduction**
The goals for sustainable development, named the SDGs, means a total of 17 general
targets with a highlight on sustainable development aspects related to:
Education, poverty, and equality.
_•_
The changes in climate, infrastructure, water, and land.
_•_
Production and consumption.
_•_
The target for these SDGs is the year 2030.
_About Sustainability Education_
UNESCO shares knowledge, produces information, and provides policy support plus
technical guidance to its Member States. It implements projects related to the SDGs. It
also acts as an advocate with one objective: that governments are able to provide quality
Climate Change Education (CCE) [1].
Education for sustainable development is an exciting new field. The objective is to
highlight the connections between the environment, society, and economy. This contributes
to a sustainable future [2]:
-----
_Sustainability 2023, 15, 16205_ 2 of 19
1. Students are able to understand and apply the concepts (at a basic level) and principles
related to sustainability.
2. Students know about sustainability viewed as a triad: economic plus ecological plus
social systems.
Today, the blockchain is an important and new technology. Its impact is related to:
The economy, as it can transform interactions (it is about social interactions);
_•_
Public institutions;
_•_
Our relationship with the environment [3].
_•_
How does blockchain technology contribute to achieving the SDGs? The answer can
be considered as follows:
1. Blockchain technology has apps that regard sustainable goals in different ways.
2. It offers resilience and security capabilities, without any implications that may harm
third parties.
3. The blockchain guarantees precision in the surveillance of actions [4]. We can use
smart contracts, which validate the actions that occur within the blockchain.
The paradigm shift that started with blockchain technology can be seen in the world’s
biggest companies like Google, IMB, and Meta (formerly Facebook), which are adopting
this technology to build new decentralized applications or to integrate it into already
existing products to enhance security, transparency, and innovation.
Decentralized applications are software apps that live and run on the blockchain
instead of on a single server. These applications can benefit from security (most blockchains
use cryptographic algorithms to secure data), user privacy, and a lack of censorship.
Blockchain technology is most commonly used in the financial and banking industries,
with protocols like AAVE allowing you to stake digital assets like cryptocurrencies and
stable coins to receive incentives and borrow assets against your collateral. The advantages
of blockchains can also benefit the educational industry:
1. Shah et al. [5] presented a combined approach in which they integrated a blockchain
for storing student information, used a machine learning algorithm to forecast the
potential job roles for students post-graduation, and evaluated the outcomes using
various machine learning methods.
2. Zhang et al. [6] presented an innovative approach aimed at improving the management of
teaching information in higher education with the integration of blockchain technology.
3. Gresch et al. [7] presented a fresh technique for enhancing the transparency of educational certificates with the application of blockchain technology. They integrated
blockchain networks at every stage of the system, ensuring the secure transmission of
student certificates.
People often donate money to charities because they wish to give back to society
and their communities, because they believe in a certain cause, or because they have
experienced some traumatic events and the donations of others have been helpful to them
or their loved ones. Some of the main causes that people donate money to support include,
for example:
_•_ Disaster relief: to assist those affected by natural disasters like earthquakes, hurricanes,
and floods.
_•_ Health causes: to assist those in need of financial assistance for surgeries or to support
research into diseases such as cancer, Alzheimer’s, or Parkinson’s.
Animal shelters.
_•_
_•_ Education: to support local schools and universities as well as to provide scholarships
and school supplies for students in need.
With the development of technology, donations can now be made using a variety of
channels, including social media, online donation platforms, and SMS. The Blackbaud
Institute conducted research on charitable giving [8], and it estimated that the total amount
of donations made in the United States in 2021 was around USD 46.4 billion, of which
-----
_Sustainability 2023, 15, 16205_ 3 of 19
USD 2.9 billion came from online donations, representing a 42% increase in overall online
giving since 2019. The problem with these traditional ways of donating money to charities
is that they lack transparency. Centralized organizations, also known as banks, handle
the processing of donated money, and most of the time, the donors do not have a clear
understanding of how their contributions are being utilized. Furthermore, the distribution
of the raised funds may be slow and bureaucratic, with funds often taking a long time to
reach the intended beneficiaries.
The technology of distributed and decentralized networks has advanced significantly
over the past few years, particularly blockchain technology. Numerous public decentralized
blockchain protocols like Bitcoin, Ethereum, Elrond, and Polygon have been developed and are
used by millions of users daily. This paper proposes an alternative solution to the conventional
methods of donating funds to charities, which uses a decentralized blockchain protocol and
allows users to transparently see how all the contributions are being utilized to support each
cause. This paper is divided into seven sections: the second presents the motivation and the
weaknesses of the existing centralized methods; the third presents the main concepts and
technologies (such as blockchain, smart contracts, wallets, Ethereum, and Polygon); the fourth
describes the proposed solution and its implementation; the fifth illustrates a short demo of
the platform; the sixth discusses the security of the smart contract; and the final section wraps
up this article and discusses some ideas for future development.
**2. Motivation**
To design the solution and a smart contract, we started with the problems of the
existing tools that we want to solve or improve, namely:
High fees: By using a public blockchain protocol, such as Polygon, we can eliminate
_•_
the commissions applied by donation collection platforms. However, in order to
interact with the platform, in particular, with its smart contract, users have to pay
fees imposed by the Polygon network to secure it. These fees are significantly lower
compared with existing platforms. For example, if we compare the fees of 2.9% + USD
0.30 for each transaction on the GoFundMe platform with the fees paid to the Polygon
network for a donation of USD 100, we obtain the following results:
GoFundMe: out of USD 100, USD 3.2 is tax-represented and USD 96.8 reaches
�
beneficiaries
Proposed solution: out of USD 100 paid to MATIC, the full amount will reach
�
the balance of the contract, with the user paying separately a transaction fee of
MATIC 0.00015112, which means less than USD 0.001 (calculated taking into
account the network load at that time and using the price of 0.61 USD/MATIC
on 11 June 2023).
Bureaucracy and delays: Some platforms, such as DonorsChoose, use a platform
_•_
campaign verification process to ensure that they follow some rules and standards.
Due to the public nature of the smart contract, anyone can interact with it and create
new campaigns. After the completion and expiration of the campaign deadline, their
initiators can immediately use the amounts collected, without any further delays.
_•_ Lack of transparency: This is a sensitive and difficult issue to address. Although most
platforms provide users with a transparent way to track the progress of donations
after the end of the campaign and the transfer of money to beneficiaries, there is no
longer a possibility to follow how they are used to support the case. The proposed
solution implements, with smart contracts, a functionality in which the initiators of the
campaigns must provide a description of the transaction, the address of the recipient,
and the amount they intend to use. For example, there may be a campaign that
aims to raise donations for the purchase of beds to equip a ward in a hospital. After
the end of the campaign, after the necessary funds are collected, the initiator of the
campaign can use the platform to send the necessary amount to the bed merchant. This
can be performed using a form that completes the description of the transaction (for
example, the name of the beneficiary hospital and the number of beds purchased can be
-----
_Sustainability 2023, 15, 16205_ 4 of 19
mentioned), the amount of the transaction, and the address of the consignee (in the case
of the above-mentioned example, the address of the bed trader, which can theoretically
be verified on the official website of the trader). All these transactions are public and
available both on the blockchain and within the platform, providing transparency to
donors and the opportunity to monitor the progress and use of donations.
Geographical limitation: As mentioned above, platforms are generally only available
_•_
in certain countries, such as the United States and some countries in Europe or Asia.
Due to the decentralized idea of the blockchain, the proposed solution is available
regardless of the geographical region in which the users are located, provided that the
law of the country in which the activity takes place allows the holding and trading of
cryptocurrencies [9].
**3. Main Concepts and Technologies**
_3.1. Blockchain_
Blockchain technology is considered by universities as an approach to improving the
teaching and learning process. It encourages the participation of all stakeholders like
(1) Undergraduates;
(2) Professors;
(3) Family members [10].
Blockchain technology is based on Distributed Ledger Technology (DLT), which enables direct transactions between users without the need for intermediaries or a centralized
authority to oversee them [11]. Transactions are validated with a consensus mechanism
within an interconnected network of computers.
What is a blockchain? In 1991, Stuard Haber and W. S. Stornetta published a paper
titled “How to Time-Stamp a Digital Document” [12], in which they proposed a method
for digitally time-stamping documents using hash functions, digital signatures, and data
stored in blocks. This paper is considered to be the first description of the blockchain
concept. Now, we refer to the term “blockchain” as a distributed database or ledger that is
shared among the nodes of a computer network and stores data into blocks that are chained
together using different consensus algorithms. Being open and distributed, the blockchain
provides immutability, security, and transparency.
In 2008, Satoshi Nakamoto published a paper titled “Bitcoin: A Peer-to-Peer Electronic
Cash System”, in which he proposed a decentralized financial instrument using a digital
currency called Bitcoin. It proposes a “peer-to-peer network using proof-of-work to record
a public history of transactions” [13].
There are several organizations with connections in the development of a blockchain,
such as:
IBM is the most involved and a principal investor.
_•_
Mastercard is another organization that has over 100 blockchain patents filed. This
_•_
company uses technology to increase protection against fraud and to reduce transaction costs [14].
According to Investopedia [15], the three biggest blockchain companies are
� Coinbase Global Inc. (San Francisco, CA, USA)—COIN;
� Canaan Inc. (Beijing, China)—CAN;
� Galaxy Digital Holdings Ltd. (New York, NY, USA)—BRPHF.
Last, but not least, some authors used blockchain technology to develop the BookChain
project—a secure library book for storing and sharing in academic institutions. For details,
please visit [16]. Another application of blockchain technology is an electronic voting
system—for more information, see [17].
-----
_Sustainability 2023, 15, 16205_ 5 of 19
_3.2. Cryptocurrency_
A cryptocurrency is a decentralized digital currency that uses cryptography to secure
transactions and that lives on a blockchain, which may be interpreted as a public digital
ledger that records all transactions made using the cryptocurrency. In contrast with traditional fiat currencies, cryptocurrencies are not issued, regulated, or backed by any financial
institution. Instead, the transactions are verified and approved by a network of users that
uses various consensus algorithms such as Proof-of-Work or Proof-of-Stake.
The popularity of blockchain technology and cryptocurrencies increased considerably
[after the launch of Bitcoin in 2009. As stated on CoinMarketCap’s website (https://](https://coinmarketcap.com)
[coinmarketcap.com, accessed on 10 September 2023) [18], on 23 April 2023, there were](https://coinmarketcap.com)
23,562 cryptocurrencies with a total market capitalization of USD 1.17 trillion. According
[to a report conducted by Crypto.com (for more information see https://crypto.com/,](https://crypto.com/)
accessed on 10 September 2023), on-chain data analysis revealed that the total number of
global crypto owners reached 425 million in December 2022 [19].
In addition, according to the “Developer Report” by Electric Capital, there are 23,343 active developers in crypto monthly.
_3.3. Polygon_
The blockchain Trilemma [20] covers the challenges faced by developers in building
a blockchain that is secure, decentralized, and scalable without sacrificing any of these
characteristics.
Even though it sacrificed its scalability, the Ethereum Foundation has focused its
efforts on building a decentralized and secure blockchain. As a result, transactions can
be slow and expensive. According to the Etherscan website (see [21]), on 24 April 2023,
the average gas price was around Gwei 44, which means that a simple transaction on the
Ethereum network costs about USD 1.78 and takes roughly three minutes to complete.
As stated on their website, “Polygon is a Layer 2 scaling solution” (see [22]). As a
Layer-2 protocol, Polygon intends to increase transaction speed and reduce costs for users
rather than duplicate Ethereum’s functionalities.
The native cryptocurrency of the Polygon Network is MATIC. As a comparison, the
current number of transactions made on Ethereum per second is around 11, while on
Polygon, it is 34. However, Polygon promises the potential of over 7000 transactions per
second. If we look at the cost and completion time of a simple transaction, the average gas
price on 24 April 2023 was Gwei 439.6, which means that a transaction costs about USD
0.00823 and takes between 30 and 60 s to complete. Compared with Ethereum, this is 3 to
6 times faster and 215 times less expensive.
_3.4. Sustainable Development Goals_
Some important SDGs (selection from a set of 17 Goals) include:
Goal No. 1: NO POVERTY;
_•_
Goal No. 4: QUALITY in EDUCATION;
_•_
Goal No. 5: GENDER EQUALITY;
_•_
Goal No. 7: CLEAN ENERGY and AFFORDABLE ENERGY;
_•_
Goal No. 11: SUSTAINABLE CITIES AND COMMUNITIES.
_•_
**4. Application for Social and Educational Causes**
The web platform in this study is designed to facilitate interaction between blockchain
and the deployed smart contract by creating an intuitive and user-friendly interface to create
new campaigns and donate money to support the causes. The platform is developed using
Next.JS (NodeJS [23] is also a solution for developers) framework, which is a React framework
that enables the creation of full-stack web applications (for more information, see [24]).
The modules of the application are described below.
-----
_Sustainability 2023, 15, 16205_ 6 of 19
_4.1. Frontend_
The frontend part of the application is developed using React, HTML, and CSS, and
its role is to provide an interactive and intuitive interface for users to easily interact with
the platform. Also, this module facilitates communication between the user, the smart
contract, and the backend of the application. Using the frontend, users can view and donate
cryptocurrencies to already-running campaigns, create a new campaign, and claim raised
funds for their own campaigns.
_4.2. Backend_
The backend part of the application is developed using JavaScript language, and its
role is to handle the server-side logic, data storage, and communication with other systems.
_4.3. IPFS module_
4.3.1. The IPFS
The IPFS (InterPlanetary File System) is “a peer-to-peer hypertext protocol designed
to preserve and develop the knowledge of humanity by transforming the web into an
up-to-date, resilient and more open platform” [25]. It was created to solve the problems of
scalability, redundancy, and censorship associated with traditional architectures for storage
and file transfer. This protocol is ideal for the proposed platform because we can use
it to store large files outside the blockchain network, such as pictures, and to store only
immutable and permanent links to those files on the blockchain.
4.3.2. The IPFS Module
The platform interface provides the ability to add a representative picture for each
campaign. In the contract, there is no possibility of storing pictures for campaigns. The
classic solution would be to add the pictures to a database (more info about databases
and SQL in [26]), but this would limit the decentralized nature of the proposed solution.
To avoid using a centralized database specific to the platform, we chose to use the IPFS
protocol to store campaign images and saved the reference to the picture uploaded on IPFS
in the contract for each campaign. Subsequently, this reference is used by the application
interface to retrieve the picture from the IPFS and display it to users.
_4.4. Wallet Connect_
In order to use the platform, it is mandatory to have a wallet. When a wallet is created
on the blockchain, it generates two paired keys: a public key that is used for identification
and a private key that is used for authorization (e.g., for signing transactions). In the new
Web 3.0 iteration, decentralized applications will now authenticate users by using wallets.
[For implementing the connection to a crypto wallet, RainbowKit (https://www.](https://www.rainbowkit.com)
[rainbowkit.com, accessed on 15 of September 2023 [27]) is used, which is a React library](https://www.rainbowkit.com)
that facilitates the wallet connection to decentralized applications.
_4.5. Smart Contracts_
Smart contracts are programs that are stored on a blockchain and run when predetermined conditions are achieved. Smart contracts are used to automate the running of an
agreement. There is no involvement of intermediaries or time loss. Smart contracts can
automate workflows. They can trigger the next action when conditions are achieved [28].
A smart contract works as a simple “if/when...then...” statement. These statements are
inserted into the code of the blockchain. When the transaction is finished, the blockchain
is then updated. The conclusion refers to the fact that the transaction cannot be changed.
Another important aspect here is that only parties who have been granted permission can
view the results (based on [28]). The concept of “smart contracts” was first introduced
in the early 1990s by computer scientist Nick Szabo in his work titled “Formalizing and
Securing Relationships on Public Networks” [29]. Bitcoin is the first blockchain that uses
-----
_Sustainability 2023, 15, 16205_ 7 of 19
this new concept, with Vitalik Buterin affirming in the Ethereum whitepaper that “Bitcoin
protocol actually does facilitate a weak version of a concept of “smart contracts”” [30].
Nowadays, smart contracts are implemented in numerous blockchain protocols, such
as Ethereum [21,31] and Elrond [32], and can be seen as self-executing programs that ensure
that the terms of an agreement are respected or fulfilled without the need for trust among
the involved parties.
The smart contract for our applications is developed using Solidity. It is an objectoriented high-level language used for smart contracts. Ethereum is considered the most
known and secure blockchain, with a total of more than 500,000 validators as of December
2022. The main disadvantage of it is that it has a low level of scalability; it can only process
10–30 transactions per second, resulting in high gas fees. To solve the scalability issue and
to avoid the high gas fees of the Ethereum network, which is seen as a Layer-1 blockchain,
we chose to use the Polygon network. It is an overlay of Ethereum, also known as a Layer-2
protocol, that offers higher speed and lower gas fees. This Layer-2 protocol can handle
around 700 TPS, and the gas fees are around USD 0.01 per transaction. The contract is
deployed on the Polygon public blockchain, and users can interact with it using either the
platform or directly on the blockchain using the Polygon scan website. The smart contract
was developed by the authors without any additional cost.
Figure 1 shows the scheme of the contract and the data structures used. The data
structures used are as follows:
Donation: for each donation, this structure retains the donated value and the address
_•_
of the donor;
_•_ Transaction: for each transaction made by the campaign initiators, this structure retains
the donated value, the address of the recipient, and a description of the transaction;
_•_ Campaign: this structure is used to retain information relevant to a campaign, namely:
an ID, name, objective, balance, address of the originator, number of donations, a
deadline saved in UNIX timestamp format (UNIX timestamp = number of seconds past
from 1 January 1970 until now), a description of the campaign, a flag to know whether
or not the campaign is over, a reference to the image of the campaign uploaded to the
IPFS, and the final amount raised.
**Figure 1. Smart contract and data structures used.**
The structure of an intelligent contract is similar to the structure of a class in an objectoriented language such as C++ or Java. The “Crowdfunding” contract encapsulates the
main functionality of the crowdfunding platform and provides the necessary functions and
-----
_Sustainability 2023, 15, 16205_ 8 of 19
mapping for the management of campaigns, donations, and transactions. It contains the
following attributes and methods:
1. Attributes:
Owner: the address of the contract owner;
_•_
_•_ indexOfCampaign: the global variable used to assign a unique ID to each campaign;
Campaigns: the map structure that associates the ID of a campaign with the
_•_
structure of the campaign;
userDonationPerCampaign: mapping that retains the amount donated by each
_•_
donor for each campaign;
donationsPerCampaign: mapping that retains donations for each campaign;
_•_
_•_ transactionsPerCampaign: mapping that retains transactions performed for each
campaign after its completion.
2. Methods:
createCampaign: a function that allows users to create a new campaign, check
_•_
the deadline to ensure it is in the future, register the campaign in the campaigns
mapping, and increment indexOfCampaigns;
Donate: allows users to donate to a campaign, check the amount donated, up
_•_
date the campaign balance, record the donation in the DonationsPercampaign
mapping, and increase the number of donations for the campaign;
_•_ endCampaign: allows the owner of a campaign to end the campaign, check that
the campaign is still active and the deadline has been exceeded, and update the
status of the campaign;
useFunds: allows the owner of a campaign to use the funds collected, check
_•_
various conditions such as the fact that the campaign has been completed, the
caller is the owner, the balance of the campaign is not zero, the amount used is
positive, etc., update the campaign balance, record the transaction in mapping
transactionsPerCampaign, and transfer the funds to the recipient;
getCampaign: getter function that returns information about a campaign;
_•_
getBalanceOfContract: getter function that returns the balance of the contract;
_•_
_•_ getUserDonationPerCampaign: getter function that returns the amount donated
by a specific user for a specified campaign;
getCampaigns: returns a list of “Campaign” structures that represent all the
_•_
campaigns created so far;
_•_ getDonationsPerCampaign: returns a list of “Donation” structures that represent
all registered donations for a specified campaign;
getTransactionsPerCampaign: returns a list of “Transaction” structures that
_•_
represent all transactions performed for a specified campaign.
The following diagram—see Figure 2, represents how the crowdfunding platform
is structured:
-----
_Sustainability 2023, 15, 16205_ 9 of 19
**Figure 2. The architecture of the application.**
**5. A Short Demo for Educational Campaigns**
_An Introduction to Ethereum_
Ethereum is a technology that is home to digital money, global payments, and
applications [33].
Ethereum is a decentralized open-source blockchain platform that natively supports
smart contracts. It was first introduced in 2014 by Vitalik Buterin in his paper entitled “Ethereum: A Next-Generation Smart Contract and Decentralized Application Platform” [30] and later launched in 2015. When the Ethereum network originally started, it
relied on a proof-of-work consensus system that let miners compete by solving mathematical problems to add new blocks to the chain and earn Ether as a reward. Later, in 2022,
Ethereum switched to a proof-of-stake algorithm. In this new proof-of-stake model, miners
stake their own Ether into a smart contract and use it as collateral in order to validate new
blocks and earn rewards in Ether proportional to the amount they staked.
Ether (ETH) is the native cryptocurrency of the Ethereum network. As stated on their
website, “Ether is the main internal crypto-fuel of Ethereum” [33], and it is used as payment
for transaction fees when users interact with the network or as collateral for staking in
order to secure the network and earn rewards.
Regarding Ethereum, two important concepts can be discussed:
1. Accountability in Ethereum—for details, see [34].
2. Anonymity in Ethereum—for details, see [35].
In this section, we will illustrate how you can use the platform to create a charity
or educational campaign, donate to different causes, and claim the raised funds after the
campaign goal is reached. Blockchain technology can be used for developing applications
for other uses such as voting applications in non-profit systems (universities, communities,
etc.). These applications can be described by use case. Next, we detail a short demo for
educational campaigns.
**_Step 1: When you first access the platform—see Figure 3, you will be redirected to the_**
Donate page where you can see all the ongoing campaigns.
-----
_Sustainability 2023, 15, 16205_ 10 of 19
**Figure 3. Donate page where all ongoing campaigns are displayed.**
**_Step 2: If you click on the “Create campaign” button, a form for creating a new_**
campaign will be displayed—see Figure 4.
**Figure 4. Form for creating a new campaign.**
**_Step 3: After you fill the form by adding a title, telling your story, and setting the goal_**
(in MATIC) and the deadline for your campaign, at the bottom of the form, you will see a
_Sustainability 2023,message that informs you that you should first connect your wallet in order to create a new 15, x FOR PEER REVIEW_ 11 of 20
campaign—see Figure 5. You can connect your wallet using the button at the top right corner.
**Figure 5. Filled form for creating a new campaign.Figure 5. Filled form for creating a new campaign.** **Comme**
**_Step 4: After you click on the “Connect wallet” button—see Figure 6 a new modal_**
-----
_Sustainability 2023, 15, 16205_ 11 of 19
**_Step 4: After you click on the “Connect wallet” button—see Figure 6, a new modal_**
will appear, from which you can choose the wallet provider you want to use.
**Figure 6. Modal for choosing which wallet provider you want to use.**
**_Step 5: After you connect your wallet, in the right top corner, you will see some_**
information regarding your wallet like the balance of your account and a part of your
public address. Also, a submit button will appear on the bottom of the form. After clicking
on the “Submit new campaign” button, you will be asked to sign a transaction because
this way you interact with the smart contract, which is deployed on the Polygon Mumbai
testnet. Figures 7–9 show some other aspects related to a new campaign.
**Figure 7. Form for creating a new campaign with the submit button active.**
**Figure 8. Signing the transaction for creating a new campaign.**
-----
_Sustainability 2023, 15, 16205_ 12 of 19
**Figure 9. Display message after the campaign was successfully created.**
**_Step 6: After successfully creating the campaign, you can either close the modal or_**
choose to go to the campaign page. We chose to close the modal and go to the “Donate”
page—see Figures 10 and 11, where we were able to see the newly created campaign.
**Figure 10. Donate page displaying the newly created campaign—Part 1.**
**Figure 11. Donate page displaying the newly created campaign—Part 2.**
-----
_Sustainability 2023, 15, 16205_ 13 of 19
**_Step 7: You can now click on the newly created campaign, and a new page will open_**
displaying relevant information for the campaign like the amount raised, the number of days
left, the number of backers, the number of donations, the story, and a list with all the donations.
**_Step 8: Now, you can support this campaign. First, you need to fill in the amount of_**
MATIC you want to donate and after that, you can click on the “Donate” button. You will
be asked again to sign a transaction and after the transaction is complete, a message will be
displayed. Some other aspect related to the campaign can be viewed in the Figures 12–19.
**_Step 9: After closing the modal, we now can see that the campaign raised 100% of the_**
goal. Also, we can see the list of donations.
**_Step 10: Now, if we scroll down, we can see an “End Campaign” section. This section_**
only appears when you are the owner of the contract. After successfully raising all the
funds needed for your campaign or after the deadline has passed, you can click on the
“Finish campaign” button to end the campaign and claim the raised amount. You will be
asked again to sign a transaction, and after the transaction is finished, a new message will
appear on the screen.
**Figure 12. The Donate page ready for a donation of MATIC 0.5.**
**Figure 13. Signing a transaction for donating to the campaign.**
-----
_Sustainability 2023, 15, 16205_ 14 of 19
**Figure 14. Display message after the donation was successful.**
**Figure 15. Campaign page with the updated state.**
-----
_Sustainability 2023, 15, 16205_ 15 of 19
**Figure 16. Campaign page with the “End Campaign” section.**
**Figure 17. Signing the transaction for finishing the campaign and claiming the raised funds.**
-----
_Sustainability 2023, 15, 16205_ 16 of 19
**Figure 18. Display message after ending the campaign successfully.**
**Figure 19. The Polygon scan—the transparent view of the crowdfunding platform.**
All the interactions with the smart contract can also be found on Polygon scan, which
provides users with a transparent view of what is happening on the crowdfunding platform.
**6. Security of the Smart Contract—Unit Testing**
Unit tests are a form of software testing that focuses on verifying the individual
functionality of its smallest components, called units. These can be functions, methods, or
classes of an application. The main purpose of unit tests is to validate the correct behavior
and functionality of these units. You can develop your apps by using [36] and monitor
them by using [37].
In the blockchain environment, a single mistake can compromise the security of the
contract and implicitly of its funds. Once a contract is loaded on a blockchain, it becomes
public and immutable, and any errors in the contract can no longer be resolved. These
errors can lead to contract vulnerabilities that can be exploited by malicious users, and for
this reason, contract testing is a necessary step.
The Hardhat development environment also provides support for testing contracts
by writing unit tests. In Figure 20, the coverage percentage of the code with the unit
tests written for the verification of the contract is presented. It measures the coverage of
instructions, decision-making branches, functions, and lines in the code. The percentage
for covering branches is 94.74% because we failed to simulate the case where the transfer of
cryptocurrencies between the contract and an external address fails.
-----
_Sustainability 2023, 15, 16205_ 17 of 19
**Figure 20. Percentage of coverage of written unit tests.**
**7. Contributions, Future Work, Limitations and Conclusions**
_7.1. Contributions_
This platform aims to demonstrate the potential of blockchain technology and decentralized applications, provide a decentralized and transparent system for creating and
managing crowdfunding campaigns, and enable greater trust and accountability between
donors and recipients in the field of charitable giving to support educational causes. Some
contributions include:
Bibliographic research on the paper topic: Polygon, Ethereum, blockchain, smart
_•_
contracts, etc.;
The design of the platform (including the architecture of the platform);
_•_
_•_ The implementation of the web application, which consists of the frontend and backend;
The implementation of smart contracts for the platform;
_•_
A short demo for charity and educational campaigns.
_•_
This project tries to innovate by providing a decentralized solution for raising funds
for schools, students, and the educational field in general. It uses the latest technologies
like Next.JS for the web platform development and also combines blockchain protocols like
Polygon, which is the network where the smart contract is deployed, and IPFS, where the
platform stores images to not limit the decentralized character of the proposed solution.
_7.2. Future Work_
The platform’s next stage is to offer a brand-new, improved user interface. Also,
we plan to integrate multiple blockchains and cryptocurrencies, explore the use of nonfungible tokens (NFTs) as a means of incentivizing donations, and integrate smart contracts
to automate and streamline fund distribution. Blockchains can reshape the educational
system as we know it. In this context, blockchains provide significant benefits that can
secure processes. Blockchains create security and trust, as they eliminate the need for an
intermediary to validate transactions.
_7.3. Limitations_
Generally speaking, the desire for objectives like decentralization, transparency, and
privacy has limitations. Also, the platform’s user interface and experience are still in their
early stages. The user interface is on the second version and could also be improved.
Regarding the application, it is not fully decentralized due to some data that are not
stored on the blockchain, and we aim for them to be improved or corrected in case of errors.
The application stores information like images or descriptions of the charities and any other
relevant data.
_7.4. Conclusions_
The Sustainable Development Goals (more info in [38])—SDGs—are a set of 17 general
targets focused on sustainable development issues. The following is a selection from this
set of 17 goals:
Goal no 2: Zero hunger;
_•_
Goal no 3: Good health plus well-being;
_•_
Goal no 6: Clean water plus sanitation;
_•_
Goal no 8: Decent work plus economic growth;
_•_
-----
_Sustainability 2023, 15, 16205_ 18 of 19
Goal no 9: Industry, innovation, and infrastructure;
_•_
Goal no 10: Reduced inequalities;
_•_
Goal no 12: Responsible consumption plus production;
_•_
Goal no 13: Climate action;
_•_
Goal no 14: Life below the water;
_•_
Goal no 15: Life on the land;
_•_
Goal no 16: Justice, peace, and strong institutions;
_•_
Goal no 17: Partnerships for goals.
_•_
According to this direction, in this paper, we propose a decentralized web3 application
that utilizes blockchain technology and addresses the problems of traditional donation
systems by creating a platform where people can create and manage charitable campaigns
to support educational causes, all of which are performed by interacting with a smart
contract deployed on a public blockchain. The platform leverages decentralized protocols
and smart contracts to ensure secure and transparent transactions, enabling donors to
track the utilization of their contributions and ensure their funds reach their intended
beneficiaries. The main contributions are highlighted in a separate section. The limitations
and future works are also provided. The software application can be used for now in
education, but not only, and it will have some new features in the near future.
**Author Contributions: Methodology, B.T. and G.-M.A.; Software, G.-M.A.; Validation, G.-M.A.;**
Investigation, B.T.; Resources, B.T.; Writing—original draft, G.-M.A.; Writing—review & editing, B.T.;
Visualization, G.-M.A.; Project administration, B.T.; Funding acquisition, B.T. All authors have read
and agreed to the published version of the manuscript.
**Funding: This research was funded by POLITEHNICA BUCHAREST.**
**Institutional Review Board Statement: Not applicable.**
**Informed Consent Statement: Informed consent was obtained from all subjects involved in the study.**
**Data Availability Statement: The data presented in this study are available on request from the**
corresponding author.
**Conflicts of Interest: The authors declare no conflict of interest.**
**References**
1. [Education for Sustainable Development. Available online: https://www.unesco.org/en/education-sustainable-development](https://www.unesco.org/en/education-sustainable-development)
(accessed on 2 May 2023).
2. [What Is Sustainability Education? 2021. Available online: https://online.sou.edu/degrees/education/msed/curriculum-and-](https://online.sou.edu/degrees/education/msed/curriculum-and-instruction-stem/what-is-sustainability-edu/)
[instruction-stem/what-is-sustainability-edu/ (accessed on 2 May 2023).](https://online.sou.edu/degrees/education/msed/curriculum-and-instruction-stem/what-is-sustainability-edu/)
3. [Sirimanne, S.N.; Freire, C. How Blockchain Can Power Sustainable Development. 2021. Available online: https://unctad.org/](https://unctad.org/news/how-blockchain-can-power-sustainable-development)
[news/how-blockchain-can-power-sustainable-development (accessed on 2 May 2023).](https://unctad.org/news/how-blockchain-can-power-sustainable-development)
4. [Blockchain and Its Impact on the SDGs (Sustainability Objectives). Available online: https://icommunity.io/en/blockchain-sdgs/](https://icommunity.io/en/blockchain-sdgs/)
(accessed on 2 May 2023).
5. Shah, D.; Patel, D.; Adesara, J.; Hingu, P.; Shah, M. Integrating machine learning and blockchain to develop a system to veto the
[forgeries and provide efficient results in education sector. Vis. Comput. Ind. Biomed. Art 2021, 4, 18. [CrossRef] [PubMed]](https://doi.org/10.1186/s42492-021-00084-y)
6. Zhang, L.; Ma, Z.; Ji, X.; Wang, C. Blockchain: Application in the System of Teaching Informatization Management of Higher
Education. In Proceedings of the 2020 3rd International Conference on Smart BlockChain (SmartBlock), Zhengzhou, China, 23–25
[October 2020; pp. 185–190. [CrossRef]](https://doi.org/10.1109/SmartBlock52591.2020.00041)
7. Gresch, J.; Rodrigues, B.; Scheid, E.; Kanhere, S.S.; Stiller, B. The Proposal of a Blockchain-Based Architecture for Transparent
Certificate Handling. In Business Information Systems Workshops. BIS 2018; Abramowicz, W., Paschke, A., Eds.; Lecture Notes in
[Business Information Processing; Springer: Cham, Switzerland, 2019; Volume 339. [CrossRef]](https://doi.org/10.1007/978-3-030-04849-5_16)
8. MacLaughlin, S.; Perrotti, E.; Thomson, A. Charitable Giving Report. Blackbaud Institute. February 2022. Available online:
[https://institute.blackbaud.com/wp-content/uploads/2022/03/BBI_CGR_2022.pdf (accessed on 24 April 2023).](https://institute.blackbaud.com/wp-content/uploads/2022/03/BBI_CGR_2022.pdf)
9. [The World’s Leading Cryptocurrency Platform. Available online: https://crypto.com/ (accessed on 2 May 2023).](https://crypto.com/)
10. Chinnasamy, P.; Ramani, D.R.; Ayyasamy, R.K.; Jebamani, B.J.A.; Dhanasekaran, S.; Praveena, V. Applications of Blockchain
Technology in Modern Education System—Systematic Review. In Proceedings of the 2023 International Conference on Computer
[Communication and Informatics (ICCCI), Coimbatore, India, 23–25 January 2023. [CrossRef]](https://doi.org/10.1109/ICCCI56745.2023.10128381)
-----
_Sustainability 2023, 15, 16205_ 19 of 19
11. [Michael, J.; Cohn, A.; Butcher, J.R. Blockchain technology. Journal 2018, 1, 35–45. Available online: https://www.steptoe.com/a/](https://www.steptoe.com/a/web/171269/3ZEKzc/lit-febmar18-feature-blockchain.pdf)
[web/171269/3ZEKzc/lit-febmar18-feature-blockchain.pdf (accessed on 20 April 2023).](https://www.steptoe.com/a/web/171269/3ZEKzc/lit-febmar18-feature-blockchain.pdf)
12. Haber, S.; Stornetta, W.S. How to Time-Stamp a Digital Document; Springer: Berlin/Heidelberg, Germany, 1991; pp. 437–455.
13. [Nakamoto, S. Bitcoin: A Peer-to-Peer Electronic Cash System. 2008. Available online: https://bitcoin.org/bitcoin.pdf (accessed](https://bitcoin.org/bitcoin.pdf)
on 20 April 2023).
14. Cernian, A.; Vlasceanu, E.; Tiganoaia, B.; Iftemi, A. Deploying blockchain technology for storing digital diplomas. In Proceedings
of the CSCS23: The 23th International Conference on Control Systems and Computer Science, Bucharest, Romania, 26–28 May
2021; IEEE Computer Society: New York, NY, USA, 2021; pp. 322–328.
15. [6 Biggest Blockchain Companies. Available online: https://www.investopedia.com/10-biggest-blockchain-companies-5213784](https://www.investopedia.com/10-biggest-blockchain-companies-5213784)
(accessed on 2 May 2023).
16. Chinnasamy, P.; Elumalai, A.; Ayyasamy, R.K.; Kavya, S.P.; Dhanasekaran, S.; Kiran, A. BookChain: A Secure Library Book
Storing and Sharing in Academic Institutions using Blockchain Technology. In Proceedings of the 2023 International Conference
[on Computer Communication and Informatics (ICCCI), Coimbatore, India, 23–25 January 2023. [CrossRef]](https://doi.org/10.1109/ICCCI56745.2023.10128571)
17. Dai, H.-N.; Wu, J.; Wang, H. Blockchain for Electronic Voting System—Review and Open Research Challenges. Sensors 2021, 21,
[5874. [CrossRef]](https://doi.org/10.3390/s21175874)
18. [Coin Market Cap. Available online: https://coinmarketcap.com/ (accessed on 2 May 2023).](https://coinmarketcap.com/)
19. [Crypto Market Sizing. January 2023. Available online: https://content-hub-static.crypto.com/wp-content/uploads/2023/01/](https://content-hub-static.crypto.com/wp-content/uploads/2023/01/Cryptodotcom_Crypto_Market_Sizing_Jan2023-1.pdf)
[Cryptodotcom_Crypto_Market_Sizing_Jan2023-1.pdf (accessed on 24 April 2023).](https://content-hub-static.crypto.com/wp-content/uploads/2023/01/Cryptodotcom_Crypto_Market_Sizing_Jan2023-1.pdf)
20. [Musharraf, M. What Is the Blockchain Trilemma? Available online: https://www.ledger.com/academy/what-is-the-blockchain-](https://www.ledger.com/academy/what-is-the-blockchain-trilemma)
[trilemma (accessed on 1 May 2023).](https://www.ledger.com/academy/what-is-the-blockchain-trilemma)
21. [Ethereum Gas Tracker. Available online: https://etherscan.io/gastracker (accessed on 30 April 2023).](https://etherscan.io/gastracker)
22. [What Is Polygon. Available online: https://wiki.polygon.technology/docs/home/polygon-basics/what-is-polygon (accessed](https://wiki.polygon.technology/docs/home/polygon-basics/what-is-polygon)
on 30 April 2023).
23. [Next-Generation Node.Js and TypeScript ORM. Available online: https://www.prisma.io/ (accessed on 20 April 2023).](https://www.prisma.io/)
24. [React Framework for the Web. Available online: https://nextjs.org/ (accessed on 20 April 2023).](https://nextjs.org/)
25. [The Official Website of IPFS. Available online: https://ipfs.tech/ (accessed on 29 August 2023).](https://ipfs.tech/)
26. [PostgreSQL: The World’s Most Advanced Open Source Relational Database. Available online: https://www.postgresql.org/](https://www.postgresql.org/)
(accessed on 20 April 2023).
27. [RainbowKit—The Best Way to Connect a Wallet. Available online: https://www.rainbowkit.com/ (accessed on 2 May 2023).](https://www.rainbowkit.com/)
28. [IBM. Available online: https://www.ibm.com/topics/smart-contracts (accessed on 1 August 2023).](https://www.ibm.com/topics/smart-contracts)
29. [Szabo, N. Formalizing and securing relationships on public networks. First Monday 1997, 2. [CrossRef]](https://doi.org/10.5210/fm.v2i9.548)
30. Buterin, V. A next-generation smart contract and decentralized application platform. White Pap. 2014, 3, 1–36.
31. [The Official Website of Ethereum. Available online: https://ethereum.org/ (accessed on 2 May 2023).](https://ethereum.org/)
32. [The Official Website of the MultiversX. Available online: https://multiversx.com/ (accessed on 2 May 2023).](https://multiversx.com/)
33. [Ethereum Whitepaper. Available online: https://ethereum.org/en/whitepaper/ (accessed on 2 May 2023).](https://ethereum.org/en/whitepaper/)
34. [Gramoli, V. What Is Blockchain Accountability, and Why It Matters? Available online: https://www.redbelly.network/blog/](https://www.redbelly.network/blog/what-is-blockchain-accountability-and-why-it-matters)
[what-is-blockchain-accountability-and-why-it-matters (accessed on 1 October 2023).](https://www.redbelly.network/blog/what-is-blockchain-accountability-and-why-it-matters)
35. [Pseudonymity and Anonymity: Be Untraceable in the Blockchain World. Available online: https://www.immunebytes.com/](https://www.immunebytes.com/blog/pseudonymity-and-anonymity-be-untraceable-in-the-blockchain-world/)
[blog/pseudonymity-and-anonymity-be-untraceable-in-the-blockchain-world/ (accessed on 1 October 2023).](https://www.immunebytes.com/blog/pseudonymity-and-anonymity-be-untraceable-in-the-blockchain-world/)
36. [Electric Capital Developer Report. Available online: https://github.com/electric-capital/developer-reports/blob/master/dev_](https://github.com/electric-capital/developer-reports/blob/master/dev_report_2022.pdf)
[report_2022.pdf (accessed on 1 May 2023).](https://github.com/electric-capital/developer-reports/blob/master/dev_report_2022.pdf)
37. [Bring Your Code, We’ll Handle the Rest. Available online: https://railway.app/ (accessed on 20 April 2023).](https://railway.app/)
38. [Population and the Sustainable Development Goals. Available online: https://populationmatters.org/un-sdgs/?gclid=Cj0](https://populationmatters.org/un-sdgs/?gclid=Cj0KCQjwsIejBhDOARIsANYqkD1vyLpVSL9Xf_bsHFT_MCGZxFW99a4Tm6byzQzf5mAaGWf3_HoVRCgaAg8jEALw_wcB)
[KCQjwsIejBhDOARIsANYqkD1vyLpVSL9Xf_bsHFT_MCGZxFW99a4Tm6byzQzf5mAaGWf3_HoVRCgaAg8jEALw_wcB (ac-](https://populationmatters.org/un-sdgs/?gclid=Cj0KCQjwsIejBhDOARIsANYqkD1vyLpVSL9Xf_bsHFT_MCGZxFW99a4Tm6byzQzf5mAaGWf3_HoVRCgaAg8jEALw_wcB)
cessed on 2 May 2023).
**Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual**
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3390/su152316205?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/su152316205, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "GOLD",
"url": "https://www.mdpi.com/2071-1050/15/23/16205/pdf?version=1700659792"
}
| 2,023
|
[
"JournalArticle"
] | true
| 2023-11-22T00:00:00
|
[
{
"paperId": "80260a4af770a022ccb4955b0999075dc908643a",
"title": "Blockchain for Electronic Voting System—Review and Open Research Challenges"
},
{
"paperId": "7ec86f51a93d646e5e4cf15d4ee48d9570021adc",
"title": "Integrating machine learning and blockchain to develop a system to veto the forgeries and provide efficient results in education sector"
},
{
"paperId": "1192f9947017ac30295fbc1621fccee7e7bbafd3",
"title": "Blockchain Technology"
},
{
"paperId": "5b4cf1e37954ccd1ca6b315986d45904f9d2f636",
"title": "Formalizing and Securing Relationships on Public Networks"
},
{
"paperId": "a85721c6d540e02fde873fbb07af3dab3b8c0452",
"title": "Business Information Systems Workshops: BIS 2019 International Workshops, Seville, Spain, June 26–28, 2019, Revised Papers"
},
{
"paperId": "0dbb8a54ca5066b82fa086bbf5db4c54b947719a",
"title": "A NEXT GENERATION SMART CONTRACT & DECENTRALIZED APPLICATION PLATFORM"
}
] | 12,895
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Engineering",
"source": "s2-fos-model"
},
{
"category": "Environmental Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/020293b1e58f0d7aefdf0d5c72e95fa7fc01fc2e
|
[
"Computer Science"
] | 0.879753
|
Distributed Strategy for Optimal Dispatch of Unbalanced Three-Phase Islanded Microgrids
|
020293b1e58f0d7aefdf0d5c72e95fa7fc01fc2e
|
IEEE Transactions on Smart Grid
|
[
{
"authorId": "30811424",
"name": "P. Vergara"
},
{
"authorId": "38506412",
"name": "J. Rey"
},
{
"authorId": "1814090",
"name": "H. Shaker"
},
{
"authorId": "2255353261",
"name": "J. Guerrero"
},
{
"authorId": "9114734",
"name": "B. Jørgensen"
},
{
"authorId": "134292628",
"name": "L. D. da Silva"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"IEEE Trans Smart Grid"
],
"alternate_urls": null,
"id": "1c2f3998-b5ca-48ca-9991-94b71c71ecb7",
"issn": "1949-3053",
"name": "IEEE Transactions on Smart Grid",
"type": "journal",
"url": "http://ieeexplore.ieee.org/servlet/opac?punumber=5165411"
}
|
This paper presents a distributed strategy for the optimal dispatch of islanded microgrids, modeled as unbalanced three-phase electrical distribution systems. To set the dispatch of the distributed generation (DG) units, an optimal generation problem is stated and solved distributively based on primal–dual constrained decomposition and a first-order consensus protocol, where units can communicate only with their neighbors. Thus, convergence is guaranteed under the common convexity assumptions. The islanded microgrid operates with the standard hierarchical control scheme, where two control modes are considered for the DG units: a voltage control mode, with an active droop control loop, and a power control mode, which allows setting the output power in advance. To assess the effectiveness and flexibility of the proposed approach, simulations were performed in a 25-bus unbalanced three-phase microgrid. According to the obtained results, the proposed strategy achieves a lower cost solution when compared with a centralized approach based on a static droop framework, with a considerable reduction on the communication system complexity. Additionally, it corrects the mismatch between generation and consumption even during the execution of the optimization process, responding to changes in the load consumption, renewable generation, and unexpected faults in units.
|
### University of Southern Denmark
Distributed Strategy for Optimal Dispatch of Unbalanced Three-Phase Islanded Microgrids
Vergara Barrios, Pedro Pablo ; Rey-López, Juan Manuel; Shaker, Hamid Reza; Guerrero, Josep M.; Jørgensen, Bo Nørregaard; da Silva, Luiz Carlos Pereira
Published in:
IEEE Transactions on Smart Grid
DOI:
[10.1109/TSG.2018.2820748](https://doi.org/10.1109/TSG.2018.2820748)
Publication date:
2019
Document version:
Accepted manuscript
Citation for pulished version (APA):
Vergara Barrios, P. P., Rey-López, J. M., Shaker, H. R., Guerrero, J. M., Jørgensen, B. N., & da Silva, L. C. P.
(2019). Distributed Strategy for Optimal Dispatch of Unbalanced Three-Phase Islanded Microgrids. IEEE
[Transactions on Smart Grid, 10(3), 3210-3225. https://doi.org/10.1109/TSG.2018.2820748](https://doi.org/10.1109/TSG.2018.2820748)
[Go to publication entry in University of Southern Denmark's Research Portal](https://portal.findresearcher.sdu.dk/en/publications/09899419-7445-46c7-8b29-56d3b83ca744)
Terms of use
This work is brought to you by the University of Southern Denmark.
Unless otherwise specified it has been shared according to the terms for self-archiving.
If no other license is stated, these terms apply:
- You may download this work for personal use only.
- You may not further distribute the material or use it for any profit-making activity or commercial gain
- You may freely distribute the URL identifying this open access version
If you believe that this document breaches copyright please contact us providing details and we will investigate your claim.
Please direct all enquiries to puresupport@bib.sdu.dk
-----
## Distributed Strategy for Optimal Dispatch of Unbalanced Three-Phase Islanded Microgrids
#### Pedro P. Vergara, Juan M. Rey, Hamid R. Shaker, Josep M. Guerrero, Fellow, IEEE, Bo N. Jørgensen, Luiz C. P. da Silva.
Abstract—This paper presents a distributed strategy for the
optimal dispatch of islanded microgrids, modeled as unbalanced
three-phase electrical distribution systems (EDS). To set the
dispatch of the distributed generation (DG) units, an optimal
generation problem is stated and solved distributively based
on primal-dual constrained decomposition and a first-order
consensus protocol, where units can communicate only with their
neighbors. Thus, convergence is guaranteed under the common
convexity assumptions. The islanded microgrid operates with the
standard hierarchical control scheme, where two control modes
are considered for the DG units: a voltage control mode (VCM),
with an active droop control loop, and a power control mode
(PCM), which allows setting the output power in advance. To
assess the effectiveness and flexibility of the proposed approach,
simulations were performed in a 25-bus unbalanced three-phase
microgrid. According to the obtained results, the proposed
strategy achieves a lower cost solution when compared with a
centralized approach based on a static droop framework, with a
considerable reduction on the communication system complexity.
Additionally, it corrects the mismatch between generation and
consumption even during the execution of the optimization process, responding to changes in the load consumption, renewable
generation and unexpected faults in units.
Index Terms—Consensus algorithm, distributed dispatch, optimal power flow, nonlinear programming, three-phase microgrid.
NOTATION
Sets:
F Set of phases {A, B, C}
G Set of DG units, G = G1 ∪G2
G1 Set of DG units operating in PCM, G1 ⊂G
G2 Set of DG units operating in VCM, G2 ⊂G
L Set of lines
N Set of nodes of the EDS
Om Set of operational constraints of the DG unit m
W Set of wind turbines (WTs) units
Indexes:
φ, ψ Phases φ ∈F and ψ ∈F
mn Line mn ∈L
This work was supported by the São Paulo Research Foundation (FAPESP).
Research Grants: 2015/09136-8 and 2016/04164-6.
Pedro P. Vergara and Luiz C. P. da Silva are with the Department of Systems
and Energy, UNICAMP, University of Campinas, 13083-852 Campinas, São
Paulo, Brazil (emails: {pedropa,lui}@fee.dsee.unicamp.br).
Juan M. Rey is with Escuela de Ingenierías Eléctrica, Electrónica y
de Telecomunicaciones (E3T), Universidad Industrial de Santander (UIS),
680002 Bucaramanga, Colombia (e-mail: juanmrey@uis.edu.co).
Josep M. Guerrero is with the Department of Energy Technology, Aalborg
University, Aalborg DK-9220, Denmark (e-mail: joz@et.aau.dk).
Pedro P. Vergara, Hamid R. Shaker and Bo N. Jørgensen are with the Center
for Energy Informatics, University of Southern Denmark, Odense DK-5230,
D k ( il { b h h b j}@ i d dk)
m, n Node m ∈N and n ∈N
Parameters:
αm Constant parameter associated to the DGs operation
cost
βm Linear parameter associated to the DGs operation
cost
∆tD Length for the discretization of the operational time
∆ω Angular frequency deviation
∆V Voltage magnitude deviation
Dm[P] Active power droop gain of DG units in VCM
Dm[Q] Reactive power droop gain of DG units
ε Parameter to control the converge of the active droop
protocol
εˆ Parameter to control the converge of the frequency
reference protocol
γm Quadratic parameter associated to the DGs operation
cost
λ Dual variable associated with the active power balance constraint
λm Local estimation of λ by the DG unit m
κ Parameter to control the converge of consensus protocol
Pm[W] Expected active power generation of the WTs
G
P m Maximum active generation limit of the DG units
P [G]m Minimum active generation limit of the DG units
Pm,φ[D] Active load consumption
G
Qm Maximum reactive generation limit of the DG units
Q[G] Minimum reactive generation limit of the DG units
m
Q[D]m,φ Reactive load consumption
V Maximum voltage magnitude
V Minimum voltage magnitude
V0 Nominal voltage magnitude
ω0 Nominal angular frequency
ωm Frequency reference of the DG units in VCM
Zmn,φ,ψ Line impedance
′ ′
Zmn,φ,ψ [Transformed line impedance, defined as][ Z]mn,φ,ψ [=]
Zmn,φ,ψ θψ − θφ
Continuous Variables:
Pm[G] Total active output power of the DG units
Pm[G][0] Total scheduled active of the DG units
Pmn,φ Active power flow in line mn at phase φ
Q[G] Total reactive generation power of the DG units
-----
Qmn,φ Reactive power flow in line mn at phase φ
Smn,φ Apparent power of line mn at phase φ
Smn,φ[L] Apparent power losses in line mn at phase φ
Vm,φ Voltage magnitude of nodes
ω Frequency of the system
Remark: Through the paper, it is assumed that the DG unit
m ∈G (and equivalently the WT m ∈W), it is connected to
the node m ∈N of the EDS.
I. INTRODUCTION
RADITIONALLY, the optimal dispatch of a microgrid
is performed in a centralized way, where a system op# T
erator gathers all the operational and technical information of
the distributed generation (DG) units, aiming to define the
generation dispatch that minimizes the overall cost [1]. Centralized optimal strategies for microgrids have been proposed
in [2]–[4]. Due to the way information is exchanged with
the central operator, these approaches require high bandwidth
communication infrastructures and high-levels of connectivity,
increasing the complexity of their implementation, specially
considering that the number of DG units can be large.
Moreover, these approaches do not show privacy preserving
characteristics, considering that units can belong to different
owners, which might not be interested in sharing private operational information. In contrast, distributed approaches offer
features that make them an interesting alternative, including
scalability, adaptability, privacy preserving and robustness,
allowing to respond to changes in the number of operating
units, unexpected increase in renewable generation or load
consumption, among others [5].
Recently, distributed approaches have drawn a lot of attention in the technical literature [6]. In general, two main
groups can be identified: (i) the approaches based on consensus
algorithms and (ii) the approaches based on local updating
rules. In all these, the objective is to define the generation
dispatch of each DG unit locally, limiting the amount of
information that is exchanged between the DG units.
For the first group, in [7] an iterative algorithm is developed
based on the incremental cost principle. This correspond to
the consensus variable. In these works, a DG leader unit
is required to balance the generation and load consumption.
In [8], two consensus algorithms are executed in parallel
to estimate locally the mismatch between generation and
load consumption. In [9], a term is added to the consensus
algorithm using only local information based on the nodal
power balance equation, which plays the role of a gradient.
In [10], [11] a modified consensus algorithm with finitetime convergence characteristics is presented, while in [12]
a distributed gradient-based algorithm is developed, taken
the derivative of the cost function of each DG unit as the
consensus variable.
In the second group, simple updating rules are developed.
These rules are continuously executed in an iterative procedure
aiming to define the operational schedule of each unit until a
convergence criterion is reached. For instance, in [13]–[17],
the iterative rule is defined to be proportional to the active
power mismatch between generation and load consumption
In addition to this, in [14], a proportional term based on the
marginal cost is also considered. Thus, units with low marginal
cost will increase their output power faster than high cost
generators. As the active power mismatch is a global variable,
and to be able to estimate it locally, in [13] local measurements
of frequency deviation are used, while in [14] and [17], a
complex communication procedure between neighbor units is
considered.
The main drawback of the above-discussed works [7]– [17],
is that they assume that all the generators and loads are
connected to one bus, ignoring the underlying operation of
the electrical distribution system (EDS). In general, in these
works it is assumed that a balancing mechanism operates i. e.,
a leader unit supply the required active power to correct the
mismatch between generation and load consumption; all this
while the optimization algorithm is executed. In an actual
operation, this is not a practical assumption since the mismatch
between the generation and load consumption is corrected in
a faster speed of response (normally, in the order of seconds)
by the lower level controllers [18]. Moreover, as in islanded
operation the DG units are responsible for providing the
frequency and voltage magnitude references for the system,
if the optimal dispatch does not consider the control operation
of the DG units, the system might operate with a higher
frequency or voltage deviation, and consequently, the optimal
schedule might not be technically feasible. In this regard, in
[5], [19], a distributed approach including the operation of the
EDS was developed. However, a centralized communication
infrastructure is still required, while the unbalanced operation
of the microgrid is not taken into account.
Considering this, a distributed strategy for the optimal
dispatch of islanded microgrids is presented in this paper.
The microgrid is modeled as an unbalanced three-phase EDS,
operating within a hierarchical control scheme. To define the
active dispatch of the DG units, an optimal generation problem
is stated and solved distributively using a first-order consensus protocol, where units can communicate only with their
neighbors. This strategy is based on primal-dual constrained
decomposition theory, in order to distribute the problem among
the units and take into account locally their technical operational requirements. Thus, convergence is guaranteed under
the common convexity assumptions. Additionally, two control
modes are considered for the DG units: a voltage control mode
(VCM), with an active droop control loop, and a power control
mode (PCM), which allows setting the output power of the
unit in advance. To assess the effectiveness and flexibility of
the proposed approach, simulations were performed in a 25bus microgrid for different case of studies. According to the
results, the proposed strategy achieves a lower cost solution
when compared with a centralized approach, with a considerable reduction on the communication system complexity,
responding to changes in the load consumption, renewable
generation, and unexpected faults in units.
Among all the features previously discussed of the proposed
distributed strategy, the main contributions of this paper can
be summarized as follows:
- The proposed strategy considers the control modes of
the DG units (VCM and PCM) in the optimization
-----
approach. In this context, as the units in VCM operate
with a droop control loop, these are responsible for
correcting the active power mismatch between generation
and consumption after any load or renewable generation
increase/decrease, and more importantly, during the execution of the optimization algorithm.
- The proposed strategy is considered to operate within the
standard hierarchical control framework for microgrids.
Thus, it is ensured that the dynamics of the optimization
algorithm and the primary control layer (implemented
with droop control) are decoupled, which helps to maintain the stability of the system. Moreover, it considers
a correction protocol, in order to operate in steady-state
with a lower frequency deviation.
II. CONTROL AND OPERATION OF MICROGRIDS
Microgrids can operate in two modes: grid-connected or
islanded mode. In grid-connected mode, the frequency and
voltage magnitude references are provided by the main grid,
while in islanded mode, these must be provided by the DG
units [20]. Since the operation of microgrids deals with issues
from different technical areas, time scales and infrastructure
levels, the hierarchical control scheme has been widely accepted as the standard solution [18].
In general, the hierarchical control scheme comprises three
different and well defined levels: (i) a primary level, the
fastest level, responsible for the local control of the DG units,
generally based on droop control, which does not require
communications; (ii) a secondary level, which deals with the
deviation at steady-state conditions of the frequency and the
voltage magnitude due to the operation of the primary level;
and (iii) a tertiary level, the slowest level, responsible for
the economical operation of the system, implemented through
a dispatch algorithm, generally based on the solution of an
optimization problem.
A. Primary and Secondary Control Level
Regarding the primary control level, DG units can operate
in two different control modes: power control mode (PCM)
and voltage control mode (VCM) [21]. In islanded operation,
at least one unit is required operating in VCM to define the
frequency and voltage magnitude reference of the EDS [18].
Hence, if the unit operates in PCM, the output power can be
set at the schedule value defined by the tertiary control level,
i.e.,
Pm[G] [=][ P][ G]m [0][,] ∀m ∈G1. (1)
In this case, the output power is independent of the state of
the EDS. Different from this, if the unit operate in VCM, its
output power cannot be set in advance, since this unit operates
with a droop control loop. Therefore, all units in VCM share
the remaining active power mismatch between generation and
consumption, in inverse proportion to their active droop gain
(Dm[P] [). The droop operation of a unit in VCM mode can be]
represented using the expression,
ω = ωm − Dm[P] [P][ G]m [,] ∀m ∈G2, (2)
where Pm[G] [is the total active output power of the unit. A]
schematic representation of both control modes is shown in
PCM
ω0
ωm
VCM
ω
Dm[P]
Pm[G] [=][ P][ G]m [0] Pm[G] Pm[G][0]
Figure 1. Control modes of the DG units: VCM and PCM. The line in
the VCM indicates the direction of variation in Pm[G] [when][ D]m[P] [is decreased.]
Additionally, ωm can be modified in order to reduce the frequency deviation.
Fig. 1. The droop gain Dm[P] [reflects the slope of the][ ω][ −] [P]
curve. Thus, the total output power of units in VCM (i.e., Pm[G][)]
can be set to their scheduled value (Pm[G][0][), tuning][ D]m[P] [; all this]
in order to minimize the overall generation cost.
The main difference between these control modes is related
to the operation of the control loops and how these set the
output power in steady-state conditions. Thus, they are essentially independent, which means that each DG unit can decide
its operation mode (see Sec. IV-C for a further discussion).
Implementation and stability issues related to the transition
between both control modes are discussed in detail in [21].
As for the reactive power Q[G]m[, in both control modes all]
the DG units share the reactive power consumption defining
their output voltage magnitude using the expression,
Vm,φ = V0 − Dm[Q] [Q][G]m[,] ∀m ∈G. (3)
This droop control is based on the assumption that the output impedance of the DG unit is inductive, which is valid for
synchronous-based and the majority of inverted-based units,
coupled to the EDS with an inductor filter. Nevertheless, in
case of output non-inductive impedance, control strategies that
aims to decouple the active and reactive power regulation can
be implemented, e.g. virtual output impedance strategies [22].
Regarding the secondary control level, its main function is
related to the definition of the frequency and voltage reference
i. e., ωm and V0. This is done in order to reduce the frequency
and voltage deviation in steady-state conditions [23], and as
shown in Fig. 1. The operation of the secondary control level
can be seen as a correction process, which operates with a
lower speed of response than the primary control, in order to
maintain their dynamics decoupled.
B. Tertiary Control Level
Regarding the tertiary level, to define the operational schedule of the DG units, an optimization problem is formulated
and solved. The formulation of this problem must account
for all the operational constraints of units, while the total
load consumption is supplied with minimum generation cost.
In general, this problem is known as the optimal generation
problem, which can be stated using the formulation given by
(4)–(6), for a microgrid comprising DG units, WT units and
loads.
G0min �� fm(Pm[G][0] [)]� (4)
-----
subject to,
� Pm[G][0] [+] � Pm[W] [=] �
m∈G m∈W m∈N
� Pm,φ[D] (5)
φ∈F
Thus, after replacing (9) into (8), and some re-arrange, the
dynamics of the consensus variable of each DG unit can be
updated using (10), which can be seen as the weighted average
of its current state and the current state of its neighbors units.
xm(k + 1) = � cmnxn(k). (10)
n∈Nm
To reach consensus under dynamics in (10), the consensus
matrix C = [cmn] can be defined as [14],
P [G]m [≤] [P][ G]m [0] [≤] [P] Gm ∀m ∈G. (6)
In the above formulation, the objective function in (4)
aims to minimize the overall generation cost, where fm(Pm[G][0][)]
models the generation cost of each DG unit, which can be
approximated with a quadratic function [1], such as,
fm(Pm[G][0] [) =][ γ][m][(][P][ G]m [0] [)][2][ +][ β][m][P][ G]m [0] [+][ α][m][,] ∀m ∈G, (7)
where usually γm holds a positive value, which yields convexity of the generation cost function.
For the operational constraints, the active power balance
in the EDS (neglecting power losses) is modeled in (5), as
a function of the three-phase output power of the DG units
(Pm[G][), WT units (][P][ W]m [) and the load consumption (][P]m,φ[ D] [); while]
constraints in (6) models the generation limits of the DG units.
III. DISTRIBUTED OPTIMAL STRATEGY
In this section, a description of the communication topology
of the DG units seen as a graph is discussed. Additionally, the
consensus algorithm used is introduced. Then, the distributed
optimal strategy is presented. Finally, an overview of the
proposed approach is discussed.
A. First-Order Consensus Algorithm
Let the graph G = (V, E, A) describes the communication
topology of the DG units. For this graph, the set of nodes V
represents the set of DG units, while the set of edges E ⊂
V × V represents the set of communication links between the
DG units. Considering this, an adjacency matrix A = [amn],
with non-negative adjacency elements amn, can be defined
for the microgrid. The adjacency elements associated with the
communication links (or edges of the graph) are positive, i.e.,
amn = 1 ∀(m, n) ∈E, and otherwise, amn = 0. Additionally,
Nm is defined as the set of neighbors of the DG unit m, i.e.,
the set of DG units that can exchange information with unit m.
Finally, the cardinality (i.e., the size) of the set Nm is deifned
as dm.
Define for the DG unit m a generic variable xm ∈ R, and
named it as the consensus variable. The consensus variable
represents the quantity in which all the DG units want to agree
(in Sec. III-B, the consensus variable defined corresponds to
λ, i.e., the incremental cost variable). Thus, it can be said that
the DG unit m and n agree if and only if xm = xn. Moreover,
it can be said that all the DG units have reached consensus if
and only if xm = xn ∀m, n ∈V.
The dynamics of the consensus variable xm for each DG
unit can be described by the discrete-time model in (8), where
k is an iteration counter.
xm(k + 1) = xm(k) + um(k). (8)
It can be shown that under the protocol in (9), all the DG
units reach consensus when k →∞ [24], where C = [cmn]
is known as the consensus matrix.
um(k) = � cmn(xn(k) − xm(k)). (9)
1Notice that Om as defined in (6), is closed due that P Gm [take values within]
the range P [G]m [≤] [P][ G]m [≤] [P] Gm[. Additionally, it is convex since it is described]
b t f li ti
cmn =
�1/(dm + 1) if n ∈Nm ∪{m},
(11)
0 if n ̸∈Nm.
Such definition leads to a row-stochastic (i.e., row sum of
1), as required according [25]. It is important to highlight
that the notion of neighborhood used here is related to the
existence of a communication link between the DG units.
The protocol in (10) is known as the first-order consensus
algorithm, and its speed of convergence depends on the
level of connectivity of the communication topology of the
DG units. Nevertheless, convergence is guarantee as long as
the communication topology fulfills the design requirements
discussed in Sec. III-E.
B. Distributed Optimal Dispatch Strategy
The optimal generation problem, as formulated in Section II-B, can be solved using a distributed optimization
approach taking advantage of its structure. In this, the only
constraint that couple the problem among all the units is the
active power balance in (5). Moreover, the set of operational
constraints of the DG units, given by (6), defines a closed and
convex set[1] Om, ∀m ∈G, in such a way that Om ∩On =
∅, ∀m, n ∈G; or in other words, the operational constraints of
the DG unit m are independent of those of unit n. In this case,
only generation limit constraints are considered. However,
other operational constraints such as prohibited operational
zones can be added to the set Om without modifying the
proposed optimization strategy.
Based on this, a distributed strategy can be developed.
Firstly, define the Lagrangian function L(Pm[G][0] [, λ][)][ as]
L(Pm[G][0] [, λ][) =] � fm(Pm[G][0][)]
m∈G
+ λ [�] � Pm,φ[D] [−] � Pm[G][0] [−] � Pm[W]
m∈N φ∈F m∈G m∈W
, (12)
where λ corresponds to the dual variable associated to constraint in (5). The optimal solution, which defines the active
power dispatch of each DG unit, must meet the first optimality
condition, which can be expressed as
∂L(·) m [)]
= [df][m][(][P][ G][0] − λ = 0, ∀m ∈G (13)
∂Pm[G][0] dPm[G][0]
or equivalently [26],
min
Pm[G][0] [∈O]m
�
fm(Pm[G][0][)][ −] [λP][ G]m [0]
�
. (14)
-----
λn(k)
λn(k)
|Col1|Stage I k = k + 1 Consensus Algorithm Ag ent m Agent n Agent p|
|---|---|
Pm[G][, ω] Define Pm[G][0][(][k][)][ solving (14)]2 DefineAgent Pn[G][0] n,[(][k][)] ∀[ solving (14)]n ∈G1 Vm,φ
Vm,φ Define Dm[P] [(][k][)][ using (16)]
(a) DG unit operating with VCM control.
(b) DG unit operating with PCM control.
Figure 2. Structure of a DG unit seen as an agent. The black dashed lines
represent exchange of information between different agents, while the red
dashed lines represent local measurements.
Therefore, to define its scheduled active power, i.e., Pm[G][0][,]
each DG unit m ∈G solve (14) locally. The only global
information required to solve (14) corresponds to λ, which
can be estimated locally by each unit. To do this, variable λm
is introduced and defined as the local estimation of the dual
variable by unit m. From the economic operation of power
systems, λm can be seen as the incremental cost of the DG
units. Hence, the minimum cost dispatch is reached when all
units have the same incremental cost value [7]. This condition
is equivalent to state that λm = λn, ∀m, n ∈G, which suggests
that variable λm can be defined as the consensus variable.
Therefore, to estimate λm, this paper proposes that each unit
execute locally the iterative consensus protocol given by,
|Col1|DP m(k), ωm(k) P G0 m (k)|Col3|
|---|---|---|
||Microgrid||
||||
||||
Figure 3. Flowchart of the proposed distributed dispatch strategy composed
of Stage I and Stage II.
Dm[P] [(][k][ + 1) =][ D]m[P] [(][k][) +][ ε][(][P][ G]m [−] [P][ G]m [0][)][,][ ∀][m][ ∈G][2][,] (16)
λm(k+1) = � cmnλn(k)+κ(Pm[G][−][P][ G]m [0][)][,][ ∀][m][ ∈G][,][ (15)]
n∈Nm
where κ is a parameter that controls the convergence of the
protocol. Notice that for units operating in PCM, the second
term in (15) is reduced to zero, due to (1). The rationale of
protocol in (15) can be understood if each DG unit is modeled
as an independent agent, with functionalities such as acquiring
local measurements, exchange information with its neighboring agents and define its active power schedule independently.
A representation of the structure and information flow of a DG
unit as an agent for both control modes, is shown in Fig. 2.
Notice that each DG unit has two active power variables:
Pm[G][, which stands for the active output power, and][ P][ G]m [0][,]
which stands for the scheduled active output power. These
two variables must not be confused: Pm[G][0] is obtained for all
the DG units solving the problem in (14), and corresponds
to the output power that minimizes the total generation cost,
while Pm[G] [is the actual power that the DG unit is supplying]
to the microgrid. Thus, the active output power of a DG unit
operating in PCM, can be directly set considering (1). On
the other hand, as the output power of units operating in
VCM cannot be directly set (i.e., Pm[G] [cannot take directly the]
value given by Pm[G][0][), the active droop gain (][D]m[P] [) is modified]
iteratively using the protocol
where ε controls the convergence. This protocol guarantees
that the active output power of the DG units in VCM (Pm[G][),]
defined through the droop expression in (2), equals the dispatched power (Pm[G][0][), defined cooperatively by all the DG]
units.
Nevertheless, when Dm[P] [is modified, the system might reach]
a steady-state with a frequency different from the nominal
frequency value, given by ω0. To reduce this deviation, each
unit updates its frequency reference (ωm) in (2), using the
following protocol,
ωm(k + 1) = ωm(k) + ˆε(ω0 − ω), ∀m ∈G2. (17)
This protocol guarantees that the system will operate with
a lower frequency deviation in steady-state. A theoretical
convergence analysis of protocols (15), (16) and (17), is
presented in the Appendix.
As for the reactive power, and as shown in Fig. 2, information about the voltage at the node of connection is required for
all units in order to define their reactive output power using
(3). In this case, reactive power has not been considered in the
optimization strategy, as it does not incur in any cost [27].
C. Overview of the Distributed Strategy
Fig. 3 shows the flowchart of the proposed distributed
dispatch strategy, composed of two stages: one to run the
consensus protocol (Stage I) and other to run the optimization
algorithm (Stage II). At each iteration k, each stage can be
explained as follows:
Stage I: All DG units execute locally their consensus
protocols, as explained in Section III-B, in order to calculate
λm(k), i.e., the local estimation of variable λ. Here, it is
assumed that the exchange of information between the DG
units is done synchronously, i.e., the time-delay of the communication process is not considered. Additionally, recall that
the exchange of information between units depends on the
-----
communication topology, as explained in Sec. III-A and shown
in Fig. 3.
Stage II: Here, each DG unit solves the problem in (14)
independently, in order to define Pm[G][0][(][k][)][, i.e., the optimal]
dispatch that minimizes the overall generation cost for the
current λm(k). To be able to solve the sub-problem in (14),
each DG unit requires: λm(k) (previously calculated in the
Stage I), and its own operational data i.e., parameters of the
cost function (αm, βm, γm), maximum and minimum active
generation capacity (P m, P m). Additionally, the units in VCM
update their parameters Dm[P] [(][k][)][ and][ ω][m][(][k][)][, using (16) and]
(17), respectively.
It is important to highlight that, as the consensus protocol
of units operating in VCM considers the current output power
though the droop control loop, the proposed strategy takes
into account the active power losses, even if these were not
considered in the formulation stated in Sec. II-B.
Notice that, in order to update protocols (16) and (17) in
Stage II, operational information of the EDS, such as the
frequency (ω) and the current output power of the DG units
in VCM (Pm[G][), is required (see Fig. 2). This information can]
be obtained by the DG unit through local measurements, as
explained in [12].
D. Operation within the Hierarchical Control Framework
To better understand the operation of the proposed strategy
within the hierarchical control scheme, consider the illustrative
example of the dynamics of the microgrid shown in Fig. 4.
The hierarchical control is activated at t1. Before this time it
is considered that the microgrid is in steady-state operation,
and the DG unit m has output power Pm[G][. After][ t][1][, the]
active droop gain Dm[P] [is modified, forcing the unit in VCM]
to supply Pm[G][0] (previously defined as a predetermined value
of the tertiary control). This process is performed by the
primary control level, which operates with the fastest time
of response, denominated as TP . Due to the operation of the
primary control, the frequency is modified, as shown in Fig. 4b.
This frequency deviation is corrected by the secondary level,
which acts in a slower time scale compared with the primary
level. For this reason, its time of response TS is greater than
TP, which helps to decouple its dynamics. Finally, the tertiary
level operates defining the new scheduled active power value
Pm[G][0] with the slowest time scale TT . Thus, the update of the
dispatch variables is done once the frequency recovery process
is completed.
In this context, considering the operation of the proposed
distributed strategy, Stage I and II perform the functions of the
tertiary and secondary control, defining the optimal schedule
of all the DG units (i.e., Pm[G][0][), as well as the parameters of]
the droop control to reduce the frequency deviation in steadystate conditions (i.e., Dm[P] [,][ ω][m][). Considering this, and aiming]
to maintain decoupled the dynamics of Stage I and II and the
primary control, the response time of the proposed strategy
(named as TD in Fig. 4a) should be selected to have a value
greater than TS, but lower than TT, i.e., TS ≤TD ≤TT .
The selected value will depend on the speed of response
desired for the system. Usually, TS takes values near to
30 s or lower [23] [28] [29] while T can take values
|Col1|Col2|Col3|TD|Col5|Col6|
|---|---|---|---|---|---|
||P mG · · ·||D|P mG0||
||||TT|||
||||TS|||
|||TP||||
|||||||
t1 t2 t3 t4
(a) Active output power of the DG unit m operating in VCM.
ω0 ω3
ω2
ω1
t3
t1
t2
t4
|(a) Active outpu|ut power of|f the DG unit m op|perating in VCM|M.|
|---|---|---|---|---|
||||||
|ω1 · · ·|ω2|ω3|||
|t1 t2 t3 t4|||||
(b) Angular frequency of the microgrid.
Figure 4. Illustrative example of the dynamics of the primary, secondary and
tertiary control level. The convergence time of the optimization algorithm in
the proposed strategy is limited by the time length TD.
between 5 to 15 minutes [18]. Notice also that TD limits
the maximum processing time of the proposed strategy in a
practical implementation.
E. Design Considerations
The selection of the parameters κ, ε and ˆε in protocols,
(15), (16) and (17), can affect the speed of convergence of
the distributed strategy. Higher values for these parameters
can lead to a faster convergence. However, a trade-off exists
so that if they are set too large, an oscillatory behavior can
be observed due to the excessively fast update of λm(k) in
(15). Additionally, the number of DG units also affects the
choice of these parameters. It is possible to observe that, after
a load or WT generation increase/decrease, the higher number
of DG units in VCM, the lower the active power that each
DG unit supply to reduce the mismatch between generation
and load consumption; which means that the output power of
the DG unit (Pm[G][) might not be too far from its new schedule]
value (Pm[G][0][). Therefore, the distributed approach can converge]
faster. This analysis suggests that the control parameters κ, ε
and ˆε should be set inversely proportional to the number of
DG units, N . Thus, the following heuristic rules can be used,
κ = 100/N, ε = 1/100N, εˆ = 1/N. (18)
Although these rules do not estimate an optimal value for the
convergence parameters κ, ǫ and ˆǫ, these have shown a good
performance in different scenarios, as described in Sec. IV-F.
The design of the communication topology can follow
multiple criteria in order to define how the DG units exchange
information, including: geographical localization, closeness in
the electrical network, tolerance to links failure, among others [7], [12]. Nevertheless, the convergence of the distributed
strategy is guarantee as long as the graph G, that models
the communication topology, meets, at least, the next design
criteria:
-----
Table I
DG UNITS INFORMATION
m γm[$/kW[2]] βm[$/kW] αm[$] P [G]m[[kW]] P [G]m[[kW]] Q[G]m[[kvar]] Q[G]m[[kvar]]
8 0.444 0.111 0.0 90 900 -180 540
13 0.264 0.067 0.0 150 1500 -300 900
19 0.400 0.100 0.0 120 1200 -120 720
22 0.500 0.125 0.0 80 800 -160 480
25 0.250 0.063 0.0 160 1600 -240 960
- There exists a path that links any DG unit m to any DG
unit n. This condition guarantee that all the DG units are
connected.
- The consensus matrix C = [cmn] is balanced, which can
be obtained with bidirectional communications links.
The use of the proposed strategy, as based on a distributed
approach, reduces the dependence on the communication system complexity when compared with a centralized approach.
In fact, in a centralized approach, a high level of connectivity
is required due to the central operator, reducing its robustness
and reliability since a single point of failure exists. Moreover,
in case of a link failure, the corresponding DG unit will be
isolated, preventing its control. In this context, in the proposed
strategy the communication topology can be designed to be
robust in case of any link failure, following, for instance, the
design rule presented in [12]. Additionally, as there is no leader
unit, the system will continue its operation in case of a unit
fault, as will be shown in Sec. IV-D.
IV. SIMULATION RESULTS AND DISCUSSION
The proposed strategy was tested in the unbalanced 25bus microgrid shown in Fig. 5a. In total, five DG units and
two WT units are considered. For comparison purposes, the
two communication topology shown in Fig. 5 are considered.
These are selected as they correspond to the more common
topologies used in literature to test distributed algorithms [7].
Nevertheless, other topologies can be tested as well, as long
as they meet the design requirements described in Sec. III-E.
The information of the DG units is shown in Table I. The
reactive droop gain of all units were defined as in (19), while
at initialization (i.e., at k = 0), the active droop gain of the
units operating in VCM were defined as in (20). This definition
allows the DG units operating in VCM achieve proper active
power sharing according to their power ratings [30].
G
Dm[Q] [= (][V][ −] [V][ )][/][2][Q]m[,] ∀m ∈G (19)
G
Dm[P] [(][k][) = ∆][ω/P] m[,] ∀m ∈G2|k = 0. (20)
Additionally, V, V and ∆ω/2π were set to 0.94 p.u.,
1.05 p.u. and 0.1 Hz, respectively. The parameter κ was set in
20, while ε and ˆε were defined to be 0.02 and 0.2, respectively,
using (18). The active power in all the protocols is in p.u.,
using 1000 kW as the nominal base. Initially, the WT units
are not operating. The units G13, G19, G25 operate in VCM,
while units G8, G22 operate in PCM. The distributed model
was implemented in AMPL and solved with IPOPT [31], using
a computer with an Intel i7-4749 processor and 16 GB RAM.
As explained in Sec. III-C, at each iteration k, information
related to the the frequency (ω) and the current output power
21 22
G22
19 20 23 24 25
G19 G25
18
1 2 3 4 5
W5
13
16 G13
8 6 7 9 10 11 12
G8
15 14 17 W17
(a) 25-bus three-phase microgrid.
G22 G22
G25 G25
G19 G19
G13 G13
G8 G8
(b) Ring topology. (c) Tree topology.
Figure 5. Microgrid and the communication topology presented as a graph.
The data of the system and load demand of each node can be found in [32].
of the DG units in VCM (Pm[G][) is required to update protocols]
(16) and (17). To simulate this measurement process, in this
paper, an optimal power flow (OPF) formulation is used, as
explained next in Sec. IV-A.
A. Simulation of the Measurement Process
In order to simulate the physical response of the microgrid
during the optimization process at each iteration k, an optimal
power flow formulation is solved. In practical cases, the
response in the microgrid variables is measured using sensors
locally implemented in each DG unit, thus, this formulation
is not necessary.
The use of this formulation is based on two facts: first,
the dynamics of the primary control level is faster than the
dynamics of the higher control levels (or equivalently, TD ≫
TP in Fig. 4); and second, to define Dm[P] [in (16),][ P][ G]m [is obtained]
filtering the measured instantaneous active power with a lowpass filter [23]. This means that, at the end of the time length
TD, the system has reached a steady-state condition, which
can be estimated using a power flow model [30].
The unbalanced three-phase islanded OPF formulation is
given by the non-linear optimization problem in (21)–(33),
which is based on the work presented in [33]. Notice that
in this formulation there are not control variables, thus, its
solution is equivalent to the one provided by a three-phase
power flow tool, such as OpenDSS or GridLabD.
� R{Smn,φ[L] [}] (21)
φ∈F [,]
min
�
mn∈L
where Smn,φ[L] [is defined as,]
Smn,φ. (22)
Smn,φ[L] [=]
′
mn,φ,ψ[S]mn,ψ[∗]
[�] [Z] Vm φVm ψ
-----
Subject to:
� Plm,φ − �
lm∈L mn∈L
�Pmn,φ + R{Smn,φ[L] [}]� + � Pm[G][0][(][k][)][/][3]
m∈G1
+ � Pm,φ[G] [=][ P][ D]m,φ [−] � Pm[W] [/][3]
m∈G2 m∈W
∀m ∈N, ∀φ ∈F (23)
� Qlm,φ − �
lm∈L mn∈L
�Qmn,φ + I{Smn,φ[L] [}]� + � Q[G]m,φ,t
m∈G
= Q[D]m,φ ∀m ∈N, ∀φ ∈F (24)
Vm,φ[2] [−] [V][ 2]n,φ [= 2] �
ψ∈F
�R{Zmn,φ,ψ′ [}][P][mn,ψ] [+]
G1 G2 G3 P [D]
Figure 6. Microgrid used for the simple numerical example. DG units G1
and G2 operate in VCM, while the DG unit G3 operate in PCM. The total
load demand is defined to be 400 kW.
Table II
DG UNITS INFORMATION FOR THE NUMERICAL EXAMPLE
m γm[$/kW[2]] βm[$/kW] αm[$] P [G]m[[kW]] P [G]m[[kW]]
1 0.20 0.25 0.0 40 400
2 0.50 0.15 0.0 20 200
3 0.30 0.22 0.0 25 250
and VCM). The active droop expression of units in VCM is
considered using (28), while the reactive droop expression is
considered in (29) for all units (in PCM and VCM). Constraint
in (30) models the electromotive force of synchronous-based
DG units, which is represented by the balanced voltages
magnitudes at their internal nodes [2]. Constraint in (31)
enforces the maximum and minimum limits for the voltage
magnitudes. Finally, the total active generation power limits
are defined by (32) for units in VCM, while the total reactive
generation limits are defined by (33) for all units. For a more
detailed discussion of this power flow formulation and the DG
units modeling, see [33].
B. Numerical Example
In order to illustrate the proposed strategy, a simple numerical case is presented in this section, applying the distributed
approach, which is summarized in (40) to (45) in the Appendix. This simple case is composed of three DG units, the
DG unit G1 and G2 operate in VCM, while the DG unit G3
operates in PCM. The total load consumption is defined as 400
kW, while the system is considered to be lossless, as shown
in Fig. 6. The DG units parameters are shown in Table II,
meanwhile it is considered that all the DG units exchange
information with the remaining, defining the consensus matrix
to be equal to C = [cmn] = 1/3, ∀(n, m) ∈V, as explained
in Sec. III-A.
The first iterations for this simple case are shown in Table III. For k = 0, all the DG units define the local estimation
of the incremental cost as λ1(0) = λ2(0) = λ3(0) = 0. To
define Pm[G][0][(0)][, each DG unit solves the problem in (14). This]
can be solved analytically, giving the next expression,
P [G]m if (λm(k) − βm)/(2γm) < P [G]m[,]
Pm[G][0][(][k][) =] P Gm if (λm(k) − βm)/(2γm) > P Gm[,]
λm(2kγ) −m βm Otherwise.
(34)
Thus, using the current local estimation of λm(k) and (34),
each DG unit defines its optimal active power as P1[G][0](0) = 40,
P2[G][0](0) = 20 and P3[G][0](0) = 25, which corresponds to their
minimum active power generation. Notice in Table III, that as
the unit G3 operates in PCM, its output power can be defined
to be its scheduled value i e P [G][0](0) P [G](0) 25 For the
2
� Zmn,φ,ψ′∗ [S][mn,φ]
ψ∈F
������
∀mn ∈L, ∀φ ∈F (25)
I{Zmn,φ,ψ′ [}][Q][mn,ψ]� − Vm,φ1[2]
������
Pm[G] [=] � Pm,ψ[G] ∀m ∈G2 (26)
ψ∈F
Q[G]m [=] � Q[G]m,ψ ∀m ∈G (27)
ψ∈F
ω = ωm(k) − Dm[P] [(][k][)][P][ G]m ∀m ∈G2 (28)
Vm,φ = V0 − Dm[Q] [Q][G]m ∀m ∈G (29)
Vm,φ = Vm,ψ ∀m ∈G2, ∀φ, ψ ∈F (30)
V ≤ Vm,φ ≤ V ∀m ∈N, ∀φ ∈F (31)
P [G]m [≤] [P][ G]m [≤] [P] Gm ∀m ∈G2 (32)
G
Q[G]m [≤] [Q]m[G] [≤] [Q]m ∀m ∈G (33)
In the above formulation, the objective function in (21) aims
to minimize the total active power losses of the EDS. This
ensure that the solution of the power flow formulation matches
the steady-state operation of the unbalanced microgrid, considering the active droop loop of the DG units operating in
VCM and the reactive droop loop of all units [30], [33].
The EDS is modeled by (23)–(25), derived as a function
of the active and reactive power flow through lines, i.e.,
Smn,φ = Pmn,φ + jQmn,φ. The line impedance is considered to
be constant. Additionally, a transformation is introduced and
′
defined as Zmn,φ,ψ [=][ Z][mn,φ,ψ] [θ][ψ] [−] [θ][φ][. Notice that due to this]
′
definition Zmn,φ,ψ [is not symmetric as][ Z][mn,φ,ψ][. Constraints]
(23) and (24) model the active and reactive power balance,
respectively, considering the output power of the DG units
operating in VCM and PCM, and the balanced power of the
WTs.
In (23), the three-phase output power of DG units in PCM
is modeled as a balanced constant power injection, which
value was previously defined in the Stage II (i.e. Pm[G][0][(][k][)][, see]
Fig. 3); while the three-phase output power of units operating
in VCM is considered as a power flow variable. Equation (25)
models the voltage magnitude drop in the lines. In (26), the
three-phase output power of units in VCM is modeled as a
function of their output power per phase, while (27) models
the three phase reactive output power of all units (in PCM
-----
units in VCM, the DG current output power can be obtained
using power flow models or analytic solutions, if the size of
the problem allows it. For real operation, the estimation of the
current output power is based on local real-time measurements.
In this case, the current output power of the units in VCM can
be estimated as[2],
3 [(][k][))][D]2[P] [(][k][)]
P1[G][(][k][) = (][P][ D][ −] [P][ G], (35)
D1[P] [(][k][) +][ D]2[P] [(][k][)]
3 [(][k][))][D]1[P] [(][k][)]
P2[G][(][k][) = (][P][ D][ −] [P][ G] . (36)
D1[P] [(][k][) +][ D]2[P] [(][k][)]
10
Tree Topology
Ring Topology
9
8
7
0 100 200 300 400 500 600 700 800 900
Figure 7. Total generation cost of the DG units considering the ring and tree
topology shown in Fig. 5. The x-axis represents k, i.e., the iteration counter.
|Col1|Tree T Ring T|opology opology|Col4|Col5|Col6|Col7|Col8|Col9|Col10|Col11|
|---|---|---|---|---|---|---|---|---|---|---|
||||||||||||
||||||||||||
||||||||||||
Using these expressions, for k = 0, the current output power
of the units in VCM are given by P1[G][(0) = 250][ and][ P][ G]2 [(0) =]
125. Additionally, D1[P] [(0)][ and][ D]2[P] [(0)][, are defined as using]
the standard expression, which is given in (20), defining the
values[3] of D1[P] [(0) = 1][.][57][ and][ D]2[P] [(0) = 3][.][14][.]
For k = 1, first all the DG units update their local estimation
of the incremental cost, using the expression in (15). For the
unit G1, (15) gives[4],
λ1(1) = 1/3 λ1(0) + 1/3 λ2(0) + 1/3 λ3(0)
+ 33.33(P1[G][(0)][ −] [P][ G]1 [0](0)) = 7.00. (37)
The factor 33.33 represents κ, calculated as explained in
Sec. III-E. For the remaining DG units, the same procedure
is done, giving: λ2(1) = 3.5 and λ3(1) = 0. Once the local
estimation of the incremental cost are obtained, each DG unit
uses the expression in (34) to obtain their optimal schedule,
which gives: P1[G][0](1) = 40, P2[G][0](1) = 20 and P3[G][0](1) = 25.
Then, the droop gains of units in VCM are updated, using the
expression in (16), giving,
D1[P] [(1) =][ D]1[P] [(0)]
+ 0.0033(P1[G][(0)][ −] [P][ G]1 [0](0)) = 1.5707 (38)
The factor 0.0033 represents ε. Applying the same procedure
for G2, gives D2[P] [(1) = 3][.][1404][. Finally, the current output]
power of the units in VCM are estimated using the expressions
in (35) and (36), which gives: P1[G][(1) = 249][.][97][ and][ P][ G]2 [(1) =]
125.03. Notice in Table III, and as described in Sec. III-B, the
units in VCM always maintain the power balance, supplying
the total amount of load, even during the execution of the
proposed distributed strategy.
This procedure can be applied iteratively, reducing the
total generation cost, as can be seen in Table III. The final
solution will converge to the optimal solution of the centralized
problem, given by P1[G][0] = 193.47, P2[G][0] = 77.49 and
P3[G][0] = 129.03, with a total generation cost of 15572.25 $.
C. Case I: Initialization, Validation and Comparison
Fig. 7 shows the total generation cost, while Fig. 8 and
Fig. 9 show the local estimation of the incremental cost
variable (λm) and the total active output power of each DG
unit, respectively; all during operation. The iteration counter
2A detailed derivation of this solution can be found in [20].
3D1P [(0) = 2][π][ ·][ 0][.][1][/][0][.][4 = 1][.][5700][.]
4In all the protocols, the active power is in p.u., using 1000 kW as the
i l b
k, in Fig. 7 to Fig. 9 (and the remaining ones), should be seen
as a discretization of the operational time, using a time length
of ∆tD (see Sec. IV-F). Additionally, it is assumed that there
is not changes in the operational conditions (increase/decrease
of load and WT generation) until the proposed strategy reaches
the optimal solution. This is done in order to assess its
convergence properties.
Initially, the value of λm was set to zero at each DG unit.
Due to this, the output power of units in PCM is set at their
minimum value, as a result of the problem stated in (14). This
is shown in the early iterations in Fig. 9d. Simultaneously, the
units operating in VCM correct the active power mismatch,
supplying the remaining active power, as can be seen in Fig 9a
to Fig 9c. After some iterations, the units operating in VCM
increase their local estimation of λm, as shown in Fig. 8, this
value is then distributed through the communication topology,
and as a consequence, the output powers of units operating in
PCM is increased. The units operating in VCM decrease their
output powers to respond to the increase in the output powers
of the units operating in PCM. For this, the active gain of the
droop loops are modified as shown in Fig. 10. Moreover, due
to the way the consensus protocol is defined in (15) for units
operating in VCM, their output powers will converge to the
optimal value defined through the solution of (14). Therefore,
through this cooperative procedure, the total generation cost
is reduced as the strategy reaches the optimal solution, as can
be seen in Fig. 7.
Regarding the reactive power, as all units operate with
a droop-based loop, the reactive generation is distributed
proportionally according to the rating of each DG unit, as
shown in Fig. 11. Notice that, although the reactive power
dispatch is not considered in the proposed strategy, the voltage
constraints are considered through the definition of the reactive
droop gain in (19), which guarantees that the voltage in the
microgrid will operate within the maximum and minimum
allowed values, as it is shown in Fig. 12, before and after
the convergence of the iteration process. As for the frequency,
Fig. 13 compares the frequency of the system with and without
considering the protocol in (17). Thus, when protocol in (17) is
considered, the system operates with a lower frequency deviation. This is accomplished after each DG unit defines locally
their frequency reference ωm, using local measurements of the
system’s frequency, as shown also in Fig. 13.
The solution that the proposed distributed strategy reaches
is independent of the communication topology of the DG
units. This can be seen in Fig. 7, where the total cost is the
same for both communication topologies. Moreover, notice
that, although the results for the dual variables displayed in
Fig 8 are obtained considering the ring topology shown in
-----
Table III
FIRST ITERATIONS OF THE NUMERICAL EXAMPLE
k P1[G][(][k][)] P2[G][(][k][)] P3[G][(][k][)] � P Gm [(][k][)] D1[P] [(][k][)] D2[P] [(][k][)] λ1(k) λ2(k) λ3(k) P1[G][0] (k) P2[G][0] (k) P3[G][0] (k) $
0 250.00 125.00 25.00 400.0 1.5700 3.1400 0.000 0.000 0.000 40.00 20.00 25.00 20586.75
1 249.97 125.03 25.00 400.0 1.5707 3.1404 7.000 3.500 0.000 40.00 20.00 25.00 20587.44
2 249.94 125.06 25.00 400.0 1.5714 3.1407 10.499 7.001 3.500 40.00 20.00 25.00 20588.14
3 249.92 125.08 25.00 400.0 1.5721 3.1411 13.998 10.502 7.000 40.00 20.00 25.00 20588.83
4 249.89 125.11 25.00 400.0 1.5728 3.1414 17.497 14.003 10.500 43.12 20.00 25.00 20589.53
5 249.86 125.14 25.00 400.0 1.5735 3.1418 20.892 17.504 14.000 51.61 20.00 25.00 20590.21
6 247.34 123.92 28.74 400.0 1.5741 3.1421 24.074 20.970 17.465 59.56 20.82 28.74 20247.74
7 243.58 122.06 34.36 400.0 1.5748 3.1424 27.096 24.273 20.836 67.11 24.12 34.36 19756.54
8 239.97 120.29 39.75 400.0 1.5754 3.1428 29.950 27.333 24.068 74.25 27.18 39.75 19311.92
9 236.56 118.61 44.83 400.0 1.5759 3.1431 32.641 30.221 27.117 80.98 30.07 44.83 18916.08
10 233.35 117.03 49.62 400.0 1.5764 3.1434 35.179 32.944 29.993 87.32 32.79 49.62 18563.74
0.15
0.1
G13 G19 G25
0.05
460 500
450 400
440 300
296 298 300 302 304 306 308 730 740 750 760 770 780 790
0
0 100 200 300 400 500 600 700 800 900
Figure 10. Active droop gain of units operating in VCM during operation.
The x-axis represents k, i.e., the iteration counter.
G8 G13 G19 G22 G25
|λ8 λ13 λ19 λ22 λ25 600|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|Col11|Col12|Col13|Col14|Col15|
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
|400 200 0 0 100 200 300 400 500 600 700 800 900 500 400 300|||||||||||||||
||||||||||||||||
||||||||||||||||
||||||||||||||||
||||||||||||||||
Figure 8. Local estimation of the incremental cost variable (or dual variable)
at each DG unit using the ring topology in Fig.5b. The x-axis represents k,
i.e., the iteration counter.
1500 Pm[G] Pm[G][0]
1000
500
800
600
400
200
00
|Col1|P mG|P mG0|Col4|Col5|Col6|Col7|Col8|Col9|
|---|---|---|---|---|---|---|---|---|
||||||||||
||||||||||
100
100
900
900
300 400 500 600
(a) Active power of unit G13
300 400 500 600
(b) Active power of unit G19
700
700
200
m[G]
200
800
800
1000
500
00
|Col1|Col2|Col3|(a) Active|e power of|f unit G13|Col7|Col8|Col9|
|---|---|---|---|---|---|---|---|---|
||P mG|P mG0|||||||
||||||||||
1500
1000
500
00
|Col1|P mG|P mG0|Col4|Col5|Col6|Col7|Col8|Col9|Col10|
|---|---|---|---|---|---|---|---|---|---|
|||||||||||
|||||||||||
900
900
300 400 500 600
(c) Active power of unit G25
|Col1|Col2|Col3|(c) Active|e power of|Col6|f unit G25|Col8|Col9|Col10|Col11|
|---|---|---|---|---|---|---|---|---|---|---|
|G8|G22|W5|+ W17||||||||
||||||||||||
||||||||||||
300
200
22
200
800
800
0
0 100 200 300 400 500 600 700 800 900
Figure 11. Total reactive output power of all the DG units. The x-axis
represents k, i.e., the iteration counter.
Notice that the formulation of the optimal generation problem in Sec. II-B is independent on the operational mode of the
DG units (VCM or PCM), suggesting that the optimal solution
is also independent on these operation modes. However, as
units operating in VCM share the power losses in proportion
to the their ratings, the final solution will depend on the set of
units operating in VCM and in PCM. To show this, Table IV
compares the optimal solution for different cases of modes
of operation. In all cases, the communication topology used
was the ring topology. According to these results, the solution
obtained have the same total generation cost, but different
active power losses and incremental cost variable. The same
total generation cost is due to the cost of the power losses,
which is negligible when compared with the cost of supplying
the total active load consumption.
The selection of the units operating in VCM is an important
issue, as these units are responsible for automatically correct
the mismatch between generation and load consumption after
any change in the operating conditions and, more importantly,
during the optimization process. Therefore, its selection should
be based on their maximum generation capacity. In this
context, if the generation capacity of these units is lower than
the total load demand the droop control cannot guarantee
1000
500
00
100
8
100
600
700
700
500
400
(d) Active power of units G8, G22 and W5+W17
Figure 9. Total active output power of DG and WT units. The x-axis
represents k, i.e., the iteration counter. In (a) to (c), the blue line represents
Pm[G][, i.e., the real active power of the DG units, while the dashed red line]
represents Pm[G][0] [, i.e., the optimal scheduled active power of each DG unit.]
After convergence, the real output and the optimal scheduled power are the
same.
Fig. 5b, the same results are obtained if the tree topology is
considered instead. In this sense, the main difference is related
to the speed of convergence of the consensus algorithm, which
depends on the connectivity (i.e., the number of links) of the
communication topology
-----
Phase A Phase B Phase C
Table IV
COMPARISON OF THE DISTRIBUTED STRATEGY FOR DIFFERENT CASES
Distributed Strategy Centralized [33]
1.05
1
0.95
1.05
Units in PCM G8, G22 G8 – –
G13, G19 G13, G19, G8, G13, G19, G8, G13, G19,
Units in VCM G25 G22, G25 G22, G25 G22, G25
|Col1|Col2|Col3|Col4|Col5|
|---|---|---|---|---|
||||||
||||||
||||||
25
20
5
10 15
(a) At iteration k = 1
Total Cost [10[5]$] 7.369 7.369 7.369 7.413
Total Losses [kW] 32.77 32.79 33.04 33.54
Frequency [Hz] 60.00 60.00 60.00 59.97
1
max{λm} [$/kW] 450.495 450.43 450.328 –
min{λm} [$/kW] 450.058 450.095 450.185 –
0.95
5 10 15 20 25
(b) At iteration k = 300
Figure 12. Voltage profile of the microgrid for all the phases before and after
the convergence process.
|Col1|Col2|Col3|Col4|Col5|
|---|---|---|---|---|
||||||
||||||
||||||
||||||
||||||
60.1
60.05
60
59.95
59.9
0 100 200 300 400 500 600 700 800 900
Figure 13. Frequency and frequency reference for all units during operation.
The x-axis represents k, i.e., the iteration counter.
|Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|
|---|---|---|---|---|---|---|---|---|
||||||||||
|||ωm(k|)||ω(k|) with the|protocol i|n (17)|
||||ω(k|) without|the prot|ocol in (17|)||
||||||||||
||||||||||
a feasible operation, especially in the early iterations of the
optimization process and after any load increase or WT
generation reduction.
Table IV also shows a comparison with the solution obtained
using the centralized strategy in [33], which considers a static
droop operation framework, i.e., the active and reactive droop
gains are not modified during operation, and they are defined
based on the Standard IEEE 1547.7 [34]. In this case, as the
strategy proposed in [33] is centralized, all the information
regarding the microgrid, the DG units and loads is available
in advance, gathered by a high-level and central operator.
According to these results, the total generation cost and the
power losses are reduced 0.6% and 2.4%, when comparing
the distributed with the centralized solution. Notice that the
static definition for the active droop gain (Dm[P] [) used by the]
centralized solution does not consider the economic operation
of the DG units, but only the generation capacity instead,
sharing the active load among all the units in proportion to
their ratings [33]. In contrast, the proposed approach defines
the active droop gains based on the solution of the economical
dispatch problem, and consequently, a lower cost solution is
obtained. Regarding the frequency, the microgrid operates with
the nominal value in steady-state in all cases, while in the
solution obtained by the centralized strategy a deviation of
0.05% was observed. These results show the effectiveness
of the proposed strategy when compared with a centralized
strategy.
D. Case II: Time-varying Conditions
min{Vm,φ} [p.u.] 0.9466 0.9460 0.9430 0.9430
case, units operating in VCM respond automatically correcting
the active power mismatch in the EDS, as can be seen in
Fig 9. Due to this, the local estimation of the incremental
cost variable (λm) are increased, which cause that units in
PCM respond to the new operational condition, increasing
their output active powers. After some iterations, the system
reaches a new optimal operational state. At k = 450, both
WTs are dispatched, as shown in Fig 9d. This increase in
the renewable generation creates an active power mismatch
(generation is higher than consumption), causing that units
operating in VCM respond automatically and reaching a new
operational condition characterized by a lower value of the
incremental cost variable, as shown in Fig 8.
At k = 600, unit G22, which is operating in PCM,
unexpectedly is turned off and all its communication links are
disabled, simulating a fault. Here, it is assumed that the WTs
generation maintains the same value in order to assess only the
impact due to the DG units fault. This fault creates an active
power mismatch (generation is lower than consumption), that
leads to an immediate response of the units operating in VCM.
In fact, this is one of the main advantages of the proposed
strategy, since some units operate in VCM mode (i.e., with
an active droop loop), they are responsible for automatically
reduce the active power mismatch between generation and
consumption after any change in the operational conditions
of the microgrid, using only local information, increasing the
robustness and reliability during operation.
Finally, at k = 750, unit G22 restores its operation and
the system reaches the same optimal operational point before
the fault. Notice, however, after unit G22 is turned on, the
units operating in VCM have a different value of active droop
gains, as shown in Fig. 10. This is due to the fact that the
output power of units operating in VCM, defined through the
active droop loops, depends on the rates Dm[G] [/D]n[G][,][ ∀][m, n][ ∈G][2][,]
which in this case are the same before and after the simulated
fault of unit G22.
To assess the flexibility of the proposed distributed strategy
under time-varying conditions, different unexpected changes in
the operational conditions of the microgrid are analyzed. Here,
the communication topology used was also the ring topology.
At k 300 the load demand is increased by 10% In this
E. Case III: Impact of the Communication Topology on the
Performance
In this section, the impact of the communication topology
on the performance of the proposed strategy is assessed. To
do this, the communication topologies shown in Fig. 14 are
considered, in addition to the ring and tree topologies shown
in Fig. 5. All the parameters, units operating in VCM and
PCM are defined as in Case I
-----
G22 G22 G22
G19 G25 G19 G25 G19 G25
G13 G13 G13
G8 G8 G8
(a) Fully Connected (b) Leader (c) Robust
Figure 14. Additional communication topologies used to test the proposed
strategy. The robust topology was designed using the rule proposed in [12].
Ring Tree Fully Leader Robust
10
9
8
7
0 50 100 150
Figure 15. Comparison of the convergence of the total generation cost of the
DG units for different communication topologies. The x-axis represents k,
i.e., the iteration counter.
The convergence of the proposed strategy can be affected
by the connectivity level of the DG units, as it is based on a
first-order consensus algorithm. This level of connectivity can
be measured by the coefficient β, defined by the relationship
between the number of links over the number of DG units.
For the considered topologies, this takes the value of 1 and
0.8 for the ring and tree topologies, respectively (see Fig. 5);
and 2, 0.8 and 1.2 for the fully connected, leader and robust
topologies, respectively (see Fig. 14). Thus, it is expected
that communication topologies with higher value of β, reach
consensus faster. However, according to the results shown in
Fig. 15, the tree and robust topology have better performance
(i.e., they converge faster) when compared with the fully
connected topology, which has the higher β. Considering
these results, it is possible to conclude that although the level
of connectivity β and the speed of convergence are closely
related, there is no an inversely proportional relationship
between them. For this reason, it is not possible to know
exactly which topology will have the fastest convergence speed
based exclusively on the values of β. These results are in
agreement with those presented in [7], where a first-order
consensus algorithm was also studied. Finally, it is important
to add that in the simulations the proposed strategy defines
the same optimal dispatch for all the DG units, regardless the
communication topology used.
time-step to discretize the operational time (∆tD). This time
should also include the time to perform all the measurements
and exchange the data between the DG units. In this context,
this low computational time is one of the main advantages of
the proposed distributed strategy, which is a consequence of
the reduction of the size of the optimization problem solved
by each DG unit. Notice that in all cases, all the DG units
have reached consensus in less than 400 iterations, which
means that the proposed approach requires approximately 12 s
to converge to the optimal solution (or 40 s, if 0.1 s is
used for ∆tD), if all the DG units execute the optimization
process in parallel, as expected in practical implementations.
In fact, based on the hierarchical control approach discussed
in Sec. III-D, the maximum time for the distributed approach
to reach the optimal solution is actually limited by TD, which
can take values in the order of minutes. Thus, the proposed
algorithm is sufficiently fast and suitable for implementation.
Finally, it is important to highlight that the results shown
in Fig. 16 were obtained using the proposed heuristic rules in
(18), showing their good performance for these cases, as the
distributed strategy properly reaches the optimal solution.
V. CONCLUSION
F. Case IV: Scalability and Computational Time
In this section, simulations with 3, 7 and 10 DG units were
carried out, in addition to the case with 5 DG units presented
in Sec. IV-C. This is done in order to assess the scalability
performance of the proposed strategy. In all the simulations,
the communication topology used was the ring topology, while
the units operating in VCM and PCM are selected following
the discussion presented in Sec. IV-C.
Fig. 16 shows the active output power of the DG units
in VCM and PCM in all cases. For these simulations, the
maximum computational time required by one DG unit to
solve the problem (14) in one iteration k, is near to 0.030 s.
Based on this a conservative value of 0 1 s can be used as the
In this paper, a distributed strategy for optimal dispatch of
unbalanced three-phase islanded microgrids was presented. To
define the generation dispatch of the DG units that minimize
the overall generation cost, an optimization problem is stated
and solved distributively based on primal-dual constrained
decomposition and a first-order consensus algorithm. Two
operational modes are considered for the DG units: VCM
and PCM. Comprehensive simulations and comparison were
given to show the effectiveness and flexibility of the proposed
distributed approach.
According to the obtained results, the proposed strategy
achieves a lower cost solution, when compared with the
standard centralized approach based on a static droop framework, since the solution of the economic dispatch is used
to define the active droop gains; while the frequency deviation is reduced in steady-state, using a local correction
term. Additionally, as units in VCM operate with an active
droop loop, they are responsible for automatically reduce
the active power mismatch between the generation and the
consumption, after any change in the operational conditions of
the microgrid and, more importantly, during the optimization
process. Finally, as the proposed strategy is considered to
operate within the standard hierarchical control framework,
the dynamics of the primary level (implemented with droop
control) and the dispatch layer (implemented in the Stage I
and II) are decoupled, which helps to maintain the stability of
the system.
APPENDIX
CONVERGENCE ANALYSIS OF THE PROPOSED
DISTRIBUTED STRATEGY
To study the convergence of the proposed distributed strategy, the next conditions are assumed to hold,
(C1) Condition 1: The problem in (4)–(6) is technically
feasible i e the DG units in VCM have enough generation
-----
G13 G25
600
500
400
1600
1500
1400
1300
1200
0
1000
500
1000
800
600
G5 G13 G13 G25
G19 G25
G1 G5 G12 G13
400
0 50 100 150 200 250 300 350 400
400
300
200
100
0
0
50
300
0 50 100 150 200 250 300
200
400
400
300
300
100
100
150
150
200
G19
250
250
350
350
350
400
50
G8 G15 G17 G22
0
0
50
50
300
200
100
0
0
400
400
G8 G17 G22
100 150 200 250 300
(b) Case with 7 DG units
350
100 150 200 250 300
(c) Case with 10 DG units
350
(a) Case with 3 DG units
Figure 16. Active output power of the DG units for the cases with 3, 7 and 10 units. Upper: Units in VCM. Lower: Units in PCM. The x-axis represents k,
i.e., the iteration counter.
capacity to correct the active power mismatch between generation and consumption during the optimization process. This
can be written mathematically as,
� P Gm [≫] � � Pm,φ[D] [−] � Pm[W] [.] (39)
m∈G2 m∈N φ∈F m∈W
(C2) Condition 2: All the DG units exchange information
following a pre-defined communication topology, which defines the consensus matrix C = [cmn], designed as explained
in Sec. III-E.
(C3) Condition 3: All the DG units gather the required
data, process and update their protocols synchronously and in
parallel.
Thus, the next proposition can be defined,
(P1) Proposition 1: The proposed distributed strategy converge monotonically if conditions C1–C3 holds.
To prove Proposition 1, first, recall the proposed distributed
strategy.
For each iteration k, and each DG unit m ∈G, apply
sequentially:
λm(k) λm(k + 1) λm(k + 2)
≈ λm(k)/2γm ≈ λm(k + 1)/2γm
≈ λm(k + 2)/2γm
P [G]m P Gm
Figure 17. Schematic representation of L(Pm[G][0] [, λ]m[(][k][))][ in (48), as a]
function of Pm[G][0] for different values of λm(k), such that λm(k + 2) >
λm(k + 1) > λm(k). Pm[G][0] can be obtained as the root of ∂L(·), which
is approximately ≈ λm(k)/2γm. The areas, Pm[G] [< P][ G]m [and][ P][ G]m [> P] Gm[,]
represent the non-feasible values for Pm[G][0] [, according to the set][ O]m[.]
|Col1|Col2|λm(k)|Col4|λm(k +|+ 1) λm(k + 2|
|---|---|---|---|---|---|
|||||||
|||||||
|≈λm(k)|/2γm|≈λm(k +|1)/2γm|||
|||||≈λm(k|+ 2)/2γm|
Proof: Recall that (40) is equivalent to the definition of the
Lagrangian given in (12). Thus, (40) can be stated as,
L(Pm[G][0] [, λ][m][(][k][)) =]
γm(Pm[G][0][)][2][ +][ β][m][P][ G]m [0] [+][ α][m] [−] [λ][m][(][k][)][P][ G]m [0] [,] (46)
�
(40)
which can be re-written as,
L(·) = γm(Pm[G][0][)][2][ + (][β][m] [−] [λ][m][(][k][))][P][ G]m [0] [+][ α][m][.] (47)
Pm[G][0][(][k][) =][ arg min]
Pm[G][∈O][m]
�
fm(Pm[G][0] [)][ −] [λ][m][(][k][)][P][ G]m [0]
∆Pm[G][(][k][) =][ P][ G]m [−] [P][ G]m [0][(][k][)] (41)
λm(k + 1) = � cmnλn(k) + κ∆Pm[G][(][k][)] (42)
n∈Nm
Dm[P] [(][k][ + 1) =][ D]m[P] [(][k][) +][ ε][∆][P][ G]m [(][k][)] (43)
∆ω(k + 1) = ω0 − ω (44)
ωm(k + 1) = ωm(k) + ˆε∆ω(k + 1) (45)
In (41), Pm[G] [corresponds to the total active output power]
of the DG units, meanwhile ω in (44), corresponds to the
angular frequency of the microgrid, both obtained by the DG
unit through local measurements.
Additionally, consider the next following lemmas,
(L1) Lemma 1: The active output power of a DG unit
operating with droop control is inversely proportional to its
active droop gain, or equivalently, Pm[G] [∝] [1][/D]m[P] [.]
Proof: A technical discussion related to the operation of DG
units with droop control is presented in [20].
(L2) Lemma 2: In (40), Pm[G][0][(][k][ + 1)][ > P][ G]m [0][(][k][)][, if][ λ][m][(][k][ +]
1) > λ (k)
Considering that αm = 0, since the shut-up/shut-down cost
is not considered; and λm(k) ≫ βm in the economic dispatch
problem [35], (47) can be approximated as,
L(·) ≈ γm(Pm[G][0] [)][2][ −] [λ][m][(][k][)][P][ G]m [0] [.] (48)
In order to better understand the solution of (40), which
defines Pm[G][0][, as a function of][ λ][m][(][k][)][, Fig. 17 shows the]
second-order polynomial function given by (48), for different
values of λm(k). Notice that the solution of (40) is equivalent
to find the root of the derivative of (48), in such a way that
Pm[G][0] [∈O][m][. This root can be analytically found as,]
∂L(·)
= 2γmPm[G][0] [−] [λ][m][(][k][) = 0][,] (49)
∂Pm[G][0]
which gives that,
�
Pm[G][0] [≈] λm(k)/2γm
�
. (50)
P [G][0] ∈O
-----
Thus, as γm > 0, from (50) and Fig. 17, is possible to
conclude that Pm[G][0] [∝] [λ][m][(][k][)][, which prove L2.]
For each iteration k ∈ Z[+] ∪{0}, let construct the series
of variables (∆Pm[G][, λ]m[(][k][)][, P][ G]m[0] [(][k][)][, D]m[P] [(][k][)][,][ ∆][ω][(][k][)][, ω]m[(][k][))][, fol-]
lowing the iterative strategy in (41)–(45). Proposition 1 will be
proved, if the series defined using (41)–(45) converge and are
monotonic. To show this, let k = 0, thus λm(0) = 0, ∀m ∈G,
while Dm[P] [(0)][ is defined as in (20) and][ ω][m][(0) =][ ω][0][, as]
explained in Sec. IV. Since λm(0) = 0, then Pm[G][0][(0) =][ P][ G]m[,]
due to (50). Moreover, Pm[G] [>][ 0][, since C1 holds and the]
DG units in VCM correct the active power mismatch in the
microgrid. Thus, ∆Pm[G][(0)][ >][ 0][.]
For k = 1, λm(1) > λm(0) and Dm[P] [(1)][ > D]m[P] [(0)][, since]
∆Pm[G][(0)][ >][ 0][ and][ κ >][ 0][, ǫ >][ 0][ in (42) and (43). As]
λm(1) > λm(0), then, Pm[G][0][(1)][ > P][ G]m [0][(0)][ in (40) and due]
to L2. Moreover, Pm[G][(1)][ < P][ G]m [(0)][ since][ D]m[P] [(1)][ > D]m[P] [(0)]
and due to L1. Finally, ∆ω(1) - 0, since in islanded
droop-based microgrids, the frequency is under the nominal
value when the frequency reference is set to ω0 [33], as
for k = 0. Thus, ωm(1) > ωm(0), since ˆǫ > 0, and as a
consequence, ∆ωm(1) < ∆ω(0), which means that deviation
of the frequency of the microgrid is reduced, considering also
that C3 holds (see in Fig. 1 ω when ωm increase).
Because of the iterative nature of (41)–(45), it can be
verified that the following monotonic series exits for k =
2, 3, 4, ...,
{Pm[G][(][k][)][}][ :][ P][ G]m [(0)][ > P][ G]m [(1)][ >][ · · ·][ > P][ G]m
{Pm[G][0][(][k][)][}][ :][ P][ G]m [0][(0)][ < P][ G]m [0][(1)][ <][ · · ·][ < P] Gm
{∆Pm[G][(][k][)][}][ : ∆][P][ G]m [(0)][ >][ ∆][P][ G]m [(1)][ >][ · · ·][ >][ 0]
{λm(k)} : λm(0) < λm(1) < · · · < λ
{Dm[P] [(][k][)][}][ :][ D]m[P] [(0)][ < D]m[P] [(1)][ <][ · · ·][ < D]m[P] [(][k][)]
{∆ω(k)} : ∆ω(0) > ∆ω(1) > · · · > 0
{ωm(k)} : ωm(0) < ωm(1) < · · · < ωm(k).
In the last series, P [G]m [is a bound of][ {][P]m[ G][(][k][)][}][, while]
G
P m [is a bound of][ {][P]m[ G][0][(][k][)][}][. Moreover, the series][ {][P]m[ G][(][k][)][}]
is monotonically decreasing, while the series {Pm[G][0] [(][k][)][}][ is]
monotonically increasing. Due this, the series {∆Pm[G][(][k][)][}][ is]
monotonically decreasing and bounded by 0. These series can
be technically interpreted as, during the optimization process,
the DG units can not supply lower than P [G]m [and the optimal]
G
dispatch can not be greater than P m[. The fact that the series]
{∆Pm[G][(][k][)][}][ is decreasing imply that][ P][ G]m [converge to][ P][ G]m [0][.]
The series {λm(k)}, {Dm[P] [(][k][)][}][ and][ {][ω][m][(][k][)][}][ are not]
bounded, but they are limited since their definition in (42),
(43) and (45), are a function of ∆Pm[G][(][k][)][ and][ ∆][ω][(][k][)][, which]
are bounded and monotonically converge to 0. Additionally, as
C2 holds, all the λm(k) converge to λ through the consensus
matrix C = [cmn].
Thus, based on the fact that as all these series are monotonic
and limited, P1 is proved to be valid.
REFERENCES
[1] J. Zhu, Optimization of Power System Operation. USA: Wiley-IEEE,
2009
[2] D. Olivares, C. Cañizares, and M. Kazerani, “A centralized energy
management system for isolated microgrids,” IEEE Trans. Smart Grid,
vol. 5, no. 4, pp. 1864–1875, July 2014.
[3] K. H. Youssef, “Power quality constrained optimal management of unbalanced smart microgrids during scheduled multiple transitions between
grid-connected and islanded modes,” IEEE Trans. Smart Grid, vol. 8,
no. 1, pp. 457–464, Jan. 2017.
[4] P. P. Vergara, J. C. Lopez, L. C. da Silva, and M. J. Rider, “Securityconstrained optimal energy management system for three-phase residential microgrids,” Electr. Pow. Syst. Res., vol. 146, pp. 371–382, May
2017.
[5] W. Shi, X. Xie, C. C. Chu, and R. Gadh, “Distributed optimal energy
management in microgrids,” in IEEE Trans. Smart Grid, vol. 6, no. 3,
May 2015, pp. 1137–1146.
[6] Y. Wang, S. Wang, and L. Wu, “Distributed optimization approaches for
emerging power systems operation: A review,” Electr. Pow. Syst. Res.,
vol. 144, pp. 127 – 135, 2017.
[7] Z. Zhang and M. Y. Chow, “Convergence analysis of the incremental cost
consensus algorithm under different communication network topologies
in a smart grid,” IEEE Trans. Power Systems, vol. 27, no. 4, pp. 1761–
1768, Nov. 2012.
[8] S. Yang, S. Tan, and J. X. Xu, “Consensus based approach for economic
dispatch problem in a smart grid,” IEEE Trans. Power Systems, vol. 28,
no. 4, pp. 4416–4426, Nov 2013.
[9] G. Hug, S. Kar, and C. Wu, “Consensus and innovations approach for
distributed multiagent coordination in a microgrid,” IEEE Trans. Smart
Grid, vol. 6, no. 4, pp. 1893–1903, July 2015.
[10] G. Chen, J. Ren, and E. N. Feng, “Distributed finite-time economic
dispatch of a network of energy resources,” IEEE Trans. Smart Grid, to
be published, 2016.
[11] F. Guo, C. Wen, J. Mao, and Y. D. Song, “Distributed economic dispatch
for smart grids with random wind power,” IEEE Trans. Smart Grid,
vol. 7, no. 3, pp. 1572–1583, May 2016.
[12] W. Zhang, W. Liu, X. Wang, L. Liu, and F. Ferrese, “Online optimal
generation control based on constrained distributed gradient algorithm,”
IEEE Trans. Power Systems, vol. 30, no. 1, pp. 35–45, Jan 2015.
[13] R. Mudumbai, S. Dasgupta, and B. B. Cho, “Distributed control for
optimal economic dispatch of a network of heterogeneous power generators,” IEEE Trans. Power Systems, vol. 27, no. 4, pp. 1750–1760, Nov
2012.
[14] G. Binetti, A. Davoudi, F. L. Lewis, D. Naso, and B. Turchiano, “Distributed consensus-based economic dispatch with transmission losses,”
IEEE Trans. Power Systems, vol. 29, no. 4, pp. 1711–1720, July 2014.
[15] Y. Xu and Z. Li, “Distributed optimal resource management based on
the consensus algorithm in a microgrid,” IEEE Trans. on Ind. Electr.,
vol. 62, no. 4, pp. 2584–2592, April 2015.
[16] G. Chen, F. L. Lewis, E. N. Feng, and Y. Song, “Distributed optimal
active power control of multiple generation systems,” IEEE Trans. Ind.
Electr., vol. 62, no. 11, pp. 7079–7090, Nov 2015.
[17] H. Pourbabak, J. Luo, T. Chen, and W. Su, “A novel consensus-based
distributed algorithm for economic dispatch based on local estimation
of power mismatch,” IEEE Trans. Smart Grid, to be published, 2017.
[18] J. M. Guerrero, J. C. Vasquez, J. Matas, L. G. de Vicuña, and M. Castilla,
“Hierarchical control of droop-controlled AC and DC microgrids - A
general approach toward standardization,” IEEE Trans. Ind. Electron.,
vol. 58, no. 1, pp. 158–172, Jan. 2011.
[19] W. J. Ma, J. Wang, V. Gupta, and C. Chen, “Distributed energy
management for networked microgrids using online alternating direction
method of multipliers with regret,” IEEE Trans. Smart Grid, to be
published, 2016.
[20] S. J. Ahn, J. W. Park, I. Y. Chung, S. I. Moon, S. H. Kang, and
S. R. Nam, “Power-sharing method of multiple distributed generators
considering control modes and configurations of a microgrid,” IEEE
Trans. Power Delivery, vol. 25, no. 3, pp. 2007–2016, July 2010.
[21] D. Wu, F. Tang, T. Dragicevic, J. C. Vasquez, and J. M. Guerrero, “A
control architecture to coordinate renewable energy sources and energy
storage systems in islanded microgrids,” IEEE Trans. Smart Grid, vol. 6,
no. 3, pp. 1156–1166, May 2015.
[22] J. Matas, M. Castilla, L. G. d. Vicuña, J. Miret, and J. C. Vasquez,
“Virtual impedance loop for droop-controlled single-phase parallel inverters using a second-order general-integrator scheme,” IEEE Trans.
Power Electronics, vol. 25, no. 12, pp. 2993–3002, Dec 2010.
[23] J. M. Rey, P. Marti, M. Velasco, J. Miret, and M. Castilla, “Secondary
switched control with no communications for islanded microgrids,”
IEEE T I d El t i l 64 11 8534 8545 N 2017
-----
[24] R. Olfati-Saber and R. M. Murray, “Consensus problems in networks of
agents with switching topology and time-delays,” IEEE Trans. Automatic
Control, vol. 49, no. 9, pp. 1520–1533, Sep. 2004.
[25] R. Olfati-Saber, J. A. Fax, and R. M. Murray, “Consensus and cooperation in networked multi-agent systems,” Proceedings of the IEEE,
vol. 95, no. 1, pp. 215–233, Jan 2007.
[26] M. Zhu and S. Martinez, “On distributed convex optimization under
inequality and equality constraints,” IEEE Trans. Automatic Control,
vol. 57, no. 1, pp. 151–164, Jan. 2012.
[27] F. Chen, M. Chen, Q. Li, K. Meng, Y. Zheng, J. M. Guerrero, and
D. Abbott, “Cost-based droop schemes for economic dispatch in islanded
microgrids,” IEEE Trans. Smart Grid, vol. 8, no. 1, pp. 63–74, Jan. 2017.
[28] Q. Shafiee, C. Stefanovic, T. Dragicevic, P. Popovski, J. C. Vasquez,
and J. M. Guerrero, “Robust networked control scheme for distributed
secondary control of islanded microgrids,” IEEE Trans. Ind. Electronics,
vol. 61, no. 10, pp. 5363–5374, Oct 2014.
[29] X. Lu, X. Yu, J. Lai, Y. Wang, and J. M. Guerrero, “A novel distributed
secondary coordination control approach for islanded microgrids,” IEEE
Trans. Smart Grid, to be published, 2017.
[30] M. M. A. Abdelaziz, H. E. Farag, E. F. El-Saadany, and Y. Mohamed,
“A novel and generalized three-phase power flow algorithm for islanded
microgrids using a newton trust region method,” IEEE Trans. Power
Systems, vol. 28, no. 1, pp. 190–201, Feb. 2013.
[31] A. Wachter and L. T. Biegler, “On the implementation of an interiorpoint filter line-search algorithm for large-scale nonlinear programming,”
Math. Prog., vol. 106, no. 1, pp. 25–57, 2006.
[32] G. K. V. Raju and P. R. Bijwe, “Efficient reconfiguration of balanced
and unbalanced distribution systems for loss minimisation,” IET Gen.
Trans. Distr., vol. 2, no. 1, pp. 7–12, Jan. 2008.
[33] P. P. Vergara, J. C. Lopez, M. J. Rider, and L. C. P. da Silva, “Optimal
operation of unbalanced three-phase islanded droop-based microgrids,”
IEEE Trans. Smart Grid, to be published, 2017.
[34] Standard 1547: IEEE Standard for Interconnecting Distributed Resources with Electric Power Systems,, IEEE Interconnecting Committee
Std., 2013.
[35] A. Wood and B. Wollenberg, Power Generation, Operation and Control.
New York, NY, USA: Wiley, 1996.
Pedro P. Vergara was born in Barranquilla, Colombia in 1990. He received
the B.Sc. degree in electronic engineering from the Universidad Industrial
de Santander, Bucaramanga, Colombia, in 2012, and the M.Sc. degree in
electrical engineering from University of Campinas, UNICAMP, Campinas,
Brazil, in 2015. He is currently working toward the Ph.D. degree in electrical
engineering at the University of Campinas and at the University of Southern
Denmark, SDU, Denmark, as part of a double degree program between
UNICAMP and SDU. His current research interests include development
of methodologies for the optimization, planning, and control of electrical
distribution systems with high penetration of distributed generation and
renewable energy systems.
Juan M. Rey was born in Bucaramanga, Colombia in 1989. He received
the B.S. in electrical engineering from Universidad Industrial de Santander,
Bucaramanga, Colombia, in 2012. He is currently working toward the Ph.D.
degree in the Department of Electronic Engineering, Technical University
of Catalonia, Spain. Since 2013, he has been with the Electrical, Electronic
and Telecommunications Engineering School (E3T), Universidad Industrial de
Santander, Bucaramanga Colombia, where he is currently an Assistant Professor. His research interest are power electronics and control for distributed
ti d i id
Hamid R. Shaker received his PhD in 2010 from Aalborg University,
Denmark. He has been a visiting researcher at MIT, a post-doctoral researcher
and an assistant professor at Aalborg University within 2009-2013 and an
associate professor at Norwegian University of Science and Technology
(NTNU), Norway within 2013-2014. Since 2014, he has been an associate
professor at Center for Energy Informatics, University of Southern Denmark.
His research interests are in the area of fault detection and diagnosis, processes
monitoring, modeling and control with applications in energy technology. His
contributions have been reported in more than 90 journal and conference
publications. He serves three journals as a member of editorial board and has
been IPC member for several conferences.
Josep M. Guerrero (S’01-M’04-SM’08-FM’15) received the B.S. degree in
telecommunications engineering, the M.S. degree in electronics engineering,
and the Ph.D. degree in power electronics from the Technical University of
Catalonia, Barcelona, in 1997, 2000 and 2003, respectively. Since 2011, he has
been a Full Professor with the Department of Energy Technology, Aalborg
University, Denmark, where he is responsible for the Microgrid Research
Program (www.microgrids.et.aau.dk). His research interests are oriented to
different microgrid aspects, including power electronics, distributed energystorage systems, hierarchical and cooperative control, energy management
systems, smart metering and the internet of things for AC/DC microgrid
clusters and islanded minigrids.
Prof. Guerrero is an Associate Editor for a number of IEEE TRANSACTIONS. He received the best paper award of the IEEE Transactions on Energy
Conversion for the period 2014-2015, and the best paper prize of IEEEPES in 2015. As well, he received the best paper award of the Journal of
Power Electronics in 2016. In 2014, 2015, 2016, and 2017 he was awarded
by Thomson Reuters as Highly Cited Researcher, and in 2015 he was elevated
as IEEE Fellow for his contributions on distributed power systems and
microgrids.
Bo Nørregaard Jørgensen is founder and head of Center for Energy Informatics at the University of Southern Denmark. Center for Energy Informatics
is an interdisciplinary research center focusing on innovative solutions for facilitating the transition towards a smart sustainable energy system. The center’s
research is conducted in close collaboration with industrial partners, public
bodies, and government agencies. As head of center, Dr. Jørgensen represent
University of Denmark at national and international events, in advisory boards
and government reference committees. He is appointed member of the Danish
Academy of Technical Science. Dr. Jørgensen research focuses on integration
and management of demand-side flexibility with supply-side fluctuations, from
the business and technological perspectives. He holds a Ph.D. in Computer
Science from the University of Southern Denmark, a M.Sc. and a B.Sc. in
Computer System Engineering from Odense University, Denmark.
Luiz C. P. da Silva graduated in electrical engineering in Federal University
of Goias, Goias, Brazil, in 1995 and received the M.Sc. and Ph.D. degrees in
electrical engineering from the University of Campinas, UNICAMP, Campinas, Brazil, in 1997 and 2001, respectively. From 1999 to 2000, he was
visiting Ph.D. student at the University of Alberta, Edmonton, AB, Canada.
Currently, he is an Associate Professor at the University of Campinas,
UNICAMP, Campinas, Brazil. His research interests are power system transmission and distribution.
-----
|
{
"disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1109/TSG.2018.2820748?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/TSG.2018.2820748, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "other-oa",
"status": "GREEN",
"url": "https://findresearcher.sdu.dk/ws/files/160389217/Distributed_Strategy_for_Optimal_Dispatch_of_Unbalanced_Three_Phase_Islanded_Microgrids.pdf"
}
| 2,019
|
[
"JournalArticle"
] | true
| 2019-05-01T00:00:00
|
[] | 26,050
|
en
|
[
{
"category": "Engineering",
"source": "s2-fos-model"
},
{
"category": "Environmental Science",
"source": "s2-fos-model"
},
{
"category": "Economics",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/02038327a403f44d76766bdb875612bc733fb499
|
[] | 0.845957
|
A Novel Distributed Consensus-Based Approach to Solve the Economic Dispatch Problem Incorporating the Valve-Point Effect and Solar Energy Sources
|
02038327a403f44d76766bdb875612bc733fb499
|
Energies
|
[
{
"authorId": "2077526466",
"name": "Muhammad Moin"
},
{
"authorId": "2120578249",
"name": "Waqas Ahmed"
},
{
"authorId": "144811267",
"name": "M. Rehan"
},
{
"authorId": "2069560928",
"name": "Muhammad Iqbal"
},
{
"authorId": "34734215",
"name": "N. Ullah"
},
{
"authorId": "9236658",
"name": "Kamran Zeb"
},
{
"authorId": "35702336",
"name": "Waqar Uddin"
}
] |
{
"alternate_issns": null,
"alternate_names": null,
"alternate_urls": [
"http://nbn-resolving.de/urn/resolver.pl?urn=urn:nbn:ch:bel-155563",
"https://www.mdpi.com/journal/energies",
"http://www.mdpi.com/journal/energies"
],
"id": "1cd505d9-195d-4f99-b91c-169e872644d4",
"issn": "1996-1073",
"name": "Energies",
"type": "journal",
"url": "http://www.e-helvetica.nb.admin.ch/directAccess?callnumber=bel-155563"
}
|
This research focused on the design of a distributed approach using consensus theory to find an optimal solution of the economic dispatch problem (EDP) by considering the quadratic cost function along with the valve-point effect of generators and renewable energy systems (RESs). A distributed consensus approach is presented for the optimal economic dispatch under a complex valve-point effect by accounting for solar energy in addition to conventional power plants. By employing the beta distribution function and communication topology between generators, a new optimality condition for the dispatch problem was formulated. A novel distributed updation law for generation by considering the communication between generators was provided to deal with the valve-point effect. The convergence of the proposed updation law was proved analytically using Lyapunov stability and graph theory. An algorithm for ensuring a distributed economic dispatch via conventional power plants, integrated with solar energy, was addressed. To the best of the authors’ knowledge, a distributed nonlinear EDP approach for dealing with the valve-point loading issue via nonlinear incremental costs has been addressed for the first time. The designed approach was simulated for benchmark systems with and without a generation capacity constraint, and the results were compared with the existing centralized and distributed strategies.
|
# energies
_Article_
## A Novel Distributed Consensus-Based Approach to Solve the Economic Dispatch Problem Incorporating the Valve-Point Effect and Solar Energy Sources
**Muhammad Moin** **[1], Waqas Ahmed** **[1], Muhammad Rehan** **[1,]*, Muhammad Iqbal** **[2], Nasim Ullah** **[3]** **,**
**Kamran Zeb** **[4,]*** **and Waqar Uddin** **[5]**
1 Department of Electrical Engineering, Pakistan Institute of Engineering and Applied Sciences (PIEAS),
Islamabad 45650, Pakistan
2 Department of Computer Science, National University of Technology (NUTECH),
Islamabad 44000, Pakistan
3 Department of Electrical Engineering, College of Engineering, Taif University, P.O. Box 11099,
Taif 21944, Saudi Arabia
4 School of Electrical Engineering and Computer Science, National University of Sciences and Technology,
Islamabad 44000, Pakistan
5 Department of Electrical Engineering, National University of Technology (NUTECH),
Islamabad 44000, Pakistan
***** Correspondence: rehanqau@gamil.com (M.R.); kamran.zeb@seecs.edu.pk (K.Z.)
**Citation: Moin, M.; Ahmed, W.;**
Rehan, M.; Iqbal, M.; Ullah, N.; Zeb,
K.; Uddin, W. A Novel Distributed
Consensus-Based Approach to Solve
the Economic Dispatch Problem
Incorporating the Valve-Point Effect
and Solar Energy Sources. Energies
**[2023, 16, 447. https://doi.org/](https://doi.org/10.3390/en16010447)**
[10.3390/en16010447](https://doi.org/10.3390/en16010447)
Academic Editors: David Borge-Diez
and Donato Morea
Received: 9 November 2022
Revised: 10 December 2022
Accepted: 21 December 2022
Published: 30 December 2022
**Copyright:** © 2022 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
[Attribution (CC BY) license (https://](https://creativecommons.org/licenses/by/4.0/)
[creativecommons.org/licenses/by/](https://creativecommons.org/licenses/by/4.0/)
4.0/).
**Abstract: This research focused on the design of a distributed approach using consensus theory to**
find an optimal solution of the economic dispatch problem (EDP) by considering the quadratic cost
function along with the valve-point effect of generators and renewable energy systems (RESs). A
distributed consensus approach is presented for the optimal economic dispatch under a complex
valve-point effect by accounting for solar energy in addition to conventional power plants. By
employing the beta distribution function and communication topology between generators, a new
optimality condition for the dispatch problem was formulated. A novel distributed updation law for
generation by considering the communication between generators was provided to deal with the
valve-point effect. The convergence of the proposed updation law was proved analytically using
Lyapunov stability and graph theory. An algorithm for ensuring a distributed economic dispatch via
conventional power plants, integrated with solar energy, was addressed. To the best of the authors’
knowledge, a distributed nonlinear EDP approach for dealing with the valve-point loading issue
via nonlinear incremental costs has been addressed for the first time. The designed approach was
simulated for benchmark systems with and without a generation capacity constraint, and the results
were compared with the existing centralized and distributed strategies.
**Keywords: consensus; distributed algorithm; economic dispatch problem; renewable energy sources;**
incremental cost; non-smooth cost function; optimization; valve-point loading effect
**1. Introduction**
In recent years, great attention has been given to the study and development of optimization techniques; see, for instance [1–5]. One of the fundamental optimization problems
in power systems is deciding the output power of generation facilities that minimizes the
total generation cost, which is commonly referred to as the economic dispatch problem
(EDP). The EDP has been widely investigated since the advent of computers, and efforts
have been focused on developing centralized optimization algorithms [6,7]. Particle swarm
optimization (PSO) is the most popular among other metaheuristic techniques, despite the
fact that it may not converge to an optimal solution in the case of the non-convex power
system optimization problem [8]. Inspired by PSO, economic dispatch algorithms were
investigated by considering generation constraints [9] and wind power uncertainty [10].
-----
_Energies 2023, 16, 447_ 2 of 23
The consideration of the valve-point effect (VPE), resulting from the sequential opening
of control valves in thermal power plants, makes the cost function highly nonlinear. Due
to the VPE, some ripples float over the cost function, which may be modeled as rectified
sine waves. Different techniques are well-established in the literature for solving complex
EDP considering the VPE. A genetic algorithm with a multi-parent crossover solution for
the EDP with the VPE was presented in [11]. The coalescence of incremental rates and bee
colony optimization methods were used in [12]. The authors in [13] used the iterative piecewise linear function approximation and mixed integer programming to find an optimal
solution, and the obtained solution was then improved using the nonlinear programming
models. In [14] (see also [15]), a multi-population-based differential evolution algorithm
was applied to optimize the cost function with the VPE. All of these approaches for solving
the EDP with the VPE are centralized and require a central controller to receive information
from available nodes.
Emerging technologies of renewable energy resources (RESs), such as solar energy,
wind energy, and hydro-power, have influenced researchers to devise methods to solve the
EDP, considering integrated power plants. Authors in [16] have exploited PSO, Newton–
Raphson, and binary integer programming methods for finding a combined optimized
solution for solar integrated power systems. The work of [17] considered a modified
genetic algorithm for the consideration of thermal power cost optimization along with
wind–solar constraints for a reduction in toxic emissions. The concept of a multi-generation
system based on photovoltaic cells along with a battery system for the cost of energy
optimization was revealed in [18]. To attain a low-carbon economic dispatch, through
the consideration of bio-gas, wind, and solar sources, the work of [19] considered the
stochastic optimization approach. The methods of [20,21] accounted for low-carbon energy
optimization under various constraints by considering uncertainties in solar irradiance
and energy efficiency, respectively. The major common concern in the above-mentioned
algorithms [11–14,17–21] is that these methods apply a central dispatching facility, which
gathers data of all generation nodes and gives a dispatch command to all nodes accordingly.
The centralized approaches have several concerns, such as a single-point of failure (if the
central node fails), system insecurity as the central processor can be vulnerable to cyberattacks), and time-delays (due to the communication of all nodes with a central dispatch
center). In addition, these centralized optimization methods have privacy of data issues in
a competitive environment, increase the business of the main server due to requests from
all generating nodes, and have computational issues due to a central facility. Owing to
these shortcomings, efforts have been devoted in the recent era to investigate distributed
techniques, as observed in [22–29].
Recently, the cooperative control of multiagent systems (MASs) has been widely
investigated and the EDP has transformed into the consensus of MASs. Some recent works
on applying consensus theory to resolve the EDP in a distributed manner were discussed
in [30–35]. Authors in [36] showed that the distributed EDP is solvable, and an optimal
solution can be obtained if the incremental costs (ICs) of all generation facilities reach an
agreement. In [37], a fully distributed control strategy was designed using two-level control
through an upper level for discovering the reference of optimal power generation and a
lower level for reference tracking. The method in [38] utilized stochastic programming
along with robust and distributed optimization methods to minimize the overall cost of all
generation units, including uncertain and intermittent renewable generations. The work
in [39] developed a distributed scheme via an alternating direction method of multipliers
for resolving the EDP. To address communication delays, it was studied in [40] that a
discrete-time consensus approach should be adopted because information flows discretely
through the underlying communication network. A distributed consensus strategy for EDP
with communication delays was presented in [41]. Adaptive consensus-based strategies for
EDP under communication uncertainties were designed in [42,43]. Based on the literature
review, a brief detail of different areas considered in the existing works is provided in
Table 1. Most of the attention in the above-mentioned literature is paid to minimizing a
-----
_Energies 2023, 16, 447_ 3 of 23
quadratic cost function, which is a smooth and convex function. An attempt to solve
EDP-VPE using a distributed consensus approach was presented in [44], where piecewise linear approximation was used for each nonlinear region. Approximation results
in a loss of information, and the consideration of multiple regions makes this approach
highly conservative.
This paper deals with a distributed cooperative optimization (rather than the conventional central optimization) approach for the economic dispatch by considering thermal
generators under the VPE and a solar energy system for attaining low-carbon footprints.
A new algorithm for dispatching the powers economically by employing the beta distribution function for solar irradiance and by considering a smart-grid via cooperation
and communication between generators through graph theory has been revealed. Here,
a consensus-based distributed algorithm was designed to solve the EDP with a quadratic
cost function and VPE, which takes the generator’s output power as the consensus update
variable and local power mismatch as the feedback variable. It was shown that updating the
generators’ output power in the consensus-based optimization protocol ultimately results
in a consensus of the proposed modified ICs with the VPE under an initial supply–demand
balance assumption according to RESs. The authors further improved the distributed algorithm to deal with the generation capacity constraint by adding a power limit compensation
factor and by omitting the initial supply–demand balance restriction. It was shown that
the proposed algorithms are able to solve the EDP with or without the generator capacity
constraint, while the power demand and supply is balanced in addition to the consideration
of RESs. The novel contributions of the presented work are four-fold:
1. _Optimality Condition under VPE: A new optimality condition for the EDP under the_
VPE of power plants, integrated with solar energy (for the distributed optimization
case), was revealed via the Lagrangian method. In contrast to existing conditions [2,
30,33,36,42,43,45,46], the proposed conditions employ modified ICs with the VPE,
and can be applied to more complicated scenarios of the EDP for considering the VPE.
2. _Distributed Dispatching Strategy: A novel distributed approach for the optimal solution_
of the EDP under the VPE and solar energy is proposed. To the best of our knowledge,
a distributed method by considering the communication topology between generators,
without requiring a central dispatch facility, under the nonlinear handling of the VPE,
has been provided for the first time. In contrast to central methods [11–14,17–21,36,47,
48], the proposed distributed approach applies a smart-grid concept for cooperation
between agents, which supports plug-and-play, privacy of data, a simple generatorlevel handling of the dispatch, and better security against cyber attacks. As opposed to
existing centralized strategies in [11–14,17–21], the design of a distributed consensus
algorithm avoids single-point failure, ensures the minimum interaction between
nodes, reduces the computation burden, reduces lags due to the central facility and
promotes the flexible use of communication resources.
3. _Convergence of Algorithm: An analytical convergence analysis of the proposed method_
was performed under VPE constraints, in contrast to the conventional distributed
methods [2,30,33,36,42,43,45,46]. The optimal convergence of the proposed approach
was guaranteed via analysis through Lyapunov stability theory, dynamics of modified
ICs, modified ICs consensus, generation dynamics analysis, and properties of graph
theory, which are non-trivial in the analysis.
4. _Consideration of Clean Energy: The integration of solar energy sources with conventional_
thermal power plants has a substantial influence on the cost and emission reduction,
which was considered in this study, in contrast with the conventional (distributed)
methods [2,30,33,36,42,43,45,46]. The incorporation of green energy sources has a
favorable ecological impact and helps conventionally fuelled power plants to achieve
better carbon trade-offs, resulting in lower carbon penalties imposed by environmental
regulatory authorities. Furthermore, the application of renewable energy plays an
important role in stabilizing state GDP because fuel imports are cut significantly.
-----
_Energies 2023, 16, 447_ 4 of 23
Based on these contributions, the proposed approach can be applied for attaining the
advantages of the distributed EDP (rather than the central EDP), along with the challenges
of the VPE constraint and low-carbon footprints. However, the adaptation of this approach
will require smart infrastructure at generating units, including communication devices,
smart meters, and real-time computational facilities. The simulation was accomplished
on two benchmark test systems, i.e., a ten-unit system and forty-unit system, to validate
the theoretical results, and a comparison was provided with the existing centralized and
distributed approaches. In comparison to [36,47,48], the proposed consensus algorithm
gives a better optimal cost and requires less CPU time.
The remaining paper is organized as follows. In Section 2, the mathematical background of algebraic graphs and consensus in MASs is reviewed. The description of the
problem is provided in Section 3. In Section 4, a distributed algorithm for the EDP considering the VPE, with and without the generation capacity constraint, is proposed. In Section 5,
simulation results and comparisons are provided to validate the effectiveness of the algorithm. Finally, a conclusion is provided to conclude the article.
**Table 1. Area of research considered in existing works.**
**Area of Research Considered** **Works** **Limitations**
Methods with VPE [11–14] Mostly central optimization
Methods concerning RESs [17–21] Mostly central optimization
Distributed EDP methods [2,30,33,36,42,43,45,46] Mostly ignore VPE and RESs
**2. Preliminaries**
Before presenting a detailed analysis of the proposed algorithm, a mathematical
background of algebraic graph theory and the consensus of first-order MASs is provided.
_2.1. Graph Theory_
In a networked system, agents are represented as nodes and the communication
between nodes is represented by edges. A graph is defined as G = _V, E_, where V is the
_{_ _}_
set of nodes, and E is the set of edges. An undirected edge Eij in the network is denoted
by an unordered pair of vertices (vi, vj). The degree of a vertex in an undirected graph
is the total number of edges associated with it. For simplicity, it is assumed that there
are no self-loops and that the graph is connected [36]. Two important associated matrices
with graphs are the adjacency matrix A = [aij]N×N and Laplacian matrix L = [lij]N×N. We
consider that aij = aji = 1 if i and j are connected; otherwise, aij = 0. The entries of the
Laplacian matrix are taken as lij = −aij, i ̸= j and lii = − ∑[N]j=1,j̸=i _[a][ij][, which ensures the]_
diffusion that ∑[N]j=1 _[l][ij][ =][ 0. The following lemma is required to prove the main results.]_
**Lemma 1 ([36]). 1.** _The Laplacian matrix for a connected undirected graph has a zero eigenvalue_
_and the remaining eigenvalues are positive._
_2._ _The second least eigenvalue of the Laplacian matrix, denoted by λo(L), validates the following_
_condition: λo(L) ≤_ _[x]x[T][T][Lx]x_ [.]
_2.2. Consensus of First-Order MASs_
The consensus protocol in MASs is defined as follows [49].
-----
_Energies 2023, 16, 447_ 5 of 23
_x˙i(t) = ui(t),_
_N_
_ui(t) =_ ∑ _aij(xj(t) −_ _xi(t))_
_j=1,j̸=i_
_N_
= − ∑ _lijxj(t),_
_j=1_
(1)
where ui(t) is referred to as the control signal, xi(t) is the state vector, which can represent
a physical quantity, aij is the adjacency matrix entries, and lij is the Laplacian matrix entries.
Consensus in multi-agents is achieved if the following holds.
lim (2)
_t→∞_ _[∥]_ _[x][i][(][t][)][ −]_ _[x][j][(][t][)][ ∥][=][ 0,][ ∀][i][,][ j][ =][ 1, 2,][ · · ·][,][ N][.]_
An interesting result on the consensus of multi-agents is established in [50] as follows.
**Lemma 2. Consensus in multi-agents can be achieved for a connected undirected graph if the**
_following condition holds._
lim (3)
_t→∞_ _[∥]_ _[x][i][(][t][)][ −]_ _[x][∗][(][t][)][ ∥][=][ 0,][ ∀][i][,][ j][ =][ 1, 2,][ · · ·][,][ N][,]_
_where x[∗](t) =_ _N[1]_ [∑]k[N]=1 _[x][k][(][t][)][ represents the average value of states of all agents.]_
**3. System Description**
We assumed a network of N generating facilities working cooperatively to achieve
an optimal power dispatch in a power system or smart-grid. To this end, a quadratic cost
function without the VPE for each generation facility was assumed, which is given as
follows.
_Ci = ai + biPi + ciPi[2][.]_ (4)
Thermal power plants apply a stream to run turbines, which are controlled sequentially
through the opening of stream valves. This opening of valves is needed to increase the
generation of a unit. However, the effect of this valve opening (namely, VPE) causes a
nonlinear rippling effect at the cost function. Hence, a practical generating unit cannot
have a simple quadratic cost function, leading to a highly nonlinear EDP. Including the
VPE into the quadratic cost function leads to the following.
_Ci[vpe]_ = ai + biPi + ciPi[2] [+][ |][e][i] [sin][(][ f][i][(][P]i[min] _−_ _Pi))|,_ (5)
where ai, bi, ci, ei, fi > 0 are cost function coefficients, Pi represents the output power of the
_ith generator, Pi[min]_ is the lower bound of the generation capacity and |ei sin( fi(Pi[min] _−_ _Pi))|_
is the VPE in the cost function. The difference in cost functions (4) and (5) is depicted in
Figure 1.
The below mathematical strategy may be employed to estimate the expense of photovoltaic energy (PE) production.
_CSC = ∑sNSU=1_ _s_ RP,s × MiGs. (6)
Under this scenario, CSC represents the cost of solar energy, whereas NSUs and RP,s
represent the number of solar panels and power, respectively. It is evident from Figure 1 that
(4) is a convex function whereas (5) is a nonlinear, non-smooth, and non-convex function,
-----
_Energies 2023, 16, 447_ 6 of 23
which, in turn, inherits the difficulty in devising an optimization algorithm to solve the
EDP subject to the VPE. The total cost of the power generation is given by
_N_
_CT[vpe]_ = (∑ _Ci[vpe]) + CSC._ (7)
_i=1_
**Cost function without valve point effect**
**Pmin** **Pmax**
**Power Output (MW)**
**Figure 1. The cost function with and without valve-point effect.**
The research objective was to minimize the total generation cost by considering the
valve-point loading effect under the constraint that the power demand and generation
must be balanced; that is,
_N_
min ∑ _Ci[vpe]_
_i=1_
_N_
s.t. PD = ∑ _Pi + RP,s,_
_i=1_
(8)
where PD is the total power demand. Sunlight rays, surrounding temperatures, and the
efficiency characteristics of the photovoltaic panel all have a substantial effect on solar
power production. Here, we incorporated the beta distribution function (BDF) to calculate
the energy production, and the BDF was used to describe solar energy mathematically.
```
D(F + G)
D(F)D(G)
```
_[×][ B][F][−][1][(][1][ −]_ `[B][)][G][−][1]`
_f or 0 ⩽_ `B ⩽` 1, F ⩾ 0, G ⩾ 0
0 _Otherwise_
_BDFβ(B) =_
(9)
where D and G are the parameters of BDFβ. We can write this function in terms of mean X
and standard deviation Z.
� `X(X + 1)` �
`F = X` 1, (10)
_−_
```
Z[2]
```
�� `X(X + 1)` ��
`Y = (1` `X)` 1 . (11)
_−_ _−_
```
Z[2]
```
-----
_Energies 2023, 16, 447_ 7 of 23
As said before, the following model can be used to predict how solar radiation and
ambient temperature would affect the solar output.
_RP(t) = Nsrs × Nparl[RP(SC) ×_ _[R]Srad[(][t][)].[rad]SC_ _× [1 −_ Θ × (Ucel − _Ucel.SC)]]],_ (12)
_Ucel = Uambt +_ _[R]R[(]rad[t][)].[rad]stc_ _× (Unrml.temp −_ 20). (13)
**Assumption 1. The communication topology between generators is connected.**
**Assumption 2. The initial condition of generators is such that ∑i[N]=1** _[P][i][(][0][) +][ R][P,s][ =][ P][D][.]_
An important constraint for generators is the capacity constraint, which is given by
_Pi[min]_ _≤_ _Pi ≤_ _Pi[max], where Pi[min]_ and Pi[max] represent the minimum and maximum generation
limits of the ith generator.
**4. Main Results**
Before presenting the main algorithm, conventional and proposed definitions of IC for
generators are given.
**Definition 1. The incremental cost of the ith generator (by ignoring the VPE) is given by**
_ηi =_ _[∂][C][i]_ = bi + 2ciPi, i = 1, · · ·, N. (14)
_∂Pi_
**Definition 2. The incremental cost of the ith generator by incorporating the VPE has the form**
_ηi,_ _f =_ _[∂][C]∂Pi[vpe]i_ = bi + 2ciPi − _fi(gi)ei cos ( fi(Pi[min]_ _−_ _Pi)),_ (15)
_where gi = sin( fi(Pi[min]_ _−_ _Pi))._
For dealing with the VPE, we applied the modified definition of ICs in Definition 2.
Based on this modified definition, the EDP was resolved via the application of ηi, _f rather_
than conventional ηi. Equation (15) can also be written in a convenient form as
_ηi,_ _f =_ _[∂][C]i[vpe]_ = ηi + φi, (16)
_∂Pi_
where φi = − _fi(gi)ei cos( fi(Pi[min]_ _−_ _Pi))._
Note that the above condition provides the relation between the conventional IC and
the modified IC for the issue of the VPE. The proposed Definition 2 can be interesting as it
can be applied to deal with the EDP for addressing the non-convex valve-point loading
effect.
**Remark 1. An expression for IC with the VPE was derived in the recent interesting and motivating**
_study of [44]. This condition is given as ηi,_ _f = bi + 2ciPi + fiei cos(mod( fi(Pi[min]_ _−_ _Pi), π)),_
_which is also equivalent to the present case of (15). However, the expression (15) is more convenient_
_than the above condition as the signum function is better to understand, realize, and implement. It is_
_also even easier to approximate than the MOD function. Due to this difficulty in [44], the definition_
_provided in [44] for IC with the VPE is based on a piece-wise linear approximation of the mentioned_
_MOD-based expression. The resultant approach for this approximation is conservative due to the_
_loss of information owing to linearization. Furthermore, it is also difficult to design and implement_
_due to the consideration of several regions. The switching between these regions may also cause a_
-----
_Energies 2023, 16, 447_ 8 of 23
_discontinuous operation, which can be fatal. The present work is based on the nonlinear and more_
_relevant Definition 2, which does not have conservatism as observed in [44]._
_4.1. Proposed Optimality Condition_
The optimization problem (8) can have an optimal solution if the conditions in
Lemma 3 are satisfied.
**Lemma 3. The optimal solution of EDP with the VPE and RESs as in (8) can be obtained if**
_ηi + φi = ηj + φj_ (17)
_and_
_N_
### ∑ Pi + RP,s = PD. (18)
_i=1_
**Proof. Using the Lagrange multiplier method, the Lagrange function for (8) was con-**
structed as
_N_ _N_
_L(Pi, λ) =_ ∑ _Ci[vpe]_ + λ(PD − ∑ _Pi −_ RP,s), (19)
_i=1_ _i=1_
where λ is the Lagrange multiplier. By the application of (5), we attain
_N_ _N_
_L(Pi, λ) =_ ∑ _ai + biPi + ciPi[2]_ [+][ |][e][i] [sin][(][ f][i][(][P]i[min] _−_ _Pi))| + λ(PD −_ ∑ _Pi −_ RP,s). (20)
_i=1_ _i=1_
Differentiating L(Pi, λ) with respect to Pi leads to
_∂L_
_∂Pi_ = bi + 2ciPi − _fi(gi)ei cos( fi(Pi[min]_ _−_ _Pi)) −_ _λ._ (21)
Putting the derivative equal to zero for achieving an optimality condition, we have
_ηi + φi_ _λ = 0,_
_−_ (22)
_ηi + φi = λ._
The above equation shows that all IC with the VPE should be equal to a constant. Therefore,
we can say that
_ηi + φi = ηj + φj, ∀i, j = 1, · · ·, N._ (23)
In addition, taking the derivative of L(Pi, λ) with respect to the Lagrange multiplier
produces
_∂L_ _N_
_∂λ_ [=][ P][D][ −] ∑ _Pi −_ RP,s. (24)
_i=1_
Putting the derivative equal to zero leads to
_N_
### ∑ Pi + RP,s = PD. (25)
_i=1_
This completes our proof.
**Remark 2. The conventional distributed IC consensus method [36] (see also [45,46]) does not**
_consider the VPE. Therefore, it has φi = 0, ∀i = 1, · · ·, N. By using this condition in the proposed_
_optimality condition of Lemma 3, the generalized optimal condition in (17) reduces to_
_ηi = ηj, ∀i, j = 1, · · ·, N._ (26)
-----
_Energies 2023, 16, 447_ 9 of 23
_Hence, the proposed condition in Lemma 3 is the generalization of the conventional condition._
_Our approach supports the use of the VPE for attaining coherency between generators for an effective_
_cost minimization._
_4.2. Proposed Consensus-Based Optimization Protocol_
IC with the VPE contains nonlinearity, which is difficult to handle and update in
a consensus protocol. Therefore, we proposed a novel consensus-based optimization
protocol using power generation Pi and updated it to reach the consensus of ICs with the
VPE. The designed consensus protocol is as follows.
_N_
_P˙i =c_ ∑ _aij(bi + 2ciPi −_ _fi(gi)ei cos( fi(Pi[min]_ _−_ _Pi))_ (27)
_j=1_
_−_ _bj −_ 2cjPj + fj(gj)ej cos( fj(Pj[min] _−_ _Pj))),_
with the initial condition ∑i[N]=1 _[P][i][(][0][) +][ R][P,s][ =][ P][D][. For the novel proposed method (][27][),]_
the following condition in Theorem 1 provides the optimal solution of the EDP (8).
**Theorem 1. Consider N distributed generators with generations Pi, ∀i = 1, · · ·, N, with indi-**
_vidual cost functions (5)–(6) under the VPE, connected via a graph of Assumption 1, validating_
_Assumption 2. The proposed optimization protocol (27) for c < 0 under 2ci > fi[2][e][i][ will ensure the]_
_optimal convergence of Pi to Pi[∗][, where P]i[∗]_ _[is an optimal solution of the problem (][8][).]_
**Proof. Using the cost functions in (5)–(6), IC with the VPE is calculated as in (15). Expand-**
ing (15) leads to
_bi + 2ciPi −_ _fiei cos( fi(Pi[min]_ _−_ _Pi)), gi > 0,_
_bi + 2ciPi,_ _gi = 0,_ (28)
_bi + 2ciPi + fiei cos( fi(Pi[min]_ _−_ _Pi)), gi < 0._
_ηi,_ _f =_
Taking the time-derivative, we have
2ciP[˙]i − _fi[2][e][i][ sin][(][ f][i][(][P]i[min]_ _−_ _Pi))P[˙]i,_ _g > 0,_
2ciP[˙]i, _g = 0,_ (29)
2ciP[˙]i + fi[2][e][i][ sin][(][ f][i][(][P]i[min] _−_ _Pi))P[˙]i,_ _g < 0._
_η˙i,_ _f =_
After combining all of these piece-wise functions, we have a generalized dynamics of IC
with the VPE as follows.
_η˙i,_ _f = (2ci −_ _fi[2][e][i][g][i][(][g][i][))][ ˙][P][i][.]_ (30)
Equation (30) can also be written as
_η˙i,_ _f = s(t, Pi)P[˙]i,_ (31)
where s(t, Pi) = 2ci − _fi[2][e][i][g][i][(][g][i][)][. It is important to note that the following condition must]_
be satisfied for a guaranteed consensus (which can be relaxed, to be discussed later)—
2ci > fi[2][e][i][—to make][ s][(][t][,][ P][i][)][ >][ 0. From (][31][), we have]
_P˙i =_ _η˙i,_ _f_ (32)
_s(t, Pi)_ [,]
which indicates that the dynamics of IC with the VPE and dynamics of power generation
depend on each other; that is, _P[˙]i ∝_ _η˙i,_ _f . By multiplying s(t, Pi) on both sides in (27) and_
-----
_Energies 2023, 16, 447_ 10 of 23
writing in terms of IC with the VPE, we can convert the generation dynamics into dynamics
of IC with the VPE via
_N_
_η˙i,_ _f = cs(t, Pi)_ ∑ _aij(ηi,_ _f −_ _ηj,_ _f )._ (33)
_j=1_
This indicates that the design of the EDP protocol using Pi can ultimately result in the
consensus of ICs with the VPE. In (33), s(t, Pi) is a time-dependent variable. This variable
can be transformed into a linear parameter variable (LPV) model as follows [51].
_s(t, Pi) = Θi, where Θi ∈_ [Θmin, Θmax]. (34)
Hence, by the application of LPV model, the relation (33) becomes
_η˙i,_ _f = cΘi_
_N_
### ∑ aij(ηi, f − ηj, f ). (35)
_j=1_
Now, we develop the consensus error dynamics of ICs with the VPE. Let the error
_εi = ηi,_ _f −_ _η¯ as the consensus error, where ¯η = ∑[N]j=1_ Θ1jΘ _[η][j][,]_ _[f][ and][ Θ][ =][ ∑]i[N]=1_ Θ1i [. As per]
Lemma 2, the consensus between ICs with the VPE will be achieved if this consensus error
converges to zero. For constructing the error dynamics, we take the time-derivative of this
error as follows.
_N_ 1
_ε˙i = ˙ηi,_ _f −_ _j∑=1_ ΘjΘ _[η][˙]_ _[j][,]_ _[f][ .]_ (36)
Applying (35) leads to
_ε˙i = cΘi_
_N_
### ∑ aij(ηi.j − ηj, f ) − Θ[c]
_j=1_
_N_
### ∑ ajk(ηj, f − ηk, f ). (37)
_k=1_
_N_
### ∑
_j=1_
The term ∑[N]j=1 [∑]k[N]=1 _[a][jk][(][η][j][,]_ _[f][ −]_ _[η][k][,]_ _[f][ )][ reduces to zero, and we are left with]_
_ε˙i = cΘi_
_N_
### ∑ aij(ηi, f − ηj, f ), (38)
_j=1_
which can be further written as
_ε˙i = cΘi_
_N_
### ∑ aij(ηi, f − η¯ + ¯η − ηj, f ). (39)
_j=1_
The compact form of the error dynamics is attained as follows.
_N_
### ∑ lijε j. (40)
_j=1_
_ε˙i = cΘi_
_N_
### ∑ aij(εi − ε j) = cΘi
_j=1_
After attaining the error dynamics for ICs with the VPE, we show that this error
converges to the origin. This convergence is required to attain the first optimization
condition in Lemma 3. In addition, we also show that the supply–demand condition
also holds. The conditions for the consensus of ICs with the VPE and supply–demand
balance are investigated in Appendix A. By the application of Lemma 3, the proposed
consensus-based optimization protocol (27) guarantees the convergence of Pi to the optimal
solution Pi[∗] [of (][8][). This completes the proof.]
**Remark 3. To solve the optimization problem using the consensus protocol designed according to**
_Theorem 1, a Lagrangian method approach was used to derive the optimal conditions for the issue_
_of the VPE. Since the optimization problem is non-convex, this implies that there may be multiple_
-----
_Energies 2023, 16, 447_ 11 of 23
_optimal solutions based on the initial point. This necessitates that the initial point should be chosen_
_carefully to drive the solution to an optimum value. Therefore, we suggest applying this algorithm_
_for fine-tuning. The conventional distributed optimization methods, by ignoring the VPE, can be_
_applied for the initial solution, while the presented method can be used for fine tuning. Moreover,_
_if different operational constraints are considered, then these constraints will drive the solution_
_towards the global one._
**Remark 4. In Theorem 1, a distributed consensus-based algorithm is designed to dispatch power**
_in a distributed manner in the presence of the VPE and RESs. This is different from conventional_
_distributed strategies as they only consider a quadratic cost function [30,33,36,42,43,45,46]._
**Remark 5. Conventional distributed approaches use IC as the consensus protocol variable [30,33,**
_36,42,43,45,46]. In our approach, a modified IC with the VPE was taken as the consensus variable._
_In addition, the protocol’s update variable was also different (power generation Pi). The inclusion of_
_the VPE in ICs and variation in the protocol update variable for (27) led us to apply the proposed_
_distributed approach for a complex objective function with the valve-point loading effect._
**Remark 6. In this approach, the LPV model was used to transform a time-dependent variable**
_through s(t, Pi) = Θi, where Θi ∈_ [Θmin, Θmax] to reach the consensus of ICs with the VPE.
_The proposed optimization protocol is different from the existing studies, as it contains highly_
_nonlinear terms as fi(gi)ei cos( fi(Pi[min]_ _−_ _Pi)) and fj(gj)ej cos( fj(Pj[min]_ _−_ _Pj)), rather than linear_
_terms as in [30,33,36,42,43,45,46]. These terms appeared due to a novel distributed optimization_
_scenario of the VPE, which was considered in the present study. It should also be noted that_
_optimization analysis for a highly nonlinear protocol (27) is also a challenging research task._
_The presented proof required the generation and IC dynamics with valve-point nonlinearities, LPV_
_modeling, and LPV-based modified IC dynamics. Even the presented Lyapunov function and stability_
_analysis are based on the LPV parameter Θi._
**Remark 7. In the presented EDP approach of Theorem 1, we require 2ci > fi[2][e][i][ making][ s][(][t][,][ P][i][)][ >]**
0, which is a limitation of the proposed method. As s(t, Pi) = 2ci − _fi[2][e][i][g][i][(][g][i][)][ for][ g][i][ =][ sin][(][ f][i][(][P]i[min]_ _−_
_Pi)), the sign of gi can be either positive or negative (with unity gain). Usually, we have ci > 0, and_
_the further negative sign of gi will also contribute towards s(t, Pi) > 0. Therefore, the term s(t, Pi)_
_can have a positive value for most of the time, even if 2ci > fi[2][e][i][ is not validated. The expected]_
_values of Θi for i = 1, ..., N can be positive, resulting in the consensus of expected values of the_
_modified ICs. A simulation study is also provided in the next section to demonstrate the relaxation_
_of the constraint 2ci > fi[2][e][i][. The simulation comparison demonstrated that the presented approach]_
_is still better than the conventional distributed optimization schemes._
**Remark 8. The problem of an optimal dispatch under the complex nonlinear VPE without any**
_linearization is formulated in the framework of distributed consensus-based optimization. To the best_
_of our knowledge, a nonlinear consensus-based distributed approach for the EDP under the VPE for_
_smart-grid applications has been formulated for the first time. The proof of convergence analysis was_
_provided, which is a non-trivial research problem for a distributed strategy. The problem becomes_
_complicated as a central processor and the collection of information to the central unit were relaxed_
_in our study._
To solve the EDP subject to the VPE in a distributed manner, the proposed distributed
algorithm is summarized in steps in Algorithm 1. The proposed approaches in Theorem 1
and Algorithm 1 will remain valid as long as Assumption 2 is valid from the communication
graph topology point of view. However, if a graph has more connections, the convergence
of the algorithm can be faster. It should also be noted that the convergence of the proposed
optimization protocol (27) can be improved by increasing the magnitude of c; however, it
can also amplify the noise effects.
-----
_Energies 2023, 16, 447_ 12 of 23
**Algorithm 1: Algorithm to solve EDP with VPE and RESs**
**Input: PD −** RP,s, aij
**Output: Pi**
**1 Initialize generator parameters: ai, bi, ci, ei, fi, Pi[min], Pi[max], and tolerance τ.**
**2 Set initial generations according to ∑i[N]=1** _[P][i][(][0][) +][ R][P,s][ =][ P][D][.]_
**3 Choose c < 0.**
**4 while | ∑** _aij(ηi,_ _f −_ _ηj,_ _f )| > τ do_
**5** Each unit computes IC with VPE given by
_ηi,_ _f = bi + 2ciPi −_ _fi(gi)ei cos( fi(Pi[min]_ _−_ _Pi))._
**6** All generation units share
_bi + 2ciPi −_ _fi(gi)ei cos( fi(Pi[min]_ _−_ _Pi))_
with neighbours according to underlying communication topology.
**7** Each generator updates Pi according to (27).
**8 End If | ∑** _aij(ηi,_ _f −_ _ηj,_ _f )| ≤_ _τ._
_4.3. Extension to Generator’s Capacity Constraints_
The consensus protocol in (27) does not take the generator’s capacity constraint into
account, and is hence unable to solve the EDP with the VPE in the presence of the capacity
limit constraint. For this protocol to be able to solve this optimization problem, a power
limit compensation factor along with a conditional statement for regulating the generation
constraint was added. The proposed protocol (27) can be modified as follows.
_P˙i = 0,_ if Pi ≤ _Pi[min],_
_N_
_P˙i = c_ _j∑=1_ _aij(ηi,_ _f −_ _ηj,_ _f ) + δi,_ ifPi[max] _≥_ _Pi ≥_ _Pi[min],_ (41)
_P˙i = 0,_ if Pi ≥ _Pi[max],_
where δ = −c0∆Pi, and c0 > 0. The term ∆Pi represents an estimation of the power
mismatch for the ith generation facility, computed via local knowledge. The estimate of
the local power mismatch can be determined by the use of the local communication of
neighboring units.
**5. Simulation Results and Discussions**
_5.1. Simulation_
In this subsection, the designed distributed algorithm is simulated, with and without
the generation capacity constraint, to validate the results of the designed strategy. The simulations were carried out on an Intel Core i7 − 3520M CPU @ 2.90 GHz processor equipped
with 4 GB RAM. For the sake of numerical simulation, two benchmark test systems were
selected. One was the ten-unit system with PD = 2000 MW, and the other was the fortyunit system with PD = 10500 MW. The data set for both test systems was taken from [48].
The unit data for the ten-unit system is depicted in Table 2. The communication topology graph for generators in the case of the ten-unit system is shown in Figure 2. For the
forty-unit system, a randomly generated connected graph was considered.
-----
_Energies 2023, 16, 447_ 13 of 23
**Figure 2. Communication topology graph for a ten-unit system.**
**Table 2. Unit data for ten-unit system.**
**Unit** **_Pi[min]_** **_Pi[max]_** **_ai_** **_bi_** **_ci_** **_ei_** **_fi_**
**1** 10 55 1000.403 40.5407 0.12951 33 0.0174
**2** 20 80 950.606 39.5804 0.10908 25 0.0178
**3** 47 120 900.705 36.5104 0.12511 32 0.0162
**4** 20 130 800.705 39.5104 0.12111 30 0.0168
**5** 50 160 756.799 38.5390 0.15247 30 0.0148
**6** 70 240 451.325 46.1592 0.10587 20 0.0163
**7** 60 300 1243.531 38.3055 0.03546 20 0.0152
**8** 70 340 1049.998 40.3965 0.02803 30 0.0128
**9** 135 470 1658.569 36.3278 0.02111 60 0.0136
**10** 150 470 1356.659 38.2704 0.01799 40 0.0141
5.1.1. Simulation on Ten-Unit System without Generation Constraint
In this case, there is no generation capacity constraint imposed on the generation
units and the initial condition is set such that ∑i[N]=1 _[P][i][(][0][) =][ P][D][. The consensus protocol (][27][)]_
was used. The parameter for the optimization protocol (27) was selected as c = 0.1 by
_−_
virtue of Theorem 1. The total output power and generators’ active power are plotted in
Figures 3 and 4, respectively. Figure 5 shows that ICs with the VPE reach consensus. The
optimal output power of each generation unit with the optimal cost and CPU time is given
in Table 3.
2200
Power generation
2150
2100
2050
2000
1950
1900
1850
1800
0 50 100 150 200
Time(sec)
**Figure 3. Total active power output without capacity constraint.**
**Figure 2. Communication topology graph for a ten-unit system.**
**Table 2. Unit data for ten-unit system.**
**Unit** **_Pi[min]_** **_Pi[max]_** **_ai_** **_bi_** **_ci_** **_ei_**
**1** 10 55 1000.403 40.5407 0.12951 33
-----
_Energies 2023, 16, 447_ 14 of 23
600
unit1
unit2
500
unit3
unit4
400 unit5
unit6
unit7
300 unit8
unit9
unit10
200
100
0
0 50 100 150 200
Time(sec)
**Figure 4. Output power of ten generation nodes without capacity constraint.**
100
unit1
unit2
90 unit3
unit4
unit5
80 unit6
unit7
unit8
unit9
70
unit10
60
50
0 50 100 150 200
Time(sec)
**Figure 5. Consensus of ICs with the VPE.**
**Table 3. Optimal output power of generation units and total cost in case of no capacity limits.**
**Quantity** **Optimal Results**
**_P1 (MW)_** 64.06
**_P2 (MW)_** 80.42
**_P3 (MW)_** 80.85
**_P4 (MW)_** 72.98
**_P5 (MW)_** 60.23
**_P6 (MW)_** 53.10
**_P7 (MW)_** 266.66
**_P8 (MW)_** 311.62
**_P9 (MW)_** 494.37
**_P10 (MW)_** 515.71
**Total Generation (MW)** 2000
**Cost ×10[5]** **($/MWh)** 1.06
**CPU Time (s)** 1.7
-----
_Energies 2023, 16, 447_ 15 of 23
5.1.2. Simulation on Ten-Unit System Using Improved Algorithm with Capacity Constraint
In this case, the improved distributed algorithm (41) is applied on a ten-unit system
with a capacity constraint. In addition, the initial condition is not restricted to be equal to
_PD. Again, c = −0.1 was selected, and we chose c0 = 2 for the modified approach (41)._
The total active output power and generation units’ output power are plotted in Figures 6
and 7, respectively. The initial condition on the total power generation was taken to be
1830 MW. The simulation shows that the algorithm is able to solve the EDP considering
the generation capacity constraint and initial conditions other than PD. Figure 8 illustrates
the IC with the VPE. These ICs tend to reach consensus up until when the generation of a
generator is not saturated due to the capacity constraint, and therefore consensus is not
fully achieved. Individual generations are restricted with generation capacity limits, which
restricts generators in achieving complete consensus in the modified ICs. It can be seen in
Figure 8 that some generation units tried to achieve consensus while few could not, due to
the generation capacity limit, referring to Figure 7.
2200
Power generation
2150
2100
2050
2000
1950
1900
1850
1800
0 50 100 150 200
Time(sec)
**Figure 6. Total active power output using improved consensus protocol considering capacity constraint.**
500
unit1
unit2
400 unit3
unit4
unit5
300 unit6
unit7
unit8
unit9
200
unit10
100
0
0 50 100 150 200
Time(sec)
**Figure 7. Output power of ten generation nodes for improved consensus protocol considering**
capacity constraint.
-----
_Energies 2023, 16, 447_ 16 of 23
100
unit1
unit2
90
unit3
unit4
80 unit5
unit6
unit7
70 unit8
unit9
unit10
60
50
40
0 50 100 150 200
Time(sec)
**Figure 8. Incremental cost consensus in case of generation capacity limit constraint.**
5.1.3. Simulation on Forty-Unit System under RESs
To discuss the validity of the proposed method on large-scale systems, the proposed
consensus protocol (41) was applied to a forty-unit system [48] in the presence of a capacity
constraint. Additionally, the present work also considered the renewable energy sources
in this simulation. We considered a share of 500MW from renewable sources, leading to
RP,s = 500 MW. The initial conditions were taken balanced such that ∑i[N]=1 _[P][i][(][0][) +][ R][P,s][ =]_
_PD. The communication topology and adjacency matrix for this system was generated_
randomly using a standard uniform distribution on MATLAB for incorporating a random
behavior. Again, c = −0.1 was selected, and c0 = 2 was chosen. The total power, which
was ∑i[N]=1 _[P][i][(][t][)][, and the individual generation of each conventional unit are shown in]_
Figures 9 and 10, respectively. In Figure 11, the modified ICs are plotted for the sake
of analysis. The results show that most of the units achieved consensus, whereas the
remaining units attained partial consensus due to saturation to the maximum upper limit
of power generation as imposed by the generator capacity constraint. Hence, the proposed
approach can be applied to a large-scale system with capacity, non-convex, and renewable
energy constraints.
10[4]
1.07
Total Power
1.065
1.06
1.055
1.05
1.045
1.04
1.035
1.03
0 5 10 15 20 25
Time(sec)
**Figure 9. Total power generation in case of forty-unit system.**
-----
_Energies 2023, 16, 447_ 17 of 23
**Figure 10. Individual power generation of forty units.**
**Figure 11. Modified IC consensus for forty-unit system.**
_5.2. Discussion and Comparison_
5.2.1. Comparison with Centralized Algorithms
To authenticate the proposed distributed consensus-based algorithm, a comparison
between the existing centralized strategies to solve EDP-VPE and the proposed strategy
is presented in Table 4. The results obtained from the proposed algorithm are compared
in Table 4, along with those obtained from multi-objective differential evolution (MODE)
in [47] and new global particle swarm optimization (NGPSO) in [48]. For the comparison
study, we considered the case of the capacity constraint and used the approach of the
consensus protocol (41). It can be seen that the proposed strategy gives a comparable cost
(because the central methods are multi-objective schemes) with the advantage of solving
the problem in a distributed manner.
We also provided the expected time for a single node, as the previous CPU time was
computed via the central processing unit. As the generating units are working independently in the proposed work, we can roughly compute the time of the individual node by
dividing the CPU time by the total number of units in our case. Hence, the time in our
case will be further reduced due to the use of distributed computing facilities. Note that
the communication delays in central methods and the business of central processor issues
are also eliminated in our approach. In addition, the proposed approach is not prone to
single-point failure and is resilient against attacks due to its distributed nature compared
to [11–14], [17–21], and [36,47,48]. For launching a cyber attack, an expensive attack for
blocking all generating units will be needed, rather than considering the central unit only.
In addition, the proposed approach is flexible for increasing the number of generating units,
-----
_Energies 2023, 16, 447_ 18 of 23
as it will not require an enhancement of the communication and computational powers of
the central facility. With these advantages, the proposed approach can be a better choice
than the conventional central methods.
**Table 4. Generation, cost, and CPU time comparison for ten-unit system (PD = 2000 MW).**
**Type** **Centralized** **Centralized** **Distributed**
**Quantity** **MODE [47]** **NGPSO [48]** **Proposed**
_P1 (MW)_ 55.00 55.00 55.00
_P2 (MW)_ 79.81 80.00 80.00
_P3 (MW)_ 106.82 106.94 62.42
_P4 (MW)_ 102.83 100.58 87.35
_P5 (MW)_ 82.24 81.50 160.00
_P6 (MW)_ 80.44 83.02 69.99
_P7 (MW)_ 300.00 300.00 300.00
_P8 (MW)_ 340.00 340.00 340.00
_P9 (MW)_ 470.00 470.00 470.00
_P10 (MW)_ 469.90 470.00 375.38
**Cost ×10[5]** **($/MWh)** 1.1150 1.1149 1.082
**CPU time (s)** 9.42 – 2.00
**Time for one node (s)** - – 0.2 approx.
5.2.2. Comparison with Distributed Methods
The consensus protocol from [36], one of the fundamental distributed consensusbased strategies to solve EDP, was applied on a ten-unit system. This comparison study
investigated the optimization protocol (27) under an unconstrained environment. For the
sake of comparison, we took ei to be 100 times larger than that of Table 2, and no generation
constraint was imposed in this experiment. The large value of ei was accounted for, as we
wanted to attempt to solve the EDP with th eVPE in case of a violation of the constraint
2ci > fi[2][e][i][. The obtained power generation][ P][i][ was used to calculate the cost from (][4][) and (][5][),]
and then the obtained results were compared with the results of the proposed algorithm in
Table 5.
It is shown in Table 5 that the optimized cost for [36] with cost function (5) is more than
the optimized cost with cost function (4), which is actually logical because cost function
(4) is an ideal approximation of the fuel cost and does not incorporate the VPE. It is
also evident that the proposed algorithm gives better optimal results compared to [36]
when considering the VPE. In contrast to conventional methods [2,30,33,36,42,43,45,46],
the presented approach considered the effect of RESs for the forty-unit system. In contrast to
conventional distributed methods [2,30,33,36,42,43,45,46], the proposed approach considers
the highly nonlinear VPE constraint and employs low-carbon energy sources in the form of
solar energy. In addition to these two technical advantages, the theoretical convergence
analysis of the proposed method via the stability theory of MASs was performed in the
presence of new constraints through complex Lyapuov, graph theory, and dynamical
analysis formulation, which improves the reliability of the proposed method.
-----
_Energies 2023, 16, 447_ 19 of 23
**Table 5. Comparison with distributed approach.**
**Quantity** **Existing Protocol** **Proposed Protocol**
_P1 (MW)_ 64.29 10.00
_P2 (MW)_ 80.73 200.55
_P3 (MW)_ 82.66 47.00
_P4 (MW)_ 73.00 206.99
_P5 (MW)_ 61.17 50.02
_P6 (MW)_ 52.11 164.46
_P7 (MW)_ 266.32 266.71
_P8 (MW)_ 299.61 315.45
_P9 (MW)_ 494.20 366.00
_P10 (MW)_ 525.91 372.81
Cost without VPE ×10[5] ($/MWh) 1.058 –
Cost with VPE ×10[5] ($/MWh) 1.257 1.144
Total generation (MW) 2000 2000
_5.3. CPU Time_
To emphasize the fact that the proposed approach can solve the optimization problem significantly more quickly than the existing centralized methods, the CPU time was
calculated for all test systems. Due to the distributed framework of optimization, the computation time was significantly reduced compared to central methods. This is shown and
compared in Table 4. In addition, Table 6 is provided, which compares the CPU time for
the proposed approach as applied on different benchmark test systems. The authors want
to emphasize the fact that these CPU times were calculated for the whole simulation time,
and should not be confused with the convergence time of the ICs. In addition, it should be
noted that these simulations were conducted on a central processor. When this algorithm is
implemented in real-time on a distributed controller in the framework of MASs, the CPU
time will be much shorter than those reported in the article.
**Table 6. CPU time comparison.**
**Test System with Approach** **CPU Time (s)**
Ten-unit unconstrained 1.7
Ten-unit constrained 2.0
Forty-unit unconstrained 5.4
Forty-unit constrained 13.5
Recently, some Lyapunov and energy function methods were reported for a better convergence analysis as in [52–55]. In the future, these methods can be applied for investigating
comprehensive convergence properties.
**6. Conclusions**
This paper considered a distributed optimization approach for the EDP under the
VPE and solar energy constraints over a communication topology. The generators were
assumed to be equipped with smart devices, such as transmitters, receivers, and real-time
computational facilities. The proposed strategy applied power generation as an updation
variable and modified ICs as consensus variables for dealing with cost optimization under
clean energy sources by accounting for solar energy distribution properties. In contrast
with the conventional central optimization methods, the proposed distributed approach is
-----
_Energies 2023, 16, 447_ 20 of 23
cooperative, resilient against cyber attacks, not limited to one-point failure, does not have
delays due to the dispatch center, and does not have a server business issue with respect to
the central unit. In addition, it can be easily extended for increasing the number of units
and requires less computational effort due to it having a simple algorithm and the division
of the algorithm at several nodes. Compared with the existing distributed approaches,
the designed distributed consensus protocol deals with the highly nonlinear constraint
of the VPE and incorporates a solar energy system for attaining low-carbon footprints.
Simulation results for medium-scale and large-scale systems were performed along with
a comparison with central and distributed methods. With respect to central methods,
the CPU time for the proposed algorithm was found to be quite better. Compared with
the existing distributed methods, our approach provides a better optimal cost due to the
consideration of the VPE constraint. In future, a more practical approach for considering
a realistic network reconfiguration, including the sizing and allocation of the distributed
energy hubs, will be considered for a distributed optimization framework.
**Author Contributions: M.M. written the initail version of manuscript. W.A. and M.R. completed the**
final version of manuscript. W.A., M.I., N.U. and K.Z. conceived of the idea. M.M., W.A. N.U. and
K.Z. developed the theoretical framework. M.R. and W.U. verified the analytical methods. M.M. and
M.R. performed the simulation results. All authors have read and agreed to the published version of
the manuscript.
**Funding: This Research was Supported by Adaptive Controller Design and Validation of Electric**
Vehicle Charger (Project No. NUST-22-41-45).
**Data Availability Statement: The data used in this study is included in the article. Further inquiries**
can be directed to the corresponding authors.
**Conflicts of Interest: The authors declare no conflict of interest.**
**Appendix A. Consensus and Supply–Demand Conditions**
We took support from Lyapunov stability theory, and considered the following Lyapunov function [56,57]:
_N_ _ε[T]i_ _[ε][i]_
_V =_ ∑ . (A1)
_i=1_ 2Θi
Note that Θi is a positive scalar because of 2ci > fi[2][e][i][, leading to][ s][(][t][,][ P][i][)][ >][ 0, and resulting]
in the LPV parameter Θi > 0. Taking the time-derivative of V gives
_N_
_V˙_ = ∑
_i=1_
_ε[T]i_ _[ε][˙][i]_
. (A2)
Θi
Applying (40) leads to
_N_ _N_
_V˙_ = ∑ _ε[T]i_ _[c]_ ∑ _lijε_ _j._ (A3)
_i=1_ _j=1_
The expansion and evaluation of these sums along with e[T] = [ε1, ε2, ..., ε _N] imply that_
_V˙_ = e[T]cLe. (A4)
By application of Lemma 1, we have
_V˙_ _≤_ _cλo(L)e[T]e._ (A5)
Since c < 0, _V[˙]_ _< 0 is made. This implies that ICs with the valve-point loading effect_
reach consensus with each other. Hence, the first optimality condition in Lemma 3 has
been validated.
-----
_Energies 2023, 16, 447_ 21 of 23
To assure the second condition of Lemma 3 related to the supply–demand balance, we
move towards the generation dynamics. Substituting (32) into (33) leads to the generation
dynamics for the ith generator as follows.
_N_
_P˙i = c_ ∑ _aij(ηi,_ _f −_ _ηj,_ _f )._ (A6)
_j=1_
For achieving total generation dynamics, we applied the summation to the above-mentioned
_ith generation to achieve_
_N_ _N_ _N_
### ∑ P˙i = c ∑ ∑ aij(ηi, f − ηj, f ). (A7)
_i=1_ _i=1_ _j=1_
The expansion and evaluation of these sums reduce the right side to zero. Therefore,
the total generation dynamics will follow
_N_
### ∑ P˙i = 0. (A8)
_i=1_
Equation (A8) implies that ∑i[N]=1 _[P][i][ remains constant during the dispatch process. Hence,]_
∑i[N]=1 _[P][i][ +][ R][P,s][ =][ P][D][, and the second optimality condition in Lemma 3 also holds.]_
**References**
1. Samia, C.; Houssem, J. The Use of a Heuristic Optimization Method to Improve the Design of a Discrete-time Gain Scheduling
[Control Int. J. Control Autom. Syst. 2021, 19, 1836–1846. [CrossRef]](http://doi.org/10.1007/s12555-019-0774-1)
2. Meraihi, Y.; Gabis, A.B.; Mirjalili, S.; Ramdane-Cherif, A. Grasshopper Optimization Algorithm: Theory, Variants, and Applica[tions. IEEE Access 2021, 9, 50001–50024. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2021.3067597)
3. Perng, J.W.; Kuo, Y.C.; Lu, K.C. Design of the PID controller for hydro-turbines based on optimization algorithms. Int. J. Control
_[Autom. Syst. 2020, 18, 1758–1770. [CrossRef]](http://dx.doi.org/10.1007/s12555-019-0254-7)_
4. Awal, M.A.; Masud, M.; Hossain, M.S.; Bulbul, A.A.M.; Mahmud, S.M.H.; Bairagi, A.K. A Novel Bayesian Optimization-Based
[Machine Learning Framework for COVID-19 Detection From Inpatient Facility Data. IEEE Access 2021, 9, 10263–10281. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2021.3050852)
[[PubMed]](http://www.ncbi.nlm.nih.gov/pubmed/34786301)
5. Nelem, A.T.; Ele, P.; Ndiaye, P.A.; Essiane, S.N.; Pesdjock, M.J.P. Dynamic Optimization of Switching States of an Hybrid Power
[Network. Int. J. Control Autom. Syst. 2021, 19, 2468–2478. [CrossRef]](http://dx.doi.org/10.1007/s12555-020-0088-3)
6. Xia, X.; Elaiw, A. Optimal dynamic economic dispatch of generation: A review. Electr. Power Syst. Res. 2010, 80, 975–986.
[[CrossRef]](http://dx.doi.org/10.1016/j.epsr.2009.12.012)
7. Attaviriyanupap, P.; Kita, H.; Tanaka, E.; Hasegawa, J. A Hybrid EP and SQP for Dynamic Economic Dispatch with Nonsmooth
[Fuel Cost Function. IEEE Power Eng. Rev. 2002, 22, 77. [CrossRef]](http://dx.doi.org/10.1109/MPER.2002.4312139)
8. Abbas, G.; Gu, J.; Farooq, U.; Asad, M.U.; El-Hawary, M. Solution of an economic dispatch problem through particle swarm
[optimization: A detailed survey-part I. IEEE Access 2017, 5, 15105–15141. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2017.2723862)
9. Gaing, Z.L. Particle swarm optimization to solving the economic dispatch considering the generator constraints. IEEE Trans.
_[Power Syst. 2003, 18, 1187–1195. [CrossRef]](http://dx.doi.org/10.1109/TPWRS.2003.814889)_
10. Yao, F.; Dong, Z.Y.; Meng, K.; Xu, Z.; Iu, H.H.C.; Wong, K.P. Quantum-inspired particle swarm optimization for power system
[operations considering wind power uncertainty and carbon tax in Australia. IEEE Trans. Ind. Inf. 2012, 8, 880–888. [CrossRef]](http://dx.doi.org/10.1109/TII.2012.2210431)
11. Pulluri, H.; Vyshnavi, M.; Shraddha, P.; Priya, B.S.; Hari, T.S. Genetic Algorithm with Multi-Parent Crossover Solution for
Economic Dispatch with Valve Point Loading Effects. In Innovations in Electrical and Electronics Engineering; Springer: Singapore,
2020; pp. 429–438.
12. Khamsen, W.; Takeang, C.; Aunban, P. Hybrid method for solving the non smooth cost function economic dispatch problem. Int.
_[J. Electr. Comput. Eng. 2020, 10, 609. [CrossRef]](http://dx.doi.org/10.11591/ijece.v10i1.pp609-616)_
13. Sharifzadeh, H. Sharp formulations of nonconvex piecewise linear functions to solve the economic dispatch problem with
[valve-point effects. Int. J. Electr. Power Energy Syst. 2021, 127, 106603. [CrossRef]](http://dx.doi.org/10.1016/j.ijepes.2020.106603)
14. Li, X.; Zhang, H.; Lu, Z. A differential evolution algorithm based on multi-population for economic dispatch problems with
[valve-point effects. IEEE Access 2019, 7, 95585–95609. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2019.2927574)
15. Sakthivel, V.P.; Goh, H.H.; Srikrishna, S.; Sathya, P.D.; Abdul Rahim, S.K. Multi-Objective Squirrel Search Algorithm for
Multi-Area Economic Environmental Dispatch With Multiple Fuels and Valve Point Effects. IEEE Access 2020, 9, 3988–4007.
[[CrossRef]](http://dx.doi.org/10.1109/ACCESS.2020.3046257)
16. Khan, N.A.; Sidhu, G.A.S.; Gao, F. Optimizing combined emission economic dispatch for solar integrated power systems. IEEE
_[Access 2016, 4, 3340–3348. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2016.2587665)_
-----
_Energies 2023, 16, 447_ 22 of 23
17. Kumar Dey, S.; Prasad Dash, D.; Basu, M. Application of NSGA-II for environmental constraint economic dispatch of thermal[wind-solar power system. Renew. Energy Focus 2022, 43, 239–245. [CrossRef]](http://dx.doi.org/10.1016/j.ref.2022.08.008)
18. Zhang, C.; Xia, J.; Guo, X.; Huang, C.; Lin, P.; Zhang, X. Multi-optimal design and dispatch for a grid-connected solar photovoltaicbased multigeneration energy system through economic, energy and environmental assessment. Sol. Energy 2022, 243, 393–409.
[[CrossRef]](http://dx.doi.org/10.1016/j.solener.2022.08.016)
19. Bakirtzis, A.; Petridis, V.; Kazarlis, S. Genetic algorithm solution to the economic dispatch problem. IEE Proc.-Gener. Transm.
_[Distrib. 1994, 141, 377–382. [CrossRef]](http://dx.doi.org/10.1049/ip-gtd:19941211)_
20. Kahvecio˘glu, G.; Morton, D.P.; Wagner, M.J. Dispatch optimization of a concentrating solar power system under uncertain solar
[irradiance and energy prices. Appl. Energy 2022, 326, 119978. [CrossRef]](http://dx.doi.org/10.1016/j.apenergy.2022.119978)
21. Xu, Y.; Song, Y.; Deng, Y.; Liu, Z.; Guo, X.; Zhao, D. Low-carbon economic dispatch of integrated energy system considering the
[uncertainty of energy efficiency. Energy Rep. 2023, 9, 1003–1010. [CrossRef]](http://dx.doi.org/10.1016/j.egyr.2022.11.102)
22. Hua, M.; Ding, H.; Yao, X.Y.; Zhang, X. Distributed Fixed-time Formation-containment Control for Multiple Euler-Lagrange
[Systems with Directed Graphs. Int. J. Control Autom. Syst. 2021, 19, 837–849. [CrossRef]](http://dx.doi.org/10.1007/s12555-020-0106-5)
23. Nezami, Z.; Zamanifar, K.; Djemame, K.; Pournaras, E. Decentralized Edge-to-Cloud Load Balancing: Service Placement for the
[Internet of Things. IEEE Access 2021, 9, 64983–65000. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2021.3074962)
24. Jameel, A.; Rehan, M.; Hong, K.S.; Iqbal, N. Distributed adaptive consensus control of Lipschitz nonlinear multiagent systems
[using output feedback. Int. J. Control 2016, 89, 2336–2349. [CrossRef]](http://dx.doi.org/10.1080/00207179.2016.1155755)
25. Fu, H.; Cui, B.; Zhuang, B.; Zhang, J. Anti-collision and Obstacle Avoidance of Mobile Sensor-plus-actuator Networks over
[Distributed Parameter Systems with Time-varying Delay. Int. J. Control Autom. Syst. 2021, 19, 2373–2384. [CrossRef]](http://dx.doi.org/10.1007/s12555-020-0317-9)
26. Tang, X.; Li, M.; Wei, S.; Ding, B. Event-triggered Synchronous Distributed Model Predictive Control for Multi-agent Systems. Int.
_[J. Control Autom. Syst. 2021, 19, 1273–1282. [CrossRef]](http://dx.doi.org/10.1007/s12555-019-0795-9)_
27. Li, S.; Ai, W.; Wu, J.; Feng, Q. A fixed-time distributed algorithm for least square solutions of linear equations. Int. J. Control
_[Autom. Syst. 2021, 19, 1311–1318. [CrossRef]](http://dx.doi.org/10.1007/s12555-020-0096-3)_
28. Zhang, Q.; Gong, Z.; Yang, Z.; Chen, Z. Distributed convex optimization for flocking of nonlinear multiagent systems. Int. J.
_[Control Autom. Syst. 2019, 17, 1177–1183. [CrossRef]](http://dx.doi.org/10.1007/s12555-018-0191-x)_
29. Hu, C.; Meng, Z.; Qu, G.; Shin, H.S.; Tsourdos, A. Distributed cooperative path planning for tracking ground moving target by
[multiple fixed-wing UAVs via DMPC-GVD in urban environment. Int. J. Control Autom. Syst. 2021, 19, 823–836. [CrossRef]](http://dx.doi.org/10.1007/s12555-019-0625-0)
30. Wang, A.; Liu, W. Distributed incremental cost consensus-based optimization algorithms for economic dispatch in a microgrid.
_[IEEE Access 2020, 8, 12933–12941. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2020.2966078)_
31. Wang, R.; Li, Q.; Zhang, B.; Wang, L. Distributed consensus based algorithm for economic dispatch in a microgrid. IEEE Trans.
_[Smart Grid 2019, 10, 3630–3640. [CrossRef]](http://dx.doi.org/10.1109/TSG.2018.2833108)_
32. Zhang, Z.; Ying, X.; Chow, M.Y. Decentralizing the economic dispatch problem using a two-level incremental cost consensus
algorithm in a smart grid environment. In Proceedings of the 2011 North American Power Symposium, Boston, MA, USA, 4–6
[August 2011; pp. 1–7. [CrossRef]](http://dx.doi.org/10.1109/NAPS.2011.6025103)
33. Zhang, Z.; Chow, M.Y. Incremental cost consensus algorithm in a smart grid environment. In Proceedings of the 2011 IEEE
Power and Energy Society General Meeting, Detroit, MI, USA, 24–28 July 2011; pp. 1–6.
34. Tang, Z.; Hill, D.J.; Liu, T. A novel consensus-based economic dispatch for microgrids. IEEE Trans. Smart Grid 2018, 9, 3920–3922.
[[CrossRef]](http://dx.doi.org/10.1109/TSG.2018.2835657)
35. Zhang, Z.; Chow, M.Y. Convergence Analysis of the Incremental Cost Consensus Algorithm Under Different Communication
[Network Topologies in a Smart Grid. IEEE Trans. Power Syst. 2012, 27, 1761–1768. [CrossRef]](http://dx.doi.org/10.1109/TPWRS.2012.2188912)
36. Yu, W.; Li, C.; Yu, X.; Wen, G.; Lü, J. Distributed consensus strategy for economic power dispatch in a smart grid. In Proceedings
of the 2015 10th Asian Control Conference (ASCC), Kota Kinabalu, Malaysia, 31 May–3 June 2015; pp. 1–6.
37. Xu, Y.; Li, Z. Distributed Optimal Resource Management Based on the Consensus Algorithm in a Microgrid. IEEE Trans. Ind.
_[Electron. 2015, 62, 2584–2592. [CrossRef]](http://dx.doi.org/10.1109/TIE.2014.2356171)_
38. Chang, X.; Xu, Y.; Gu, W.; Sun, H.; Chow, M.Y.; Yi, Z. Accelerated Distributed Hybrid Stochastic/Robust Energy Management of
[Smart Grids. IEEE Trans. Ind. Inf. 2021, 17, 5335–5347. [CrossRef]](http://dx.doi.org/10.1109/TII.2020.3022412)
39. Li, P.; Hu, J. An ADMM based distributed finite-time algorithm for economic dispatch problems. IEEE Access 2018, 6, 30969–30976.
[[CrossRef]](http://dx.doi.org/10.1109/ACCESS.2018.2837663)
40. Zhang, Z.; Chow, M.Y. The influence of time delays on decentralized economic dispatch by using incremental cost consensus
algorithm. In Control and Optimization Methods for Electric Smart Grids; Springer: New York, NY, USA, 2012; pp. 313–326.
41. Zhu, Y.; Yu, W.; Wen, G. Distributed consensus strategy for economic power dispatch in a smart grid with communication time
delays. In Proceedings of the 2016 IEEE International Conference on Industrial Technology (ICIT), Taipei, Taiwan, 14–17 March
2016; pp. 1384–1389.
42. Wen, G.; Yu, W.; Yu, X.; Cao, J. Designing adaptive consensus-based scheme for economic dispatch of smart grid. In Proceedings
of the 2016 Eighth International Conference on Advanced Computational Intelligence (ICACI), Chiang Mai, Thailand, 14–16
February 2016; pp. 236–241.
43. Wen, G.; Yu, X.; Liu, Z.W.; Yu, W. Adaptive consensus-based robust strategy for economic dispatch of smart grids subject to
[communication uncertainties. IEEE Trans. Ind. Inf. 2017, 14, 2484–2496. [CrossRef]](http://dx.doi.org/10.1109/TII.2017.2772088)
-----
_Energies 2023, 16, 447_ 23 of 23
44. Zhou, Y.; Zhu, S.; Chen, Q. Distributed Prescribed Finite Time Consensus Scheme for Economic Dispatch of Smart Grids with the
[Valve Point Effect. Complexity 2020, 2020, 5476846. [CrossRef]](http://dx.doi.org/10.1155/2020/5476846)
45. Yu, M.; Song, C.; Feng, S.; Tan, W. A consensus approach for economic dispatch problem in a microgrid with random delay
[effects. Int. J. Electr. Power Energy Syst. 2020, 118, 105794. [CrossRef]](http://dx.doi.org/10.1016/j.ijepes.2019.105794)
46. Chen, W.; Li, T. Distributed Economic Dispatch for Energy Internet Based on Multiagent Consensus Control IEEE Trans. Autom.
_[Control 2021, 66, 137–152. [CrossRef]](http://dx.doi.org/10.1109/TAC.2020.2979749)_
47. Basu, M. Economic environmental dispatch using multi-objective differential evolution. Appl. Soft Comput. 2011, 11, 2845–2853.
[[CrossRef]](http://dx.doi.org/10.1016/j.asoc.2010.11.014)
48. Zou, D.; Li, S.; Li, Z.; Kong, X. A new global particle swarm optimization for the economic emission dispatch with or without
[transmission losses. Energy Convers. Manag. 2017, 139, 45–70. [CrossRef]](http://dx.doi.org/10.1016/j.enconman.2017.02.035)
49. Olfati-Saber, R.; Murray, R.M. Consensus problems in networks of agents with switching topology and time-delays. IEEE Trans.
_[Autom. Control 2004, 49, 1520–1533. [CrossRef]](http://dx.doi.org/10.1109/TAC.2004.834113)_
50. Ren, W.; Beard, R.W. Distributed Consensus in Multi-Vehicle Cooperative Control; Springer: Berlin/Heidelberg, Germany, 2008.
51. Kamran, M.A.; Hong, K.S. Linear parameter-varying model and adaptive filtering technique for detecting neuronal activities: An
[fNIRS study. J. Neural Eng. 2013, 10, 056002. [CrossRef]](http://dx.doi.org/10.1088/1741-2560/10/5/056002)
52. Yao, Z.; Zhou, P.; Zhu, Z.; Ma, J. Phase synchronization between a light-dependent neuron and a thermosensitive neuron.
_[Neurocomputing 2021, 423, 518–534. [CrossRef]](http://dx.doi.org/10.1016/j.neucom.2020.09.083)_
53. Xu, L.; Qi, G.; Ma, J. Modeling of memristor-based Hindmarsh-Rose neuron and its dynamical analyses using energy method.
_[Appl. Math. Model. 2022, 101, 503–516. [CrossRef]](http://dx.doi.org/10.1016/j.apm.2021.09.003)_
54. [Zhou, P.; Hu, X.; Zhu, Z.; Ma, J. What is the most suitable Lyapunov function? Chaos Solitons Fractals 2021, 150, 111154. [CrossRef]](http://dx.doi.org/10.1016/j.chaos.2021.111154)
55. Ahmad, S.; Rehan, M.; Hong, K.S. Observer-based robust control of one-sided Lipschitz nonlinear systems. ISA Trans. 2016,
_[65, 230–240. [CrossRef]](http://dx.doi.org/10.1016/j.isatra.2016.08.010)_
56. Rehan, M.; Ahmad, S.; Hong, K.S. Novel results on observer-based control of one-sided Lipschitz systems under input saturation.
_[Eur. J. Control 2020, 53, 29–42. [CrossRef]](http://dx.doi.org/10.1016/j.ejcon.2019.10.007)_
57. Hussain, M.; Rehan, M.; Ahn, C.K.; Hong, K.S.; Saqib, N.u. Simultaneous design of AWC and nonlinear controller for uncertain
[nonlinear systems under input saturation. Int. J. Robust Nonlinear Control 2019, 29, 2877–2897. [CrossRef]](http://dx.doi.org/10.1002/rnc.4529)
**Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual**
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3390/en16010447?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/en16010447, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "GOLD",
"url": "https://www.mdpi.com/1996-1073/16/1/447/pdf?version=1672888489"
}
| 2,022
|
[] | true
| 2022-12-30T00:00:00
|
[
{
"paperId": "e77c908ad1e5bcabaa8bf8a204cbabb8a47c3724",
"title": "Low-carbon economic dispatch of integrated energy system considering the uncertainty of energy efficiency"
},
{
"paperId": "903edff32ae88ba67bd5199d335ed48b209a8c74",
"title": "Dispatch optimization of a concentrating solar power system under uncertain solar irradiance and energy prices"
},
{
"paperId": "f5bc57b736a2f3f3efe4483004cbfb8f315f3db7",
"title": "Application of NSGA-II for Environmental Constraint Economic Dispatch of Thermal-Wind-Solar Power System"
},
{
"paperId": "2730e1f3d9f939bab0bb2895a9180760d66a3d86",
"title": "Multi-optimal design and dispatch for a grid-connected solar photovoltaic-based multigeneration energy system through economic, energy and environmental assessment"
},
{
"paperId": "d3307bc0c48318dba4f0acbae3a2d2431756d9d2",
"title": "Modeling of memristor-based Hindmarsh-Rose neuron and its dynamical analyses using energy method"
},
{
"paperId": "23d882676f08b6a3862d623da839312ea998d584",
"title": "What is the most suitable Lyapunov function"
},
{
"paperId": "e2043116c2609c1d9bcaf3ad751417a40297527d",
"title": "Accelerated Distributed Hybrid Stochastic/Robust Energy Management of Smart Grids"
},
{
"paperId": "ab640b4e41574811376b481c2b7a47d88e53ae48",
"title": "Dynamic Optimization of Switching States of an Hybrid Power Network"
},
{
"paperId": "2c7c48090bd56b49ed08aaddc67807628abe30ae",
"title": "Anti-collision and Obstacle Avoidance of Mobile Sensor-plus-actuator Networks over Distributed Parameter Systems with Time-varying Delay"
},
{
"paperId": "c2d3dda981852d7a5f4e8e077843ed2461eb9c96",
"title": "The Use of a Heuristic Optimization Method to Improve the Design of a Discrete-time Gain Scheduling Control"
},
{
"paperId": "ddb41d61c745d6e5f4a2ff27987c5c8ae0090efe",
"title": "A Novel Bayesian Optimization-Based Machine Learning Framework for COVID-19 Detection From Inpatient Facility Data"
},
{
"paperId": "d6f1e247ffe03183e8fb14c2a09df931b3f8dbb5",
"title": "Phase synchronization between a light-dependent neuron and a thermosensitive neuron"
},
{
"paperId": "8a17832ba4062bf3854933b9e46a4a4834f8468b",
"title": "Sharp formulations of nonconvex piecewise linear functions to solve the economic dispatch problem with valve-point effects"
},
{
"paperId": "d60fad1c905393d61c0ab725192acaa4172bac82",
"title": "Event-triggered Synchronous Distributed Model Predictive Control for Multi-agent Systems"
},
{
"paperId": "60a067fafbe7f3239530fdeb35f48ed778363ce7",
"title": "Distributed Fixed-time Formation-containment Control for Multiple Euler-Lagrange Systems with Directed Graphs"
},
{
"paperId": "0b9a5b49e2d15428cd97bddcd67da53bd2c873c3",
"title": "Distributed Cooperative Path Planning for Tracking Ground Moving Target by Multiple Fixed-wing UAVs via DMPC-GVD in Urban Environment"
},
{
"paperId": "983e65bfd76c6f18c5847609cdbad7bd94695f59",
"title": "Distributed Prescribed Finite Time Consensus Scheme for Economic Dispatch of Smart Grids with the Valve Point Effect"
},
{
"paperId": "cf4fcae3667b9443d23dcc3fe21425eadb012b4c",
"title": "A consensus approach for economic dispatch problem in a microgrid with random delay effects"
},
{
"paperId": "525defc6b7bd9629802eaa5be823c6b23b50cefe",
"title": "Decentralized Edge-to-Cloud Load Balancing: Service Placement for the Internet of Things"
},
{
"paperId": "23748090bf03f129e552f431facfe70bae69b37e",
"title": "Novel results on observer-based control of one-sided Lipschitz systems under input saturation"
},
{
"paperId": "18d665651b841ef8aa9f0bf8e6775927041554da",
"title": "Design of the PID Controller for Hydro-turbines Based on Optimization Algorithms"
},
{
"paperId": "8c42dbc9c094432b1dd892ad833a37b3788dfc54",
"title": "Hybrid method for solving the non smooth cost function economic dispatch problem"
},
{
"paperId": "499ecfdf7fabeb66e1b367bb4b579225ca620782",
"title": "Distributed Incremental Cost Consensus-Based Optimization Algorithms for Economic Dispatch in a Microgrid"
},
{
"paperId": "dc4c63f276883a095b2d6436db6e45efcb79fd6b",
"title": "Distributed Consensus Based Algorithm for Economic Dispatch in a Microgrid"
},
{
"paperId": "6b1c707c68a1d1885bd0e418fb697a466381be05",
"title": "A Fixed-time Distributed Algorithm for Least Square Solutions of Linear Equations"
},
{
"paperId": "3cc9842af90668ade143880b65f1fd9ea1a0f929",
"title": "Distributed Convex Optimization for Flocking of Nonlinear Multi-agent Systems"
},
{
"paperId": "624716e76bc51fb3ef73941d034ee000200d0d23",
"title": "Simultaneous design of AWC and nonlinear controller for uncertain nonlinear systems under input saturation"
},
{
"paperId": "8c9ac526021d13fda27220c98e2894a27ba446a6",
"title": "Distributed Economic Dispatch for Energy Internet Based on Multiagent Consensus Control"
},
{
"paperId": "c4bdc7b9f8aae36de0678b206d9a72f62580759b",
"title": "Adaptive Consensus-Based Robust Strategy for Economic Dispatch of Smart Grids Subject to Communication Uncertainties"
},
{
"paperId": "a9dd1498b4fe4a0da5a838431d4abe3e8648bd90",
"title": "A Novel Consensus-Based Economic Dispatch for Microgrids"
},
{
"paperId": "b722707389c1b10599b802866e22ec5a81ba5108",
"title": "A new global particle swarm optimization for the economic emission dispatch with or without transmission losses"
},
{
"paperId": "10f908862b4f80bd0394403309821a742e539bf4",
"title": "Observer-based robust control of one-sided Lipschitz nonlinear systems."
},
{
"paperId": "674a19cdeef6a2059356cc265d0c170c8336ed2b",
"title": "Optimizing Combined Emission Economic Dispatch for Solar Integrated Power Systems"
},
{
"paperId": "00812d9e653d4954552a4dcce987ea7de51c31b4",
"title": "Distributed adaptive consensus control of Lipschitz nonlinear multi-agent systems using output feedback"
},
{
"paperId": "1260805dc54fb86c893f94df3298622e1f1c2c39",
"title": "Distributed Optimal Resource Management Based on the Consensus Algorithm in a Microgrid"
},
{
"paperId": "6b93640439df1f5f8bfc32657a947831b1211b74",
"title": "Linear parameter-varying model and adaptive filtering technique for detecting neuronal activities: an fNIRS study"
},
{
"paperId": "95023389a834a6b9fca3df5899fac96d477753eb",
"title": "Quantum-Inspired Particle Swarm Optimization for Power System Operations Considering Wind Power Uncertainty and Carbon Tax in Australia"
},
{
"paperId": "d7c06dca27dc6329b5f061f38f9004f521dd3cda",
"title": "Convergence Analysis of the Incremental Cost Consensus Algorithm Under Different Communication Network Topologies in a Smart Grid"
},
{
"paperId": "974dd86ab3c07673ca4e78ff73cc68d0fda6fc12",
"title": "Economic environmental dispatch using multi-objective differential evolution"
},
{
"paperId": "a5072af724cce320e6aca60840a3cbee35b76cbc",
"title": "Optimal dynamic economic dispatch of generation: A review"
},
{
"paperId": "bbadd4370cb353cd31a260668824f27a41c9af56",
"title": "Distributed Consensus in Multi-vehicle Cooperative Control - Theory and Applications"
},
{
"paperId": "9839ed2281ba4b589bf88c7e4acc48c9fa6fb933",
"title": "Consensus problems in networks of agents with switching topology and time-delays"
},
{
"paperId": "4ed4c873b592eecdf4c662d95e0f17be136c5409",
"title": "Particle swarm optimization to solving the economic dispatch considering the generator constraints"
},
{
"paperId": "86e87db2dab958f1bd5877dc7d5b8105d6e31e46",
"title": "A Hybrid EP and SQP for Dynamic Economic Dispatch with Nonsmooth Fuel Cost Function"
},
{
"paperId": "0f7af097fe558e98686349be82ad2ae90104a71e",
"title": "Genetic algorithm solution to the economic dispatch problem"
},
{
"paperId": "d44ee8614d118bc1d68a59869e35352df0d2bc08",
"title": "Multi-Objective Squirrel Search Algorithm for Multi-Area Economic Environmental Dispatch With Multiple Fuels and Valve Point Effects"
},
{
"paperId": "f194a613e67ee4d8ba1a24faf4edf0d8d8b1ab0e",
"title": "Grasshopper Optimization Algorithm: Theory, Variants, and Applications"
},
{
"paperId": "8c64b9e0dc3c842a1b68ffecb609d45a5d09941b",
"title": "A Differential Evolution Algorithm Based on Multi-Population for Economic Dispatch Problems With Valve-Point Effects"
},
{
"paperId": "5accccb060936cec582020e562844924d45f1135",
"title": "An ADMM Based Distributed Finite-Time Algorithm for Economic Dispatch Problems"
},
{
"paperId": "09d320c1922384fd9c1cee2aa92ea3bae0ebad06",
"title": "Solution of an Economic Dispatch Problem Through Particle Swarm Optimization: A Detailed Survey - Part I"
}
] | 20,780
|
en
|
[
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Environmental Science",
"source": "s2-fos-model"
},
{
"category": "Engineering",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/020558f09e952f682805891d6c1393b0d3b2be5c
|
[] | 0.822493
|
An Incentive Mechanism for Energy Internet of Things Based on Blockchain and Stackelberg Game
|
020558f09e952f682805891d6c1393b0d3b2be5c
|
International Journal of Engineering
|
[
{
"authorId": "2217243090",
"name": "H. Zhou"
},
{
"authorId": "2112953148",
"name": "J. Gong"
},
{
"authorId": "2054501316",
"name": "W. Bao"
},
{
"authorId": "2107786910",
"name": "Q. Liu"
}
] |
{
"alternate_issns": [
"1985-2312",
"1728-1431",
"2423-7167",
"1735-9244",
"1728-144X",
"2545-417X"
],
"alternate_names": [
"International Journal of Engineering&Technology",
"Int j eng Trans basic",
"International journal of engineering. Transactions A: basics",
"Int J Eng"
],
"alternate_urls": [
"http://ijet.ielas.org/",
"http://www.cscjournals.org/",
"http://www.cscjournals.org/journals/IJE/description.php?JCode=IJE"
],
"id": "e6019940-2d4f-4a6e-95b4-1c2735d629c2",
"issn": "1025-2495",
"name": "International Journal of Engineering",
"type": "journal",
"url": "http://www.ije.ir/"
}
|
In the Internet of Everything era, the Energy Internet of Things (IoT), as a typical application of IoT technology, has been extensively studied. Meanwhile, blockchain technology and energy IoT can be coordinated and complementary. The energy IoT is diversified and has a high transaction demand. it is an issue worthy of research to discuss the impact of the energy IoT environment on the performance of blockchain consensus algorithms and guarantee blockchain stability in energy IoT environment. In the research, an incentive mechanism based on Stackelberg game is proposed for the network scenario involving multiple roadside units and user nodes. The proposed strategy is analyzed through the Matlab simulation platform. The simulation results show that the proposed scheme can effectively protect the interests of blockchain users and miners. It also can improve the security and stability of the blockchain-based energy IoT system. Moreover, the numerical results not only verify the model feasibility. It also shows that when there are many blockchain miners, the model performance is fine. However, when the number of miners reaches a certain value, there will be unobvious growth. Furthermore, it is
|
IJE TRANSACTIONS B: Applications Vol. 36, No. 08, (August 2023) 1468-1477
# International Journal of Engineering
J o u r n a l H o m e p a g e : w w w . i j e . i r
## An Incentive Mechanism for Energy Internet of Things Based on Blockchain and Stackelberg Game
H. Zhou[a], J. Gong[*b], W. Bao[b], Q. Liu[b]
_a Department of Electromechanical and Information Engineering, Changde Vocational Technical College, Changde, China_
_b School of Communication and Information Engineering, Chongqing University of Posts and Telecommunications, Chongqing, China_
_P A P E R I N F O_
_Paper history:_
Received 01 November 2022
Received in revised form 18 April 2023
Accepted 19 April 2023
_Keywords:_
_Blockchain_
_Energy Internet of Things_
_Incentive Mechanism_
_Stackelberg Game_
**NOMENCLATURE**
_A B S T R A C T_
In the Internet of Everything era, the Energy Internet of Things (IoT), as a typical application of IoT
technology, has been extensively studied. Meanwhile, blockchain technology and energy IoT can be
coordinated and complementary. The energy IoT is diversified and has a high transaction demand. it is
an issue worthy of research to discuss the impact of the energy IoT environment on the performance of
blockchain consensus algorithms and guarantee blockchain stability in energy IoT environment. In the
research, an incentive mechanism based on Stackelberg game is proposed for the network scenario
involving multiple roadside units and user nodes. The proposed strategy is analyzed through the Matlab
simulation platform. The simulation results show that the proposed scheme can effectively protect the
interests of blockchain users and miners. It also can improve the security and stability of the blockchainbased energy IoT system. Moreover, the numerical results not only verify the model feasibility. It also
shows that when there are many blockchain miners, the model performance is fine. However, when the
number of miners reaches a certain value, there will be unobvious growth. Furthermore, it is also
confirmed that the wireless energy IoT environment will also create a certain impact on the game model.
**_doi: 10.5829/ije.2023.36.08b.07_**
_N_ The set of miner nodes _TPS_ _dag_ The number of transactions verified per second in the
blockchain network
_Ts ( )_ User's response time _U_ _r*_ The optimal total reward
_Tv ( )_ The transaction verification delay * The optimal equilibrium point
_Tw ( )_ The queuing and service time _U_ _l*_ Benefit function
_U_ _l_ The user's benefit function (2xU)r2 The second derivative of U with respect tor _x_
_f ( )i_ The satisfaction function of blockchain users ( )i The verification delay of the transaction under high load.
The weight factor of the response time function, _c_ The computing and storage cost in each transaction
_Ll_ A convex function with respect to ( ) The ideal response time demand
_x*_ The optimal pricing strategy that can maximize U . r
**1. INTRODUCTION[1]**
Energy IoT is a new energy internet system based on
cutting-edge technologies such as 5G and artificial
intelligence, combined with energy. According to the
complementary mode of different energy sources, energy
internet greatly promotes the linkage between electricity,
fossil, and heat energy sources with the help of internet
*Corresponding Author Email: _[junquan123gong@163.com](mailto:junquan123gong@163.com)_ (J. Gong)
technology [1]. Meanwhile, blockchain technology and
energy IoT can be coordinated and complementary in
integrated development. This complement is mainly
reflected in decentralization, collaborative autonomy,
marketization, and smart contracts.
As a cutting-edge technology, blockchain deeply
integrates a series of emerging computer technologies
such as distributed data storage, P2P (peer-to-peer)
Please cite this article as: H. Zhou, J. Gong, W. Bao, Q. Liu, An Incentive Mechanism for Energy Internet of Things Based on Blockchain and
Stackelberg Game, International Journal of Engineering, Transactions B: Applications, Vol. 36, No. 08, (2023), 1468-1477
-----
H. Zhou et al. / IJE TRANSACTIONS B: Applications Vol. 36, No. 08, (August 2023) 1468-1477 1469
transmission, consensus mechanism, encryption
algorithm, and so on. It also displays distinct application
characteristics of decentralization, openness and
transparency, traceability, and tamper-proofness [2]. The
application value and application scenarios of blockchain
technology in the field of energy IoT have been deeply
discussed in a large number of studies. Zhao et al. [3]
summarized and introduced the development status of
blockchain energy application engineering at home and
abroad. And it has provided reliable development ideas
and suggestions for the engineering application of
blockchain technology in China's energy field. Zhang et
al. [4] comprehensively and systematically sorted out the
application dimensions of blockchain technology in the
energy Internet. The key role of blockchain technology
in the field of energy Internet has been elaborated in
detail from the perspectives of energy, information and
value. Fernández-Caramés et al. [5] described the
demand for blockchain technology in the IoT field and
the impact of its application on the development of the
modern IoT. Doshi and Varghese [6] examined how
renewable energy and AI-powered IoT can be used to
improve agriculture. The paper explores how to use
technologies to optimize crop yield, reduce water
consumption and improve the efficiency of the
agricultural industry. The authors also discussed
potential challenges and solutions to ensure successful
implementation of smart agriculture. Wang and Liu [7]
presented an energy efficient optimization method for
smart-IoT data centers based on task arrival. The authors
proposed a task scheduling algorithm to minimize energy
consumption while ensuring system performance. The
algorithm dynamically assigns tasks to different nodes
based on task arrival, system load, and energy
consumption. This approach is compared with existing
scheduling algorithms. The results show that this method
improves energy efficiency while maintaining system
performance.
However, the most concerned challenge is that the
current performance of the traditional blockchain cannot
meet the needs of high-frequency data usage. The
traditional single-chain structure results in a limited
number of transactions that can be processed in a
consensus cycle. This cannot meet the dynamic
scalability requirements for performance of blockchain
technology in the actual production. Therefore, for the
scalability of blockchain, a distributed ledger based on
DAG is proposed, which greatly improves the system
performance under high concurrency. How to balance the
response strategies of each participant to protect the
interests of blockchain users, miners and the system is a
problem worth studying.
Game theory is a mathematical model for the study of
strategic interactions between rational decision makers
[6, 7]. It can be used to analyze the strategies of nodes
and the interactions between nodes. Due to the power of
game theory, it is one of the new trends of future
development to use game theory to solve the optimization
problem in blockchain. The optimization problem,
especially the CAP theory problem in current blockchain
[8], is namely impossible triangle: decentralization,
scalability and security. Secondly, the Stackelberg game
model is generally widely used to solve the pricing
problem between service providers and users [9, 10]. For
wireless environments like Energy IoT, the work of end
users needs to rely on the purchase of computing
resources from edge computing networks. Modeling the
interaction between the two using Stackelberg games is a
problem worth investigating for system optimization.
Nejati and Faraji [11] dealed with the issue of actuator
fault detection and isolation for a helicopter unmanned
aerial vehicle. The authors proposed a methodology
based on the observer and residual generation technique
to detect and isolate actuator faults in real-time [11].
Khosravian and Maghsoudi [12] discussed the design of
an intelligent controller for station keeping, attitude
control, and path tracking of a quadrotor using recursive
neural networks. The authors proposed a control scheme
based on the fusion of multiple recursive neural networks
for precise control of the quadrotor [12]. Xiong et al. [13]
discussed about cloud computing and pricing
management for blockchain networks. Wei et al. [14] also
investigated on application of blockchain for uncertainity
in energy pricing and market pricing for the enegy
sectors.
Given the basis of game theory and the problems
faced in this paper, this paper proposes a Stackelberg
game-based incentive mechanism based on the DAG
consensus mechanism. The game model simulates the
interaction between blockchain users and miners,
verifying the existence of the game balance point. The
simulation results show that the algorithm can effectively
improve the system security and stability. Specifically, it
aims to improve the system security by encouraging
miners to join the blockchain network, while meeting the
needs of blockchain users. The rest of the paper is
organized as follows. Section 2 introduces the related
problems and system models. Section 3 introduces the
best solution analysis and leader analysis. In section 4,
the simulation results are analyzed and the system
performance is evaluated numerically. Finally, section 5
summarizes this paper. The research objective of this
paper is to propose an incentive mechanism based on
Stackelberg game model to simulate the transaction
behavior between blockchain users and miners. The
proposed scheme can effectively protect the interests of
blockchain users and miners. The security and stability of
the blockchain-based energy IoT system has been
improved.
-----
1470 H. Zhou et al. / IJE TRANSACTIONS B: Applications Vol. 36, No. 08, (August 2023) 1468-1477
**2. PROBLEM DESCRIPTION AND NETWORK**
**MODEL**
**2. 1. System Model** Our model consists of two
entities: 1. blockchain user, namely solar inverter,
vehicle, etc.; 2. Blockchain consensus node, namely
roadside unit with computing and storage capabilities,
also known as miner, as shown in Figure 1. It is
noteworthy that in DAG, miners do not need a lot of
computing resources in mining, just needing to verify
every collected transaction. This is referred to as mining
behavior in this paper. Blockchain users deliver
transactions to miner nodes through wireless channels.
Wireless channels require all blockchain users in the area
covered by miners' nodes to compete with each other.
Miner nodes communicate with each other via wired
channels, run DAG consensus algorithms, validate and
store the collected transactions. This consumes
computing and storage resources. Due to the selfishness
of nodes themselves, this is unfair for miners. Therefore,
to maintain the normal operation of the blockchain
system, it is reasonable for miners to charge certain
transaction fees from blockchain users. For blockchain
users, the transaction verification will cause new delays,
so the process from publication to confirmation of
transactions in the blockchain will go through two stages:
delivery and verification.
The blockchain network model considered in this
study consists of multiple blockchain user clusters, each
of which receives data by a miner node. Where,
_N_ = 1,..., _Nc_ represents the set of miner nodes. The
number of blockchain users within the coverage area of
each miner follows Poisson distribution, and the
transaction arrival rate of users is i i, _N_ . Moreover,
each user has an independent satisfaction function whose
value is related to its own response time needs and the
miner's pricing x of the transaction. In the blockchainbased energy IoT, the user's response time _Ts ( )_ is
composed of two parts. The first part is the queuing and
service time in the wireless phase _Tw( )_ = _Tq_ ( ) +Tst ( ),
and the other part is the transaction verification delay
_Tv ( )_, namely:
_Ts_ ( ) = _Tw( )_ +Tv ( ) (1)
After joining the blockchain network, the user
response time is more affected by the verification delay.
The delay _Tv ( )_ for transactions to be validated at miner
nodes is the time it takes for the cumulative weight of
blockchain transactions to reach the weight threshold.
Due to the directed acyclic graph property in DAG, the
verification delay is proportional to the transaction
generation rate . It means that blockchain users need to
generate more transactions to meet the lower response
delay requirements.
Here, in view of the queuing process in the first stage,
this paper only considers the transaction verification
delay under stable high load. According to the
description in DAG white paper, the change process of
verification delay with transaction arrival rate can be
expressed as:
_Tv_ ( ) = 0.352D ln(4LsN Dc2 ) + 2W W TL- _s(Nac2)_ (2)
Since this study only considers the block verification
process during the high load phase, we need to add a
restriction on the transaction generation rate, i.e.:
_N_ 1
## i (3)
_i_ =1 _N Dc_
where, _N represents the mean value of the distributed_
blockchain user nodes. Meanwhile, it should be made
clear that in the transaction delivery, the wireless channel
capacity is limited. Therefore, the wireless channel will
restrict the transaction delivery after the service intensity
1 . Therefore, this section sets restriction 1,
which can be specifically expressed as follows:
_m_
i _E T[_ _st_ ] (4)
**2. 2. Analysis of Stackelberg Game Model Problem**
To encourage blockchain miners to share their computing
resources, more miners are motivated to participate in the
blockchain consensus to improve the system security.
The system has the authority to require blockchain users
to pay a fee for each transaction. And it allows blockchain
users to have different needs for response time.
Therefore, there is a non-cooperative game between
blockchain users and miners. In this paper, an incentive
mechanism based on Stackelberg game model is
proposed to simulate the interaction between blockchain
users and miner nodes. Where, the set of blockchain
miners is the leader and blockchain users are the
followers. Miners charge transaction fees at the expense
of computing and storage resources, while blockchain
```
区块链共识节点Blockchain
```
Blockchain user区块链用户 consensus node（矿工）
Transaction
```
交易上传upload
```
Block 区块传播
propagation
Block 区块验证
propagation
**Figure 1. Game Model**
-----
H. Zhou et al. / IJE TRANSACTIONS B: Applications Vol. 36, No. 08, (August 2023) 1468-1477 1471
users have higher demand for system response time. This
paper mainly uses a game theory model to maximize the
benefits of blockchain users and miner nodes. And it
verifies the existence of equilibrium points in this game.
**2. 2. 1. Benefit Function of Blockchain Users In**
terms of blockchain users, its benefit function includes
satisfaction function of response time and incentive cost,
namely transaction cost. The response time here
represents the verification delay of transactions in the
DAG network. Due to the different load of transactions
arriving in the network, transactions have different
verification delays. Therefore, the user's benefit function
can be defined in the following equation:
_Ul_ = _f_ (,1 2,...,i ) − i _x_ (5)
In general, logarithmic function is used to evaluate
user satisfaction [11]. Therefore, in this paper, the
satisfaction function of blockchain users with respect to
response time is expressed as follows:
_f_ (i ) = log 1 + _g_ (( _i_ )) (6)
where, represents the weight factor of the response
time function, and ( )i represents the verification delay
of the transaction under high load. It has been calculated
previously. It can be known that ( )i is a function
inversely proportional to the transaction rate i . Let
term is the computing and storage cost in each transaction
_c . This paper assumes that each user has the same_
transaction request, i.e. i .
In general, the benefit functions of leaders and
followers are expressed as follows:
Leader : max _Ul_,
1 _m_
s.t
_NN Dc_ _E T[_ _st_ ]
Followers : maxx _U_ _r_,
(9)
s.t
_N cc_ _x_ _xmax_
1
_g (_ ( _i_ ) = () i )
, so that can be clearly understood
**3. ANALYSIS OF OPTIMAL SOLUTION**
According to the Stackelberg game model proposed in
section 2.2, both blockchain users and miners are rational
users who want to maximize their revenues. If one party
achieves the maximum revenue, it will damage the other
party's revenues and eventually lead to game breakdown.
Therefore, an equilibrium point must be found so that
both buyers and sellers can accept it. In the model, firstly,
blockchain miners fix the price of each transaction on the
basis of their own cost function to gain the optimal total
reward _U from their own strategy space. Secondly, r*_
blockchain users choose respective response time
strategy according to the pricing of miners. In this
section, backward induction [15, 16] will be used to first
analyze the benefit function of the following blockchain
user, especially the verification delay, to obtain the
* _U_ _l*_
optimal equilibrium point and benefit function of
the blockchain user. Then, analysis will be made on the
optimal equilibrium point _x and benefit function *_ _U of r*_
the leading blockchain miner. Finally, in the distributed
environment, the optimal solution can be obtained with
the help of our proposed iterative update function.
Therefore, definition 1 can be obtained based on the
above analysis.
Definition 1: Let the policy set of blockchain users be
_R_ = 1,...,i, and the policy set of miners be
Equation (6).
Through the above analysis, the expression of user
benefit function can be rewrite as follows:
_Ul_ = log 1 + _g_ (( _i_ )) − i _x_ (7)
**2. 2. 2. Benefit Function of Blockchain Mine For**
the blockchain miners, their benefit function is defined as
the charged transaction fees minus the cost of computing
resources consumed per transaction. Miners aim to help
blockchain users verify and store valid transactions and
charge transaction fees _x for each transaction, thereby_
maximizing revenue. Mathematically, the optimization
problem can be expressed as follows:
_U_ _r_ = _TPS_ _dag_ ( _Nxc_ − _c)_ (8)
_U_ _l_ (i*,,R x) _U_ _l_ (, _R−i_, _x)_, −i indicates the user policy
_C_ = x1,..., _x_ _j_ . When _[x]_ is fixed, if
*
meets
−i
where, _TPS_ _dag_ =2LsNc represents the system throughput
in the blockchain network under the wireless channel
service strength 1 . That is, the number of transactions
verified per second in the blockchain network. In
Equation (8), the first term represents the average
verification revenue of all blockchain miners. The second
− _j_
_λ x*,_ - )
miner strategy set excluding. Then, the strategy[(]
is the optimal equilibrium point of the non-cooperative
Stackelberg game.
set excluding
i* . Meanwhile, when is fixed, if
_x*_
meets _U_ _r_ ( _x C*j_,, ) _U_ _r_ ( _x C,_ − _j_, ),
_x*j_ − _j_ represents the
-----
1472 H. Zhou et al. / IJE TRANSACTIONS B: Applications Vol. 36, No. 08, (August 2023) 1468-1477
**3. 1. Follower Analysis Through backward**
induction, first the benefit maximization strategy of the
follower blockchain user is analyzed. For the benefit
function of blockchain users, its derivative is as follows:
t, _t_
instantaneous values of iteration parameters at time
_t can be calculated by solving Equations (11) and (12)_
simultaneously, as shown in Equation (14). _[t]_ represents
the index of iteration times.
Ul = _W2−L NW Ts_ (c2a ) +− _x_
2 2
( U)l2 = − W −W T( _a2)1L N+s_ 2Wc−LsW TN( _c2a)_ 2
(10)
(14)
From the analysis of the above two expressions
combined with the derivation in section 3, the second
derivative 2U2l 0 of _U can be concluded.l_ _U is l_
clearly convex function with respect to []. Due to the
constraint conditions in Equation (9), generally Lagrange
multiplier method is used to solve the optimization
problem. After substituting the constraint conditions into
the benefit function, the following expression can be
obtained.
_t_ _E T[mst_ ] _xt_ −log 1 + 2WL N−s _W Tc2(_ _E Ta_ )[mst ]
=
_m_ 1
−
_E T[_ _st_ ] _NN Dc_
1 _t_ _W_ −W T( _a_ )
t = _NN Dc_ _x_ −log 1 + 2L NNs _c3D_
_m_ 1
−
_E T[_ _st_ ] _NN Dc_
1
_Ll_ (,, ) = log 1 + _g_ (( )) − x − −
_NN D_
_c_
(11)
_m_
− −
_E T[_ _st_ ]
Based on this, the KKT condition can be obtained as
shown in Equation (12). Where, * represents the optimal
solution.
**3. 2. Leader Analysis On the basis of the optimal**
strategy of the following blockchain user, the second step
of backward induction method is to use the obtained
optimal strategy solution of the follower and substitute it
into the leader's utility function. Then the first order and
second derivative analysis is used in the Stackelberg
game to find the optimal strategy _x of the leading *_
blockchain miner.
For the blockchain consensus node, based on
backward induction, the second derivative of _U with r_
respect to _[x]_ can be expressed as follows:
To prove the existence of extreme values of U, the r
concavity and convexity must be analyzed first.
Therefore, to further solve the first and second
derivatives of [] with respect to _[x]_, the following
expression can obtained:
(2xU)r2 = 2Ls 2 x + ( _x_ − _N cc_ ) (x2)2
(15)
* * − _m_ = 0,
_E T[_ _st_ ]
* 1 − * =0,
_NN D_
_c_
By analyzing the above equation, the first and second
2
x 0, (x)2 0
derivatives in Equation (15) can be
obtained. Finally, through the above analysis, it can be
_xmax_ = ( _x2−N c+c_ _v)3_ + ( _x_ −2+ _v)2_
obtained that when,
2xU2r 0 if _x_ xmin, _xmax_ , and the benefit function _U_ _r_
of blockchain miners is a convex function with respect to
_x ._
x = − ( _x_ −+ _v)2_
2 2
( _x)2_ = ( _x_ −+ _v)3_
*
_m_
− 0,
_E T[_ _st_ ]
1 − * 0,
_NN D_
_c_
* 0,* 0,* 0.
(16)
Ll (,, ) = 0
Let , then the optimal policy
blockchain users can be obtained.
* = _x_ − _v*_ +* − _W2−L NW Ts_ (c2a )
(12)
*
of
(13)
* _x v,_ - ,
It is noteworthy that is a function of,
_x v,_ - ,
which means that the corresponding is the
*
information necessary to get . In addition, the
-----
H. Zhou et al. / IJE TRANSACTIONS B: Applications Vol. 36, No. 08, (August 2023) 1468-1477 1473
U _r_ =2Ls + _x_ − _N cc_ = 0
Therefore, when x x , the
optimal strategy price _x of blockchain miners can be *_
obtained, namely:
_x*_ = (4− 4N c vc ) - − 4* + 4Nc* _c_ +2
+(2N cc − 2)v* + 2− 2Nc* _c_ − (2− 2N cc )−1
(17)
where, _x has a negative solution, which does not meet *_
the conditions and will not be discussed here. Meanwhile,
as can be seen from Equation (17); _x is a closed *_
,v*,*
expression related to . Therefore, to solve this
,v*,*
equation, the game strategies of both parties in
the previous round must be obtained first.
However, in a distributed environment, since the two
sides of the game are non-cooperative, neither the
blockchain miner nor the user knows the optimal strategy
of the other. Therefore, this paper uses the classical
iterative method [17] to find the optimal solution, and this
process is shown in Algorithm 1.
In Algorithm 1, if the iterative convergence condition
is not met, the value calculated in this round will be used
as the initial value for the next round of update, and this
process will be repeated until _[x][,][ ]converge._
The above analysis, on the basis of definition 1,
demonstrates that the optimal solution is the unique
equilibrium solution by proving 1 and 2.
Proof 1: For blockchain user, when the transaction
globally optimal. In particular, it is proved in section 3.1
that 2U2l 0, 2L2l = 2U2l 0 under KKT, so _Ll is a_
convex function with respect to [], which meets the
contents of Definition 1.
Proof 2: For blockchain miners, when the user gets
the ideal response time demand [( )], the optimal trading
strategy [] can be obtained. As proved in section 3.2,
under the condition [(]2xU)r2 0, _x is the optimal pricing *_
strategy that can maximize U . r
**4. PERFORMANCE EVALUATION**
In this paper, an incentive scheme based on Stackelberg
game is proposed for the network scenario involving
multiple roadside units and user nodes. The proposed
strategy is analyzed through the Matlab simulation
platform. The following will first explain the scenario
setting of simulation verification. The specific simulation
parameters are shown in Table 1.
In this section, the system performance is evaluated
numerically from three aspects. First, the update process
of blockchain user and miner policies with the number of
iterations is examined. Second, the influence of user
distribution on benefit function in the energy IoT
scenario is considered. Third, as the number of
blockchain miners increases, the change trend of the
benefit function is analyzed.
Miners, as leaders, first have the authority to
formulate pricing strategies. This is to update respective
strategies for following blockchain users on the basis of
miners' strategies to meet their own response time
requirements. Figure 2 represents the iterative update
process of transaction pricing for blockchain miners. In
this figure, transaction price decreases with an increase
in the number of iterations, which ultimately converges
to a stable value. This is because only when the
transaction price _[x]_ is lower, blockchain users will choose
**TABLE 1. Simulation Parameters of the Game Model**
**Parameter** **Value (range)**
DAG transaction broadcast delay D 1×10[-2]s
DAG verification threshold W 800
DAG transaction weight 3
Wireless transmission transaction threshold m 32
Algorithm convergence accuracy ε 10[-8]
Weight factor θ 1
Mining cost in transaction c 10[-2]
price _[x]_ is fixed,
* makes the user benefit function _U_ _l_
**Algorithm 1 Iterative update algorithm**
Input: initial value _x t_, _t_, J _t_ [, convergence accuracy] [][, ]
other parameter values of energy IoT.
Output: convergent _t x,_ - , - [. ]
1: The number of initialization iterations _t=0_ ; the flag bit
flag=flase, the initial value of _x, t_ U _r_ = _U_ _rt_ +1 −U _rt_ ,
denotes the convergence accuracy;
2：while (!flag)
3: The blockchain user gets _x from the blockchain miner t_
and updates it into t (xt ) ;
4. The blockchain miner obtains the updated t [ from the ]
DAG network and substitutes it into Equation (17);
5: Updatet, _t_ [ according to Equation (14); ]
6： if ( Ur )
7： flag=true；
8： _x*=,x t_ - = t ；
9： t=t+1；
10：endwhile；
11：return _t x,_ - , - [；]
-----
1474 H. Zhou et al. / IJE TRANSACTIONS B: Applications Vol. 36, No. 08, (August 2023) 1468-1477
**Figure 2. Update Process of Transaction Price Strategy with**
the Number of Iterations
to increase transaction arrival rate strategy []. Although
transaction price falls, a greater number of transactions in
the network will make miner's total revenue increase. In
addition, as the number of miners increases, so does the
ability to collect transactions in the network. Therefore,
despite the low transaction price, the miners' revenue can
still be guaranteed.
Figure 3 shows the change trend of the transaction
demand rate of blockchain users with the number of
iterations under different number of miners. Similarly, it
can be seen from the figure that with an increase in the
number of iterations, the transaction demand rate of
blockchain users increases and finally enters a stable
state. This is because as the number of miners increases,
the transaction price decreases, which exactly encourages
blockchain users to demand faster transaction rates.
Where, verification delay _Tv ( )_ is a function of ,
which represents the transaction verification delay of
blockchain users. As can be seen from Figure 4, with an
increase in the number of iterations, the value of _Tv ( )_
will gradually decrease, which is consistent with the
analysis result in Figure 3. Since _Tv ( )_ is inversely
proportional to [], when [] increases, the user’s
verification delay will decrease. Consequently, the
benefit function of the user is guaranteed, and eventually,
_Tv ( )_ will tend to a stable value.
Figures 5 and 6 show the impact of the number of miners
on the benefit functions of blockchain users and miners
themselves. According to the figure, as the number of
miners increases, the benefit function of blockchain users
and miners will also increase. This is because more
miners can process more transactions per unit time. That
is, the number of transactions participating in the
consensus process per unit time increases. This leads to
the decrease of verification delay, the improvement of
**Figure 3. Update process of demand strategy for transaction**
arrival rate with the number of iterations
**Figure 4. Update Process of Transaction Verification Delay**
with the Number of Iterations
**Figure 5. Benefit Function of Blockchain Users**
blockchain users' satisfaction, and an increase in
transaction demand. For blockchain miners, although the
-----
H. Zhou et al. / IJE TRANSACTIONS B: Applications Vol. 36, No. 08, (August 2023) 1468-1477 1475
transaction price is falling, more transactions in the
system can also ensure that miners gain decent revenue.
This paper also shows distribution comparison of four
groups of blockchain users in Figures 5 and 6. It can be
seen that, due to the limitation of wireless environment,
under greater blockchain user distribution area, that is,
greater [] value, the benefit function will be greater.
However, despite the continuous increase in []value, the
=0.15, =0.20
difference between the two curves in the
=0.05, =0.10
figure is obviously less than that of .
This is because the dense distribution of blockchain users
will lead to the continuous decline of transaction delivery
efficiency in the wireless environment. This will slow
down the growth in the number of transactions in the
network, reducing benefits for blockchain users and
miners.
Through the above simulation, it can be concluded
that the incentive mechanism proposed in this paper not
only encourages miners to join the blockchain network.
This increases the system stability and meets the response
time requirements of blockchain users. This is the
purpose of this algorithm, namely, not only guaranteeing
the interests of both parties of the game, but also
improving the distributed stability of the system.
**6. DISCUSSION**
The proposed incentive mechanism based on Stackelberg
game has numerically proved to be beneficial for both
blockchain users and miners. Simulation results have
shown that the proposed scheme can effectively protect
the interests of blockchain users and miners, and improve
the security and stability of the blockchain-based energy
IoT system. This conclusion is supported by the results
of several studies. For example, a survey conducted by
Liu et al. [8] on blockchain on the use of game theory to
**Figure 6. Benefit Function of Blockchain Miners**
analyze the incentives of different participants in an
energy blockchain system found that the incentive
mechanism proposed in their study was able to balance
the interests of energy producers, consumers, and miners.
Similarly, a study by Sun et al. [18] investigated on the
impact of game theory on the security of blockchainbased energy trading systems, and found that gametheoretic approaches can effectively enhance the security
of energy trading systems. Moreover, a study by Dong et
al. [19] on the use of game theory to optimize the
performance of blockchain-based energy trading systems
found that the game-theoretic approach can effectively
improve the performance of blockchain-based energy
trading systems. These studies all provide evidence that
the proposed incentive mechanism based on Stackelberg
game can protect the interests of blockchain users and
miners, and improve the security and stability of the
blockchain-based energy IoT system.
**6. CONCLUSION**
In this paper, the Stackelberg game is used to coordinate
the needs of blockchain users and miners. Blockchain
users can upload data to the DAG blockchain by paying
a fee to blockchain miners. Miners can gain revenue by
charging transaction fees. Through the game, on the one
hand, the revenue of the whole blockchain miners can be
guaranteed, and on the other hand, the response time
demand of blockchain users can be guaranteed. The
numerical results not only verify the model feasibility,
but also show that when there are many blockchain
miners, the model performance is fine, but when the
number of miners reaches a certain value, there will be
unobvious growth. Furthermore, the wireless energy IoT
environment can be confirmed that it will also create a
certain impact on the game model. The simulation results
also show that with an increase in the number of miners,
the benefit function of blockchain users and miners will
also increase. This is because more miners can process
more transactions per unit time. This can reduce
verification delay, improve blockchain users'
satisfaction, and an increase in the transaction demand.
For blockchain miners, although the transaction price is
falling, more transactions in the system can also ensure
that miners gain decent revenue. Overall, the results of
this study show that the proposed incentive scheme based
on the Stackelberg game model can effectively protect
the interests of blockchain users and miners, and improve
the security and stability of the blockchain-based energy
IoT system.
This research has several limitations. First, it only
focuses on the game model between blockchain users and
miners, and does not consider the impact of other factors
on the system performance. Second, the simulation
parameters are only applied in the energy IoT
-----
1476 H. Zhou et al. / IJE TRANSACTIONS B: Applications Vol. 36, No. 08, (August 2023) 1468-1477
environment. There is no discussion on the application of
the proposed model in other scenarios. Third, the game
model in this paper only considers the response time
requirements of blockchain users, and does not consider
the resource utilization efficiency of blockchain miners.
To further improve the system performance, there is still
a lot of work to be done in the future. First, the game
model should be extended to consider the resource
utilization efficiency of blockchain miners. Second, the
game model should consider the impact of other factors
on system performance such as network latency,
transaction broadcast delay, etc. Third, the application of
the proposed model should be further extended to other
scenarios. Finally, additional research should be done to
explore other incentive mechanisms for blockchain
networks.
**7. FUNDINGS**
The research is supported by: Basic and Advanced
Research Projects of CSTC (No. cstc2019jcyj
zdxmX0008); the Science and Technology Research
Program of Chongqing Municipal Education
Commission (No. KJZD-K201900605); The project of
Chongqing big data application and Development
Administration Bureau (No.22-30); The project of
Housing and Urban-rural Development Commission of
Chongqing Municipality (No.2021-0-104).
**8. REFERENCES**
1. Rifkin, J., "The third industrial revolution: How lateral power is
transforming energy, the economy, and the world, Macmillan,
(2011).
2. Nakamoto, S., "Bitcoin: A peer-to-peer electronic cash system",
**_Decentralized Business Review, (2008), 21260._**
3. Zhao, Y., Peng, K., Xu, B.-y. and Liu, Y., "Status and prospect of
pilot project of energy blockchain", **_Automation of Electric_**
**_Power Systems, Vol. 43, No. 7, (2019), 14-22._**
4. Zhang, N., Wang, Y., Kang, C., Cheng, J. and He, D.,
"Blockchain technique in the energy internet: Preliminary
research framework and typical applications", Proceedings of the
**_CSEE, Vol. 36, No. 15, (2016), 4011-4022._**
5. Fernández-Caramés, T.M. and Fraga-Lamas, P., "A review on the
use of blockchain for the internet of things", Ieee Access, Vol. 6,
(2018), 32979-33001. doi: 10.1109/ACCESS.2018.2842685.
6. Doshi, M. and Varghese, A., Smart agriculture using renewable
energy and ai-powered iot, in Ai, edge and iot-based smart
agriculture. 2022, Elsevier.205-225.
7. Wang, B. and Liu, F., "Task arrival based energy efficient
optimization in smart-iot data center", Mathematical Biosciences
**_and Engineering, Vol. 18, No. 3, (2021), 2713-2732. doi:_**
10.3934/mbe.2021138.
8. Liu, Z., Luong, N.C., Wang, W., Niyato, D., Wang, P., Liang, Y.C. and Kim, D.I., "A survey on blockchain: A game theoretical
perspective", **_IEEE Access, Vol. 7, (2019), 47615-47643. doi:_**
10.1109/ACCESS.2019.2909924.
9. Saad, W., Han, Z., Debbah, M., Hjorungnes, A. and Basar, T.,
"Coalitional game theory for communication networks", **_Ieee_**
**_Signal Processing Magazine, Vol. 26, No. 5, (2009), 77-97. doi:_**
10.1109/MSP.2009.000000.
10. Hakak, S., Khan, W.Z., Gilkar, G.A., Imran, M. and Guizani, N.,
"Securing smart cities through blockchain technology:
Architecture, requirements, and challenges", **_IEEE Network,_**
Vol. 34, No. 1, (2020), 8-14. doi: 10.1109/MNET.001.1900178.
11. Nejati, Z. and Faraji, A., "Actuator fault detection and isolation
for helicopter unmanned arial vehicle in the present of
disturbance", **_International_** **_Journal_** **_of_** **_Engineering,_**
**_Transactions C: Aspects, Vol. 34, No. 3, (2021), 676-681. doi:_**
10.5829/IJE.2021.34.03C.12.
12. Khosravian, E. and Maghsoudi, H., "Design of an intelligent
controller for station keeping, attitude control, and path tracking
of a quadrotor using recursive neural networks", **_International_**
**_Journal of Engineering,_** **_Transactions B: Applications, Vol. 32,_**
No. 5, (2019), 747-758. doi: 10.5829/ije.2019.32.05b.17.
13. Xiong, Z., Feng, S., Wang, W., Niyato, D., Wang, P. and Han, Z.,
"Cloud/fog computing resource management and pricing for
blockchain networks", IEEE Internet of Things Journal, Vol.
6, No. 3, (2018), 4585-4600. doi: 10.1109/JIOT.2018.2871706.
14. Wei, W., Liu, F. and Mei, S., "Energy pricing and dispatch for
smart grid retailers under demand response and market price
uncertainty", IEEE Transactions on Smart Grid, Vol. 6, No. 3,
(2014), 1364-1374. doi: 10.1109/TSG.2014.2376522.
15. Yang, D., Xue, G., Fang, X. and Tang, J., "Incentive mechanisms
for crowdsensing: Crowdsourcing with smartphones",
**_IEEE/ACM Transactions on Networking, Vol. 24, No. 3,_**
(2015), 1732-1744. doi: 10.1109/TNET.2015.2421897.
16. Hedges, J., "Backward induction for repeated games", arXiv
Preprint arXiv:1804.07074, (2018).
[https://doi.org/10.48550/arXiv.1804.07074](https://doi.org/10.48550/arXiv.1804.07074)
17. Cao, B., Xia, S., Han, J. and Li, Y., "A distributed game
methodology for crowdsensing in uncertain wireless scenario",
**_IEEE Transactions on Mobile Computing, Vol. 19, No. 1,_**
(2019), 15-28. doi: 10.1109/TMC.2019.2892953.
18. Sun, J., Wu, C. and Ye, J., "Blockchain-based automated
container cloud security enhancement system", in 2020 IEEE
international conference on smart cloud (SmartCloud), IEEE.
(2020), 1-6.
19. Dong, J., Song, C., Liu, S., Yin, H., Zheng, H. and Li, Y.,
"Decentralized peer-to-peer energy trading strategy in energy
blockchain environment: A game-theoretic approach", **_Applied_**
**_Energy,_** Vol. 325, (2022), 119852.
[https://doi.org/10.1016/j.apenergy.2022.119852](https://doi.org/10.1016/j.apenergy.2022.119852)
-----
H. Zhou et al. / IJE TRANSACTIONS B: Applications Vol. 36, No. 08, (August 2023) 1468-1477 1477
**COPYRIGHTS**
©2023 The author(s). This is an open access article distributed under the terms of the Creative Commons
Attribution (CC BY 4.0), which permits unrestricted use, distribution, and reproduction in any medium, as
long as the original authors and source are cited. No permission is required from the authors or the publishers .
Persian Abstract
**چکیده**
، به عنوان یک کاربرد معمولی فناوری اینترنت اشیا، به طور گسترده مورد مطالعه قرار گرفته است. در همین حال، فناوری(IoT) در عصر اینترنت همه چیز، اینترنت اشیا انرژی
.بالک چین و انرژی اینترنت اشیا می توانند هماهنگ و مکمل یکدیگر باشند. انرژی اینترنت اشیا متنوع است و تقاضای تراکنش باالیی دارد بحث در مورد تأثیر م حیط اینترنت
اشیا انرژی بر عملکرد الگوریتم های اجماع بالک چین و تضمین ثبات بالک چین در محیط اینترنت اشیا انرژی، موضوعی است که ارزش تحقیق د ارد. در این تحقیق، یک
ها ی کاربر پیشنهاد شده است. استراتژی پیشنهادی از طریق پلتای و گره برای سناریوی شبکه شامل چندین واحد کنار جادهStackelberg مکانیسم انگیزشی مبتنی بر بازی
تحلیل می شود. نتایج شبیه سازی نشان می دهد که طرح پیشنهادی می تواند به طور موثر از منافع کاربران بالک چین و ماینرها محافظت کند . همچنینMatlab فرم شبیه سازی
سنجی مد ل را تأیید می کنند. همچنین نشان می دهد کهمی تواند امنیت و ثبات سیستم اینترنت اشیاء مبتنی بر بالک چین را بهبود بخشد. عالوه بر این، نتایج عددی نه تنها امکان
وقتی ماینرهای بالک چین زیادی وجود دارد، عملکرد مدل خوب است. با این حال، زمانی که تعداد ماینرها به مقدار مشخصی برسد، رشد نامشخ .صی وجود خواهد داشت
عالوه بر این، همچنین تایید شده است که محیط اینترنت اشیا انرژی بی سیم .نیز تاثیر خاصی بر مدل بازی ایجاد خواهد کرد
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.5829/ije.2023.36.08b.07?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.5829/ije.2023.36.08b.07, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "GOLD",
"url": "https://www.ije.ir/article_170108_0606d7f89e99e10eed7dcd50415a9b55.pdf"
}
| 2,023
|
[
"JournalArticle"
] | true
| null |
[
{
"paperId": "5044316f1a8510873c9c3ce2bb53f2fdd6102854",
"title": "Decentralized peer-to-peer energy trading strategy in energy blockchain environment: A game-theoretic approach"
},
{
"paperId": "b16153b44542775da53036b811ac829405b3bcb4",
"title": "Task arrival based energy efficient optimization in smart-IoT data center."
},
{
"paperId": "f66b0a487551452aa73def8ca3e43d5775dc7245",
"title": "Actuator Fault Detection and Isolation for Helicopter Unmanned Arial Vehicle in the Present of Disturbance"
},
{
"paperId": "58de6c3adb6f303ceb691c8a99a3c1aa6c410511",
"title": "Blockchain-based Automated Container Cloud Security Enhancement System"
},
{
"paperId": "a59b3a1ff98afe3f98ec89fb3027a9e8bcf62866",
"title": "A Distributed Game Methodology for Crowdsensing in Uncertain Wireless Scenario"
},
{
"paperId": "ee59451deedd0c7532b5b4c34340a615c1d76312",
"title": "Securing Smart Cities through Blockchain Technology: Architecture, Requirements, and Challenges"
},
{
"paperId": "00ed0d22e38f54ed3e0e5c2a73b582ffda55ab3e",
"title": "Design of an Intelligent Controller for Station Keeping, Attitude Control, and Path Tracking of a Quadrotor Using Recursive Neural Networks"
},
{
"paperId": "7ff71f9b85ca18554d56228edb7271d439214707",
"title": "A Survey on Blockchain: A Game Theoretical Perspective"
},
{
"paperId": "02458904f9bd718bd8c6a1a36e9847ad83b0410b",
"title": "A Review on the Use of Blockchain for the Internet of Things"
},
{
"paperId": "e173f6a03d74ca2192d6cbddd1a09632e882d41d",
"title": "Backward induction for repeated games"
},
{
"paperId": "b3f538f13b441c969a998350e63efaded86223e6",
"title": "Cloud/Fog Computing Resource Management and Pricing for Blockchain Networks"
},
{
"paperId": "d8403984bff42604cce78876efb31c11321bfed8",
"title": "Incentive Mechanisms for Crowdsensing: Crowdsourcing With Smartphones"
},
{
"paperId": "acade013bcbcadcafc0498c0a5ce97facb600804",
"title": "Energy Pricing and Dispatch for Smart Grid Retailers Under Demand Response and Market Price Uncertainty"
},
{
"paperId": null,
"title": "A game-theoretic approach\", Applied Energy, Vol"
},
{
"paperId": "f8458ab0d8988bcd7a74c63992f358d64a49babf",
"title": "Smart agriculture using renewable energy and AI-powered IoT"
},
{
"paperId": null,
"title": "Status and prospect of pilot project of energy blockchain"
},
{
"paperId": "18cf728c1f9e9ad19d1354add882c1da48f7291a",
"title": "Blockchain Technique in the Energy Internet: Preliminary Research Framework and Typical Applications"
},
{
"paperId": "b10f53fbb737f5a244202eda8a2ab20d4b3336f2",
"title": "The third industrial revolution : how lateral power is transforming energy, the economy, and the world"
},
{
"paperId": "4e9ec92a90c5d571d2f1d496f8df01f0a8f38596",
"title": "Bitcoin: A Peer-to-Peer Electronic Cash System"
},
{
"paperId": "41bab193bd1d5432d3784d077d3ef3f078168dac",
"title": "IEEE Signal Processing Magazine, Special Issue on Game Theory, to appear, 2009 Coalitional Game Theory for Communication Networks: A Tutorial"
}
] | 12,802
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Medicine",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/0205b4b033b326a2bfc28ec21fccbd91df47ab2f
|
[
"Computer Science"
] | 0.862838
|
Using Visualization to Build Transparency in a Healthcare Blockchain Application
|
0205b4b033b326a2bfc28ec21fccbd91df47ab2f
|
Sustainability
|
[
{
"authorId": "145467727",
"name": "Jesús Peral"
},
{
"authorId": "2056867384",
"name": "E. Gallego"
},
{
"authorId": "2060054848",
"name": "D. Gil"
},
{
"authorId": "1413908809",
"name": "M. Tanniru"
},
{
"authorId": "2368536",
"name": "P. Khambekar"
}
] |
{
"alternate_issns": null,
"alternate_names": null,
"alternate_urls": [
"http://mdpi.com/journal/sustainability",
"http://nbn-resolving.de/urn/resolver.pl?urn=urn:nbn:ch:bel-172127"
],
"id": "8775599f-4f9a-45f0-900e-7f4de68e6843",
"issn": "2071-1050",
"name": "Sustainability",
"type": "journal",
"url": "http://www.e-helvetica.nb.admin.ch/directAccess?callnumber=bel-172127"
}
|
With patients demanding services to control their own health conditions, hospitals are looking to build agility in delivering care by extending their reach into patient and partner ecosystems and sharing relevant patient data to support care continuity. However, sharing patient data with several external stakeholders outside a hospital network calls for the development of a digital platform that is trusted by both hospitals and stakeholders, given that there is often no single entity supporting such coordination. In this paper, we propose a methodology that uses a blockchain architecture to address the technical challenge of linking disparate systems used by multiple stakeholders and the social challenge of engendering trust by using visualization to bring about transparency in the way in which data are shared. We illustrate this methodology using a pilot implementation. The paper concludes with a discussion and directions for future research and makes some concluding comments.
|
## sustainability
_Article_
# Using Visualization to Build Transparency in a Healthcare Blockchain Application
**Jesús Peral** **[1,]*, Eduardo Gallego** **[1,2], David Gil** **[2]** **, Mohan Tanniru** **[3]** **and Prashant Khambekar** **[4]**
1 Department of Software and Computing Systems, University of Alicante,
03690 Alicante, Spain; ejgl2@alu.ua.es
2 Department of Computer Technology and Computation, University of Alicante,
03690 Alicante, Spain; dgil@dtic.ua.es
3 College of Public Health, University of Arizona, Phoenix, AZ 85006, USA; tanniru@oakland.edu
4 Harbinger Systems, Philadelphia, PA 19103, USA; Prashant.Khambekar@harbingergroup.com
***** Correspondence: jperal@dlsi.ua.es; Tel.: +34-96-590-3772
Received: 12 July 2020; Accepted: 17 August 2020; Published: 20 August 2020
[����������](https://www.mdpi.com/2071-1050/12/17/6768?type=check_update&version=2)
**�������**
**Abstract: With patients demanding services to control their own health conditions, hospitals are**
looking to build agility in delivering care by extending their reach into patient and partner ecosystems
and sharing relevant patient data to support care continuity. However, sharing patient data with
several external stakeholders outside a hospital network calls for the development of a digital platform
that is trusted by both hospitals and stakeholders, given that there is often no single entity supporting
such coordination. In this paper, we propose a methodology that uses a blockchain architecture to
address the technical challenge of linking disparate systems used by multiple stakeholders and the
social challenge of engendering trust by using visualization to bring about transparency in the way
in which data are shared. We illustrate this methodology using a pilot implementation. The paper
concludes with a discussion and directions for future research and makes some concluding comments.
**Keywords: blockchain; IoT; secure transaction; health; file sharing; visualization**
**1. Introduction**
In today’s digital age, advanced technologies are continually altering customer expectations of
services delivered and requiring that organizations build “agility” within their internal operations
by using an agile organizational model of structure and governance [1]. The agile model supports
the exploration of innovative service value propositions and the use of a mix of internal and external
resources to evaluate these innovations to fulfill customer value [2,3]. Such a model is also used to
support evaluation, adaptation, and learning to improve organizational capacity to sustain value as
customer expectations change [4,5]. One can argue that “agile” organizations are indeed sustainable
organizations, as they continue to meet the current needs of customers by using external resources and
conserve their own resources to address future customer needs. In this paper, we focus on hospitals
that are responsible for supporting continuity of care for patients outside the hospital.
Hospitals are extending patient care using several care facilities (e.g., urgent care facilities,
ambulatory care facilities, etc.) [6,7] and are helping patients self-manage their care using multiple
technologies [8–11]. This calls for hospitals to build agility to leverage the resources of external partners
and motivate patients to self-manage their health in a tightly regulated and resource constrained
environment. This means that patient data are generated by multiple stakeholder systems (partners
and patients) that use several advanced technologies, such as internet of things (IoT), mobile apps,
digital exchanges, and social media, and such data have to be understood, collected, integrated,
and shared by all involved in the support of patient care. Unless there is a public health crisis
-----
_Sustainability 2020, 12, 6768_ 2 of 20
(e.g., COVID-19) that calls for public health agencies to coordinate significant disruptions to economic,
health, and social conditions [12], opioid addiction that calls for tracking drug distribution [13],
or chronic care management of high risk patients to reduce hospital readmissions [14–16], there is little
incentive for hospitals to coordinate patient data sharing outside their hospital networks. This calls for
a distributed digital platform that is either coordinated by a trusted third party or an architecture that
ensures trust for everyone to contribute and use the data shared.
Blockchain technology has been suggested in prior research as a platform when there is no trusted
coordinator to support data sharing. It supports peer-to-peer connectivity among various stakeholders
using agreed upon protocols about who can participate in such data sharing. Using characteristics
such as immutability and auditability, it is considered a viable and trusted platform to share data when
there is no central entity coordinating such sharing activities. In healthcare, establishing trust is both a
technical challenge (i.e., ensuring the integrity of data shared by multiple stakeholder systems and
making it available for impact on care) and a social challenge (i.e., ensuring transparency to engender
confidence that the mechanism used to share data addresses confidentiality). For example, a system
that monitors patients’ vital signs and uses an algorithm to generate a metric used to track patient
conditions has to be trusted for its integrity. Patients’ questions sent to peers and clinicians may
be anonymized to share with peers for comment to ensure confidentiality and identified and made
available to patients quickly for treatment adaptation. This paper proposes a methodology that uses
blockchain technology as a digital platform with a visualization feature added to address both the
technical and social challenges.
This paper is organized as follows. Section 2 provides prior research on the use of blockchain in
multiple domains as well as in healthcare. Section 3 discusses the methodology that creates visibility
for both the creation of the data and its movement within the network to address the challenges.
A case study to illustrate a pilot implementation is explained in Section 4. Section 5 discusses,
in detail, an implementation methodology, and Section 6 includes discussion, future research directions,
and limitations. Section 7 provides some concluding comments.
**2. Background**
Blockchain applications can be categorized by domain-financial or non-financial [17],
since cryptocurrencies represent many but not all of the applications using blockchain technology.
These applications can also be classified by the version of technology used (i.e., 1.0, 2.0 and 3.0) [18,19].
Along application domains, they can classified by application type (e.g., financial, healthcare, business
and industrial, education, etc.), business issue focus (e.g., governance, privacy and security, etc.),
or technical issue focus (e.g., integrity verification, IoT, data management, etc.) [20]. Application of
blockchain in healthcare has been more recent [21,22], and, as discussed earlier, trust in sharing sensitive
healthcare information among several actors outside a hospital system has been a challenge [23].
However, the mechanisms embedded in the distributed ledger technology associated with blockchain
technology may be able to address this challenge [21,22,24–26]. In other words, if healthcare
organizations are to become agile in meeting patient needs outside a hospital, the digital platform has
to address some of the technical challenges, such as ensuring security, interoperability, data sharing,
and mobility, if it is to engender trust [23]. Let us take each of these in more detail.
(1) Security
Existing methods used to protect and secure patient medical records have not been effective [27,28].
While access controls and authentication of records are widely used in ensure integrity, confidentiality,
and accessible of medication information [26,29,30], their implementation becomes a challenge once
systems are extended outside a hospital [31,32]. The encryption of data among Electronic Medical
Record (EMR) and stakeholder systems is useful, but this leads to problems when there are many
different encryption standards [33,34]. With no single technology platform addressing the security
challenges [35], a distributed platform that allows local control of the data at each node but ensures
-----
_Sustainability 2020, 12, 6768_ 3 of 20
security as it moves across a distributed platform may be a solution. Blockchain technology, which has a
uniform method to encrypt the data transferred, public–private keys for the authentication of users who
transfer the data, and validation of those who decrypt the data for use, can be effective in addressing
security when data are shared by several stakeholders [20,26,36,37].
(2) Interoperability
Sharing data among multiple stakeholder systems, such as apps or intelligent agents, or multiple
people, such as messages sent using mobile phones, requires having a uniform method to collect
disparate sources of data and a centralized database for all to share. With no single entity coordinating
such a shared database, a blockchain architecture can allow each partner to upload data for sharing and
use using certain agreed upon protocols about who can contribute and access data, with embedded
security, controlled redundancy, and auditability [23].
(3) Data Sharing
Data sharing in healthcare is critical, as patient care is remotely managed at various locations
(at home, at partner sites, or at hospitals) and must be shared with others to support continuity of care.
Moreover, the data gathered at each site may be in a different form [24,33,38]. Blockchain technology
allows for each partner connected to the network to share data either directly or indirectly using a
secure link. In some cases, data are stored elsewhere (e.g., when the data are large, as with image
scans, or in narrative form, as with doctors’ notes), and associated links can be used for data access.
In summary, blockchain technology allows the sharing of multiple data types without forcing a single
data normalization method.
(4) Mobility (IoMT)
As patients become mobile and must access their data when and where they need it, its portability
is critical. With more devices such as smart phones and sensors (IoT) connected to the Internet, data are
collected from these devices [39–41] have to be effectively integrated. This concept is often referred
to as digital mobility or Internet of Medical Things (IoMT) [42–44]. With a blockchain’s ability to
connect with any partner (human or machine) with permission to share data with others, such mobility
is feasible.
_2.1. Blockchain in Healthcare_
Blockchain technology has begun to see applications that extend care to patients outside a hospital.
Traditionally, EMRs are used to manage patient data within a hospital system, and their use has
grown significantly [10,45,46]. However, as hospitals try to extend care to patients outside the hospital,
and with partners and patients using a myriad of systems, the challenge is one of interoperability.
Blockchains can provide a gateway for data sharing among these systems by addressing the four key
areas of importance discussed above: security, interoperability, data sharing, and mobility. For example,
OmniPHR (Omnipresent Personal Health Record) has been proposed as a distributed model to
integrate personal health records for patients and hospitals to access and use [38], and MedRec
(decentralized record management system to handle EMRs) is being developed as a component of
a hospital EMR system [24]. A framework for EMR data sharing for cancer patients is proposed by
Dubovitskaya et al. [47], and a decentralized platform that provides a secure, fast, and transparent
exchange of a single version of a patient’s data are provided by Medicalchain [48].
Other applications include HealthChain, which leverages blockchain technology to support
the sharing of patients’ medical data [49,50]. MediBchain is another patient centric healthcare data
management system that enables patient data sharing using cryptographic techniques [51]. Borioli and
Couturier [52] discuss the potential of blockchain to conduct clinical trials using smart contracts,
and Mamoshina et al. [53] propose a roadmap for decentralizing the personal health data ecosystem for
drug discovery, biomarker development, and preventative healthcare. The use of microscopy sensors
-----
_Sustainability 2020, 12, 6768_ 4 of 20
that take an image of fingernails for identity authentication was proposed by Lee at al. [54] to protect
data privacy, and an Ethereum protocol that remotely monitors and manages patients using data from
sensors and smart devices and smart contracts was presented by Griggs et al. [27]. MeDShare (a system
that addresses the issue of medical data sharing) is a blockchain-based system that is used to provide
data provenance, auditing, and control of shared medical data in cloud repositories and to monitor
malicious use of these data [28].
The goal of all these applications is to support operational continuity as care is extended outside a
hospital so that patient data can be accessed by doctors, hospitals, laboratories, pharmacists, insurers,
etc., and strategic support (e.g., analysis for treatment adherence to change diagnoses or treatment
plans, at an individual level over time or at an aggregate level for discovering patterns, possibly using
big data analytics). To address these two types of support, one may consider two different blockchain
architectures: one blockchain with parallel computing capabilities and big data analytics for strategic
support, and another blockchain to support operational continuity that includes data integration,
secure identity management, and a trust supporting data sharing component [55]. Each of these
blockchains still leverages the blockchain properties of authentication, confidentiality, accountability,
and data sharing among those using the networks. In other words, operational continuity leads to
data collection (or surveillance of patient–partner activities), and strategy support is used to leverage
these data for analysis and to refine care processes.
_2.2. Increasing Trust through Visibility_
While the discussion thus far demonstrates the role of blockchain technology in addressing a
number of technical challenges to ensure trust in the way data are collected from disparate systems and
shared to ensure integrity and confidentiality, there is still the issue of the social challenge: Will those who
have to adopt the system trust the system enough to contribute to it? Transparency through visualization
to enhance trust has been discussed in the literature. For example, transparency of the supply chain
is viewed as critical to engender trust among the participating stakeholders [56], and visualization
is often used to communicate information to groups with varying technical backgrounds, especially
when there are opportunities for misrepresentation of the data [57]. In some cases, interactive graphics
are used to make static reports dynamic, so that individuals can understand the data by seeing such
data at various levels of granularity [58]. Dashboards with drill-down capabilities have been used
by many organizations to improve both transparency and accountability, especially when clinical
decisions and administrative decisions lead to conflicts [59].
Visualizations has also been used to debug software and help with understanding the reasoning
processes of forward-chaining rule-based expert systems [60], as well as when individuals are engaged
in global software development to ensure that workflows that are generating data to influence a
project can be monitored [61]. Today, when data are manipulated by multiple entities including robots,
designing human-like and visualization-based transparency is critical to map the processes used to
manipulate data so it can match an individual’s mental models [62] and reduce the cognitive burden by
helping with external anchoring, information foraging, and cognitive offloading [63]. The methodology
discussed here uses “visualization” to improve the trustworthiness of those sharing the data using the
blockchain architecture, thus addressing both technical and social challenges.
**3. The Proposed Methodology: A Blockchain-Based Solution**
In this section, we present our general methodology where a blockchain architecture is used
to visually show how data are shared by users as it moves among various nodes in the network.
The architecture uses two web applications: one to create the data for the blockchain and the other to
visualize the network to improve transparency and build trust. The application supports the sharing of
data files (PDF, text, images, etc.) between different nodes, so that a user will have the ability to visually
see the files as they are sent and received, ensuring the existence, order, and immutability of these files.
Specifically, we will illustrate the process used when permission is granted for some data by the patient
-----
_Sustainability 2020, 12, 6768_ 5 of 20
and the subsequent movement of these data along the network to support transparency. To achieve
the stated objectives, the methodology uses two features: blockchain technology and visualization
techniques. This methodology is technology agnostic, i.e., different blockchain technologies can be
used for application implementation. The methodology can be summarized as follows:
_Sustainability 2020 FOR PEER REVIEW_ 5 of 20
(1) Create the blockchain with the different network nodes, where each node corresponds to different
1) Create the blockchain with the different network nodes, where each node corresponds to users who will participate in data sharing. In our case study the nodes correspond to patients
different users who will participate in data sharing. In our case study the nodes correspond to who decide to share their files as well as the buyers of information from these files.
patients who decide to share their files as well as the buyers of information from these files.
(2) Manage the transactions generated by different nodes. Here, we will focus on authentication, file
2) Manage the transactions generated by different nodes. Here, we will focus on authentication,
transfer, and visualization. These transactions are combined with other transactions to create a
file transfer, and visualization. These transactions are combined with other transactions to create
new block.
a new block.
(3) Configure and customize the information to be visualized after choosing a tool for the
3) Configure and customize the information to be visualized after choosing a tool for the network
network visualization.
visualization.
(4) Connect or integrate the blockchain with the visualization tool.
4) Connect or integrate the blockchain with the visualization tool.
(5)5) Demonstrate the visualization of how nodes are interacting during a transaction. Demonstrate the visualization of how nodes are interacting during a transaction.
Figure 1 shows at a high level how the transactions are managed within the blockchain network.
Figure 1 shows at a high level how the transactions are managed within the blockchain network.
**Figure 1. Methodology for transactions management within the blockchain network.**
**Figure 1. Methodology for transactions management within the blockchain network.**
**Figure 1. Methodology for transactions management within the blockchain network.**
Figure 2 shows the basic blockchain structure. A blockchain is a data structure in which the
information contained is grouped into sets of blocks. Each block has information on the previous block,Figure 2 shows the basic blockchain structure. A blockchain is a data structure in which the
and, using cryptographic techniques, this information can only be repudiated or edited by modifyinginformation contained is grouped into sets of blocks. Each block has information on the previous
all subsequent blocks. The information stored in each block includes: (a) records or transactions,block, and, using cryptographic techniques, this information can only be repudiated or edited by
(b) information about the block, and (c) a link to the previous block through a digital signature (hash).modifying all subsequent blocks. The information stored in each block includes: (a) records or
Each block has a specific and unmovable place within the chain, since each block contains informationtransactions, (b) information about the block, and (c) a link to the previous block through a digital
on the previous block as a hash. The entire chain is stored at each of the nodes that make up thesignature (hash). Each block has a specific and unmovable place within the chain, since each block
network, so that all network participants have an exact copy of it. When a new record is created,contains information on the previous block as a hash. The entire chain is stored at each of the nodes
it is verified and validated by all the nodes that form the network and then added to a new blockthat make up the network, so that all network participants have an exact copy of it. When a new
record is created, it is verified and validated by all the nodes that form the network and then added
and linked to the chain. Each node uses different types of certificates and digital signatures to verify
to a new block and linked to the chain. Each node uses different types of certificates and digital
information, as well as to validate transactions and data stored in the blockchain.
signatures to verify information, as well as to validate transactions and data stored in the blockchain.
-----
record is created, it is verified and validated by all the nodes that form the network and then added
_Sustainability 2020, 12, 6768_ 6 of 20
to a new block and linked to the chain. Each node uses different types of certificates and digital
signatures to verify information, as well as to validate transactions and data stored in the blockchain.
**Figure 2.Figure 2. Basic structure of a blockchain.Basic structure of a blockchain.**
Despite having introduced in this section the generic methodology based in blockchain, not all
environments or organizations may use this methodology shown here. As Hebert et al. [64] point
out, varying levels of security threats specific to a blockchain may call for an integrated multi-staged
architecture. For the healthcare application chosen here, where a patient stores his or her data
and provides access to these data to others, the methodology used here is considered appropriate.
The data shared is tamper-proof because of the immutability property, and participants of the network
are a-priori authenticated as trusted partners to share data using the network (i.e., permissioned
blockchain). The application also uses the proposed transparency feature to enable users to see who
accessed the data, and the blockchain keeps a record of every time data are accessed with a time stamp,
thus providing an audit trail. The following section will discuss the application.
**4. Case Study**
The case study presented here allows patients to share their health data, including diagnoses
and treatments, and gives research organizations access to these data for payment. The transparency
of access and its purpose ensures that payments are made for the right purpose and accurately,
while protecting patients’ rights over their data. The payments increase patients’ willingness to share
their data for research purposes, and research institutions will benefit by paying a small amount to
gather a large amount of patient data to support analysis. While such a payment of small amounts to
patients may be viewed as too complicated [65], research organizations today spend large sums of
money to solicit patient participation in clinical trials and a large proportion of this money goes to
intermediaries. Within a blockchain, the patient has more control over their data and can monetize the
data by selling it directly to potential buyers.
As the blockchain can track every access, the payment can be coupled with access, thus leading to
immediacy and accuracy. Such transparency can lead to increased patient participation and improve
the quality of clinical trials. Some systems have used token mechanisms for payments and several
blockchains have their own crypto tokens. However, the tradability of tokens with fiat currency,
liquidity, and the handling of inflationary pressures makes their use complicated [66]. Therefore,
the system described here uses fiat currency (US dollars), thus creating a determined value for
each transaction.
_Processes for Uploading and Accessing Health Data_
The main roles in the blockchain-based system are patients, caregivers, and buyers. The caregivers
monitor patients for care continuity (e.g., doctors, external care providers, family members, etc.).
The case study here, however, will focus on the interaction between patients and buyers, where patients
upload health data and allow buyers to download and read the data after purchasing it. The patient
data can be in the form of a Continuity of Care Document (CCD) or Fast Healthcare Interoperability
Resources (FHIR). Patients get these data either from hospitals and clinics or can upload it from their
-----
improve the quality of clinical trials. Some systems have used token mechanisms for payments and
_Sustainability 2020, 12, 6768_ 7 of 20
several blockchains have their own crypto tokens. However, the tradability of tokens with fiat
currency, liquidity, and the handling of inflationary pressures makes their use complicated [66].
own devices (e.g., via Fitbit devices). These data can either relate to a patient visit to a care centerTherefore, the system described here uses fiat currency (US dollars), thus creating a determined value
(encounter) or an episode related to his healthfor each transaction. /wellness.
The data are uploaded one record at a time by the system front-end and is stored in the blockchain.
_Processes for Uploading and Accessing Health Data_
The metadata about that data are stored in local storage, and this can include the nature of the data
uploaded. Depending on the size of the data, this can take a couple of minutes. The patient is informedThe main roles in the blockchain-based system are patients, caregivers, and buyers. The
once the data are uploaded and stored. This is shown in Figurecaregivers monitor patients for care continuity (e.g., doctors, external care providers, family 3.
members, etc.). The case study here, however, will focus on the interaction between patients and Patients publish the names of the files they wanted to share. When a buyer wants to purchase
data, they are shown dibuyers, where patients upload health data and allow buyers to download and read the data after fferent types of data and the corresponding information (e.g., time range).
Subsequently, once the buyer decides to purchase some data, the system determines the owner of thepurchasing it. The patient data can be in the form of a Continuity of Care Document (CCD) or Fast
Healthcare Interoperability Resources (FHIR). Patients get these data either from hospitals and clinics
data and checks whether permission was provided. If permission was not already provided, the system
or can upload it from their own devices (e.g., via Fitbit devices). These data can either relate to a
informs the patient of the buyer request and the incentive offered by the buyer. If the patient provides
patient visit to a care center (encounter) or an episode related to his health/wellness.
permission, then the system stores the permission (for one patient, one buyer, and one piece of data) in
The data are uploaded one record at a time by the system front-end and is stored in the
the blockchain. It notifies the buyer of the permission, so the buyer can request the data to be read.Sustainability blockchain. The metadata about that data are stored in local storage, and this can include the nature 2020 FOR PEER REVIEW 7 of 20
The system stores the data access and deducts the payment from the buyer and credits it to the patient.
of the data uploaded. Depending on the size of the data, this can take a couple of minutes. The patient
The payments made are accumulated with each buyer read. These interactions are shown in FigurePatients publish the names of the files they wanted to share. When a buyer wants to purchase 4.
is informed once the data are uploaded and stored. This is shown in Figure 3.
data, they are shown different types of data and the corresponding information (e.g., time range).
Subsequently, once the buyer decides to purchase some data, the system determines the owner of the
data and checks whether permission was provided. If permission was not already provided, the
system informs the patient of the buyer request and the incentive offered by the buyer. If the patient
provides permission, then the system stores the permission (for one patient, one buyer, and one piece
of data) in the blockchain. It notifies the buyer of the permission, so the buyer can request the data to
be read. The system stores the data access and deducts the payment from the buyer and credits it to
the patient. The payments made are accumulated with each buyer read. These interactions are shown
in Figure 4. **Figure 3. Process for uploading of health data by patient.**
**Figure 3. Process for uploading of health data by patient.**
**Figure 4. Process for access to health data by buyer.**
**Figure 4. Process for access to health data by buyer.**
The blockchain keeps health data, permission data, and the monetary amounts belonging to both
The blockchain keeps health data, permission data, and the monetary amounts belonging to both
the patient and the buyer. Based on the size of each block, these patient–buyer transactions can be
the patient and the buyer. Based on the size of each block, these patient–buyer transactions can be
spread across many blocks, as shown in Figure 5. The next section will discuss the implementation.
spread across many blocks, as shown in Figure 5. The next section will discuss the implementation.
Subsequently, once the buyer decides to purchase some data, the system determines the owner of the
data and checks whether permission was provided. If permission was not already provided, the
system informs the patient of the buyer request and the incentive offered by the buyer. If the patient
provides permission, then the system stores the permission (for one patient, one buyer, and one piece
of data) in the blockchain. It notifies the buyer of the permission, so the buyer can request the data to
be read. The system stores the data access and deducts the payment from the buyer and credits it to
the patient. The payments made are accumulated with each buyer read. These interactions are shown
-----
The blockchain keeps health data, permission data, and the monetary amounts belonging to both
_Sustainability 2020, 12, 6768_ 8 of 20
the patient and the buyer. Based on the size of each block, these patient–buyer transactions can be
spread across many blocks, as shown in Figure 5. The next section will discuss the implementation.
**Figure 5. Figure 5.Patient’s data and information across blocks. Patient’s data and information across blocks.**
**5. Implementation5. Implementation**
In this section, we will discuss the implementation of the case. It is assumed that permission isIn this section, we will discuss the implementation of the case. It is assumed that permission is
granted by the patient, his or her data are sent to the buyer, and these transactions are tracked.granted by the patient, his or her data are sent to the buyer, and these transactions are tracked.
The technologies considered for the development of this application were Corda R3, HyperledgerThe technologies considered for the development of this application were Corda R3,
Fabric and Ethereum. After studying these three technologies, Hyperledger Fabric [Hyperledger Fabric and Ethereum. After studying these three technologies, Hyperledger Fabric [67] 67] was chosen for
its robustness and the privacy it offers for the stored information compared to several competitors.
It is also configurable, guarantees security, interoperability, and data sharing. Inside the Hyperledger
family, there is also Hyperledger Sawtooth with a different consensus algorithm and a different mode
of execution. For the purpose of this study, Hyperledger Fabric was chosen because its Explorer is
much easier to use than the Explorer that comes with Hyperledger Sawtooth. The main challenge in
the implementation was the integration of the different used technologies such as Hyperledger Fabric,
Hyperledger Explorer or Vue.js; implementation details are shown in the following subsections.
_5.1. Blockchain Creation_
Each node in the network (associated with the users: patient, buyer, etc.) will be created using
Vue.js. Different templates will be created for viewing files, sending files, and support authentication.
When the permission is provided by the patient, the corresponding transaction and the subsequent
block is created.
The first step calls for the downloading the blockchain platform using the latest version of
Hyperledger Fabric from the official repository, unzip it and access the first-network folder to check
accuracy of the download. Once in the folder and is running correctly, the message shown in Figure A1
will be displayed.
_5.2. Transaction Management: Patient Permission, File Transmission and Block Creation_
Once the permission is granted by the patient to share his or her data, the application will check
that the recipient is in the system and the file is in the right format. Then, the file will be encrypted in
base64. Base64 is a method of encoding and decoding binary data (e.g., HTML, CSS, text documents or
images) [68]. After encrypting the file, the Application Programming Interface (API) endpoint will be
called to upload the file to the blockchain. Subsequently, a “json” file with the user’s credentials and
the encrypted document is sent to the API. This information becomes part of the transaction and will
be converted into a block, as shown in Figure 6.
-----
be called to upload the file to the blockchain. Subsequently, a “json” file with the user’s credentials
_Sustainability 2020, 12, 6768_ 9 of 20
and the encrypted document is sent to the API. This information becomes part of the transaction and
will be converted into a block, as shown in Figure 6.
**Figure 6.Figure 6. Description of information block.Description of information block.**
_Sustainability 2020 FOR PEER REVIEW_ 9 of 20
_5.3. File Reception_
_5.3. File Reception_
In this step, the receiver (the buyer) can download the shared files. When the receiver logs on
In this step, the receiver (the buyer) can download the shared files. When the receiver logs on to
to the home page and clicks “View my received documents”, a screen (as shown in Figure 7) will
the home page and clicks “View my received documents”, a screen (as shown in Figure 7) will appear.
appear. The recipient user will be able to download the documents needed, and these are ordered from
The recipient user will be able to download the documents needed, and these are ordered from the
the most recent to the oldest, showing the sender, the send date, and the ID of the sender. When the
most recent to the oldest, showing the sender, the send date, and the ID of the sender. When the
receiver clicks on “Download”, the file will be decrypted using base64 and then downloaded.
receiver clicks on “Download”, the file will be decrypted using base64 and then downloaded.
**Figure 7. Figure 7.Received files. Received files.**
_5.4. Visualization Configuration and Connection with the Blockchain Network_
_5.4. Visualization Configuration and Connection with the Blockchain Network_
Hyperledger Explorer will be used for the display of the network using React.js. It offers default
Hyperledger Explorer will be used for the display of the network using React.js. It offers default
templates ready to be launched or edited, and it provides several graphics to customize the templates
templates ready to be launched or edited, and it provides several graphics to customize the templates
for visualization. Such a method of sharing documents and using visualization to track its flow isfor visualization. Such a method of sharing documents and using visualization to track its flow is
useful in healthcare to build transparency and gain the trust of all actors involved. There are potentiallyuseful in healthcare to build transparency and gain the trust of all actors involved. There are
other applications where such transparency is needed to ensure user adoption of blockchain technologypotentially other applications where such transparency is needed to ensure user adoption of
blockchain technology for sharing data. The rest of the section will discuss some of the
implementation details such as the installation, configuration, and visualization of Hyperledger
Explorer.
-----
_Sustainability 2020, 12, 6768_ 10 of 20
for sharing data. The rest of the section will discuss some of the implementation details such as the
installation, configuration, and visualization of Hyperledger Explorer.
5.4.1. Installation and Configuration
For the Installation, the first step is to download the latest version of Hyperledger Explorer from
the official repository, followed by downloading PostgreSQL packages, and running the database
services to make sure the database has been installed correctly (as shown in Figure A2).
Once installed, the next step will be to authorize Hyperledger Explorer to access the network in
Fabric (Configuration). In the “app” folder inside the main folder of “blockchain-explorer”, the file
“explorerconfig.json” should be modified (Figure A3).
In “platform”, the fabric platform is used. In “PostgreSQL”, the database credentials will be
detailed. To connect Explorer with Fabric, access “blockchain-explorer/app/platform/fabric” where the
file “config.json” will be modified. The goal here is to define the connection with Fabric (Figure A4).
The name of the blockchain network in our case is set to “first-network”.
Finally, we open the json file located at:
```
/blockchain-explorer/app/platform/fabric/connection-profile/first-network.json
```
Then, we update “adminPrivateKey”, “signedCert” and “path” with the corresponding routes of
the Fabric network for visualization (Figure A5).
Once Fabric and Explorer are connected, the last commands (Figure 8) are executed to build the
project, which contains our case study:
_Sustainability 2020 FOR PEER REVIEW_ 10 of 20
`./main.sh install` (To make the build of the project.)
`./main.sh clean` (To clean up unnecessary files that were installed with the
previous command.)
`./main.sh test` (To test the REST API as well as the interface components,
it generates a document reporting errors.)
**Figure 8. Commands needed to run Fabric.**
**Figure 8. Commands needed to run Fabric.**
5.4.2. Visualization
5.4.2. Visualization
The final step is to visualize the blockchain network from an analytical point of view. For this
The final step is to visualize the blockchain network from an analytical point of view. For this
purpose, it is necessary to modify some packages of Hyperledger Explorer. Its structure is shown in
purpose, it is necessary to modify some packages of Hyperledger Explorer. Its structure is shown in
Figure A6.
Figure A6.
In order to customize Hyperledger Explorer, the default code of the official package must be
In order to customize Hyperledger Explorer, the default code of the official package must be
modified. It is developed with React.js and Redux frameworks. Therefore, to edit the components it is
modified. It is developed with React.js and Redux frameworks. Therefore, to edit the components it
necessary to access the folder “is necessary to access the folder “/blockchain-explorer/client/src/components” and edit the /blockchain-explorer/client/src/components” and edit the components
that are required. Here, we have only modified Charts, as it supports visualization. Figurecomponents that are required. Here, we have only modified Charts, as it supports visualization. 9 shows the
dashboard of Explorer, including a set of panels with the current configuration.Figure 9 shows the dashboard of Explorer, including a set of panels with the current configuration.
-----
is necessary to access the folder “/blockchain-explorer/client/src/components” and edit the
_Sustainability 2020, 12, 6768_ 11 of 20
components that are required. Here, we have only modified Charts, as it supports visualization.
Figure 9 shows the dashboard of Explorer, including a set of panels with the current configuration.
**Figure 9. Dashboard of Hyperledger Explorer.**
**Figure 9. Dashboard of Hyperledger Explorer.**
On the top panel, we can see that the network has eight blocks (from Block 0 to Block 7; the genesis
On the top panel, we can see that the network has eight blocks (from Block 0 to Block 7; the
block is a configuration block for a specific Hyperledger Fabric channel and contains no data) with
genesis block is a configuration block for a specific Hyperledger Fabric channel and contains no data)
eight transactions (one transaction per block). There are four nodes representing four diwith eight transactions (one transaction per block). There are four nodes representing four different fferent users
registered on the network. In this case study, there are zero chaincodes since no smart contracts wereusers registered on the network. In this case study, there are zero chaincodes since no smart contracts
created. Chaincode refers to the code for executing programs in the blockchain. These codes or smartwere created. Chaincode refers to the code for executing programs in the blockchain. These codes or
contracts signify a particular mini agreement that gets automatically triggered when the conditionsmart contracts signify a particular mini agreement that gets automatically triggered when the
values align to the required set of conditions. The word chaincode is a simple phrase to indicate thatcondition values align to the required set of conditions. The word chaincode is a simple phrase to
the code is related to the blockchain.indicate that the code is related to the blockchain.
Below the top panel are the list of Peers on the left and the network traffic on the right. Peers are
network elements that help maintain the network and verify and approve transactions. They also
provide methods for interacting with the network, such as creating different APIs.
The component on the lower left shows the blockchain. It shows the last block added (Block 7).
Each block has three different fields:
- Channel Name: The name of the channel through which the block has been created. A channel
is a mechanism by which a set of components of a blockchain network interact and exchange
information. They provide privacy to the network. There can be different channels, and users can
access one or another, depending on how their permissions are configured.
- Datahash: This is an encrypted code that contains all the information of the block. Here, you can
find information about the sender of the file, the receiver of the file, and the file itself.
- Number of Tx: This represents the number of transactions per block.
To the right of this last component is Transactions by Organization, an entity that has access to
different channels and shows how network participants are grouped according to their privileges.
Finally, it is important to question the suitability of approaches similar to ours for inherently
decentralized architectures such as distributed ledgers or blockchains, where processing, storage,
and control flow are shared among many equal participants. Van Landuyt et al. [69] performed
an analysis of blockchain security and the privacy of data it supports with other threat-modeling
approaches discussed in the literature and their findings identify areas for future improvements needed
for threat-modeling approaches.
-----
control flow are shared among many equal participants. Van Landuyt et al. [69] performed an
_Sustainabilityanalysis of blockchain security and the privacy of data it supports with other threat-modeling 2020, 12, 6768_ 12 of 20
approaches discussed in the literature and their findings identify areas for future improvements
needed for threat-modeling approaches.
_5.5. User Study_
_5.5. User Study A user study was carried out to determine the features important to users with respect to the_
visualization model and implementation. The total number of users within the authors’ research groupA user study was carried out to determine the features important to users with respect to the
performing the study were 11: two full professors, three associate professors, three PhD students,visualization model and implementation. The total number of users within the authors’ research
and three degree students. Figuregroup performing the study were 11: two full professors, three associate professors, three PhD 10 shows the results from the users’ responses. The users indicated
that security and a user-friendly nature are the most important features. The preliminary results showstudents, and three degree students. Figure 10 shows the results from the users’ responses. The users
that transparency in data sharing is important for user participation when there is no single trustedindicated that security and a user-friendly nature are the most important features. The preliminary
coordinating entity that users can rely on.results show that transparency in data sharing is important for user participation when there is no
single trusted coordinating entity that users can rely on.
## Main feature
Security
**20% [5%]**
**45%** User friendly
**30%**
Easy integration into
other systems
Others ( open
source, ...)
(a) Main feature to users with respect to the
visualization model and implementation.
## Second feature
**10%**
**25%**
**30%**
**35%**
(b) Second feature to users with respect to the
visualization model and implementation.
**Figure 10.Figure 10. User study results.User study results.**
In summary, the proposed digital platform can be used in any healthcare application where thereIn summary, the proposed digital platform can be used in any healthcare application where
are multiple actors (hospital, patients, external clinical and non-clinical care providers) sharing selectthere are multiple actors (hospital, patients, external clinical and non-clinical care providers) sharing
data among each other to support care. The content of the file (or resource) to be shared and who itselect data among each other to support care. The content of the file (or resource) to be shared and
should be sent to is determined by the client (patient, provider, etc.), and the blockchain architecturewho it should be sent to is determined by the client (patient, provider, etc.), and the blockchain
supports interoperability among a number of distributed systems outside a hospital’s own EMR.architecture supports interoperability among a number of distributed systems outside a hospital’s
With the immutability of data stored and the authenticity of those accessing the data, the architectureown EMR. With the immutability of data stored and the authenticity of those accessing the data, the
ensures that those who are designated to receive the data are indeed the ones who are accessing the
data. More importantly, by visually tracking the movement of data files, the users can see and interpret
the activity. This is a key contribution of this paper.
**6. Discussion**
The implementation can be generalized to share different types of files based on the application
context. For example, users may cast their vote on an issue or in an election and see how these
are pooled by an authentic node on the network for compilation. Similarly, in today’s COVID-19
environment, data from various test facilities and hospitals can be tracked for the number of people
infected (or testing positive) and the number of hospitalizations and deaths for public health officials
to develop regional patterns. With some of the demographic or geographical data of each node
stored outside the blockchain, it can reduce the data redundancy but provide access to interpret the
data traffic within the network. Moreover, blockchains using smart contracts can provide alerts in
appropriate nodes based on data analysis. For example, an alert can be sent to a public health node on
the network when the number of positive cases coming from that region exceeds a threshold for its
regional population, so that it can develop alternative preventive practices. Similarly, it can trigger an
alert to an emergency management vehicle station node when a hospital within its area has exceeded
its hospitalization capacity, so that patients can be diverted to another hospital. Furthermore, some of
these partners, such as emergency management vehicle stations or public health agencies, can be
-----
_Sustainability 2020, 12, 6768_ 13 of 20
outside the blockchain if they are primarily receiving alerts or aggregated data to reduce network
complexity, or else in a separate blockchain that is used for receiving such alerts.
_6.1. Future Research Directions_
When care is moved outside a hospital and with a number of actors sharing different types of data
at varying frequencies, future research needs to explore certain heuristics or algorithmic models to
segment the digital platform that may include a mix of centralized and distributed networks. Each of
these networks is synchronized to ensure that data moving within and across these networks are not
lost. The larger the network, the greater the technical challenge of managing the actors and the data
they share and the more complex the social challenge of aligning the goals of these actors. In addition,
the distributed actors using blockchain must eventually interact with other actors (e.g., hospitals) who
operate centrally coordinated patient health records or a government agency that regulates the type of
data shared. This leads to three different possible research directions.
Addressing the technical challenge: Are there ways to decide when to segment the data based on
frequency of use and the size of the data shared? Given the redundancy embedded in the way in which
the blockchain replicates the data, decision rules may guide the size of the data to be shared, the type
of data shared (e.g., images vis-à-vis text), the frequency of data sharing (e.g., once a month with a few
nodes or real time for tracking infections) and, of course, the number of nodes who need access to
these data. This may lead to the creation of subnetworks, which also are relevant in addressing the
complexity of the social challenge.
Addressing the social challenge: Healthcare outside a hospital is supported by many different
actors, such as clinical actors like pharmacies or testing labs, non-clinical actors like social workers or
care givers of patients at home, or researchers who analyze data for treatment adherence or disease
patterns. The motivation of these users to use such a platform to share data and the transparency they
need to enhance their trust in using the system may vary. Therefore, having different networks support
clinical actors, non-clinical actors, and analysis may lead to reducing the goal alignment complexity
and help mitigate the need for visualization and associated complexities in system design. Moreover,
many of these subgroups have varying levels of interaction with hospitals, thus creating the need for
different gateways for data sharing with the hospital EMR, an issue which is discussed next.
Gateways to centralized systems: Hospitals and government agencies still drive much of healthcare
around the world, and the type of data integration they need with external actors varies. For example,
central public health agencies of regions or countries need aggregated data from hospitals and other
external care providers like test facilities to track disease conditions, except during health emergencies
when real time data access is critical. Similarly, hospitals may need certain data in real time from clinical
actors outside the system like pharmacies to control over-prescription or use of drugs, whereas they
need periodic data from social workers on patient adherence to treatment protocols. This means that
each blockchain network may have to decide which centralized systems will become nodes and how
data are aggregated and sent to these nodes based on pre-defined criteria. In some cases, the centralized
systems may be part of a separate network, with the blockchain network of the distributed actors
simply connected to the centralized system network to ensure the integrity of each.
_6.2. Limitations_
We reviewed the state of the art of both the challenges and opportunities offered by the blockchain
technology-based solution in terms of modeling problems in general and in healthcare in particular. It is
important to emphasize that although the technology itself is not new, the fundamental contribution of
the paper here is the use of visualization to make blockchain use transparent. This highly model-driven
and flexible methodology provides an integration with existing technologies, highlights various
challenges and opportunities when they are integrated with a blockchain with IoT [70] and suggests
improvements to support decentralization and scalability, identity (identification of every device),
autonomy, reliability (verifying the data authenticity), security (validation by smart contracts among
-----
_Sustainability 2020, 12, 6768_ 14 of 20
other services), market of services (interesting solutions for an IoT ecosystem of services and data
marketplaces), and secure code deployment (significant advantage of blockchain secure-immutable
storage). Similarly, the survey [71] reviews blockchain challenges and opportunities and indicates a wide
spectrum of blockchain applications extending from cryptocurrency, financial services, risk management
and internet of things (IoT) to public and social services. The authors conducted a comprehensive
survey on the blockchain technology with a focus on taxonomy, algorithms, applications, and technical
challenges as well as recent advances to address some of these challenges.
Another important issue within the blockchain framework is cryptocurrencies, as they are an
emerging economic force, but there are concerns about their security. The reason for this is due to
the complex collusion cases and new threat vectors that could be missed by conventional security
assessment strategies. Almashaqbeh et al. [72] propose an ABC: an Asset-Based Cryptocurrency-focused
threat modeling framework, which demonstrates the effectiveness of some real-world use cases.
Finally, as we have observed in Section 5.5, the user study that has been carried out has the
usual limitations of a preliminary study. For this reason, it will be necessary to extend it to a study
with more users with different profiles in order to evaluate our proposal in a more exhaustive and
comprehensive way and thus make it more general purpose. Finally, it would be necessary to compare
our proposal with similar cases, where visualization is not present, to demonstrate the advantages of
our methodology in gaining user trust to use a blockchain to share data.
**7. Conclusions**
This paper illustrates the use of a digital platform based on an underlying blockchain technology
architecture to support data sharing by patients with external partners. It brings to surface the
mechanism used by blockchain technology to send and receive data in a secure manner to engender
trust among those sharing the data. Such transparency is key if the digital platform is to motivate
patients, who are unfamiliar with the technology, to share their data with others who are willing to
provide a service. Ultimately, the ease of use supported by interoperability among different patient
and partner systems and the transparency with regard to how the data are shared among patients and
partners are both critical for enhancing the external resources used to sustain care outside a hospital.
**Author Contributions: Conceptualization, J.P., D.G., M.T. and P.K.; methodology, J.P., E.G., D.G., M.T. and P.K.;**
software, E.G.; validation, E.G. and P.K.; formal analysis, J.P., D.G., M.T. and P.K.; investigation, J.P., E.G., D.G.,
M.T. and P.K.; resources, E.G.; data curation, E.G.; writing—original draft preparation, J.P., E.G., D.G., M.T. and
P.K.; writing—review and editing, J.P., D.G. and M.T.; visualization, E.G.; supervision, J.P., D.G., M.T. and P.K.;
project administration, J.P., D.G. and M.T.; funding acquisition, J.P. and D.G. All authors have read and agreed to
the published version of the manuscript.
**Funding: This study has been partially funded by the ECLIPSE-UA project (RTI2018-094283-B-C32).**
**Conflicts of Interest: The authors declare no conflict of interest.**
-----
to the published version of the manuscript.
_SustainabilityFunding: This study has been partially funded by the ECLIPSE-UA project (RTI2018-094283-B-C32). 2020, 12, 6768_ 15 of 20
**Funding: This study has been partially funded by the ECLIPSE-UA project (RTI2018-094283-B-C32).**
**Conflicts of Interest: The authors declare no conflict of interest.**
**Conflicts of Interest: The authors declare no conflict of interest.**
**Appendix A Screenshots of Terminal Windows**
**Appendix A. Screenshots of Terminal Windows**
**Appendix A. Screenshots of Terminal Windows**
**Figure A1.Figure A1. Successful blockchain creation.Successful blockchain creation.**
**Figure A1. Successful blockchain creation.**
_Sustainability 2020 FOR PEER REVIEW Figure A2.Figure A2. Check for correct database creation.Check for correct database creation._ 15 of 20
**Figure A2. Check for correct database creation.**
**Figure A3.Figure A3. Hyperledger Explorer access to the network in Fabric.Hyperledger Explorer access to the network in Fabric.**
-----
_Sustainability 2020, 12, 6768_ 16 of 20
**Figure A3. Hyperledger Explorer access to the network in Fabric.**
_Sustainability 2020 FOR PEER REVIEW_ 16 of 20
_Sustainability 2020 FOR PEER REVIEW Figure A4.Figure A4. Connection with Fabric: network name.Connection with Fabric: network name._ 16 of 20
**Figure A5.Figure A5. Figure A5. Connection with Fabric: path settings.Connection with Fabric: path settings. Connection with Fabric: path settings.**
_Sustainability 2020 FOR PEER REVIEW Figure A4.Figure A4. Connection with Fabric: network name.Connection with Fabric: network name._ 16 of 20
**References**
**2020 FOR PEER REVIEW**
**Figure A6.Figure A6. Hyperledger Explorer structure.Hyperledger Explorer structure.**
**Figure A6. Hyperledger Explorer structure.**
1. Aghina, W.; De Smet, A.; Weerda, K. Agility: It rhymes with stability; McKinsey Quarterly, December 2015.
**References**
Available online: https://www.mckinsey.com/business-functions/organization/our-insights/agility-it
1. Aghina, W.; De Smet, A.; Weerda, K. Agility: It rhymes with stability; McKinsey Quarterly, December 2015. rhymes-with-stability (accessed on 6 June 2020).
Available 2. Bossert, O.; Laartz, J.; Ramsey, T.J. Running your company at two speeds; McKinsey Quarterly, December online: https://www.mckinsey.com/business-functions/organization/our-insights/agility-it
2014. Available online: https://www.mckinsey.com/business-functions/mckinsey-digital/our
rhymes-with-stability (accessed on 6 June 2020)
-----
_Sustainability 2020, 12, 6768_ 17 of 20
**References**
1. Aghina, W.; De Smet, A.; Weerda, K. Agility: It rhymes with stability. McKinsey Quarterly. December 2015.
[Available online: https://www.mckinsey.com/business-functions/organization/our-insights/agility-it-rhymes-](https://www.mckinsey.com/business-functions/organization/our-insights/agility-it-rhymes-with-stability)
[with-stability (accessed on 6 June 2020).](https://www.mckinsey.com/business-functions/organization/our-insights/agility-it-rhymes-with-stability)
2. Bossert, O.; Laartz, J.; Ramsey, T.J. Running your company at two speeds. McKinsey Quarterly. December 2014.
[Available online: https://www.mckinsey.com/business-functions/mckinsey-digital/our-insights/running-](https://www.mckinsey.com/business-functions/mckinsey-digital/our-insights/running-your-company-at-two-speeds)
[your-company-at-two-speeds (accessed on 6 June 2020).](https://www.mckinsey.com/business-functions/mckinsey-digital/our-insights/running-your-company-at-two-speeds)
3. Lusch, R.F.; Nambisan, S. Service Innovation: A Service-Dominant Logic Perspective. MIS Quart. 2015, 39,
155–175.
4. Tanniru, M.; Khuntia, J.; Weiner, J. Hospital Leadership in Support of Digital Transformation. Pac. Asia J.
_[Assoc. Inf. Syst. 2018, 10, 1–24. [CrossRef]](http://dx.doi.org/10.17705/1PAIS.10301)_
5. Vargo, S.L.; Lusch, R.F. Service-Dominant Logic: What it is, What it is not, What it might be. In The
_Service-Dominant Logic of Marketing: Dialog, Debate, and Directions; Vargo, S.L., Lusch, R.F., Eds.; M.E. Sharpe:_
Armonk, NY, USA, 2006; pp. 43–55.
6. Nuckols, T.K.; Fingar, K.R.; Barrett, M.; Steiner, C.A.; Stocks, C.; Owens, P.L. The shifting landscape in
utilization of inpatient, observation, and emergency department services across payers. J. Hosp. Med. 2017,
_[12, 443–446. [CrossRef]](http://dx.doi.org/10.12788/jhm.2751)_
7. Sanborn, B.J. Outpatient shift, digitization, shortages, and telehealth will shape healthcare systems of
[tomorrow. Healthcare Finance. May 2018. Available online: https://www.healthcarefinancenews.com/](https://www.healthcarefinancenews.com/news/outpatient-shift-digitization-shortages-and-telehealth-will-shape-healthcare-systems-tomorrow)
[news/outpatient-shift-digitization-shortages-and-telehealth-will-shape-healthcare-systems-tomorrow](https://www.healthcarefinancenews.com/news/outpatient-shift-digitization-shortages-and-telehealth-will-shape-healthcare-systems-tomorrow)
(accessed on 6 June 2020).
8. Tanniru, M.; Sandhu, K. Engagement leading to empowerment-Digital Innovation Strategies for Patient
Care Continuity. In Proceedings of the Australasian Conference on Information Systems, Sydney, Australia,
[January 2018; Available online: https://aisel.aisnet.org/cgi/viewcontent.cgi?article=1039&context=acis2018](https://aisel.aisnet.org/cgi/viewcontent.cgi?article=1039&context=acis2018)
(accessed on 6 June 2020).
9. Johnston, L.; Zemanek, J.; Reeve, M.J.; Grills, N. The evidence for using mHealth technologies for diabetes
management in low- and middle-income countries. J. Hosp. Manag. Health Policy 2018, 2. Available online:
[http://jhmhp.amegroups.com/article/view/4403/5182 (accessed on 6 June 2020).](http://jhmhp.amegroups.com/article/view/4403/5182)
10. Agarwal, R.; Gao, G.; DesRoches, C.M.; Jha, A.K. Research Commentary—The Digital Transformation of
[Healthcare: Current Status and the Road Ahead. Inform. Syst. Res. 2010, 21, 796–809. [CrossRef]](http://dx.doi.org/10.1287/isre.1100.0327)
11. Ghosh, K.; Khuntia, J.; Chawla, S.; Deng, X. Media Reinforcement for Psychological Empowerment in
[Chronic Disease Management. Commun. Assoc. Inf. Syst. 2014, 34, 419–438. [CrossRef]](http://dx.doi.org/10.17705/1CAIS.03422)
12. Tanniru, M. Transforming public health using value lens and extended partner networks. Learn. Health Syst.
**[2020. Available online: https://onlinelibrary.wiley.com/doi/full/10.1002/lrh2.10234 (accessed on 6 June 2020).](https://onlinelibrary.wiley.com/doi/full/10.1002/lrh2.10234)**
13. Dowell, D.; Haegerich, T.; Chou, R. No Shortcuts to Safer Opioid Prescribing. N. Engl. J. Med. 2019,
_[380, 2285–2287. [CrossRef]](http://dx.doi.org/10.1056/NEJMp1904190)_
14. Rahman, A.; Rashid, M.; Kernec, J.L.; Phillippe, B.; Barnes, S.J.; Fioranell, F.; Yang, S.; Romain, O.; Abbasi, Q.;
Loukas, G.; et al. A Secure Occupational Therapy Framework for Monitoring Cancer Patients Quality of Life.
_[Sensors 2019, 19, 5258. [CrossRef]](http://dx.doi.org/10.3390/s19235258)_
15. Boots, L.M.M.; de Vugt, M.E.; van Knippenberg, R.J.M.; Kempen, G.I.J.M.; Verhey, F. A Systematic Review
of Internet-Based Supportive Interventions for Caregivers of Patients with Dementia. Int. J. Geriatr. Psych.
**[2013, 29, 331–344. [CrossRef]](http://dx.doi.org/10.1002/gps.4016)**
16. Carretero, S.; Stewart, J.; Centeno, C. Information and Communication Technologies for Informal Carers
and Paid Assistants: Benefits from Micro-, Meso-, And Macro-Levels. Eur. J. Ageing. 2015, 12, 163–173.
[[CrossRef]](http://dx.doi.org/10.1007/s10433-015-0333-4)
17. Crosby, M.; Nachiappan; Pattanayak, P.; Verma, S.; Kalyanaraman, V. Blockchain technology: Beyond bitcoin.
_[Appl. Innov. Rev. 2016, 2, 6–19. Available online: http://scet.berkeley.edu/wp-content/uploads/AIR-2016-](http://scet.berkeley.edu/wp-content/uploads/AIR-2016-Blockchain.pdf)_
[Blockchain.pdf (accessed on 6 June 2020).](http://scet.berkeley.edu/wp-content/uploads/AIR-2016-Blockchain.pdf)
18. Swan, M. Blockchain: Blueprint for a new economy, 1st ed.; McGovern, T., Ed.; O’Reilly Media: Sebastopol, CA,
USA, 2015.
19. Zhao, J.L.; Fan, S.; Yan, J. Overview of business innovations and research opportunities in blockchain and
[introduction to the special issue. Financ. Innov. 2016, 2, 28. [CrossRef]](http://dx.doi.org/10.1186/s40854-016-0049-2)
-----
_Sustainability 2020, 12, 6768_ 18 of 20
20. Casino, F.; Dasaklis, T.K.; Patsakis, C. A systematic literature review of blockchain-based applications:
Current status, classification, and open issues. Telemat. Inform. 2019, 36, 55–81.
21. Beninger, P.; Ibara, M.A. Pharmacovigilance and Biomedical Informatics: A Model for Future Development.
_[Clin. Ther. 2016, 38, 2514–2525. [CrossRef]](http://dx.doi.org/10.1016/j.clinthera.2016.11.006)_
22. Sankar, L.S.; Sindhu, M.; Sethumadhavan, M. Survey of consensus protocols on blockchain applications.
In Proceedings of the 2017 4th International Conference on Advanced Computing and Communication
[Systems (ICACCS), Coimbatore, India, 6–7 January 2017; pp. 1–5. [CrossRef]](http://dx.doi.org/10.1109/ICACCS.2017.8014672)
23. McGhin, T.; Choo, K.K.R.; Liu, C.Z.; He, D. Blockchain in healthcare applications: Research challenges and
[opportunities. J. Netw. Comput. Appl. 2019, 135, 62–75. [CrossRef]](http://dx.doi.org/10.1016/j.jnca.2019.02.027)
24. Azaria, A.; Ekblaw, A.; Vieira, T.; Lippman, A. MedRec: Using blockchain for medical data access and
permission management. In Proceedings of the 2016 2nd International Conference on Open and Big Data
[(OBD), Vienna, Austria, 22–24 August 2016; pp. 25–30. [CrossRef]](http://dx.doi.org/10.1109/OBD.2016.11)
25. Yue, X.; Wang, H.; Jin, D.; Li, M.; Jiang, W. Healthcare Data Gateways: Found Healthcare Intelligence on
[Blockchain with Novel Privacy Risk Control. J. Med. Syst. 2016, 40, 218. [CrossRef]](http://dx.doi.org/10.1007/s10916-016-0574-6)
26. Yüksel, B.; Küpçü, A.; Özkasap, Ö. Research issues for privacy and security of electronic health services.
_[Future Gener. Comput. Syst. 2017, 68, 1–13. [CrossRef]](http://dx.doi.org/10.1016/j.future.2016.08.011)_
27. Griggs, K.N.; Ossipova, O.; Kohlios, C.P.; Baccarini, A.N.; Howson, E.A.; Hayajneh, T. Healthcare Blockchain
System Using Smart Contracts for Secure Automated Remote Patient Monitoring. J. Med. Syst. 2018, 42, 1–7.
[[CrossRef]](http://dx.doi.org/10.1007/s10916-018-0982-x)
28. Xia, Q.; Sifah, E.B.; Asamoah, K.O.; Gao, J.; Du, X.; Guizani, M. MeDShare: Trust-Less Medical Data Sharing
[among Cloud Service Providers via Blockchain. IEEE Access 2017, 5, 14757–14767. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2017.2730843)
29. Abouelmehdi, K.; Beni-Hssane, A.; Khaloufi, H.; Saadi, M. Big data security and privacy in healthcare:
[A Review. Procedia Comput. Sci. 2017, 113, 73–80. [CrossRef]](http://dx.doi.org/10.1016/j.procs.2017.08.292)
30. Small, A.; Wainwright, D. Privacy and security of electronic patient records – Tailoring multimethodology to
explore the socio-political problems associated with Role Based Access Control systems. Eur. J. Oper. Res.
**2018, 265, 344–360.**
31. Kshetri, N. Blockchain’s roles in strengthening cybersecurity and protecting privacy. Telecommun. Policy
**[2017, 41, 1027–1038. [CrossRef]](http://dx.doi.org/10.1016/j.telpol.2017.09.003)**
32. Ramani, V.; Kumar, T.; Bracken, A.; Liyanage, M.; Ylianttila, M. Secure and Efficient Data Accessibility in
Blockchain Based Healthcare Systems. In Proceedings of the 2018 IEEE Global Communications Conference
[(GLOBECOM), Abu Dhabi, UAE, 9–13 December 2018; pp. 206–212. [CrossRef]](http://dx.doi.org/10.1109/GLOCOM.2018.8647221)
33. Khan, S.I.; Hoque, A.S.M.L. Privacy and security problems of national health data warehouse: A convenient
solution for developing countries. In Proceedings of the 2016 International Conference on Networking
[Systems and Security (NSysS), Dhaka, Bangladesh, 7–9 January 2016; pp. 1–6. [CrossRef]](http://dx.doi.org/10.1109/NSysS.2016.7400708)
34. Rajkomar, A.; Oren, E.; Chen, K.; Dai, A.M.; Hajaj, N.; Hardt, M.; Liu, P.J.; Liu, X.; Marcus, J.; Sun, M.; et al.
[Scalable and accurate deep learning with electronic health records. Npj Digital Med. 2018, 1, 1–10. [CrossRef]](http://dx.doi.org/10.1038/s41746-018-0029-1)
35. Esposito, C.; De Santis, A.; Tortora, G.; Chang, H.; Choo, K.K.R. Blockchain: A Panacea for Healthcare
[Cloud-Based Data Security and Privacy? IEEE Cloud Comput. 2018, 5, 31–37. [CrossRef]](http://dx.doi.org/10.1109/MCC.2018.011791712)
36. Dagher, G.G.; Mohler, J.; Milojkovic, M.; Marella, P.B. Ancile: Privacy-preserving framework for access
control and interoperability of electronic health records using blockchain technology. Sustain. Cities Soc.
**[2018, 39, 283–297. [CrossRef]](http://dx.doi.org/10.1016/j.scs.2018.02.014)**
37. Ekblaw, A.; Azaria, A.; Halamka, J.D.; Lippman, A. A Case Study for Blockchain in Healthcare: “MedRec”
prototype for electronic health records and medical research data. White Paper, reprinted from 2nd
[International Conference on Open & Big Data, 22–24 August 2016, 1–13. Available online: https://www.](https://www.media.mit.edu/publications/medrec-whitepaper/)
[media.mit.edu/publications/medrec-whitepaper/ (accessed on 6 June 2020).](https://www.media.mit.edu/publications/medrec-whitepaper/)
38. Roehrs, A.; da Costa, C.A.; da Rosa Righi, R. OmniPHR: A distributed architecture model to integrate
[personal health records. J. Biomed. Inform. 2017, 71, 70–81. [CrossRef]](http://dx.doi.org/10.1016/j.jbi.2017.05.012)
39. Kotz, D.; Gunter, C.A.; Kumar, S.; Weiner, J.P. Privacy and Security in Mobile Health: A Research Agenda.
_[Computer 2016, 49, 22–30. [CrossRef]](http://dx.doi.org/10.1109/MC.2016.185)_
40. Sahi, M.; Abbas, H.; Saleem, K.; Yang, X.; Derhab, A.; Orgun, M.; Iqbal, W.; Rashid, L.; Yaseen, A. Privacy
preservation in e-healthcare environments: State of the art and future directions. IEEE Access 2018, 6, 464–478.
[[CrossRef]](http://dx.doi.org/10.1109/ACCESS.2017.2767561)
-----
_Sustainability 2020, 12, 6768_ 19 of 20
41. Zhang, Y.; Liu, T.; Li, K.; Zhang, J. Improved visual correlation analysis for multidimensional data. J. Visual
_[Lang. Comput. 2017, 41, 121–132. [CrossRef]](http://dx.doi.org/10.1016/j.jvlc.2017.03.005)_
42. Cecil, J.; Gupta, A.; Pirela-Cruz, M.; Ramanathan, P. An IoMT based cyber training framework for orthopedic
[surgery using Next Generation Internet technologies. Inf. Med. Unlocked 2018, 12, 128–137. [CrossRef]](http://dx.doi.org/10.1016/j.imu.2018.05.002)
43. Joyia, G.J.; Liaqat, R.M.; Farooq, A.; Rehman, S. Internet of Medical Things (IOMT): Applications, Benefits and
[Future Challenges in Healthcare Domain. J. Commun. 2017, 12, 240–247. [CrossRef]](http://dx.doi.org/10.12720/jcm.12.4.240-247)
44. Pilkington, M. Can Blockchain Improve Healthcare Management? Consumer Medical Electronics and the
[IoMT. SSRN 2017, 1–13. Available online: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3025393](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3025393)
(accessed on 6 June 2020).
45. Hoy, M.B. An introduction to the blockchain and its implications for libraries and medicine. Med. Ref. Serv.
_Q. 2017, 36, 273–279._
46. Kuo, T.T.; Kim, H.E.; Ohno-Machado, L. Blockchain distributed ledger technologies for biomedical and
health care applications. J. Am. Med. Inform. Assoc. 2017, 24, 1211–1220.
47. Dubovitskaya, A.; Xu, Z.; Ryu, S.; Schumacher, M.; Wang, F. How blockchain could empower ehealth:
[An application for radiation oncology. Lect. Notes Comput. Sci. 2017, 10494, 3–6. [CrossRef]](http://dx.doi.org/10.1007/978-3-319-67186-4_1)
48. [Medicalchain. 2018. Whitepaper 2.1. Available online: https://medicalchain.com/Medicalchain-Whitepaper-](https://medicalchain.com/Medicalchain-Whitepaper-EN.pdf)
[EN.pdf (accessed on 6 June 2020).](https://medicalchain.com/Medicalchain-Whitepaper-EN.pdf)
49. Peterson, K.; Deeduvanu, R.; Kanjamala, P.; Boles, K. A blockchain-based approach to health
information exchange networks. In Proc. NIST Workshop Blockchain Healthcare. 2016.
[Available online: https://pdfs.semanticscholar.org/c1b1/89c81b6fda71a471adec11cfe72f6067c1ad.pdf?_ga=2.](https://pdfs.semanticscholar.org/c1b1/89c81b6fda71a471adec11cfe72f6067c1ad.pdf?_ga=2.151693625.186512466.1597711686-512884126.1594177302)
[151693625.186512466.1597711686-512884126.1594177302 (accessed on 6 June 2020).](https://pdfs.semanticscholar.org/c1b1/89c81b6fda71a471adec11cfe72f6067c1ad.pdf?_ga=2.151693625.186512466.1597711686-512884126.1594177302)
50. Ahram, T.; Sargolzaei, A.; Sargolzaei, S.; Daniels, J.; Amaba, B. Blockchain technology innovations.
In Proceedings of the 2017 IEEE Technology and Engineering Management Society Conference (TEMSCON),
[Santa Clara, CA, USA, 8–10 June 2017; pp. 137–141. [CrossRef]](http://dx.doi.org/10.1109/TEMSCON.2017.7998367)
51. Al Omar, A.; Rahman, M.S.; Basu, A.; Kiyomoto, S. MediBchain: A blockchain based privacy preserving
platform for healthcare data. In Proceedings of the 10th International Conference on Security, Privacy
and Anonymity in Computation, Communication and Storage, Guangzhou, China, 12–15 December 2017;
[pp. 534–543. [CrossRef]](http://dx.doi.org/10.1007/978-3-319-72395-2_49)
52. Borioli, G.S.; Couturier, J. How blockchain technology can improve the outcomes of clinical trials. Br. J.
_Health Care Manag. 2018, 24, 156–162._
53. Mamoshina, P.; Ojomoko, L.; Yanovich, Y.; Ostrovski, A.; Botezatu, A.; Prikhodko, P.; Izumchenko, E.;
Aliper, A.; Romantsov, K.; Zhebrak, A.; et al. Converging blockchain and next-generation artificial
intelligence technologies to decentralize and accelerate biomedical research and healthcare. Oncotarget 2018,
_9, 5665–5690._
54. Lee, S.H.; Yang, C.S. Fingernail analysis management system using microscopy sensor and blockchain
[technology. Int. J. Distrib. Sens. Netw. 2018, 14, 1–13. [CrossRef]](http://dx.doi.org/10.1177/1550147718767044)
55. Shae, Z.; Tsai, J.J.P. On the Design of a Blockchain Platform for Clinical Trial and Precision. In Proceedings of
the 2017 IEEE 37th International Conference on Distributed Computing Systems (ICDCS), Atlanta, GA, USA,
[5–8 June 2017; pp. 1972–1980. [CrossRef]](http://dx.doi.org/10.1109/ICDCS.2017.61)
56. Frentrup, M.; Theuvsen, L. Transparency in supply chains: Is trust a limiting factor? Presented at
the 99th EAAE Seminar Trust and Risk in Business Networks, Bonn, Germany, 8–10 February 2006;
[Available online: https://www.semanticscholar.org/paper/Transparency-in-supply-chains%3A-is-trust-a-](https://www.semanticscholar.org/paper/Transparency-in-supply-chains%3A-is-trust-a-limiting-Frentrup-Theuvsen/8fa45abcce7d8231c756af48810c6c797fea1f2a)
[limiting-Frentrup-Theuvsen/8fa45abcce7d8231c756af48810c6c797fea1f2a (accessed on 6 June 2020).](https://www.semanticscholar.org/paper/Transparency-in-supply-chains%3A-is-trust-a-limiting-Frentrup-Theuvsen/8fa45abcce7d8231c756af48810c6c797fea1f2a)
57. Weissgerber, T.L.; Milic, N.M.; Winham, S.J.; Garovic, V.D. Beyond bar and line graphs: Time for a new data
[presentation paradigm. PLoS Biol. 2015, 13, 1–10. [CrossRef]](http://dx.doi.org/10.1371/journal.pbio.1002128)
58. Weissgerber, T.L.; Garovic, V.D.; Savic, M.; Winham, S.J.; Milic, N.M. From Static to Interactive: Transforming
[Data Visualization to Improve Transparency. PLoS Biol. 2016, 14, 1–8. [CrossRef]](http://dx.doi.org/10.1371/journal.pbio.1002484)
59. Weiner, J.; Balijepally, V.; Tanniru, M. Integrating Strategic to Operational Decision-Making using Data-Driven
Dashboard Implementation: The Case of St. Joseph Mercy Oakland Hospital. J. Healthc. Manag. 2015,
_[60, 319–331. [CrossRef]](http://dx.doi.org/10.1097/00115514-201509000-00005)_
60. Selig, W.J.; Johannes, J.D. Reasoning visualization in expert systems-the applicability of algorithm animation
techniques. In Proceedings of the 3rd international conference on Industrial and engineering applications of
[artificial intelligence and expert systems, New York, NY, USA, 15–18 July 1990; pp. 457–466. [CrossRef]](http://dx.doi.org/10.1145/98784.98869)
-----
_Sustainability 2020, 12, 6768_ 20 of 20
61. Trainer, E.H. Supporting Trust in Globally Distributed Software Teams: The Impact of Visualized Collaborative
Traces on Perceived Trustworthiness. Doctoral Dissertation, University of California, La Jolla, CA, USA, 2012.
62. Perlmutter, L.; Kernfeld, E.; Cakmak, M. Situated language understanding with human-like and
visualization-based transparency. In Proceedings of the Robotics: Science and Systems Conference,
[Ann Arbor, MI, USA, 18–22 June 2016; pp. 1–10. [CrossRef]](http://dx.doi.org/10.15607/RSS.2016.XII.040)
63. Liu, Z.; Stasko, J. Mental Models, Visual Reasoning and Interaction in Information Visualization: A Top-down
[Perspective. IEEE Trans. Vis. Comput. Graph. 2010, 16, 999–1008. [CrossRef]](http://dx.doi.org/10.1109/TVCG.2010.177)
64. Hebert, C.; Di Cerbo, F. Secure blockchain in the enterprise: A methodology. Perv. Mob. Comput. 2019,
_[59, 101038. [CrossRef]](http://dx.doi.org/10.1016/J.PMCJ.2019.101038)_
65. Yaraghi, N. Who should profit from the sale of patient data? The Brookings Institution. Available
[online: https://www.brookings.edu/blog/techtank/2018/11/19/who-should-profit-from-the-sale-of-patient-](https://www.brookings.edu/blog/techtank/2018/11/19/who-should-profit-from-the-sale-of-patient-data/)
[data/ (accessed on 6 June 2020).](https://www.brookings.edu/blog/techtank/2018/11/19/who-should-profit-from-the-sale-of-patient-data/)
66. Hertzog, E.; Benartzi, G.; Benartzi, G.; Bancor Protocol. Continuous Liquidity for Cryptographic Tokens
[through their Smart Contracts. Available online: https://storage.googleapis.com/website-bancor/2018/04/](https://storage.googleapis.com/website-bancor/2018/04/01ba8253-bancor_protocol_whitepaper_en.pdf)
[01ba8253-bancor_protocol_whitepaper_en.pdf (accessed on 6 June 2020).](https://storage.googleapis.com/website-bancor/2018/04/01ba8253-bancor_protocol_whitepaper_en.pdf)
67. [A Blockchain Platform for the Enterprise. Hyperledger. Available online: https://hyperledger-fabric.](https://hyperledger-fabric.readthedocs.io/en/release-2.0/)
[readthedocs.io/en/release-2.0/ (accessed on 15 May 2020).](https://hyperledger-fabric.readthedocs.io/en/release-2.0/)
68. Josefsson, S. The Base16, Base32, and Base64 Data Encodings; RFC 4648, The Internet Society. 2006. Available
[online: https://www.semanticscholar.org/paper/The-Base16%2C-Base32%2C-and-Base64-Data-Encodings-](https://www.semanticscholar.org/paper/The-Base16%2C-Base32%2C-and-Base64-Data-Encodings-Josefsson/2718f599c9bbb96aecd81180167d10dcf9c65c47)
[Josefsson/2718f599c9bbb96aecd81180167d10dcf9c65c47 (accessed on 6 June 2020).](https://www.semanticscholar.org/paper/The-Base16%2C-Base32%2C-and-Base64-Data-Encodings-Josefsson/2718f599c9bbb96aecd81180167d10dcf9c65c47)
69. Van Landuyt, D.; Sion, L.; Vandeloo, E.; Joosen, W. On the Applicability of Security and Privacy Threat
Modeling for Blockchain Applications. In Computer Security; Katsikas, S., Cuppens, F., Cuppens, N.,
Lambrinoudakis, C., Kalloniatis, C., Mylopoulos, J., Antón, A., Gritzalis, S., Pallas, F., Pohle, J., et al., Eds.;
[Springer: Cham, Switzerland, 2020; Volume 11980, pp. 195–203. [CrossRef]](http://dx.doi.org/10.1007/978-3-030-42048-2_13)
70. AnaReyna, A.; Martín, C.; Chen, J.; Soler, E.; Díaz, M. On blockchain and its integration with IoT–Challenges
[and opportunities. Future Gener. Comput. Syst. 2018, 88, 173–190. [CrossRef]](http://dx.doi.org/10.1016/j.future.2018.05.046)
71. Zheng, Z.; Xie, S. Blockchain challenges and opportunities: A Survey. Int. J. Web Grid. Serv. 2018, 14, 352–375.
[[CrossRef]](http://dx.doi.org/10.1504/IJWGS.2018.10016848)
72. Almashaqbeh, G.; Bishop, A.; Cappos, J. ABC: A Cryptocurrency-Focused Threat Modeling Framework.
In Proceedings of the IEEE INFOCOM 2019-IEEE Conference on Computer Communications Workshops
[(INFOCOM WKSHPS), Paris, France, 29 April–2 May 2019; pp. 859–864. [CrossRef]](http://dx.doi.org/10.1109/INFCOMW.2019.8845101)
© 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
[(CC BY) license (http://creativecommons.org/licenses/by/4.0/).](http://creativecommons.org/licenses/by/4.0/.)
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3390/su12176768?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/su12176768, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "GOLD",
"url": "https://www.mdpi.com/2071-1050/12/17/6768/pdf?version=1598255404"
}
| 2,020
|
[] | true
| 2020-08-20T00:00:00
|
[
{
"paperId": "ce50d80f9ad070fb9a4b66e98cf787402804177f",
"title": "Can Blockchain Improve Healthcare Management?"
},
{
"paperId": "85a2d6936a792572bb2a2bb9a2687f584b5ec623",
"title": "Transforming public health using value lens and extended partner networks"
},
{
"paperId": "31a9d3bf8bebd0ce6ad95c72b16c6164f431db34",
"title": "A Secure Occupational Therapy Framework for Monitoring Cancer Patients’ Quality of Life"
},
{
"paperId": "d53231838c5887c3a5b1e05ff7567606f0f36c69",
"title": "Secure blockchain in the enterprise: A methodology"
},
{
"paperId": "489cf88a2d399edf065fc0d4cea95f1e0a8a7f88",
"title": "On the Applicability of Security and Privacy Threat Modeling for Blockchain Applications"
},
{
"paperId": "d0b74583708a96ed7562ab023aaad1261f25241d",
"title": "Engagement Leading to Empowerment-Digital Innovation Strategies for Patient Care Continuity"
},
{
"paperId": "c1e98fa629cd65080afd919eecb8628829fe66ce",
"title": "Blockchain in healthcare applications: Research challenges and opportunities"
},
{
"paperId": "3b652bc4e0ca3397357bd8a28587cdf41004796a",
"title": "No Shortcuts to Safer Opioid Prescribing."
},
{
"paperId": "dbf53d232e5282261ab628e75717d6f0b983384e",
"title": "ABC: A Cryptocurrency-Focused Threat Modeling Framework"
},
{
"paperId": "4c0945cb52d0734b25ecea49e3ae1c1b243fca66",
"title": "A systematic literature review of blockchain-based applications: Current status, classification and open issues"
},
{
"paperId": "41db775646e5909053126a13ced6c35199377f3f",
"title": "Secure and Efficient Data Accessibility in Blockchain Based Healthcare Systems"
},
{
"paperId": "305edd92f237f8e0c583a809504dcec7e204d632",
"title": "Blockchain challenges and opportunities: a survey"
},
{
"paperId": "7b55b3f41446c613453b14cbc5ea48ae4f634eeb",
"title": "Hospital Leadership in Support of Digital Transformation"
},
{
"paperId": "0c2ad254173631422348b28b6d9045e144180786",
"title": "The evidence for using mHealth technologies for diabetes management in low- and middle-income countries"
},
{
"paperId": "6d661299a8207a4bff536494cec201acee3c6c1c",
"title": "Healthcare Blockchain System Using Smart Contracts for Secure Automated Remote Patient Monitoring"
},
{
"paperId": "863dff6fea7811e6c2b76b3eb64eee84ee280b33",
"title": "Ancile: Privacy-Preserving Framework for Access Control and Interoperability of Electronic Health Records Using Blockchain Technology"
},
{
"paperId": "950c5a5c6dd3e8e338160fef2d0ff1b3c092de57",
"title": "Fingernail analysis management system using microscopy sensor and blockchain technology"
},
{
"paperId": "c078bf5f00a1bfb8df6bda9bc7fee6bfea6f5cbb",
"title": "Blockchain: A Panacea for Healthcare Cloud-Based Data Security and Privacy?"
},
{
"paperId": "baad851f534d38ffef5732149089a4a46b31e707",
"title": "How blockchain technology can improve the outcomes of clinical trials"
},
{
"paperId": "5d949d5f2c48572ff385da5576b2c10ab7142089",
"title": "Privacy and security of electronic patient records - Tailoring multimethodology to explore the socio-political problems associated with Role Based Access Control systems"
},
{
"paperId": "1f7eec4c76963a4ba7516ca00e6a2f855667b3f2",
"title": "Scalable and accurate deep learning with electronic health records"
},
{
"paperId": "87c259303857ef81593b2abd49106b9b5ce89b48",
"title": "MediBchain: A Blockchain Based Privacy Preserving Platform for Healthcare Data"
},
{
"paperId": "7f1c3c97a639a93796e935add3665c3ef329c0c8",
"title": "Blockchain's roles in strengthening cybersecurity and protecting privacy"
},
{
"paperId": "5bbc4181e073ec6b3ec894a35eacdc6a67e8c3a3",
"title": "Blockchain distributed ledger technologies for biomedical and health care applications"
},
{
"paperId": "b756c3c1f4d02421680c5ad3420bc11d5d3baeca",
"title": "How Blockchain Could Empower eHealth: An Application for Radiation Oncology - (Extended Abstract)"
},
{
"paperId": "4697609882b87d138455a60882608b293bb4e3ca",
"title": "Can Blockchain Improve Healthcare Management? Consumer Medical Electronics and the IoMT"
},
{
"paperId": "8f00a9e60c7bdb8ac473ee7a09bf1b051b33393d",
"title": "Improved visual correlation analysis for multidimensional data"
},
{
"paperId": "49af9119c09b97af977595b011afd8a3f588412d",
"title": "MeDShare: Trust-Less Medical Data Sharing Among Cloud Service Providers via Blockchain"
},
{
"paperId": "996d2697e16db08b6cfa89cadb924da534ddf3dd",
"title": "An Introduction to the Blockchain and Its Implications for Libraries and Medicine"
},
{
"paperId": "6536e22778df1d8b371cd8ff263145713e7bffd9",
"title": "OmniPHR: A distributed architecture model to integrate personal health records"
},
{
"paperId": "9092a7802f6e56dd5b6d1be30c8b5588a22e53fe",
"title": "Blockchain technology innovations"
},
{
"paperId": "c082bbc681566f6761729837a97a0ebab92a2ce8",
"title": "On the Design of a Blockchain Platform for Clinical Trial and Precision Medicine"
},
{
"paperId": "e11466449176d78b209e7ee32a03f5ddfc68cd8e",
"title": "The Shifting Landscape in Utilization of Inpatient, Observation, and Emergency Department Services Across Payers"
},
{
"paperId": "ef6e7f49c81767695d3048e6fc43d3d0a8188f72",
"title": "Research issues for privacy and security of electronic health services"
},
{
"paperId": "b66602ad0b663e93bea5765c4a56d83b7ef1c2dc",
"title": "Pharmacovigilance and Biomedical Informatics: A Model for Future Development."
},
{
"paperId": "c7083cfa963f9a58a27cbae152295b3a555bcc5a",
"title": "Overview of business innovations and research opportunities in blockchain and introduction to the special issue"
},
{
"paperId": "208735a6c437b8ae3efba01693c3e8a06289c3dd",
"title": "Healthcare Data Gateways: Found Healthcare Intelligence on Blockchain with Novel Privacy Risk Control"
},
{
"paperId": "bd8a307efcffbf57d2e5c3c23577de44d883d865",
"title": "MedRec: Using Blockchain for Medical Data Access and Permission Management"
},
{
"paperId": "3a1690ecb06986c3b5f66a81034c76e76845a80e",
"title": "Situated Language Understanding with Human-like and Visualization-Based Transparency"
},
{
"paperId": "538dd43365c1b6629f055df06f694fa53c969405",
"title": "Privacy and Security in Mobile Health: A Research Agenda"
},
{
"paperId": "457080ccecbfc681f1520cee65d8f94d39887a3e",
"title": "From Static to Interactive: Transforming Data Visualization to Improve Transparency"
},
{
"paperId": "b0f2c8ce920a63dfff803be0abe583ff03ee083e",
"title": "Converging blockchain and next-generation artificial intelligence technologies to decentralize and accelerate biomedical research and healthcare"
},
{
"paperId": "85ad29c9c76585ad33b1a339ac96abe8230ef4ae",
"title": "Integrating Strategic and Operational Decision Making Using Data‐Driven Dashboards: The Case of St. Joseph Mercy Oakland Hospital"
},
{
"paperId": "4d02488921e248b8a0cfe70f7d60c0606d3954c4",
"title": "Beyond Bar and Line Graphs: Time for a New Data Presentation Paradigm"
},
{
"paperId": "5d1d25e8e7c0abbac2b9189dbe7b9b67fee5fda7",
"title": "Service Innovation: A Service-Dominant Logic Perspective"
},
{
"paperId": "97fddbbfd681bce9eeb8e0a013353b4d5b2ba0db",
"title": "Blockchain: Blueprint for a New Economy"
},
{
"paperId": "d5d0c5c8bd73aa20d3d332ba2d6b0a2c0848ee62",
"title": "Information and communication technologies for informal carers and paid assistants: benefits from micro-, meso-, and macro-levels"
},
{
"paperId": "75dfeed53405a480817a668e966508156b33ef5d",
"title": "Service-Dominant Logic: What It Is, What It Is Not, What It Might Be"
},
{
"paperId": "d8d2d004d3ce7b5d036e5ec5587b9682cd53c714",
"title": "A systematic review of Internet‐based supportive interventions for caregivers of patients with dementia"
},
{
"paperId": "162bd6eadcdebdbc76c39a78c39333b6d763f31f",
"title": "Mental Models, Visual Reasoning and Interaction in Information Visualization: A Top-down Perspective"
},
{
"paperId": "61209d903884e8bc800e6f9ccdf61f26501dc257",
"title": "The Base16, Base32, and Base64 Data Encodings"
},
{
"paperId": "452f658a73f86a2220ac9d1163f6b01b686e428b",
"title": "Reasoning visualization in expert systems—the applicability of algorithm animation techniques"
},
{
"paperId": null,
"title": "https://hyperledger-fabric"
},
{
"paperId": null,
"title": "Beyond bitcoin"
},
{
"paperId": null,
"title": "Outpatient shift, digitization, shortages, and telehealth will shape healthcare systems of tomorrow"
},
{
"paperId": null,
"title": "Who should profit from the sale of patient data? The Brookings Institution"
},
{
"paperId": null,
"title": "A Blockchain Platform for the Enterprise"
},
{
"paperId": "36c8624c210b8e0accfc03ec68542d8ebf45cb12",
"title": "WITHDRAWN: An IoMT-based Cyber Training Framework for Orthopedic Surgery using Next Generation Internet Technologies"
},
{
"paperId": "aa156cc213266649b6809ac00be087a296f41143",
"title": "Privacy Preservation in e-Healthcare Environments: State of the Art and Future Directions"
},
{
"paperId": "33a0e8d78ccef17adaba0d9df3446c5d466c36c9",
"title": "Bancor Protocol Continuous Liquidity for Cryptographic Tokens through their Smart Contracts"
},
{
"paperId": null,
"title": "Medicalchain"
},
{
"paperId": "a3f9127171ada4280b5a8f6803b1348515b37816",
"title": "Big data security and privacy in healthcare: A Review"
},
{
"paperId": "5fac7d95cbae8f435b482bc819353f27d342373c",
"title": "Internet of Medical Things (IOMT): Applications, Benefits and Future Challenges in Healthcare Domain"
},
{
"paperId": "00113e81ef3a179d74d988d72329d306eae78525",
"title": "Survey of consensus protocols on blockchain applications"
},
{
"paperId": "de0398060e9a5f525099447b4d8093b5480a41b3",
"title": "Agility : It rhymes with stability"
},
{
"paperId": "deaa536affe5da5c87536281508f2d9ef83e60d7",
"title": "Privacy and security problems of national health data warehouse: a convenient solution for developing countries"
},
{
"paperId": "3ed0db58a7aec7bafc2aa14ca550031b9f7021d5",
"title": "A Case Study for Blockchain in Healthcare : “ MedRec ” prototype for electronic health records and medical research data"
},
{
"paperId": "c1b189c81b6fda71a471adec11cfe72f6067c1ad",
"title": "A Blockchain-Based Approach to Health Information Exchange Networks"
},
{
"paperId": "4027dcacc881fcc33f009ec6f18629b1295b1be4",
"title": "Media Reinforcement for Psychological Empowerment in Chronic Disease Management"
},
{
"paperId": "628bb3b1472745ccfd1bd430e9f237f3b9dbf7e4",
"title": "Supporting trust in globally distributed software teams: the impact of visualized collaborative traces on perceived trustworthiness"
},
{
"paperId": "cf8b96dabf484fc28049eebd3afc86e50bba45c8",
"title": "The Digital Transformation of Healthcare: Current Status and the Road Ahead"
},
{
"paperId": "8fa45abcce7d8231c756af48810c6c797fea1f2a",
"title": "Transparency in Supply Chains: Is Trust a Limiting Factor?"
},
{
"paperId": null,
"title": "This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license"
},
{
"paperId": null,
"title": "Whitepaper 2.1"
},
{
"paperId": null,
"title": "Running your company at two speeds ; McKinsey Quarterly , December 2014"
},
{
"paperId": null,
"title": "Manage the transactions generated by di ff erent nodes"
},
{
"paperId": null,
"title": "Create the blockchain with the di ff erent network nodes, where each node corresponds to di ff erent users who will participate in data sharing"
}
] | 20,689
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Engineering",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/02071107d0474cbd1f4077016d3014e1c2c9974e
|
[
"Computer Science"
] | 0.88378
|
Trust but Verify: Cryptographic Data Privacy for Mobility Management
|
02071107d0474cbd1f4077016d3014e1c2c9974e
|
IEEE Transactions on Control of Network Systems
|
[
{
"authorId": "5716768",
"name": "Matthew W. Tsao"
},
{
"authorId": "37577482",
"name": "Kaidi Yang"
},
{
"authorId": "92858908",
"name": "Stephen Zoepf"
},
{
"authorId": "1696085",
"name": "M. Pavone"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"IEEE Trans Control Netw Syst"
],
"alternate_urls": null,
"id": "0d3564f0-947d-4124-b171-400399406075",
"issn": "2325-5870",
"name": "IEEE Transactions on Control of Network Systems",
"type": "journal",
"url": "http://ieeexplore.ieee.org/servlet/opac?punumber=6509490"
}
|
The era of big data has brought with it a richer understanding of user behavior through massive datasets, which can help organizations optimize the quality of their services. In the context of transportation research, mobility data can provide municipal authorities (MAs) with insights on how to operate, regulate, or improve the transportation network. Mobility data, however, may contain sensitive information about end users and trade secrets of mobility providers (MPs). Due to this data privacy concern, MPs may be reluctant to contribute their datasets to MA. Using ideas from cryptography, we propose an interactive protocol between an MA and an MP, in which MA obtains insights from mobility data without MP having to reveal its trade secrets or sensitive data of its users. This is accomplished in two steps: 1) a commitment step and 2) a computation step. In the first step, Merkle commitments and aggregated traffic measurements are used to generate a cryptographic commitment. In the second step, MP extracts insights from the data and sends them to MA. Using the commitment and zero-knowledge proofs, MA can certify that the information received from MP is accurate, without needing to directly inspect the mobility data. We also present a differentially private version of the protocol that is suitable for the large query regime. The protocol is verifiable for both MA and MP in the sense that dishonesty from one party can be detected by the other. The protocol can be readily extended to the more general setting with multiple MPs via secure multiparty computation.
|
## Trust but Verify: Cryptographic Data Privacy for Mobility Management
##### Matthew Tsao Stanford University
```
mwtsao@stanford.edu
Stephen Zoepf Lacuna Technologies
stephen.zoepf@lacuna.ai
```
##### Kaidi Yang Stanford University
```
kaidi.yang@stanford.edu
Marco Pavone Stanford University
pavone@stanford.edu
```
##### November 16, 2021
**Abstract**
The era of Big Data has brought with it a richer understanding of user behavior through massive data sets, which can help organizations optimize the quality of their services. In the context
of transportation research, mobility data can provide Municipal Authorities (MA) with insights
on how to operate, regulate, or improve the transportation network. Mobility data, however,
may contain sensitive information about end users and trade secrets of Mobility Providers (MP).
Due to this data privacy concern, MPs may be reluctant to contribute their datasets to MA.
Using ideas from cryptography, we propose an interactive protocol between a MA and a MP in
which MA obtains insights from mobility data without MP having to reveal its trade secrets or
sensitive data of its users. This is accomplished in two steps: a commitment step, and a computation step. In the first step, Merkle commitments and aggregated traffic measurements are
used to generate a cryptographic commitment. In the second step, MP extracts insights from
the data and sends them to MA. Using the commitment and zero-knowledge proofs, MA can
certify that the information received from MP is accurate, without needing to directly inspect
the mobility data. We also present a differentially private version of the protocol that is suitable
for the large query regime. The protocol is verifiable for both MA and MP in the sense that
dishonesty from one party can be detected by the other. The protocol can be readily extended
to the more general setting with multiple MPs via secure multi-party computation.
This research was supported by the National Science Foundation under CAREER Award CMMI-1454737. K.
Yang would like to acknowledge the support of the Swiss National Science Foundation (SNSF) Postdoc Mobility
Fellowship (P400P2 199332).
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
### Contents
**1** **Introduction** **4**
1.1 Statement of Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
**2** **Model & Problem Description** **7**
2.1 Transportation Network Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Objective: Privacy for Mobility Management (PMM) . . . . . . . . . . . . . . . . . . 8
2.2.1 Regulation Compliance for Mobility Providers . . . . . . . . . . . . . . . . . . 10
2.2.2 Transportation Infrastructure Development Projects . . . . . . . . . . . . . . 10
2.2.3 Congestion Pricing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
**3** **A high level description of the protocol** **12**
**4** **The Protocol** **13**
4.1 Protocol Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
4.2 Ensuring accuracy of σ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4.2.1 Rider Witness: Detecting underreported demand . . . . . . . . . . . . . . . . 17
4.2.2 Aggregated Roadside Audits: Detecting overreported demand . . . . . . . . . 18
4.2.3 Implementation details for ARA . . . . . . . . . . . . . . . . . . . . . . . . . 20
**5** **Discussion** **21**
**6** **Conclusion** **22**
**A Incorporating Differential Privacy for the Large Query Regime** **26**
A.1 Goal: Differential Privacy without Trust . . . . . . . . . . . . . . . . . . . . . . . . . 27
A.2 A Differentially Private version of the protocol . . . . . . . . . . . . . . . . . . . . . 27
**B Supplementary Material** **30**
B.1 Mobility Provider Serving Demand . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
B.2 Mobility Provider Serving Demand (Steady State) . . . . . . . . . . . . . . . . . . . 31
B.3 Cryptographic Tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
B.3.1 Cryptographic Hash Functions . . . . . . . . . . . . . . . . . . . . . . . . . . 31
B.3.2 Cryptographic Commitments . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
B.3.3 Merkle Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
B.3.4 Merkle Proofs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
B.3.5 Digital Signatures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
B.3.6 Public Key Encryption . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
B.3.7 Zero Knowledge Proofs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
B.3.8 zk-SNARKs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
B.4 Implementation Details and Examples . . . . . . . . . . . . . . . . . . . . . . . . . . 36
B.4.1 Obtaining ridehailing period activity . . . . . . . . . . . . . . . . . . . . . . . 36
B.4.2 Evaluating contributions to congestion . . . . . . . . . . . . . . . . . . . . . . 38
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
B.5 Necessity of Assumption 1 for Verifiability . . . . . . . . . . . . . . . . . . . . . . . . 38
B.6 Roadside Audits with fewer sensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
B.6.1 Security Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
B.7 Establishing Verifiability and Differential Privacy for Appendix A . . . . . . . . . . . 41
B.8 More Details on Congestion Pricing . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
B.9 Efficacy of Merkle Proofs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
### 1 Introduction
The rise of mobility as a service, smart vehicles and smart cities is revolutionizing transportation industries all over the world. Mobility management, which entails operation, regulation, and
innovation of transportation systems, can leverage mobility data to improve the efficiency, safety,
accessibility, and adaptability of transportation systems far beyond what was previously achievable.
The analysis and sharing of mobility data, however, introduces two key concerns. The first concern
is data privacy; sharing mobility data can introduce privacy risks to end users that comprise the
datasets. The second concern is credibility; in situations where data is not shared, how can the
correctness of numerical studies be verified? These concerns motivate the need for data analysis
tools for transportation systems which are both privacy preserving and verifiable.
The data privacy issue in transportation is a consequence of the trade-off between data availability and data privacy. While user data can be used to inform infrastructure improvement, equity
and green initiatives, the data may contain sensitive user information and trade secrets of mobility
providers. As a result, end users and mobility providers may be reluctant to share their data with
city authorities. Cities have recently begun mandating micromobility providers to share detailed
trajectory data of all trips, arguing that the data is needed to enforce equity or environmental
objectives. Some mobility providers argued that while names and other directly identifiable information may not be included in the data, trajectory data can still reveal schedules, routines and
habits of the city’s inhabitants. The mobility providers’ concern over the release of anonymized
data is justified. [1] showed that any attempt to release anonymized data either fails to provide
anonymity, or there are low-sensitivity attributes of the original dataset that cannot be determined
from the published version. In general, anonymization is increasingly easily defeated by the very
techniques that are being developed for many legitimate applications of big data [2]. Such disputes
highlight the need for privacy-preserving data analysis tools in transportation.
A communication scheme between a sender and a receiver is verifiable if it enables the receiver
to determine whether the message or report it receives is an accurate representation of the truth.
When the objectives of mobility providers and policy makers are not aligned, one party may benefit
from misreporting data or other information, giving rise to verifiability issues in transportation. An
example of this is Greyball software [3]. Mobility providers developed Greyball software to deny
service or display misleading information to targeted users. It was originally developed to protect
their drivers from oppressive authorities in foreign countries, by misreporting driver location to
accounts that were believed to belong to the oppressive authorities. However, mobility providers
also used Greyball to hide their activity from authorities in the United States when their operations
were scrutinized. Another example of verifiability issues is third party wage calculation apps [4].
Drivers, frustrated by instances of being underpaid, created an app to confirm whether the pay
was consistent with the length and duration of each trip. Such incidents highlight the need for
verifiable data analysis tools in transportation.
##### 1.1 Statement of Contributions
In this paper we propose a protocol between a Municipal Authority and a Mobility Provider that
enables the Mobility Provider to send insights from its data to the Municipal Authority in a privacypreserving and verifiable manner. In contrast to non-interactive data sharing mechanisms (which
are currently used by most municipalities) where a Municipal Authority is provided an aggregated
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
Figure 1: The Mobility Provider can answer the Municipal Authority’s data-related mobility queries
in a verifiable way without needing to share the data. The absence of data sharing in the protocol
reduces the chance that a malicious third party intercepts and uses the data for nefarious privacyinvasive purposes.
and anonymized version of the data to analyze, our proposed protocol is an interactive mechanism
where a Municipal Authority sends queries and Mobility Providers give responses. By sharing
responses to queries rather than the entire dataset, interactive mechanisms circumvent the data
anonymization challenges faced by non-interactive approaches [1, 2].
Our proposed protocol, depicted in Figure 1, has three main steps. In the first step, the Mobility
Provider uses its data to produce a data identifier which it sends to the Municipal Authority. The
Municipal Authority can then send its data query to the Mobility Provider in the second step. In
the third step, the Mobility Provider sends its response along with a zero knowledge proof. The
Municipal Authority can use the zero knowledge proof to check that the response is consistent with
the identifier, i.e., the response was computed from the same data that was used to create the
identifier. If the Municipal Authority has multiple queries, steps 2 and 3 are repeated.
The protocol uses cryptographic commitments and aggregated traffic measurements to ensure
that the identifier is properly computed from the true mobility data. In particular, any deviation
from the protocol by one party can be detected by the other, making the protocol strategyproof for
both parties. Given that the identifier is properly computed, the zero knowledge proof then enables
the Municipal Authority to verify the correctness of the response without needing to directly inspect
the mobility data. Since the Municipal Authority never needs to inspect the mobility data, the
protocol is privacy-preserving.
The protocol can be extended to the more general case of multiple Mobility Providers, each
with a piece of the total mobility data. This is done by including a secure multi-party computation
in step 3 of the protocol. Answering a large number of queries with our protocol can lead to privacy
issues since it was shown in [5] that a dataset can be reconstructed from many accurate statistical
measurements. To address this concern, we generalize the protocol to enable differentially private
responses from the Mobility Provider in large query regimes.
##### 1.2 Organization
This paper is organized as follows. The remainder of the introduction discusses academic work
related to privacy and verifiability in transportation networks. In Section 2 we introduce a mathematical model of transportation networks and use it to formulate the data privacy problem for
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
Mobility Management. We provide a high level intuitive description of our proposed protocol in
Section 3. In Section 4 we provide a full technical description of our protocol. We discuss some
of the technical nuances of the protocol and their implications in Section 5. We summarize our
work and identify important areas for future research in Section 6. In Appendix A we present a
differentially private extension of the protocol that is suitable for the large query regime.
##### 1.3 Related Work
Within the academic literature, this work is related to the following four fields: misbehavior detection in cooperative intelligent transportation networks, data privacy in transportation systems,
differential privacy, and secure multi-party computation. We briefly discuss how this work complements ideas from these fields.
Cooperative intelligent transportation networks (cITS) aim to provide benefits to the safety,
efficiency, and adaptability of transportation networks by having individual vehicles share their
information. As with all decentralized systems, security and robustness against malicious agents
is essential for practical deployment. As such, misbehavior detection in cITS have been studied
extensively [6]. Misbehavior detection techniques often rely on honest agents acting as referees, and
are able to detect misbehavior in the honest majority setting. Watchdog is one such protocol [7, 8]
which uses peer-to-peer refereeing. The protocol uses a public key infrastructure (PKI) to assign a
persisting identity to each node in the network, and derives a reputation for each node based on its
historical behavior. Our objective in this work is also detection of misbehavior, but in a different
setting. In our setting, while the mobility network is comprised of many agents (customers and
drivers), there is a single entity (the Mobility Provider, e.g., a ridehailing service) who is responsible
for the storage and analysis of trip data. As such, the concept of honest majority does not apply
to our setting. Furthermore, [8] does not address the issue of data privacy; indeed, PKIs can often
expose the users’ identities, especially if an attacker cross-references the network traffic with other
traffic records.
Privacy in intelligent transportation systems is often implemented by using non-interactive
anonymization (e.g., data aggregation), cryptographic tools or differential privacy. Providing
anonymity in non-interactive data analysis mechanisms is challenging [1, 2] and thus data aggregation alone is often not enough to provide privacy. From the cryptography side, to address
the lack of anonymity provided by blockchains like Bitcoin and Ethereum, zero knowledge proofs
[9] were deployed in blockchains like Zcash [10] to provide fully confidential transactions. In the
context of transportation, zero knowledge proofs have been proposed for privacy-preserving vehicle
authentication to EV charging services [11], and privacy-preserving driver authentication to customers in ridehailing applications [12]. These privacy-preserving authentication systems rely on a
trusted third party to distribute and manage certificates.
Differential privacy is an interactive mechanism for data privacy which uses randomized responses to hide user-specific information [1]. For any query, the data collector provides a randomized
response, where two datasets which differ in only one entry produce statistically indistinguishable
outputs. Due to this randomization, there is a trade-off between the accuracy of the response and
the level of privacy provided. Randomization is necessary to preserve privacy in the large query
regime as demonstrated by [5] which showed that a dataset can be reconstructed from many accurate statistical measurements. The standard model of differential privacy, however, relies on a
_trusted data collector to apply the appropriate randomized response to queries. This is problematic_
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
in situations where the data collector is not trusted. A local model of differential privacy where
users perturb their data before sending it to the data collector has received significant attention
due to trust concerns [13]. However mobility providers often record exact details about user trips,
making local differential privacy unsuitable for current mobility applications (See Remark 15). Instead, we believe cryptographic techniques can be used to address trust concerns. There are also
more general concerns about trust; downstream applications of data queries can lead to conflicts
of interest and encourage strategic behavior.
Secure Multi-Party Computation (MPC) is a technique whereby several players, each possessing
private data, can jointly compute a function on their collective data without any player having to
reveal their data to other players [14]. MPC achieves confidentiality by applying Shamir’s Secret
Sharing [15] to inputs and intermediate results. In its base form, MPC is secure against honest-butcurious adversaries, which follow the protocol, but may try to do additional calculations to learn
the private data of other players. In general, security against active malicious adversaries, which
deviate from the protocol arbitrarily, requires a trusted third party to perform verified secret sharing
[16]. In verified secret sharing, the trusted third party creates initial cryptographic commitments
for each player’s private data. The commitments do not leak any information about the data, and
allows honest players to detect misbehavior using zero knowledge proofs. MPC is a very promising
tool for our problem, but a trusted third party able to eliminate strategic behavior does not yet
exist in the transportation industry, therefore a key objective of this work is to develop mechanisms
to defend against strategic behavior.
_In Summary - Our goal in this work is to develop a protocol that enables a mobility provider to_
share insights from its data to a municipal authority in a privacy-preserving and verifiable manner.
Existing work in accountability and misbehavior detection focus on networks with many agents and
rely on honest majority. Such assumptions, however, are not realistic for interactions between a
municipal authority and a few mobility providers. We thus turn our attention to differential privacy
and secure multi-party computation which provide data privacy but require honesty of participating
parties. To address this, we develop mechanisms based on cryptography and aggregated roadside
measurements to detect dishonest behavior.
### 2 Model & Problem Description
In this section we present a model for a city’s transportation network and formulate a data Privacy
for Mobility Management (PMM) problem. Section 2.1 introduces a mathematical representation
of a city’s transportation network along with the demand and mobility providers. In Section 2.2
we formalize the notion of data privacy using secure multi-party computation, and introduce assumptions on user behavior that we will need to construct verifiable protocols. We then formally
introduce the PMM problem and describe several transportation problems that can be formulated
in the PMM framework.
##### 2.1 Transportation Network Model
_Transportation Network - Consider the transportation network of a city, which we represent as_
a directed graph G = (V, E, f ) where vertices represent street intersections and edges represent
roads. For each road e ∈ _E we use an increasing differentiable convex function fe : R+ →_ R+ to
denote the travel cost (which may depend on travel time, distance, and emissions). of the road as
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
a function of the number of vehicles on the road. We will use n := _V_ and m := _E_ to denote the
_|_ _|_ _|_ _|_
total number of vertices and edges in G respectively. Time is represented in discrete timesteps of
size ∆t. The operation horizon is comprised of T + 1 timesteps as := 0, ∆t, 2∆t, ..., T ∆t .
_T_ _{_ _}_
_Mobility Provider - A Mobility Provider (MP) is responsible for serving the transportation de-_
mand. It does so by choosing a routing x of its vehicles within the transportation network. The
routing must satisfy multi-commodity network flow constraints (see Supplementary Material SM-I
and SM-II for explicit descriptions of these constraints) and the MP will choose a feasible flow
that maximizes its utility function JMP. Some examples of MPs are ridehailing companies, bus
companies, train companies, and micromobility (i.e., bikes & scooters) companies.
_Transportation Demand Data - The MP’s demand data is a list of completed trips Λ := {λ1, ..., λq},_
where λi contains the following basic metadata about the ith trip:
Pickup location, Dropoff location, Request time, Match time (i.e., the time at which the user
is matched to a driver), Pickup time, Dropoff time, Driver wage, Trip fare, Trip trajectory
(i.e., the vehicle’s trajectory from the time the vehicle is matched to the rider until the time
the rider is dropped off at their destination), Properties of the service vehicle.
For locations i, j _V and a timestep t, we use Λ(i, j, t) to denote the number of users in the data_
_∈_
set who request transit from location i to location j at time t.
**Remark 1 (Multiple Mobility Providers). We can consider settings where there are multiple mo-**
bility providers, MP1, MP2, ..., MPℓ, where Λj is the demand data of MPj. The demand data set
for the whole city is thus Λ = ∪j[ℓ]=1[Λ][j][.]
_Ridehailing Periods - For MPs that operate ridehailing services, a ridehailing vehicle’s trajectory_
is often divided into three different periods (with Period 0 often ignored):
_Period 0: The vehicle is not online with a platform. The driver may be using the vehicle_
personally.
_Period 1: The vehicle is vacant and has not yet been assigned to a rider._
_Period 2: The vehicle is vacant, but it has been assigned to a rider, and is en route to pickup._
_Period 3: The vehicle is driving a rider from its pickup location to its dropoff location._
##### 2.2 Objective: Privacy for Mobility Management (PMM)
In the data Privacy for Mobility Management (PMM) problem, a Municipal Authority (MA) wants
to compute a function g(Λ) on the travel demand, where g(Λ) is some property of Λ that can
inform MA on how to improve public policies. There are two main obstacles to address: privacy
and verifiability.
Privacy issues arise since trip information may contain sensitive customer information as well
as trade secrets of Mobility Providers (MP). For this reason MPs may be reluctant to contribute
their data for MA’s computation of g(Λ). This motivates the following notion of privacy:
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
**Definition 1 (Privacy in Multi-Party Computation). Suppose MP1, ...MPℓ** serve the demands
Λ1, ..., Λℓ respectively, and we denote Λ = ∪i[ℓ]=1[Λ][i][. We say a protocol for computing][ g][(Λ) between]
a MA and several MPs is privacy preserving if
1. MA learns nothing about Λ beyond the value of g(Λ).
2. For any pair i ̸= j, MPi learns nothing about Λj beyond the value of g(Λ).
Verifiability issues arise if there is incentive misalignment between the players. In particular, if
the MA or a MP can increase their utility by deviating from the protocol, then the computation of
_g(Λ) may be inaccurate. To address this issue, we need the protocol to be verifiable, as described_
by Definition 2. The following assumption is necessary to ensure accurate reporting of demand
(See Supplementary Material SM-V for more details):
**Assumption 1 (Strategic Behavior). We assume in this work that drivers and customers of the**
_transportation network will behave honestly (by this we mean they will always follow the protocol),_
_but MA and MPs may act strategically to maximize their own utility functions._
**Definition 2 (Verifiable Protocol). A protocol for computing g(Λ) is verifiable under Assumption 1**
if:
1. Any deviation from the protocol by the MA can be detected by the MPs provided that all
riders and drivers act honestly (i.e., follow the protocol).
2. Any deviation from the protocol by an MP can be detected by the MA provided that all
riders and drivers act honestly.
Our objective in this paper is to present a PMM protocol, which is defined below.
**Definition 3 (PMM Protocol). A PMM protocol between a MA and MP1, ...MPℓ** can, given
any function g, compute g(Λ) for MA while ensuring privacy and verifiability as described by
Definitions 1 and 2 respectively.
**Remark 2 (Admissible Queries and Differential Privacy). While a PMM protocol hides all infor-**
mation about Λ beyond the value of g(Λ), g(Λ) itself may contain sensitive information about Λ.
The extreme case would be if g is the identity function, i.e., g(Λ) = Λ. In such a case, the MPs
should reject the request to protect the privacy of its customers. More generally, MPs should reject
functions g if g(Λ) is highly correlated with sensitive information in Λ. The precise details as to
which functions g are deemed acceptable queries must be decided upon beforehand by MA and the
MPs together.
Differential privacy mechanisms provide a principled way to address the sensitivity of g by
having MPs include noise in the computation of g(Λ). If the noise distribution is chosen according
to both the desired privacy level and the sensitivity of g to its inputs, then the output is differentially
private. Note that this privacy is not for free; the noise reduces the accuracy of the output. The
precise choice of noise distribution is important for both the privacy and accuracy of this method,
so ensuring that the randomization step is conducted properly in the face of strategic MAs and MPs
is essential. This can be done with a combination of coinflipping protocols and secure multi-party
computation, which we describe in Appendix A.
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
**Remark 3 (A note on computational complexity). The applications we consider in this work**
do not impose strict requirements on computation times of protocols. Regulation checks can be
conducted daily or weekly, and infrastructure improvement initiatives are seldom more frequent
than one per week. The low frequency of such queries gives plenty of time to compute a solution.
For this reason, we do not expect the computational complexity of the solution to be an issue.
We now present some important social decision making problems that can be formulated within
the PMM framework.
**2.2.1** **Regulation Compliance for Mobility Providers**
Suppose MA wants to check whether a MP is operating within a set of regulations ρ1, ..., ρk. The
metadata contained within each trip includes request time, match time, pickup time, dropoff time,
and trip trajectory, which can be used to check regulation compliance. If we define the function
_ρi(Λ) to be 1 if and only if regulation i is satisfied, and 0 otherwise, then regulation compliance_
can be determined from the function g(Λ) := [�]t[k]=1 _[ρ][t][(Λ). Below are some examples of regulations]_
that can be enforced using trip metadata.
**Example 1 (Waiting Time Equity). MP is not discriminating against certain requests due to the**
pickup or droppoff locations. Specifically, the difference in average waiting time among different
regions should not exceed a specified regulatory threshold.
**Example 2 (Congestion Contribution Limit). The contribution of MP vehicles (in Period 2 or 3)**
to congestion should not exceed a specified regulatory threshold.
**Example 3 (Accurate Reporting of Period 2 Miles). A ridehailing driver’s pay per mile/minute**
depends on which period they are in. In particular, the earning rate for period 2 is often greater
than that of period 1. For this reason, mobility providers are incentivized to report period 2 activity
as period 1 activity. To protect ridehailing drivers, accurate reporting of period 2 activity should
be enforced.
**Example 4 (Emissions Limit). The collective emission rate of MP vehicles in Phases 2 and 3.**
should not exceed a specified regulatory threshold. MP emissions can be computed from the
metadata of served trips, in particular the trajectory and vehicle make and model.
See Supplementary Material SM-IV for further details on formulating the above examples within
the PMM framework.
**2.2.2** **Transportation Infrastructure Development Projects**
_Transportation Infrastructure Improvment Projects - A Municipal Authority (MA) measures the_
efficiency of the current transportation network via a concave social welfare function JMA(x). The
MA wants to make improvements to the network G through infrastructure improvement projects.
Below are some examples of such projects.
**Example 5 (Building new roads). The MA builds new roads Enew so the set of roads is now**
_E ∪_ _Enew, i.e., G now has more edges._
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
**Example 6 (Building Train tracks). The MA builds new train routes. Train routes differ from**
roads in that the travel time is independent of the number of passengers, i.e., there is no congestion
effect.
**Example 7 (Adding lanes to existing roads). The MA adds more lanes to some roads E[′]** _E. As_
_⊂_
a consequence, the shape of fe will change for each e ∈ _E[′]._
**Example 8 (Adjusting Speed limits). Similar to adding more lanes, adjusting the speed limit of**
a road will change its delay function.
_Evaluation of Projects - We measure the utility of a project using a Social Optimization Problem_
(SOP). An infrastructure improvement project θ makes changes to the transit network, so let Gθ
denote the transit network obtained by implementing θ. The routing problem ROUTE(θ, Λ) associated with θ is the optimal way to serve requests in Gθ as measured by MP’s objective function
_JMP. Letting Sθ,Λ be the set of flows satisfying multi-commodity network flow constraints (See Sup-_
plementary Material SM-I and SM-II for time-varying and steady state formulations respectively).
for the graph Gθ and demand Λ, ROUTE(θ, Λ) is given by
max JMP(x) (ROUTE(θ, Λ))
s.t. x ∈ _Sθ,Λ._
**Definition 4 (The Infrastructure Development Selection Problem). Suppose there are k infras-**
tructure improvement projects Θ := {θ1, θ2, ..., θk} available, but the city only has the budget for
one project. The city will want to implement the project that yields the most utility, which is
determined by the following optimization problem.
�
�
argmax _JMA_
1≤i≤k
argmax _JMP(x)_
_x∈Sθi,Λ_
_._ (SOP(Θ, Λ))
In the context of PMM, the function g associated with the infrastructure development selection
problem is g(Λ) := SOP(Θ, Λ).
**2.2.3** **Congestion Pricing**
Some ridehailing services allow drivers to choose the route they take when delivering customers.
When individual drivers prioritize minimizing their own travel time and disregard the negative
externalities they place on other travelers, the resulting user equilibrium can experience significantly more congestion than the social optimum. In these cases, the total travel time of the user
equilibrium is larger than that of the social optimum. This gap, known as the price of anarchy, is
well studied in the congestion games literature.
Congestion pricing addresses this issue by using road tolls to incentivize self-interested drivers
to choose routes so that the total travel time of all users is minimized. The desired road tolls
depend on the demand Λ, so MA would need help from MPs to compute the prices. Congestion
pricing can be formulated in the PMM framework through the query function gcp described in (2).
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
When the travel cost is the same as travel time, the prices can be obtained from the Traffic
Assignment Problem [17]:
�
min _xefe(xe)_ (1)
_e∈E_
� �
s.t. x = _x[od]_
_o∈V_ _d∈V_
_x[od]_ 0 _o_ _V, d_ _V_
_⪰_ _∀_ _∈_ _∈_
� _x[od](u,v)_ _[−]_ _[x]([od]v,u)_ [= Λ(][o, d][)] �1[u=o] − 1[u=d]� _∀u ∈_ _V_
(u,v)∈E
where x[od]e [denotes the traffic flow from][ o][ to][ d][ that uses edge][ e][. The objective measures the sum of]
the travel times of all requests in Λ. The desired prices are then given by:
_gcp(Λ) :=_ �x[∗]e[f]e[′][(][x][∗]e[)]�
(2)
_e∈E_ [where][ x][∗] [solves (1)][.]
See Supplementary Material SM-VIII for more details on congestion pricing.
### 3 A high level description of the protocol
We focus our discussion on the case where there is one MP. The protocol we will present can be
generalized to the multiple MP setting through secure Multi-party Computation [14]. The simplest
way for MA to obtain g(Λ) is via a non-interactive protocol where MP sends Λ to MA. MA could
then compute g(Λ) and any other attributes of Λ that it wants to know. This simple procedure,
however, does not satisfy data privacy, since MA now has full access to the demand Λ.
To address this concern, one could use an interactive protocol where MA sends a description of
the function g to MP, MP then computes g(Λ) and sends it to MA. This protocol does not require
MP to share the demand Λ. The problem with this approach is that there is no way for MA to
check whether MP computed g(Λ) properly, i.e., this approach is not verifiable. This is problematic
if there is an incentive for MP to act strategically, e.g., if MP wants to maximize its own revenue,
rather than social utility.
In this paper we present a verifiable interactive protocol, which allows MA to check whether or
not the message it receives from MP is in fact g(Λ). This will result in a protocol where MA is able
to obtain g(Λ) without requiring MP to reveal any information about Λ beyond the value of g(Λ).
First, we describe a non-confidential way to compute g(Λ). We will discuss how to make it
confidential in the next paragraph. MP will send a commitment σ = MCommit(Λ, r) of Λ to MA.
This commitment will enable MA to certify that the result given to it by MP is computed using
the true demand Λ. The commitment is confidential, meaning it reveals nothing about Λ, and is
binding, meaning that it will be inconsistent with any other demand Λ[′] = Λ. Now suppose MP
_̸_
computes a message z = g(Λ). To convince MA that the calculation is correct, MP will construct
a witness w := (Λ, r). When MA receives the message z and witness w, it will compute C(σ, z, w),
where C is an evaluation algorithm. C(σ, z, w) evaluates to True if
1. Rider Witness and Aggregated Roadside Audit checks are satisfied. (σ was reported honestly)
2. MCommit(Λ, r) = σ. (Λ is the demand that was used to compute σ).
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
3. g(Λ) = z (g was evaluated properly.)
If any of these conditions are not met, C(σ, z, w) will evaluate to False. Finally, MA will accept
the message z only if C(σ, z, w) = True.
The approach presented in the previous paragraph is not privacy-preserving because the witness
_w being sent from MP to MA includes the demand Λ. Fortunately, we can use zero knowledge proofs_
to obtain privacy. Given an arithmetic circuit C (which in our case is the evaluation algorithm C),
it is possible for one entity (the prover) to convince another entity (the verifier) that it knows an
input z, w so that C(σ, z, w) = True without revealing what w is. This is done by constructing
a zero knowledge proof π from (z, w) and sending (z, π) to the verifier instead of sending (z, w).
MA can then check whether π is a valid proof for z. The proof π is zero knowledge in the sense
that it is computationally intractable to deduce anything about w from π, aside from the fact
_C(σ, z, w) = True._ For our application, the prover will be MP who is trying to convince the
verifier, which is MA, that it computed g(Λ) correctly.
This protocol requires MP to send a commitment of the true demand data to MA. This is
problematic if MP has incentive to be dishonest, i.e., provide a commitment corresponding to a
different dataset. To ensure this does not happen, our protocol uses a Rider Witness incentive to
prevent MP from underreporting demand, and Aggregated Roadside Audits to prevent MP from
overreporting demand. These two mechanisms establish the verifiability of the protocol, since, as
seen in first requirement of C, MA will reject the message if either of these mechanisms detect
dishonesty.
_In Summary - We present a verifiable interactive protocol. First, MP sends a commitment_
of the demand to MA, which ensures that the report is computed using the true demand. The
correctness of this commitment is enforced by Rider Witness and Aggregated Roadside Audits.
MA then announces the function g that it wants to evaluate. MP computes a message z _g(Λ)_
_←_
and constructs a witness w to the correctness of z. Since w in general contains sensitive information,
it cannot be used directly to convince MA to accept the message z. MP computes a zero knowledge
proof π of the correctness of z from w, and sends the message z and proof π to MA. MA accepts z
if π is a valid zero knowledge proof for z.
_Implementation - To implement our protocol we will use several tools from cryptography. The_
commitment σ is implemented as a Merkle commitment. For computing zero knowledge proofs,
we will need a zk-SNARK that doesn’t require a trusted setup. PLONK [18], Sonic [19], and
Marlin [20] using a DARK based polynomial commitment schemes described in [21, 22]. Other
options include Bulletproofs [23] and Spartan [24]. The cryptographic tools used in the protocol
are reviewed in Supplementary Material SM-III.
### 4 The Protocol
In this section we present our protocol for the PMM problem described in Section 2.2. For clarity
and simplicity of exposition we will focus on the case where there is one Mobility Provider. The
single MP case can be extended to the multiple MP case via secure multi-party computation [14].
We present the protocol, which is illustrated in Figure 2, in Section 4.1. In Section 4.2 we discuss
mechanisms used to ensure verifiability of the protocol.
The protocol uses the following cryptographic primitives: hash functions, commitment schemes,
Merkle trees, public key encryption and zero knowledge proofs. Hash functions map data of ar
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
bitrary size to fixed size messages, often used to provide succinct identifiers for large datasets.
Commitment schemes are a form of verifiable data sharing where a receiver can reserve data from
a sender, obtain the data at a later point, and verify that the data was not changed between the
reservation and reception times. A Merkle tree is a particular commitment scheme we will use. In
public key encryption, every member of a communication network is endowed with a public key
and a private key. The public key is like a mailbox which tells senders how to reach the member,
and the secret key is the key to the mailbox, so messages can be viewed only by their intended
recipients. Zero knowledge proofs, as discussed in Section 3, enable a prover to convince a verifier
that it knows a solution to a mathematical puzzle without directly revealing its solution. For a more
detailed description of these concepts, we refer the reader to the Supplementary Material SM-III,
where we provide a self-contained introduction of the cryptographic tools used in this work.
##### 4.1 Protocol Description
The protocol entails 6 stages:
_Stage 0: (Data Collection) MP serves the demand Λ and builds a Merkle Tree TΛ of the demand_
it serves. MP publishes the root of TΛ, which is denoted as σ := MCommit(Λ, r) so that MA, all
riders and all drivers have access to σ. Here r is the set of nonces used to make the commitment
confidential.
_Stage 1: (Integrity Checks) MA instantiates Rider Witness and Aggregated Roadside Audits to_
ensure that σ was computed using the true demand Λ. The description of these mechanisms can
be found in Section 4.2.
_Stage 2: (Message Specifications) MA specifies to MP the function g it wants to compute._
_Stage 3: (zk-SNARK Construction) MA constructs an evaluation algorithm C for the function g._
_σ, z are public parameters of C, and the input to C is a witness of the form w = (Λw, rw, cw), where_
_rw is a set of nonces, Λw is a demand matrix, and cw is an optional input that may depend on g_
(See Remark 5). C does the following:
1. Checks whether the Rider Witness and Aggregated Roadside Audit tests are satisfied (This
checks that σ was reported honestly),
2. Checks whether MCommit(Λw, rw) = σ (This determines whether the provided demand Λw
is the same as the demand that created σ),
3. Checks whether g(Λw) = z (This checks that the message z is computed properly from Λw).
_C will evaluate to True if and only if all of those checks pass. Now, using one of the schemes from_
[18, 19, 20, 23, 24], MA will create a zk-SNARK (S, V, P ) for C. S is a set of public parameters
that describes the circuit C, P is a prover function which MP will use to construct a proof, and
_V is a verification function which MA will use to verify the correctness the MP’s proof. It sends_
_C, (S, V, P_ ), g to MP.
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
Figure 2: A block diagram of the communication between MA and MP.
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
_Stage 4: (Function Evaluation) If the request g is not a privacy-invasive function (see Remark 2),_
MP will compute a message z = g(Λ) and construct a witness w := (Λ, r, cw) to the correctness of z.
_Stage 5: (Creating a Zero Knowledge Proof) MP uses the zk-SNARK’s prover function P to con-_
struct a proof π := P (σ, z, w) that certifies the calculation of z. MP sends z, π to MA.
_Stage 6: (zk-SNARK Verification) MA uses the zk-SNARK’s verification function V (σ, z, π) to_
check whether MP is giving a properly computed message. If this is the case, MA accepts the
message z.
**Remark 4 (Computational Gains via Commit-then-Prove). Steps 2) and 3) of the evaluation**
circuit C involve different types of computation. This heterogeneity can introduce computational
overhead in the zk-SNARK. Commit-and-Prove zk-SNARKs [25, 26] are designed to handle computational heterogeneities, however existing implementations require a trusted setup.
**Remark 5 (Verifying solutions to convex optimization problems). If g(Λw) is the solution to**
a convex optimization problem parameterized by Λw, (e.g., g(Λw) = SOP(Θ, Λw) or congestion
pricing gcp(Λw)), then computing g(Λw) within the evaluation algorithm C may cause C to be
a large circuit, thus making evaluation of C computationally expensive. Fortunately, this can be
avoided by leveraging the structure of convex problems. If z = g(Λw), we can include the optimal
primal and dual variables associated with z in the optional input cw. This way, checking the
optimality of z can be done by checking that cw satisfy the KKT conditions rather than needing
to re-solve the problem.
##### 4.2 Ensuring accuracy of σ
The protocol presented in the previous section requires MP to share a commitment to the true
demand Λ. However, scenarios exist where the MP may face direct or indirect incentives to
misreport demand, such as per-ride fees, congestion charges, or other regulations that may constrain MP operations. In this section we present mechanisms to ensure that MP submits a commitment σ = MCommit(Λ, r) corresponding to the true demand Λ rather than a commitment
_σ[′]_ = MCommit(Λ[′], r) corresponding to some other demand Λ[′]. Specifically, we present Rider Witness and Aggregated Roadside Audits which detect underreporting and overreporting of demand
respectively.
The Rider Witness mechanism described in Section 4.2.1 prevents MP from omitting real trips
from its commitment. Under the Rider Witness mechanism, each rider is given a receipt for
their trip signed by MP. By signing a receipt, the trip is recognized as genuine by MP. Since
_σ[′]_ = MCommit(Λ[′], r) is a Merkle commitment, for each λ[′] Λ[′], MP can provide a proof that λ[′] is
_∈_
included in the calculation of σ[′]. Conversely, if λ Λ[′], MP is unable to forge a valid proof to claim
_̸∈_
that λ is included in the calculation of σ[′]. Therefore if there exists a genuine trip λ Λ that is not
_∈_
included in Λ[′], then that rider can report its receipt to MA. MP cannot provide a proof that λ was
included, and since the receipt of λ is signed by MP, this is evidence that MP omitted a genuine
trip from σ[′]. If this happens, MP is fined, and the reporting rider is rewarded.
The Aggregated Roadside Audit mechanism described in Section 4.2.2 prevents MP from adding
fictitious trips into its commitment. Due to Rider Witness, MP will not omit genuine trips, so
_σ[′]_ = MCommit(Λ[′], r) where Λ ⊆ Λ[′]. Recall that the trip metadata includes the trajectory. If Λ[′]
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
contains fictitious trips, then the road usage reported by Λ[′] will be greater than what happens in
reality. Thus if MA measures the number of passenger carrying vehicles that traverse each road,
then it will be able to detect if MP has included fictitious trips. However, auditing every road
can lead to privacy violations. Therefore, the audits are aggregated so that MA obtains the total
volume of passenger carrying traffic in the entire network, but not the per-road traffic information.
**4.2.1** **Rider Witness: Detecting underreported demand**
In this section, we present a Rider Witness mechanism to detect omission or tampering of the
demand Λ. Concretely, if a MP sends to MA a Merkle commitment σ[′] = MCommit(Λ[′], r) which
underreports demand, i.e., Λ Λ[′] is non-empty, then Rider Witness will enable MA to detect this.
_\_
MA can impose fines or other penalties when such detection occurs to deter MP from underreporting
the demand.
_Rider Witness Incentive Mechanism - At the beginning of Stage 0 (Data Collection) of the_
protocol, MP constructs a public key and private key pair (pkmp, skmp) to use for digital signatures.
The payment process is as follows: When the ith customer is delivered to their destination, the
customer will send a random nonce ri to MP. MP will respond with a receipt (H(ri||λi), σi), where
_σi := sign(skmp, H(ri||λi)) is a digital signature certifying that MP recognizes λi as an official ride_
(here || represents concatenation of binary strings). Here H is SHA256, so that H(ri||λi) is a
cryptographic commitment to the trip λi. The customer is required to pay the trip fare only if
verify(pkmp, H(ri||λi), σi) = True, i.e., they received a valid receipt.
**Definition 5 (Rider Witness Test). Given a commitment σ[′]** reported by MP to MA, each rider
who was served by MP requests a Merkle proof that their ride is included in the computation of
_σ[′]. If there exists a valid[1]_ ride receipt (H(ri||λi), σi) for which MP cannot provide a Merkle proof,
then the customer associated with λi will report (H(ri||λi), σi) to MA. MA checks if σi is a valid
signature for H(ri||λi), and if so, directly asks MP for a Merkle Proof that λi is included in the
computation of σ[′]. If MP is unable to provide the proof, then σ[′] fails the Rider Witness Test.
**Observation 1 (Efficacy of Rider Witness). Under Assumption 1, if MP submits a commitment**
_σ[′]_ = MCommit(Λ[′], r) which omits a ride, i.e., Λ Λ[′] _is non-empty, then σ[′]_ _will fail the Rider_
_\_
_Witness Test._
_Proof of Observation 1. If Λ ̸⊆_ Λ[′], then there exists some λi which is in Λ but not Λ[′]. Suppose
Alice was the rider served by ride λi. Forging a proof that λi ∈ Λ[′] requires finding a hash collision
for the hash function used in the Merkle commitment. Since MCommit is implemented using
a cryptographic hash function (e.g., SHA256), it is computationally intractable to find a hash
collision, and thus MP will be unable to forge a valid proof that λi ∈ Λ[′].
If MP does not provide Alice a valid proof within a reasonable amount of time (e.g., several hours), Alice can then report (H(ri||λi), σi) to MA. This reporting does not compromise
Alice’s privacy due to the hiding property of cryptographic hash functions. MA will check whether
verify(pkmp, H(ri||λi), σi) = True, and if so, means that λi is recognized as a genuine trip by MP.
MA will directly ask MP for a Merkle proof that H(ri||λi) ∈ _TΛ. Since MP cannot provide a valid_
proof, this is evidence that a genuine trip was omitted in the computation of σ[′], and hence σ[′] will
fail the Rider Witness test.
1In the sense that verify(pkmp, H(ri||λi), σi) = True.
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
**Remark 6 (Tamperproof Property). We note that Rider Witness also prevents the MP from**
altering the data associated with genuine rides. If MP makes changes to λi ∈ Λ resulting in some
_λ[′]i[, then by collision resistance of][ H][, it is computationally infeasible to find][ r][′][ so that][ H][(][r][i][||][λ][i][) =]_
_H(ri[′][||][λ]i[′][). If such a change is made, then][ H][(][r]i[′][||][λ]i[′][) is included into the computation of][ σ][′][ instead]_
of H(ri||λi). This means (H(ri||λi), σi) becomes a valid witness that data tampering has occurred.
**Remark 7 (Receipts are Unforgeable). Note that it is not possible for a rider to report a fake ride**
_λ[′]_ Λ to MA. This is because the corresponding signature σ[′] cannot be forged without knowing
_̸∈_
MP’s secret key skmp. Therefore, assuming skmp is only known to MP, only genuine trips can be
reported.
**Remark 8 (Honesty of riders). The Rider Witness mechanism assumes that riders are honest, i.e.,**
they will not collude with MP by accepting invalid receipts.
**4.2.2** **Aggregated Roadside Audits: Detecting overreported demand**
In this section we present an Aggregated Roadside Audit (ARA) mechanism to detect overreporting
of demand. Concretely, if MP announces a commitment σ[′] = MCommit(Λ[′], r), where Λ[′] is a strict
superset of Λ (i.e., Λ[′] Λ is non-empty), then ARA will enable MA to detect this. Thus between
_\_
ARA and Rider Witness, MA can detect if MP commits to a demand that is not Λ.
_Aggregated Roadside Audits - Due to the Rider Witness mechanism, we can assume that MP_
submits a commitment σ[′] computed from Λ[′] satisfying Λ Λ[′], i.e., Λ[′] is a superset of Λ. For an
_⊆_
edge e _E and a demand Λ, define_
_∈_
_ϕ(e, Λ) :=_ � 1[λ traverses e] (3)
_λ∈Λ_
to be the number of trips that traversed e during passenger pickup (Period 2) or passenger delivery
(Period 3). Since trip route is provided in the trip metadata, ϕ(e, Λ) can be computed from Λ.
**Definition 6 (ARA Test). The Aggregated Roadside Audit places a sensor on every road to**
conduct an audit on each road e _E to measure ϕ(e, Λ). These values are then aggregated as_
_∈_
_φ :=_ [�]e∈E _[ϕ][(][e,][ Λ). A witness][ w][ = (Λ][w][, r][w][, c][w][) passes the ARA test if and only if]_
�
_ϕ(e, Λw) = φ._ (ARA)
_e∈E_
**Observation 2 (Efficacy of Aggregated Roadside Audits). Under Assumption 1, if MP submits a**
_commitment σ[′]_ = MCommit(Λ[′], r) to a strict superset of the demand, i.e., Λ Λ[′], then any proof
_⊂_
_submitted by MP will either be inconsistent with σ[′]_ _or will fail the ARA test. Hence MP cannot_
_overreport demand._
_Proof of Observation 2. Suppose Λ[′]_ is a strict superset of Λ, which means that there exists some
_λ[′]_ Λ[′] Λ. Then there must exist some e[′] _E for which ϕ(e[′], Λ[′]) > ϕ(e[′], Λ). In particular, any_
_∈_ _\_ _∈_
edge in the trip route of λ[′] will satisfy this condition. With the inclusion of the ARA test, MP is
unable to provide a valid witness for MA’s evaluation algorithm C (and as a consequence, will be
unable to produce a valid zero knowledge proof) for the following reason:
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
Figure 3: An example of ARA. The true demand is Λ, which results in traffic shown on the left.
Here ϕ(eij, Λ) is the total number of trips in Λ that use the edge from i to j. Suppose MP submits
a commitment to Λ[′] = Λ _λ[′]_, i.e., inserts a fake trip λ[′] into the commitment. In this example,
_∪{_ _}_
_λ[′]_ is a fake trip from 5 to 2 that MP claims was served via the route {e56, e63, e31, e12} (shown in
red on the right). λ[′] increases the total traffic on the roads e56, e63, e31, e12 and as a result, we have
�
_e∈E_ _[ϕ][(][e,][ Λ][′][) =][ φ][ + 4.]_
1. MCommit is a collision-resistant function (since it is built using a cryptographic hash function
_H), so because σ[′]_ = MCommit(Λ[′], r), it is computationally intractable for MP to find Λ[′′] = Λ̸ _[′]_
and nonce values r[′′] so that MCommit(Λ[′′], r[′′]) = σ[′]. Therefore, in order to satisfy condition 2
of C (see Stage 3 of Section 4.1), MP’s witness must choose Λw to be Λ[′].
2. However, Λ[′] will not pass the ARA test. To see this, note that (a) Λ Λ[′] implies that
_⊆_
_ϕ(e, Λ)_ _ϕ(e, Λ[′]) for all e_ _E. Furthermore, (b) there exists an edge e[′]_ where the inequality
_≤_ _∈_
is strict, i.e., ϕ(e[′], Λ) < ϕ(e[′], Λ[′]). From this, we see that
� �
_φ =_ _ϕ(e, Λ) = ϕ(e[′], Λ) +_ _ϕ(e, Λ)_
_e∈E_ _e∈Λ,e≠_ _e[′]_
(a)
�
_ϕ(e[′], Λ) +_ _ϕ(e, Λ[′])_
_≤_
_e∈Λ,e≠_ _e[′]_
(b) �
_< ϕ(e[′], Λ[′]) +_ _ϕ(e, Λ[′])_
_e∈Λ,e≠_ _e[′]_
�
= _ϕ(e, Λ[′]),_
_e∈E_
i.e., if the witness passes condition 2 of C, then it will fail the ARA test.
Therefore the value of φ can be used to detect fictitious rides. See Figure 3 for a visualization of
ARA. In the following remark, we present a variant of ARA that is robust to measurement errors.
**Remark 9 (Error Tolerance in ARA). Trip trajectories are often recorded via GPS, so GPS**
errors can lead to inconsistencies between ARA sensor measurements and reported trajectories. To
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
prevent an honest MP from failing the ARA test due to GPS errors, one can use an error tolerant
version of the ARA test defined below
�
_φ −_ _ϕ(e, Λw)_ _≤_ _ϵφ_
����� _e∈E_ �����
where ϵ [0, 1] is a tuneable tolerance parameter to account for GPS errors while still detecting
_∈_
non-negligible overreporting of demand.
**Remark 10 (Honesty of Drivers). The correctness of ARA presented in Observation 2 assumes**
that drivers are honest when declaring their current period to ARA sensors, e.g., a driver who is
in period 3 will not report themselves as period 1 or 2.
Two challenges that arise in the computation of φ are privacy and honesty, which are described
below.
**Remark 11 (Privacy-Preserving computation of φ). The na¨ıve way to compute φ is for MA to**
collect the values ϕ(e, Λ) from each road. This, however, can compromise data privacy. Indeed, if
there is only 1 request in Λ, then measuring the number of customer carrying vehicles that traverse
each link exposes the trip route of that request: Edges that are traversed 1 time are in the route,
and edges that are traversed 0 times are not. More generally, observing ϕ(e, Λ) on all roads e _E_
_∈_
exposes trip routes to or from very unpopular locations.
**Remark 12 (Honest computation of φ). It is essential that MA acts truthfully when taking**
measurement and computing φ in ARA, otherwise MP will be wrongfully accused of dishonesty.
Fortunately, the ARA sensors can use public key encryption to share their data with each other
to compute φ in a privacy-preserving and honest way so that MA cannot learn ϕ(e, Λ) for any
_e_ _E even if it tries to eavesdrop on the communication between the sensors. After φ has been_
_∈_
sent to MA and the protocol has finished, the data on the sensors should be erased. We describe
this process in Section 4.2.3.
**4.2.3** **Implementation details for ARA**
In this section we describe the implementation details of ARA to ensure that the computation of
_φ is both privacy-preserving and accurate._
_ARA Sensors - To implement ARA, MA designs a sensor to detect MP vehicles. Concretely,_
the sensor records the current period of all MP vehicles that pass by. For communication, the
sensor will generate a random public and private key pair, and share its public key with the other
sensors. The sensor should have hardware to enable it to encrypt and decrypt messages it sends
and receives, respectively. To ensure honest auditing by MA, these sensors are inspected by MP
to ensure that they detect MP vehicles properly, key generation, encryption and decryption are
functioning properly, and that there are no other functionalities. Once the sensors have passed the
inspection, the following storage and communication restrictions are placed on them:
1. The device can only transmit data if it receives permission from both MP and MA.
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
Figure 4: An ARA sensor records the vehicle ID number, vehicle period and current timestamp of
each ridehailing vehicle that traverses the road. The dashed line around the sensor represents a
communication restriction: The sensor data can only be accessed with the consent of both parties.
2. The device can only transmit to addresses (i.e., public keys) that are on its sender whitelist.
The sender whitelist is managed by both MA and MP, i.e., an address can only be added
with the permission of MA and MP.
3. The device can only receive data from addresses that are on its receiver whitelist. The receiver
whitelist is managed by both MA and MP.
4. The device’s storage can be remotely erased with permission from both MP and MA.
_Deployment - To conduct ARA, a sensor is placed on every road, and will record the timestamp_
and period information of MP vehicles that pass by during the operation period. During operation,
both the sender and receiver whitelists should be empty. As a consequence, MA cannot retrieve
the sensor data. After the operation period ends and MP has sent a commitment σ to MA, MP
and MA conduct a coin flipping protocol to choose one sensor at random to elect as a leader (the
leader is elected randomly for the sake of robustness. If the leader is the same every time, then the
system would be unable to function if this sensor malfunctions or is compromised in any way). A
coin flipping protocol is a procedure where several parties can generate unbiased random bits. The
leader sensor’s public key is added to the whitelist of all other sensors, and all sensors are added
to the leader’s receiving whitelist. Each sensor then encrypts and sends its data under the leader
sensor’s public key. Since the MA does not know the leader sensor’s secret key, it cannot decrypt
the data even if it intercepts the ciphertexts. The addresses of MA and MP are then added to the
leader’s sender whitelist. The leader sensor decrypts the data, computes φ and reports the result
to both MA and MP. Once the protocol is over, the sender and receiver whitelists of all sensors
are cleared, and MA and MP both give permission for the sensors to delete their data. Figure 4
illustrates the sensor setup for ARA.
### 5 Discussion
The protocol requires minimal computational resources from the MA. Indeed, the computation of
_g(Λ), and all data analysis therein, is conducted by the MPs. The MA only needs to construct an_
evaluation circuit C and zk-SNARK (S, V, P ) for each of their queries g. In terms of data storage,
the MA only needs to store the commitments σ to the demand and the total recorded volume of
MP traffic φ for each data collecting period. If the Merkle Trees are built using the SHA256 hash
function, then σ is only 256 bits, and is thus easy to store. φ is a single integer, which is also easy
to store.
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
On the other hand, the hardware requirements for the Aggregated Roadside Audits may be
difficult for cities to implement, as placing a sensor on every road in the city will be expensive. To
address this concern, we present an alternative mechanism known as Randomized Roadside Audits
(RRA) in Supplementary Material SM-VI . RRA is able to use fewer sensors by randomly sampling
the roads to be audited, however as a tradeoff for using fewer sensors, overreported demand will
only be detected probabilistically. See Supplementary Material SM-VI for more details.
There is a trade-off between privacy and diagnosis when using zero knowledge proofs. In the
event that the zk-SNARK’s verification function fails, i.e., V (σ, z, π) = False, we know that z is
not a valid message, but we do not know why it is invalid. Specifically, V (σ, z, π) does not specify
which step of the evaluation algorithm C failed (See Stage 3 of Section 4.1). Thus in order to
determine whether the failure was due to integrity checks, inconsistency between Λ and σ, or a
mistake in the computation of g, further investigation would be required. Thus, while the zero
knowledge proof enables us to check the correctness of z without directly inspecting the data, it
does not provide any diagnosis in the event that z is invalid.
Multi-party computation is a natural way to generalize the proposed protocol to the multiple
MP setting. In such a case, the demand Λ = ∪i[k]=1[Λ][k][ is the disjoint union of Λ][1][, ...,][ Λ][k][, where Λ][i]
is the demand served by the ith MP, and is hence the private data of the ith MP. Multi-party
computation is a procedure by which several players can compute a function over their combined
data without any player learning the private data of other players. In the context of PMM with
multiple MPs, the MPs are the players and their private data is the Λi’s. In stage 0, each MP
would send to MA a commitment to its demand data, and the computation of z and π in stages
4 and 5 would be done using secure multi-party computation. Verifiability is established using
Rider Witness and ARA, as is done in the single MP case. See [27] and multiparty.org for an
open-source implementation of multi-party computation.
### 6 Conclusion
In this paper we presented an interactive protocol that enables a Municipal Authority to obtain
insights from the data of Mobility Providers in a verifiable and privacy-preserving way. During the
protocol, a Municipal Authority submits queries and a Mobility Provider computes responses based
on its mobility data. The protocol is privacy-preserving in the sense that the Municipal Authority
learns nothing about the dataset beyond the answer to its query. The protocol is verifiable in
the sense that any deviation from the protocol’s instructions by one party can be detected by
the other. Verifiability is achieved by using cryptographic commitments and aggregated roadside
measurements, and data privacy is achieved using zero knowledge proofs. We showed that the
protocol can be generalized to a setting with multiple Mobility Providers using secure multi-party
computation. We present a differentially private version of the protocol in Appendix A to address
situations where the Municipal Authority has many queries.
There are several interesting and important directions for future work. First, while this work
accounts for strategic behavior of the Municipal Authority and Mobility Providers, it assumes
that drivers and customers will act honestly. A more general model which also accounts for potential strategic behavior of drivers and customers would be of great value and interest. Second,
while secure multi-party computation can be used to generalize the protocol to settings with multiple Mobility Providers, generic tools for secure multi-party computation introduce computational
and communication overhead. Developing specialized multi-party computation tools for mobility
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
related queries is thus of significant practical interest. Finally, we suspect there are other applications for this protocol in transportation research beyond city planning and regulation enforcement
that could be investigated.
### References
[1] C. Dwork, F. McSherry, K. Nissim, and A. D. Smith, “Calibrating noise to sensitivity in private
data analysis,” in Theory of Cryptography Conference, vol. 3876, pp. 265–284, Springer, 2006.
[2] PCAST, “Big data and privacy: A technological perspective,” 2014.
[3] M. Isaac, “How uber deceives the authorities worldwide,” New York Times, Mar 2017.
[4] S. Szymkowski, “Google removes app that calculated if uber drivers were underpaid,” Road_Show, Feb 2021._
[5] I. Dinur and K. Nissim, “Revealing information while preserving privacy,” in ACM Symposium
_on Principles of Database Systems, pp. 202–210, ACM, 2003._
[6] R. W. van der Heijden, S. Dietzel, T. Leinm¨uller, and F. Kargl, “Survey on misbehavior
detection in cooperative intelligent transportation systems,” IEEE Commun. Surv. Tutorials,
vol. 21, no. 1, pp. 779–811, 2019.
[7] S. Marti, T. J. Giuli, K. Lai, and M. Baker, “Mitigating routing misbehavior in mobile ad hoc
networks,” in Proceedings of the 6th Annual International Conference on Mobile Computing
_and Networking, 2000._
[8] J. Hortelano, J. C. Ruiz, and P. Manzoni, “Evaluating the usefulness of watchdogs for intrusion
detection in vanets,” in 2010 IEEE International Conference on Communications Workshops,
pp. 1–5, 2010.
[9] S. Goldwasser, S. Micali, and C. Rackoff, “The knowledge complexity of interactive proof
systems,” SIAM J. Comput., vol. 18, no. 1, pp. 186–208, 1989.
[10] E. B. Sasson, A. Chiesa, C. Garman, M. Green, I. Miers, E. Tromer, and M. Virza, “Zerocash:
Decentralized anonymous payments from bitcoin,” in 2014 IEEE Symposium on Security and
_Privacy, pp. 459–474, 2014._
[11] D. Gabay, K. Akkaya, and M. Cebe, “Privacy-preserving authentication scheme for connected
electric vehicles using blockchain and zero knowledge proofs,” IEEE Transactions on Vehicular
_Technology, vol. 69, no. 6, pp. 5760–5772, 2020._
[12] W. Li, C. Meese, H. Guo, and M. Nejad, “Blockchain-enabled identity verification for safe
ridesharing leveraging zero-knowledge proof,” arXiv preprint arXiv:2010.14037, 2020.
[13] S. P. Kasiviswanathan, H. K. Lee, K. Nissim, S. Raskhodnikova, and A. D. Smith, “What can
we learn privately?,” SIAM J. Comput., vol. 40, no. 3, pp. 793–826, 2011.
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
[14] O. Goldreich, S. Micali, and A. Wigderson, “How to play any mental game or A completeness
theorem for protocols with honest majority,” in STOC 1987, New York, New York, USA (A. V.
Aho, ed.), pp. 218–229, ACM, 1987.
[15] A. Shamir, “How to share a secret,” Commun. ACM, vol. 22, no. 11, pp. 612–613, 1979.
[16] B. Chor, S. Goldwasser, S. Micali, and B. Awerbuch, “Verifiable secret sharing and achieving
simultaneity in the presence of faults,” in FOCS 1985, pp. 383–395, IEEE Computer Society,
1985.
[17] Y. Sheffi, Urban Transportation Networks: Equilibrium Analysis with Mathematical Program_ming Methods. Prentice-Hall, Englewood Cliffs, New Jersey, 1 ed., 1985._
[18] A. Gabizon, Z. J. Williamson, and O. Ciobotaru, “PLONK: permutations over lagrangebases for oecumenical noninteractive arguments of knowledge,” IACR Cryptol. ePrint Arch.,
vol. 2019, p. 953, 2019.
[19] M. Maller, S. Bowe, M. Kohlweiss, and S. Meiklejohn, “Sonic: Zero-knowledge snarks from
linear-size universal and updatable structured reference strings,” in ACM Conference on Com_puter and Communications Security, CCS, pp. 2111–2128, ACM, 2019._
[20] A. Chiesa, Y. Hu, M. Maller, P. Mishra, N. Vesely, and N. P. Ward, “Marlin: Preprocessing
zksnarks with universal and updatable SRS,” in Advances in Cryptology - EUROCRYPT 2020,
vol. 12105 of Lecture Notes in Computer Science, pp. 738–768, Springer, 2020.
[21] B. B¨unz, B. Fisch, and A. Szepieniec, “Transparent snarks from DARK compilers,” in Advances
_in Cryptology - EUROCRYPT 2020, vol. 12105 of Lecture Notes in Computer Science, pp. 677–_
706, Springer, 2020.
[22] A. R. Block, J. Holmgren, A. Rosen, R. D. Rothblum, and P. Soni, “Time- and spaceefficient arguments from groups of unknown order,” in Advances in Cryptology - CRYPTO
_2021, vol. 12828 of Lecture Notes in Computer Science, pp. 123–152, Springer, 2021._
[23] B. B¨unz, J. Bootle, D. Boneh, A. Poelstra, P. Wuille, and G. Maxwell, “Bulletproofs: Short
proofs for confidential transactions and more,” in IEEE Symposium on Security and Privacy,
_SP, pp. 315–334, IEEE Computer Society, 2018._
[24] S. T. V. Setty, “Spartan: Efficient and general-purpose zksnarks without trusted setup,” in
_Advances in Cryptology - CRYPTO 2020, vol. 12172 of Lecture Notes in Computer Science,_
pp. 704–737, Springer, 2020.
[25] M. Campanelli, D. Fiore, and A. Querol, “Legosnark: Modular design and composition of succinct zero-knowledge proofs,” in ACM Conference on Computer and Communications Security,
_CCS, pp. 2075–2092, ACM, 2019._
[26] M. Campanelli, A. Faonio, D. Fiore, A. Querol, and H. Rodr´ıguez, “Lunar: a toolbox for more
efficient universal and updatable zksnarks and commit-and-prove extensions,” IACR Cryptol.
_ePrint Arch., p. 1069, 2020._
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
[27] A. Lapets, F. Jansen, K. D. Albab, R. Issa, L. Qin, M. Varia, and A. Bestavros, “Accessible
privacy-preserving web-based data analysis for assessing and addressing economic inequalities,”
in Conference on Computing and Sustainable Societies, ACM, 2018.
[28] M. Blum, “Coin flipping by telephone - A protocol for solving impossible problems,” in COM_PCON’82, pp. 133–137, IEEE Computer Society, 1982._
[29] T. P. Pedersen, “Non-interactive and information-theoretic secure verifiable secret sharing,”
in Advances in Cryptology - CRYPTO (J. Feigenbaum, ed.), vol. 576 of Lecture Notes in
_Computer Science, pp. 129–140, Springer, 1991._
[30] A. Narayanan, J. Bonneau, E. W. Felten, A. Miller, and S. Goldfeder, Bitcoin and Cryptocur_rency Technologies - A Comprehensive Introduction. Princeton University Press, 2016._
[31] D. Boneh and V. Shoup, A Graduate Course in Applied Cryptography. 2020.
[[32] V. Buterin, “Zk rollup.” Available Online, 2016.](https://ethresear.ch/t/on-chain-scaling-to-potentially-500-tx-sec-through-mass-tx-validation/3477)
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
### A Incorporating Differential Privacy for the Large Query Regime
One potential concern with the protocol described in Section 4 arises in the large query regime. It
was shown in [5] that a dataset can be reconstructed from many accurate statistical measurements.
One way to address this is to set a limit on the number of times the MA can query the data for
a given time period. Such a restriction would not lead to data scarcity since the MP is collecting
new data daily. Differential privacy offers a principled way to determine how many times MA
should query a dataset (see Remark 13). Differentially private mechanisms address the result of
[5] by reducing the accuracy of the responses to queries, i.e., responding to a query g with a noisy
version of g(Λ). In this section we describe how the protocol from section 4 can be generalized
to facilitate verifiable and differentially private responses from MP. To this end we first define
differential privacy.
**Definition 7 (Datasets and Adjacency). A dataset Λ is a set of datapoints. In the context of**
transportation demand, a datapoint is the metadata corresponding to a single trip. We say two
datasets Λ, Λ[′] are adjacent if either (a) Λ Λ[′] with Λ[′] containing exactly 1 more datapoint than
_⊂_
Λ, or (b) Λ[′] Λ with Λ containing exactly 1 more datapoint than Λ[′].
_⊂_
**Definition 8 (Differential Privacy). Let** be a σ-algebra on a space Ω. A mechanism M : Ω
_F_ _D →_
is (ϵ, δ)-differentially private if for any two adjacent datasets Λ, Λ[′] and any -measurable event
_∈D_ _F_
_S,_
P (M (Λ) ∈ _S) ≤_ _e[ϵ]P_ �M (Λ[′]) ∈ _S�_ + δ.
In words, the output of a (ϵ, δ)-differentially private mechanism on Λ is statistically indistinguishable from the output of the mechanism on Λ _λ_ for any single datapoint λ Λ. Since Λ
_∪{_ _}_ _̸∈_
does not contain λ, M (Λ) does not reveal any information about λ. Since M (Λ _λ_ ) is statistically
_∪{_ _}_
indistinguishable from M (Λ), M (Λ _λ_ ) does not reveal much about λ.
_∪{_ _}_
**Example 9 (Laplace Mechanism for Vote Tallying). Suppose a city is trying to decide whether to**
expand its railways or expand its roads based on a majority vote from its citizens. The dataset is
Λ := {λ1, ..., λn} where λi is a boolean which is 0 if the ith citizen prefers the railway and 1 if the ith
citizen prefers the roads. To implement majority vote, the city needs to compute g(Λ) := [�]i[n]=1 _[λ][i][.]_
The Laplace Mechanism achieves (ϵ, 0)-differential privacy for this computation via
_Mlaplace(Λ) := Y +_
_n_
�
_λi,_
_i=1_
where Y has the discrete Laplace distribution: for any k ∈ Z, P[Y = k] ∝ _e[−][ϵ][|][k][|]. To see why this_
achieves (ϵ, 0)-differential privacy, for any 1 _j_ _n, note that_
_≤_ _≤_
P[M (Λ) = k] _i=1_ _[λ][i][|]_
P[M (Λ \ {λj}) = k] [=][ e]e[−][−][ϵ][ϵ][|][|][k][k][−][−][�][�][n]i≠ _j_ _[λ][i][| ≤]_ _[e][ϵλ][j][ ≤]_ _[e][ϵ][.]_
Note that the noise distribution for Y depends only on ϵ, and is independent of n, the size of the
dataset.
**Remark 13 (Privacy Budget). By composition rules, the result of k queries to a (ϵ, 0)-differentially**
private mechanism is (kϵ, 0)-differentially private. Thus a dataset should only be used to answer k
separate (ϵ, 0)-differentially private queries if e[kϵ] is sufficiently close to 1.
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
##### A.1 Goal: Differential Privacy without Trust
Given a query function g from MA, let M be an polynomial-time computable (ϵ, δ)-differentially
private mechanism for computing g. For a given dataset Λ we can represent the random variable
_M_ (Λ) with a function _g(Λ, Z) where Z_ 0, 1 represents the random bits used by M . Here
� _∈{_ _}[v]_
_v is an upper bound on the number of random bits needed for the computation of M_ . By its
construction, _g(Λ, Z) is (ϵ, δ)-differentially private if Z is drawn uniformly at random over_ 0, 1 .
� _{_ _}[v]_
Therefore differential privacy is achieved if MP draws Z uniformly at random over 0, 1 and
_{_ _}[v]_
sends _g(Λ, Z) to MA. However, as mentioned in Assumption 1, we are studying a model where_
�
MP can act strategically. Thus we cannot assume that MP will sample Z uniformly at random if
there is some other distribution over Z that leads to a more favorable outcome for MP. We revisit
Example 9 to illustrate this concern.
**Example 10 (Dishonest Vote Tallying). Consider the setting from Example 9.** The Laplace
mechanism can be represented as
�
_,_
_g(Λ, Z) := Y +_
�
_n_
�
_λi, where Y = Flaplace[−][1]_
_i=1_
� int(Z)
2[v]
where int(Z) is the integer whose binary representation is the bits of Z. Here Flaplace[−][1] [is the]
inverse cumulative distribution for the discrete Laplace distribution. Thus Flaplace[−][1] [(int(][Z][)][/][2][v][) is]
an application of inverse transform sampling that converts a uniform random variable Z into a
random variable Y with a discrete Laplace distribution. Suppose the MP has a ridehailing service
and would thus prefer an upgrade to city roads over an upgrade to the railway system. If this is the
case, choosing Z so that _g(Λ, Z) > n/2 (as opposed to choosing Z randomly) is a weakly dominant_
�
strategy for MP, even if g(Λ) < n/2 and a majority of the citizens prefer railway upgrades.
Thus we need a way to verify that the randomness Z used in MP’s evaluation of g(Λ, Z) has
the correct distribution. We will now show how the protocol can be adjusted to accommodate this,
and as a consequence, enable verifiable differentially private data queries for MA.
**Remark 14 (MA provided randomness). One natural attempt to ensure that Z is uniformly**
random is to have MA specify Z. However, this destroys the differential privacy, since for some
mechanisms (including the Laplace mechanism) g(Λ) can be computed from _g(Λ, Z) and Z. Also,_
�
it is not clear a priori whether such a setup is strategyproof for MA.
##### A.2 A Differentially Private version of the protocol
In this section, we present modifications to the protocol from Section 4.1 that enables verifiable
differentially private responses from MP. At a high level, the MA and MP jointly determine the
random bits Z via a coin flipping protocol [28]. The zk-SNARK can then be modified to ensure
that _g(Λ, Z) is computed correctly. The protocol has a total of 6 stages which are described below._
�
_Stage 0: (Data Collection) MP builds a Merkle Tree TΛ of the demand Λ that it serves. It computes_
a commitment σ := MCommit(Λ, r) to this demand. Additionally, MP samples Zmp uniformly at
random from {0, 1}[v] and computes a Pedersen commitment [29] zmp := Commit(Zmp, rmp). The
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
Pederson commitment scheme is a secure commitment scheme which is perfectly hiding and computationally binding. MP sends both σ, zmp to MA.
_Stage 1: (Integrity Checks) Same as in Section 4.1._
_Stage 2: (Message Specifications) MA specifies the function g it wants to compute. Additionally,_
MA samples Zma uniformly at random from {0, 1}[v] and specifies a differentially private mechanism
_g for the computation of g._
�
_Stage 3: (zk-SNARK Construction) MA constructs an evaluation circuit C for the function_ _g._
�
The public parameters of C are σ, zmp, Zma, z and the input to C is a witness of the form w =
(Λw, rw, cw, Zmp,w, rmp,w). C does the following:
1. Checks whether the Rider Witness and Aggregated Roadside Audit tests are satisfied,
2. Checks whether MCommit(Λw, rw) = σ,
3. Checks whether Commit(Zmp,w, rmp,w) = zmp,
4. Checks whether �g(Λw, Zma ⊕ _Zmp,w) = z. (Here ⊕_ is bit-wise XOR.)
_C will return True if and only if all of these checks pass. MA constructs a zk-SNARK (S, V, P_ ) for
_C and sends g, �g, Zma, C, (S, V, P_ ) to MP.
_Stage 4: (Function Evaluation) If_ _g is a differentially private mechanism for computing g, then MP_
�
computes a message z = �g(Λ, Zma ⊕ _Zmp) and a witness w := (Λ, r, cw, Zmp, rmp) to the correctness_
of z.
_Stage 5: (Creating a Zero Knowledge Proof) Same as in Section 4.1._
_Stage 6: (zk-SNARK Verification) Same as in Section 4.1._
In Supplementary Material SM-VII we show that this protocol has the following two desirable
features that enable verifiable and differentially private responses from MP to MA queries.
1. Verifiability - If the MA receives a valid proof from MP, then it can be sure that the corresponding message is indeed �g(Λ, Zma ⊕ _Zmp)._
2. Differential Privacy - The MP’s output is differentially private with respect to the dataset Λ
if at least one of Zma, Zmp is sampled uniformly at random.
**Remark 15 (A note on Local Differential Privacy). Local Differential Privacy [13] addresses the**
setting where the data collector is untrusted. Differential privacy is achieved by users adding noise
to their data before sending it to the data collector. This is in contrast to the setting we study
here where an untrusted data collector has the clean data of many users. We chose to study the
latter model due to the way current mobility companies collect high resolution data on the trips
they serve. Additionally, local differential privacy requires users to add noise to their data so
they become statistically indistinguishable from one another. In the context of transportation, this
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
means the noisy data of users will be statistically indistinguishable from one another, even if they
have very different travel preferences. This level of noise significantly reduces the accuracy of any
computation done on the data.
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
### B Supplementary Material
##### B.1 Mobility Provider Serving Demand
For a given discretization of time T := {0, ∆t, 2∆t, ..., T ∆t}, the demand Λ ∈ R[n][×][n][×][T] can be
represented as a 3-dimensional matrix (e.g., a 3-Tensor) where Λ(i, j, t) represents the number of
riders who request transit from i to j at time t. We use τij to represent the time it takes to travel
from i to j.
To serve the demand from i to j, the MP chooses passenger carrying flows x[ij] _∈_ R[mT]+ where
_x[ij]t_ [(][u, v][) is the number of passenger carrying trips from][ i][ to][ j][ that enter the road (][u, v][) at time][ t][.]
Such vehicles will exit the road at time t + τuv. There is also a rebalancing flow r ∈ R[mT] which
represents the movement of vacant vehicles that are re-positioning themselves to better align with
future demand. Concretely, rt(u, v) is the number of vacant vehicles which enter road (u, v) at time
_t. The initial condition is y ∈_ R[n]+[, where][ y][i] [denotes the number of vehicles at location][ i][ at time 0.]
The Mobility Provider’s routing strategy is thus x := ��x[ij][�] � which satisfies the
(i,j)∈V ×V _[, r]_
following multi-commodity network flow constraints:
(4)
�
rt(u, v) + _x[ij]t_ [(][u, v][)]
(i,j)∈V ×V
�
=
_v:(u,v)∈E_
�
_v:(v,u)∈E_
�
rt−τvu(v, u) + _x[ij]t−τvu[(][v, u][)]_
(i,j)∈V ×V
for all (u, t) _V_ [T ]
_∈_ _×_
� �
_x[ij]t_ [(][u, v][) =] _x[ij]t−τvu[(][v, u][) for all (][i, j][)][ ∈]_ _[V][ ×][ V, t][ ∈]_ [[][T] []][, u][ ̸∈{][i, j][}] (5)
_v:(u,v)∈E_ _v:(v,u)∈E_
_t_
�
Λ(i, j, τ ) for all (i, j, t) _V_ _V_ [T ] (6)
_∈_ _×_ _×_
_τ_ =0
_≤_
_t_
�
_τ_ =0
� �
_x[ij]τ_ [(][i, v][)][ −] _x[ij]τ_ _−τvi[(][v, i][)]_
_v:(i,v)∈E_ _v:(v,i)∈E_
_x[ij]t_ [(][j, v][) = 0 for all (][i, j][)][ ∈] _[V][ ×][ V, t][ ∈]_ [[][T] []][,][ (][j, v][)][ ∈] _[E.]_ (7)
�
_x[ij]0_ [=][ y][i][ for all][ i][ ∈] _[V.]_ (8)
_j:(i,j)∈E_
Here (4) represents conservation of vehicles, (5),(6),(7) enforce pickup and dropoff constraints
according to the demand Λ, and (8) enforces initial conditions.
The utility received by the Mobility provider (e.g., total revenue) from implementing flow x for
a given demand Λ is JMP(x; Λ). An optimal routing algorithm for demand Λ is a solution to the
following optimization problem.
maximize _JMP(x; Λ)_
_x_
s.t. (4), (5), (6), (7), (8).
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
##### B.2 Mobility Provider Serving Demand (Steady State)
In a steady state model, the demand can be represented as Λ ∈ R[n]+[×][n], a matrix where Λ(i, j)
represents the rate at which riders request transit from node i to node j.
For each origin-destination pair (i, j) _V_ _V, the MP serves the demand Λ(i, j) by choosing a_
_∈_ _×_
�� � �
passenger carrying flow x[ij]p _[∈]_ [R]+[m] [and a rebalancing flow][ x][r] _[∈]_ [R]+[m] [so that][ x][ :=] _x[ij]p_ (i,j)∈V ×V _[, x][r]_
satisfy the multi-commodity network flow constraints with demand Λ:
�
xr(v, u) + _x[ij]p_ [(][v, u][)]
(v,u)∈V ×V
for all u _V_
_∈_
(9)
�
=
_v:(v,u)∈E_
�
_v:(u,v)∈E_
�
xr(u, v) + _x[ij]p_ [(][u, v][)]
(i,j)∈V ×V
Λ(i, j)1[u=j] + � _x[ij]p_ [(][u, v][) = Λ(][i, j][)][1][u=i] [+] � _xr(v, u) + x[ij]p_ [(][v, u][) for all (][i, j][)][ ∈] _[V][ ×][ V, u][ ∈]_ _[V.]_
_v:(u,v)∈E_ _v:(v,u)∈E_
(10)
Here (9) represents conservation of flow and (10) enforces pickup and dropoff constraints according
to the demand Λ.
The utility received by the Mobility provider (e.g., total revenue) from implementing flow x is
_JMP(x). Therefore the Mobility Provider will choose x according to the following program._
maximize _JMP(x)_
_x_
s.t. (9), (10)
##### B.3 Cryptographic Tools
In this section we introduce existing cryptographic tools that are used in the protocol. The contents
of this section are discussed in greater detail in [30, 31, 9, 10, 18]. Throughout this paper, we use
_r_ _x to denote the concatenation of r and x._
_||_
**B.3.1** **Cryptographic Hash Functions**
**Definition 9 (Cryptographic Hash Functions). A function H is a d-bit cryptographic hash function**
if it is a mapping from binary strings of arbitrary length to 0, 1 and has the following properties:
_{_ _}[d]_
1. It is deterministic.
2. It is efficient to compute.
3. H is collision resistant - For sufficiently large d, it is computationally intractable to find
distinct inputs x1, x2 so that H(x1) = H(x2).
4. H is hiding - If r is a sufficiently long random string (256 bits is often sufficient), then it is
computationally intractable to deduce anything about x by observing H(r _x)._
_||_
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
Property 3 is called collision resistance and enables the hash function to be used as a digital
fingerprint. Indeed, since it is unlikely that two files will have the same hash value, H(x) can serve
as a unique identifier for x. We refer the interested reader to [30] for further details on cryptographic
hash functions.
[SHA256 is a widely used collision resistant hash function which has extensive applications](https://web.archive.org/web/20160330153520/http://www.staff.science.uu.nl/~werkh108/docs/study/Y5_07_08/infocry/project/Cryp08.pdf)
including but not limited to establishing secure communication channels, computing checksums for
online downloads, and computing proof-of-work in Bitcoin.
**B.3.2** **Cryptographic Commitments**
Cryptographic commitment schemes are tamper-proof communication protocols between two parties: a sender and a receiver. In a commitment scheme, the sender chooses (i.e., commits to) a
message. At a later time, the sender reveals the message to the receiver, and the receiver can be
sure that the message it received is the same as the original message chosen by the sender.
_Intuition - We can think of a commitment scheme as follows: A sender places a message into a_
box and locks the box with a key. The sender then gives the locked box to the receiver. Once the
sender has given the box away, the sender can no longer change the message inside the box. At
this point, the receiver, who does not have the key, cannot open the box to read the message. At
a later time, the sender can give the key to the receiver, allowing the receiver to read the message.
A commitment scheme is specified by a message space, nonce space, commitment space,
_M_ _R_ _X_
a commitment function commit : and a verification function verify : 0, 1 .
_M×R →X_ _M×R×X →{_ _}_
Creating a commitment to a message m happens in two steps:
_∈M_
1. Commitment Step - The sender computes σ := commit(m, r) for some r and gives σ to
_∈R_
the receiver.
2. Reveal Step - At some later time, the sender gives m, r to the receiver who accepts m as the
original message if and only if verify(m, r, σ) := 1[commit(m,r)=σ] evaluates to 1.
A secure commitment scheme has two important properties:
1. Binding - If σ is a commitment to a value m, it is computationally intractable to find m[′], r[′]
so that m[′] = m and commit(m[′], r[′]) = σ. Hence σ binds the committer to the value m.
_̸_
2. Hiding - It is computationally intractable to learn anything about m from σ.
Cryptographic hash functions can be used to build secure commitment schemes. To do this,
given a cryptographic hash function H, we define
commit(m, r) := H(r||m) and verify(m, r, σ) := 1[H(r||m)=σ].
The security of this commitment scheme comes from the properties of H. The binding property
of this commitment scheme follows directly from collision resistance of H. Furthermore, if r is
chosen uniformly at random from, then the commitment scheme is hiding due to the hiding
_R_
property of H.
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
**B.3.3** **Merkle Trees**
A Merkle tree is a data structure that is used to create commitments to a collection of items
_M := {m0, ..., mq−1}. A Merkle Tree has two main features:_
1. The root of the tree contains a hiding commitment to the entire collection M .
2. The root can also serve as a commitment to each item m _M_ . Furthermore, the proof that
_∈_
_m is a leaf of the tree reveals nothing about the other items in M and has length O(log q),_
where q is the total number of leaves in the tree.
A Merkle tree can be constructed from a cryptographic hash function. Concretely, given a
cryptographic hash function H and a collection of items m0, ..., mq−1, construct a binary tree with
these items as the leaves.
The leaves of the Merkle Tree are the zeroth level h0,0, h0,1, ..., h0,q−1, where h0,i = mi. The
next level has the same number of nodes h1,0, ..., h1,q−1 defined by h1,i = H(ri||mi) where ri
is a random nonce. Level k where k 2, has half as many nodes as level k 1, defined by
_≥_ _−_
_hk,i = H(hk−1,2i||hk−1,2i+1). Figure 5 illustrates an example of a Merkle tree. In total there are_
_ℓt +1 levels where ℓq := ⌈log2 q⌉_ +1. With this notation, hℓq,0 is the value at the root of the Merkle
Tree.
The root of a Merkle tree hℓq,0 is a commitment to the entire collection due to collision resistance
of H. To commit to the data M, the committer will generate r0, ..., rq−1, compute the Merkle Tree,
and announce hℓq,0. In the reveal step, the committer can announce {(mi, ri)}i[q]=0[−][1][, and anyone can]
then compute the resulting Merkle tree and confirm that the root is equal to hℓq,0.
**B.3.4** **Merkle Proofs**
The root also serves as a commitment to each mi _M_ . Suppose someone who knows mi wants a
_∈_
proof that mi is a leaf in the Merkle tree. A proof π(mi) can be constructed from the Merkle tree.
Furthermore, this proof reveals nothing about the other items {mj}j≠ _i._
Define x0, x1, ...xℓq recursively as:
_x0 := i_
_xj :=_ � _xj−1_ � for 1 ≤ _j ≤_ _ℓq._
2
With this notation, �hj,xj �ℓjq=0 [is the path from][ m][i][ to the root of the Merkle Tree. The Merkle]
proof for mi is denoted as π(mi) and is given by
_π(mi) := {ri} ∪_ �sibling(hj,xj )�ℓjq=1 _[,][ where]_
sibling(hi,j) := � _hhi,ji,j−+11_ ifif j j is even, is odd.
See Supplementary Material SM-IX for details on the binding and hiding properties of Merkle
commitments, and how to verify the correctness of Merkle proofs.
**Definition 10 (Merkle Commitment). Given a data set M = {m1, ..., mt} and a set of random**
nonce values r = {r1, ..., rt}, we use MCommit(M, r) to denote the root of the Merkle Tree constructed from the data M and random nonces r.
We refer the interested reader to Section 8.9 of [31] for more details on Merkle Trees.
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
Figure 5: An example of a Merkle tree containing 8 items. Each item mi is a leaf node and has
one parent which is H(ri||mi), where ri is a random hiding nonce. All other internal nodes are
computed by applying H to the concatenation of its children.
**B.3.5** **Digital Signatures**
A digital signature scheme is comprised by three functions: Gen, sign, verify. Gen() is a random
function that produces valid public and private key pairs (pk, sk). Given a message, a signature
is produced using the secret key via σ = sign(sk, m). The authenticity of the signature is checked
using the public key via verify(pk, m, σ). A secure digital signature scheme has two properties:
1. Correctness - For a valid key pair (pk, sk) obtained from Gen() and any message m, we have
verify(pk, m, sign(sk, m)) = True.
2. Secure - Given a public key pk, if the corresponding secret key sk is unknown, then it is
computationally intractable to forge a signature on any message. Specifically, if sk has never
been used to sign a message m[′], then without knowledge of sk, it is computationally intractable
to find (m[′], σ[′]) so that verify(pk, m[′], σ[′]) = True.
We refer the interested reader to Section 13 of [31] for more details on digital signatures.
**B.3.6** **Public Key Encryption**
A public key encryption scheme is specified by three functions: a key generation function, an
encryption function E, and a decryption function D. In a public key encryption scheme, each user
has a public key and private key denoted (pk, sk), produced by the key generation function. As
the name suggests, the public key pk is known to everyone, while each secret key sk is known only
by its owner. Encryption is done using public keys, and decryption is done using secret keys. To
send a message m to Bob, one would encrypt m using Bob’s public key via c = E(pkBob, m). Then
Bob would decrypt the message via D(skBob, c). A secure Public Key Encryption scheme has two
properties:
1. Correctness - For every valid key pair (pk, sk) and any message m, we have m = D(sk, E(pk, m)),
i.e., the intended recipient receives the correct message upon decryption.
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
2. Secure - For a public key pk and any message m, if the corresponding secret key sk is not
known, then it is computationally intractable to deduce anything about m from the ciphertext
_E(pk, m)._
The appeal of public key encryption is that users do not have to have a shared common key in
order to send encrypted messages to one another. We refer the interested reader to Part II of [31]
for more details on Public Key Encryption.
**B.3.7** **Zero Knowledge Proofs**
A zero knowledge proof for a mathematical problem is a technique whereby one party (the prover)
can convince another party (the verifier) that it knows a solution w to the problem without revealing
any information about w other than the fact that it is a solution. Before discussing zero knowledge
proofs further, we must first introduce proof systems.
**Definition 11 (Proof System). Consider an arithmetic circuit C :** 0, 1, and the
_X × W →{_ _}_
following optimization problem: For a fixed x, find a w so that C(x, w) = 0. Here x
_∈X_ _∈W_
is part of the problem statement, and w is a solution candidate. Consider a tuple of functions
(S, V, P ) where
1. S is a preprocessing function that takes as input C, x and outputs public parameters pp.
2. P is a prover function that takes as input pp, x, w and produces a proof π.
3. V is a verification function that takes as input pp, x, π and outputs either 0 or 1 corresponding
to whether the proof π is invalid or valid respectively.
The tuple (S, V, P ) is a proof system for C if it satisfies the following properties:
1. Completeness - If C(x, w) = 0, then V (pp, x, P (pp, x, w)) should evaluate to 1; i.e., the verifier
should accept proofs constructed from valid solutions w.
2. Proof of Knowledge - If V (pp, x, π) = 1, then whoever constructed π must have known a w
satisfying C(x, w) = 0.
With this definition in hand, we can now define zero knowledge proof systems.
**Definition 12 (Zero Knowledge Proof Systems). Consider a proof system (S, V, P** ) for the problem
of finding w so that C(x, w) = 0. (S, V, P ) is a zero knowledge proof system if it is computationally
intractable to learn anything about w from π := P (pp, x, w). If this is the case, then π is a zero
knowledge proof.
Zero knowledge proofs were first proposed by [9], but the prover and verifier functions were not
optimized to be computationally efficient. In the next section, we present zk-SNARKs, which are
computationally efficient zero knowledge proof systems.
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
**B.3.8** **zk-SNARKs**
In this section we introduce Succinct Non-interactive Arguments of Knowledge (SNARK). SNARKs
are proof systems where proofs are short, and both the construction and verification of proofs are
computationally efficient.
**Definition 13 (Succinct Non-interactive Argument of Knowledge (SNARK)). Consider the prob-**
lem of finding w so that C(x, w) = 0, where C is an arithmetic circuit with n logic gates. A
_∈W_
proof system (S, V, P ) is a SNARK if
1. The runtime of the prover P is _O(n),_
[�]
2. The length of a proof computed by P is O(log n),
3. The runtime of the verifier V is O(log n).
**Definition 14 (zk-SNARK). If a SNARK (S, V, P** ) is also a zero knowledge proof system, then it
is a zk-SNARK.
The Zcash cryptocurrency, which provides fully confidential transactions, was the first setting
where zk-SNARKs have been used in the field [10]. zk-SNARKs have also been deployed in the
zk-rollup procedure which increases the transaction throughput of the Ethereum blockchain [32].
For PMM we will need a zk-SNARK that does not require a trusted setup. PLONK [18], Sonic
[19], and Marlin [20] using a DARK based polynomial commitment scheme described in [21, 22].
Other options include Bulletproofs [23] and Spartan [24].
##### B.4 Implementation Details and Examples
In this section, we show how driver period information in ridehailing services, and a mobility
provider’s impact on congestion can be obtained from the protocol. Both cases involve specifying
characteristics of the query function g and trip metadata that enable the desired information to be
computed by the protocol.
**B.4.1** **Obtaining ridehailing period activity**
As discussed in Example 3, the pay rate of ridehailing drivers depends on the period they are in.
Ridehailing companies use period 2 to tell users that they are matched and a driver is en route,
thereby reducing the likelihood that the user leaves the system out of impatience. Due to this utility,
period 2 has a higher pay rate than period 1. There is thus a financial incentive for ridehailing
companies to report period 2 activity as period 1 activity so that they can have improved user
retention while keeping operations costs low. Accurate period information is thus important to
protect the wages of ridehailing drivers.
We achieve accurate period information by including digital signatures in the trip metadata.
Recall that the trip metadata includes the request time, match time, pickup time, and dropoff time
of the request. The period 2 and period 3 activity associated with a trip can be deduced from these
timestamps, as shown in Figure 6. Furthermore, Rider Witness and ARA ensure that reporting
the true demand is a dominant strategy for the ridehailing operator.
Therefore to ensure accurate period information, it is sufficient to ensure that the aforementioned timesteps are recorded correctly. For period 2 accuracy, we need to ensure that the match
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
#### Request Match Pickup Dropoff
Time Time Time Time
Phase 2 Phase 3
Wait Time
Figure 6: The timesteps within the trip metadata determine the Period 2 and Period 3 activity of
the vehicle that serves this trip.
time and pickup time are recorded properly for each trip. To do this, we will use digital signatures.
To notify a user that they have been matched, the ridehailing operator will send (mpt, σpt), where:
_mpt = You have been matched to vehicle vehID at time currtime,_
_σpt = sign(skmp, mpt)._
The user will only consider the message mpt as genuine if it is accompanied by a valid signature σpt.
Therefore, telling a user they are matched (and thus reducing the likelihood that this user cancels
their trip) requires the ridehailing company to provide an irrefutable and unforgeable declaration
of the match time in the form of (mpt, σpt). The message and signature (mpt, σpt) is then included
in the trip metadata to certify the trip’s match time. The same can be done for the pickup time,
and as a result, ensure accurate reporting of all period 2 activity.
The accuracy of period 3 activity can be ensured by ensuring that pickup time and dropoff time
are recorded correctly.
To implement driver wage inspection through the protocol, the query function g would be
_gwage(Λ) :=_ � 1[w(λ) = fwage(λ)],
_λ∈Λ_
where w(λ) is the driver wage of ride λ, and fwage is the MP’s wage formula which may depend on
the period and trajectory information contained in the trip metadata of λ. Note that gwage(Λ) = 1
if and only if all drivers were paid properly, and is 0 otherwise.
**Remark 16 (Evaluating Waiting Time Equity). Using the idea from Section B.4.1, one can also**
evaluate the equity of waiting times throughout the network. It is clear from Figure 6 that the
wait time can be determined by the request time and pickup time, both of which can be found in
the trip metadata. The trip metadata also includes the pickup location and dropoff location, so
the average wait time as a function of pickup location, dropoff location, both pickup and dropoff
locations, can all be computed from the trip metadata.
To implement a waiting time evaluation through the protocol, the Municipal Authority would
specify a fairness threshold τ . The query function g is then designed to output 1 if and only if the
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
average waiting time across locations does not vary by more than the pre-specified threshold, and
outputs 0 otherwise. Concretely, if we want to enforce wait time equity across pickup regions, we
could do this with the function
_gwait(Λ) =_ � 1 [|τi − _τj| ≤_ _τ_ ]
_i,j∈V_
where τi is the average wait time for requests in region i.
**B.4.2** **Evaluating contributions to congestion**
The trip metadata contains the trip trajectory which can be used to evaluate a ridehailing fleet’s
contribution to congestion. The trip trajectory provides the location of the service vehicle as a
function of time, which provides two important insights. First, the trip trajectories can be used
to determine how many ridehailing vehicles are on a particular road at any given time. Second,
from a trajectory one can compute the amount of time the vehicle spends on each road within
the trip path. Thus the average travel time for a road can be calculated, which can then be used
to estimate the total traffic flow on the road using traffic models. Combining these two pieces of
information, the fraction of a road’s total traffic that is ridehailing vehicles can be computed from
the trip metadata.
##### B.5 Necessity of Assumption 1 for Verifiability
In this section we show that Assumption 1 is necessary for verifiable queries on mobility data under
the natural assumption that MA does not have surveillance in the interior of MP vehicles. This
assumption on limited surveillance ability of MA is formalized in Assumption 2.
**Assumption 2. There does not exist a practical way for MA to determine whether a MP vehicle**
_is carrying a customer or not, without directly tracking all customers. In particular, MA cannot_
_determine the period information of MP vehicles._
Note that MA can obtain phase information from the drivers or from MP, but in the absence
of Assumption 1, drivers and MPs may act strategically, and may not be trustworthy.
The following result shows that under Assumption 2, if the drivers or riders are willing to
collude with the MP, then the MP can misreport properties of its mobility demand in a way that
is undetectable by the MA.
**Observation 3 (Necessity of Assumption 1 for Strategyproofness). Under Assumption 2, the**
_following events are undetectable by MA, even if MA can track all MP vehicles (i.e., knows the_
_location of each MP vehicle at any time):_
_1. If drivers collude with MP, then MP can overreport demand._
_2. If riders and drivers collude with MP, then MP can underreport demand, or misreport at-_
_tributes of the demand._
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
_Proof of Observation 3. By Assumption 2, MA cannot distinguish between MP vehicles in period 1_
and MP vehicles in period 2 or 3. Suppose the drivers are willing to collude with MP. If MP wants
to overreport demand from some origin i to some destination j, it can have some drivers drive from
_i to j without a passenger. This will lead to period 1 traffic from i to j, however the drivers will_
report themselves in period 3 to MA. This way, even if MA is able to track the MP vehicles, the
reported period information from the drivers will be consistent with the demand report from MP.
Now suppose both riders and drivers are willing to collude with MP. If the MP wants to
underreport demand from i to j, they can have some drivers who are serving passengers from i to
_j report themselves in period 1 to MA._
So in the absence of Assumption 1, the MA has no way of checking whether the messages it
receives from MP are computed from the true demand.
**Remark 17 (Tracking Users is also insufficient). Even if the MA is able to track users and thus**
determine whether a MP vehicle has a passenger, this still does not prevent overreporting of demand.
In this case, MP can hire people to hail rides from specific trips if it wants to overreport demand.
##### B.6 Roadside Audits with fewer sensors
In this section we present the Randomized Roadside Audits (RRA) mechanism. Like ARA, the
RRA detects overreporting of demand by conducting road audits. Where ARA places sensors on
every road, RRA places sensors on a small subset of randomly selected roads, enabling it to use
fewer sensors.
The sensors used in RRA are similar to the sensors used in ARA described in Section 4.2.3,
with the following differences:
1. Each sensor has its own pair of public and secret keys (pks, sks) for digital signatures. Everyone
knows pks, but sks is contained in a Hardware Security Module within the sensor so that it
is impossible to extract sks from the sensor, but it is still possible to sign messages using sks.
2. Each sensor now records its own location using GPS.
First, the MA and MP agree on a list of public keys belonging to the sensors. In particular,
they must agree on the number of sensors being deployed in the network. Let mp be the number
of sensors being deployed (recall that m is the number of roads in the network). We will focus on
the case where p (0, 1). If p > 1, then there are enough sensors to implement ARA.
_∈_
During the data-collection period, the MA will place the sensors inside vehicles which are driven
by its employees. We assume that MP cannot determine which vehicles are carrying sensors. In
practice, MA can have much more than mp employees driving around in the network, but only mp
of them will have sensors.
The data collection period is divided up into many rounds (e.g., a round could be 1 hour long).
In each round, MA will sample a random set of mp roads. Each vehicle with a sensor is assigned
to one of these roads, where they will stay (i.e., parked on the side of the street) to measure the
MP traffic that pass by them. The sensor will record a measurement u which specifies the time,
the period and the location of every MP vehicle that passes by. It will then sign the message with
its secret key via σu := sign(sks, u). In particular, a sensor assigned to road e ∈ _E in round t will_
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
be able to determine ϕt(e, Λ), which is the total number of period 2 or period 3 MP vehicles that
traverse e in round t, formally described below.
_ϕt(e, Λ) :=_ � 1[λ traverses e in round t].
_λ∈Λ_
As was the case in ARA, the sensors have a communication constraint that prevents them
from transmitting their data unless both MA and MP give permission. Therefore during the data
collection period, MP does not know where the sensors are. Once the data collection period is over,
both MA and MP give permission to collect the data from the sensors.
**Definition 15 (RRA Test). The RRA test checks whether the road usage on sampled roads is**
consistent with the demand reported by MP. Concretely, a witness w = (Λw, rw, cw) passes the
RRA test if ϕt(e, Λ) = ϕt(e, Λw) for all pairs (e, t) such that e was sampled in round t.
**Observation 4 (Efficacy of Random Roadside Audits). Under Assumption 1, if MP submits a**
_commitment σ[′]_ = MCommit(Λ[′], r) to a strict superset of the demand, i.e., Λ Λ[′], then with
_⊂_
_probability at least p, any proof submitted by MP will either be inconsistent with σ[′]_ _or will fail the_
_RRA test. Hence overreporting of demand will be detected with positive probability._
_Proof of Observation 4. We use a similar analysis to ARA. Suppose MP overreports the demand,_
i.e., submits a commitment σ[′] = MCommit(Λ[′], r) where Λ[′] is a strict superset of Λ. Then there
exists λ[′] Λ[′] Λ. Let e[′] be any road in the trip trajectory of λ[′]. and t(λ[′], e[′]) be the round in
_∈_ _\_
which trip λ[′] traverses e[′]. We then have ϕt(λ′,e′)(e[′], Λ) < ϕt(λ′,e′)(e[′], Λ[′]). If e[′] is audited in round
_t(λ[′], e[′]), then σ[′]_ will be inconsistent with the roadside audit measurements, and will fail the RRA
test. Since MA samples mp roads to audit uniformly randomly in each round, and there are a total
of m roads, the probability that e[′] is chosen in round t(λ[′], e[′]) is p. Since overreporting is detected
only probabilistically, in the event that it is detected, MA should fine MP so that MP’s expected
utility is reduced if it overreports demand.
**Remark 18 (Comparing RRA to ARA). When compared to ARA, RRA uses fewer sensors. This,**
however, is not without drawbacks, since RRA detects demand overreporting only probabilistically.
Thus in RRA the MA needs to fine the MP in the event that demand overreporting is detected.
In particular, the fine should be chosen so that the MP’s expected utility is decreased if it decides
to overreport demand. Concretely, suppose Uh, Ud are the utilities received by MP when acting
honestly and dishonestly respectively. Since dishonesty is detected with probability p, the fine F
must satisfy
_Uh > (1 −_ _p)Ud −_ _pF =⇒_ _F >_ _p[1]_ [(][U][d][ −] _[U][h][)][ −]_ _[U][d][.]_
If MA is using very few sensors or if Ud is much larger than Uh, then F needs to be very large.
A large fine, however can be difficult to implement. Recall from Section 4.2.2 that inconsistencies
between demand metadata and roadside measurements due to GPS errors can occur even if all
parties are honest. If such errors occur, then an MP would incur a large fine even if it behaves
honestly. For this reason, even an honest MP may not want to participate in the protocol. One
could use an error tolerant version of RRA, but for large F the tolerance parameter ϵ would need
to be large, enabling a dishonest MP to overreport demand while remaining within the tolerance
parameter.
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
**B.6.1** **Security Discussion**
We now make several remarks regarding the two sensor modifications we made for RRA. First,
the signatures generated by the sensors’ secret keys ensure that MA cannot fabricate or otherwise
tamper with the sensor’s data. This is important because the sensors are in the possession of the
MA and its employees. Even if the MA manages to change the data in the sensor’s storage, it
cannot produce the corresponding signatures for the altered data since it does not know the secret
key, which is protected by a Hardware Security Module.
Second, the sensor’s location data is essential to prevent MA from conducting relay attacks. A
relay attack is as follows: Suppose Alice and Bob are both MA employees. Alice has a sensor in
her car. Bob does not have a sensor in his car, but he wants to collect data as if he had a sensor
in his car. The MA can give Bob an unofficial sensor (this sensor does not have a valid public and
secret key recognized by the MP), allowing Bob to detect signals from MP vehicles. Since Bob’s
sensor does not have an official secret key, he cannot obtain a valid signature for his measurements.
To get the signatures, Bob sends the detected signal to Alice, and Alice relays the signal to her
sensor, which will sign the measurement and record it. In this manner, the MA is able to get
official measurements and signatures on Bob’s road even though he does not have an official sensor.
Fortunately, this attack is thwarted if the sensor knows its own location. If a sensor receives a
measurement whose location is very different than its location, then it will reject the message, thus
thwarting the relay attack.[2]
##### B.7 Establishing Verifiability and Differential Privacy for Appendix A
Verifiability is established by steps 1, 3 and 4 of C. Based on the analysis in Section 4.2, a witness
satisfies step 1 of C if and only if Λw = Λ, i.e., the demand is reported honestly. Since the Pedersen
commitment scheme is secure, it is computationally binding, meaning that it is computationally
intractable for MP to find Zmp[′] _[, r]mp[′]_ [with][ Z][mp] _[̸][=][ Z]mp[′]_ [and][ Commit][(][Z]mp[′] _[, r]mp[′]_ [) =][ z][mp][. So in order for]
the MP’s witness to pass step 3 of C, it must have Zmp,w = Zmp. Given steps 1, 3 have passed, step
4 ensures that the message z is indeed equal to �g(Λ, Zma ⊕ _Zmp), which establishes verifiability._
To establish differential privacy, we need to show two things: (a) MA does not know Zma ⊕ _Zmp_
(see Remark 14) and (b) Zma ⊕ _Zmp is uniformly distributed over {0, 1}[v], even if MA and MP_
are acting strategically. To this end, we consider a game between MA and MP with actions
_Zma, Zmp ∈{0, 1}[v]_ and outcome Zma ⊕ _Zmp ∈{0, 1}[v]. We will show that the strategy profile where_
both Zma, Zmp are independently sampled uniformly at random is a Nash equilibrium, meaning
that differential privacy is achieved as long as at least one party is honest.
To show that independent uniform random sampling of both Zma, Zmp is a Nash equilibrium,
we first need to show that Zma, Zmp are independent. In the protocol Zmp is sampled first, and a
Pedersen commitment zmp is sent to MA. Since Pedersen commitments are perfectly hiding, the
distribution of zmp does not depend on Zmp. So even if MA samples Zma based on the value of zmp,
the result will be independent of Zmp. Now that we have established independence of Zmp, Zma,
we make use of the following observation.
**Observation 5 (One Time Pad). Suppose Zma, Zmp are independent random variables. If Zma is**
_uniformly distributed over {0, 1}[v], then Zma ⊕_ _Zmp is uniformly distributed over {0, 1}[v], regardless_
2The measurements can be protected by authenticated encryption so that relayers (e.g., Bob) cannot modify the
messages (i.e., changing the vehicle position part of the measurement)
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
_of how Zmp is sampled. If Zmp is uniformly distributed over {0, 1}[v], then Zma ⊕_ _Zmp is uniformly_
_distributed over {0, 1}[v], regardless of how Zma is sampled._
Observation 5 says that if Zma, Zmp are independent, the distribution of Zma _Zmp does not_
_⊕_
depend on Zmp if Zma is uniformly random, and vice versa. Hence independent uniform sampling
of Zma, Zmp is a Nash equilibrium, establishing condition (b). To establish (a), if at least one party
is honest, then we can assume without loss of generality that both parties are acting according to
the Nash equilibrium. By observation 5, this means the marginal distribution of Zma _Zmp and_
_⊕_
the conditional distribution of Zma _Zmp given Zma are both uniform. In particular, MA does not_
_⊕_
learn anything about Zma _Zmp from Zma._
_⊕_
##### B.8 More Details on Congestion Pricing
When the travel cost is the same as travel time, the prices can be obtained from the following
optimization problem
�
min _xefe(xe)_
_e∈E_
� �
s.t. x = _x[od]_
_o∈V_ _d∈V_
_x[od]_ 0 _o_ _V, d_ _V_
_⪰_ _∀_ _∈_ _∈_
� _x[od](u,v)_ _[−]_ _[x]([od]v,u)_ [= Λ(][o, d][)] �1[u=o] − 1[u=d]� _∀u ∈_ _V_
(u,v)∈E
where x[od]e is the traffic flow from o to d that uses edge e, and x is the total traffic flow. Λ is the
travel demand where Λ(o, d) is the rate at which users require transport from o to d. Here the
objective measures the sum of the travel times of all requests in Λ.
Let x[∗] be a solution to (1). By first order optimality conditions[3] of x[∗], for any origin-destination
pair (i, j), and any two paths p1, p2 from i to j with non-zero flow, we have
� _∂_ � _∂_
_xefe(xe)_ =
_e∈p1_ _∂xe_ ����xe=x[∗]e _e[′]∈p2_ _∂xe′_ _[x][e][′][f][e][′][(][x][e][′][)]����xe′_ =x[∗]e[′]
� �
=⇒ _fe(x[∗]e[) +][ x][∗]e[f]e[′][(][x][∗]e[) =]_ _fe′(x[∗]e[′][) +][ x][∗]e[′][f]e[′][′][(][x][∗]e[′][)][.]_ (11)
_e∈p1_ _e[′]∈p2_
In order to realize x[∗] as a user equilibrium, the costs of p1, p2 should be the same so that no user
has an incentive to change their strategy. This can be achieved by setting the toll for each road e
as pe := x[∗]e[f]e[′][(][x][∗]e[). By doing so, from (11) we can see that the cost (travel time plus toll) for the]
two paths will be equal.
In the context of PMM, the function g associated with congestion pricing is
_gcp(Λ) :=_ �x[∗]e[f]e[′][(][x][∗]e[)]�e∈E [where][ x][∗] [solves (1)][.]
3i.e., it should be impossible to decrease the objective function by reallocating flow from p1 to p2 or vice versa.
-----
Cryptographic Data Privacy for Mobility Management Tsao, Yang, Zoepf and Pavone
##### B.9 Efficacy of Merkle Proofs
To verify the proof π(mi) = {ri} ∪ �sibling(hj,xj )�ℓjq=1 [for membership of][ m][i][, the recipient of the]
proof would compute v1, v2, ..., vℓq−1 recursively via:
_v1 := H(ri||mi)_
_vj :=_ � _H(vj−1||sibling(hj−1,xj−1))_ if xj−1 is even, for 1 ≤ _j < ℓq._
_H(sibling(hj−1,xj−1)||vj−1)_ if xj−1 is odd,
By the construction of the Merkle tree, vj = hj,xj, and so in particular the Merkle Proof is valid if
and only if vℓq is equal to the root, i.e., vℓq = hℓq,0.
Since there are q leaves in the binary tree, pi has at most log2 q vertices in it, and each hash is
_d bits, so the length of π is at most d log2 q._
By collision resistance of H, it is intractable to forge a proof if mi is not in the tree, and since
hiding nonces are used when hashing the items, the proof reveals nothing about the other items in
the tree.
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/2104.07768, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "publisher-specific-oa",
"status": "BRONZE",
"url": "https://doi.org/10.1109/tcns.2022.3141027"
}
| 2,021
|
[
"JournalArticle"
] | true
| 2021-04-15T00:00:00
|
[
{
"paperId": "a5f325bd45f33872560042a63a06ae26f1c52b27",
"title": "Blockchain-Enabled Identity Verification for Safe Ridesharing Leveraging Zero-Knowledge Proof"
},
{
"paperId": "a876aeff0d2d0bef08738ac063ef6c48f7eb8789",
"title": "Spartan: Efficient and general-purpose zkSNARKs without trusted setup"
},
{
"paperId": "7a5b83e86384eef4eb9285032d1ee6761473ba42",
"title": "Blockchain for the Internet of Vehicles Towards Intelligent Transportation Systems: A Survey"
},
{
"paperId": "ea6facc5e26fb6ca80a54930dc49158ca9dd7f60",
"title": "Verifiable and Privacy-Preserving Traffic Flow Statistics for Advanced Traffic Management Systems"
},
{
"paperId": "70ef5d5fd45518d0ab498fa23f15d80e326fbd64",
"title": "A differential privacy-based privacy-preserving data publishing algorithm for transit smart card data"
},
{
"paperId": "6a346977fe0eb181c1de9ba636235727ea4ecc80",
"title": "Marlin: Preprocessing zkSNARKs with Universal and Updatable SRS"
},
{
"paperId": "c71edf106571a39c9cf6d2ace05b63b4c66bf72a",
"title": "Transparent SNARKs from DARK Compilers"
},
{
"paperId": "6a2d8d174ecd081d2b4ce0f09480751f5445b3e9",
"title": "A Blockchain-Assisted Intelligent Transportation System Promoting Data Services with Privacy Protection"
},
{
"paperId": "83d68f99b53491d52d68e7d2a3e49fd7e02e73ed",
"title": "Optimal privacy control for transport network data sharing"
},
{
"paperId": "4e5a0594e5c35a37df1818cd9c0fdbfac968fdc4",
"title": "Privacy-Preserving Authentication Scheme for Connected Electric Vehicles Using Blockchain and Zero Knowledge Proofs"
},
{
"paperId": "fd575712e32943402ade9c75e8439ba860af48fb",
"title": "A Review of Blockchain-Based Systems in Transportation"
},
{
"paperId": "2ac167785e327e4e74b754265c6fd92542963353",
"title": "Sonic: Zero-Knowledge SNARKs from Linear-Size Universal and Updatable Structured Reference Strings"
},
{
"paperId": "153798f293f693b12a2c8dc9300ba117b18248f5",
"title": "LegoSNARK: Modular Design and Composition of Succinct Zero-Knowledge Proofs"
},
{
"paperId": "6cc50a2fb28f7c8f810a5e1c919686e2d6bf9ed2",
"title": "A Privacy-Preserving Trust Model Based on Blockchain for VANETs"
},
{
"paperId": "cd61d69c58565c9f2a9baee6c7b780553d80627f",
"title": "Accessible Privacy-Preserving Web-Based Data Analysis for Assessing and Addressing Economic Inequalities"
},
{
"paperId": "31d5acdca6c2543a191c7ce3ca27c4e357a81be2",
"title": "Bulletproofs: Short Proofs for Confidential Transactions and More"
},
{
"paperId": "2313d55f4e7739de529e966521018c9043c3fe45",
"title": "Survey on Misbehavior Detection in Cooperative Intelligent Transportation Systems"
},
{
"paperId": "80b826d4effc89e347cd243905ab4b63193907ba",
"title": "Racial and Gender Discrimination in Transportation Network Companies"
},
{
"paperId": "c2de5385c197aab309abd859b36bee1362147688",
"title": "Bitcoin and Cryptocurrency Technologies - A Comprehensive Introduction"
},
{
"paperId": "3797d924fc5f832b72a84c512e92021906793965",
"title": "Zerocash: Decentralized Anonymous Payments from Bitcoin"
},
{
"paperId": "773d73400f9b453e28949b316e65a1954d19f44c",
"title": "Privacy Protection Method for Fine-Grained Urban Traffic Modeling Using Mobile Sensors"
},
{
"paperId": "4ce8f55b89d2622ca0709f5e1e7f467057ca0ccf",
"title": "Evaluating the Usefulness of Watchdogs for Intrusion Detection in VANETs"
},
{
"paperId": "c8f98af75931260b0a58d5aed9b62ee0f6bf2f23",
"title": "Challenges in teaching a graduate course in applied cryptography"
},
{
"paperId": "8c23ea0ed7badd70a8e26dcea73f2d673cc0c74d",
"title": "What Can We Learn Privately?"
},
{
"paperId": "e4ce10063cd25447dcde75c2d9ce327446ced952",
"title": "Calibrating Noise to Sensitivity in Private Data Analysis"
},
{
"paperId": "61b66a8324742a09d259a24f98effbb1fbfec9b2",
"title": "Revealing information while preserving privacy"
},
{
"paperId": "afc0af9aa4462fbb52f8e85859ad6e445ed1e9a2",
"title": "Mitigating routing misbehavior in mobile ad hoc networks"
},
{
"paperId": "26747053b4bc759f4517f2d570b7c0227d116c2c",
"title": "Non-Interactive and Information-Theoretic Secure Verifiable Secret Sharing"
},
{
"paperId": "9ccf7b6cb32cf89752a35bd910555adac54773e0",
"title": "The Knowledge Complexity of Interactive Proof Systems"
},
{
"paperId": "df2473061df11b76cebb7400c50246d0b354390c",
"title": "How to play ANY mental game"
},
{
"paperId": "8f3282f4141f3a096f821a19aeeaf0f9f6c491f6",
"title": "Verifiable secret sharing and achieving simultaneity in the presence of faults"
},
{
"paperId": "88abb2cda4f2a57499a717966ac4fbe9a993027a",
"title": "How to share a secret"
},
{
"paperId": "9700307a2d35a594929d44edc49fcde7226ed663",
"title": "Time- and Space-Efficient Arguments from Groups of Unknown Order"
},
{
"paperId": null,
"title": "Uber must face lawsuit over ‘woefully inadequate"
},
{
"paperId": "9e0f026d02ed411889cb2b14efb390b7661924b8",
"title": "Lunar: a Toolbox for More Efficient Universal and Updatable zkSNARKs and Commit-and-Prove Extensions"
},
{
"paperId": "d928b78ea85cae93d3ca0bfabe47bf954db55e7a",
"title": "PLONK: Permutations over Lagrange-bases for Oecumenical Noninteractive arguments of Knowledge"
},
{
"paperId": null,
"title": "“Zk rollup.”"
},
{
"paperId": "0dbb8a54ca5066b82fa086bbf5db4c54b947719a",
"title": "A NEXT GENERATION SMART CONTRACT & DECENTRALIZED APPLICATION PLATFORM"
},
{
"paperId": null,
"title": "Big data and privacy: A technological perspective"
},
{
"paperId": "b7f039f3e24404dfdae60547ec1a26df671675aa",
"title": "A Fistful of Bitcoins Characterizing Payments Among Men with No Names"
},
{
"paperId": "1e34435375dcfc6381200ac6d23c48037b3fa5a9",
"title": "Available online at:"
},
{
"paperId": "4e9ec92a90c5d571d2f1d496f8df01f0a8f38596",
"title": "Bitcoin: A Peer-to-Peer Electronic Cash System"
},
{
"paperId": "f3ca3e173ea3cfa0f8a74e4c68970e8a281d95eb",
"title": "Urban Transportation Networks: Equilibrium Analysis With Mathematical Programming Methods"
},
{
"paperId": "a513c22df84d752391f050fa8e004ba2630409d4",
"title": "Coin flipping by telephone a protocol for solving impossible problems"
},
{
"paperId": null,
"title": "How uber deceives the authorities worldwide"
},
{
"paperId": null,
"title": "If riders and drivers collude with MP, then MP can underreport demand, or misreport attributes of the demand"
},
{
"paperId": null,
"title": "H is hiding - If r is a sufficiently long random string (256 bits is often sufficient), then it is computationally intractable to deduce anything about x by observing H ( r || x ). 31"
},
{
"paperId": null,
"title": "Secure - Given a public key pk"
},
{
"paperId": null,
"title": "The tuple ( S, V, P ) is a proof system for C if it satisfies the"
},
{
"paperId": null,
"title": "H is collision resistant - For sufficiently large d , it is computationally intractable to find distinct inputs x 1 , x 2 so that H ( x 1 ) = H ( x 2 )"
},
{
"paperId": "f48281cf5f701777bc1d664dcb7eb5e669e089e1",
"title": "Lunar : a Toolbox for More Efficient Universal and Updatable zkSNARKs and Commit-and-Prove Extensions"
},
{
"paperId": null,
"title": "Cryptographic Data Privacy for Mobility Management"
},
{
"paperId": null,
"title": "Data sharing: What's the worst that could happen? Government Technology"
},
{
"paperId": null,
"title": "Google removes app that calculated if uber drivers were underpaid"
}
] | 33,226
|
en
|
[
{
"category": "Engineering",
"source": "external"
},
{
"category": "Engineering",
"source": "s2-fos-model"
},
{
"category": "Environmental Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/02081d2aaf6c56a33e89743aa88faafa64171819
|
[
"Engineering"
] | 0.8888
|
Distributed Energy Storage Control for Dynamic Load Impact Mitigation
|
02081d2aaf6c56a33e89743aa88faafa64171819
|
[
{
"authorId": "30554145",
"name": "Maximilian J. Zangs"
},
{
"authorId": "2055232521",
"name": "P. B. Adams"
},
{
"authorId": "46515838",
"name": "Timur Yunusov"
},
{
"authorId": "2678149",
"name": "W. Holderbaum"
},
{
"authorId": "2124923",
"name": "B. Potter"
}
] |
{
"alternate_issns": null,
"alternate_names": null,
"alternate_urls": null,
"id": null,
"issn": null,
"name": null,
"type": null,
"url": null
}
|
The future uptake of electric vehicles (EV) in low-voltage distribution networks can cause increased voltage violations and thermal overloading of network assets, especially in networks with limited headroom at times of high or peak demand. To address this problem, this paper proposes a distributed battery energy storage solution, controlled using an additive increase multiplicative decrease (AIMD) algorithm. The improved algorithm (AIMD+) uses local bus voltage measurements and a reference voltage threshold to determine the additive increase parameter and to control the charging, as well as discharging rate of the battery. The used voltage threshold is dependent on the network topology and is calculated using power flow analysis tools, with peak demand equally allocated amongst all loads. Simulations were performed on the IEEE LV European Test feeder and a number of real U.K. suburban power distribution network models, together with European demand data and a realistic electric vehicle charging model. The performance of the standard AIMD algorithm with a fixed voltage threshold and the proposed AIMD+ algorithm with the reference voltage profile are compared. Results show that, compared to the standard AIMD case, the proposed AIMD+ algorithm further improves the network’s voltage profiles, reduces thermal overload occurrences and ensures a more equal battery utilisation.
|
# energies
_Article_
## Distributed Energy Storage Control for Dynamic Load Impact Mitigation
**Maximilian J. Zangs** **[†], Peter B. E. Adams** **[†], Timur Yunusov, William Holderbaum ***
**and Ben A. Potter**
School of Systems Engineering, University of Reading, Whiteknights Campus, Reading RG6 6AY, UK;
m.j.zangs@pgr.reading.ac.uk (M.J.Z.); p.b.e.adams@pgr.reading.ac.uk (P.B.E.A.);
t.yunusov@reading.ac.uk (T.Y.); b.a.potter@reading.ac.uk (B.A.P.)
***** Correspondence: w.holderbaum@reading.ac.uk; Tel.: +44-118-378-6086; Fax: +44-118-975-1994
† These authors contributed equally to this work.
Academic Editor: Rui Xiong
Received: 31 January 2016; Accepted: 4 August 2016; Published: 17 August 2016
**Abstract: The future uptake of electric vehicles (EV) in low-voltage distribution networks can cause**
increased voltage violations and thermal overloading of network assets, especially in networks with
limited headroom at times of high or peak demand. To address this problem, this paper proposes
a distributed battery energy storage solution, controlled using an additive increase multiplicative
decrease (AIMD) algorithm. The improved algorithm (AIMD+) uses local bus voltage measurements
and a reference voltage threshold to determine the additive increase parameter and to control the
charging, as well as discharging rate of the battery. The used voltage threshold is dependent on
the network topology and is calculated using power flow analysis tools, with peak demand equally
allocated amongst all loads. Simulations were performed on the IEEE LV European Test feeder and a
number of real U.K. suburban power distribution network models, together with European demand
data and a realistic electric vehicle charging model. The performance of the standard AIMD algorithm
with a fixed voltage threshold and the proposed AIMD+ algorithm with the reference voltage profile
are compared. Results show that, compared to the standard AIMD case, the proposed AIMD+
algorithm further improves the network’s voltage profiles, reduces thermal overload occurrences and
ensures a more equal battery utilisation.
**Keywords: battery storage; distributed control; electric vehicles; additive increase multiplicative**
decrease (AIMD); voltage control; smart grid
**1. Introduction**
The adoption of electric vehicles (EV) is seen as a potential solution to the decarbonisation of
future transport networks, offsetting emissions from conventional internal combustion engine vehicles.
The current rate of EV uptake is anticipated to increase with improved driving range, reduced cost of
purchase and greater emphasis on leading an environmentally-friendly lifestyle [1]. It is predicted
that by 2030, there will be three million plug-in hybrid electric vehicles (PHEV) and EVs sold in Great
Britain and Northern Ireland [2], and it is expected that by 2020, every tenth car in the United Kingdom
will be electrically powered [3]. It is anticipated that the majority of PHEV/EV will be charged at home,
putting additional stress on the existing local low voltage distribution network, which must then cater
for the increased demand in energy [4,5]. Uncontrolled charging of multiple PHEV/EV can raise the
daily peak power demand, which leads to: increased transmission line losses, higher voltage drops,
equipment overload, damage and failure [6–9]. Accommodating the increased demand and mitigation
of such failures is a major area of research interest, with the focus mainly placed on the coordinating
and support of home charging.
-----
_Energies 2016, 9, 647_ 2 of 20
Demand Side Management (DSM) strategies for Distributed Energy Resources (DER), aim to
alleviate the impacts of PHEV/EV home-charging and are a favoured solution. Mohsenian-Rad et al.
in [10] developed a distributed DSM algorithm that implicitly controls the operation of loads, based on
game theory and the network operator’s ability to dynamically adjust energy prices. Focusing on
financial incentive-driven DSM strategies, in [11], a Time-Of-Use (TOU) tariff and real-time load
management strategy was proposed, where disruptive charging is avoided by allocating higher prices
to times of peak demand. Financial incentives have also become a drive towards optimising the
operation of Battery Energy Storage Solutions (BESS) and Distributed Generation (DG) when including
PHEV/EV into the problem formulation [12].
Research focused on grid support has been driven by the need to deliver long-term savings and
to avoid the immediate costs and disruption of network reinforcements and upgrades. This area of
research proposes the implementation of alternative solutions to support the adoption of low carbon
technologies, such as EVs, heat pumps and the electrification of consumer products. To reduce the
resulting increased peak demand, Mohsenian-Rad et al. developed an approach of direct interaction
between grid and consumer to achieve valley-filling, by means of dynamic game theory [10]. In [13],
a Multi-Agent System (MAS) was used to manage flexible loads for the minimisation of cost in a
dynamic game. The use of aggregators has been proposed to allow the participation of a number of
small providers to participate in network support, such as grid frequency response [14–16]. Yet without
the availability of power demand forecasts, real-time control needs to be implemented.
Real-time DSM can either be implemented in a centralised or distributed control approach.
In the former, a central controller relays control signals to its aggregated DERs, whereas the latter
allows each DER to control itself. A common form of controlling DERs in this mode of operation is
set-point control [17]. Using set-point control on multiple identically-configured DERs would yield
optimal operation conditions if each DER’s control parameters (e.g., bus voltage) were shared. In a
system without sharing network information, DER control algorithms have to be improved to prevent,
for example, devices located furthest from the substation from being used more frequently than others.
This paper therefore presents an individualised BESS control algorithm that lets distributed
batteries respond to fluctuations in real-time local bus voltage readings. The proposed algorithm is
based on the robust Additive Increase Multiplicative Decrease (AIMD) type algorithm, yet implements
a set-point adjustment based on the location of the controlled BESS. It will be shown how these
home-connected batteries can mitigate the impact of additional loads (i.e., EV uptake), whilst assuring
that all BESS are cycled equally.
The key contribution of this work can be summarised as a novel distributed battery storage
algorithm for mitigating the negative impact of dynamic load uptake on the low-voltage network.
This algorithm uses an individualised set-point control to regulate bi-directional battery power
flow and, for convergence, extends the traditional AIMD algorithm. As a result, the developed
battery control method reduces voltage deviation, over-currents and the inequality of battery usage.
Reducing this usage inequality leads to a homogeneous usage of all of the distributed batteries
and, hence, prevents unequal degradation rates and unfair device utilisation.
The remainder of this paper is organised as follows: Section 2 gives some background to related
work on AIMD algorithms on which this research is based. Section 3 outlines the EV, network and
storage models used in the research. Additionally, it explains the assumptions that accommodate and
justify these models. Section 4 elaborates on the proposed AIMD control algorithm (AIMD+). Next,
Section 5 details the implementation and scenarios used for a set of test cases. For later comparison,
this section also outlines a set of comparison metrics. Section 6 presents and discusses the results,
followed by the conclusion in Section 7.
**2. Related Work**
Existing literature addresses the usage of energy storage units in low-voltage distribution
networks to assure voltage security [18–22]. An approach used by, e.g., Mokhtari et al. in [21]
-----
_Energies 2016, 9, 647_ 3 of 20
relies on bus voltage and network load measurements to prevent system overloads. Yet, these kinds of
storage control systems do require communication infrastructures to relay the network information
and control instructions. This requirement has also been addressed in the comprehensive review
on storage allocation and application methods by Hatziargyriou et al. [23]. In the presented work,
a control algorithm is proposed that removes the need for such an inter-BESS communication, since it
only uses local voltage measurements to infer the network operation. Yet, to prevent conflicting
device behaviour, the underlying coordination mechanism is of particular importance. Assuring
convergence, the AIMD algorithm is perfectly suited for such coordinated control.
Originally, AIMD algorithms were applied to congestion management in communications
networks using the TCP protocol [24], to maximise utilisation while ensuring a fair allocation of
data throughput amongst a number of competing users [25]. AIMD-type algorithms have previously
been applied to power sharing scenarios in low voltage distribution networks, where the limited
resource is the availability of power from the substation’s transformer.
For instance, such an algorithm was first proposed for EV charging by Stüdli et al. [26], requiring a
one-way communications infrastructure to broadcast a “capacity event” [27,28]. Later, their work was
further developed to include vehicle-to-grid applications with reactive power support [29]. The battery
control algorithm proposed in this paper builds upon the algorithm used by Mareels et al. [30],
where EV charging was organised by including bidirectional power flow and the use of a reference
voltage profile derived from network models. Similar to the work by Xia et al. [31], who utilised
local voltage measurements to adjust the charging rate, only voltage measurements at the batteries’
connection sites were used in this work to control the batteries’ operations.
Previous research is therefore extended by the work presented here, as previous work has only
utilised common set-point thresholds for controlling each of the DERs. The approach proposed
in this paper ensures that unavoidable voltage drops along the feeder do not skew the control
decisions, and voltage oscillations caused by demand variation are taken into control considerations.
In contrast to previous work, where substation monitoring was used to inform control units of the
transformer’s present operational capacity, the proposed AIMD+ algorithm does not require this
information and, hence, does not require such an extensive communications infrastructure.
**3. System Modelling**
In this section, the underlying assumptions to validate the research are addressed. Next, a model
to describe EV charging behaviour is explained. This is followed by a model of the BESS. Finally,
the network models used to simulate the power distribution networks are explained.
_3.1. Assumptions_
For this work, several underlying assumption were made to obtain the models:
1. The uptake of EVs is assumed to increase and, hence, to have a significant impact on the normal
operation of the low voltage distribution network. This assumption is based on a well-established
prediction that the majority of EV charging will take place at home [32].
2. The transition from internal combustion engine-powered vehicles to EVs is assumed to not impact
the users’ driving behaviour. Similar to [33], this assumption allows the utilisation of recent
vehicle mobility data [34] to generate leaving, driving and arriving probabilities, from which the
EV charging demand can be determined.
3. The transition to low carbon technologies will increase the variability of electricity demand,
and therefore, grid-supporting devices, such as BESS, are anticipated to play a more important
role [35]. Hence, alongside a high uptake of EVs, an increased adoption of distributed BESS
devices is assumed.
4. It is assumed that BESS solutions, or more specifically battery energy storage solutions, start
the simulations at 50% SOC and are not 100% efficient at storing and releasing electrical energy,
as in [36]. Additionally, its utilisation will degrade the energy storage capability and performance
-----
_Energies 2016, 9, 647_ 4 of 20
over time, as shown in [37]. Therefore, the requirements for equal and fair storage usage is of
high importance.
5. It is assumed that the load profiles provided by the IEEE Power and Energy Society (PES) are
sufficient as base load profiles for all simulations.
_3.2. Electric Vehicle Charging Behaviour_
From publicly-available car mobility data [33,34] an empirical model was developed to capture
the underlying driving behaviour. The raw data, nr(t), represents the probabilities of starting a trip
during a 15-min period of a weekday. Three continuous normal distribution functions, each defined as:
�
1
_nˆ_ _x(t) = βx_ _√_
_σx_
�
exp
2π
_−_
�t/24 − _µx�2_
2σx[2]
where t = [0, 24] (1)
were used to represent vehicles leaving in the morning, _nˆ_ _m(t), lunch time,_ _nˆ_ _l(t), and in_
the evening, ˆne(t). The aggregate probability of these three functions was optimised using a
Generalised Reduced Gradient (GRG) algorithm to fit the original data. In order to represent a
symmetric commuting behaviour, i.e., vehicles departing in the morning and returning during the
evening, an equality amongst the three probabilities was defined as follows:
� 24
0 = [ ˆnm(t) + ˆnl(t) − _nˆ_ _e(t)] dt_ (2)
0
The resulting parameters from the GRG fitting of the three distribution functions are tabulated
in Table 1. Additionally, the resulting departure probabilities, as well as the reference data nr(t) are
shown in Figure 1.
**Table 1. Parameters for normal distributions.**
**Equation ˆnx(t)** **_µx (Mean)_** **_σx (SD)_** **_βx (Weight)_**
_nˆ_ _m(t)_ 0.3049 0.0488 0.00206
_nˆ_ _l(t)_ 0.4666 0.0829 0.00314
_nˆ_ _e(t)_ 0.7042 0.0970 0.00521
**Figure 1. The probability of starting a trip at a particular time during a weekday, extrapolated into**
three normal distributions (RMS error: 9.482%).
Statistical data capturing the probability distribution of a trip being of a certain distance were
also extracted from the dataset. This was done for both the weekdays wwd(d) and weekends wwe(d).
The Weibull function was chosen to be fitted against the extracted probability distributions and is
defined as:
-----
_Energies 2016, 9, 647_ 5 of 20
_wˆ_ _x(d) :=_
_kx_ � _d_ �kx−1 exp � � _d_ �kx [�] if d 0
_γx_ _γx_ _−_ _γx_ _≥_ (3)
0 if d < 0
Performing the curve fitting using the GRG optimisation algorithm, a weekday trip
distance distribution, _wˆ_ _wd(d), and a weekend trip distribution,_ _wˆ_ _we(d), could be estimated._
The computed function parameters for these two estimated distribution functions are tabulated
in Table 2. Their resulting probability distributions are plotted for comparison against the real data,
_wwd(d) and wwe(d), in Figure 2._
**Table 2. Parameters for Weibull distributions.**
**Equation ˆwx(d)** **_γx (Scale)_** **_kx (Shape)_**
_wˆ_ _wd(t)_ 15.462 0.6182
_wˆ_ _we(t)_ 38.406 0.4653
**Figure 2. The probability of a trip being of a particular distance during a weekday, extrapolated into a**
Weibull distribution (RMS error: 3.791%).
In addition to these probabilities, an average driving speed of 56 kmh (35 mph) and an average
driving energy efficiency of 0.1305 kWh/kmh (0.21 kWh/mph) are taken from [38]. Using the predicted
driving distance and average driving speed with the driving energy efficiency, it is possible to estimate
an EV’s energy demand upon arrival. Starting to charge from this arrival time until the energy
demand has been met allows the generation of an estimated charging profile of a single EV. To do this,
a maximum charging power of the U.K.’s average household circuit rating (i.e., 7.4 kW) and an
immediate disconnection of the EV upon charge completion were assumed [39].
Generating several of those charging profiles and aggregating them produces an estimated
charging demand for an entire fleet of EVs. To provide an example, charge demand profiles for 50 EVs
were generated, aggregated and plotted in Figure 3. This plot shows the expected magnitude and
variability in energy demand that is required to charge several EVs at consumers’ homes based on the
vehicles’ daily usage.
This model’s EV charging behaviour has been implemented to reflect EV demand if applied today
without widespread smart charging infrastructure. It does therefore reflect the worst case scenario.
Future smart-charging schemes would mitigate the currently present collective EV charging spike,
yet the implementation and validation of available smart-charging schemes lies beyond the scope of
this paper. This model’s data were used to feed additional demand into the power network models,
which are outlined in the next section.
-----
_Energies 2016, 9, 647_ 6 of 20
**Figure 3. Excerpt from the aggregated 50 EVs; charging powers that were each generated from the**
empirical models.
_3.3. Battery Modelling_
For this work, a well-established model that has been used in previous publications by this
research group was used [36,40,41]. This model consists of a battery with a self-discharge loss that is
dependent on the current battery’s State Of Charge (SOC) and an energy conversion loss to represent
the energy lost when charging or discharging this battery. A complete list of all notations that are used
for this battery model is included in Table 3.
**Table 3. Table of the notation used in this section.**
**Parameter** **Description**
ine Pbat(t) Battery power at time t
_SOC(t)_ Battery state of charge at time t
_δSOC(t)_ Change in SOC during time period τ
_µ_ Self-discharge loss factor
_η_ Energy conversion efficiency
_SOCmin_ Minimum rated SOC for limited battery operation
_SOCmax_ Maximum rated SOC for limited battery operation
_C_ Battery capacity
_Pmax_ Power rating of battery
When an ideal battery charges or discharges, the change in SOC is related by the
battery power, Pbat. When sampling battery operation at a regular period, τ, then the energy transferred
into the battery can be described as Pbat(t)τ. The change in SOC for this ideal battery, δSOC, is therefore
defined as:
_δSOC(t) :=_ _[P][bat][(][t][)][τ]_ = SOC(t) − SOC(t − _τ)_ (4)
_C_
The self-discharge loss is added to this ideal battery model to represent the continual loss of
energy in the battery typical of chemical energy storage. This self-discharge loss, δSOC,sel f -discharge,
is proportional to the current SOC and is determined using the self-discharge loss factor, µ:
_δSOC,sel f_ -discharge(t) := µSOC(t) (5)
Additionally, to represent the losses in the power electronics and energy conversion process,
an energy conversion loss, δSOC,conversion, is defined. This loss is proportional to the rate at which the
battery’s SOC changes, by using the energy conversion efficiency, ˆη as follows:
_δSOC,conversion(t) := ˆηδSOC(t)_ (6)
-----
_Energies 2016, 9, 647_ 7 of 20
Here, the conversion losses in the power electronics are reflected as an asymmetric efficiency,
which depends on the direction of the flow of energy. This is done by charging the battery at
a lower power when consuming energy and discharging it more quickly when releasing energy.
Mathematically, this can be represented as:
_ηˆ =_
�
_η_ if δSOC(t) ≥ 0
(7)
_η1_ if δSOC(t) < 0
When substituting the self-discharge loss and conversion losses, respectively δSOC,sel f -discharge
and δSOC,conversion, into the SOC evolution equation, the full battery model can be summarised
as follows:
SOC(t) : = δSOC(t − _τ) −_ _δSOC,sel f_ -discharge(t − _τ) −_ _δSOC,conversion(t)_
(8)
= (1 − _µ)δSOC(t −_ _τ) −_ _ηδˆ_ _SOC(t)_
In addition, both the SOC and the Pbat are constrained due to the device’s maximum and
minimum energy storage capabilities, respectively SOCmax and SOCmin, and maximum charge and
discharge rate, Pmax. These limitations are captured in Equations (9) and (10), respectively.
SOCmin ≤ SOC(t) ≤ SOCmax (9)
_|Pbat(t)| ≤_ Pmax (10)
_3.4. Network Models_
To simulate the low-voltage energy distribution networks, the Open Distribution System
Simulator (OpenDSS) developed by the Electronic Power Research Institute (EPRI) was used. It requires
element-based network models, including line, load and transformer information, and generates
realistic power flow results.
(a) (b)
**Figure 4. Sample Open Distribution System Simulator (OpenDSS) power flow plots of the used power**
networks. Consumers are indicated as red crosses and 11/0.416-kV substations are marked with a green
square. (a) IEEE Power and Energy Society (PES) EU Low Voltage Test Feeder plot; (b) Scottish and
Southern Energy Power Distribution (SSE-PD) Common Information Model (CIM) (UK) feeder plot.
Simulations were conducted using the IEEE’s European Low Voltage Test Feeder [42]
and six detailed U.K. feeder models, that are based on real power distribution networks and provided
by Scottish and Southern Energy Power Distribution (SSE-PD). The SSE-PD circuit models were
provided as Common Information Models (CIM) during the collaboration on the New Thames Valley
-----
_Energies 2016, 9, 647_ 8 of 20
Vision Project Project (NTVV) [43]. An example of the IEEE EU LV Test feeder and a U.K. feeder
provided by SSE-PD are shown in Figure 4a,b, respectively. A summary of these model’s parameters is
given in the Table 4.
**Table 4. Network model parameters.**
**IEEE EU**
**Parameter** **SSE-PD LV Feeders**
**LV Test Feeder**
Network number 1 [1] 2 [1] 3 4 5 6 7
Number of customers 55 56 53 91 59 88 37
Median load per customer (VA) 227 227 231 241 224 237 237
Maximum load per customer (kVA) 16.8 16.8 16.8 19.5 16.8 19.5 16.8
Customer connection Single-phase Single-phase
Median substation load (kVA) 24.4 24.9 23.9 41.9 25.6 38.9 16.3
Maximum load per customer (kVA) 72.6 72.7 72.2 92.9 73.5 89.6 60.5
Three-phase Three-phase
Feeder line model
implicit-neutral explicit-neutral
1 These networks are shown in Figure 4.
Throughout this paper, all excerpt and time series results were extracted from experiments with
the IEEE EU LV Test feeder (i.e., Network No. 1). All concluding results are based on an aggregation
of all networks to include network diversity in the analysis.
The model-derived EV data and IEEE EU LV Test feeder consumer demand profiles were used in
all simulations. The resultant demand profiles represent the total daily electricity demand of households
with EVs. These profiles were sampled at τ = 1 min. The OpenDSS simulation environment was
controlled using MATLAB, achieved through OpenDSS’s Common Object Model (COM) interface and
accessible using Microsoft’s ActiveX server bridge.
**4. Storage Control**
In this section, the control of the energy storage system is explained. Firstly, the additive
increase multiplicative decrease algorithm is presented, and its decision mechanism is explained
in full. Then, the voltage referencing, used for AIMD+, is outlined.
_4.1. Additive Increase Multiplicative Decrease_
The proposed distributed battery storage control is shown in Algorithm 1. The parameter
_α denotes the size of the power’s additive increase step, and β denotes the size of the multiplicative_
decrease step. It is worth mentioning that α linearly increases and β exponentially decreases,
both charging and discharging powers, where discharging power is represented as a negative
power flow, i.e., energy released by the battery. The constants Vmax and Vthr are the maximum
historic voltage value and the set-point threshold used to regulate the total demand. In the case
when the total demand is too high, the local voltages will fall below Vthr, and the batteries reduce
their charging power and start discharging. This behaviour reduces total demand on the feeder.
At simulation start, Vmax is set to the nominal voltage of the substation transformer, i.e., 240 V, and Vthr
is set to a fraction of Vmax, which was found by solving a balanced power flow analysis. The variable
_V(t) is the battery’s local bus voltage, and Pmax denotes the maximum charging/discharging power_
of the battery. The charging and discharging power of the batteries is increased in proportion to
the available headroom on the network, which is inferred from the local voltage measurement V(t),
to avoid any sudden overloading of the substation transformer.
-----
_Energies 2016, 9, 647_ 9 of 20
**Algorithm 1 Compute battery power.**
1: R(t) = (V(t) − _Vthr)/(Vmax −_ _Vthr)_ _▷_ Defines the rate for the current voltage reading
2: if V(t) ≥ _Vthr then_ _▷_ Given the voltage levels are nominal...
3: **if SOC < SOCmax then** _▷_ ...and the battery is not fully charged...
4: _P(t) = P(t −_ _τ) + αPmaxR(t)_ _▷_ ...increase the charging power
5: **else** _▷_ If the battery has fully charged...
6: _P(t) = 0_ _▷_ ...shut off
7: **end if**
8: **if P(t) < 0 then** _▷_ If the battery has been discharging...
9: _P(t) = βP(t −_ _τ)_ _▷_ ...reduce the discharging power by β
10: **end if**
11: else _▷_ If voltage levels are not nominal...
12: **if SOC > SOCmin then** _▷_ ...and battery is charged sufficiently...
13: _P(t) = P(t −_ _τ) + αPmaxR(t)_ _▷_ ...increase discharge power
14: **else** _▷_ If the battery is not sufficiently charged...
15: _P(t) = 0_ _▷_ ...shut off
16: **end if**
17: **if P(t) > 0 then** _▷_ If the battery has been charging...
18: _P(t) = βP(t −_ _τ)_ _▷_ ...reduce the charging power by β
19: **end if**
20: end if
21: P(t) = signum(P(t)) × min{|P(t)|, Pmax} _▷_ Limit the power to battery specifications
The algorithm itself, as shown in Algorithm 1, contains two decision levels. The first determines
whether the network is over- or under-loaded by comparing the local bus voltage, V(t), to the battery’s
set-point threshold, Vthr. In the event that the network is not under high load, the battery’s SOC is
compared to its operation limit to check whether the battery can charge, i.e., SOC < SOCmax. If there is
enough charging capacity left, then the battery’s charging power is linearly increased following Line 4.
If the battery was previously discharging, the related discharging power is exponentially reduced
(Line 9) to reflect the multiplicative decrease.
The second decision level is entered when the network is under load. Here, the discharging power
is linearly increased if the battery has enough energy stored, i.e., SOC > SOCmin (Line 13). Additionally,
if the battery was previously charging, then its charging power is multiplicatively reduced (Line 18).
The direction of the charging/discharging power adjustment is determined by the first decision level,
as well as the threshold proximity ratio R(t). As the battery’s bus voltage, V(t), approaches the
threshold voltage, Vthr, this ratio tends to zero and, hence, stops the battery operation. Therefore,
oscillatory hunting is effectively mitigated. The last step of the algorithm (Line 21) assures that the
battery charge/discharge power is within its device rating.
_4.2. Reference Voltage Profile_
When using a fixed voltage threshold, the difference in the location and load of each customer
results in the over-utilisation of batteries located at the feeder end. Similar to Papaioannou et al. [44],
yet for the control of BESS instead of EV charging, a reference voltage profile is proposed, which is
produced by performing a power flow analysis of the network under maximum demand. An example
of a fixed threshold and reference voltage profile is shown in Figure 5.
In the AIMD+, consumers located at the head of the feeder are allocated a higher voltage threshold,
while those towards the end of the feeder have similar voltage thresholds to that of the fixed threshold.
This replicates the expected voltage drop along the length of the feeder, hence resulting in a more equal
utilisation of battery storage units that are located at those distances. The voltage threshold is set in
such a way as to limit the maximum voltage drop to 3% at the end of the feeder.
-----
_Energies 2016, 9, 647_ 10 of 20
**Figure 5. A plot showing the difference between the fixed voltage threshold (AIMD) and the reference**
voltage profile (AIMD+).
**5. Scenarios and Comparison Metrics**
In this section, several scenarios are explained that were used to test the performance of the battery
control algorithm. Following that is the definition of three comparison metrics. These metrics quantify
the improvements caused by the different algorithms in comparison to the worst case scenario.
_5.1. Test Cases and Scenarios_
In all simulations, the EVs plug-in on arrival and charge at their nominal charging rate until
fully charged. The BESS devices were chosen to have a capacity of 7 kWh with a maximum power
rating of 2 kW (battery specifications are based on the Tesla Powerwall [45]). Four excerpt cases were
defined with different levels of EV and storage uptakes, these are as follows:
**A** A baseline scenario, where only household demand is used.
**B** A worst case scenario, in which EV uptake is 100% and no BESS is used.
**C** An AIMD scenario, in which EV uptake is 100% and each household has a battery energy
storage device. Here, each battery was controlled using the AIMD algorithm using a fixed
voltage threshold.
**D** An AIMD+ scenario, in which EV uptake is 100%, and each household has a battery energy
storage device. Here, each battery was controlled using the AIMD+ algorithm using the optimised
reference voltage profile.
A storage uptake of 100% was adopted to represent the worst case scenario. In addition to the four
defined scenarios, a full set of simulations was performed with EV and storage uptake combinations
of 0% to 100% in steps of 10%.
_5.2. Performance Metric Definition_
To obtain comparable performance metrics, three parameters are defined. These parameters
capture the improvements in voltage violation mitigation, line overload reduction and the equality of
battery usage. All excerpt performance metrics were calculated based on simulations from the IEEE
EU LV Test feeder for reproducibility.
5.2.1. Parameter for Voltage Improvement
The first parameters are ζC[∗] [and][ ζ]D[∗] [for, respectively, Cases C and D, and calculate the magnitude]
of the voltage level improvement by comparing two voltage frequency distributions. More specifically,
they find the difference between these probability distributions and compute a weighted sum. Here,
the weighting, δ[∗](v), emphasises the voltage level improvements that deviate further from the nominal
substation voltage Vss. If the resulting weighted sum is negative, then the obtained voltage frequency
-----
_Energies 2016, 9, 647_ 11 of 20
distribution was improved in comparison to the associated worst case scenario. In contrast, a positive
number would indicate a worse outcome. The performance metric ζC[∗] [is defined as follows.]
_Vmax_
_ζC[∗]_ [:][=] ∑ _δ[∗](v) [PB(v) −_ _PC(v)]_ (11)
_v=Vmin_
Here, Vmin is the lowest recorded voltage, and Vmax is the highest recorded voltage. PB(v) is
the voltage probability distribution of the worst case scenario (Case B), and PC(v) is the voltage
probability distributions of Case C (i.e., the case with maximum EV and AIMD storage uptake).
Similarly, the parameter ζD[∗] [therefore compares Case D, i.e., the AIMD+ case, with Case B.]
The aforementioned factor, δ[∗](v), scales down the summation in Equation (11) for voltages within
the nominal operating band, where no voltage violations take place. Voltage violations on the other
hand are scaled up to increase their impact on the summation. This scaling was produced using a
linear function, with its minimum at Vss, that is defined as:
_VssVss−−Vlowv_ if v ≤ _Vss_ (12)
_Vhighv−V−ssVss_ otherwise
_δ[∗](v) :=_
_Vlow and Vhigh are defined as the lower and upper limits of the nominal operation voltage band,_
respectively. In general, the proposed voltage comparison parameter, ζ[∗], shows an improvement in
voltage distribution when it is negative, whereas a positive value implies a voltage distribution with
more voltage violations.
5.2.2. Parameter for Line Overload Reduction
Similar to measuring the voltage level improvements, all line utilisation probability distributions
between the storage and worst case scenarios were compared. This follows a similar equation to before,
but uses a different scaling factor, as described in Equation (11):
_Cmax_
_ζC[∗∗]_ [:][=] ∑ _δ[∗∗](c) [PC(c) −_ _PB(c)]_ (13)
_c=0_
Here, Cmax is the highest line utilisation. PB(c) and PC(c) present the line utilisation probability
distributions for Cases B and C, respectively, and δ[∗∗](c) is the associated scaling factor. Since the
relationship between line current and ohmic losses is quadratic, this scaling factor is defined as an
exponential function that amplifies the impact of line currents beyond the line’s nominal rating.
_δ[∗∗](c) =_
� 1−Cc _min_ �2 if c ≥ _Cmin_ (14)
0 otherwise
The capacity scale modifier, Cmin, defines from where the scaling should start and has been set
to 0.5 for this work as only line utilisation above 0.5 p.u. was considered. Therefore, a reduction in
line overloads would give a negative ζ[∗∗], whereas a positive value implies a higher line utilisation,
i.e., worse results.
5.2.3. Parameter for the Improvement of Battery Cycling
The final metric, ζ[∗∗∗], gives an indication of the inequality of battery cycling (one battery cycle is
defined as a full discharge and charge of the battery at maximum operating power, i.e., Pmax) across
-----
_Energies 2016, 9, 647_ 12 of 20
all battery units. It does this by computing the the ratio between the peak and mean battery cycling.
This Peak-to-Average Ratio (PAR) of batteries’ cycling is defined in the following equation.
max |CC|
_ζC[∗∗∗]_ := _B[−][1]_ ∑b[B]=1 ��cCb �� (15)
Here, B is the number of batteries, and cC[b] [is the total cycling of battery][ b][ during Scenario C.][ C][C][ is]
a vector of R≥[B] 0 [that contains all batteries’ cycling values, i.e.,][ c]C[b] _[∈]_ _[C][C][. Equally, the battery cycling]_
for Scenario D would be captured by ζD[∗∗∗][. In the unlikely event of an equal cycling of all batteries,]
_ζ[∗∗∗]_ will have a value of one. Yet, as batteries are operated differently, the value of ζ[∗∗∗] is likely to be
greater than one. Therefore, a resulting PAR closer to one implies a more equal and therefore fairer
utilisation of the deployed batteries.
**6. Results and Discussion**
In this section, the results are outlined that were generated from all simulations. In each of
the three subsections, the performances of the AIMD and AIMD+ algorithm are compared against
each other. To do so, the performance metrics outlined in Section 5.2 were used. In the following
subsections, results from the four test cases defined as A, B, C and D in Section 5.1 are explained
first, then the results from the full analysis over the large range of EV and battery storage uptake is
presented. In the end, these results are summarised and discussed.
_6.1. Voltage Violation Analysis_
For the comparison of voltage improvements, results compared the algorithms’ performances at
reducing bus voltage variation; particularly by increasing the lowest recorded bus voltage. Each load’s
bus voltage was recorded, from which a sample voltage profile, Figure 6, was extracted, where the
bus voltage fluctuation over time becomes apparent. It can be seen that the introduction of EVs has
significantly lowered the line-to-neutral voltage. Adding energy BESS devices did raise the voltage
levels during times of peak demand, as can be seen between 17:00 and 21:00, where the AIMD+
algorithm has elevated voltages further than the AIMD scenario. To obtain a better understanding
of the level of improvement, the voltage frequency distribution of all buses along the feeder was
generated and plotted in a histogram in Figure 7.
**Figure 6. Recorded voltage profile at the bus of the customer closest to the substation over the period**
of one day with a certain uptake in EV and battery storage devices using a moving average over a
window of 5 min. Here, Case A is blue; Case B is red; Case C is yellow; and Case D is violet.
In this histogram, the voltage probability distributions for all four cases were normalised and
plotted against each other. Here, the previously seen drop in voltages by introducing EVs is recorded
as a shift in the voltage distribution. This voltage drop is mitigated by the introduction of the
storage solutions, since the probability distribution is shifted towards higher voltage bands. For the
-----
_Energies 2016, 9, 647_ 13 of 20
IEEE EU LV Test feeder, the AIMD+-controlled batteries outperform the AIMD devices as the resulting
_ζC[∗]_ [is greater than][ ζ]D[∗] [.]
**Figure 7. Voltage probability distribution of all loads’ buses for certain uptakes of EV and battery**
storage devices. Here, Case A is blue; Case B is red; Case C is yellow; and Case D is violet; with
_ζC[∗]_ [=][ −][0.153 and][ ζ]D[∗] [=][ −][0.135.]
To gain a full understanding of the performance of the AIMD and AIMD+ algorithms, a full sweep
of EV and BESS uptake combinations was simulated on all available power distribution networks.
The resulting parameters were averaged and plotted in Figure 8.
(a) (b)
**Figure 8.** Comparison of voltage improvement indices (i.e., ζ[∗]) for (a) AIMD and (b) AIMD+.
(a) ζC[∗] [indices (AIMD); (b)][ ζ]D[∗] [indices (AIMD+).]
These figures show that the AIMD+ control algorithm reduces voltage deviation more effectively
as the uptake in storage and EVs increases. For low storage uptake, the AIMD algorithm does not
perform as strongly since more ζC[∗] [values are positive and larger than their corresponding][ ζ]D[∗] [value.]
This becomes more apparent when averaging all ζC[∗] [and][ ζ]D[∗] [values for their common storage uptake]
and across all EV uptakes. The resulting averaged metrics are plotted in Figure 9.
In this last figure, it can be seen how the sole impact of BESS uptake reflects in a continuing
improvement of voltage levels. In fact, both compared algorithms improved the bus voltage,
which coincides with the findings in the case studies. On average, this is the case for all BESS uptakes,
as ζC[∗] _[≈]_ _[ζ]D[∗]_ [. Nonetheless, it should be noted that the AIMD+ algorithm had reduced the frequency of]
severe voltage deviations in comparison to the AIMD algorithm and is more effective during scenarios
with lower BESS uptake.
-----
_Energies 2016, 9, 647_ 14 of 20
**Figure 9.** Average ζC[∗] [(AIMD) and][ ζ]D[∗] [(AIMD+) values recorded against the corresponding]
storage uptake.
_6.2. Line Overload Analysis_
Similar to the voltage improvement analysis, a frequency distribution of the line utilisation
was generated. Figure 10 shows a probability distribution of the per unit (1 p.u. represents a 100%
line usage, i.e., a line current of the same value as the line’s nominal current rating) current in all lines,
for each of the four scenarios. The corresponding ζC[∗∗] [and][ ζ]D[∗∗] [values for the AIMD and AIMD+]
storage deployment have also been included in the figure’s caption. In this figure, the observed high
probability of line over-utilisation confirms that the used test network is of insufficient capacity to
cater for the chosen EV uptake.
**Figure 10. Line utilisation probability distribution of all lines in the simulated feeder for certain uptakes**
of EV and battery storage devices. Here, Case A is blue; Case B is red; Case C is yellow; and Case D is
violet; with ζC[∗∗] [=][ −][0.360 and][ ζ]D[∗∗] [=][ −][0.518.]
Here, the AIMD+ controlled storage devices yielded a noticeable reduction in line overloads.
This improvement is apparent through the compressed width of the probability distribution and the
negative ζD[∗∗] [value. In contrast, the AIMD controlled storage devices do not fully utilise the line capacity]
as effectively, which leads to a positive value of ζC[∗∗][. To evaluate the line utilisation improvement]
across all simulations, the full range of EV and storage uptake was evaluated. The resulting plots are
shown in Figure 11.
In these figures, it can be seen how the performance metrics change as EV uptake and storage
uptake increase. For the AIMD-controlled BESS, the resulting ζC[∗∗] [values are distributed around zero,]
whereas the AIMD+ algorithm achieved mostly negative values of ζD[∗∗][. These negative values confirm]
the better usage of available line capacity. This becomes particularly noticeable for scenarios where
very low EV uptake is combined with larger BESS uptake. Here, AIMD-controlled storage devices
commence their initial charge simultaneously. As they are located closer to the substation, they do not
measure a sufficient bus voltage offset to regulate down their charging power. This behaviour causes a
number of line overloads at the very beginning of the simulated days. The AIMD+ algorithm on the
other hand, with its adjusted thresholds, is more responsive to non-optimal network operation and,
therefore, increases the charging rate gradually.
-----
_Energies 2016, 9, 647_ 15 of 20
(a) (b)
**Figure 11.** Comparison of line utilisation improvement indices for (a) AIMD and (b) AIMD+.
(a) ζC[∗∗] [indices (AIMD); (b)][ ζ]D[∗∗] [indices (AIMD+).]
This gradual adjustment is based on the fact that the bus voltages in the AIMD+ algorithm
are closer to their nominal voltages (i.e., bus voltages found by simulating the feeder with
its equally-distributed nominal load) than they are in the conventional AIMD case. A greater
voltage disparity, which is the case in AIMD, causes a prolonged additive adjustment to the
battery’s power. This prolonged adjustment is particularly apparent for batteries situated at the
bottom of the feeder, as their voltage measurements deviate the furthest from the substation voltage
level. AIMD+ on the other hand prevents this behaviour by setting the voltage threshold based on
the network’s nominal voltage drop, which is dependent on the distance between the BESS and its
feeding substation. As a result, the set-point voltage thresholds at the bottom of the feeder are lower
than those closer to the substation. Hence, the additive power adjustment is equalised along the entire
feeder. Therefore, by applying these individualised control thresholds, the sensitivity of the algorithm
is corrected, whilst successfully mitigating the severity of line overloads.
Averaging the ζC[∗∗] [and][ ζ]D[∗∗] [values over all EV uptakes gives a clearer indication of performance,]
as this is now the only variable in the performance analysis. The result is plotted in Figure 12.
Here, the hypothesis that AIMD-controlled energy storage devices do not improve line utilisation
is confirmed. In contrast, the AIMD+-controlled devices succeed at effectively reducing line overloads.
This is also demonstrated by the values of ζC[∗∗][, which remain positive yet close to zero, whereas][ ζ]D[∗∗]
decreases with increasing uptake of battery storage devices.
**Figure 12.** Average ζC[∗∗] (AIMD) and ζD[∗∗] [(AIMD+) values recorded against the corresponding]
storage uptake.
Whilst the deployment of energy storage has often been seen as a possible solution to defer
network reinforcements, the presented results show that this is not always the case. In fact,
the importance of choosing an appropriate control algorithm outweighs the availability of the
-----
_Energies 2016, 9, 647_ 16 of 20
energy storage itself. This becomes particularly apparent when energy storage devices need to
recharge their injected energy for times of peak demand. For the AIMD case, this recharging was
not controlled sufficiently, which led to higher line currents. The proposed AIMD+ algorithm
was not as susceptible to this kind of behaviour, as it has been designed to take battery location
into account. This immunity and well-controlled power flow caused little to no additional strain on
the network’s equipment, allowing the deployed storage devices to also provide voltage support.
_6.3. Battery Utilisation Analysis_
In this part of the analysis, the batteries’ fairness of usage was evaluated. The battery power
profiles were recorded; excerpts are plotted in Figure 13 and are arranged by distance from
the substation.
(a) (b)
**Figure 13. Battery power profiles of each load’s battery storage device over four days for (a) AIMD and**
(b) AIMD+. (a) Case C, 60% EV and 100% AIMD (kW); (b) Case D, 60% EV and 100% AIMD+ (kW).
In this figure, it can be seen that only half of the deployed storage devices were active in Case C
(AIMD control), whereas all devices are utilised in Case D (AIMD+ control). From the recorded
battery SOC profiles, the net cycling of each battery was computed and divided by the duration of
the simulation, giving an average daily cycling value. This value is plotted for each load in Figure 14a.
The corresponding statistical analysis is presented in Figure 14b.
(a) (b)
**Figure 14. Each load’s battery cycling compared for (a) 60% EV and 100% AIMD and AIMD+ uptake**
and (b) in a statistical context. Here, ζC[∗∗∗] = 3.89 and ζD[∗∗∗] = 2.54. (a) Battery cycling for each load;
(b) statistic.
These two plots show the under-usage of AIMD controlled batteries, as well as the variance in
battery usage under AIMD and AIMD+ control. In fact, under AIMD control, 20 out of 55 batteries
experienced a cycling of less than 10% per day, whereas the remaining devices were utilised fully.
-----
_Energies 2016, 9, 647_ 17 of 20
This discrepancy causes the ζC[∗∗∗] value to be noticeably larger than ζD[∗∗∗][. A more detailed comparison]
is given when plotting the Peak-to-Average Ratios (PAR) of the batteries’ daily cycling over the full
range of EV and storage uptake scenarios; these plots are shown in Figure 15. Section 5.2.3 gives the
detail on the PAR, ζ[∗∗∗].
(a) (b)
**Figure 15. Peak-to-Average Ratios (PAR) of the battery cycling profiles of each load’s battery storage**
device over four days for (a) AIMD and (b) AIMD+. (a) Cycling PAR for AIMD; (b) cycling PAR
for AIMD+.
The figure shows that for any EV uptake scenario, AIMD-controlled energy storage units were
cycled less equally than the AIMD+ controlled devices. Results also show that with a low EV uptake,
both the AIMD and AIMD+ algorithm performed worse; yet improved as EV uptake increased.
Averaging the PARs for all batteries’ SOC profiles over all EV uptake percentages yields a clear
performance difference between AIMD and AIMD+. These resulting PARs, i.e., the ζC[∗∗∗] and ζD[∗∗∗]
values for their corresponding storage uptake percentages, are presented in Figure 16.
**Figure 16. The performance index ζC[∗∗∗]** for AIMD storage and ζD[∗∗∗] for AIMD+ storage control against
storage uptake.
Although the AIMD controlled batteries were, on average, cycled less than the batteries controlled
by the proposed AIMD+ algorithm, looking at the average produces a distorted understanding of
the performance. In fact, as more than half of the assigned AIMD BESS devices never partook in the
network control, a lower average cycling was expected to begin with. The variation in cycling across
all batteries, or the cycling PAR, reveals the difference between usage and effective usage. A lower
ratio indicates a better usage of the deployed batteries.
**7. Conclusions**
In this paper, an algorithm is proposed for distributed battery energy storage, in order to mitigate
the negative impact of highly variable uncontrolled loads, such as the charging of EVs. The improved
-----
_Energies 2016, 9, 647_ 18 of 20
AIMD algorithm uses local bus voltage measurements and implements a reference voltage profile,
derived from power flow analysis of the distribution network, for its set-point control. Taking
the distance to the feeding substation into account allowed optimising the algorithm’s parameters
for each BESS. Simulations were performed on the IEEE EU LV Test feeder and a set of real U.K.
suburban network models. Comparisons were made of the standard AIMD algorithm with a fixed
voltage threshold against the proposed AIMD+ algorithm using a reference voltage threshold. A set
of European demand profiles and a realistic EV travel model were used to feed load data into
the simulations.
For all conducted simulations, the performance of the energy storage units was improved by
using the proposed AIMD+ algorithm instead of traditional AIMD control. The improved algorithm
resulted in a reduction of voltage variation and an increased utilisation of available line capacity, which
also reduced the frequency of line overloads. Additionally, the same algorithm equalised the cycling
and utilisation of battery energy storage, making better use of the deployed battery assets. To take
this work further, future work will also consider distributed generation, such as photovoltaic panels,
smart-charging EV uptake, as well as decentralised methods for determining voltage reference values,
so no prior network knowledge is required.
**Acknowledgments: The authors would like to thank SSE-PD for providing their network information for the**
utilised U.K. feeder models and also Miss Catriona Scrivener for proofreading this manuscript.
**Author Contributions: Maximilian J. Zangs and Peter B. E. Adams contributed equally to this piece of work and**
were supervised by William Holderbaum and Ben A. Potter. Timur Yunusov has provided technical input and
feedback throughout.
**Conflicts of Interest: The authors declare no conflict of interest.**
**References**
1. Shah, V.; Booream-Phelps, J. F.I.T.T. for Investors: Crossing the Chasm; Technical Report; Deutsche Bank Market
Research: Frankfurt am Main, Germany, 2015.
2. Department for Business Enterprise and Regulatory (DBER); Department for Transport (DfT). Investigation
_into the Scope for the Transport Sector to Switch to Electric Vehicles and Plug-in Hybrid Vehicles; Technical Report;_
Department for Business Enterprise and Regulatory (DBER), Department for Transport (DfT): London, UK,
2008.
3. Ecolane, University of Aberdeen. Pathways to High Penetration of Electric Vehicles; Technical Report; Ecolane,
University of Aberdeen: Bristol, UK, 2013.
4. Clement-Nyns, K.; Haesen, E.; Driesen, J. The impact of Charging plug-in hybrid electric vehicles on a
residential distribution grid. IEEE Trans. Power Syst. 2010, 25, 371–380.
5. Fernández, L.P.; San Román, T.G.; Cossent, R.; Domingo, C.M.; Frías, P. Assessment of the impact of plug-in
electric vehicles on distribution networks. IEEE Trans. Power Syst. 2011, 26, 206–213.
6. Hadley, S.W.; Tsvetkova, A.A. Potential Impacts of Plug-in Hybrid Electric Vehicles on Regional
Power Generation. Electr. J. 2009, 22, 56–68.
7. Putrus, G.; Suwanapingkarl, P.; Johnston, D.; Bentley, E.; Narayana, M. Impact of electric vehicles on power
distribution networks. In Proceedings of the IEEE Vehicle Power and Propulsion Conference, Dearborn,
MI, USA, 7–10 September 2009; pp. 827–831.
8. Pillai, J.R.; Bak-Jensen, B. Vehicle-to-grid systems for frequency regulation in an islanded Danish
distribution network. In Proceedings of the IEEE Vehicle Power and Propulsion Conference (VPPC),
Lille, France, 1–3 September 2010.
9. Zhou, K.; Cai, L. Randomized PHEV Charging Under Distribution Grid Constraints. IEEE Trans. Smart Grid
**2014, 5, 879–887.**
10. Mohsenian-Rad, A.H.; Wong, V.W.S.; Jatskevich, J.; Schober, R.; Leon-Garcia, A. Autonomous demand-side
management based on game-theoretic energy consumption scheduling for the future smart grid. IEEE Trans.
_Smart Grid 2010, 1, 320–331._
-----
_Energies 2016, 9, 647_ 19 of 20
11. Deilami, S.; Masoum, A.S. Real-time coordination of plug-in electric vehicle charging in smart grids to
minimize power losses and improve voltage profile. IEEE Trans. Smart Grid 2011, 2, 456–467.
12. Masoum, A.S.; Deilami, S.; Member, S.; Masoum, M.A.S. Fuzzy Approach for Online Coordination of Plug-In
Electric Vehicle Charging in Smart Grid. IEEE Trans. Sustain. Energy 2015, 6, 1112–1121.
13. Karfopoulos, E.L.; Hatziargyriou, N.D. A Multi-Agent System for Controlled Charging of a Large Population
of Electric Vehicles. IEEE Trans. Power Syst. 2013, 28, 1196–1204.
14. Wu, C.; Mohsenian-Rad, H.; Huang, J. Vehicle-to-aggregator interaction game. IEEE Trans. Smart Grid 2012,
_3, 434–442._
15. Samadi, P.; Mohsenian-Rad, H.; Schober, R.; Wong, V.W.S. Advanced Demand Side Management for the
Future Smart Grid Using Mechanism Design. IEEE Trans. Smart Grid 2012, 3, 1170–1180.
16. Xu, N.Z.; Chung, C.Y. Challenges in Future Competition of Electric Vehicle Charging Management
and Solutions. IEEE Trans. Smart Grid 2015, 6, 1323–1331.
17. Leadbetter, J.; Swan, L. Battery storage system for residential electricity peak demand shaving. Energy Build.
**2012, 55, 685–692.**
18. Sugihara, H.; Yokoyama, K.; Saeki, O.; Tsuji, K.; Funaki, T. Economic and efficient voltage management
using customer-owned energy storage systems in a distribution network with high penetration of
photovoltaic systems. IEEE Trans. Power Syst. 2013, 28, 102–111.
19. Toledo, O.M.; Oliveira, D.; Diniz, A.; Martins, J.H.; Vale, M.H.M. Methodology for Evaluation of Grid-Tie
Connection of Distributed Energy Resources-Case Study with Photovoltaic and Energy Storage. IEEE Trans.
_Power Syst. 2013, 28, 1132–1139._
20. Marra, F.; Yang, G.Y.; Fawzy, Y.T.; Træholt, C.; Larsen, E.; Garcia-Valle, R.; Jensen, M.M. Improvement of
local voltage in feeders with photovoltaic using electric vehicles. IEEE Trans. Power Syst. 2013, 28, 3515–3516.
21. Mokhtari, G.; Nourbakhsh, G.; Ghosh, A. Smart coordination of energy storage units (ESUs) for voltage and
loading management in distribution networks. IEEE Trans. Power Syst. 2013, 28, 4812–4820.
22. Atia, R.; Yamada, N. Sizing and Analysis of Renewable Energy and Battery Systems in Residential Microgrids.
_IEEE Trans. Smart Grid 2016, 7, 1204–1213._
23. Hatziargyriou, N.D.; Škrlec, D.; Capuder, T.; Georgilakis, P.S.; Zidar, M. Review of energy storage allocation
in power distribution networks: Applications, methods and future research. IET Gener. Transmi. Distrib.
**2015, 10, 1–8.**
24. Chiu, D.M.; Rain, R. Analysis of the increase and decrease algorithms for congestion avoidance in
computer networks. Comput. Netw. ISDN Syst. 1989, 17, 1–14.
25. Wirth, F.; Stuedli, S.; Yu, J.Y.; Corless, M.; Shorten, R. IBM Research Report: Nonhomogeneous Place-Dependent
_Markov Chains, Unsynchronised AIMD, and Network Utility Maximization; Technical Report; IBM: New York,_
NY, USA, 2014.
26. Stüdli, S.; Crisostomi, E.; Middleton, R.; Shorten, R. A flexible distributed framework for realising electric
and plug-in hybrid vehicle charging policies. Int. J. Control 2012, 85, 1130–1145.
27. Studli, S.; Griggs, W.; Crisostomi, E.; Shorten, R. On Optimality Criteria for Reverse Charging of
Electric Vehicles. IEEE Trans. Intell. Transp. Syst. 2014, 15, 451–456.
28. Stüdli, S.; Crisostomi, E.; Middleton, R.; Shorten, R. Optimal real-time distributed V2G and G2V management
of electric vehicles. Int. J. Control 2014, 87, 1153–1162.
29. Stüdli, S.; Crisostomi, E.; Middleton, R.; Braslavsky, J.; Shorten, R. Distributed Load Management Using
Additive Increase Multiplicative Decrease Based Techniques. In Plug in Electric Vehicles in Smart Grids;
Springer: Singapore, Singapore, 2015; pp. 173–202.
30. Mareels, I.; Alpcan, T.; Brazil, M.; de Hoog, J.; Thomas, D.A. A distributed electric vehicle charging
management algorithm using only local measurements. In Proceedings of the IEEE PES Innovative Smart
Grid Technologies Conference (ISGT), Washington, DC, USA, 19–22 February 2014; pp. 1–5.
31. Xia, L.; Hoog, J.D.; Alpcan, T.; Brazil, M. _Electric Vehicle Charging: A Noncooperative Game Using_
_Local Measurements; The International Federation of Automatic Control: Cape Town, South Africa, 2014;_
pp. 5426–5431.
32. Munkhammar, J.; Bishop, J.D.; Sarralde, J.J.; Tian, W.; Choudhary, R. Household electricity use, electric vehicle
home-charging and distributed photovoltaic power production in the city of Westminster. Energy Build.
**2015, 86, 439–448.**
-----
_Energies 2016, 9, 647_ 20 of 20
33. Dallinger, D.; Wietschel, M. Grid integration of intermittent renewable energy sources using price-responsive
plug-in electric vehicles. Renew. Sustain. Energy Rev. 2012, 16, 3370–3382.
34. Institut für angewandte Sozialwissenschaft GmbH; Deutsches Zentrum für Luft-und Raumfahrt e.V. Mobilität
_in Deutschland 2008; Technical Report; Mobilität in Deutschland: Bonn und Berlin, Germany, 2008.[¨]_
35. National Grid. Future Energy Scenarios 2015; Technical Report; National Grid: Warwick, UK, July 2015.
36. Rowe, M.; Member, S.; Yunusov, T.; Member, S.; Haben, S.; Singleton, C.; Holderbaum, W.; Potter, B. A Peak
Reduction Scheduling Algorithm for Storage Devices on the Low Voltage Network. IEEE Trans. Smart Grid
**2014, 5, 2115–2124.**
37. Laresgoiti, I.; Käbitz, S.; Ecker, M.; Sauer, D.U. Modeling mechanical degradation in lithium ion batteries
during cycling: Solid electrolyte interphase fracture. J. Power Sour. 2015, 300, 112–122.
38. Government Digital Service. Vehicle Free-Flow Speeds (SPE01); Government Digital Service: London, UK,
2013.
39. Office for Low Emission Vehicles. _Electric Vehicle Homecharging Scheme—Guidance for Manufacturers_
_and Installers; Technical Report; Office for Low Emission Vehicles: London, UK, 2016._
40. Rowe, M.; Holderbaum, W.; Potter, B. Control methodologies: Peak reduction algorithms for DNO owned
storage devices on the Low Voltage network. In Proceedings of the 4th IEEE/PES Innovative Smart Grid
Technologies Europe (ISGT), Lyngby, Denmark, 6–9 October 2013; pp. 1–5.
41. Rowe, M.; Yunusov, T.; Haben, S.; Holderbaum, W.; Potter, B. The real-time optimisation of DNO owned
storage devices on the LV network for peak reduction. Energies 2014, 7, 3537–3560.
42. Society, I.P. Energy, European Low Voltage Test Feeder. Available online:
http://ewh.ieee.org/soc/pes/dsacom/testfeeders/ (accessed on 31 January 2016).
43. Thames Valley Vision—Project Library—Published Documents. Available online:
http://thamesvalleyvision.co.uk/project-library/published-documents/ (accessed on 31 January
2016).
44. Papaioannou, I.T.; Purvins, A.; Demoulias, C.S. Reactive power consumption in photovoltaic inverters:
A novel configuration for voltage regulation in low-voltage radial feeders with no need for central control.
_Prog. Photovolt. Res. Appl. 2015, 23, 611–619._
45. Tesla Motors Inc. Tesla Powerwall; Tesla Motors Inc.: Fremont, CA, USA, 2015.
_⃝c_ 2016 by the authors; licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
[(CC-BY) license (http://creativecommons.org/licenses/by/4.0/).](http://creativecommons.org/licenses/by/4.0/)
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3390/EN9080647?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/EN9080647, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "GOLD",
"url": "https://www.mdpi.com/1996-1073/9/8/647/pdf?version=1471431770"
}
| 2,016
|
[] | true
| 2016-08-17T00:00:00
|
[
{
"paperId": "f92dbeeeff4c7c01616c4ebbbeea94cb79e3a009",
"title": "Service"
},
{
"paperId": "bea991075c2d63ff016a05298daa2aee58ef32bb",
"title": "Review of energy storage allocation in power distribution networks: applications, methods and future research"
},
{
"paperId": "d2c2d2430a7b21ab6448983dd8948bf28d2c1165",
"title": "Sizing and Analysis of Renewable Energy and Battery Systems in Residential Microgrids"
},
{
"paperId": "4bfb725c697d5ad666d84f5fea7adb73819db862",
"title": "Modeling mechanical degradation in lithium ion batteries during cycling: Solid electrolyte interphase fracture"
},
{
"paperId": "18e110a5314c4d8a652bbf980ecda63b4ec7c8db",
"title": "Fuzzy Approach for Online Coordination of Plug-In Electric Vehicle Charging in Smart Grid"
},
{
"paperId": "53cf2ba8428417db7296ab8edc4bae922e5a6246",
"title": "Challenges in Future Competition of Electric Vehicle Charging Management and Solutions"
},
{
"paperId": "74c6066ca2e388b88f85c3f3ee7a4a6016968115",
"title": "Reactive power consumption in photovoltaic inverters: a novel configuration for voltage regulation in low‐voltage radial feeders with no need for central control"
},
{
"paperId": "e280a50a565feb8bffe6f121cd689eae516caccd",
"title": "A Peak Reduction Scheduling Algorithm for Storage Devices on the Low Voltage Network"
},
{
"paperId": "c7dbc255e51d2693d581850184f178f5c41985bb",
"title": "The Real-Time Optimisation of DNO Owned Storage Devices on the LV Network for Peak Reduction"
},
{
"paperId": "c79e09ec37d3ced5fe30e5d41ff2e826b5da1457",
"title": "A distributed electric vehicle charging management algorithm using only local measurements"
},
{
"paperId": "20c178d152b3fab13710d3dd7ae1d8d4c0be2f28",
"title": "Nonhomogeneous Place-Dependent Markov Chains, Unsynchronised AIMD, and Network Utility Maximization"
},
{
"paperId": "22016e5e4871e52a377b927fe4e376775bd96821",
"title": "Randomized PHEV Charging Under Distribution Grid Constraints"
},
{
"paperId": "e31aca41cda405f371bfb1f71725ce3638fa9d84",
"title": "Optimal real-time distributed V2G and G2V management of electric vehicles"
},
{
"paperId": "75edf98c614dbe54b330502000b986846a08ca1e",
"title": "On Optimality Criteria for Reverse Charging of Electric Vehicles"
},
{
"paperId": "9cd40acac793e054efaed364af3c3a36e315472e",
"title": "Control methodologies: Peak reduction algorithms for DNO owned storage devices on the Low Voltage network"
},
{
"paperId": "304ea789633461c2ed89c078743b431ad9ceefde",
"title": "Smart Coordination of Energy Storage Units (ESUs) for Voltage and Loading Management in Distribution Networks"
},
{
"paperId": "00e3337142859baf13ecdcf8dd3189d832f316b3",
"title": "A Multi-Agent System for Controlled Charging of a Large Population of Electric Vehicles"
},
{
"paperId": "00bbd9ef8ad23302ebea5f3893f436a24a62c16a",
"title": "Methodology for evaluation of grid-tie connection of distributed energy resources - Case study with photovoltaic and energy storage"
},
{
"paperId": "b5d6abdef987fc9cf776fc57a6fdb0237a3f0057",
"title": "Improvement of Local Voltage in Feeders With Photovoltaic Using Electric Vehicles"
},
{
"paperId": "6db10cf37bcfe71c37dcaf166114ae2f0a082c34",
"title": "Economic and Efficient Voltage Management Using Customer-Owned Energy Storage Systems in a Distribution Network With High Penetration of Photovoltaic Systems"
},
{
"paperId": "9c3d2cce5a973fa5b4023c5baa492223a6e636c7",
"title": "Battery storage system for residential electricity peak demand shaving"
},
{
"paperId": "0c386db6f124431e753b5fbba79cab32a07ed1cc",
"title": "Advanced Demand Side Management for the Future Smart Grid Using Mechanism Design"
},
{
"paperId": "0fd0ef5c807d6cca4c4012108dcb229f8019c758",
"title": "A flexible distributed framework for realising electric and plug-in hybrid vehicle charging policies"
},
{
"paperId": "7d546948366df251e4007e446570ae7b40cb6afe",
"title": "Grid integration of intermittent renewable energy sources using price-responsive plug-in electric vehicles"
},
{
"paperId": "2d69bb478b740fa70e4f2e0da55cdc8c67f60489",
"title": "Vehicle-to-Aggregator Interaction Game"
},
{
"paperId": "8a59e11d3e42d200cb4892194f2b018b1952e806",
"title": "Real-Time Coordination of Plug-In Electric Vehicle Charging in Smart Grids to Minimize Power Losses and Improve Voltage Profile"
},
{
"paperId": "eeff9128a6f0fec8d2af1afa8241253b59458297",
"title": "Assessment of the Impact of Plug-in Electric Vehicles on Distribution Networks"
},
{
"paperId": "021d2b357c2d4fb0000f93970732bdef961ac6d4",
"title": "Autonomous Demand-Side Management Based on Game-Theoretic Energy Consumption Scheduling for the Future Smart Grid"
},
{
"paperId": "783fad50d6a366fb48f0343d233d14b5ed5ca501",
"title": "Vehicle-to-grid systems for frequency regulation in an Islanded Danish distribution network"
},
{
"paperId": "88d9ed51d8bf9c230052dcadad0b675588e46981",
"title": "The Impact of Charging Plug-In Hybrid Electric Vehicles on a Residential Distribution Grid"
},
{
"paperId": "3fe0ddbe3dbcf8cd0bc00923755f56385e65e8fd",
"title": "Potential Impacts of Plug-in Hybrid Electric Vehicles on Regional Power Generation"
},
{
"paperId": "4239b92d7d159c483dbd1f933be2d4795e44b4f5",
"title": "Impact of electric vehicles on power distribution networks"
},
{
"paperId": "805d0da469da6ba7571ee75732ab66202aaea9e0",
"title": "Analysis of the Increase and Decrease Algorithms for Congestion Avoidance in Computer Networks"
},
{
"paperId": "c37157c4b031f8e5e9669135e226127bf10c4ed3",
"title": "Distributed Load Management Using Additive Increase Multiplicative Decrease Based Techniques"
},
{
"paperId": "3f996d1353aaad418807ac2f5401f97da7d2058a",
"title": "Household electricity use, electric vehicle home-charging and distributed photovoltaic power production in the city of Westminster"
},
{
"paperId": "cc953dffedaf5347dfd1cfe0b458a44dd046594e",
"title": "Electric Vehicle Charging: A Noncooperative Game Using Local Measurements"
},
{
"paperId": null,
"title": "Autonomous 399 demand-side management based on game-theoretic energy consumption scheduling for the future 400 smart grid"
},
{
"paperId": null,
"title": "F.I.T.T. for Investors: Crossing the Chasm"
},
{
"paperId": null,
"title": "Energy , European Low Voltage Test Feeder Thames Valley Vision — Project Library — Published Documents"
},
{
"paperId": null,
"title": "Energy, European Low Voltage Test Feeder"
},
{
"paperId": null,
"title": "This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC-BY) license"
},
{
"paperId": null,
"title": "Pathways to High Penetration of Electric Vehicles"
},
{
"paperId": null,
"title": "European Low Voltage Test Feeder"
},
{
"paperId": null,
"title": "Electric Vehicle Charging: A Noncooperative Game Using Local Measurements; The International Federation of Automatic Control: Cape"
},
{
"paperId": null,
"title": "Thames Valley Vision-Project Library-Published Documents"
},
{
"paperId": null,
"title": "Electric Vehicle Homecharging Scheme-Guidance for Manufacturers and Installers"
},
{
"paperId": null,
"title": "Investigation into the Scope for the Transport Sector to Switch to Electric Vehicles and Plug-in Hybrid Vehicles"
},
{
"paperId": null,
"title": "for Investors: Crossing the Chasm"
},
{
"paperId": null,
"title": "Government Digital Service Vehicle Free-Flow Speeds (SPE01); Government Digital Service"
},
{
"paperId": null,
"title": "This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC-BY) license (http://creativecommons"
},
{
"paperId": null,
"title": "Household electricity 430 use, electric vehicle home-charging and distributed photovoltaic power production in the city of 431 Westminster"
},
{
"paperId": null,
"title": "Office for Low Emission Vehicles Electric Vehicle Homecharging Scheme—Guidance for Manufacturers and Installers"
},
{
"paperId": null,
"title": "Impact of electric 392 vehicles on power distribution networks"
}
] | 15,377
|
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Mathematics",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/02086f491c51a3222da753b7d2b7100fffa01e0b
|
[
"Computer Science"
] | 0.813728
|
Compact McEliece Keys from Goppa Codes
|
02086f491c51a3222da753b7d2b7100fffa01e0b
|
IACR Cryptology ePrint Archive
|
[
{
"authorId": "2237912",
"name": "Rafael Misoczki"
},
{
"authorId": "1698800",
"name": "Paulo L. Barreto"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"IACR Cryptol eprint Arch"
],
"alternate_urls": null,
"id": "166fd2b5-a928-4a98-a449-3b90935cc101",
"issn": null,
"name": "IACR Cryptology ePrint Archive",
"type": "journal",
"url": "http://eprint.iacr.org/"
}
| null |
# Compact McEliece Keys from Goppa Codes
Rafael Misoczki and Paulo S.L.M. Barreto[⋆]
Departamento de Engenharia de Computação e Sistemas Digitais (PCS),
Escola Politécnica, Universidade de São Paulo, Brazil
{rmisoczki,pbarreto}@larc.usp.br
**Abstract. The classical McEliece cryptosystem is built upon the class**
of Goppa codes, which remains secure to this date in contrast to many
other families of codes but leads to very large public keys. Previous proposals to obtain short McEliece keys have primarily centered around
replacing that class by other families of codes, most of which were shown
to contain weaknesses, and at the cost of reducing in half the capability
of error correction. In this paper we describe a simple way to reduce
significantly the key size in McEliece and related cryptosystems using a
subclass of Goppa codes, while also improving the efficiency of cryptographic operations to _O[˜](n) time, and keeping the capability of correcting_
the full designed number of errors in the binary case.
## 1 Introduction
Quantum computers can potentially break most if not all conventional cryptosystems actually deployed in practice, namely, all systems based on the integer
factorization problem (like RSA) or the discrete logarithm problem (like traditional or elliptic curve Diffie-Hellman and DSA, and also all of pairing-based
cryptography).
Certain classical cryptosystems, inspired on computational problems of a na
ture entirely different from the above and potentially much harder to solve,
remain largely unaffected by the threat of quantum computing, and have thus
been called quantum-resistant or, more suggestively, ‘post-quantum’ cryptosystems. These include lattice-based cryptosystems and syndrome-based cryptosystems like McEliece [16] and Niederreiter [19]. Such systems usually have even
a speed advantage over conventional schemes; for instance, both McEliece and
Niederreiter encryption over a code of length n has time complexity O(n[2]), while
Diffie-Hellman/DSA and (private exponent) RSA with n-bit keys have time complexity O(n[3]). On the other hand, they are plagued by very large keys compared
to their conventional counterparts.
It is therefore of utmost importance to seek ways to reduce the key sizes for
post-quantum cryptosystems while keeping their security level. The first steps
_⋆_ Supported by the Brazilian National Council for Scientific and Technological De
velopment (CNPq) under research productivity grant 312005/2006-7 and universal
grant 485317/2007-9, and by the Science Foundation Ireland (SFI) as E. T. S. Walton
Award fellow under grant 07/W.1/I1824.
M.J. Jacobson Jr., V. Rijmen, and R. Safavi-Naini (Eds.): SAC 2009, LNCS 5867, pp. 376–392, 2009.
_⃝c_ Springer-Verlag Berlin Heidelberg 2009
-----
Compact McEliece Keys from Goppa Codes 377
toward this goal were taken by Monico et al. using low density parity-check
codes [18], by Gaborit using quasi-cyclic codes [8], and by Baldi and Chiaraluce
using a combination of both [1].
However, these proposals were all shown to contain weaknesses [22]. In those
proposals the trapdoor is protected essentially by no other means than a private
permutation of the underlying code. The attack strategy consists of obtaining
a solvable system of linear equations that the components of the permutation
matrix must satisfy, and was successfully mounted due to the very constrained
nature of the secret permutation (since it has to preserve the quasi-cyclic structure of the result) and the fact that the secret code is a subcode of a public
code.
A dedicated fix to the problems in [1] is proposed in [2]. More recently, Berger
et al. [3] showed how to circumvent the drawbacks of Gaborit’s original scheme
and remove the weaknesses pointed out in [22] by means of two techniques:
1. Extracting block-shortened public codes from very large private codes, ex
ploiting Wieschebrink’s theorem on the NP-completeness of distinguishing
punctured codes [29];
2. Working with subfield subcodes over an intermediate subfield between the
base field and the extension field of the original code.
These two techniques were successfully applied to quasi-cyclic codes, yet we will
see that their applicability is not restricted to that class.
**Our contribution: In this paper we propose the class of quasi-dyadic Goppa**
codes, which admit a very compact parity-check or a generator matrix representation, for efficiently instantiating syndrome-based cryptosystems. We stress
that we are not proposing any new cryptosystem, but rather a technique to
obtain efficient parameters and algorithms for such systems, current or future.
In contrast to many other proposed families of codes [10,11,22,27], Goppa codes
have withstood cryptanalysis quite well, and despite considerable progress in the
area [14,26] (see also [6] for a survey) they remain essentially unscathed since they
were suggested with the very first syndrome-based cryptosystem known, namely,
the original McEliece scheme. Our method produces McEliece-type keys that are
up to a factor t = O[˜](n) smaller than keys produced from generic t-error correcting Goppa codes of length n in characteristic 2. In the binary case it also retains
the ability of correcting the full designed number of errors rather than just half as
many, a feature that is missing in all previous attempts at constructing compact
codes for cryptographic purposes, including [3]. Moreover, the complexity of all
typical cryptographic operations become _O[˜](n); specifically, under the common_
cryptographic setting t = O(n/ lg n), code generation, encryption and decryption
all have asymptotic complexity O(n lg n).
The remainder of this paper is organized as follows. Section 2 introduces some
basic concepts of coding theory. In section 3 we describe our proposal of using
binary Goppa codes in quasi-dyadic form, and how to build them. We consider
hardness issues in Section 4, and efficiency issues, including guidelines on how
to choose parameters, in Section 5. We conclude in Section 6.
-----
378 R. Misoczki and P.S.L.M. Barreto
## 2 Preliminaries
In what follows all vector and matrix indices are numbered from zero onwards.
**Definition 1. Given a ring R and a vector h = (h0, . . ., hn−1) ∈R[n], the dyadic**
_matrix ∆(h) ∈R[n][×][n]_ _is the symmetric matrix with components ∆ij = hi⊕j where_
_⊕_ _stands for bitwise exclusive-or on the binary representations of the indices. The_
_sequence h is called its signature. The set of dyadic n × n matrices over R is_
_denoted ∆(R[n]). Given t > 0, ∆(t, h) denotes ∆(h) truncated to its first t rows._
One can recursively characterize a dyadic matrix when n is a power of 2: any
1 × 1 matrix is dyadic, and for k > 0 any 2[k] _× 2[k]_ dyadic matrix M has the form
�
_M =_
� _A B_
_B A_
where A and B are 2[k][−][1] _× 2[k][−][1]_ dyadic matrices. It is not hard to see that the
signature of a dyadic matrix coincides with its first row. Dyadic matrices form
a commutative subring of R[n][×][n] as long as R is commutative [12].
**Definition 2. A dyadic permutation is a dyadic matrix Π** _[i]_ _∈_ _∆({0, 1}[n]) whose_
_signature is the i-th row of the identity matrix._
A dyadic permutation is clearly an involution, i.e. (Π _[i])[2]_ = I. The i-th row (or
equivalently the i-th column) of the dyadic matrix defined by a signature h can
be written ∆(h)i = hΠ _[i]._
**Definition 3. A quasi-dyadic matrix is a (possibly non-dyadic) block matrix**
_whose component blocks are dyadic submatrices._
Quasi-dyadic matrices are at the core of our proposal. We will be mainly concerned with the case R = Fq, the finite field with q (a prime power) elements.
**Definition 4. Given two disjoint sequences z = (z0, . . ., zt−1) ∈** F[t]q _[and][ L][ =]_
(L0, . . ., Ln−1) ∈ F[n]q _[of distinct elements, the][ Cauchy matrix][ C][(][z, L][)][ is the][ t]_ _[×]_ _[n]_
_matrix with elements Cij = 1/(zi_ _Lj), i.e._
_−_
1 1
_. . ._
_z0 −_ _L0_ _z0 −_ _Ln−1_
_..._ _..._ _..._
1 1
_. . ._
_zt−1 −_ _L0_ _zt−1 −_ _Ln−1_
⎤
⎥⎥⎥⎥⎦
_C(z, L) =_
⎡
⎢⎢⎢⎢⎣
_._
Cauchy matrices have the property that all of their submatrices are nonsingular [25]. Notice that, in general, Cauchy matrices are not dyadic and vice-versa,
although the intersection of these two classes is non-empty in characteristic 2.
**Definition 5. Given t > 0 and a sequence L = (L0, . . ., Ln−1) ∈** F[n]q _[, the][ Van-]_
dermonde matrix vdm(t, L) is the t × n matrix with elements Vij = L[i]j[.]
-----
Compact McEliece Keys from Goppa Codes 379
**Definition 6. Given a sequence L = (L0, . . ., Ln−1) ∈** F[n]q _[of distinct elements]_
_and a sequence D = (D0, . . ., Dn−1) ∈_ F[n]q _[of nonzero elements, the][ General-]_
ized Reed-Solomon code GRSr(L, D) is the [n, k, r] linear error-correcting code
_defined by the parity-check matrix_
_H = vdm(r_ 1, L) diag(D).
_−_ _·_
_An alternant code is a subfield subcode of a Generalized Reed-Solomon code._
Let p be a prime power, let q = p[d] for some d, and let Fq = Fp[x]/b(x) for
some irreducible polynomial b(x) ∈ Fp[x] of degree d. Given a code specified
by a parity-check matrix H ∈ F[t]q[×][n], the trace construction derives from it an
Fp-subfield subcode by writing the Fp coefficients of each Fq component of H
onto d successive rows of a parity-check matrix Td(H) ∈ F[dt]p _[×][n]_ for the subcode.
The related co-trace parity-check matrix Td[′][(][H][)][ ∈] [F][dt]p _[×][n], equivalent to Td(H)_
by a left permutation, is obtained from H by writing the Fp coefficients of terms
of equal degree from all components on a column of H onto successive rows of
_Td[′][(][H][)][.]_
Thus, given elements ui(x) = ui,0 + · · · + ui,d−1x[d][−][1] _∈_ Fq = Fp[x]/b(x),
the trace construction maps a column (u0, . . ., ut−1)[T] from H to the column
(u0,0, . . ., u0,d−1; . . . ; ut−1,0, . . ., ut−1,d−1)[T] on the trace matrix Td(H), and to
the column (u0,0, . . ., ut−1,0; . . . ; u0,d−1, . . ., ut−1,d−1)[T] on the co-trace matrix
_Td[′][(][H][)][.]_
Finally, one of the most important families of linear error-correcting codes for
cryptographic purposes is that of Goppa codes:
**Definition 7. Given a prime power p, q = p[d]** _for some d, a sequence L =_
(L0, . . ., Ln−1) ∈ F[n]q _[of distinct elements and a polynomial][ g][(][x][)][ ∈]_ [F][q][[][x][]][ of]
_degree t such that g(Li) ̸= 0 for 0 ⩽_ _i < n, the Goppa code Γ_ (L, g) over
Fp is the alternant code over Fp corresponding to GRSt(L, D) where D =
(g(L0)[−][1], . . ., g(Ln−1)[−][1]), and its minimum distance is at least 2t + 1.
An irreducible Goppa code in characteristic 2 can correct up to t errors using Patterson’s algorithm [23], or slightly more using Bernstein’s list decoding
method [5], and t errors can still be corrected by suitable decoding algorithms if
the generator g(x) is not irreducible[1]. In all other cases no algorithm is known
that can correct more than t/2 errors (or just a few more).
## 3 Goppa Codes in Cauchy and Dyadic Form
A property of Goppa codes that is central to our proposal is that they admit a
parity-check matrix in Cauchy form:
1 For instance, one can equivalently view the binary Goppa code as the alternant code
defined by the generator polynomial g[2](x), in which case any alternant decoder will
decode t errors. We are grateful to Nicolas Sendrier for pointing this out.
-----
380 R. Misoczki and P.S.L.M. Barreto
**Theorem 1 ([28]). The Goppa code generated by a monic polynomial g(x) =**
(x − _z0) . . . (x −_ _zt−1) without multiple zeros admits a parity-check matrix of the_
_form H = C(z, L), i.e. Hij = 1/(zi_ _Lj), 0 ⩽_ _i < t, 0 ⩽_ _j < n._
_−_
This theorem (also appearing in [15, Ch. 12, §3, Pr. 5]) is entirely general when
one considers the factorization of the Goppa polynomial over its splitting field,
in which case a single root of g is enough to completely characterize the code.
For simplicity, we will restrict our attention to the case where all roots of that
polynomial are in the field Fq itself.
**3.1** **Building a Binary Goppa Code in Dyadic Form**
We now show how to build a binary Goppa code that admits a parity-check
matrix in dyadic form. To this end we seek a way to construct dyadic Cauchy
matrices. The following theorem characterizes all matrices of this kind.
**Theorem 2. Let H ∈** F[n]q _[×][n]_ _with n > 1 be simultaneously a dyadic matrix_
_H = ∆(h) for some h ∈_ F[n]q _[and a Cauchy matrix][ H][ =][ C][(][z, L][)][ for two dis-]_
_joint sequences z ∈_ F[n]q _[and][ L][ ∈]_ [F]q[n] _[of distinct elements. Then][ F][q]_ _[is a field of]_
_characteristic 2, h satisfies_
1
_hi⊕j_
= [1]
_hi_
+ [1]
_hj_
+ [1]
_h0_
_,_ (1)
_and zi = 1/hi + ω, Lj = 1/hj + 1/h0 + ω for some ω ∈_ Fq.
_Proof. Since a dyadic matrix is symmetric, the sequences that define it must_
satisfy 1/(zi − _Lj) = 1/(zj −_ _Li), hence Lj = zi + Li −_ _zj for all i and j. Then_
_zi + Li must be a constant α, and taking i = 0 in particular this simplifies to_
_Lj = α−zj. Substituting back into the definition Mij = 1/(zi_ _−Lj) one sees that_
_Hij = 1/(zi + zj + α). But dyadic matrices also have constant diagonal, namely,_
_Hii = 1/(2zi + α) = h0. This is only possible if all zi are equal (contradicting_
the definition of a Cauchy matrix), or else if the characteristic of the field is 2,
as claimed.
In this case we see that α = 1/h0, and hence Hij = 1/(zi + zj + 1/h0).
Plugging in the definition Hij = hi⊕j we get 1/Hij = 1/hi⊕j = zi + zj + 1/h0,
and taking j = 0 in particular this yields 1/hi = zi + z0 + 1/h0, or simply
_zi = 1/hi + 1/h0 + z0. Substituting back one obtains 1/hi⊕j = zi + zj + 1/h0 =_
1/hi + 1/h0 + z0 + 1/hj + 1/h0 + z0 + 1/h0 = 1/hi + 1/hj + 1/h0, as expected.
Finally, define ω = 1/h0 + z0 and substitute into the derived relations zi =
1/hi +1/h0 + _z0 and Lj = α_ _−_ _zj to get zi = 1/hi +_ _ω and Lj = 1/hj +1/h0 +_ _ω,_
as desired. _⊓⊔_
Therefore all we need is a method to solve Equation 1. The technique we propose
consists of simply choosing distinct nonzero h0 and hi at random where i scans
all powers of two smaller than n, and setting all other values as
1
_hi+j ←_ 1
+ [1]
_hi_ _hj_
+ [1]
_h0_
-----
Compact McEliece Keys from Goppa Codes 381
for 0 < j < i (so that i + j = i ⊕ _j), as long as this value is well-defined._
Algorithm 1 captures this idea. Since each element of the signature h is assigned
a value exactly once, its running time is O(n) steps. The notation u _←$_ _U means_
that variable u is uniformly sampled at random from set U . For convenience
we also define the essence of h to be the sequence ηs = 1/h2s + 1/h0 for s =
0, . . ., ⌈lg n⌉− 1 together with η⌈lg n⌉ = 1/h0, so that, for i = [�]k[⌈][lg]=0[ n][⌉−][1] _ik2[k],_
1/hi = η⌈lg n⌉ + [�]k[⌈][lg]=0[ n][⌉−][1] _ikηk._
**Algorithm 1. Constructing a binary Goppa code in dyadic form**
Input: q (a power of 2), n ⩽ _q/2, t._
Output: Support L, generator polynomial g, dyadic parity-check matrix H for a bi
nary Goppa code Γ (L, g) of length n and design distance 2t + 1 over Fq, and the
essence η of the signature of H.
1: U ← Fq \ {0}
_▷_ Choose the dyadic signature (h0, . . ., hn−1). N.B. Whenever hj with j > 0 is taken
from U, so is 1/(1/hj + 1/h0) to prevent a potential spurious intersection between
_z and L._
2: h0 _←$_ _U_
3: η⌈lg n⌉ _←_ 1/h0
4: U ← _U \ {h0}_
5: for s ← 0 to ⌈lg n⌉− 1 do
6:7: _ih ←i_ _←$2U[s]_
8: _ηs ←_ 1/hi + 1/h0
9: _U ←_ _U \ {hi, 1/(1/hi + 1/h0)}_
10: **for j ←** 1 to i − 1 do
11: _hi+j ←_ 1/(1/hi + 1/hj + 1/h0)
12: _U ←_ _U \ {hi+j_ _, 1/(1/hi+j + 1/h0)}_
13: **end for**
14: end for
15: ω _←$_ Fq
_▷_ Assemble the Goppa generator polynomial:
16: for i ← 0 to t − 1 do
17: _zi ←_ 1/hi + ω
18: end for
19: g(x) ← [�]i[t]=0[−][1] [(][x][ −] _[z][i][)]_
_▷_ Compute the support:
20: for j ← 0 to n − 1 do
21: _Lj ←_ 1/hj + 1/h0 + ω
22: end for
23: h ← (h0, . . ., hn−1)
24: H ← _∆(t, h)_
25: return L, g, H, η
**Theorem 3. Algorithm 1 produces up to** [�]i[⌈]=0[lg][ n][⌉] (q − 2[i]) Goppa codes in dyadic
_form._
-----
382 R. Misoczki and P.S.L.M. Barreto
_Proof. Each dyadic signature produced by Algorithm 1 is entirely determined_
by the values h0 and h2s for s = 0, . . ., ⌈lg n⌉− 1 chosen at steps 2 and 7 (ω
only produces equivalent codes). Along the loop at line 5, exactly 2i = 2[s][+1]
elements are erased from U, corresponding to the choices of h2s . . . h2s+1−1. At
the end of that loop, 2 + 2 [�]ℓ[s]=0 [2][ℓ] [= 2][s][+2][ elements have been erased in total.]
Hence at the beginning of each step of the loop only 2[s][+1] elements had been
erased from U, i.e. there are q − 2[s][+1] elements in U to choose h2s from, and
_q −_ 1 possibilities for h0. Therefore this construction potentially yields up to
(q − 1) [�]s[⌈][lg]=0[ n][⌉−][1] (q − 2[s][+1]) = [�]i[⌈]=0[lg][ n][⌉] (q − 2[i]) possible codes. _⊓⊔_
Theorem 3 actually establishes the number of distinct essences of dyadic signatures corresponding to Cauchy matrices. The roots of the Goppa polynomial are
completely specified by the first ⌈lg t⌉ elements of the essence η together with
_η⌈lg n⌉, namely, zi = η⌈lg n⌉_ + [�][⌈]k[lg]=0[ t][⌉−][1] _ikηk, disregarding the ω term which is im-_
plicit in the choice of η⌈lg n⌉. We see that any permutation of the essence elements
_η0, . . ., η⌈lg t⌉−1 only changes the order of those roots. Since the Goppa polyno-_
mial itself is defined by its roots regardless of their order, the total number of
possible Goppa polynomials is therefore ��⌈i=0lg t⌉ [(][q][ −] [2][i][)]� _/⌈lg t⌉! ≈_ (q−t)�⌈lgq t⌉�.
For n ≈ _q/2 the number of dyadic codes can be approximated by q[m]Q = 2[m][2]Q_
where Q = [�]i[∞]=1 [(1][ −] [1][/][2][i][)][ ≈] [0][.][2887881][. We will also see that the number]
of quasi-dyadic codes, which we describe next and propose for cryptographic
applications, is larger than this. Before we proceed, however, it is interesting to
notice that one of the reasons the attack proposed in [22] succeeds against certain
quasi-cyclic codes, besides the constrained structure of the applied permutation,
is that those schemes start from a known BCH or Reed-Solomon code which is
unique up to the choice of a primitive element from the underlying finite field.
Thus, in those proposals an initial code over F2m is at best chosen from a set
of O(2[m]) codes. In comparison, we start from a secret code sampled from a
much larger family of O(2[m][2] ) codes. For instance, while those proposals have
only 2[15] starting points over F216, our scheme can sample a family with more
than 2[254] codes over the same field. The main protection of the hidden trapdoor
is, of course, the block puncturing process and the more complex blockwise
permutation of the initial secret code, as detailed next.
**3.2** **Constructing Quasi-Dyadic, Permuted Subfield Subcodes**
To complete the construction it is necessary to choose a compact generator
matrix for the subfield subcode. Although the parity check matrix H built by
Algorithm 1 is dyadic over Fq, the usual trace construction leads to a generator
of the dual code that most probably violates the dyadic symmetry. However, by
representing each field element to a basis of Fq over the subfield Fp, one can view
_H as a superposition of d = [Fq : Fp] distinct dyadic matrices over Fp, and each_
of them can be stored in a separate dyadic signature.
A cryptosystem cannot be securely defined on a Goppa code specified directly
by a parity-check matrix in Cauchy form, since this would immediately reveal
-----
Compact McEliece Keys from Goppa Codes 383
the Goppa polynomial g(x): it suffices to solve the overdefined linear system
_zi −_ _Lj = 1/Hij consisting of tn equations in t + n unknowns._
Algorithm 1 generates fully dyadic codes. We now show how to integrate the
techniques of Berger et al. with Algorithm 1 so as to build quasi-dyadic subfield
subcodes whose parity-check matrix is a non-dyadic matrix composed of blocks
of dyadic submatrices. The principle to follow here is to select, permute, and
_scale the columns of the original parity-check matrix so as to preserve quasi-_
dyadicity in the target subfield subcode and the distribution of introduced errors
in cryptosystems. A similar process yields a generator matrix in convenient quasidyadic, systematic form.
For the desired security level (see the discussion in Section 5.1), choose p = 2[s]
for some s, q = p[d] = 2[m] for some d with m = ds, a code length n and a design
number of correctable errors t such that n = ℓt for some ℓ> d. For simplicity
we assume that t is a power of 2, but the following construction method can be
modified to work with other values.
Run Algorithm 1 to produce a code over Fq whose length N ≫ _n is a large_
multiple of t not exceeding the largest possible length q/2, so that the constructed
_t × N parity-check matrix_ _H[ˆ] can be viewed as a sequence of N/t dyadic blocks_
[B0 | · · · | BN/t−1] of size t × t each. Select uniformly at random ℓ distinct
blocks Bi0 _, . . ., Biℓ−1 in any order from_ _H[ˆ]_, together with ℓ dyadic permutations
_Π_ _[j][0]_ _, . . ., Π_ _[j][ℓ][−][1]_ of size t × t and ℓ nonzero scale factors σ0, . . ., σℓ−1 ∈ Fp. Let
_Hˆ_ _[′]_ = [Bi0 _Π_ _[j]0 | · · · | Biℓ−1Π_ _[j]ℓ−1_ ] ∈ (F[t]q[×][t])[ℓ] and Σ = diag(σ0It, . . ., σℓ−1It) ∈
(F[t]p[×][t])[ℓ][×][ℓ]. Compute the co-trace matrix H _[′]_ = Td[′][( ˆ][H] _[′][Σ][) =][ T][ ′]d[( ˆ][H]_ _[′][)][Σ][ ∈]_ [(][F]p[t][×][t])[d][×][ℓ]
and finally the systematic form H of H _[′]. Notice that, if the systematic form_
of Td[′][( ˆ][H] _[′][)][ is][ H][0][, then][ H][ =][ U][ −][1][H][0][V][ where][ U][ = diag(][σ][0][I][t][, . . ., σ][ℓ][−][d][−][1][I][t][)][ and]_
_V = diag(σℓ−dIt, . . ., σℓ−1It)._
The resulting parity-check matrix defines a code of length n and dimension
_k = n−_ _dt over Fp, and since all block operations performed during the Gaussian_
elimination are carried out in the ring ∆(F[t]p[)][, the result still consists of dyadic]
submatrices which can be represented by a signature of length t. Hence the
whole matrix can be stored in an area a factor t smaller than a general matrix.
However, the dyadic submatrices that appear in this process are not necessarily
nonsingular, as they are not associated to a Cauchy matrix anymore; should
all the submatrices on a column be found to be singular (above or below the
diagonal, according to the direction of this process) so that no pivot is possible,
the whole block containing that column may be replaced by another block Bj′
chosen at random from _H[ˆ] as above._
The trapdoor information consisting of the essence η of h, the sequence
(i0, . . ., iℓ−1) of blocks, the sequence (j0, . . ., jℓ−1) of dyadic permutation identifiers, and the sequence of scale factors (σ0, . . ., σℓ−1), relates the public code
defined by H with the private code defined by _H[ˆ]_ . The space occupied by the
trapdoor information is thus m[2] + ℓ lg N + ℓs bits. If one starts with the largest
possible N = 2[m][−][1], this simplifies to the maximal size of m[2] + ℓ(m _−_ 1 + s) bits.
The total space occupied by the essential part of the resulting generator (or
parity-check) matrix over Fp is dt × (n − _dt)/t = dk Fp elements, or mk bits – a_
-----
384 R. Misoczki and P.S.L.M. Barreto
factor t better than plain Goppa codes, which occupy k(n − _k) = mkt bits. Had_
_t not been chosen to be a power of 2, say, t = 2[u]v where v > 1 is odd, the cost_
of multiplying t × t matrices would be in general O(2[u]uv[3]) rather than simply
_O(2[u]u), and the final parity-check matrix would be compressed by only a factor_
2[u].
For each code produced by Algorithm 1, the number of codes generated by
this construction is �N/tℓ � _× ℓ! × t[ℓ]_ _× (r −_ 1)[ℓ], hence �N/tℓ � _× ℓ! × t[ℓ]_ _× (r −_ 1)[ℓ] _×_
�i⌈=0lg N _⌉_ (q − 2[i]) codes are possible in principle.
**3.3** **A Toy Example**
Let F25 = F2[u]/(u[5] + u[2] + 1). The dyadic signature
_h = (u[20], u[3], u[6], u[28], u[9], u[29], u[4], u[22], u[12], u[5], u[10], u[2], u[24], u[26], u[25], u[15])_
and the offset ω = _u[21]_ define a 2-error correcting binary Goppa code
of length N = 16 with g(x) = (x − _u[12])(x −_ _u[15]) and support L_ =
(u[21], u[29], u[19], u[26], u[6], u[16], u[7], u[5], u[25], u[3], u[11], u[28], u[27], u[9], u[22], u[2]). The associated
parity-check matrix built according to Theorem 1 is
_,_
_Hˆ =_
� _u20 u3 u6 u28 u9 u29 u4 u22 u12 u5 u10 u2 u24 u26 u25 u15_
_u[3]_ _u[20]_ _u[28]_ _u[6]_ _u[29]_ _u[9]_ _u[22]_ _u[4]_ _u[5]_ _u[12]_ _u[2]_ _u[10]_ _u[26]_ _u[24]_ _u[15]_ _u[25]_
�
with eight 2 × 2 blocks B0, . . ., B7 as indicated. From this we extract the shortened, rearranged and permuted sequence _H[ˆ]_ _[′]_ = [B7Π [0] _| B5Π_ [1] _| B1Π_ [0] _| B2Π_ [1] _|_
_B3Π_ [0] _| B6Π_ [1] _| B4Π_ [0]] (because in this example the subfield is the base field
itself, all scale factors have to be 1), i.e.:
_u[25]_ _u[15]_ _u[2]_ _u[10]_ _u[6]_ _u[28]_ _u[29]_ _u[9]_ _u[4]_ _u[22]_ _u[26]_ _u[24]_ _u[12]_ _u[5]_
_u[15]_ _u[25]_ _u[10]_ _u[2]_ _u[28]_ _u[6]_ _u[9]_ _u[29]_ _u[22]_ _u[4]_ _u[24]_ _u[26]_ _u[5]_ _u[12]_
�
_,_
_Hˆ =_
�
whose co-trace matrix over F2 has the systematic form:
0 1 0 1 1 0 0 0 0 0 0 0 0 0
1 0 1 0 0 1 0 0 0 0 0 0 0 0
0 1 0 0 0 0 1 0 0 0 0 0 0 0
1 0 0 0 0 0 0 1 0 0 0 0 0 0
0 0 1 1 0 0 0 0 1 0 0 0 0 0
0 0 1 1 0 0 0 0 0 1 0 0 0 0
0 1 1 0 0 0 0 0 0 0 1 0 0 0
1 0 0 1 0 0 0 0 0 0 0 1 0 0
1 1 0 0 0 0 0 0 0 0 0 0 1 0
1 1 0 0 0 0 0 0 0 0 0 0 0 1
⎤
⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦
_H =_
⎡
⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣
= [M [T] _| In−k],_
from which one readily obtains the k _×_ _n = 4_ _×_ 14 generator matrix in systematic
form:
⎡ 1 0 0 0 0 1 0 1 0 0 0 1 1 1 ⎤
0 1 0 0 1 0 1 0 0 0 1 0 1 1
_G =_ 0 0 1 0 0 1 0 0 1 1 1 0 0 0 = [Ik | M ],
⎢⎢⎣ ⎥⎥⎦
0 0 0 1 1 0 0 0 1 1 0 1 0 0
-----
Compact McEliece Keys from Goppa Codes 385
where both G and H share the essential part M :
**0 1 0 1 0 0 0 1 1 1**
1 0 1 0 0 0 1 0 1 1
**0 1 0 0 1 1 1 0 0 0**
1 0 0 0 1 1 0 1 0 0
⎤
_,_
⎥⎥⎦
_M =_
⎡
⎢⎢⎣
which is entirely specified by the elements in boldface and can thus be stored in
20 bits instead of, respectively, 4 · 14 = 56 and 10 · 14 = 140 bits.
## 4 Assessing the Hardness of Decoding Quasi-Dyadic Codes
The original McEliece (or, for that matter, the original Niederreiter) schemes are
perhaps better described as a candidate trapdoor one-way functions rather than
full-fledged public-key encryption schemes. Such functions are used in cryptography in many different settings, each with different security requirements, and
we do not consider such applications in this paper. Instead we focus purely on
the question of inverting the trapdoor function, in other words, decoding.
As we pointed out in Section 1, the well-studied class of Goppa codes remains
one of the best choices to instantiate McEliece-like schemes. Although our proposal is ultimately based on Goppa codes, one may wonder whether or not the
highly composite nature of the Goppa generator polynomial g(x), or the peculiar structure of the quasi-dyadic parity-check and generator matrices, leak any
information that might facilitate decoding without knowledge of the trapdoor.
Yet, any alternant code can be written in Goppa-like fashion by using the
diagonal component of its default parity-check matrix (see Definition 6) to interpolate a generating polynomial (not necessarily of degree t) that is composite
with high probability. We are not aware of any way this fact could be used to
facilitate decoding without full knowledge of the code structure, and clearly any
result in this direction would affect most of the alternant codes proposed for
cryptographic purposes to date.
Otmani et al.’s attack against quasi-cyclic codes [22] could be modified to
work against Goppa codes in dyadic form. For this reason we adopt the same
countermeasures proposed by Berger et al. to thwart it for cyclic codes, namely,
working with a block-shortened subcode of a very large code as described in
Section 3.2. This idea also build upon the work of Wieschebrink [29] who proved
that deciding whether a code is equivalent to a shortened code is NP-complete.
In our case, the result is to hide the Cauchy structure of the private code in a
general dyadic structure, rather than disguising a quasi-cyclic code as another
one with the same symmetry.
We now give a reduction of the problem of decoding the particular class of
quasi-dyadic codes to the well-studied syndrome decoding problem, classical in
coding theory and known to be NP-complete [4].
**Definition 8 (Syndrome decoding). Let Fq be a finite field, and let (H, w, s)**
_be a triple consisting of a matrix H ∈_ F[r]q[×][n], an integer w < n, and a vector
-----
386 R. Misoczki and P.S.L.M. Barreto
_s ∈_ F[r]q[. Does there exist a vector][ e][ ∈] [F][n]q _[of Hamming weight][ wt][(][e][)][ ⩽]_ _[w][ such that]_
_He[T]_ = s[T]?
The corresponding problem for quasi-dyadic matrices reads:
**Definition 9 (Quasi-Dyadic Syndrome Decoding). Let Fq be a finite field,**
_and let (H, w, s) be a triple consisting of a quasi-dyadic matrix H ∈_ _∆(F[ℓ]q[)][r][×][n][,]_
_an integer w < ℓn, and a vector s ∈_ F[ℓr]q _[. Does there exist a vector][ e][ ∈]_ [F]q[ℓn] _of_
_Hamming weight wt(e) ⩽_ _w such that He[T]_ = s[T]?
**Theorem 4. The quasi-dyadic syndrome decoding problem (QD-SDP) is poly-**
_nomially equivalent to the syndrome decoding problem (SDP). In other words,_
_decoding quasi-dyadic codes is as hard in the worst case as decoding general codes._
_Proof. The QD-SDP, being an instance of the SDP restricted to a particular_
class of codes, is clearly a decision problem in NP.
Consider now a generic instance (H _[′], w[′], s[′]) ∈_ F[r]q[×][n] _× Z × F[r]q_ [of the SDP.]
Assume one is given an oracle that solves the QD-SDP over ∆(F[ℓ]q[)][ for some]
given ℓ> 0. Let vℓ _∈_ F[ℓ]q [be the all-one vector, i.e.][ (][v][ℓ][)][j][ = 1][ for all][ j][. Define]
the quasi-dyadic matrix H = H _[′]_ _⊗_ _Iℓ_ _∈_ _∆(F[ℓ]q[)][r][×][n][ with blocks][ H][ij]_ [=][ H]ij[′] _[I][ℓ][, the]_
vector s = s[′] _⊗_ _vℓ_ _∈_ (F[ℓ]q[)][r][ with blocks][ s][i] [=][ s][′]i[v][ℓ][, and][ w][ =][ ℓw][′][. It is evident that]
the instance (H, w, s) ∈ _∆(F[ℓ]q[)][r][×][n]_ _[×][Z][×][(][F][ℓ]q[)][r][ of the QD-SDP can be constructed]_
in polynomial time.
Assume now that there exists e ∈ F[ℓn]q of Hamming weight wt(e) ⩽ _w such_
that He[T] = s[T]. For all 0 ⩽ _i < ℓ, let e[′]i_ _[∈]_ [F]q[n] [be the vector with elements]
(e[′]i[)][j][ =][ e][i][+][jℓ][,][ 0][ ⩽] _[j < n][, so that the][ e][′]i_ [are interleaved to compose][ e][. Obviously]
at least one of the e[′]i [has Hamming weight not exceeding][ w/ℓ] [=][ w][′][, and by the]
construction of H any of them satisfies He[′]i[T] = s[′][T], constituting a solution to
the given instance of the SDP. This effectively reduces the SDP to the QD-SDP
for any given ℓ in polynomial time. Thus, the QD-SDP itself is NP-complete. _⊓⊔_
Although this theorem does not say anything about hardness in the average case,
it nevertheless strengthens our claim that the family of codes we propose is in
principle no less suitable for cryptographic applications than a generic code, in
the sense that, should the QD-SDP problem turn out to be feasible in the worst
case, then all coding-based cryptosystems would definitely be ruled out, regardless of which code is used to instantiate them. Incidentally, the expected running
time of all known algorithms for the SDP (and the QD-SDP) is exponential, so
there is empirical evidence that the average case is also very hard. We stress,
however, that particular cryptosystems based on quasi-dyadic codes will usually
depend on more specific security assumptions, whose assessment transcends the
scope of this paper.
## 5 Efficiency Considerations
Due to their simple structure the matrices in our proposal can be held on a
simple vector not only for long-term storage or transmission, but for processing
as well.
-----
Compact McEliece Keys from Goppa Codes 387
The operation of multiplying a vector by a (quasi-)dyadic matrix is at the
core of McEliece encryption. The fast Walsh-Hadamard transform (FWHT) [12]
approach for dyadic convolution via lifting[2] to characteristic 0 leads to the
asymptotic complexity O(n lg n) for this operation and hence also for encoding. Sarwate’s decoding method [24] sets the asymptotic cost of that operation
at roughly O(n lg n) as well for the typical cryptographic setting t = O(n/ lg n).
Inversion, on the other hand, can be carried out in O(n) steps: one can show
by induction that a binary dyadic matrix ∆(h) of dimension n satisfies ∆[2] =
([�]i _[h][i][)][2][I][, and hence its inverse, when it exists, is][ ∆][−][1][ = (][�]i_ _[h][i][)][−][2][∆][, which]_
can be computed in O(n) steps since it is entirely determined by its first row.
Converting a quasi-dyadic matrix to systematic (echelon) form involves a
Gaussian elimination incurring about d[2]ℓ products of dyadic t × t submatrices, implying a complexity O(d[2]ℓt lg t) = O(d[2]n lg n), and hence the overall cost
of formatting is O(n lg n) as long as d is a small constant, which is indeed the
case in practice since maximum size reduction is achieved when Fp is a large
proper subfield of Fq (see Section 5.1). Notice that, contrary to systems based
on quasi-circulant matrices [8, Proposition 3.4], our proposal does not require
a lengthy process, involving expensive O(n[3]) matrix rank computations to construct a generator matrix in suitable form, often larger than one would expect
for a code of the given dimension.
Table 1 summarizes the asymptotic complexities of code generation (mainly
due to systematic formatting), encoding and decoding, which coincide with the
complexities of key generation, encryption and decryption of typical cryptosystems based on codes.
**Table 1. Operation complexity relative to the code length n**
operation generic ours
Code generation O(n[3]) O(n lg n)
Encode/Decode O(n[2]) O(n lg n)
**5.1** **Suggested Parameters**
Several trade-offs are possible when choosing parameters for a particular application. One may wish to minimize the key size, or increase speed, or simplify the
underlying arithmetic, or attaining a balance between them. We present here
some non-exhaustive combinations. The number of errors is always a power of 2
to enable maximum size reduction.
Table 2 shows the influence of varying the subfield degree while keeping fixed
the approximate security level and the number of design errors. In general, codes
over larger subfields allow for smaller keys as already indicated in [3]. For these
parameters the number of possible codes ranges from 2[392] to 2[731].
2 We are grateful to Dan Bernstein for suggesting the lifting technique to emulate the
FWHT in characteristic 2.
-----
388 R. Misoczki and P.S.L.M. Barreto
**Table 2. Sample parameters for a fixed number of errors (t = 128) and approximately**
128-bit security level, using a subcode over the subfield F2[s] of F216
_s_ _n_ _k_ size (bits)
1 4096 2048 32768
2 2560 1536 24576
4 1408 896 14336
8 768 512 8192
Table 3 displays a different trade-off whereby the key size and the subfield
are kept constant at the cost of varying the number of errors and the code
length. The estimated security level on column ‘level’ refers to the approximate
logarithmic cost of the best known attack according to the guidelines in [7].
**Table 3. Sample parameters for a fixed key size (8192 bits, corresponding to k = 512),**
using a subcode over the subfield F28 of F216
_n_ _t_ level
640 64 102
768 128 136
1024 256 168
One more trade-off is obtained by defining the subfield subcode over the base
field itself, following the common practice for generic codes. The corresponding
settings[3] are summarised on Table 4.
**Table 4. Sample parameters for a subcode over the base subfield F2 of F216**
level _n_ _k_ _t_ size (bits)
80 2304 1280 64 20480
112 3584 1536 128 24576
128 4096 2048 128 32768
192 7168 3072 256 49152
256 8192 4096 256 65536
Table 5 contains a variety of balanced parameters for practical security lev
els. Although we do not recommend these for actual deployment before further
analysis is carried out, these parameters were chosen to stress the possibilities
of our proposal while giving a realistic impression of what one might indeed
3 The actual security levels computed according to the attack strategy in [7] for the
parameters suggested in Table 4 are, respectively, 84.3, 112.3, 136.5, 216.0, and 265.1.
We are grateful to Christiane Peters for kindly providing these estimates.
-----
Compact McEliece Keys from Goppa Codes 389
adopt in practice. The target security level, roughly corresponding to the estimated logarithmic cost of the best known attack according to the guidelines
in [7], is shown on the ‘level’ column. The ‘size’ column contains the amount of
bits effectively needed to store a quasi-dyadic generator or parity-check matrix
in systematic form. The size of a corresponding systematic matrix for a generic
Goppa code at roughly the same security level as suggested in [7] is given on
column ‘generic’. The ‘shrink’ column contains the size ratio between such a
generic matrix and a matching quasi-dyadic matrix. The ‘RSA’ column lists the
typical size of a (quantum-susceptible) RSA modulus at the specified security
level (more accurate RSA estimates can be found in [20,21]). To assess our results against what can be achieved by other post-quantum settings, column ‘QC’
lists key sizes for quasi-cyclic codes of approximately the specified security level
(although not necessarily for the same code length, dimension, and distance)
as suggested in [3], column ‘LDPC’ does the same for (quasi-cyclic) low-density
parity-check codes as discussed in [2], and finally the ‘NTRU’ column contains
the range (from size-optimal to speed-optimal) of NTRU key sizes as suggested
in the draft IEEE 1363.1 standard [13]. For these very compact parameters the
number of possible codes ranges between 2[346] and 2[392], less than those of Table 2
but still very large.
**Table 5. Sample parameters for a subcode over the subfield F28 of F216**
level _n_ _k_ _t_ size generic shrink RSA QC LDPC NTRU
80 512 256 128 4096 460647 112 1024 6750 49152 –
112 640 384 128 6144 1047600 170 2048 14880 – 4411–7249
128 768 512 128 8192 1537536 188 3072 20400 – 4939–8371
192 1280 768 256 12288 4185415 340 7680 – – 7447–11957
256 1536 1024 256 16384 7667855 468 15360 – – 11957–16489
For the parameters on Table 5, we observed the timings on Table 6 (measured
in ms) for generic Goppa codes and quasi-dyadic (QD) codes, and also for RSA
to assess the efficiency relative to a very common pre-quantum cryptosystem. We
made no serious attempt at optimizing the implementation, which was done in
C++ and tested on an AMD Turion 64X2 2.4 GHz. Benchmarks for RSA-15360
were omitted due to the enormous time needed to generate suitable parameters.
**Table 6. Benchmarks for typical parameters**
level generation encoding decoding
RSA generic QD RSA generic QD RSA generic QD
80 563 375 17.2 0.431 0.736 0.817 15.61 1.016 3.685
112 1971 1320 18.7 1.548 1.696 1.233 110.34 2.123 4.463
128 4998 2196 20.5 3.467 2.433 1.575 349.91 3.312 5.261
192 628183 13482 47.6 22.320 6.872 4.695 5094.10 8.822 17.783
256 – 27161 54.8 – 12.176 6.353 – 15.156 21.182
-----
390 R. Misoczki and P.S.L.M. Barreto
## 6 Conclusion and Further Research
We have described how to generate Goppa codes in quasi-dyadic form suitable for
cryptographic applications. Key sizes for a typical McEliece-like cryptosystem
are roughly a factor t = O[˜](n) smaller than generic Goppa codes, and keys
can be kept in this compact size not only for storing and transmission but for
processing as well. In the binary case these codes can correct the full design
number of errors. This brings the size of cryptographic keys to within a factor 4
or less of equivalent RSA keys, comparable to NTRU keys. Our work provides
an alternative to conventional cyclic and quasi-cyclic codes, and benefits from
the same trapdoor-hiding techniques proposed by Wieschebrink in general [29],
and by Berger et al. for that family of codes [3].
The complexity of all operations in McEliece and related cryptosystems is
reduced to O(n lg n). Other cryptosystems can also benefit from dyadic codes,
e.g. entity identification and certain digital signatures for which double circulant
codes have been proposed [9] could use dyadic codes instead, even random ones
without a Goppa trapdoor. One further line of research is whether one can securely combine the techniques in [2] with ours to define quasi-dyadic, low-density
parity-check (QD-LDPC) codes that are suitable for cryptographic purposes and
potentially even shorter than plain quasi-dyadic codes.
Interestingly, it is equally possible to define lattice-based cryptosystems with
short keys using dyadic lattices entirely analogous to ideal (cyclic) lattices as
proposed by Micciancio [17], and achieving comparable size reduction. We leave
this line of inquiry for future research since it falls outside the scope of this
paper.
## Acknowledgments
We are most grateful and deeply indebted to Marco Baldi, Dan Bernstein, PierreLouis Cayrel, Philippe Gaborit, Steven Galbraith, Robert Niebuhr, Christiane
Peters, Nicolas Sendrier, and the anonymous reviewers for their valuable comments and feedback during the preparation of this work.
## References
1. Baldi, M., Chiaraluce, F.: Cryptanalysis of a new instance of McEliece cryptosys
tem based on QC-LDPC code. In: IEEE International Symposium on Information
Theory – ISIT 2007, Nice, France, pp. 2591–2595. IEEE, Los Alamitos (2007)
2. Baldi, M., Chiaraluce, F., Bodrato, M.: A new analysis of the mcEliece cryptosys
tem based on QC-LDPC codes. In: Ostrovsky, R., De Prisco, R., Visconti, I. (eds.)
SCN 2008. LNCS, vol. 5229, pp. 246–262. Springer, Heidelberg (2008)
3. Berger, T.P., Cayrel, P.-L., Gaborit, P., Otmani, A.: Reducing key length of
the McEliece cryptosystem. In: Preneel, B. (ed.) AFRICACRYPT 2009. LNCS,
vol. 5580, pp. 77–97. Springer, Heidelberg (2009),
[http://www.unilim.fr/pages_perso/philippe.gaborit/reducing.pdf](http://www.unilim.fr/pages_perso/philippe.gaborit/reducing.pdf)
-----
Compact McEliece Keys from Goppa Codes 391
4. Berlekamp, E., McEliece, R., van Tilborg, H.: On the inherent intractability of
certain coding problems. IEEE Transactions on Information Theory 24(3), 384–
386 (1978)
5. Bernstein, D.J.: List decoding for binary Goppa codes (2008) (preprint),
[http://cr.yp.to/papers.html#goppalist](http://cr.yp.to/papers.html#goppalist)
6. Bernstein, D.J., Buchmann, J., Dahmen, E.: Post-Quantum Cryptography.
Springer, Heidelberg (2008)
7. Bernstein, D.J., Lange, T., Peters, C.: Attacking and defending the mcEliece cryp
tosystem. In: Buchmann, J., Ding, J. (eds.) PQCrypto 2008. LNCS, vol. 5299, pp.
31–46. Springer, Heidelberg (2008),
[http://www.springerlink.com/content/68v69185x478p53g](http://www.springerlink.com/content/68v69185x478p53g)
8. Gaborit, P.: Shorter keys for code based cryptography. In: International Workshop
on Coding and Cryptography – WCC 2005, Bergen, Norway, pp. 81–91. ACM
Press, New York (2005)
9. Gaborit, P., Girault, M.: Lightweight code-based authentication and signature. In:
IEEE International Symposium on Information Theory – ISIT 2007, Nice, France,
pp. 191–195. IEEE, Los Alamitos (2007)
10. Gibson, J.K.: Severely denting the Gabidulin version of the McEliece public key
cryptosystem. Designs, Codes and Cryptography 6(1), 37–45 (1995)
11. Gibson, J.K.: The security of the Gabidulin public key cryptosystem. In: Maurer,
U.M. (ed.) EUROCRYPT 1996. LNCS, vol. 1070, pp. 212–223. Springer, Heidelberg (1996)
12. Gulamhusein, M.N.: Simple matrix-theory proof of the discrete dyadic convolution
theorem. Electronics Letters 9(10), 238–239 (1973)
13. IEEE P1363 Working Group. IEEE 1363-1: Standard Specifications for Public-Key
Cryptographic Techniques Based on Hard Problems over Lattices, Draft (2009),
[http://grouper.ieee.org/groups/1363/lattPK/index.html](http://grouper.ieee.org/groups/1363/lattPK/index.html)
14. Loidreau, P., Sendrier, N.: Some weak keys in McEliece public-key cryptosystem.
In: IEEE International Symposium on Information Theory – ISIT 1998, Boston,
USA, p. 382. IEEE, Los Alamitos (1998)
15. MacWilliams, F.J., Sloane, N.J.A.: The theory of error-correcting codes. North
Holland Mathematical Library, vol. 16 (1977)
16. McEliece, R.: A public-key cryptosystem based on algebraic coding theory. The
Deep Space Network Progress Report, DSN PR 42–44 (1978),
[http://ipnpr.jpl.nasa.gov/progressreport2/42-44/44N.PDF](http://ipnpr.jpl.nasa.gov/progressreport2/42-44/44N.PDF)
17. Micciancio, D.: Generalized compact knapsacks, cyclic lattices, and efficient one
way functions. Computational Complexity 16(4), 365–411 (2007)
18. Monico, C., Rosenthal, J., Shokrollahi, A.: Using low density parity check codes
in the McEliece cryptosystem. In: IEEE International Symposium on Information
Theory – ISIT 2000, Sorrento, Italy, p. 215. IEEE, Los Alamitos (2000)
19. Niederreiter, H.: Knapsack-type cryptosystems and algebraic coding theory. Prob
lems of Control and Information Theory 15(2), 159–166 (1986)
20. European Network of Excellence in Cryptology (ECRYPT). ECRYPT yearly re
port on algorithms and keysizes (2007-2008). D.SPA.28 Rev. 1.1, IST-2002-507932
ECRYPT, 07/2008 (2008),
[http://www.ecrypt.eu.org/ecrypt1/documents/D.SPA.28-1.1.pdf](http://www.ecrypt.eu.org/ecrypt1/documents/D.SPA.28-1.1.pdf)
21. National Institute of Standards and Technology (NIST). Recommendation for key
management – part 1: General (2007),
[http://csrc.nist.gov/publications/nistpubs/800-57/](http://csrc.nist.gov/publications/nistpubs/800-57/sp800-57-Part1-revised2_Mar08-2007.pdf)
[sp800-57-Part1-revised2_Mar08-2007.pdf](http://csrc.nist.gov/publications/nistpubs/800-57/sp800-57-Part1-revised2_Mar08-2007.pdf)
-----
392 R. Misoczki and P.S.L.M. Barreto
22. Otmani, A., Tillich, J.-P., Dallot, L.: Cryptanalysis of two McEliece cryptosystems
based on quasi-cyclic codes (2008) (preprint),
[http://arxiv.org/abs/0804.0409v2](http://arxiv.org/abs/0804.0409v2)
23. Patterson, N.J.: The algebraic decoding of Goppa codes. IEEE Transactions on
Information Theory 21(2), 203–207 (1975)
24. Sarwate, D.V.: On the complexity of decoding Goppa codes. IEEE Transactions
on Information Theory 23(4), 515–516 (1977)
25. Schechter, S.: On the inversion of certain matrices. Mathematical Tables and Other
Aids to Computation 13(66), 73–77 (1959),
[http://www.jstor.org/stable/2001955](http://www.jstor.org/stable/2001955)
26. Sendrier, N.: Finding the permutation between equivalent linear codes: the support
splitting algorithm. IEEE Transactions on Information Theory 46(4), 1193–1203
(2000)
27. Sidelnikov, V., Shestakov, S.: On cryptosystems based on generalized Reed
Solomon codes. Discrete Mathematics 4(3), 57–63 (1992)
28. Tzeng, K.K., Zimmermann, K.: On extending Goppa codes to cyclic codes. IEEE
Transactions on Information Theory 21, 721–726 (1975)
29. Wieschebrink, C.: Two NP-complete problems in coding theory with an application
in code based cryptography. In: IEEE International Symposium on Information
Theory – ISIT 2006, Seattle, USA, pp. 1733–1737. IEEE, Los Alamitos (2006)
-----
|
{
"disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references', 'abstract'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1007/978-3-642-05445-7_24?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1007/978-3-642-05445-7_24, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "BRONZE",
"url": "https://link.springer.com/content/pdf/10.1007/978-3-642-05445-7_24.pdf"
}
| 2,009
|
[
"JournalArticle"
] | true
| 2009-11-04T00:00:00
|
[] | 15,683
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Engineering",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/020b47164f980cf8f1c447a85554fe0217c96be6
|
[
"Computer Science"
] | 0.874104
|
A LIGHT-WEIGHT MUTUAL AUTHENTICATION AND KEY-EXCHANGE PROTOCOL BASED ON ELLIPTICAL CURVE CRYPTOGAPHY FOR ENERGY-CONSTRAINED DEVICES
|
020b47164f980cf8f1c447a85554fe0217c96be6
|
[
{
"authorId": "2275441",
"name": "K. Yow"
},
{
"authorId": "2064749211",
"name": "Amol Dabholkar"
}
] |
{
"alternate_issns": null,
"alternate_names": null,
"alternate_urls": null,
"id": null,
"issn": null,
"name": null,
"type": null,
"url": null
}
| null |
A LIGHT-WEIGHT MUTUAL AUTHENTICATION AND KEY-EXCHANGE PROTOCOL BASED
ON ELLIPTICAL CURVE CRYPTOGAPHY FOR ENERGY-CONSTRAINED DEVICES
Kin Choong Yow and Amol Dabholkar
School of Computer Engineering,
Nanyang Technological University,
Singapore 639798,
email: kcyow@ntu.edu.sg, amold@pmail.ntu.edu.sg
**Abstract**
Wireless devices are characterized by low computational power and memory. Hence security protocols
dealing with these devices have to be designed to give minimal computational and memory load. We
present an efficient authentication and key exchange protocol for low-end wireless clients and high end
servers, which is overall nearly three times as fast as comparable protocols. The basic idea of our
protocol is to use symmetric key encryption in place of public key encryption wherever possible.
**1.** **Introduction**
Wireless environment is by its very nature insecure and limited in bandwidth [1]. For example, an
adversary could easily tap into wireless communication channels or jam other people’s devices.
Furthermore personal wireless devices are generally of low power, low in computational abilities and
low in memory. These reasons have prevented a simple migration of cryptographic protocols from
fixed networks to wireless networks for authentication and security [2].
The first phase of a typical secure transaction or communication is a secure key exchange where both
parties decide on a common secret-key known only to them. The security of the rest of the transaction
then depends on this first step and if there is a compromise at this phase, the rest of the communication
may be compromised.
In this paper we will look at protocols for key exchange and ways for devices to authenticate each
other. We will be looking specifically at protocols that combine these two objectives, which are called
Mutually Authenticated Key Exchange Protocols or MAKEPs. The goal of these protocols is to provide
the communicating parties with some assurance that they know each others’ true identity and at the
same time have a common key known only to them [3].
We will first look into the design strategies for a typical MAKEP and study two recently proposed
protocols that deal with similar scenarios. We will then present our protocol and discuss its design and
security. Finally we will compare implementations of the protocols based on elliptic curve
cryptography using PDA clients.
**2.** **Design Strategies**
In this section we will briefly analyze the designs of two MAKEPs designed for low-powered clients
and discuss their advantages and disadvantages.
10.5121/ijnsa.2010.2210 123
-----
_2.1 Server-Specific MAKEP_
This protocol was proposed in [2] as an efficient MAKEP for communication between a low powered
wireless client and a powerful server. The main feature of this MAKEP is that it eliminates any public
key cryptographic operations on the client side. The client side performs only symmetric key
operations. Symmetric key cryptography is much faster than public key cryptography [4]. Hence the
protocol executes much faster on the client side, than it would have if public key operations were
involved.
The protocol relies on server-specific certificates of the form given by _Cert = <IDAB_ A, {KA}PKB,
SigTA(IDA, {KA}PKB)> which is a certificate from the low powered wireless client _Alice to the high_
powered server Bob.
Here KA is Alice’s long-lived symmetric secret key. PKB is the well known public key of B. {KA}PKB
means _Alice’s secret key KA is encrypted with_ _Bob’s public key PKB. SigTA is a signature from a_
trusted authority TA, and IDA is Alice’s identity.
Alice
_Bob_
rKAAЄ (random)Zq PK, SK
{rA}KA, _Cert_ _AB_
Verfify Cert[B]A
{rA, rB, IDB}KA rB Є (random)Zq
{rB}KA
σ = rA ⊕ rB σ = rA ⊕ rB
Figure 1: Server-Specific MAKEP protocol
On receiving the first message from _Alice,_ _Bob verifies the TA’s signature and obtains_ _Alice’s long_
lived symmetric key KA, by decrypting {KA}PKB from the certificate by using his private key SKB. This
step also authenticates _Alice._ _Bob then encrypts a new random value_ _rB along with_ _Alice’s random_
value _rA and his identity IDB, with_ _Alice’s symmetric key KA. This step authenticates_ _Bob. Now the_
session key σ is calculated as rA ⊕ _rB._
The protocol completes in 3 steps with hardly any pre-computations. Another major advantage is that
the client uses only symmetric-key cryptography with her secret key KA. This increases its speed and
efficiency as compared to other comparable MAKEPs that use public-key cryptography. The freshness
of the session key is maintained by using the random numbers _rA and_ _rB. Unlike the Jakobsson-_
124
-----
Pointcheval protocol [5], where only the client computes the session key σ, here we use client and
server contributions, rA and rB respectively, to calculate σ. This increases the strength of the key.
However, there are some disadvantages to this protocol in its raw form –
1. The protocol needs server specific certificates. For n servers it will need n distinct certificates. This
creates scalability issues.
2. The client’s long lived secret key KA is known to the server. If the server is malicious it can create
nonexistent sessions and there is no way for Alice to deny these occurred, i.e. a malicious server can
impersonate its clients to create false runs of the protocol.
3. Even if the server is trustworthy, it has access to the client’s secret key KA. If the server’s security
is compromised in any way it implies the client’s security is also compromised since a hacker can
get KA from Bob’s memory.
These problems can be taken care of by our proposed protocol described in the next section.
_2.2 Client-Server MAKEP_
This protocol solves the scalability and other problems of the Server-Specific protocol by using publickey cryptography on the client side as well as the server side [3]. However, the public-key
computations are kept to a minimum and most of the costly operations are done on the server side.
_Alice chooses a random value_ _rA as its contribution towards creating the shared key, and encrypts it_
using Bob’s well known public key PKB. Alice also computes a random value ‘b’ and sends β = g[b] to
_Bob._
_Bob replies by sending his shared key contribution_ _rB, encrypted with_ _Alice’s contribution_ _rA. This_
authenticates Bob to Alice. Since rA had been encrypted under Bob’s private key, only Bob will know
_rA, and rB encrypted under rA implies that Bob is replying to the current session. Alice sends the value y_
= ah(σ) + b mod q (q is the prime number used to define the field Fq) to Bob where σ is computed from
_rA and rB. Bob checks if g[y] = (g[a])[h(][σ][)]β. Now g[y] = g[(ah(][σ][) + b)] = (g[a])[h(][σ][)]g[b] = (g[a])[h(][σ][)]β. Bob knows g[a] from_
the certificate and he has received β in the previous message. Thus Alice is authenticated as the secret
key portion of ‘a’ in the value of y can only be computed by Alice, and the β value effectively binds the
earlier and later messages, as Alice will compute a new value of ‘b’ for each session.
125
-----
Figure 2: Client Server MAKEP
The third step is necessary because Alice has not been authenticated to Bob. Note that in the first step,
when _Alice sends “CertA,_ β, x” to _Bob, this does not authenticate_ _Alice, as anyone could send this_
message by capturing previous protocol runs. The dotted line shows the binding between the first and
the third step by b and β. The third step combines the secret key (a) and the session random (b) value of
_Alice to authenticate her to Bob. If Alice was authenticated in the first step itself, the third step will not_
be necessary.
This protocol has been shown to be secure against attacks by using the Bellare-Rogaway [6] model.
Since it uses general certificates there is no scaling problem like the server-specific MAKEP. Most of
the computationally expensive operations like Certificate verification are carried out on the server-side.
**3.** **Our Proposed Protocol**
We observe that by using the keys of the client and the server to make a Diffie-Hellman [7] key pair in
the first message, we can have a protocol that can be executed in just 2 steps. This removes the need of
a separate binding ‘b’ and β between the first and the last message, and further removes the need for the
server to perform exponential calculations to verify and authenticate the client..
The protocol is shown in figure 3. Let _g be the generator of the field over which the DH problem is_
intractable. The client Alice has a private key ‘a’ and a public key g[a] and the server Bob has a private
key ‘y’ with a public key g[y]. We assume that the client has a certificate CertA from a trusted authority
126
-----
where CertA = <IDA, g[a], SigTA(IDA, g[a] )>. Unlike the Server-Specific MAKEP algorithm in [2] we do
not use a server-specific certificate and avoid scaling issues.
Calculate σ = r ⊕ _r’_
Figure 3: A new MAKEP for wireless devices
_Alice first computes a random number r. She then calculates the DH key g([ya]) using her private key a_
and Bob’s public key _g[y]. Our protocol is unique in the sense that we propose that this DH key be_
converted to an AES key. The subsequent use of the AES symmetric key algorithm in encryption and
decryption of data leads to a substantial improvement in time.
_Alice sends her random contribution r encrypted under the AES key to Bob along with her certificate._
_Bob verifies Alice’s public key g[a] and uses it to calculate the DH key g([ya]) using his own private key y._
He also converts the DH key to an AES key and uses it to decrypt the message and get Alice’s random
number r. Bob computes his own random number r’ and uses r as an AES key to encrypt it and send it
to Alice. Alice can decrypt this message since she already knows her random r.
σ is the new session key and H is a cryptographic hash function. By using r (the client random) and r’
(the server random) we are maintaining the freshness of the session key and the g[(ya)] value is used for
authentication Alice and _Bob to each other, as well as preventing any other party from mounting any_
replay attacks or Denial of Service attacks on Alice or Bob. Since we have authenticated Alice to Bob in
the first step, we don’t need a third step like in [2].
Another variation of this protocol could be made using elliptic curve cryptography (ECC). A 1024 bit
RSA key is comparable to only a 164 bit ECC key [8] i.e. it provides the same level of security so with
the shorter key length the protocol will be much faster. In the ECC version of the protocol all the
variables will now be points on an elliptic curve E. _Alice’s private and public keys will be the field_
element a and the point Pa (obtained by point multiplication of the base point P and the secret key a),
127
-----
while Bob’s keys will be y and Py. P is a publicly known base point. Instead of encrypting r with g[(ya)]
we will encrypt it with (Pya). The rest of the protocol remains similar to the original one.
**4.** **A More Secure Version of the Protocol**
There is however a problem with the protocol based on trust issues. A weakness has been pointed out
for the Server-Specific MAKEP protocol in [9], which is the server has control over the session key.
If we look at the server side computations before the second message is sent in figure 3, we observe
that the server can decide before hand which session key (σ) will be used by calculating r’ = (r ⊕ σ)
and it can do this because it knows the random value of the client before the final session key is
calculated.
We can easily overcome this by adding a predetermined value before the random contributions before
encrypting them. By doing this, we force the participants to check if the decryption has lead to a valid
result. For example, Alice will send CertA,{IDA, h(r)}g(ya) and Bob will reply with {IDB, r’}h(r).
Decrypt {r}Y
Check h(r)
Calculate x = r ⊕ _r’_
Figure 4: A more secure version
128
-----
As can be seen from figure 4, an extra step has been added to the protocol. In the first step we send
_h(r), which is the hashed value of the client random_ _r. So now the server_ _Bob does not know the_
client’s random number before calculating his own random contribution r’. So there is no way for the
server to decide the session key _x beforehand. In the third step we send the actual client random_ _r,_
encrypted with a symmetric key derived from the DH key g[(][ya][)] and the server random r’. Note that this
will also be a symmetric key encryption and hence almost negligible in comparison to the public key
operations. So the addition of the third step will not cause any difference in the speed of the protocol.
**5.** **Implementation and Results**
We have implemented the ECC version of our protocol and an ECC version of the Client-Server
protocol. The ECC version of the Client-Server protocol has been done by replacing the exponentiation
operations with point multiplication operations, and by replacing the field generator _g with the base_
point P. By doing this we have a level playing field on which to compare both the protocols for timing
efficiency using the same key sizes and the same ECC implementation.
Our implementation of ECC is based on the one described by Rosing [10]. We have used a 206 MHz
64MB RAM HP-Compaq iPAQ PDA as the client and a 1.7 GHz P4 PC as the server.
Table 1: Comparison for the total time taken
**Field Size** **Our Protocol** **Client-Server** **Speed**
**(Bits)** **(secs)** **MAKEP** **comparison**
**(secs)**
158 4.81 15.3 3.18 times
155 4.53 12.44 2.75 times
134 3.45 9.3 2.7 times
119 2.88 7.49 2.6 times
113 2.79 7.03 2.51 times
90 1.41 4.02 2.85 times
65 1.078 2.06 1.9 times
50 1.04 1.63 1.56 times
As we can see from the graph in figure 5, our protocol is almost 1.5 times faster at low field sizes of
around 50 bits (1.04 seconds vs 1.63 seconds) and more than 3 times faster for higher bits (at 158 bits
our protocol runs at around 4.81 seconds as opposed to 15.3 seconds by the Client-Server MAKEP).
129
|Field Size (Bits)|Our Protocol (secs)|Client-Server MAKEP (secs)|Speed comparison|
|---|---|---|---|
|158|4.81|15.3|3.18 times|
|155|4.53|12.44|2.75 times|
|134|3.45|9.3|2.7 times|
|119|2.88|7.49|2.6 times|
|113|2.79|7.03|2.51 times|
|90|1.41|4.02|2.85 times|
|65|1.078|2.06|1.9 times|
|50|1.04|1.63|1.56 times|
-----
Figure 5: Graph of total time comparison
We will now analyze the two protocols by breaking them up into a pre-computation part and a run time
part. The pre-computation portion is the portion of the protocol that does not need to be executed at
runtime. The pre-computational time is given in the table 2.
Table 2: Runtime comparison
**Field Size** **Our Protocol** **Client-Server MAKEP**
**(Bits)**
Pre-comp Run Time Pre-comp Run Time
time (sec) (sec.) time (sec.) (sec.)
158 2.976 1.84 2.97 12.4
155 2.745 1.79 2.74 9.7
134 1.844 1.61 1.83 7.56
119 1.26 1.54 1.309 6.2
113 0.964 1.4 0.988 6.004
90 0.51 1.2 0.531 3.67
65 0.19 0.98 0.211 1.84
50 0.1 0.94 0.101 1.5
The actual time from table 1 minus the pre-computational time from table 2 gives us the run time of the
protocol. Figure 6 shows these values plotted in a graph. If we compare the graphs in figures 6 and 5
we can see that our protocol is still nearly 7 times faster at the higher bit size fields and 1.5 times faster
for the lower bit sizes.
130
|Field Size (Bits)|Our Protocol|Col3|Client-Server MAKEP|Col5|
|---|---|---|---|---|
||Pre-comp time (sec)|Run Time (sec.)|Pre-comp time (sec.)|Run Time (sec.)|
|158|2.976|1.84|2.97|12.4|
|155|2.745|1.79|2.74|9.7|
|134|1.844|1.61|1.83|7.56|
|119|1.26|1.54|1.309|6.2|
|113|0.964|1.4|0.988|6.004|
|90|0.51|1.2|0.531|3.67|
|65|0.19|0.98|0.211|1.84|
|50|0.1|0.94|0.101|1.5|
-----
Figure 6: Actual Run time graph
We will now look at a major improvement we have made in our protocol in the first step. This is
converting the DH key into an AES key and carrying out symmetric key encryption of the client
random value. As opposed to this, we have used El Gamal public key encryption to encrypt the client
random value in the client-server MAKEP protocol. We can see the time difference in table 3.
Table 3: Step 1 Encryption Comparison
**Field** **Size** **Our** **Protocol** **Client-Server** **MAKEP**
**(Bits)** **[AES Encrypt** _rA_ **[El Gamal Encrypt** _rA_
(msecs.)] (msecs.)]
158 13 5600
155 12 5090
134 12 3500
119 12 2700
113 13 2030
90 12 1040
65 12 400
50 12 200
As we can see from table 3, this is a very big improvement in terms of speed.
131
|Field Size (Bits)|Our Protocol [AES Encrypt r A (msecs.)]|Client-Server MAKEP [El Gamal Encrypt r A (msecs.)]|
|---|---|---|
||||
|158|13|5600|
|155|12|5090|
|134|12|3500|
|119|12|2700|
|113|13|2030|
|90|12|1040|
|65|12|400|
|50|12|200|
-----
**6.** **Conclusions**
We have proposed a new and efficient mutual authentication and key exchange protocol for low
powered clients like PDAs in wireless environments. We have analyzed two recently proposed
protocols and have implemented one of them using ECC for comparison with our protocol. We have
seen that our protocol is around 3 times as fast at key sizes of around 160 bits. Furthermore, our
protocol is scalable as it does not require server specific certificates. We use random contributions from
both the client and the server to produces a session key that ensures forward secrecy.
**References**
1. Nichols R. & Lekkas P. (2002) ‘Wireless Security: Models, Threats and Solutions’ McGraw-Hill.
2. Wong D & Chan A. (2001). _‘Mutual Authentication and Key exchange for low power wireless_
_communications’, IEEE MILCOM 2001 Conference Proceedings._
3. Wong D & Chan A. (2001). ‘Efficient and Mutually Authenticated Key Exchange for Low Power
_Computing Devices’. ASIACRYPT 2001._
4. Schneier B., ‘Applied Cryptography’, Wiley (1996).
5. Jakobsson M. & Pointcheval D. (2001) ‘Mutual Authentication for low-power mobile devices’.
Proceedings of Financial Cryptography 2001, Springer-Verlag.
6. Frankel S., Glenn R. et al. ‘RFC 3602: The AES-CBC Cipher Algorithm and Its Use with IPsec’
(Retrieved September 2004 from http://rfc3602.x42.com/.
7. Diffie W. & Hellman M., _‘New Directions in Cryptography’, IEEE Transactions on Information_
Theory, 644-654 (1976)
8. Lenstra A., & Verheul E. (2001) ‘Selecting Cryptographic Key Sizes’, Journal of Cryptology,
14(4):255-293.
9. S.-L. Ng and C. J. Mitchell, 'Comments on mutual authentication and key exchange protocols for
_low_ _power_ _wireless_ _communications',_ Retrieved October 2003 at
http://www.isg.rhul.ac.uk/~cjm/comaak.pdf,
10. Rosing, M. (1999) ‘Implementing Elliptic Curve Cryptography’. Greenwich, CT: Manning.
132
-----
|
{
"disclaimer": "Notice: The following paper fields have been elided by the publisher: {'abstract', 'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.5121/IJNSA.2010.2210?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.5121/IJNSA.2010.2210, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "BRONZE",
"url": "https://doi.org/10.5121/ijnsa.2010.2210"
}
| 2,010
|
[] | true
| 2010-04-25T00:00:00
|
[] | 5,270
|
|
en
|
[
{
"category": "Law",
"source": "s2-fos-model"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Business",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/020b4eec8eb34b2752e7f0c4072f8eedc07d1fd2
|
[] | 0.869871
|
TOOLS TO STIMULATE BLOCKCHAIN: APPLICATION OF REGULATORY SANDBOXES, SPECIAL ECONOMIC ZONES AND PUBLIC PRIVATE PARTNERSHIPS
|
020b4eec8eb34b2752e7f0c4072f8eedc07d1fd2
|
International Journal of Law in Changing World
|
[
{
"authorId": "72777661",
"name": "E. Gromova"
},
{
"authorId": "2058262052",
"name": "D. Ferreira"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"Int J Law Chang World"
],
"alternate_urls": null,
"id": "226eec69-7e47-42e4-99c5-8b17c4cd2b9c",
"issn": "2764-6068",
"name": "International Journal of Law in Changing World",
"type": "journal",
"url": null
}
|
The Blockchain technology has significant and almost limitless potential. However, today their use for implementation is associated with the problems of lack of high-quality legal regulation of this technology; technical standards for its application; investments required for its development. These problems and the search for their solutions are especially relevant now, in the context of the financial crisis. In this regard, the purpose of the article is to analyse the legal mechanisms and tools that make up special and experimental regimes, the use of which contributed to the introduction of the Blockchain technology into industrial production, identifying their features in relation to individual countries, problems associated with their implementation and finding solutions. The research is based on comparative legal and system analysis, as well as methods of legal modelling and content analysis. The author comes to the conclusion that in order to increase the attractiveness of the legal climate for the implementation of the Blockchain technology, it is necessary, first, to develop a “high-quality” legal regulation, which will be possible in the case of prior testing of an innovative product (service) based on the application of this technology in conditions of the experimental legal regime (regulatory sandbox); second, to develop standards for normative and technical regulation of this technology; third, to improve legislation on the main tools aimed at stimulating investment in the creation and implementation of digital innovations, and Blockchain technology, including - on special economic zones, public-private partnerships and state support of companies-developers of the Blockchain services for industrial production.
|
**Volume 2 Issue 1 (2023) ISSN 2764-6068**
**Research article**
**JNL:** [https://ijlcw.emnuvens.com.br/revista](https://ijlcw.emnuvens.com.br/revista)
**DOI:** [https://doi.org/10.54934/ijlcw.v2i1.48](https://doi.org/10.54934/ijlcw.v2i1.48)
# TOOLS TO STIMULATE BLOCKCHAIN: APPLICATION OF REGULATORY
SANDBOXES, SPECIAL ECONOMIC ZONES, AND PUBLIC PRIVATE
PARTNERSHIPS
**Elizaveta A. Gromova**
South Ural State University (National Research University), Russian Federation
**Daniel Brantes Ferreira**
Brazilian Centre for Meditation and Arbitration, Brazil
**Volume 2 Issue 1 (2023) ISSN 2764-6068**
**Article Information:**
Received
April 16, 2023
Approved
April 21, 2022
Accepted
May 3, 2022
Published
June 15, 2022
**Keywords:**
blockchain,
experimental regime,
regulatory sandboxes,
special economic zones,
public private partnership,
standardization,
government support
**FOR CITATION:**
**ABSTRACT**
The Blockchain technology has significant and almost limitless
potential. However, today their use for implementation is associated
with the problems of lack of high-quality legal regulation of this
technology; technical standards for its application; investments
required for its development. These problems and the search for their
solutions are especially relevant now, in the context of the financial
crisis. In this regard, the purpose of the article is to analyse the legal
mechanisms and tools that make up special and experimental regimes,
the use of which contributed to the introduction of the Blockchain
technology into industrial production, identifying their features in
relation to individual countries, problems associated with their
implementation and finding solutions. The research is based on
comparative legal and system analysis, as well as methods of legal
modelling and content analysis. The author comes to the conclusion
that in order to increase the attractiveness of the legal climate for the
implementation of the Blockchain technology, it is necessary to
develop a “high-quality” legal regulation, standards for normative and
technical regulation of this technology, and improve legislation on the
main tools aimed at stimulating investment in the Blockchain
technology.
Gromova, E. A., & Ferreira, D. B. (2023). Tools to Stimulate Blockchain: Application of Regulatory
Sandboxes, Special Economic Zones, and Public Private Partnerships. International Journal of Law in
Changing World, 2 (1), 17-36. DOI: https://doi.org/10.54934/ijlcw.v2i1.48
-----
**1.** **INTRODUCTION**
The Blockchain technology (Blockchain) was first talked about in 2008, after the publication of the
article “Bitcoin: A Peer-to-Peer Electronic Cash System”, written by a group of authors under the
pseudonym S. Nakamoto (2009). This technology (Blockchain 1.0) was recommended to be used to verify
Bitcoin transactions. Initially, this technology was used to verify financial transactions in operations with
cryptocurrencies. However, afterwards, it found its application in the framework of registration of rights
to real estate, medical data, etc. This technology can also be very effectively used for the development of
the so-called “smart” industry. At present, a number of companies in the energy, mining and
manufacturing industries are using this technology, among other needs, to ensure reliable supply chains.
For the successful implementation of this technology in industrial production, states use special and
experimental regimes (special and experimental regulation). Further, with the advent of self-executable
contracts (smart contracts) in 2013, Blockchain technology is gaining even more popularity (Borg and
Schembri, 2019). Smart contracts made it possible to implement through Blockchain technology
(Blockchain 2.0) (Aggarwal and Kumar, 2021) a diverse set of business functions related to the transfer
of information and/or values, while leaving transparent and reliably verifiable information flows[1].
Over time, the capabilities of this technology made it possible to use distributed ledgers not only
within the framework of cryptocurrencies. Since the records in the chain are stored and distributed across
the nodes of the network, they are very difficult to falsify, which makes Blockchain a safe and transparent
way to record transactions and service information. Therefore, today Blockchain is used in the field of
trade, government services, healthcare, tourism and even music[2]. So, for example, the courts of the PRC
use Blockchain to record court hearings (Tran, 2020). Japanese animation studios use Blockchain to
combat anime piracy[3]. The Blockchain technology is actively used in various industries (smart, digital
industry) (Xu et al., 2021). Programs such as IBM Blockchain are designed to improve supply chain, data
identification and management. Blockchain Foundry focuses on Blockchain-based services for
prototyping and industrial production. In manufacturing, 75% of industrial companies are expected to use
distributed ledger systems by 2024. This will reduce the cost of controlling the quality of raw materials
by 50%, and for document circulation by 40%. The share of successful cyber-attacks will be halved[4]. It is
1 White Paper «Blockchain in Trade Facilitation» ECE/TRADE/C/CEFACT/2019/9/Rev.1, available at:
https://unece.org/fileadmin/DAM/cefact/GuidanceMaterials/WhitePaperBlockchain.pdf(accessed 28.01.2023).
2 White Paper «Blockchain in Trade Facilitation» ECE/TRADE/C/CEFACT/2019/9/Rev.1, available at:
https://unece.org/fileadmin/DAM/cefact/GuidanceMaterials/WhitePaperBlockchain.pdf (accessed 28.01.2023).
3 Japan’s Blockchain Sandbox Is Paving The Way For The Fintech Future, available
at:https://www.forbes.com/sites/japan/2019/06/26/japans-blockchain-sandbox-is-paving-the-way-for-the-fintechfuture/?sh=5ef085832795 (accessed 28.01.2023).
4 Russian Blockchain, available at: https://www.cnews.ru/articles/2019-08-27_rossijskim_blokchejnrazrabotchikam (accessed
28.01.2023).
https://doi.org/10.54934/ijlcw.v2i1.48
18
-----
expected that the third generation of the Blockchain technology (Blockchain 3.0) will allow the
development of large-scale industrial applications capable of simultaneously managing many processes,
processing and storing huge amounts of data, ensuring their consistency (Xu et al., 2021).
According to a survey conducted by the World Economic Forum, if in 2015 only 0.025% of global
GDP was based on the use of Blockchain, then by 2027 this ratio is expected to jump to 10%[5]. According
to the respondents of the Deloitte survey conducted in 2020, any state is capable of losing its competitive
advantages if it does not use this technology and its implementation plays a very important role. The cost
of quality control of raw materials will be reduced by 50%, and the cost of document flow - by 40%. The
share of successful cyber-attacks will be halved[6]. The third generation of the Blockchain technology
(Blockchain 3.0) allows the development of large-scale industrial applications that can simultaneously
manage many processes, process and store huge amounts of data, ensuring their logical interconnection
and consistency (Di Francesco and Mori, 2020).
However, today states are faced with universal problems that can level the potential of Blockchain.
These are: lack of high-quality legal regulation of this technology; technical standards for its application;
investments required for its development. These problems are the main barriers to its implementation,
including in industrial production. Lack of regulatory clarity is one such barrier, according to Deloitte's
2020 Blockchain Survey[7]. Consequently, if states are unable to effectively implement this technology,
they can lose their competitive advantages (Swan, 2015).
In this regard, the purpose of the article is to analyze the mechanisms and tools that make up special
and experimental modes, the use of which would contribute to solving these problems. To achieve the
goal of the study, the national regulation of the creation and implementation of the Blockchain technology
in various areas, including industrial production, tools that help to improve the quality of legal (including
regulatory and technical) regulation of this technology, as well as to attract investments in its development.
Currently, many scientific articles and monographs have been published on the Blockchain
technology. These articles were written by representatives of various branches of science and touch on
completely different aspects of the creation and implementation of this technology (Mohamad et al., 2017;
Dong et al., 2018; Fan et al., 2018; Wu et al., 2021). The scientific works analyzed by the author can be
conditionally divided into the following groups:
1. Legal regulation of creation and implementation of the Blockchain technology.
5 Global Agenda Council on the Future of Software & Society Deep Shift Technology Tipping Points and Societal Impact,
available at: http://www3.weforum.org/docs/WEF_GAC15_Technological_Tipping_Points_report_2015.pdf (accessed
28.01.2023).
6 Russian Blockchain, available at: https://www.cnews.ru/articles/2019-08-27_rossijskim_blokchejnrazrabotchikam (accessed
28.01.2023).
7 Deloitte 2020 Blockchain Survey, available at: https://www2.deloitte.com/content/dam/insights/us/articles/6608_2020global-blockchain-survey/DI_CIR%202020%20global%20blockchain%20survey.pdf (accessed 28.01.2023).
https://doi.org/10.54934/ijlcw.v2i1.48
19
-----
2. Legal aspects of implementation of the Blockchain technology in certain areas of public life,
sectors of economy and industrial production.
3. Ways to develop and implement the Blockchain technology.
A significant number of works by authors from different countries are devoted to the legal regulation
of the Blockchain technology, which emphasizes the relevance of legal research on creation and
implementation of the Blockchain technology. As a rule, these articles are devoted to the search for the
optimal model of legal regulation for individual countries, their unions and integration formations. The
works of other researchers are devoted to defining the legal essence of this technology, trying to form its
definition, legal features and classification (Sultan et al., 2018; Bouraga, 2021).
A significant number of scientific articles are devoted to individual legal problems of creation and
implementation of the Blockchain technology, for the most part – the problem of correlation with
legislation on protection of personal data (Ivanc et al., 2016; Mohamad et al., 2017; Tatara et al., 2020;
Campanile et al., 2021). Most of the published scientific papers on this technology aim to describe the role
of the Blockchain technology in the financial market, cryptocurrencies and smart contracts (Alia et al.,
2020; De Filippi et al., 2020; Elisavetsky and Marun, 2020). At the same time, some authors turn to the
legal analysis of application of the Blockchain technology in other areas – medicine, public administration,
etc. (Mohamad et al., 2017; Dong et al., 2018; Fan et al., 2018; Roman-Belmonte et al., 2018; Joppen et
al., 2019; Balasubramaniana et al., 2021).
Some of the works analyzed by the author describe individual ways of developing and implementing
this technology. As a rule, these articles are devoted to the issues of attracting investments in its
development (Jani and Panda, 2019). Without diminishing the importance of the research carried out by
the authors, it should be noted that at the moment the author did not find a comprehensive study of special
and experimental legal regimes contributing to implementation of the Blockchain technology in various
spheres of society, as well as in the industrial production industry, as well as their constituent legal
instruments and mechanisms that would allow overcoming barriers that hinder its development and
promote its implementation.
To achieve the goals of the article, the author applied a set of methods, which included the
comparative legal and systemic method, as well as the method of legal modeling and content analysis. The
comparative legal method was used to analyze approaches to the legal regulation of the Blockchain
technology, the national legislation of the countries implementing this technology. The application of this
method made it possible to identify tools and mechanisms that help to attract investments in the
development of this technology, as well as best practices that contribute to improving the quality of legal,
including regulatory and technical, regulation of this technology.
https://doi.org/10.54934/ijlcw.v2i1.48
20
-----
The systematic method made it possible to consider the legal instruments and mechanisms that
contribute to creation and implementation of the Blockchain technology as a single system of techniques
and methods, the use of which, in aggregate, will overcome the barriers that hinder the development and
implementation of this technology. The method of content analysis made it possible to analyze the content
of individual information resources to identify existing practices for the implementation of the Blockchain
technology, as well as to attract investments in its development and state support of its developers.
**2.** **TOOLS TO STIMULATE BLOCKCHAIN**
As a rule, when states seek to achieve certain goals in a certain area of development, they apply a
special, different from the general, regulation (Podshivalov, 2018; Gromova, 2018; Ferreira and Filho,
2020; Kraljić, 2020; Nikitin and Marius, 2020; Ostanina and Titova, 2020). Introduction of a special
regulation is due to the need to achieve certain goals that cannot be achieved through general regulation.
In this regard, states use the so-called special legal regimes, which are a set of legal means aimed at
achieving a certain result. In case when it comes to the creation and implementation of the Blockchain
technology, states also apply special regulation (special and experimental regimes) that contribute to
solving these problems. They consist in a certain set of tools and mechanisms that contribute to the
achievement of the set goals and the solution of existing problems.
**2.1 Experimental legal regimes (regulatory sandboxes) for Blockchain**
The regulatory sandboxes for Blockchain services are the experimental legal regimes used by many
states today to create optimal regulation that facilitates implementation of the Blockchain technology. The
significant potential of the Blockchain technology, as well as the possible danger of its improper use,
raised the question of finding approaches to the legal regulation of this technology before modern states.
The policy ecosystem is not fully adapted to this technology, and rules and regulations would have to be
retrofitted (Gabison, 2016). In this regard, the governments of many countries have chosen an approach
aimed at creating a “breakthrough” regulation of digital innovations. Its essence is that, even in the absence
of legal regulation, business entities have the opportunity to “test” the capabilities of services and products
based on digital technologies in a real market and under state control.
For this, the state began to use regulatory sandboxes. As such, an environment controlled by the
regulator, in which entrepreneurs are given the opportunity to test the possibilities of innovative services
or products when applying certain regulatory “indulgences”, is meant. These may be the non-application
of licensing requirements, requirements for accreditation or certification to its participant. The purpose of
the regulatory sandbox is to create an environment for testing digital innovations in the absence of proper
https://doi.org/10.54934/ijlcw.v2i1.48
21
-----
legal regulation. They are used in order, first of all, to check the “viability” of a digital innovative service
(product), temporarily removing legislative barriers in the form of mandatory regulatory requirements.
The following factors are the advantages of regulatory sandboxes. First, their application enables
business entities to test innovations in a safe environment. This, in turn, helps to minimize the risks of
violating legal requirements. Second, the use of regulatory sandboxes allows regulators to examine the
“work” of new technologies “from the outside,” in a low-risk environment. This presupposes the
possibility of searching for the most appropriate ways of adapting legislation. And third, the use of
regulatory sandboxes helps to minimize the harm that can be done to consumers. This becomes possible
in connection with the provision of guarantees and additional protection methods. Thus, this mechanism
is an example of abandoning traditional regulatory approaches in favor of more flexible regulation
(Gromova and Ivanc, 2020).
Regulatory sandboxes were first introduced in 2015 as part of a government initiative to support
British digital financial innovation companies (Arner et al., 2016). This initiative enabled companies to
test the innovative products, services and business models they create in an isolated environment. The first
experience with regulatory sandboxes has been positive. The British regulatory sandbox has contributed
to the development of innovative activities of more than 500 companies, while in more than 40 of them it
has received regulatory reinforcement (Global Regulatory Sandbox Review…, 2017). The success of the
United Kingdom in creating sandboxes has led to their proliferation throughout the world. Currently,
regulatory sandboxes are used in countries such as Singapore, UAE, Australia, EU countries, China, India,
Russia, etc. The areas of application of regulatory sandboxes are usually Fintech (digital financial
technologies). This is the case, for example, in the UK, Singapore, Australia, India and the UAE (Jenik
and Lauer, 2017). Separate regulatory sandboxes in China, in turn, are created to develop not only Fintech
innovations, but the InsurTech market (digital innovations in the field of insurance). Regulatory sandboxes
in Russia can be used to test innovative services or business models in the field of medicine, transport,
education[8].
In order to be able to become a participant in the regulatory sandbox, a business entity must apply
to a state-authorized body (regulator) and provide the so-called experimental regime program. It should
8 Global Regulatory Sandbox Review: An Overview on the Impact, Challenges, and Benefits of Regulatory FinTech
Sandboxes, available at:
https://financedocbox.com/Insurance/73322297-Global-regulatory-sandbox-review.html (accessed 28.01.2021).
Regulatory Sandbox Review, available at: https://digitalchamber.org/wp-content/uploads/2017/11/Regulatory-SandboxReview_Nov-21-2017_2.pdf (accessed 28.01.2021); Regulatory Sandboxes and Financial Inclusion, available at:
https://www.cgap.org/sites/default/files/researches/documents/Working-Paper-Regulatory-Sandboxes-Oct-2017.pdf
(accessed 28.01.2023); Federal Law "On Experimental Legal Regimes for Digital Innovation” (in Russ.), available at:
https://sozd.duma.gov.ru/bill/922869-7; Global Regulatory Sandbox Review An Overview on the Impact, Challenges, and
Benefits of Regulatory FinTech Sandboxes November 21th, 2017, available at:
https://financedocbox.com/Insurance/73322297-Global-regulatory-sandbox-review.html (accessed 28.01.2023).
https://doi.org/10.54934/ijlcw.v2i1.48
22
-----
present the very innovative business model (service or product) based on digital innovation, analyze its
potential and possible risks, and identify ways to minimize them. If the submitted program is approved by
the regulator, the participant of the experimental legal regime gets the opportunity to test it in a real market
with real consumers, but with the application of certain regulatory indulgences (special regulation), within
a certain period of time. Typically, this period is from 3 to 12 months (UK, China, India, Australia) (Jenik
and Lauer, 2017). However, the legislation of certain countries, for example, the UAE, sets a period of up
to 2 years (Global Regulatory Sandbox Review…, 2017); Russia – up to 5 years (Gromova and Ivanc,
2020).
Upon the expiration of this period, the authorized body, based on the results of monitoring and
evaluating the effectiveness and efficiency of the experiment, draws conclusions about: the admissibility
of giving special regulation the properties of general regulation; the admissibility of imparting the
properties of general regulation to a special regulation in the event of amendments to the special regulation;
the inadmissibility of imparting the properties of general regulation to special regulation.
One of the development trends of this tool is the creation of regulatory sandboxes aimed at testing
services based on the Blockchain technology (Cheah et al., 2018). The World Association of Exchanges
highlighted the importance of creating regulatory sandboxes for distributed ledger technologies in
connection with the need to study the potential of this technology for the implementation of services based
on Blockchain[9]. In this regard, in some foreign countries, there are so-called “thematic” regulatory
sandboxes, the main purpose of which is to test innovative services, products and business models based
on the Blockchain technology. For example, the Government of Japan has launched a regulatory sandbox
for incubating Blockchain innovations[10].
Thailand’s regulatory sandbox is also being applied to the development of the Blockchain
technology. The examples of projects that are currently being tested in the sandbox are services using
Blockchain for letters of guarantee and cross-border funds transfers, iris identification for identity
verification, and QR code payment verification (Guide for Regulatory Sandboxes, 2018). Of particular
interest is the regulatory sandbox of the International Civil Aviation Organization (ICAO). Its goal was
the introduction of the Blockchain technology for the development of civil aviation. The ICAO Blockchain
Sandbox (2021) is a cloud-hosted network enabling different partners to work on subjects on the same
9 Exchange Body Calls for Creation Of Regulatory Sandboxes for Distributed Ledgers, available at:
https://www.finextra.com/newsarticle/29390/exchange-body-calls-for-creation-of-regulatory-sandboxes-for-distributedledgers (accessed 28.01.2023).
10 Japans blockchain sandbox is paving the way for the fintech future, available at:
https://www.forbes.com/sites/japan/2019/06/26/japans-blockchain-sandbox-is-paving-the-way-for-the-fintechfuture/?sh=5ef085832795 (accessed 28.01.2023).
https://doi.org/10.54934/ijlcw.v2i1.48
23
-----
platform. It is a blockchain infrastructure for the aviation sector. It empowers partners to create and test
services, systems, or products on a decentralized platform.
The European Union expects to launch a regulatory blockchain sandbox by 2022. The project was
initiated by the European Commission and the European Blockchain Partnership[11]. This sandbox will test
the viability of Blockchain technologies in healthcare, environment, energy and other key sectors[12]. This
means that, firstly, one of the trends in the use of regulatory sandboxes is the creation of specialized
Blockchain sandboxes. And, secondly, and importantly, Blockchain sandboxes aim to introduce the
Blockchain technology not only in the field of financial markets, but also in completely different areas.
These trends should be assessed positively, since they will contribute to the creation of an adequate and
effective legal regulation of this technology.
At the same time, the importance of correlating special regulation (granting regulatory concessions)
with fundamental human rights and consumer protection legislation should be considered. It is no
coincidence that critics of regulatory sandboxes see them as a means to circumvent consumer protection
laws. This is due to the fact that regulatory indulgences applied in testing conditions may negatively affect
the quality of services provided to consumers or otherwise violate their rights. However, as a rule, the only
special protective measure is to obtain the consent of consumers to participate in the experiment. Only a
few jurisdictions provide for liability insurance for sandbox participants and compensation in case of
violation of consumer rights[13].
In the case when it comes to testing services and products based on the use of the Blockchain
technology, it is very important to integrate the rules for participation in regulatory sandboxes with
legislation on the protection of personal data. The problem of personal data protection in the context of
special regulation applied in the framework of the regulatory sandbox is already obvious. For example, in
the Russian Federation there is an experiment on the development of artificial intelligence technologies in
Moscow. As part of this experiment, anonymized personal data of Moscow residents are transferred for
processing to artificial intelligence programs. The possibility of using anonymized personal data
significantly reduces the cost of processing them. At the same time, to date, there is no clarification in
Russian legislation about what “anonymized” personal data is and what is the mechanism of their
depersonalization (Mavrinskaya et al., 2017).
11 European Commission Launch Blockchain Regulatory Sandbox, available at: https://ec.europa.eu/digital-singlemarket/en/legal-and-regulatory-framework-blockchain (accessed 28.01.2023).
12 Legal and Regulatory Framework for Blockchain, available at: https://ec.europa.eu/digital-single-market/en/legal-andregulatory-framework-blockchain (accessed 28.01.2023).
13 Fintech regulatory sandbox, available at: https://asic.gov.au/for-business/innovation-hub/fintech-regulatory-sandbox/
(accessed 28.01.2023).
https://doi.org/10.54934/ijlcw.v2i1.48
24
-----
In the case when it comes to the use of the Blockchain technology, the issues of personal data
protection come first. After all, even outside the regulatory sandboxes, there are big problems of
convergence between the Blockchain technology and legislation on the protection of personal data. In this
regard, it is very important to work on improving the national regulatory framework and developing
international legislation in this area[14].
**2.2 Special economic zones to attract investments in development of the Blockchain technology**
There is no doubt that implementation of the Blockchain technology may require significant
investment. This is especially true for the introduction of such technology into industrial production. In
this regard, it is very important to attract investments in its creation and implementation. That is why it is
important for each country to create a favorable investment climate, including adequate legal conditions
for attracting investments into the national economy. For this, countries around the world use various
tools.
**_2.2.1 Special economic zones for Blockchain_**
These are, first of all, special economic zones. According to statistics, there are more than 5400 such
territories in 147 countries in the world (World Investment Report, 2019). Scientists note that these
territories are recognized as factors of accelerated economic growth due to their ability to influence the
intensification of trade, attracting investment, and deepening integration processes (Bost, 2019;
Veselkova, 2019). Within the boundaries of such territories, representatives of the private sector are
provided with tax and other preferences in order to stimulate investment and other entrepreneurial
activities.
The most famous example of the successful creation and operation of special economic zones is
undoubtedly China. The rise of the Chinese economy, associated, among other things, with the creation
of special economic zones, is called the “Chinese economic miracle”. In order to attract private investors,
residents were provided with various preferences, including inexpensive land, tax and customs benefits,
the possibility of repatriating profits and capital investments, exemption from export tax and a limited
license to sell goods in the domestic market [11] Creation of innovative products is the main goal of the
establishment and High-Tech Industrial Development Zones. Today, 54 such zones are successfully
operating in China. Their creation began in 1980 under the Program of the Ministry of Science and
Technology of China. The main goal of the Program was to use the technological capacity and resources
of research institutes, universities, and large and medium enterprises to develop new and high-tech
products and to expedite the commercialization of research and development [36].
14 Blockchain: Playing in the regulatory sandbox, 07 September 2016
https://www.finextra.com/blogposting/13055/blockchain-playing-in-the-regulatory-sandbox (accessed 28.01.2023).
https://doi.org/10.54934/ijlcw.v2i1.48
25
-----
Today, some countries are considering the possibility of creating and implementing the Blockchain
technology within the boundaries of special, free economic zones. For example, the Central Committee of
the Chinese Communist Party recently announced that research into the creation and implementation of
the Blockchain technology for the digital currency market will be supported within the Shenzhen special
economic zone. The Chinese government intends to use the Shenzhen special economic zone as a pilot
demonstration zone for supporting innovative applications such as digital money research and mobile
payment in Shenzhen [40].
Special conditions for the implementation of the Blockchain technology are envisaged in Georgia.
Thus, the Gldani Free Industrial Zone guaranteed the UK-based company access to electricity at
discounted rates for a brand-new, power-thirsty 40-megawatt datacenter devoted to the mining of
cryptocurrencies. Other special economic zones providing tech-companies with the special regulatory
environment they need to thrive are springing up across the globe, particularly in countries that have
embraced blockchain and cryptocurrencies. The Cagayan special economic zone in the Philippines has
licensed as many as 37 crypto exchanges since receiving a special mandate to develop the “Crypto Valley
of Asia” in May 2018.
Note that in the Russian Federation it is also planned to create a Blockchain cluster within the free
economic zone of the Republic of Crimea and the federal city of Sevastopol. The purpose of creating such
a cluster will be to attract investment in the implementation of Blockchain projects. For the development
of cryptocurrencies and Blockchain projects, they plan to use the Russian part of the territory of the
Bolshoi Ussuriysky Island (2019). It is planned to create a special administrative region with preferential
conditions for international companies planning to operate in this area[15]. It is believed that the creation of
clusters within the boundaries of special, special and free economic zones will contribute to the
development of the Blockchain technology. The operation of geographically related companies carrying
out complementary activities will have a positive effect on creation and implementation of this technology,
including in industrial production. Being a member of a cluster is strongly believed to enhance local
productivity and competitiveness. No wonder that policymakers are concerned to create, establish,
promote or just label existing interfirm networks or agglomerations of firms or industries as a cluster [23].
**2.3 Public-private partnership for Blockchain projects**
Public-private partnership (hereinafter – PPP) is an important tool that helps to attract investment in
the development of socially significant projects. This tool is actively used all over the world. There is even
a term “innovative public-private partnership” (Innovative PPP) for the creation and implementation of
15 Why Blockchain Developers are being given the VIP treatment, available at: https://www.fdiintelligence.com/article/75453
(accessed 28.01.2023).
https://doi.org/10.54934/ijlcw.v2i1.48
26
-----
digital innovations. The European Union is actively using the mechanisms of public-private partnership
for the development of infrastructure projects. EU legislation also provides for the possibility of creating
innovations such as robotics and supercomputers under contractual forms of innovative PPP. So, for
example, creation and development of robotics and artificial intelligence is one of the key areas for
development of the digital economy, it is actively taking place in foreign countries precisely on the basis
of PPP (Cyman et al., 2020). In the European Union, in particular, research in the field of robotics received
the largest funding under the innovation program Horizon 2020 based on PPP projects – about €190
million. Under another European robotics development program, SPARC, EU states are investing €700
million, and the private sector – €2.1 billion[16] in the creation of industrial robotics. In addition, the
development of another breakthrough direction of the digital industry – supercomputers (high performance
computing) in the EU countries is also carried out on the basis of PPP[17].
Note that in the United States there is a separate research development program in the field of
implementing the Blockchain technology in the electoral process. This program is based on the principles
of public-private partnership. It was initiated by the Government Blockchain Association in the USA. The
program GBA Public-Private Partnership (PPP) objectives include researching the technological,
regulatory and political issues associated with Blockchain and voting. The second phase of the program
includes developing the requirements, implementing, and deploying Blockchain-based voting solutions[18].
Another PPP project in the field of creation and implementation of the Blockchain technology has
also been launched in the United States. The Security and Software Engineering Research Center at
Georgetown University (S2ERC). S2ERC is a great example of a public-private partnership that seeks to
merge interest of the federal government and commercial innovation[19]. Another example of PPP in the
field of creation and implementation of Blockchain projects is the infrastructure project of the US
government and DeFi to create toll roads, payments for the use of which are saved under the Blockchain
program[20]. The Chinese government is also actively developing public-private partnerships in creation
and implementation of Blockchain technologies. So, one of these projects was the creation of a fund
(Xiong'An Global Blockchain Innovation Fund) for the development of Blockchain startups. At the same
16 Is Europe investing in robotics? (In Russ.), available at: http://www.robogeek.ru/analitika/evropa-vkladyvaet-dengi-vrobototehniku (accessed 28.01.2023).
17 Contractual forms of PPP for high performance computer, available at: https://ec.europa.eu/digital-single-market/en/highperformance-computing-contractual-public-private-partnership-hpc-cppp (accessed 28.01.2023).
18 Government Blockchain Association (GBA), available at: https://www.gbaglobal.org/blockchain-voting-public-privatepartnership-ppp-forming-now/ (accessed 28.01.2023).
19 Public Private Partnerships for Innovation Blockchain, available at: https://federalnewsnetwork.com/federal-techtalk/2018/01/public-private-partnerships-innovation-blockchain/ (accessed 28.01.2023).
20 DeFi Blockchain contract, available at: https://www.ledgerinsights.com/us-space-force-awards-blockchain-contract-toxage-security/ (accessed 28.01.2023).
https://doi.org/10.54934/ijlcw.v2i1.48
27
-----
time, the state's share was 25%, the remaining $ 1.2 billion are private investments of Tunlan Investment
Company[21].
With regard to the Russian Federation, it should be noted that in 2018 alone, the Federal Law "On
Public-Private, Municipal-Private Partnership" (2015) (hereinafter referred to as the Law on PPP) was
amended to allow the creation of information technology objects within the framework of PPP. The
introduction of these changes should be assessed positively. Since in the previous edition, the creation of
information technology objects was not allowed within the framework of PPP. Meanwhile, the fact that
only information technology objects can be created on the basis of PPP limits the possibility of
implementing PPP in the field of innovation. According to Art. 2 of the Federal Law "On Information,
Information Technologies and Information Protection" dated July 27, 2006 No. 149-FZ (2006),
“information technologies – processes, methods of searching, collecting, storing, processing, providing,
disseminating information and ways of implementing such processes and methods”. It seems that the term
“information technology objects” chosen by the legislator significantly limits the potential of PPPs in the
field of creating innovations. Note that within the framework of the Federal Program “Digital Economy”
it is proposed to develop a number of “end-to-end digital technologies: big data; neurotechnology and
artificial intelligence; distributed ledger systems; quantum technologies; new production technologies;
industrial internet; robotics and sensorics components; wireless technology; technologies of virtual and
augmented reality”. This list is not exhaustive and can be expanded as new technologies appear and
develop.
At the same time, if such digital technologies as artificial intelligence, neurotechnologies, wireless
communication technologies, virtual reality can be considered information technologies, and, accordingly,
created within the framework of PPP, then referring to information technology objects a whole range of
end-to-end digital technologies, such as new production technology, as well as the components of robotics,
is highly controversial. This, in turn, may affect the development opportunities within the framework of
PPP of the Blockchain technology itself. The fact is that a project implemented within the framework of
a PPP may be associated with not one, but several digital innovations. And in the event that one of them
does not “fall” under the regulation of this act, then the creation and implementation of the rest may be
questionable. It seems that the current legislation on PPP and its legal forms should be amended to allow
the creation of digital technologies within the framework of PPP. Such changes, it seems, would be more
conducive to the development of innovation and the digital economy on the basis of PPP, and, thereby,
would improve the country's competitiveness in the digital technology market (Ertz and Boily, 2019).
21 China invests $ 16 billion to develop Blockchain PPP, available at: https://cryptor.net/news/kitay-investiruet-v-blokcheyntehnologii-16-mlrd-v-ramkah-gosudarstvenno-chastnogo-partnerstva (accessed 28.01.2023).
https://doi.org/10.54934/ijlcw.v2i1.48
28
-----
**2.4 Other tools to stimulate Blockchain**
_2.4.1 Standardization for Blockchain_
In the context of intensive digitalization, the most important component of state innovation policy
today is the creation of standards in the field of digital technologies, consolidating in them the technical
aspects of the functioning of such technologies [8]. Standardization in the field of the Blockchain
technology will allow developing a universal terminology associated with this technology; will ensure the
safe use of technologies based on artificial intelligence. Moreover, the standardization of this technology
will increase the level of its interoperability with other digital technologies, which, in turn, will have a
positive impact on the development of scientific and technological progress[22]. At the same time, the
development and adoption of “ineffective” standards can constrain the development of digital
technologies. In this regard, global international cooperation and coordination on the development of
standards in the field of distributed ledger technology will be critical for the successful standardization of
digital technologies in general, ensuring fair competition, removing trade barriers and the flourishing of
innovation[23].
Today, there is an intensive development of standards in the field of the Blockchain technology.
International organizations for standardization, as well as authorized bodies of many foreign countries,
are actively involved in this process. The International Organization for Standardization (ISO) established
an international technical committee for the standardization of Blockchain and Distributed Ledger
Technologies in 2016 (ISO/TC 307 Blockchain and Distributed Ledger Technologies)[24]. The committee
includes five working groups: on Blockchain architecture and ontology, scope, security and privacy,
identification and smart contracts. The committee included 35 states, led by Australia. In March 2017, the
first Blockchain standardization roadmap was published. Standards in the field of this technology are
developed by standardization bodies of certain foreign countries. For example, in 2020, a focus group on
the application of distributed ledger technology (FG DLT), established by the International
Telecommunication Union (ITU-T) Standardization Sector, completed its work in Geneva. The Institute
of Electrical and Electronics Engineers (IEEE) is working on a series of standards for general purpose
frameworks and architectures, interoperability, core technology components, and Bockchain industry
specifications (P2418) (Blockchain standards, 2020).
22 Artificial Intelligence’s standardization helps create innovation friendly framework conditions for the technology of the
future, available at: https://www.din.de/blob/306690/f0eb72ae529d8a352e0b0923c67b6156/position-paper-artificialintelligence-english--data.pdf (accessed 28.01.2023).
23 U.S. leadership in AI: A Plan for Federal Engagement in Developing Technical Standards and Related Tools, available at:
https://www.nist.gov/sites/default/files/documents/2019/08/10/ai_standards_fedengagement_plan_9aug2019.pdf (accessed
28.01.2023).
24 Blockchain standards, available at: https://blockchain.ieee.org/standards (accessed 28.01.2023).
https://doi.org/10.54934/ijlcw.v2i1.48
29
-----
The National Institute for Standardization (NIST), within the framework of the doctrine of US
leadership in the field of digital technologies, approved the "Plan for state involvement in the development
of technical standards and related tools" in 2019 (U.S. leadership in AI…, 2019) and is actively working
to create an international standard for the Blockchain technology. So, in the fall of 2020, NIST posted the
Draft Standard NISTIR 8301, Blockchain Networks: Token Design and Management Overview, which
provides a high-level technical overview and conceptual framework of token designs and management
methods. The Draft Standard for the Application of the Blockchain Technology in Industry is also
important. The Blockchain Project for Industrial Applications Community of Interest is providing
guidelines to create a (better) synergy between end users, research community, and solution providers to
reduce complexity, cost, and delay of adoption of Blockchain technologies[25].
In modern conditions, Russia also does not stay away from world trends. To date, it has adopted two
strategically important documents in the field of technical regulation of digital technologies. One of these
is the Passport of the Digital Economy Program[26]. It provides for the development of a federal project
“Normative regulation of the digital environment”, including with a view to improving standardization
mechanisms in the field of digital technologies. In turn, in the Action Plan for “Normative Regulation” of
the “Digital Economy of the Russian Federation” program dated December 18, 2017 a set of measures is
envisaged to improve the mechanisms for standardizing digital technologies to eliminate barriers to their
use. Among the activities of this plan is amending the current legislation in order to simplify the procedures
for developing standardization documents, shorten the time for their development, accelerate the adoption
of national standardization documents based on or taking into account the standards of the most
authoritative associations and organizations.
The measures proposed in the Plan cannot be assessed unambiguously. So, on the one hand, the
establishment of the possibility of adopting standards based on or taking into account the standards of the
most authoritative associations and organizations, will contribute to a better “filling” of such documents
with technical requirements already tested in practice. On the other hand, simplifying the standardization
procedure and shortening its time frame will not in itself contribute to the development and adoption of
“working” standards in the field of the Blockchain technology. In this case, it is necessary to revise not
the quantitative, but the qualitative aspects of standardization.
25 NIST Blockchain Standardization, available at: https://www.nist.gov/blockchain
26 Russia's National Program For Digital Economy, available at: https://ac.gov.ru/en/projects/project/digital-economyprogram-implementation-42 (accessed 28.01.2023); "Plan of measures for the direction "Regulatory regulation" of the
program "Digital economy of the Russian Federation" 2017 г., available at:
http://static.government.ru/media/files/P7L0vHUjwVJPlNcHrMZQqEEeVqXACwXR.pdf (accessed 28.01.2023).
https://doi.org/10.54934/ijlcw.v2i1.48
30
-----
The need for standardization in the field of digital technologies predetermined the creation of a
technical committee for Blockchain standardization. The committee was named “Distributed Ledger
Technologies and Blockchain Hardware and Software”. Its main task is to increase the efficiency of work
on the development of the domestic regulatory and technical base in the field of distributed ledger and
Blockchain technologies. Within the framework of TC 26, methodological recommendations on
terminology were issued – MR 26.4.001-2018 “Terms and definitions in the field of chain data recording
technologies (Blockchain) and distributed ledgers” (2018).
One of the strategically important directions of the committee's work is participation in the
international standardization process on behalf of the Russian Federation, including the consideration of
application of international standards in the field of distributed ledger technology at the national level.
This is important, since the participation of this Russian committee in international standardization will
contribute to a greater extent to ensuring national interests than the usual adherence to the International
Standard, developed without the participation of representatives of the country. According to experts,
today it is obvious that there is a need to move from passive assimilation of foreign experience to the stage
of active construction of domestic developments in the field of standardization, which should significantly
strengthen Russia's position in the field of high technologies[27].
It seems that when preparing standards in the field of the Blockchain technology, a number of
important points should be taken into account. First, it is imperative for the international community to
continue to work together to standardize digital technologies. Second, at the national level, it is worth
actively involving the private sector in the standardization process; but, at the same time, not only large
and medium-sized businesses, but also small businesses. It should be remembered that small businesses
can also be actively involved in the development and application of the Blockchain technology, and, as a
rule, are more “mobile” in these matters. Third, it is recommended to involve leading research universities
and research organizations in the digital standards development process. And finally, fourthly, in order to
stimulate the participation of these entities in the development of standards in the field of the Blockchain
technology, it is also necessary to create a mechanism to compensate the costs of the latter for participation
in the development of international and national standards.
_2.4.2. Governmental support for Blockchain companies due to COVID-2019._
In the pandemic, programs to support business entities engaged in the creation and implementation
of digital technologies, including Blockchain technologies, are of particular importance. At the moment,
there are practically no government support programs for Blockchain companies. The US is an exception.
27 Standardization of the Digital Economy, available at: http://www.connect-wit.ru/standartizatsiya-tsifrovoj-ekonomikirossii.html (accessed 28.01.2023).
https://doi.org/10.54934/ijlcw.v2i1.48
31
-----
So, U.S. Small Business Administration (SBA) launched the Paycheck Protection Program (PPP). Under
this Program, more than 75 companies in the Blockchain industry received government loans totaling $
30 million. The PPP Program was created by the Trump administration during the COVID-19 outbreak to
help businesses pay their employees during the ongoing economic crisis[28] (Market Wrap…, 2021).
Note that in other countries the situation is diametrically opposite. So, for example, in the Russian
Federation, the Blockchain projects industry has not yet received any financial support from the state
(subsidies, compensation for lost rental income) (Blockchain has officially become…, 2020). According
to experts, at the moment, the infrastructure for supporting distributed ledger systems is insufficient to ensure
continuous improvement of relevant solutions. This is especially true today, in the context of the financial and
economic crisis, the imposed trade restrictions on certain technological components or ready-made solutions, the
inaccessibility of foreign capital markets and the lack of opportunities for exchanging experience with foreign
experts, as well as insufficient demand for solutions in the domestic market, provided that foreign markets are not
accessible (Blockchain will bring…, 2019)[29].
In this regard, the author believes that in order to support the developers of Blockchain services that
can find their application in industry, a number of government support measures should also be developed,
similar to the United States. If talk about Russian legislation, then it seems possible to give the opportunity
to such developers, small and medium-sized businesses, the right to receive financial, property and other
support measures in accordance with the legislation on small and medium-sized businesses.
**3. RECOMMENDATIONS**
The recommendations for improving the special and experimental modes are as follows.
First, with regard to experimental legal regimes (regulatory sandboxes), it is important to integrate
the rules of participation in Blockchain sandboxes with legislation on the protection of personal data, as
well as on the protection of consumer rights. In this regard, it is very important to work on improving the
national regulatory framework and the development of international legislation in this area, since now
many states are considering the possibility of creating interstate regulatory sandboxes in the field of testing
Blockchain services.
Second, it is very important to improve the Blockchain technology standardization process, again,
both at the international and national levels. At the state level, it is important to involve small and medium
sized businesses and other business representatives, as well as leading research universities and research
28 Market Wrap: Bitcoin Hovers Around $34.2K While Options Traders Pay Up for Possible ETH Upside, available at:
https://www.coindesk.com/market-wrap-bitcoin-hovers-35k-options-traders-pay-eth-upside (accessed 28.01.2023).
29 Blockchain will bring 16 bill to the Russian Economy, available at: https://www.cnews.ru/news/top/2019-0717_blokchejn_prineset_rossijskoj_ekonomike_16_trillionov (accessed 28.01.2023).
https://doi.org/10.54934/ijlcw.v2i1.48
32
-----
organizations in the development of the standard. In order to stimulate their participation in the
development of standards for the Blockchain technology, it is also necessary to create a mechanism to
compensate the costs of the latter for participation in the development of international and national
standards. If talk about the international level, then the joint efforts of states to create universal,
understandable and, most importantly, working standards are also very important.
Third, attention should be paid to improving legislation on special regimes aimed at attracting
private investment in the creation of Blockchain technologies for smart industry. So, it is necessary to
legally provide the possibility of creating Blockchain clusters within such territories. In addition, it seems
necessary to develop public-private partnerships in creation and implementation of the Blockchain
technology. This contributed to the development of innovation and the digital economy on the basis of
PPPs, and, thereby, made it possible to increase the competitiveness of each country in the digital
technology market. To support the developers of Blockchain services that can find their application in
industry, a number of government support measures should also be developed, including financial,
property and other.
**5. CONCLUSIONS**
Thus, for the legal support of the implementation of the Blockchain technology, including in
industrial production, states apply special and experimental regimes. As a rule, such regimes are generally
universal for most states. In this regard, it can be concluded that for successful implementation of the
Blockchain technology in industrial production, states should use certain tools and mechanisms that make
up the content of these modes. This will allow not only determining the national legal framework for
implementation of the Blockchain technology. It will also allow us to work together to create international
regulation for implementation of the Blockchain technology in industrial production, develop cooperation
in this area, share and develop the best world practices. The latter is especially important given the fact
that the instruments and mechanisms used by states that make up the content of these regimes are not
without drawbacks. These shortcomings, in turn, are due to the lack of quality legal regulation of such
instruments. It also requires a concerted effort to eliminate them, both nationally and internationally.
The combined use of these special and experimental regimes (special and experimental regulation)
will increase the attractiveness of the state jurisdiction by creating more adequate legal conditions for the
implementation of activities by investors and developers of services and products based on the Blockchain
technology. The conclusions reached by the author can be used in the development of international and
national legal foundations for the implementation of the Blockchain technology in industrial production.
In addition, the results of the study can be used as a basis for further scientific research in the field of legal
regulation of the creation and implementation of the Blockchain technology and other digital technologies.
https://doi.org/10.54934/ijlcw.v2i1.48
33
-----
**REFERENCES**
[1] Alia, O., Allyb, M., Clutterbuck, Y. (2020). The state of play of blockchain technology in the
financial services sector: A systematic literature review. _International Journal of Information_
_Management, 54, 102199._
[2] Arner, D., Barberis, J., Buckley, R.P. (2016). The Evolution of Fintech: A New Post-Crisis
Paradigm? Georgetown Journal of International Law, 47, 1271-1281.
[3] Bouraga, S. (2021). A taxonomy of blockchain consensus protocols: A survey and classification
framework. Expert Systems with Applications, 168, article number 114384.
[4] Balasubramaniana, S., Shuklab, V., Singh, J., Islamc, N., Salouma, R. (2021). A readiness
assessment framework for Blockchain adoption: A healthcare case study. _Technological_
_Forecasting and Social Change, 165, article number 120536._
[5] Borg, J.F., Schembri, T. (2019). The regulation of blockchain technology. In: J. Dewey (Ed.),
Blockchain & Cryptocurrency Regulation (pp. 188-192).
[6] Bost, F. (2019). Special economic zones: methodological issues and definition. _Transnational_
_Corporations, 26(2), 141-153._
[7] Campanile, L., MauroIacono, F., Marulli, M. (2021). Designing a GDPR compliant blockchain
based IoV distributed information tracking system. Information Processing & Management, 3(58),
article number 102511.
[8] Cyman D., Gromova E., Juchnevicius E. (2021). Regulation of Artificial Intelligence in BRICS
and the European Union. BRICS Law Journal, 8(1), Pp. 86-115.
[9] De Filippi, P., Reijersd, V., Morshed, M. (2020). Blockchain as a confidence machine: The
problem of trust & challenges of governance. Technology in Society, 62, article number 101284.
[10] Dong, Zh., Fengji, L., Gaoqi, L. (2018). Blockchain: a secure, decentralized, trusted cyber
infrastructure solution for future energy systems. _Journal of Modern Power Systems and Clean_
_Energy, 5(6), 958-967._
[11] Enright, M.J., Scott, E., Chung, K. (2005). Regional Powerhouse: The Greater Pearl River Delta
_and the Rise of China. Singapore: John Wiley & Sons (Asia), 404 p._
[12] Ertz, M., Boily, E. (2019). The rise of the digital economy: Thoughts on blockchain technology
and cryptocurrencies for the collaborative economy. International Journal of Innovation Studies,
3(4), 84-93.
[13] Elisavetsky, A., Marun, M.V. (2020). Technologies applied to conflict resolution. Understanding
it for the efficiency of ODR and its projection in Latin America [La tecnología aplicada a la
resolución de conflictos. Su comprensión para la eficiencia de las ODR y para su proyección en
Latinoamérica]. Revista Brasileira de Alternative Dispute Resolution, 3(2), 51-69.
[14] Fan, K., Wang, Sh., Ren, Y. (2018). MedBlock: Efficient and Secure Medical Data Sharing Via
Blockchain. Journal of Medical Systems, 8(42), 136-145.
[15] Ferreira, D.B., Filho, E.A. (2020). Action for Annulment of Arbitral Award: a doctrinaire and
empirical analysis of the jurisprudence of the courts of Santa Catarina, Rio de Janeiro and São
Paulo between 2015 and 2019 [Anulatória de sentença arbitral: uma análise doutrinária e empírica
da jurisprudência dos tribunais dos estados de Santa Catarina, Rio de Janeiro e São Paulo entre
2015 e 2019]. Revista Brasileira de Alternative Dispute Resolution, 3(2), 195-214.
https://doi.org/10.54934/ijlcw.v2i1.48
34
-----
[16] Gabison, G. (2016). Policy Considerations for the Blockchain Technology Public and Private
Applications. Science and Technology Law Review, 19(3), 327-350.
[17] Gromova, E., Ivanc, T. (2020). Regulatory Sandboxes (Experimental Legal Regimes) for Digital
Innovations in BRICS. BRICS Law Journal, 7(2), 10-36. https://doi.org/10.21684/2412-23432020-7-2-10-36
[18] Gromova, E. (2018). The Free Economic Zone of the Republic of Crimea and the Federal City of
Sevastopol. Russian Law Journal, 6(3), 79-99. https://doi.org/10.17589/2309-8678-2018-6-3-7999
[19] Ivanc T., Rijavec V., Kerestes T. (2016). Theoretical background of using information technology
_in evidence taking. Dimensions of evidence in European civil procedure. UK: Wolters Kluwer._
[20] Jani, N., Panda, P. (2019). Blockchain Technology: A Study of Investment and Applications
Introduction, In NCUICM-2019 International Conference on Management Winning in a VUCA
_World Edition: First Impression. Excel India Publishers._
[21] Joppen, R., Lipsmeier, A., Tewes, A., Dumitrescu, R. (2019). Evaluation of investments in the
digitalization of a production. Procedia CIRP, 81, 411-416.
[22] [Kraljić, S. (2020). New family code and the dejudicialization of divorce in Slovenia. Balkan Social](https://www.scopus.com/authid/detail.uri?authorId=36098958100)
_Science Review, 15, 157-176._
[23] Lehmann, E., Menter, M. (2018). Public cluster policy and performance. _The Journal of_
_Technology Transfer, 43(2), 558-592. https://doi.org/10.1007/s10961-017-9626-4_
[24] Mavrinskaya, T.V., Loshkaryov, A.V., Churakova, E.N. (2017). Depersonalization of Personal
Data and “Big Data” Technology (BIG DATA).
[25] Mohamad, A.M., Zaiton, H., Mohd Bahrin, O. (2017). Balancing Open Justice and Privacy Rights
in Adopting ICT in the Malaysian Courts. Advanced Science Letters, 23(8), 7996-8000.
[26] Nakamoto, S. (2009). Bitcoin: A Peer-to-Peer Electronic Cash System.
https://bitcoin.org/bitcoin.pdf.
[27] Nikitin, E., Marius, M.C. (2020). Unified Digital Law Enforcement Environment – Necessity and
Prospects for Creation in the “BRICS Countries”. BRICS Law Journal, 2(7), 66-93.
[28] Ostanina, E., Titova, E. (2020). The Protection of Consumer Rights in the Digital Economy
Conditions – the Experience of the BRICS Countries. BRICS Law Journal, 7(2), 118-147.
[29] Podshivalov, T. (2018). Protection of Property Rights Based on the Doctrine of Piercing the
Corporate Veil in the Russian Case Law. _Russian_ _Law_ _Journal,_ 6(2), 39-72.
https://doi.org/10.17589/2309-8678-2018-6-2-39-72
[30] Jenik, I., Lauer, K. (2017). Regulatory Sandboxes and Financial Inclusion.
[31] Roman-Belmonte, J.M., De la Corte-Rodriguez, H., Carlos Rodriguez-Merchan, E. (2018). How
blockchain technology can change medicine. Postgraduate Medicine, 130(4), 420-427.
[32] Swan, M. (2015). Blockchain: Blueprint for a New Economy. Newton: O'Reilly Media, Inc.
[33] Spevakov, A.G., Spevakova, S.V., Primenko, D.V. Method of data depersonalization in protected
automated information systems. _Radio Electronics, Computer Science, Control, 1, 162-168._
https://doi.org/10.15588/1607-3274-2020-1-16
[34] Sultan, K., Ruhi, U., Lakhani, R. (2018). Conceptualizing Blockchains: Characteristics &
Applications. In 11th IADIS International Conference Information Systems (pp. 49-57).
https://doi.org/10.54934/ijlcw.v2i1.48
35
-----
[35] Tatara, U., Gokceb Y., Nussbauma, B. (2020). Law versus technology: Blockchain, GDPR, and
tough tradeoffs. Computer Law & Security Review, 38, article number 10545.
[36] Zeng, Zh. (2012). _China's_ _Special_ _Economic_ _Zones and Industrial_
_Clusters. Success and Challenges:_ _Working_ _Paper._
[https://www.lincolninst.edu/sites/default/files/pubfiles/2261_1600_Zeng_WP13DZ1.pdf.](https://www.lincolninst.edu/sites/default/files/pubfiles/2261_1600_Zeng_WP13DZ1.pdf)
[37] Xu, Zh., Zhang, J., Song, Zh., Liu, Y. (2021). A scheme for intelligent blockchain-based
manufacturing industry supply chain management. _Computing, 1._ [https://doi.org/10.1007/s00607-](https://doi.org/10.1007/s00607-020-00880-z)
[020-00880-z](https://doi.org/10.1007/s00607-020-00880-z)
[38] Cheah, S., Ho, P., Pattalachinti, S. (2018). Blockchain Industries, Regulations and Policies in
Singapore. Asian Research Policy, 9(2), 83-98.
[39] Veselkova, E. (2019). Goals and Methods of Using Special Economic Zones in International
Economic Cooperation. Norwegian Journal of Development of the International Science, 37, 53-56.
https://doi.org/10.54934/ijlcw.v2i1.48
36
-----
**ABOUT THE AUTHORS**
**[Daniel Brantes](https://adrbc.com/team/daniel-brantes-ph-d/)** **[Ferreira – Ph.D., Senior Researcher, National Research South Ural](https://adrbc.com/team/daniel-brantes-ph-d/)**
State University (Russia), Professor, AMBRA University (USA), CEO, Brazilian
Centre for Mediation and Arbitration (Rio de Janeiro, Brazil)
[Email: daniel.brantes@gmail.com](mailto:daniel.brantes@gmail.com)
[ORCID ID: 0000-0003-0504-1154](https://orcid.org/0000-0003-0504-1154)
[Web of Science Researcher ID: AEN-4058-2022](http://www.webofscience.com/wos/author/record/AEN-4058-2022)
[Scopus Author ID: 56555993000](https://www.scopus.com/authid/detail.uri?authorId=56555993000)
[GoogleScholarID: 7maCMCUAAAAJ](https://scholar.google.com/citations?hl=en&user=7maCMCUAAAAJ)
**[Elizaveta Alexandrovna Gromova – Ph.D. (Law), Associate Professor, Deputy](https://www.susu.ru/ru/employee/gromova-elizaveta-aleksandrovna)**
Director of the Law Institute on international activity, Associate Professor,
Department of Entrepreneurial, Competition and Environmental Law, South Ural
State University (national research university) (Chelyabinsk, Russian Federation)
Address: 454090 Lenina Prospekt, 78, Chelyabinsk, Russian Federation
[Email: gromovaea@susu.ru](mailto:gromovaea@susu.ru)
[ORCID ID: 0000-0001-6655-8953](https://orcid.org/0000-0001-6655-8953)
[Web of Science Researcher ID: AAO-8876-2020](http://www.webofscience.com/wos/author/record/AAO-8876-2020)
[Scopus Author ID: 57208846603](https://www.scopus.com/authid/detail.uri?authorId=57208846603)
[GoogleScholarID: fDz6FkUAAAAJ&hl](https://scholar.google.com/citations?user=fDz6FkUAAAAJ&hl)
**ABOUT THIS ARTICLE**
**Conflict of interests: Authors declare no conflicting interests.**
.
https://doi.org/10.54934/ijlcw.v2i1.48
37
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.54934/ijlcw.v2i1.48?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.54934/ijlcw.v2i1.48, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "GOLD",
"url": "https://ijlcw.emnuvens.com.br/revista/article/download/48/18"
}
| 2,023
|
[
"JournalArticle"
] | true
| 2023-06-15T00:00:00
|
[] | 15,819
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Engineering",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/020c3ce8434387e9f141198493a2f68c2e25ca2c
|
[
"Computer Science"
] | 0.896482
|
An Adaptive Lightweight Security Framework Suited for IoT
|
020c3ce8434387e9f141198493a2f68c2e25ca2c
|
Internet of Things
|
[
{
"authorId": "35009042",
"name": "M. Domb"
}
] |
{
"alternate_issns": [
"2199-1073"
],
"alternate_names": [
"Internet Thing"
],
"alternate_urls": [
"https://www.sciencedirect.com/journal/internet-of-things",
"http://www.springer.com/series/11636"
],
"id": "2989732e-2668-4b47-9c29-326646a60273",
"issn": "2542-6605",
"name": "Internet of Things",
"type": null,
"url": "https://www.journals.elsevier.com/internet-of-things"
}
|
Standard security systems are widely implemented in the industry. These systems con- sume considerable computational resources. Devices in the Internet of Things [IoT] are very limited with processing capacity, memory and storage. Therefore, existing security systems are not applicable for IoT. To cope with it, we propose downsizing of existing security processes. In this chapter, we describe three areas, where we reduce the required storage space and processing power. The first is the classification process required for ongoing anomaly detection, whereby values accepted or generated by a sensor are clas- sified as valid or abnormal. We collect historic data and analyze it using machine learn ing techniques to draw a contour, where all streaming values are expected to fall within the contour space. Hence, the detailed collected data from the sensors are no longer required for real-time anomaly detection. The second area involves the implementation of the Random Forest algorithm to apply distributed and parallel processing for anomaly discovery. The third area is downsizing cryptography calculations, to fit IoT limitations without compromising security. For each area, we present experimental results support-ing our approach and implementation. as follows: We begin with an introduction followed by the relevant literature review. We then discuss rules extraction using machine learning tech -niques. We present random forest as the most suitable ML for IoT. We proceed with various improvements, utilizing RF and IoT attributes. We then outline an experiment that executes RF building and its corresponding classifications using 15 different configurations, each based on a unique combination of the number of processors and the forest size.
|
## We are IntechOpen, the world’s leading publisher of Open Access books Built by scientists, for scientists
# 6,700
Open access books available
# 180,000 195M
International authors and editors
Our authors are among the
Downloads
# 12.2%
# 154 TOP 1%
Countries delivered to most cited scientists Contributors from top 500 universities
Selection of our books indexed in the Book Citation Index
in Web of Science™ Core Collection (BKCI)
### Interested in publishing with us? Contact book.department@intechopen.com
Numbers displayed above are based on latest data collected.
For more information visit www.intechopen.com
-----
##### Chapter 2
#### An Adaptive Lightweight Security Framework Suited for IoT
##### Menachem DombMenachem Domb
Additional information is available at the end of the chapterAdditional information is available at the end of the chapter
http://dx.doi.org/10.5772/intechopen.73712
**Abstract**
Standard security systems are widely implemented in the industry. These systems consume considerable computational resources. Devices in the Internet of Things [IoT] are
very limited with processing capacity, memory and storage. Therefore, existing security
systems are not applicable for IoT. To cope with it, we propose downsizing of existing
security processes. In this chapter, we describe three areas, where we reduce the required
storage space and processing power. The first is the classification process required for
ongoing anomaly detection, whereby values accepted or generated by a sensor are classified as valid or abnormal. We collect historic data and analyze it using machine learning techniques to draw a contour, where all streaming values are expected to fall within
the contour space. Hence, the detailed collected data from the sensors are no longer
required for real-time anomaly detection. The second area involves the implementation
of the Random Forest algorithm to apply distributed and parallel processing for anomaly
discovery. The third area is downsizing cryptography calculations, to fit IoT limitations
without compromising security. For each area, we present experimental results supporting our approach and implementation.
**Keywords: IoT, anomaly detection, entropy, machine learning, random forest,**
cryptography, RSA
##### 1. Introduction
The area of the Internet of Things [IoT] is rapidly growing, raising severe security concerns to
the entire network. Due to its high traffic volume and real-time operation, a security framework is essential. The system should timely predict possible attacks and react accordingly.
Standard security systems are widely implemented in the industry. These systems consume
© 2016 The Author(s). Licensee InTech. This chapter is distributed under the terms of the Creative Commons
© 2018 The Author(s). Licensee IntechOpen. This chapter is distributed under the terms of the Creative
Attribution License (http://creativecommons.org/licenses/by/3.0), which permits unrestricted use,
Commons Attribution License (http://creativecommons.org/licenses/by/3.0), which permits unrestricted use,
distribution, and reproduction in any medium, provided the original work is properly cited.
distribution, and reproduction in any medium, provided the original work is properly cited.
|its unrestricted u|Col2|se,|
|---|---|---|
||||
|d.|||
-----
32 Internet of Things - Technology, Applications and Standardization
considerable computational resources and cannot operate in IoT devices (i.e., sensors) due
to their very limited memory and computation power. To cope with these limitations, two
alternatives come to mind, i.e., the development of novel security measures tailored to IoT
[1] or downsizing existing security processes to enable properly operation in IoT devices. We
apply the latter option as it is highly recommended to use proven algorithms, which have
been extensively analyzed and tested, while new algorithms exposes the user to vulnerability.
We introduce lightweight versions of several known security processes. We analyze each
relevant process and its corresponding limitations, and then we divide each complex and
large process into a collection of smaller processes. These small processes are distributed and
executed by sensors connected to the same network, based on its available capacity. Once all
small processes are completed, we collect the partial results and input them into a complementary process that integrates the partial results to compose the desired result. The final
result is the same as if the original process was generated. In this chapter, we describe three
areas, where we minimize the required storage space and processing power. The first is the
classification process required for ongoing anomaly detection, whereby values accepted or
generated by a sensor are classified as valid or abnormal. We collect historic data and analyze
it using machine learning techniques to draw a contour, and all streaming values are expected
to fall within the contour space. The detailed collected data are no longer required, thereby
considerably reducing the storage space. The second area involves the implementation of the
Random Forest algorithm to apply distributed and parallel processing for anomaly discovery,
resulting in the use of limited processing power. The third area is downsizing cryptography
calculations, such as RSA, a public-key cryptosystem, to fit IoT limitations. The rest of this
chapter is divided into three sections, one dedicated to each downsized area. In the last section, we conclude this chapter.
The rest of this chapter is organized as follows: In Section 2, we describe the preparation
stage of the classification process, which minimizes the need for the entire historic data and
then the anomaly detection processes using the outcome of the previous stage. In Section 3,
we describe the use of the Random Forest algorithm for distributed and parallel processing of automatic classification and anomaly detection. In Section 4, we present an improved
implementation of RSA to allow high class cryptography that runs in an IoT configuration. In
Section 5, we conclude this chapter and discuss our ongoing and future work.
##### 2. Classification framework for data streaming anomaly detection
To predict the behavior of a system, we usually examine its past data to discover common
patterns and other classification issues. This process consumes considerable computational
power and data storage. In this section, we describe an approach and a system, which requires
much less resources without compromising prediction capabilities and accuracy. It employs
three basic methods: a common behavior graph, the contour surrounding the graph, and
entropy calculation methods. When the system is about to be implemented for a specific
domain, the optimized combination of these three methods is considered, such that it fits the
unique nature of the domain and its corresponding type of data. In addition, we present a
-----
An Adaptive Lightweight Security Framework Suited for IoT 33
http://dx.doi.org/10.5772/intechopen.73712
framework and a process that will assist system designers in finding the optimal methods for
the case at hand. We use a case study to demonstrate this approach with meteorological data
collected over 15 years to classify and detect anomalies in new data.
This section is organized as follows: We begin by defining the problem, proceed with various solutions proposed in the literature, and then present our adjustable contour approach.
We then show how it is applicable for IoT. We proceed with a case study demonstrating the
build-up of the contour and how it is used for instant anomaly detection. We conclude with
a summary of the section.
**2.1. Problem definition**
The problem we attempt to solve is the optimization of the amount of sampling data collected
to maintain a proper balance between the quantity of sampling data and the information
extracted from it. The problem statement focuses on extracting concepts, methods, rules, and
measurements, so that at the end of the process, the original sampling data become redundant
and no longer need to be stored. However, to keep improving and adjusting the extracted
items to natural changes in the behavior of the sampled mechanism, we incorporate in the
approach an ongoing learning process. In addition, in the study, we concentrate on timedependent streaming sampling data, divided by fixed periods, so that we can repeat the analysis process for each period/cycle. Thus, while there are many classification algorithms using
time series sampling, the aim is not to compare the performance of yet another classifier, but
rather present a flexible method to compactly represent the data with several parameters that
can be chosen and adjusted. We suggest an independent framework that allows a flexible
adaptation of the contour to the nature of the given domain. Indeed, some of the reviewed
works, such as Reeves et al. [6], can be revised and adjusted to the problem statement and
serve as a valid alternative to the approach we present. We are striving for the best sampling
strategy given sequential data, generated from IoT devices.
The input given is a set of time series: D = {d [(1)], d [(2)], …, d [(][n][)]}, where each time series d[(i)] contains
pairs (timestamp and numeric value). The required output is an optimal set Dw = {a1, a2, …, am},
where ai can be any sampling item, such as a minimal data set, trends, graphs, measurements,
or rules, which strongly represents and supports the purpose of the original data set D.
We consider the set Dw and the full data set D as containing the same information, if they produce the same classifier. That is, if f (d) = fw (d) ∈ {−1, 1} for every new data series d, where f is a
classifier learned from D and fw is a classifier based on Dw. For instance, we can judge whether
a series of yearly temperatures represent an El Nino (EN) year or not, or whether a series of sensor data is characteristic of a suspected intrusion or not. Here, we consider two sets D and Dw
as containing the same (or similar) information if both can predict the future pattern of an initial series d. That is, we can use either D or Dw to predict a future item dn with similar accuracy.
**2.2. Literature review**
Real-world data typically contain repeated and periodic patterns. This suggests that the data
can be effectively represented and compressed using only a few coefficients of an appropriate
-----
34 Internet of Things - Technology, Applications and Standardization
basis. Mairal et al. [2] study modeling data vectors as sparse linear combinations of basic elements generating a generic dictionary and then adapt it to specific data. Jankov et al. [3] present an implementation of a real-time anomaly detection system over data streams and report
experimental results and performance tuning strategies. Vlachos et al. [4] formulate the problem of estimating lower/upper distance bounds as an optimization problem and establish the
properties of optimal solutions to develop an algorithm which obtains an exact solution to the
problem. Sakurada and Yairi [5] use auto-encoders with nonlinear dimensionality reduction
for the anomaly detection task. They demonstrate the ability to detect subtle anomalies where
linear PCA fails. Reeves et al. [6] present a multi-scale analysis to decompose time series and
to obtain sparse representations in various domains. Chilimbi and Hirzel [7] implement a
dynamic pre-fetching scheme that operates in several phases. The first is profiling, which
gathers a temporal data reference profile from a running program. Next, an algorithm extracts
hot data streams, which are data reference sequences that frequently repeat in the same order.
Then, a code is dynamically injected into appropriate program points to detect and pre-fetch
the hot data streams. Finally, the process enters the hibernation phase where the program
continues to execute with the added pre-fetch instructions. At the end, the program is deoptimized to remove the inserted checks and pre-fetch instructions and control returns to the
profiling phase. Lane and Brodley [8] claim that features can be extracted from object behavior and a domain heuristic. Experiments show that it detects anomalous conditions, and it is
able to identify a profiled user from other users. They present several techniques for reducing
70% of the storage required for user profile. Kasiviswanathan et al. [9] proposed a two-stage
approach based on detection and clustering of novel user-generated content to derive a scalable approach by using the alternating directions method to solve the resulting optimization
problems. Aldroubi et al. [10] show that for each dataset there is an optimized collection of
cells spanning the entire space and so generate the optimized sampling set.
The common underlying idea of the reviewed approaches is the definition of the problem
they are aiming to solve. The problem attempted to be solved is optimizing the size of the
collected sampling data so that it keeps the proper balance between the quantity of sampling
data and the information extracted from it.
**2.3. Contour-based approach**
Briefly, we analyze sampling data collected over several periods. We divide the period into
time-units. For example, for a period of a year, we divide it into daily time-units. For each
time-unit, we extract one value that represents it. This is done by averaging the samples collected during the time-unit. In the example, we may calculate the average value of all samples
of that day. We may also decide to select one of the samples to represent the day, e.g., the first
or last sample. We then calculate the average value for each time-unit from the collected values for the same time-unit in all periods, resulting in an average value for a given time-unit.
We repeat this process for all time-units in the period and obtain a graph that represents the
average values for an average and common period.
Assuming we have the average graph line for an average period, we now calculate the contour
around this average. The generated contour represents the standard range of values, such that
an unanalyzed period can be compared to this contour. If its graph value is completely within
-----
An Adaptive Lightweight Security Framework Suited for IoT 35
http://dx.doi.org/10.5772/intechopen.73712
the contour, the period is a standard period. If it is completely out of the contour, then it is
purely not standard. If the sections of the graph are within the contour, while others are out of
it, we use an entropy measure to calculate the overall “distance” of the given period from the
standard contour. Assuming an existing entropy threshold, we can decide whether the period
is a standard one or not. We apply the same concept at the unit level and decide whether a
specific time-unit in a period is within the standard or not. This specific check is relevant, for
example, to anomaly detection of IoT behavior.
In conclusion, the entire process is based on three key elements: the average graph per period,
the contour around the average graph, and an entropy value representing the overall distance
of a period from the contour. Each of these elements—average, contour, and entropy—can
be one of the several possibilities. For the contour, a simplistic choice would be minimum
and maximum (min-max) values. Alternatively, the SD or confidence interval (CI) could be
employed. These three elements affect each other, and every choice of such a triplet—average,
contour, and entropy—will produce a different behavior of the compressed classifier. The
object is to find the best triplet that will be able to disregard the original data after extracting
the representative contour, without compromising the ability to successfully analyze future
series. In our work, we consistently use the arithmetic average and classical entropy and focus
on finding the best contour.
_2.3.1. Finding the optimal contour_
We begin with a supervised learning approach, for classification, in which each time series
is labeled as one of two classes. To demonstrate, using the data set from the experiments,
the time series are year-long recordings of temperature samplings, labeled as positive, if the
corresponding year was an EN year, or otherwise negative. We now describe in detail the
process of building the classifier, with emphasis on finding the optimal contour.
Constructing the best contour is described in **Figure 1. We begin with raw data collected**
during N periods, where each record corresponds to a specific time-unit. These cycles have
already been classified positive or negative according to some classification criteria. These
classified cycles will later be used to determine the best contour.
The process is divided into four stages. In stage one, we use a selected average method and calculate the average graph line representing the N given cycles. This is done horizontally by calculating the average of the values related to the same time-unit across all N cycles. For example, we
calculate the average of the values for January 1st across the various years. Doing so for all timeunits will generate the average graph line. In stage two, we select several distance calculation
methods, and for each method, we construct its associated contour. This is done by calculating
the distance value for each distance method, e.g., the min-max difference, SD, and CI. Taking the
distance value, we add and subtract it from the average line to get the contour around the average. We repeat this process for all distance methods. At this stage, we have constructed several
contours around the average line. The goal now is to select the contour, which is most effective in
classifying unclassified cycles. This is done in stages three and four. In stage three, we calculate
the prediction power for each contour and select the one with the highest prediction power. This
is done by summing, for each contour, the number of cases in which its prediction was right and
calculating the average entropy of these correctly classified cycles. We do the same for wrong
-----
36 Internet of Things - Technology, Applications and Standardization
**Figure 1. Process of finding the optimal contour.**
predictions. In stage four, we use one entropy method with an associated threshold value. An
unclassified cycle with an entropy value lower than the threshold will be classified positive and
otherwise negative. For each contour, we calculate the entropy of the given classified cycles. The
result is a set of entropy values, where some are below the threshold and others are above it.
**a.** We repeat this for all classified cycles. We then sum up the number of correct predictions
and their total entropies. We do the same for wrong predictions. We then subtract the total
wrong numbers from the correct numbers. We repeat this process for all the constructed
contours and select the contour with the highest prediction power.
**b.** Calculating the entropy.
The entropy of a period, given a contour, is calculated as follows:
- Marking for every timestamp whether the cycle’s value at that timestamp is below,
within, or above the contour.
- Calculating the frequency of each of these three possibilities: below (p1), within (p2), and
above (p3)
- Using these as a ternary probability distribution, its entropy is calculated according to
the formula: p1 log[(][p1][)] + p2 log[(][p2][)] + p3 log[(][p3][)]
- The entropy measure is expected to return its minimum value at the two extreme cases:
When the cycle graph is entirely contained within the contour and when the cycle graph
lies entirely outside of the contour. All other cycles are expected to fall mostly within the
contour, and those which diverge enough from the contour, will have a high entropy
value which will lead to the right conclusion
-----
An Adaptive Lightweight Security Framework Suited for IoT 37
http://dx.doi.org/10.5772/intechopen.73712
**c.** Classifying a cycle/period
**Figure 2 describes the process of classifying unlabeled data cycles, as listed below:**
**1.** Apply the given data cycle to the contour and match it according to timestamps.
**2.** Noting for each timestamp whether the data point is below the contour, within it, or
above it.
**3.** Marking these cases respectively as −1, 0, and +1.
**4.** Calculating the frequencies of each of the three values: −1 (p1), 0 (p2), and +1(p3).
**5.** Calculating the entropy of the distribution defined by p1, p2, and p3.
**6.** Classifying as belonging to the contour, if the entropy is below the threshold determined in the learning phase.
_2.3.2. Advantages of the proposed technique_
The proposed technique has several advantages over other methods. The technique is a family of sampling methods and is defined by the three parameters described above. It is reasonable to expect that different datasets will require different parameters for the best sampling.
Different combinations can be tested and evaluated to ensure optimal treatment of the data.
The technique we propose is therefore flexible and adjustable and thus suits every given data
set. Secondly, this technique can be applied not only for classification but also for prediction
of time series.
Thirdly, the technique can be used to evaluate reliability of data online. In cases of high fluctuations or sharp changes in the cycle graph, which do not conform to either of the two class
contours, suspicion may arise that the reliability of the data has been compromised. This can
indicate that the sensor is damaged or that there has been a security breach.
Fourthly, the approach allows self-learning and automatic adjustments in cases of common
behavior changes and a new standard has been established. Lastly, occasionally, a post-mortem
may be run to check the system’s reaction to actual behavior and thereafter adjust the system
parameters accordingly.
**2.4. Anomaly detection for IoT security**
IoT devices generate time-related data, i.e., structured records containing a timestamp and
one or more numeric values. In many cases, we can identify recurrent time frames where the
system behavior has a repetitive format. Hence, IoT data have a structure to which the contour approach is highly applicable.
IoT security utilizes common data patterns and quantitative measurements. Based on the
identified patterns and measurements, we can extract logical rules that will be executed once
an exception is discovered. An exception may be any violation of predefined patterns, measurements, and other parameters, which represent normal, standard, and permitted behavior.
-----
38 Internet of Things - Technology, Applications and Standardization
**Figure 2. Classifying a cycle.**
In IoT, there is an abundance of possible patterns, starting with column level patterns up to
a super internet controlling several IoT networks. The goal is to find the methods and tools
to define standard patterns and how they can be identified. Once this is done, we can apply
the contour method. In our work, we show a two-dimensional contour. Using the same concept, we can expand it to be a multi-dimensional contour. This case is common where there
is a dependency among several columns within one record and the same applies for the case
where there are dependencies among networks of IoT systems.
**2.5. Case study**
In the following case study, we used meteorological data collected on EN years (positive
class) and NEN years (negative class) from 1980 to 1998. For the positive contours, we took
data from the EN years 1982, 1983, 1987, 1988, 1991, and 1992. All other years in the range were
NEN years. We tested three methods for generating contours: (a) max-min over all cycles;
(b) average cycle ± SD; and (c) CI.
**Figures 3 and 4 depict the contours for NEN years. Figure 3 shows the NEN contour in black**
according to the average ± SD and depicts how EN years diverge from this contour, as compared
-----
An Adaptive Lightweight Security Framework Suited for IoT 39
http://dx.doi.org/10.5772/intechopen.73712
to the NEN year—1995. The 1992 and 1988 (EN years) show clear divergence from the contour
while 1995 (a NEN) is more contained within the contour. This is nicely captured by the entropy
values, which for 1992 was 0.4266 and for 1988 was 0.3857—above the threshold, leading to the
conclusion that they are not NEN years—while for 1995, the entropy was 0.3631—significantly
lower than those of the EN years, leading to the correct conclusion that 1995 was indeed a NEN
year.
**Figure 4 shows two contours: the min-max contour and the average ± SD contour. The Y-axis**
in these graphs is the temperature value, and the X-axis is the time. Within each contour, the
year 1995 (a NEN year) is graphed. Its entropy is 0.3631 for the average SD contour and 0.2932
for the min-max contour. Both are the threshold, which leads to the correct conclusion that it
should indeed be classified as NEN.
In the case study, we compared the constructed contours, by using the average graph ± SD
and the average graph _± min-max. For the_ _SD contour, we obtained a significant entropy_
value difference between a classified EN case and a NEN case. In comparison, the min-max
contour resulted in close values of entropy for the EN cycle and the NEN cycle. Thus, the ability to differentiate between two extreme situations using entropy depends on the parameter
used to build the contour.
**2.6. Section summary**
In this section, we dealt with the classification problem of an unclassified cycle of IoT streaming data. We introduced the contour approach to draw the borders around the standard area
representing a specific class. If there was an unclassified cycle, we measured its distance from
**Figure 3. EN cycles on NEN average ± SD contour.**
-----
40 Internet of Things - Technology, Applications and Standardization
**Figure 4. NEN contours—min-max and SD.**
the contour using an entropy formula. Then, we compared the result to a predefined threshold. If the entropy value is below the threshold, the cycle is of the same class.
We propose a process for constructing the best contour that will presumably classify the correct underlying class. The process is based on three measurement methods: average, distance,
-----
An Adaptive Lightweight Security Framework Suited for IoT 41
http://dx.doi.org/10.5772/intechopen.73712
and entropy. For each method, there are several alternate formulas that we may use. Each
combination of these three methods may result in different contour and hence different
entropy value for the same unclassified cycle. We select the combination with the maximum
difference between positive and negative values.
In addition to the initial construction of the class contours from the given data, we suggest
ongoing improvements of the initial contours. Namely, we recalculate the class averages and
their contours to refine and revise the contours for improved classification performance.
In this manner, we are able to improve the contour approach, in reference to several aspects,
such as determining the minimal number of classified cycles required to define the best
contour, expanding the use of the contour to discover early trends or discover significant
changes in behavior and adjusting the contour accordingly, exploring the possibility of
dividing one cycle into several segments, and associating a different contour method to
each segment.
##### 3. Lightweight adaptive random forest for rule generation and execution
The volume of transmitted data over the various sensors continuously grows. Sensors typically are low in resources of storage, memory, and processing power. Data security and privacy are part of the major concerns and drawbacks of this growing domain. An IoT network
intrusion detection system is required to monitor and analyze the traffic and predict possible
attacks. Machine leaning techniques can automatically extract normal and abnormal patterns
from a large set of training sensors data. Due to the high volume of traffic and the need for
real-time reaction, accurate threat discovery is mandatory. This section focuses on designing a
lightweight comprehensive IoT rules generation and execution framework. It is composed of
three components, a machine learning rule discovery, a threat prediction model builder and
tools to ensure timely reaction to rules violation and unstandardized and ongoing changes in
traffic behavior. The generated detection model is expected to identify exceptions in real time
and notify the system accordingly.
We use random forest (RF) as the machine learning platform for the discovery of rules and
real-time anomaly detection. To allow RF adaptation for IoT, we propose several improvements to make it lightweight and propose a process that combines IoT network capabilities, messaging and resource sharing, to build a comprehensive and efficient IoT security
framework.
The rest of this section is organized as follows: We begin with an introduction followed by
the relevant literature review. We then discuss rules extraction using machine learning techniques. We present random forest as the most suitable ML for IoT. We proceed with various
improvements, utilizing RF and IoT attributes. We then outline an experiment that executes
RF building and its corresponding classifications using 15 different configurations, each based
on a unique combination of the number of processors and the forest size.
-----
42 Internet of Things - Technology, Applications and Standardization
**3.1. Introduction**
IoT is a network of objects, consisting of sensors, Internet, software, and exchange of data.
This generates critical issues of security, which must be addressed. Since to date there is no
standard for sensors, any system under development at this stage must consider the possibility that soon a standard will be defined, and the systems must be able to easily adjust
to it. Along with the limited processing power and the fact that the security issues must be
dealt with in real time, we realize the immediate need for a flexible and lightweight solution.
The solution should be dynamic, open, scalable, distributed and decentralized. The analysis
discovers patterns and measurements from the data, which are then translated into anomaly
detection rules associated with actions to be executed when a rule is violated. The rules are
then deployed in the IoT devices. When data are received from, or transmitted to an IoT
device, the rules are executed. If the result is positive, the corresponding action is triggered to
cope with the situation.
**3.2. Literature review**
Mansoori et al. [11] proposed a systematic process for retrieving fuzzy rules from a given
data set. To improve performance, the retrieved rules are then crystallized based on its
effectiveness and applicability. Dubois et al. [12] use Sugeno integrals, which are qualitative
criteria aggregations where it is possible to assign weights to groups of criteria. They show
how to extract if-then rules that express the selection of situations based on local evaluations
and rules to detect bad situations. Sumit-Gulwani, Hart, and Zorn [13] deal with converting data into an appropriate layout, which requires major investment in manual reformatting. The paper introduces a synthesis engine to extract structured relational data. It uses
examples to synthesize a program using an extraction language. Bharathidason et al. [23]
presented a fast and compact decision rules algorithm. The algorithm works online to learn
rule sets. It presents a technique to detect local drifts by taking advantage of the modularity
of the rule sets. Each rule monitors the evolution of performance metrics to detect a concept
drift. It provides useful information about the dynamics of the process generating data,
faster adaptation to changes, and generates more compact rule sets. Jafarzadeh et al. [15]
used averaging techniques to propose a method in which a previous algorithm for association rules mining is improved upon to specify minimum support. It uses fuzzy logic to distribute data in different clusters and then tries to provide the user with the most appropriate
threshold automatically. Limb et al. [16] used Fuzzy ARTMAP and Q learning to build a
data classification and rule mining model. To justify the classification, the model provides a
fuzzy conditional rule. Q-values are used to minimize QFAM prototyping. Mashinchi et al.
[17] proposed a granular-rules extraction method to simplify a data set into a granularrule set with unique granular rules. It performs in two stages to construct and prune the
granular rules. Yang H. et al [18] proposed an anomaly detection algorithm of Quick Access
Recorder (QAR) data, based on attribute support of a rough set. The method retains the
time characteristics of QAR data and strengthens the relation between the condition and
decision attributes. Tang [19] described an approach of data mining with Excel, using the
XLMiner add-in. This is an example of mining association rules to illustrate all the steps
-----
An Adaptive Lightweight Security Framework Suited for IoT 43
http://dx.doi.org/10.5772/intechopen.73712
of this approach. Tong S and Koller D. [20] introduced an algorithm for choosing which
instances to request next, in a setting in which the learner has access to a pool of unlabeled
instances and can request the labels for some number of them. The algorithm is based on
a theoretical motivation for using support vector machines (SVMs). Osungi et al. [21] proposed an active learning algorithm that balances exploration by dynamically adjusting the
probability to explore each step. Lang T et al. [22] proposed an active learning method for
multi-class classification. The method selects informative training compounds to optimally
support the learning progress. Bharathidason et al. [23] improved the performance and the
accuracy by including only uncorrelated high performing trees in a random forest.
The reviewed literature focuses on improvements to known rule discovery mechanisms, such
as machine learning, to transform them into lightweight systems that can be executed in limited resources settings. In most cases, the proposed solutions remain for general purposes but
can run with less required resources. We are seeking a solution that takes advantage of the
unique IoT attributes and utilizes them to build a combined comprehensive framework for
IoT security.
**3.3. Rules generation and deployment process**
The process consists of seven stages (see Figure 1). Stage 1 collects training data from the
IoT network, removes irrelevant records, and complements data in records with missing
data. In stage 2, we apply discovery techniques to extract important measurements and
patterns. Stage 3 consists of generating a rule for each measurement and pattern. In stage 4,
we evaluate the effectiveness of each rule with a set of training data. If the number of times
a rule has been executed is below a given threshold, the rule is removed from the rules set.
Next, in stage 5, we check the completeness and the integrity of the generated set of rules.
Rules that contradict another rule are removed and missing rules are added. Stage 6 runs a
simulation with the same training data with the presumption that all the designated rules
will be executed. Finally, in stage 7, we deploy the generated rules set. At this point, the
system is ready to accept the IoT traffic data in real time and automatically check it against
the set of rules.
**3.4. Extracting rules from training data**
A typical sensor record contains the sensor ID, timestamp, and one or more values per feature.
The main source for extracting rules is data collected from the concrete processes involved in
the explored domain. The significance to IoT is taking the accurate decision in real time and
react in real time to security alerts, notifications, automation, and predictive maintenance. To
ensure the completeness and the integrity of the generated set of rules, we use a consistent
multi-layer process of accumulating rules, starting with the simplest rules up to the most
complicated and multi-stage rules. Simple rules are extracted at the single feature level, and
then we proceed with rules extracted from a combination of any number of features having
a common relation, such as features of sensors sharing the same workflow. The generated
rules at this level relate to basic data such as maximum, minimum, average, standard deviation, median, and most frequent value. More complex relations, such as proportions among
-----
44 Internet of Things - Technology, Applications and Standardization
subsequent values, sequence trends, and significant patterns, require reasoning capabilities
and can be reached by machine learning and data mining techniques. The outcomes are measurements, thresholds, and patterns used to draw the corresponding decision trees. These
decision trees tend to grow fast, consuming large storage, and memory space along with
high runtime when pruning and analyzing it to find the specific rule. The depth of the tree
grows linearly with the number of variables, but the number of branches grows exponentially
with the number of states. Decision trees are useful when the number of states per variable
is limited. It becomes complicated when the state of the variables depends on a threshold or
complex computations. Communicating this rationale requires labeling every edge and then
tracing the tree path to understand the logic incorporated in it. Complex event processing
(CEP) engines are popular in IoT. They support matching time series data patterns that originate from different sources. However, they suffer from the same modeling issues as trees and
pipeline processing.
Rule engines have two major drawbacks in the context of IoT, the logic representation
is not compact and the use of it requires much processing power and time. We will cope
with these drawbacks in two ways. 1. Reduce the number of decision trees and improve
the search navigation scope, resulting in a reasonable and acceptable search time. 2.
Utilize IoT attributes and functionality to optimize the tree navigation flow and process
sharing.
In the following sections, we present the random forest machine learning and propose several
improvements where the known drawbacks are removed.
**3.5. Decision automation using random forest**
Random forest employs bootstrap aggregation for training. While the predictions of a single
tree are sensitive to noise in its training set, the average of many uncorrelated trees is not.
Bootstrap sampling is a way of decorrelating the trees by showing them different training
sets. Many trees reduce the depth and width of each tree and so save pruning and analysis
time, which suit IoT constraints.
The algorithm has two key parameters: the number of K trees to form a random forest
and the number of features F, randomly sampled features for building a decision tree. For
large and high dimensional data, a large K should be used. Estimating the performance
of random forest for one core is based on the following parameters: # trees [K], # features
[F], # rows [R], and maximum depth [D]. The estimated runtime is influenced by the number of features. Hence, keeping only the most important features lowers the number of
records and maintains the maximum depth low, which will improve the overall random
forest performance.
Random forest performance is better than the classical tree decision algorithm. However, it
may still be insufficient for IoT due to the memory space and processing power it requires.
Hence, building a lightweight RF process and utilizing IoT networking are required.
In the following section, we describe four proposals that make random-forest lightweight.
-----
An Adaptive Lightweight Security Framework Suited for IoT 45
http://dx.doi.org/10.5772/intechopen.73712
**3.6. Improving RF performance and consumption of resources**
**a.** Randomization may cause occurrence of redundant, irrelevant or even contradicting trees,
which may lead to redundant searches or even to the wrong decision. Therefore, selection
of trees with high classification accuracies leads to improved performance and better decision accuracy. A decision process is effective when the difference among the relevant alternatives is significant. RF contains many decision trees, where each of them may contribute
to the final decision. Many such trees generally require wider searches and thus expand the
decision process. On the one hand, reducing the number of the searched trees will shorten
the process but on the other hand may increase the probability of making the wrong decision. Therefore, a selection criterion for removing the “redundant” trees is required. An
initial approach is to remove similar trees as correlated trees hardly contribute to reaching
the correct decision. Thus, for effective RF decisions, we strive to remove uncorrelated trees
[14]. The correlation between two trees may be defined in various ways, such as:
**1.** Distance—we transform the tree into a sequence of values, and then we apply a hashing function on this sequence and get a score. Two trees are correlated if the difference
between the scores is below a predefined threshold.
**2.** Common components—count the number of similar components and compare.
**3.** Empirically by removing the tree and trying a vast number of cases, we will reach the
same decisions as we would if the tree was included, which means that the tree has no
effect on practical decisions.
**b.** Prioritize trees by simulation using labeled and already classified cases.
Instead of removing trees, we propose prioritizing them. The prioritization can be an empirical study of the historical use and effectiveness in true/false decisions. Another way is
to run a Monte-Carlo intensive simulation and prioritize trees accordingly.
**3.7. Prioritize trees by its threat level**
We define several security levels: low, normal, high, and emergency. For each level, we associate
the most effective trees and the order of the trees to be visited. For each network, we designate
a security manager device, which collects messages from its network devices, assesses it, and
determines the network security level. When the network is initiated, the designated level is
low. As time passes, messages arrive at the security manager device, which analyzes the input
and decides to change the security level. Then, a message is distributed requesting a security
level change. Once the level is changed, the local system activates the new tree search schedule.
**3.8. Messaging assisted, best trees selection**
MQTT is a lightweight messaging protocol, over TCP, adjusted to the IoT domain. Given
MQTT, we can utilize the IoT network itself to improve performance. We use it to transfer
messages and data from one device to another. For example, in case of a suspicious occasion
-----
46 Internet of Things - Technology, Applications and Standardization
detected by one of the sensors, using the protocol, the device sends alert messages to other
members. The messages include data strings and unique data patterns that receivers should
expect to receive and thus detect a malicious situation. The message may also include the
most effective trees that may cope with the suspected threat.
When suspicious data reach a sensor, it is analyzed locally, and the best tree sequence is
identified. This device sends a message to the security manager, containing the data with
the detected anomaly and the sequence of trees to visit and act accordingly. The messaging
protocol is an adjustment of HTTP.
**3.9. Experiment using the random forest in an IoT**
In this section, we describe a comprehensive test, simulating the building of various random
forests and then runs several classification cycles for a given set of anonymous records. We
used a computer with eight processors running the random forest PMI platform with 10–1000
trees per forest. It contained a random forest builder, an anonymous records classification
process, and a configuration tool. We sought the best configuration, suitable for the optimized
performance and accuracy of a random forest simulation. A configuration in this context is
measured by the combination of the number of processors and the number of trees in a forest. For the simulation, we used 500 anonymous records and 3350 already classified samples,
where each sample has 95 attributes. We ran 30 test cycles where each cycle represented a
unique configuration—number of processors: 2, 4, 6, 8, and 16 and the number of trees per
forest: 10, 100, 250, 500, 750, and 1000. For comparison, all test cycles used the same data set.
In cases of similar trees, we ran a process that removes similar trees. The performance of the
entire 30 test cycles is evaluated by its accuracy and processing time.
**Figures 5 and 6 show that accuracy, performance of each of the processes and combined are**
best achieved when using 10 trees per forest and 8 processors. Based on the above simulations,
it seems that for the example at hand, using a relatively small number of trees per forest and
multi-core processors is recommended for optimal performance and high accuracy. However,
this may not be the common case. Therefore, prior to implementing RF-based anomaly detection, it is recommended that a simulation test be run with the main data. In addition, we
propose a prototype of an IoT environment. The prototype is composed of one server and six
Arduino OS devices. We built two configurations, A and B. In configuration A, all the devices
are connected via WIFI 14 to the server, where the data transmission between two devices is
done through the server. The entire RF is loaded in the server while the devices have one tree
installed in them. The data flow of an incoming event in configuration A can be one of the following: 1. An event arrives at a device, the device forwards it to the server, which then runs
the RF and classifies the event. 2. An event arrives at a device and the device forwards it to the
server. The server forwards it to all devices. Each device checks the event against the appropriate local tree and sends the result to the server. The server then counts the results and sends the
reply to the sender, which acts accordingly. The flow in configuration B is as follows: An event
arrives at a device, the device propagates it to other devices, checks it against its own tree,
and propagates the results back to the sender. The sender classifies the event and acts accordingly. To test the feasibility of the prototype, we used the trees built by the simulation tool and
loaded it to the server and devices accordingly. We transmitted 500 events to the devices in
-----
An Adaptive Lightweight Security Framework Suited for IoT 47
http://dx.doi.org/10.5772/intechopen.73712
**Figure 5. Results of running the 30 classification processes.**
**Figure 6. Accuracy and combined results of running the 30 classification processes.**
a round robin schedule. The resulting accuracy level was similar to the level we found in the
previous simulation. Performance was out of the scope of the prototype stage. Nonetheless,
we did not notice streaming interruptions or delays. In future work, we intend to design and
perform consistent and comprehensive tests of the device and other similar devices. Based on
the results, we will be better able to determine which rules are to be executed in real time and
which are to be executed online or in batch mode.
##### 4. Lightweight public key cryptographic processor suited for IoT
Due to the vast number of IoT devices and high transmission volumes, a robust and adaptive cryptography system is required. However, since IoT devices have limited memory and
computation power, they are unable to execute public key cryptographic systems. To cope
-----
48 Internet of Things - Technology, Applications and Standardization
with this limitation, we propose a lightweight RSA process. A combination of symmetric
and asymmetric encryption systems is commonly used by the industry. Symmetric encryption systems require moderate computation resources and consequently are already used
in IOT. However, asymmetric public key encryption requires vast computation resources,
and as a result cannot be executed by most IOT devices. In this section, we describe a lightweight RSA encryption, where three improvements are incorporated: acceleration of modular
exponentiation calculation, parallel and distributed multi-core processing, and splitting the
original message if the message length is very long. After each part is completed, the system
collects the intermediate results and loads them into a consolidation and integration process,
which generates the result. We ran comprehensive encryption and decryption processes on
messages of various lengths. The results prove that lightweight RSA is ready to be incorporated in IoT devices.
The rest of this section outlines the relevant literature review. Then, we describe an example
of smart modular exponential calculation, which runs efficiently in an IoT architecture.
**4.1. Literature review**
Lin et al. [24] proposed the execution in parallel on CPU/GPU hybrids, of the Montgomery
algorithm, to improve RSA performance and security. Fadhil and Younis [25] proposed a
hybrid system, running RSA on multi-core CPU and multi GPU cores. For comparison purposes, they implemented variants of RSA, Crypto++, and the sequential counterpart. Multithread CPU improved performance by 6, over the sequential CPU implementation, and with
GPU, it improved 23 times over the sequential implementation. The throughput gained for
1024 bits was ~1800 msg/sec, and for 2048 bits, it was ~250 msg/sec. Yanga et al. [26] suggested
a parallel block Wiedemann algorithm in cloud to enhance the performance of GNFS and
reduce communication costs, involved in solving large and sparse linear systems over GF.
**4.2. Example of the acceleration of a modular exponentiation calculator**
The calculation of “a factor b modulo n” is the heart of RSA cryptography and is also the most
resource consuming component. Dividing this calculation into smaller parts will allow distributed and parallel processing of this calculation, where each smaller part is calculated by
one sensor and later is integrated to obtain the result of “a factor b modulo n.” The underlying
concept is the following conceptual equation: ((a mod n) * (b mod n)) mod n = (a*b) mod n.
This concept is used by the following algorithm to calculate modular exponentiation. Step 1:
Translate the input into a binary number. Step 2: Start at the rightmost digit, let k = 0, for each
positive digit calculate the value of 2^k, Step 3: Calculate mod n of the powers of two ≤ b, Step
4: Use modular multiplication properties to combine the calculated mod n values. Steps 2 and
3 can be executed in parallel by several connected sensors. The results from the sensors are
then sent to the sensor requested the encryption/decryption, to execute step 4 and obtain the
final result. Using a network of 7688 devices, we ran a comprehensive test, which proves the
feasibility of executing RSA using parallel and distributed processing.
-----
An Adaptive Lightweight Security Framework Suited for IoT 49
http://dx.doi.org/10.5772/intechopen.73712
##### 5. Conclusion
Connecting sensors to the Internet exposes the entire network to malicious penetrations.
This is due to poor computation resources in standard sensors, which do not allow the execution of robust security systems. Hence, lightweight primitive systems should be implemented in IoT. To maintain current Internet security level, we adjusted implementations of
known security concepts and mechanisms, which contribute to the security of the Internet
of things. In this chapter, we focused on three key security elements where downsizing is
feasible without compromising security: (a) Eliminating the frequent use of detailed data
in the classification process. (b) Adjusted random forest machine learning to work in a distributed and parallel mode, when building the forest and during the detection process. (c)
Adjust RSA cryptography calculations which are executed in parallel and distributed. The
proposed solutions have smaller footprints, are efficient, and in most cases demonstrate better performance. We prove that downsizing and parallel processing are the most appropriate
approaches for implementing comprehensive concepts for proper operation in constrained
environments of IoT.
We are currently working on expanding current research areas. For example, additional
improvements in RF implementation and exploring other machine learning technologies to
check its applicability to IoT anomaly detection. We are exploring other asymmetric cryptography systems to check their applicability to IoT. In parallel, we are investigating authentication methods and technologies to discover a suitable one for IoT, or we are considering
building an IoT-specific authentication.
##### Author details
Menachem Domb
Address all correspondence to: dombmnc@edu.aac.ac.il
Ashkelon Academic College, Computer Science Department, Ashkelon, Israel
##### References
[1] Aldosari HM, Snasel V, Abraham A. A new security layer for improving the security of
internet of things (IoT). Inteernational Journal of Computer Information Systems and
Industrial Management Applications. 2016;8:275-283. ISSN: 2150-7988
[2] Mairal FB, Ponce J, Sapiro G. Online dictionary learning for sparse coding. In: Proceedings
of the 26th Annual International Conference on Machine Learning. Montreal, Quebec,
Canada: ACM; 2009. pp. 689-696
-----
50 Internet of Things - Technology, Applications and Standardization
[3] Jankov D, Sikdar S, Mukherjee R, Teymourian K, Jermaine C. Real-time high performance anomaly detection over data streams: Grand challenge. In: Proceedings of the
11th ACM International Conference on Distributed and Event-based Systems. Tokyo,
Japan: ACM; 2017. pp. 292-297
[4] Vlachos M, Freris NM, Kyrillidis A. Compressive mining: Fast and optimal data mining
in the compressed domain. The VLDB Journal. 2015;24(1):1-24
[5] Sakurada M, Yairi T. Anomaly detection using autoencoders with nonlinear dimensionality reduction. In: Proceedings of the MLSDA 2014 2nd Workshop on Machine Learning
for Sensory Data Analysis. ACM; 2014. p. 4
[6] Reeves G, Liu J, Nath S, Zhao F. Managing massive time series streams with multi-scale
compressed trickles. Proceedings of the VLDB Endowment. 2009;2(1):97-108
[7] Chilimbi TM, Hirzel M. Dynamic hot data stream prefetching for general-purpose programs. In: ACM SIG-PLAN Notices. Berlin, Germany: ACM; 2002;37(5):199-209
[8] Lane T, Brodley CE. Temporal sequence learning and data reduction for anomaly detection. ACM Transactions on Information and System Security (TISSEC). 1999;2(3):295-331
[9] Kasiviswanathan SP, Melville P, Banerjee A, Sindhwani V. Emerging topic detection
using dictionary learning. In: Proceedings of the 20th ACM international conference on
Information and knowledge management. ACM; 2011. pp. 745-754
[10] Aldroubi A, Cabrelli C, Molter U. Optimal non-linear models for sparsity andsampling.
Journal of Fourier Analysis and Applications. 2008;14(5-6):793-812
[11] Mansoori EG, Zolghadri MJ, Katebi SD. SGERD: A steady-state genetic algorithm for
extracting fuzzy classification rules from data. IEEE Transactions on Fuzzy Systems.
Aug 2008;16(4):1061-1071. ISSN: 1063-6706
[12] Dubois D, Durrireu C, Prade H, Rico A, Ferro Y. Extracting decision rules from qualitative data using sugeno integral: A case-study. In: Proceedings of the 13th European
Conference, ECSQARU 2015. Compiègne, France: Springer; Jul 2015;9161:14-24. ISBN:
978-3-319-20806-0; ISSN: 0302-9743
[13] Daniel W. Gulwani S, Hart T, Zorn B. FlashRelate: extracting relational data from semistructured spreadsheets using examples. In: Proceedings of the 36th ACM SIGPLAN
Conference on Programming Language Design and Implementation. Vol. 50, Issue. 6.
New York: ACM; Jun 2015. pp. 218-228
[14] Kosina P, Gama J. Very fast decision rules for classification in data streams. Data Mining
and Knowledge Discovery. Jan 2015;29(1):168-202. ISSN: 1384-5810
[15] Jafarzadeh H, Torkashvand R, Asgari C, Amiry A. Provide a new approach for mining fuzzy association rules using apriori algorithm. Indian Journal of Science and
Technology. Apr 2015;8(S7):127-134. ISSN: 0974-6846
-----
An Adaptive Lightweight Security Framework Suited for IoT 51
http://dx.doi.org/10.5772/intechopen.73712
[16] Pourpanaha F, Peng Limb C, Mohamad Saleh J. A hybrid model of fuzzy ARTMAP and
genetic algorithm for data classification and rule extraction. Elsevier, Expert Systems
with Applications. 2016;49(7):4-85
[17] Mashinchi R, Selamat A, Ibrahim S, Krejcar O. Granular-rule extra action to simplify
data. In: Intelligent Information and Database Systems. Vol. 9012 of the Series LNCS. Mar
2015. pp. 421-429. ISBN: 978-3-319-15704-7
[18] Yang H, Xiao C, Qiao Y. Study on anomaly detection algorithm of qar data based on
attribute support of rough set. International Journal of Hybrid Information Technology.
2015;8(1):371-382. DOI: http://dx.doi.org/10.14257/ijhit.2015.8.1.33. ISSN: 1738-9968. IJHIT
Copyright ⓒ 2015 SERSC
[19] Tang H. A simple approach of data mining in excel. In: 4th International Conference
Browse Conference Publications, IEEE Xplore; 2008
[20] Tong S, Koller D. Support vector machine active learning with applications to text classification. Journal of Machine Learning Research, Leslie Pack Kaelbling. 2001:45-66
[21] Osugi T, Kun D, Scott S. Balancing exploration and exploitation: A new algorithm for
active machine learning. In: Fifth IEEE International Conference on Data Mining. IEEE
Xplore, 2005. DOI: 10.1109/ICDM.2005.33
[22] Lang T, Flachsenberg F, Luxburg U, Rarey M. Feasibility of Active Machine Learning
for Multiclass Compound Classification. 2016. PMID: 26740007. DOI: 10.1021/acs.jcim.
5b00332
[23] Bharathidason S, Jothi Venkataeswaran C. Improving classification accuracy based on
random forest model with uncorrelated high performing trees. International Journal of
Computer Applications (0975-8887). Sep 2014;101(13)
[24] Lin C, Liu J, Li C-C, Chu P-W. Parallel modulus operations in RSA encryption by CPU/
GPU hybrid computation. In: Taiwan, Conference Paper, IEEE Xplore: 29. Jan 2015. DOI:
10.1109/AsiaJCIS.2014.25
[25] Fadhil HM, Younis MI. Parallelizing RSA algorithm on multicore CPU and GPU.
International Journal of Computer Applications (0975-8887). Feb 2014;87(6)
[26] Yanga LT, Huanga G, Jun Feng B, Xua L. Parallel GNFS algorithm integrated with parallel block Wiedemann algorithm for RSA security in cloud computing. Information
Sciences, Elsevier. 2016
-----
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.5772/INTECHOPEN.73712?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.5772/INTECHOPEN.73712, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "HYBRID",
"url": "https://www.intechopen.com/citation-pdf-url/59350"
}
| 2,018
|
[
"Review"
] | true
| 2018-08-01T00:00:00
|
[
{
"paperId": "a83bc4f8ad7e2867ac5a34ccca850d1a52ca3d47",
"title": "Real-time High Performance Anomaly Detection over Data Streams: Grand Challenge"
},
{
"paperId": "8f2983c45eac6f055bc44a2f41dada29e1c46881",
"title": "Parallel GNFS algorithm integrated with parallel block Wiedemann algorithm for RSA security in cloud computing"
},
{
"paperId": "f361bb890bf1186bb4ccdecf0c684c9d5ccb8766",
"title": "A hybrid model of fuzzy ARTMAP and genetic algorithm for data classification and rule extraction"
},
{
"paperId": "49ab7c56113d526411a715c91fe80080e20750c9",
"title": "Feasibility of Active Machine Learning for Multiclass Compound Classification"
},
{
"paperId": "feae78252abb92b6066968d490a1e691f3dc0b94",
"title": "Extracting Decision Rules from Qualitative Data Using Sugeno Integral: A Case-Study"
},
{
"paperId": "8a60432910c1d4c0459bf448008ed967e25529b1",
"title": "FlashRelate: extracting relational data from semi-structured spreadsheets using examples"
},
{
"paperId": "5b3ec3254e8591728d164c2a975db1c3bbc0c257",
"title": "Provide A New Approach for Mining Fuzzy Association Rules using Apriori Algorithm"
},
{
"paperId": "0325826ae01aa5f79af71b4592af5c8d48294b0e",
"title": "Granular-Rule Extraction to Simplify Data"
},
{
"paperId": "3e911b6d87e9e6ac661d1bb04782497d0a958277",
"title": "Study on Anomaly Detection Algorithm of QAR Data Based on Attribute Support of Rough Set"
},
{
"paperId": "ee70341e75c2dffebbabd24b239cc158ad691ed1",
"title": "Anomaly Detection Using Autoencoders with Nonlinear Dimensionality Reduction"
},
{
"paperId": "e9393f9a6a6817b96c51ee49f8697f56635aebea",
"title": "Improving Classification Accuracy based on Random Forest Model with Uncorrelated High Performing Trees"
},
{
"paperId": "1552fa6de4088ed9466cef940e05d9e4d3e731b4",
"title": "Parallel Modulus Operations in RSA Encryption by CPU/GPU Hybrid Computation"
},
{
"paperId": "98f3a2ca2dfe5387da11c283f64c4c16caf2e3f6",
"title": "Compressive mining: fast and optimal data mining in the compressed domain"
},
{
"paperId": "d7aad9d68a40aed4fa215f2e26b31ac7d26c1027",
"title": "Parallelizing RSA Algorithm on Multicore CPU and GPU"
},
{
"paperId": "8465541f4bc7a0488124cad784bccba56647f94f",
"title": "Very fast decision rules for classification in data streams"
},
{
"paperId": "12a2958c923b48849cb1d06faaab357969295806",
"title": "Emerging topic detection using dictionary learning"
},
{
"paperId": "6d91ceb0e58c6f82c6901ab1861bbb91332db2c8",
"title": "Managing Massive Time Series Streams with MultiScale Compressed Trickles"
},
{
"paperId": "12439a6ff384e95ee2262ee982bc055534e30487",
"title": "Online dictionary learning for sparse coding"
},
{
"paperId": "de8230310005d9945271fce3b58a8598241f7ddc",
"title": "A Simple Approach of Data Mining in Excel"
},
{
"paperId": "f15ab3c43ebb0d6201818925eabe41a553386bf7",
"title": "SGERD: A Steady-State Genetic Algorithm for Extracting Fuzzy Classification Rules From Data"
},
{
"paperId": "aa6da31754897581238cdf1f3602bdde325c3c21",
"title": "Optimal Non-Linear Models for Sparsity and Sampling"
},
{
"paperId": "01a27d77257c47a24daaa969f258ea844b9cbff8",
"title": "Balancing exploration and exploitation: a new algorithm for active machine learning"
},
{
"paperId": "60b85b7ee655397a4d2202f9cdf6dd5e3f04f6fd",
"title": "Dynamic hot data stream prefetching for general-purpose programs"
},
{
"paperId": "a8797f1d253c75669d96e6fcceda2be3f8534e1d",
"title": "Support Vector Machine Active Learning with Applications to Text Classification"
},
{
"paperId": "62fdd44b8048848c5bd25cab55ff4913215da731",
"title": "Temporal sequence learning and data reduction for anomaly detection"
},
{
"paperId": "ee240b54cba7480f40f182c932edd06f5d5d4ab7",
"title": "A New Security Layer for Improving the security of internet of things ( IoT )"
}
] | 12,997
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Engineering",
"source": "s2-fos-model"
},
{
"category": "Environmental Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/02118711c1841d380df17c2537690bc0cedc4906
|
[
"Computer Science"
] | 0.903187
|
A Double Auction for Charging Scheduling among Vehicles Using DAG-Blockchains
|
02118711c1841d380df17c2537690bc0cedc4906
|
ACM Trans. Sens. Networks
|
[
{
"authorId": "47093617",
"name": "Jianxiong Guo"
},
{
"authorId": "3437541",
"name": "Xingjian Ding"
},
{
"authorId": "47203186",
"name": "Weili Wu"
},
{
"authorId": "2065276848",
"name": "Dingzhu Du"
}
] |
{
"alternate_issns": null,
"alternate_names": null,
"alternate_urls": null,
"id": null,
"issn": null,
"name": null,
"type": null,
"url": null
}
|
Electric Vehicles (EVs) are becoming more and more popular in our daily life, which replaces traditional fuel vehicles to reduce carbon emissions and protect the environment. EVs need to be charged, but the number of charging piles in a Charging Station (CS) is limited, and charging is usually more time-consuming than fueling. According to this scenario, we propose a secure and efficient charging scheduling system based on a Directed Acyclic Graph (DAG)-blockchain and double-auction mechanism. In a smart area, it attempts to assign EVs to the available CSs in the light of their submitted charging requests and status information. First, we design a lightweight charging scheduling framework that integrates DAG-blockchain and modern cryptography technology to ensure security and scalability during performing scheduling and completing tradings. In this process, a constrained multi-item double-auction problem is formulated because of the limited charging resources in a CS, which motivates EVs and CSs in this area to participate in the market based on their preferences and statuses. Due to this constraint, our problem is more complicated and harder to achieve truthfulness as well as system efficiency compared to the existing double-auction model. To adapt to it, we propose two algorithms, namely, Truthful Mechanism for Charging (TMC) and Efficient Mechanism for Charging (EMC), to determine an assignment between EVs and CSs and pricing strategies. Then, both theoretical analysis and numerical simulations show the correctness and effectiveness of our proposed algorithms.
|
## A Double Auction for Charging Scheduling among Vehicles Using DAG-Blockchains
#### Jianxiong Guo, Member, IEEE, Xingjian Ding, Weili Wu, Senior Member, IEEE, and Ding-Zhu Du
**_Abstract—Electric Vehicles (EVs) are becoming more and_**
**more popular in our daily life, which replaces traditional fuel**
**vehicles to reduce carbon emissions and protect the environment.**
**EVs need to be charged, but the number of charging piles**
**in a Charging Station (CS) is limited and charging is usually**
**more time-consuming than fueling. According to this scenario,**
**we propose a secure and efficient charging scheduling system**
**based on a Directed Acyclic Graph (DAG)-blockchain and double**
**auction mechanism. In a smart area, it attempts to assign EVs**
**to the available CSs in the light of their submitted charging**
**requests and status information. First, we design a lightweight**
**charging scheduling framework that integrates DAG-blockchain**
**and modern cryptography technology to ensure security and scal-**
**ability during performing scheduling and completing tradings. In**
**this process, a constrained multi-item double auction problem**
**is formulated because of the limited charging resources in a**
**CS, which motivates EVs and CSs in this area to participate**
**in the market based on their preferences and statuses. Due to**
**this constraint, our problem is more complicated and harder to**
**achieve truthfulness as well as system efficiency compared to the**
**existing double auction model. To adapt to it, we propose two**
**algorithms, namely Truthful Mechanism for Charging (TMC)**
**and Efficient Mechanism for Charging (EMC), to determine an**
**assignment between EVs and CSs and pricing strategies. Then,**
**both theoretical analysis and numerical simulations show the**
**correctness and effectiveness of our proposed algorithms.**
**_Index Terms—Electric Vehicle (EV), Charging Scheduling,_**
**DAG-based Blockchain, Constrained Multi-item Double Auction,**
**Truthfulness, System Efficiency.**
I. INTRODUCTION
O deal with the fossil energy crisis and reduce the emission of greenhouse gases, Electric Vehicles (EVs) have
# T
attracted more and more people’s attention because of their
great potential. Renewable energy will become the mainstream
way of energy supply in the near future. Compared to traditional fuel vehicles, EVs have a number of advantages such
as cost reduction, renewability, and environmental protection.
With the development of EVs, a large number of Charging
Stations (CSs) will be deployed in cities, which is different
from current gas stations. Because of the tedious storage and
transportation of gasoline, the deployment of gas stations is
Jianxiong Guo is with the Advanced Institute of Natural Sciences, Beijing
Normal University, Zhuhai 519087, China, and also with the Guangdong Key
Lab of AI and Multi-Modal Data Processing, BNU-HKBU United International College, Zhuhai 519087, China. (E-mail: jianxiongguo@bnu.edu.cn)
Xingjian Ding is with the Faculty of Information Technology, Beijing
University of Technology, Beijing 100124, China. (e-mail: dxj@bjut.edu.cn)
Weili Wu and Ding-Zhu Du are with the Department of Computer Science,
Erik Jonsson School of Engineering and Computer Science, The University of
Texas at Dallas, Richardson, TX 75080, USA. (E-mail: weiliwu@utdallas.edu;
dzdu@utdallas.edu)
_(Corresponding author: Xingjian Ding.)_
M i t i d i d
usually centralized. There is no such problem with electricity,
especially after the rise of distributed energy. It can be seen
that the deployment of CSs is more distributed, the number of
CSs is larger, and the single CS is smaller. Besides, it usually
takes more than half an hour to charge an EV, which is not
the same thing as being able to refuel in an instant. This has
also created a degree of administrative hardship.
In this paper, we consider such a scenario: In a smart
area, there is a manager that is responsible for the overall
scheduling of EVs and CSs in this area. EVs hope to complete
charging at the fastest speed and at the least cost, while
CSs hope to maximize their profits by providing charging
services. However, there are still some challenges that need
to be resolved.
On the one hand, it lacks an effective approach to ensure the security of charging trading between EVs and CSs.
Traditional centralized trading platforms depend on a trusted
third party to manage and store every transaction between
EVs and CSs. These platforms are plagued by some attacks
[1] [2] such as single point of failure, denial of service, and
privacy leakage [3]. A lot of existing research about charging
management neglects the security and privacy protection of
trading. The advent of blockchain technology has made it
possible to improve these issues. Blockchain is a kind of
decentralized ledger, which can realize security and privacy
protection through the knowledge of modern cryptography
and distributed consensus protocol. Based on that, we propose
a Blockchain-based Charging Scheduling (BCS) system to
manage the charging assignments between EVs and CSs in
a secure and efficient manner.
First, we are able to protect the contents of communication
among the entities in this system from being tampered with or
leaked by using digital encryption techniques. However, due
to its confirmation delay, limited scalability, and the inherent
uncertainty of consensus mechanisms [4] [5], traditional chainbased blockchains are not applicable to transactions with highfrequency characteristics. In addition, the incentive to miners,
transaction fees, and restricted block sizes are also important
factors that limit the use of such blockchains. To overcome the
above drawbacks, the infrastructure of our BCS system is built
on a Directed Acyclic Graph (DAG)-based blockchain [6]. It
has an asynchronous consensus process, which can make full
use of network bandwidth to improve scalability. The DAGbased blockchain has no miners, thus it reduces latency and
transaction fees, which makes frequent energy transactions
between EVs and CSs possible [6]. At the same time, it
can guarantee the same security and decentralization as the
traditional blockchain In our BCS system EVs work as light
-----
nodes, while CSs work as full nodes which are responsible for
issuing new transactions, storing, and maintaining the whole
blockchain system.
On the other hand, we have mentioned that charging is more
time-consuming than fueling. For each CS in this system, the
number of charging piles is limited. We can imagine that an
EV goes to a CS for charging, but there is no idle charging pile
at this moment that can charge it. This EV has to wait for other
EVs to finish charging or go to other CSs, which is frustrating.
For each EV in this system, it is more inclined to those CSs
that locate near it or provide better services to it. Because EVs
and CSs in this system belong to different entities, they are
driven by their own utilities. Then, a natural question is how to
assign the EV that submits a charging request in this system to
a CS with a limited number of charging piles. Based on that,
we formulate a constrained multi-item double auction model,
where EVs are buyers and CSs are sellers.
This model can be considered as an instance of a singleround multi-item double auction process [7]. For the multiitem double auction, Jin et al. [8] [9] proposed a resource
allocation problem in mobile cloud computing and designed a
truthful resource auction mechanism for the resource trading
between mobile users (buyers) and cloudlet (sellers). However,
this is a one-to-one mapping from buyers to sellers, that is,
at most one mobile user can be assigned to one cloudlet at a
moment. This scheme cannot be applied to solve our problem,
because the number of EVs that can be assigned to a CS is
more than one but less than the number of available charging
piles in this CS. In other words, this is a many-to-one mapping
from buyers to sellers. Secondly, to maximize the utilities of
CSs (sellers), the assignment and price determination depends
not only on the unit bids of EVs, but also on their charging
amounts. Due to the limited number of charging piles in each
CS as well as the different preferences and charging amounts
of each EV, our constrained multi-item double auction model is
distinguished from any existing double auction mechanisms.
It is more complex and harder to achieve truthfulness and
system efficiency. In order to solve this challenge, we design
a Truthful Mechanism for Charging (TMC) first to determine
the winners and transaction prices. By theoretical analysis,
it meets the requirements of individual rationality, budget
balance, truthfulness, and computational efficiency. Because
of achieving truthfulness, TMC sacrifices a part of system
efficiency. To improve the system efficiency, we design an
Efficient Mechanism for Charging (EMC) that increases the
number of successful trades (winning buyers) significantly
more than that in TMC. Nevertheless, EMC is not able to
ensure truthfulness for the buyers in extreme cases. The DAGbased blockchain and built-in smart contract implemented by
the double auction enable our BCS system to work distributed
and securely together.
The contributions of this paper are summarized as follows:
_• We investigate the challenges of EVs charging at present,_
integrate DAG-blockchain with charging scheduling, and
proposed a complete framework for charging scheduling
and digital trading.
_• We model the assignment between EVs and CSs as a_
constrained multi item double auction This model is
first proposed by us, and its objective and constraint
are fundamentally different from existing auctions. To
solve it, we design a TMC algorithm that can guarantee
truthfulness.
_• In order to meet the rapid response of scheduling systems,_
we design an EMC algorithm that is more efficient than
TMC at the expense of truthfulness.
_• We build a simulation environment for the BCS system,_
and verify our expected goals for the system and auction
theory through intensive simulations.
**Organizations: In Sec. II, we discuss the-state-of-art work.**
In Sec. III, we introduce the background of DAG-based
blockchain. In Sec. IV, we present our BCS system and charging scheduling framework elaborately. In Sec. V, constrained
multi-item double auction is formulated. Then, two solving
mechanisms TMC and EMC are shown in Sec.VI and VII.
Finally, we evaluate our proposed algorithms by numerical
simulations in Sec.VIII and show the conclusions in Sec. IX.
II. RELATED WORK
Recently, blockchain has been used as an effective method
to deal with the issues of transactions generated by peerto-peer (P2P) energy trading among EVs. Kang et al. [10]
exploited a P2P electricity trading system with consortium
blockchain to motivate EVs to discharge for balancing local
electricity demand. Wu et al. [11] studied energy scheduling
in office buildings through combining renewable energies and
workplace EV charging, and used stochastic programming to
manage the uncertainty of charging. Liu et al. [12] proposed an
EV participation charging scheme for a blockchain-enable system to minimize the fluctuation level of the grid network and
charging cost of EVs. Su et al. [13] designed a contract-based
energy blockchain in order to make EVs charge securely in the
smart community, while they implemented a reputation-based
delegated Byzantine fault tolerance consensus algorithm. Zhou
_et al. [14] developed a secure and efficient energy trading_
mechanism based on consortium blockchain for vehicle-to-grid
that exploited the bidirectional transfer technology of EVs to
reduce the demand-supply mismatch. Xia et al. [15] proposed
a vehicle-to-vehicle electricity scheme in blockchain-enabled
Internet of Vehicles to address the driving endurance issue of
EVs. Guo et al. [16] designed a lightweight blockchain-based
information trading framework to realize a real-time traffic
monitoring.
However, due to its confirmation delay, limited scalability, and lack of computational power, traditional chain-based
blockchains cannot be deployed in the systems associated with
EVs. The DAG-based blockchain (Tangle) [6] emerged as a
new type of blockchain, which reduces the reliance on computational power and improves the throughput significantly.
Huang et al. [17] presented a DAG-based blockchain with
a credit-based consensus mechanism for power-constrained
IoT devices. Hassija et al. [18] exploited the DAG-based
blockchain to support the increasing number of transactions in
the vehicle-to-grid network. In this paper, our BCS system is
built on DAG-based blockchain as well because of its limited
computational power and high frequency characteristics
-----
Auction theory has been widely applied in many different
fields, such as mobile crowdsensing [19] [20] [21], mobile
cloud computing [8] [9] [22], and energy trading [23] [24].
Here, we only focus on double auctions, where buyers (resp.
sellers) submit their bids (resp. asks) to an auctioneer. For
example, Borjigin et al. [25] proposed a double auction
method to maximize the profit of participants that can be
applied in service function chain routing and NFV price
adjustment. There are two classic models: McAfee double
auction [26] and Vickrey-based model [27]. They only considered homogeneous items and the Vickrey-based model cannot
satisfy the truthfulness, unfortunately. To heterogeneous items
(multi-item), Yang et al. [7] designed a truthful double auction
scheme for the cooperative communications, where the auctioneer first uses an assignment algorithm based on its design
goal to get candidates and mapping from buyer to seller, then
apply McAfee double auction to determine the winners and
corresponding transaction prices.
According to [7], Jin et al. [8] [9] proposed a truthful resource auction mechanism in mobile cloud computing, which
is the most relevant to our paper and has been introduced in
Sec. I. However, there are obvious differences in our work. In
our constrained multi-item double auction model, the number
of buyers that can be assigned to the same seller is constrained.
Furthermore, we should not only consider the unit bids of
buyers, but also consider their charging amounts. It is more
difficult to achieve truthfulness and system efficiency, which
is our main contribution to this paper.
III. BACKGROUND FOR DAG-BLOCKCHAINS
Blockchain is an emerging technology that acts as a decentralized ledger or database since Nakamoto published his
original prototype in 2008 [28], which is implemented by
modern cryptographic technologies and distributed consensus
algorithms. Because of these technical characteristics, it takes
more than half of the computational power to tamper with
transactions in the blockchain, which prevents attacks from
malicious nodes effectively. In the process of communication
and storage, All transactions are encrypted digitally, which realizes the anonymity and privacy protection of the blockchain
system. They can enable users who do not trust each other
to trade freely in a secure and reliable environment without
a third-party platform. Then, a number of real applications
flourished, such as Bitcoin [28], Ethereum [29], and Hyperledger fabric [30], all of which were based on the chain-based
blockchain.
The chain-based blockchain is shown in Fig. 1 (a), where
a block contains a certain number of transactions. All users
maintain the longest main chain jointly, and only the transactions in the main chain can be considered legal. However,
this chain-based structure is limited by its high requirement
for computational power because of its consensus mechanism
based on hashing puzzles, thereby it cannot be applied to
devices with limited power. Secondly, it generates and verifies
the blocks in a sequential and synchronous manner, resulting in
unacceptable confirmation delay and scalability. Take Bitcoin
as an example its average throughput is about 7 TPS (Trans
(a) The chain-based blockchain
(b) The DAG-based blockchain
Fig. 1. The comparison of chain-based and DAG-based blockchian, where
the square nodes in (a) are blocks and the circle nodes in (b) are transactions.
_actions Per Second) [31]. Thus, it is not suitable for those real-_
time application scenarios with high-frequency characteristics.
Here, the system we design is based on EVs which are both
resource-limited and high-frequency, hence the chain-based
structure is difficult to achieve our goal.
To improve the scalability, a new blockchain architecture
based on the Directed Acyclic Graph (DAG) was proposed,
called DAG-based blockchain or tangle [6]. The DAG-based
blockchain is shown in Fig. 1 (b), where each node represents
a transaction instead of a block. When a new transaction is
issued, it must validate two tips which are previous transactions attached in the tange but not verified by any others.
This validation is denoted by a directed edge in the tangle.
This new transaction will be verified by the other upcoming
transactions. The time required to verify a tip (confirmation
delay) depends on the tip selection algorithms [6] and the rate
of new transactions.
Generally, there is a weight associated with each transaction,
which is proportional to the difficulty of the hashing puzzle
defined by itself. When issues a new transaction, it has to find
a random number nouce such that
Hash(Transaction, Timestamp, nouce) Target (1)
_≤_
where the smaller the target implies the greater the weight.
Until now, the newly issued transaction has been completed
and waits to be verified by subsequent transactions. When
can we consider a transaction valid? This is related to its
cumulative weight. The cumulative weight of a transaction
is the weighted sum of transactions that approve it directly
or indirectly. Shown as Fig.1(b), the cumulative weight of
transaction A equals the weighted sum of transactions A,
_B, C, D, E, F_, and G. Consider a transaction, the larger
cumulative weight can only be achieved by consuming more
computational power, thereby it is more likely to be legal if
it has a larger cumulative weight. A transaction is believed to
be legal if and only if its cumulative weight exceeds a predefined threshold Suppose that most power is controlled by
-----
legal users, we can distinguish those transactions issued by
malicious users since the valid transactions will be verified by
other legal users and their cumulative weight will be larger
and larger.
Different from those synchronous consensus mechanisms
in the chain-based blockchain, the consensus process in the
DAG-based blockchain is completed in an asynchronous
approach. This eliminates an inherent defect of the chainbased blockchain that has on consensus finality because of
forking and pruning [32]. Besides, it is able to defend against
possible attacks, such as a single point of failure, Sybil attack,
lazy tips, and double-spending. According to that, the DAGbased blockchain can not only provide us with a secure and
reliable trading environment but also improve the scalability
and reduce the requirements on the computational power of
devices. Then, a lot of real systems based on DAG-based
blockchains emerged, such as IOTA [6], Byteball [33], and
Nano [34]. Finally, the devices in our charging scheduling
system are resource-limited and trade with others frequently.
Therefore, the DAG-based blockchain is an ideal choice to act
as infrastructure to achieve security and privacy protection.
IV. BLOCKCHAIN-BASED CHARGING SCHEDULING
SYSTEM
In this section, we demonstrate the overview of our
Blockchain-based Charging Scheduling (BCS) system by introducing entities and a charging scheduling framework.
_A. Entities for BCS System_
Consider a smart area, there are a certain number of CSs
with charging piles available to charge the EVs in this area.
Then, it exists a manager that is responsible for managing
entities and executing the charging scheduling between EVs
and CSs. Therefore, there are three main entities in this system
shown as follows.
_• Electric Vehicle (EV): The EVs running in this area play_
the part of requesters. When it is low on power, the EV
will request for charging services from the manager.
_• Charging Station (CS): The CSs located in this area_
play the part of providers to charge the EVs. A CS has
many charging piles, and each charging pile can charge
one EV at a time. When it is idle, the CS will inform the
manager of its status information.
_• Manager: The manager works as an energy scheduler._
Each EV sends a request about its charging preference
and each CS submits its status information on how many
charging piles are available to the manager. Then, the
manager acts as an auctioneer to perform a constrained
double auction mechanism between EVs and CSs that
assigns EVs to CSs and determines the transaction prices.
The price charged to EV and payment rewarded to CS
are determined by the transaction prices.
Thus, the BCS system can be denoted by B = {M, V, C}
where V = {V1, V2, · · · } is the set of EVs, C = {C1, C2, · · · }
is the set of CSs, and M is manager. Then, we define a
transaction T between V ∈ V and C ∈ C as the charging
Fig. 2. The architecture of blockchain-based charging scheduling system in
a smart area.
trading and digital payment record between them. This transaction is stored in the blockchain and includes pseudonyms of
_Vi and Cj, data type, transaction details, and timestamps of_
generation. In order to ensure security and privacy protection,
the transaction is encrypted by their digital signatures and the
payment is made in the form of charging coins.
_B. Architecture for BCS system_
The infrastructure of BCS system B = {M, V, C} is
established on the DAG-based blockchain, where each aforementioned entity is a node in this system. Depending on
their abilities towards computation and storage, they can be
split into two categories: light node and full node. Light
nodes have limited computational power and memory space,
thereby they are only responsible for generating transactions
together with full nodes together. Only part of their own
information is stored so that they can check it conveniently.
Full nodes usually possess powerful servers with multiple
functions, which can issue new transactions by finding a valid
nonce and validating the previous tips. Moreover, they are
responsible for storing and maintaining the entire blockchain,
namely the tangle.
The architecture of our BCS system is shown in Fig. 2. In
our BCS system B, each EV Vi ∈ V works as a light node,
and each CS Cj ∈ C works as a full node. The manager M
is a specific full node that manages the entire system to make
sure it works instead of storing and maintaining the tangle. For
the full nodes, the difficulty of hashing puzzle can be set by
modifying different target values dynamically to adapt to their
own computational power. For the manager, in addition to the
above-mentioned function that carries out constrained multiitem double auction between EVs and CSs as an auctioneer,
it has the right to add or delete nodes in real-time according
to the actual situation. For example, it can permit new CSs
to join this system and remove some malicious EVs from this
system. Besides, it can block those invalid requests from the
nodes within the system and prevent various attacks from the
nodes outside the system in advance. Also, the architecture
of our BCS system is based on the DAG-based blockchain,
which is not only distributed and reliable to defend against
attacks but also improves the throughput to be qualified for
high frequency energy trading
-----
Fig. 3. The procedure and message flow of charging scheduling between CSs
and EVs in a smart area.
In our BCS system, we use an asymmetric cryptography,
such as the elliptic curve digital signature algorithm [35], for
system initialization. Each EV Vi ∈ V registers on a trusted
authority to be a legitimate node through obtaining a unique
identification IDi that is associated with its license plate and
a certificate Ceri that is signed by the private key of authority
to certify the authenticity of this identity. After verifying its
certificate, the EV Vi can join this system and be assigned
with a public/private key pair (PKi, SKi) where its public
key works as a pseudonym that is open to all nodes and its
private key that is kept by itself. In asymmetric encryption,
the message encrypted by the public key can be decrypted by
a corresponding private key, and vice versa. After joining this
system, there is a set of wallet address {Wi(k)}k[θ]=1 [owned by]
the EV Vi, where we assume there are θ wallets for each entity.
Thus, the account of each EV Vi ∈ V that joins this system can
be denoted by AEi = {IDi, Ceri, (PKi, SKi), {Wi(k)}k[θ]=1[}][.]
Similarly, the account of each CS Cj ∈ C can be denoted by
_CEj = {IDj, Cerj, (PKj, SKj), {Wj(k)}k[θ]=1[}][.]_
_C. Charging Scheduling Framework_
The detailed procedure and message flow of charging
scheduling between EVs and CSs are shown in Fig. 3. We
assume it performs in discrete times. Here, we only consider
the situation in the time slot t 1, 2,, thus we ignore
_∈{_ _· · · }_
the time superscript t in the following variables. By adding
this superscript t, we can design online scheduling further
according to the actual demand. The specific operations can
be divided into 7 steps in detail:
**1) Request and Status: Each EV Vi** _∈_ V sends a
request message Ri that includes its bid for charging at
each CS Cj ∈ C to the manager. This request message is
denoted by ReqMsg = {PKM (SKi(Ri)), Ceri, STime}
where PKM is the public key of the manager and STime is
the timestamp of this message generation. At the same time,
each CS Cj C sends a status message Sj that includes
_∈_
its ask for serving a vehicle and the number of available
charging piles to the manager. This status message is denoted
by StaMsg = {PKM (SKj(Sj)), Cerj, STime}. Here, the
request and status message are encrypted by the manager’s
public key PKM since they are only allowed to be read by
the manager for privacy protection and fair trading
**2) Scheduling: The manager waits to receive request**
messages from EVs and status messages from CSs. After
collecting them, the manager confirms their legitimate identity
by verifying their certificates. Then, it works as a scheduler to
assign each winning EV to a CS that has unoccupied charging
piles. Besides, it determines the price charged to EV and
payment rewarded to CS as well, which is executed by the
built-in smart contract. The smart contract is implemented by
the constrained multi-item double auction model explained in
the following Sec.V and VI.
**3) Order and Assignment: The manager sends an order**
message Oi to each winning EV Vi V and an assignment
_∈_
message Aj to each winning CS Cj C. The Oi includes a
_∈_
CS that can serve to it and the price charged to it, which is
denoted by OrdMsg = {PKi(SKM (Oi)), CerM _, STime}._
The Aj includes an assignment, the set of EVs that can be
charged in Cj and the payment rewarded to it, which is
denoted by AssMsg = {PKj(SKM (Aj)), CerM _, STime}._
Here, the order and assignment message are encrypted by their
public key PKi and PKj because of permitting to be read by
themselves.
**4) Confirm: If EV Vi receives an order message from the**
manager, it implied it can be charged at the CS Cx designated
by the manager. Then, the EV Vi sends a confirm message
_Fi like “I will come on time” to the designated CS Cx. It_
is denoted by ConMsg = {PKx(SKiFi), Ceri, STime}
encrypted by Cx’s public key PKx for similar reasons.
**5) Charging: Once CS Cx receives the confirm message**
from the EV Vi, it will check whether the Vi is in its
assignment Ax. If yes, the CS Cx can provide charging service
to EV Vi before the deadline. After charging, the EV Vi
generates a new transaction Tix according to their trading
information, signs, and sends the SKi(Tix) back to Cx.
**6) Transactions: When CS Cx receives the new transaction**
from the EV Vi, it will check and sign this transaction by its
private key as well. Now, the CS Cx issues this new transaction
_SKx(SKi(TXix)) in the DAG-based blockchain. It is able to_
adjust the difficulty of the hashing puzzle by setting different
targets dynamically according to its computational power and
transaction frequency.
**7) Verification and payment: After the transaction Tix is**
issued, it will be verified to be legal in the future when its
cumulative weight is large enough. At this moment, charging
coins should be transferred from the wallet of Vi to Cx. The
coins with the price charged to Vi will be deducted from the
wallet Wi(k) and coins with the payment rewarded to Cx will
be added into the wallet Wx(k) permanently.
V. PROBLEM FORMULATION
In our BCS system B = {M, V, C}, CSs in C provide
charging piles for EVs in V that need charging. Each CS
_Ci ∈_ C has limited charging piles, where the number of
charging piles in this charging station is ki ∈ Z+. In general, CSs are distributed evenly across this smart area, also
those located in the area center are usually more crowded.
Furthermore, CSs have different charging efficiencies, where
higher efficiency means shorter charging time Thus there are
-----
two critical attributes, location and efficiency, associated with
each CS, which determine the valuation of an EV toward it.
The valuation of an EV toward a CS can be decided according
to its requirement. For example, when the battery of an EV
is very low, it values high a CS that is nearest to it. But for
an EV in a hurry, it considers both location and efficiency to
minimize its charging time. In the trading between EVs and
CSs, we aim to incentivize CSs to provide charging services
and meet the demands of EVs. To benefit both EVs and CSs,
we design a constrained multi-item double auction model that
gets a truthful assignment between EVs and CSs.
_A. Constrained Multi-Item Double Auction Model_
Shown as Fig. 3, we assume that this system runs in discrete
times. At each time step, EVs send request messages and CSs
send status messages privately to the managers. Based on the
single-round multi-item double auction model [7], EVs are
buyers and CSs are sellers in this auction. The manager M ∈ B
works as the trusted third auctioneer to assign n buyers to
_m sellers and determine the price charged to each buyer and_
payment rewarded to each seller.
The set of buyers is V = {V1, V2, · · ·, Vn} and the set of
sellers is C = {C1, C2, · · ·, Cm}. For each buyer Vi ∈ V,
its bid vector is denoted by Bi = (b[1]i _[, b]i[2][,][ · · ·][, b]i[m][)][ where][ b]i[j]_
is the unit bid (maximum buying price per unit charging) of
_Vi for charging at seller Cj_ C. Additionally, we define a
_∈_
charging vector R = (r1, r2, · · ·, rn) where ri is the charging
amount of Vi. For the sellers in C, the ask vector is denoted by
**_A = (a1, a2, · · ·, am) where aj ∈_** **_A is the unit ask (minimum_**
selling price per unit charging) of Cj. As for the number of
available charging piles in each CS, we define a vector K =
(k1, k2, · · ·, km) where kj ∈ Z+ is the number of piles that
can charge EVs in Cj ∈ C. Here, we notice that the bids of a
buyer vary with sellers since each EV has different evaluations
on CSs according to its requirements regarding location and
efficiency. However, the ask of a seller remains unchanged
among buyers because it is only concerned about payment
from charging vehicles.
By aforementioned definitions, the request message sent by
buyer Vi is denoted by Ri = (Bi, ri) and the status message
sent by seller Cj is denoted by Sj = (aj, kj). After it gets
the collection (B, R, A, K) where B = (B1; B2; · · · ; Bn),
the auctioneer determines the winning buyer set Vw ⊆ V, the
winning seller set Cw C, a mapping from Vw to Cw that is
_⊆_
_σ : {i : Vi ∈_ Vw} →{j : Cj ∈ Cw}, the unit price ˆpi charged
to buyer Vi Vw, and the unit payment ¯pj rewarded to seller
_∈_
_Cj. The assignment Aj for each Cj_ Cw is
_∈_
_Aj = {Vi ∈_ Vw : σ(i) = j} where |Aj| ≤ _kj_ (2)
because the CS Cj permits at most kj EVs to be charged at
the same time, which is the reason why this model is called
“constrained” double auction. Moreover, for each buyer Vi ∈
V, its valuation vector is denoted by Vi = (vi[1][, v]i[2][,][ · · ·][, v]i[m][)]
where vi[j] [is its unit valuation of][ V][i][ for charging at seller][ C][j][ ∈]
C. For the sellers in C, the cost vector is denoted by C =
(c1, c2, · · ·, cm) where cj ∈ **_C is the unit cost of Cj to provide_**
charging service Based on the buyer’s valuation and seller’s
Besides, in our model Ψ, we need to guarantee that the
assignment for each winning seller is not more than its
number of charging piles, that is |Aj| ≤ _kj for Cj ∈_ Cw.
_• Computational Efficiency: The auction results, includ-_
ing winning buyers, winning sellers, mapping from winning, price charged to buyer, and payment rewarded to
seller, can be obtained in polynomial time.
In addition to the above three properties, there are two
more important properties that should be satisfied strictly or
approximately.
_• Truthfulness: A double auction is truthful if every buyer_
(resp. seller) bids (resp. asks) truthfully is one of its
dominant strategies that can maximize its utility. That
is to say, no buyer can increase its utility by giving a bid
that is different from its true valuation and no seller can
increase its utility by giving an ask that is different from
its true cost. Consider our model Ψ, we have ˆui can be
maximized by bidding Bi = Vi for each Vi V and ¯uj
_∈_
can be maximized by asking aj = cj for each Cj C
_∈_
when other players do not change their strategies.
_• System Efficiency: There are a number of different_
metrics to evaluate the system efficiency of a double
auction model. The most common approach [27] is the
number of completed trades, that is the number of buyers
in the winning buyer set Vw in our model Ψ. Here, each
buyer V ∈ V will be assigned to a seller C _∈_ C
cost, the utility ˆui of winning buyer Vi Vw and the utility
_∈_
_u¯j of winning seller Cj_ Cw can be defined as follows:
_∈_
_uˆi = (vi[σ][(][i][)]_ _−_ _pˆi) · ri_ (3)
�
_u¯j = (¯pj −_ _cj) ·_ _Vi∈Aj_ _[r][i]_ (4)
Otherwise, for losing buyer Vi /∈ Vw and losing seller Cj /∈
Vw, their utilities are ˆui = 0 and ¯uj = 0. Here, the utility
_uˆi is proportional to the difference between its valuation and_
charged price, which implies the satisfaction level of Vi on its
assigned CS. The utility ¯uj is proportional to the difference
between rewarded payment and its cost, which characterizes
the profitability of Cj for providing charging service.
_B. Design Rationales_
The constrained multi-item double auction model defined in the last subsection can be denoted by Ψ =
(V, C, B, R, A, K). A valid double auction mechanism has
to meet the following three properties first, they are
_• Individual Rationality: The price charged to the winning_
buyer is not more than its bid and the payment rewarded
to the winning seller is not less than its ask. Consider
our model Ψ, we have ˆpi ≤ _b[σ]i_ [(][i][)] for each Vi ∈ Vw and
_p¯j_ _aj for each Cj_ Cw.
_≥_ _∈_
_• Budget Balance: The total price charged to all winning_
buyers is not less than the total payment rewarded to
all winning sellers, which ensures the profitability of the
auctioneer. Thus,
� � �
(5)
_Vi∈Vw_ _[p][ˆ][i][ ·][ r][i][ −]_ _Cj_ _∈Cw_ _[p][¯][j][ ·]_ _Vi∈Aj_ _[r][i][ ≥]_ [0]
-----
**Algorithm 1 TMC (V, C, B, R, A, K)**
**Input: V, C, B, R, A, K**
**Output: Vw, Cw, σ,** P[ˆ]w, P[¯]w
1: (Vc, Cc, ajψ ) ← TMC-WCD (V, C, B, A)
2: (Vw, Cw, σ, P[ˆ]w, P[¯]w) ← TMC-AP (Vc, Cc, ajψ _, R, K)_
3: return (Vw, Cw, σ, P[ˆ]w, P[¯]w)
**Algorithm 2 TMC-WCD (V, C, B, A)**
**Input: V, C, B, A**
**Output: Vc, Cc, ajϕ**
1: Vc ←∅, Cc ←∅
2: Construct a set V[′] = {Vst : b[t]s _[>][ 0][, V][s]_ _[∈]_ [V][}][ based on][ B]
3: Sort the buyers in V[′] and transfer it to an order list V[′] =
_⟨Vs1t1_ _, Vs2t2_ _, · · ·, Vsxtx_ _⟩_ such that b[t]s[1]1 _[≥]_ _[b]s[t][2]2_ _[≥· · · ≥]_ _[b]s[t][x]x_
4: Sort the sellers, get an order list C[′] = ⟨Cj1 _, Cj2_ _, · · ·, Cjm⟩_
such that aj1 _aj2_ _ajm_
5: Find the median ask ≤ _≤· · · ≤ ajϕ of C[′], ϕ =_ � _m2+1_ �
6: Find the minimum φ from V[′] such that bs[t][φ]φ[+1]+1 _[< a]jϕ_
7: V[′′] _←⟨Vs1t1_ _, Vs2t2_ _, · · ·, Vsφtφ⟩_
8: for each Vst ∈ V[′′] **do**
9: **if at < ajϕ then**
10: Vc ← Vc ∪{Vst}
11: **if Ct /∈** Cc then
12: Cc ← Cc ∪{Ct}
13: **end if**
14: **end if**
15: end for
16: return (Vc, Cc, ajϕ )
which satisfies the requirements of both buyer and seller.
To maximize it, this is in line with our original intention
of designing this system to make as many EVs as possible
to be charged. Other metrics, such as total price charged
to winning buyers, total payment rewarded to winning
seller, and profit of auctioneer, should be considered as
well based on needs. We will analyze them later.
To the truthfulness, we assume the submitted vector R and
**_K are trusted and cannot be tampered with because they can_**
be monitored by reliable hardware. Thereby, we only consider
the bids B and asks A to analyze the truthfulness of model
Ψ. When it is truthful, the double auction model avoids being
manipulated maliciously due to the fact that each player can
get the best utility by telling the truth. There is no player that
has the motivation to lie since they do not have to adapt to
others’ strategies by telling the lie for improving their utilities.
Therefore, truthfulness simplifies the strategic decisions for
players and makes sure a fair market environment, which plays
an important role in mechanism design.
VI. TRUTHFUL MECHANISM FOR CHARGING
In this section, we design a Truthful Mechanism for Charging (TMC) based on our constrained multi-item double auction
model and analyze whether it satisfies the desired properties
mentioned in Sec V B
**Algorithm 3 TMC-AP (Vc, Cc, ajϕ** _, R, K)_
**Input: Vc, Cc, ajϕ** _, R, K_
**Output: Vw, Cw, σ,** P[ˆ]w, P[¯]w
1: Vw ←∅, Cw ←∅, P[ˆ]w ←∅, P[¯]w ←∅
2: Sort the buyers in Vc, get an ordered queue Qc =
_⟨Vs1t1_ _, Vs2t2_ _, · · ·, Vsyty_ _⟩_ such that b[t]s[1]1 _[·][ r][s]1_ _[≥]_ _[b]s[t][2]2_ _[·][ r][s]2_ _[≥]_
_· · · ≥_ _b[t]s[y]y_ _[·][ r]sy_
3: Create a tentative set Hj ←∅ from each Cj ∈ Cc
4: K[′] = (k1[′] _[, k]2[′]_ _[,][ · · ·][, k]m[′]_ [)][ copied from vector][ K]
5: while Qc ̸= ∅ **do**
6: _Vsltl_ Qc.pop(0) // Obtain the first element in the
_←_
_queue Qc, then remove it from Qc._
7: **if kt[′]l** _[>][ 0][ then]_
8: Htl ← Htl ∪{Vsl _}, kt[′]l_ _[←]_ _[k]t[′]l_ _[−]_ [1]
9: _pˆsltl_ _ajϕ_
_←_
10: **else**
11: **for each Vs ∈** Htl do
12: _pˆstl ←_ max{ajϕ _, b[t]s[l]l · (rsl_ _/rs)}_
13: **end for**
14: **for each Vstl ∈** Qc do
15: Qc ← Qc\{Vstl _}_
16: **end for**
17: **end if**
18: end while
19: Vw ←{Vs : ∃j(Cj ∈ Cc ∧ _Vs ∈_ Hj)}
20: for each Vs ∈ Vw do
21: Is ←{Ct : Vs ∈ Ht, Ct ∈ Cc}
22: Find Ctm ∈ arg maxCt∈Is _{uˆst = (vs[t]_ _[−]_ _[p][ˆ][st][)][ ·][ r][s][}]_
23: _σ(s) = tm_
24: _pˆs ←_ _pˆstm_ _,_ P[ˆ]w ← P[ˆ]w ∪{pˆs}
25: **if Ctm /∈** Cw then
26: Cw ← Cw ∪{Ctm}
27: **end if**
28: end for
29: for each Ct ∈ Cw do
30: _p¯t ←_ _ajϕ_ _,_ P[¯]w ← P[¯]w ∪{p¯t} // Payments
31: end for
32: return (Vw, Cw, σ, P[ˆ]w, P[¯]w)
_A. Algorithm Design_
The process of TMC is shown in Algorithm 1, where it is
composed of two sub-processes, Winning Candidate Determination (TMC-WCD) shown in Algorithm 2 and Assignment &
Pricing (TMC-AP) shown in Algorithm 3. In TMC, we select
the winning candidates, then assign winning buyer candidates
to winning seller candidates truthfully. At the same time, the
price charged to winning buyer and payment rewarded to
winning seller can be determined.
Shown as Algorithm 2, in the process of winning candidate
determination, we sort the buyers in descending order based
on their bids for different sellers, where each buyer Vs ∈ V
is replaced with a buyer set {Vst : b[t]s _[>][ 0][, C][t]_ _[∈]_ [C][}][ in which]
buyer Vs gives a positive bid to seller Ct. The sellers are sorted
in ascending order based on their ask. The median of asks from
sellers ajϕ is selected as a threshold [8] to control the number
of buyer and seller candidates Let φ satisfy b[t][φ] _a_ and
_≥_
-----
_b[t]s[φ]φ[+1]+1_ _[< a]jϕ_ [, buyer][ V]st [will be a winning buyer candidate if]
its bid b[t]s [is not less than][ b]s[t][φ]φ [and the ask of requested seller]
_at is less than the threshold ajϕ_ . Seller Ct will be a winning
seller candidate if its ask at is less than ajϕ and there exists
at least one winning buyer candidate bidding for it.
Shown as Algorithm 3, in the process of assignment &
pricing, we sort the Vc in descending order based on their
total bids, which is the unit bids multiplied by their charging
amounts. The total bid of buyer Vst is b[t]s _[·][ r][s]_ [definitely. For]
each winning buyer candidate Vst ∈ Vc, it has met the basic
conditions for closing a deal because there is a winning seller
candidate Ct ∈ Cc with b[t]s _[> a][t]_ [and][ a][t] _[< a][j]ϕ_ [. For each]
seller Ct ∈ Cc, we assign the buyers with larger total bids to
it in priority. Thereby, there is a “tentative set” Ht associated
with each Ct Cc, which contains at most kt buyers with
_∈_
maximum total bids to Ct in Qc. It can be implemented in
line 5-18. We denoted by ˆpst the unit price charged to buyer
_Vs that gets service from Ct. Similar, the utility ˆust for each_
buyer Vs ∈ Ht can be defined as ˆust = (vs[t] _[−]_ _[p][ˆ][st][)][ ·][ r][s][;]_
Otherwise, the utility is ˆust = 0 for the buyer Cs /∈ Ht.
For example, let ⟨Vo1t, · · ·, Vozt⟩⊆ Qc be all buyers in
Qc that bid for Ct with b[t]o1 1 _oz_ _z_ [. If]
_[·][ r][o]_ _[≥· · · ≥]_ _[b][t]_ _[·][ r][o]_
_kt ≥_ _z, we have Ht = ⟨Vo1t, · · ·, Vozt⟩_ and ˆpoit = ajϕ for
each Voit ∈ Ht; Else, we have Ht = ⟨Vo1t, · · ·, Vokt _t⟩_ and
_pˆoit = max{ajϕ_ _, b[t]okt_ +1 _[·][ (][r][o]kt_ +1 _[/r][o]i_ [)][}][ for each][ V][o]i[t] _[∈]_ [H][t] [to]
guarantee truthfulness. Then, for each winning buyer Vs ∈ Vw,
it can be assigned to one of the seller in Is. The Vs selects
the optimal seller Ctm Is such that ˆustm _uˆst for each_
_∈_ _≥_
_Ct_ Is. Now, the mapping is σ(s) = tm and the charged
_∈_
price is ˆps = ˆpstm . Finally, the payment rewarded to winning
seller in Cw is given by ajϕ unanimously.
_B. Properties of TMC_
Next, we argue that our proposed TMC mechanism satisfies individual rationality, budget balance, computational
efficiency, and truthfulness.
**Lemma 1. The TMC is individually rational.**
_Proof. For each winning buyer Vs_ Vw and winning seller
_∈_
_Ct ∈_ Cw, we need to show that the charged price ˆps ≤ _b[σ]s_ [(][s][)]
and the rewarded payment ¯pt ≥ _at. According to Algorithm 2,_
it must be at < ajϕ if Ct ∈ Cw ⊆ Cc. Thereby the payment
rewarded to winning seller ¯pt = ajϕ is larger than its ask at
for each Ct ∈ Cw. Consider a winning buyer Vs ∈ Vw, the
charged price is either ˆps = ajϕ or ˆps = bs[σ]l[(][s][)] _· (rsl_ _/rs)._
_• In the first case, it must be b[σ]s_ [(][s][)] _≥_ _b[q]p[φ]φ_ [if][ V]s _[∈]_ [V]w[. The]
price charged to winning buyer ˆps = ajϕ is not more than
its bid b[σ]s [(][s][)] definitely.
_• In the second case, we have b[σ]s_ [(][s][)] _·rs ≥_ _b[t]s[l]l ·rsl according_
to Algorithm 3. The price charged to winning buyer ˆps =
_b[σ]sl[(][s][)]_ _· (rsl_ _/rs) is not more than its bid b[σ]s_ [(][s][)].
Thus, the TMC is individually rational because all winning
buyers and sellers are individually rational.
**Lemma 2. The TMC is budget balanced.**
_Proof. Shown as Algorithm 3, each winninng buyer Vs ∈_ Vw
is assigned to exact one winning seller C ∈ C hence this is
since b[t]o _[·][ r][o]_ _[≥]_ _[v]s[t]_ _[·][ r][s][. Thus we have][ ˆ][u]st[′]_ _[<][ ˆ][u][st]_ [= 0][.]
**(b) The Vs is not a winning buyer when bidding truthfully:**
According to Algorithm 3 there is no such a H for C ∈ C
a many-to-one mapping from Vw to Cw and Vw = ∪Cj _∈Cw_ _Aj._
For each mapping σ(s) = t from winning buyer Vs to winning
seller Ct, we have ˆps ≥ _ajϕ = ¯pt. Based on (5), it can be_
shown that
�
(6)
_Vi∈Vw_ [(ˆ][p][i][ −] _[p][¯][σ][(][i][)][)][ ·][ r][i][ ≥]_ [0]
Besides, it is easy to see that the assignment Aj for each
winning seller Cj ∈ Cw satisfies |Aj| ≤ _kj._
**Lemma 3. The TMC is truthful.**
_Proof. Here, we need to show the truthfulness to sellers and_
buyers one by one as follows:
For each buyer Vs V, we need to show its utility ˆus when
_∈_
giving a truthful bid Bs = Vs is not less than its corresponding
utility ˆu[′]s [when given an untruthful bid][ B]s[′] _[̸][=][ V][s][. Afterward,]_
any notation xs and x[′]s [refer to the concepts given by bid][ B][s]
and Bs[′] [respectively.]
**(a) The Vs is a winning buyer when bidding truthfully:**
According to Algorithm 3, it wins the seller Ctm such that
maximizes its utility. Thereby, the utility ˆustm _uˆst for each_
_≥_
_Ct_ Is. First, for each seller Ct Is, it implies Vs Ht and
_∈_ _∈_ _∈_
giving an untruthful bid (b[t]s[)][′][ to seller][ C][t] [cannot increase the]
utility such that ˆu[′]st _[>][ ˆ][u][st]_ [since]
_• (b[t]s[)][′][ > v]s[t][: The charged price][ ˆ][p][′]st[(= ˆ][p][st][)][ will not be]_
changed because the Vs has been in Ht when bidding
_b[t]s[(=][ v]s[t][)][. Thus, we have][ ˆ][u][′]st_ [= ˆ][u][st][.]
_• (b[t]s[)][′][ < v]s[t][: The charged price][ ˆ][p][′]st[(= ˆ][p][st][)][ will not be]_
changed if the Vs is still in Ht when bidding (b[t]s[)][′][(][< v]s[t][)][.]
Thus, we have ˆu[′]st [= ˆ][u][st][. However, when this untruthful]
bid (b[t]s[)][′][ decreases to be lower than][ ˆ][p][st][, the][ V][s] [cannot]
be in Ht and ˆu[′]st [= 0][. Thus, we have][ ˆ][u]st[′] _[<][ ˆ][u][st][.]_
Then, for each seller Ct ∈ V\Is, it can be divided into two
sub-cases where giving an untruthful bid (b[t]s[)][′][ to seller][ C][t]
cannot increase the utility such that ˆu[′]st _[>][ ˆ][u][st]_ [as well. They]
are analyzed as follows:
_• Vst /∈_ Vc: If at ≥ _ajϕ_, it is impossible to make Vst be in
Vc according to Algorithm 2 regardless of what the (b[t]s[)][′]
is. Thus, we have ˆu[′]st [= ˆ][u][st] [= 0][. If][ a][t] _[< a][j]ϕ_ [but truthful]
bid b[t]s[(=][ v]s[t][)][ < a][j]ϕ [, the][ V][s] [has to increase its bid such]
that (b[t]s[)][′][ ≥] _[a][j]ϕ_ [in order to make][ V][st] [be in][ V][c][. At this]
time, we have
_uˆ[′]st_ [= (][v]s[t] _[−]_ _[p][ˆ]st[′]_ [)][ ·][ r][s] _[≤]_ [(][v]s[t] _[−]_ _[a][j]ϕ_ [)][ ·][ r][s] _[<][ 0]_ (7)
if Ct ∈ Is when bidding (b[t]s[)][′][ untruthfully, otherwise]
_uˆ[′]st_ [= 0][. Thus, we have][ ˆ][u]st[′] _[<][ ˆ][u][st]_ [= 0][.]
_• Vst ∈_ Vc but Vs /∈ Ht: In this case, we can know that the
Ht has been full where |Ht| = kt. From this, we have
_b[t]o_ _[·][ r][o]_ _[≥]_ _[b]s[t]_ [(=][ v]s[t][)][ ·][ r][s] [where][ V][o] [has minimum value of]
_b[t]o_ _[·][ r][o]_ [among all buyers in][ H][t][. The][ V][s] [has to increase]
its bid such that (b[t]s[)][′][ ≥] _[b][t]o_ _[·][ (][r][o][/r][s][)][ in order to replace]_
_Vo in Ht. Then, the charged price will be changed to_
_pˆ[′]st_ [=][ b]o[t]
_[·][ (][r][o][/r][s][)][. The utility is]_
� �
_uˆ[′]st_ [= (][v]s[t] _[−]_ _[p][ˆ]st[′]_ [)][ ·][ r][s] [=] _vs[t]_ _[−]_ _[b]o[t]r[r]s[o]_ _· rs < 0_ (8)
-----
that has Vs ∈ Ht. For each seller Ct ∈ C, we have utility
_uˆst = 0. Similarly, we need to show giving an untruthful bid_
(b[t]s[)][′][ to seller][ C][t] [cannot increase the utility such that][ ˆ][u][′]st _[>]_
_uˆst. The analysis about it can be divided into Vst /∈_ Vc and
_Vst ∈_ Vc but Vs /∈ Ht, which are the same as the analysis for
_Ct ∈_ V\Is in part (a). Thus, we have ˆu[′]st _[≤]_ _[u][ˆ][st]_ [= 0][.]
From above, the utility of Vs to each seller Ct C satisfies
_∈_
_uˆst ≥_ _uˆ[′]st_ [definitely. By selecting the the seller such that]
maximizes its utility, we have ˆus ≥ _uˆ[′]s[. Therefore, the buyers]_
are truthful.
For each seller Ct C, we need to show its utility ¯ut when
_∈_
giving a truthful ask at = ct is not less than its corresponding
utility ¯u[′]t [when given an untruthful ask][ a]t[′] _[̸][=][ c][t][. Afterward,]_
any notation xt and x[′]t [refer to the concepts given by bid][ a][t]
and a[′]t [respectively.]
**(c) The Ct is a winning seller when asking truthfully: Accord-**
ing to Algorithm 3, its ask at(= ct) < ajϕ and at least one
buyer are assigned to it. Hence, we have |At| > 0. Its utility
can be denoted by
**Algorithm 4 EMC (V, C, B, R, A, K)**
**Input: V, C, B, R, A, K**
**Output: Vw, Cw, σ,** P[ˆ]w, P[¯]w
1: (Vc, Cc, ajψ ) ← EMC-WCD (V, C, B, A)
2: (Vw, Cw, σ, P[ˆ]w, P[¯]w) ← EMC-AP (Vc, Cc, ajψ _, R, K)_
3: return (Vw, Cw, σ, P[ˆ]w, P[¯]w)
**Lemma 4. The TMC is computationally efficient.**
�
_u¯t = (ajϕ −_ _ct) ·_
(9)
_Vi∈At_ _[r][i][ >][ 0]_
_Proof. In Algorithm 2, there are nm buyers in V[′], hence it_
takes O(nm log(nm)) and O(m log(m)) times to sort the V[′]
and C[′] respectively. The size of V[′′] in line 7 is at most nϕ,
and the number of iterations in the for-loop (line 8-15) is at
most nϕ. Consequently, the time complexity of Algorithm 2
is O(nm log(nm)). In Algorithm 3, it takes O(nϕ log(nϕ))
times to sort the Vc. The number of iterations in the while-loop
(line 5-18) is at most n. The line 12 can be executed at most
�m
_i=1_ _[k][i][ times, thus it takes][ O][(][n][ +][ �]i[m]=1_ _[k][i][)][ time to execute]_
this while-loop. Besides, there are at most n buyers in Vw
and find the best (line 22) from at most ϕ sellers, thus it takes
_O(nϕ) time to execute the for-loop (line 20-28). Consequently,_
the time complexity of Algorithm 2 is O(nϕ log(nϕ)) and
overall time complexity of TMC is O(nm log(nm)).
**Theorem 1. The TMC is individually rational, budget bal-**
_anced, truthful, and computationally efficient._
_Proof. It can be derived from Lemma 1 to Lemma 4._
VII. EFFICIENT MECHANISM FOR CHARGING
Even though TMC is able to ensure truthfulness, it sacrifices
the system efficiency. Shown as Algorithm 3, suppose the
winning buyer Vs satisfies Vs ∈ Ht1 and Vs ∈ Ht2, it can
be assigned to only one seller t ∈{t1, t2}. Thus, another
seller will have a charging pile empty, which could be used to
charge other buyers. Thus, for each winning buyer Vs ∈ Vw
with |Is| > 1, there are |Is| − 1 charging piles being wasted.
To address this drawback, we propose an Efficient Mechanism
for Charging (EMC) to improve system efficiency and ensure
its truthfulness to some extent.
_A. Algorithm Design_
The process of EMC is shown in Algorithm 4. Similar to
TMC, it is composed of Winning Candidate Determination
(EMC-WCD) and Assignment & Pricing (EMC-AP). Here, the
EMC-WCD is the same as TMC-WCD shown in Algorithm
2, which can be used to generate a winning buyer candidate
set Vc and winning seller candidate set Cc. The EMC-AP is
shown in Algorithm 5.
Shown as Algorithm 5, in the process of assignment &
pricing, we sort the winning buyer candidates in Vc in descending order based on their total bids. Then, we give priority
to assigning the buyer that can give the maximum total bid,
which is the critical step to improve the system efficiency.
At each iteration, we pop the buyer that has the maximum
total bid from Qc, denoted by Vsltl and check whether the
seller C requested by V still have available charging piles
since ¯pt = ajϕ . Consider giving an untruthful ask a[′]t[, we can]
discuss as follows:
_• a[′]t_ _[≥]_ _[a][j]ϕ_ [: It loses this auction because the new median]
ask a[′]jϕ [is not less than][ a][j]ϕ [and][ a]t[′] _[≥]_ _[a]j[′]_ _ϕ_ _[≥]_ _[a][j]ϕ_ [. Thus,]
the utility is ¯u[′]t [= 0][ <][ ¯][u][t][.]
_• a[′]t_ _[< a][j]ϕ_ [: At the time, the new median ask][ a][′]jϕ [is equal]
to ajϕ and a[′]t _[< a]j[′]_ _ϕ_ [=][ a][j]ϕ [. Moveover, the assignment]
of Ct remains unchanged, A[′]t [=][ A][t][, and the rewarded]
payment ¯p[′]t [= ¯][p][t][. According to (9), the utility when]
asking untruthfully is the same as that when asking
truthfully. Thus, we have ¯u[′]t [= ¯][u][t][.]
**(d) The Ct is not a winning seller when asking truthfully: At**
this time, its utility when asking truthfully is ¯ut = 0. Here,
we need to analyze how the Ct loses this auction. If losing
since ct _ajϕ_, we have
_≥_
_• a[′]t_ _[< a][j]ϕ_ [: The new median ask][ a][′]jϕ [is not more than][ a][j]ϕ
and a[′]t _jϕ_ _ϕ_ [. If the][ C][t][ still loses the auction,]
_[≤]_ _[a][′]_ _[≤]_ _[a][j]_
its utility ¯u[′]t [= 0][. If the][ C][t] [wins now, its utility][ ¯][u][′]t _[≤]_ [0]
because of the rewarded payment ¯p[′]t [=][ a]j[′] _ϕ_ _[≤]_ _[a][j]ϕ_ _[≤]_ _[c][t][.]_
Thus, we have ¯u[′]t _[≤]_ _[u][¯][t]_ [= 0][.]
_• a[′]t_ _[≥]_ _[a][j]ϕ_ [: It loses this auction because the new median]
ask a[′]jϕ [is equal to][ a][j]ϕ [and][ a]t[′] _[≥]_ _[a][j]ϕ_ [=][ a]j[′] _ϕ_ [. Thus, the]
utility is ¯u[′]t [= ¯][u][t] [= 0][.]
If ct < ajϕ but still loses, there are two situations that no
buyer Vst gives a bid b[t]s [such that][ b]s[t] _[≥]_ _[a][j]ϕ_ [or utility][ ˆ][u][st] [is]
not the maximum one for each Vs ∈ Vw. Now,
_• a[′]t_ _[< a][j]ϕ_ [: The above two situations still happen because]
the new median ask a[′]jϕ [=][ a][j]ϕ [and][ a]t[′] _[< a]j[′]_ _ϕ_ [. Thus, we]
have ¯u[′]t [= ¯][u][t] [= 0][.]
_• a[′]t_ _[≥]_ _[a][j]ϕ_ [: It loses this auction because the new median]
ask a[′]jϕ [is not less than][ a][j]ϕ [and][ a]t[′] _[≥]_ _[a]j[′]_ _ϕ_ _[≥]_ _[a][j]ϕ_ [. Thus,]
the utility is ¯u[′]t [= ¯][u][t] [= 0][.]
Therefore, the sellers are truthful.
In summary, both buyers and sellers cannot improve the
utility by deviating from their valuations and costs.
-----
**Algorithm 5 EMC-AP (Vc, Cc, ajϕ** _, R, K)_
**Input: Vc, Cc, ajϕ** _, R, K_
**Output: Vw, Cw, σ,** P[ˆ]w, P[¯]w
1: Vw ←∅, Cw ←∅, P[ˆ]w ←∅, P[¯]w ←∅
2: Sort the buyers in Vc, get an ordered queue Qc =
_⟨Vs1t1_ _, Vs2t2_ _, · · ·, Vsyty_ _⟩_ such that b[t]s[1]1 _[·][ r][s]1_ _[≥]_ _[b]s[t][2]2_ _[·][ r][s]2_ _[≥]_
_· · · ≥_ _b[t]s[y]y_ _[·][ r]sy_
3: K[′] = (k1[′] _[, k]2[′]_ _[,][ · · ·][, k]m[′]_ [)][ copied from vector][ K]
4: while Qc ̸= ∅ **do**
5: _Vsltl_ Qc.pop(0)
_←_
6: **if ktl > 0 then**
7: _σ(sl) = tl, kt[′]l_ _tl_
_[←]_ _[k][′]_ _[−]_ [1]
8: Vw ← Vw ∪{Vsl _}, ˆpsl ←_ _ajϕ_ _,_ P[ˆ]w ← P[ˆ]w ∪{pˆsl _}_
9: **if Ctl /∈** Cw then
10: Cw ← Cw ∪{Ctl _}_
11: **end if**
12: **for each Vslt ∈** Qc do
13: Qc ← Qc\{Vslt}
14: **end for**
15: **else**
16: _Atl ←{Vs ∈_ Vw : σ(s) = tl}
17: **for each Vs ∈** _Atl do_
18: _pˆs ←_ max{ajϕ _, b[t]s[l]l · (rsl_ _/rs)} ∈_ P[ˆ]w
19: **end for**
20: **for each Vstl ∈** Qc do
21: Qc ← Qc\{Vstl _}_
22: **end for**
23: **end if**
24: end while
25: for each Ct ∈ Cw do
26: _p¯t ←_ _ajϕ_ _,_ P[¯]w ← P[¯]w ∪{p¯t} // Payments
27: end for
28: return (Vw, Cw, σ, P[ˆ]w, P[¯]w)
If yes, kt[′]l _[>][ 0][, the buyer][ V][s]l_ [will be assigned to seller]
_Ctl_ . Also, Vsl is a winning buyer, Ctl is a winning seller,
and the price charged to buyer Vsl is given by ˆpsl = ajϕ
tentatively. Furthermore, the requests from buyer Vsl to other
sellers should be deleted from Qc since the buyer Vsl has been
assigned. If no, kt[′]l [= 0][, the buyer][ V][s]l [will not be assigned]
to seller Ctl because there is no unoccupied charging piles in
_Ctl_ . It implies that in previous iterations, the buyers in Atl
have been assigned to seller Ctl, then the price charged to
each winning buyer Vs ∈ _Atl has to be changed to its critical_
price ˆps = b[t]s[l]l (rsl _/rs). Finally, the payment rewarded to_
_·_
winning seller is given by ajϕ unanimously.
_B. Properties of EMC_
Similar to Sec. VI-B, we analyze whether our EMC mechanism satisfies the aforementioned four properties.
**Lemma 5. The EMC is individually rational.**
_Proof. It can be discussed similar to proof of Lemma 1._
**Lemma 6. The EMC is budget balanced.**
_Proof It can be discussed similar to proof of Lemma 2_
**Lemma 7. The EMC is not truthful, but the truthfulness is**
_held for sellers._
_Proof. For a buyer Vs /∈_ Vw when bidding truthfully, giving
an untruthful bid (b[t]s[)][′][ to seller][ C][t] [cannot increase the utility]
such that ˆu[′]st _[<][ ˆ][u][st]_ [= 0][, which can be divided into][ V][st] _[∈][/]_ [V][c]
(similar to the analysis for Vst /∈ Vc in part (a) of Lemma 3)
and Vst ∈ Vc. Consider the case Vst ∈ Vc, it exists a Vot ∈ Qc
with b[t]o[·][r][o] _[≥]_ _[b]s[t]_ [(=][ v]s[t][)][·][r][s] [which is the last one can be assigned]
to seller Ct in Algorithm 5. To replace Vot, the Vs has to bid a
(b[t]s[)][′][ such that][ (][b]s[t] [)][′][ ≥] _[b]o[t]_ _[·][(][r][o][/r][s][)][ and the charged price will be]_
_pˆ[′]st_ [=][ b]o[t] _[·][(][r][o][/r][s][)][. From here, we have][ ˆ][u][′]st_ _st_ _s[.]_
_[≤]_ [0][ since][ ˆ][p][′] _[≥]_ _[v][t]_
For a buyer Vs ∈ Vw, we give two examples where giving an
untruthful bid is possible to improve its utility. There are two
sellers o1 and o2 requested by the Vs lie in Qc when it bids
truthfully, thus Qc = ⟨· · ·, Vso1 _, · · ·, Vslo1_ _, · · ·, Vso2_ _, · · · ⟩_
with vs[o][1] _[·][ r][s]_ _[≥]_ _[b]s[o]l[1]_ _[·][ r][s]l_ _[≥]_ _[v]s[o][2]_ _[·][ r][s][. The results returned by]_
Algorithm 5 are σ(s) = o1, ˆps = b[o]sl[1] _l_ _[/r][s][)][, and][ k]o[′]_ 2 _[>][ 0][.]_
_[·][ (][r][s]_
At this time, its utility is ˆus = (vs[o][1] _−_ _b[o]sl[1]_ _[·][ (][r][s]l_ _[/r][s][))][ ·][ r][s][.]_
When giving an untruthful bid (b[o]s[1] [)][′][ such that][ (][b]s[o][1] [)][′][ < v]s[o][2] [,]
the results are changed to σ[′](s) = o2, ˆp[′]s [=][ a][j]ϕ [. At this time,]
its utility is ˆu[′]s [= (][v]s[o][2] _[−]_ _[a][j]ϕ_ [)][ ·][ r][s][. We cannot judge which one]
is larger since vs[o][1] _s_ and b[o]sl[1] _l_ _[/r][s][)][ ≥]_ _[a][j]ϕ_ [. If][ ˆ][u][′]s _[>][ ˆ][u][s][,]_
_[≥]_ _[v][o][2]_ _[·][ (][r][s]_
its utility can be improved by bidding untruthfully. Similarly,
it can give an untruthful bid (b[o]s[2] [)][′][ such that][ (][b]s[o][2] [)][′][ > v]s[o][1] [, the]
results are changed to σ[′](s) = o2, ˆp[′]s [=][ a][j]ϕ [as well. Thus,]
truthfulness is not held for winning buyers.
The analysis for sellers are similar to the case (c) and (d)
in proof of Lemma 3, thus truthfulness is held for sellers.
**Lemma 8. The EMC is computationally efficient.**
_Proof. It can be discussed similar to proof of Lemma 4._
**Theorem 2. The EMC is individually rational, budget bal-**
_anced, and computationally efficient, but not truthful._
_Proof. It can be derived from Lemma 5 to Lemma 8._
In EMC, we attempt to assign each buyer in Vc to a seller
in a greedy manner, thus avoiding the waste of charging piles.
Therefore, it increases the number of winning buyers (successful trades) and improves system efficiency. Even though the
winning buyers are able to improve their utilities by bidding
untruthfully, this is difficult to achieve. In our BCS system,
each player bids or asks privately. The buyer cannot have
knowledge of other players’ strategies such as other buyers’
bids and sellers’ asks. Thus, it is not able to predict whether
it will win, which seller it will be assigned to, and its charged
price. For the buyer losing the auction, it is possible to get
a negative utility when bidding untruthfully. For the buyer
winning the auction, despite the potential of improving its
utility, there is also the possibility of losing the auction when
bidding untruthfully. Obviously, they have no motivation to lie
because the risks are great. Therefore, we can say the EMC
is truthful to some extent.
_C. A Walk-through Example_
To understand our TMC and EMC algorithms clearly and
compare their difference we give a walk through example
-----
TABLE I
AN EXAMPLE WITH 5 BUYERS AND 5 SELLERS.
**_B_** _C1_ _C2_ _C3_ _C4_ _C5_ **_R_**
_V1_ 0 4 0 5 2 5
_V2_ 2 0 5 1 0 2
_V3_ 7 5 0 4 0 6
_V4_ 6 4 0 3 0 4
_V5_ 0 0 2 3 5 3
**_A_** 4 1 3 2 5 **_K_** 1 2 4 2 3
with 5 buyers and 5 sellers. The bids and charging amounts
of buyers, the asks and number of charging piles of sellers are
shown in Table I.
In TMC-WCD (EMC-WCD) according to Algorithm 2,
the median of asks is denoted by ajϕ = a3 = 3. We can
get V[′] = ⟨V31, V41, V14, V23, V32, V55, V12, V34, V42, V44, V54⟩.
By removing those Vst V[′] with at _ajϕ_, we have
_∈_ _≥_
Vc = {V14, V32, V12, V34, V42, V44, V54} and Cc = {C2, C4}.
Then, sort the Vc according to their total bids, we have
Qc = ⟨V32 = 30, V14 = 25, V34 = 24, V12 = 20, V42 =
16, V44 = 12, V54 = 9⟩.
For TMC, in TMC-AP according to Algorithm 3, we have
tentative sets H2 = {V1, V3} with ˆp12 = max{ajϕ _, b[2]4_ _[·]_
(r4/r1)} = 3.2, ˆp32 = max{ajϕ _, b[2]4_ _[·][ (][r][4][/r][3][)][}][ = 3][ and]_
H4 = {V1, V3} with ˆp14 = max{ajϕ _, b[4]4_ _[·][ (][r][4][/r][1][)][}][ = 3][,][ ˆ][p][34]_ [=]
max{ajϕ _, b[4]4_ _[·][ (][r][4][/r][3][)][}][ = 3][. For the buyer][ V][1][, its utility]_
satisfies ˆu12 = (b[2]1 _[−]_ _[p][ˆ][12][)][ ·][ r][1]_ [= 4][ <][ 10 = ˆ][u][14][. For]
the buyer V3, its utility satisfies ˆu32 > ˆu34. Thus, we have
Vw = {V1, V3}, Cw = {C2, C4}, {σ(1) = 4, σ(3) = 2},
_Pˆw = {pˆ1 = 3, ˆp3 = 3}, and ¯Pw = {p¯2 = 3, ¯p4 = 3}._
For EMC, in EMC-AP according to Algorithm 5, we assign
buyer V3 to seller C2 with ˆp3 = 3 in the first iteration,
then Qc is revised to Qc = ⟨V14 = 25, V12 = 20, V42 =
16, V44 = 12, V54 = 9⟩. Repeat it until Qc = ∅, we have
Vw = {V1, V3, V4, V5}, Cw = {C2, C4}, {σ(1) = 4, σ(3) =
2, σ(4) = 2, σ(5) = 4}, _P[ˆ]w = {pˆ1 = 3, ˆp3 = 3, ˆp4 = 3, ˆp5 =_
3}, and _P[¯]w = {p¯2 = 3, ¯p4 = 3}. From this example, we can_
see that the winning sellers in TMC are not full where there
are idle charging piles not used to charge vehicles. Therefore,
the number of successful trades |Vw| = 2 in TMC is less than
_|Vw| = 4 in EMC, which explains the reason why the system_
efficiency of EMC is better than TMC.
VIII. NUMERICAL SIMULATIONS
In this section, we implement our TMC and EMC algorithms, evaluate their performances, and verify whether they
satisfy our design rationale separately.
_A. Simulation Setup_
To simulate our TMC and EMC, we consider a smart area
B = (M, V, C) with 1000 × 1000 km[2]. There are m CSs and
_n EVs distributed uniformly in this area, where we default by_
_n_ 10 m unless otherwise specified For each CS C ∈ C its
|B|C C C C C 1 2 3 4 5|R|
|---|---|---|
|V 1 V 2 V 3 V 4 V 5|0 4 0 5 2 2 0 5 1 0 7 5 0 4 0 6 4 0 3 0 0 0 2 3 5|5 2 6 4 3|
|A K|4 1 3 2 5 1 2 4 2 3|- -|
where (xi, yi) and (xj, yj) are the coordinates of Vi and Cj.
We assume the valuation of an EV to a CS is related to their
distance. Thus, the larger d[j]i [is, the lower][ v]i[j] [is. The maximum]
_√_
distance between two entities in this area is 1000 2, thus we
_√_
assume that vi[j] [= 1][ −] _[d]i[j][/][(1000]_ 2).
_B. Simulation Results and Analysis_
To evaluate individual rationality, budget balance, and
truthfulness, we consider a smart area with m = 10
CSs and n = 100 EVs. They can be denoted by C =
_{C1, C2, · · ·, C10} and V = {V1, V2, · · ·, V100}. The median_
ask is ajϕ = 0.764 and there are five winning sellers, thus
C _{C_ _C_ _C_ _C_ _C_ _} Moreover the corresponding_
(a) TMC
(b) EMC
Fig. 4. The assignment results and individual rationality obtained by our
TMC and EMC.
number of charging piles kj is generated from {1, 2, · · ·, 10}
randomly with probability 1/10. The cost cj of CS Cj is
generated according to a uniform distribution within (0, 1].
Similarly, for each EV Vi V, its charging amount ri is
_∈_
sampled from a truncated normal distribution with mean 50
and variance 1 in interval (0, 100]. To quantify its valuation
_vi[j]_ [to CS][ C][j][, we defined the distance][ d]i[j] [between][ V][i][ and][ C][j]
according to their coordinates, that is
�
_d[j]i_ [=]
(xi − _xj)[2]_ + (yi − _yj)[2]_ (10)
-----
(a) Buyer V50 ∈ Vw (b) Seller C1 ∈ Cw
(c) Buyer V86 /∈ Vw (d) Seller C3 /∈ Cw
Fig. 5. The truthfulness of buyers and sellers in TMC.
(a) Buyer V50 ∈ Vw (b) Seller C1 ∈ Cw
(c) Buyer V86 /∈ Vw (d) Seller C3 /∈ Cw
Fig. 6. The truthfulness of buyers and sellers in EMC.
number of charging piles of CSs in Cw is given by {r1 :
3, r2 : 8, r6 : 4, r7 : 3, r10 : 8}.
**Individual Rationality: Fig. 4 shows the assignment results**
and individual rationality obtained by TMC and EMC. The
first line from the bottom is sellers (CSs) and the second
line from the bottom is buyers (EVs). Take Fig. 4 (a) as an
example, for the seller C1, there are two buyers, V50 and
_V58, assigned to it in TMC. For the mapping σ(50) = 1,_
the payment rewarded to C1 (red column) is more than the
ask of C1 (grey column) and the price charged to V50 (green
column) is less than the bid of V (blue column) Then for
Fig. 7. The running time varies with the increasing number of CSs in TMC
and EMC.
Fig. 8. The number of successful trades (winning buyers) varies with the
increasing number of CSs in TMC and EMC.
any mappings from Vw to Cw in TMC and EMC, the price
charged to the winning buyer is not more than its bid and the
payment rewarded to the winning seller is not less than its
ask, thus individual rationality is held.
**Budget Balance: According to the charged price and**
rewarded payment shown as Fig. 4, the total price charged to
all winning buyers is not less than the total payment rewarded
to all winning sellers. Furthermore, the number of buyers |Aj|
assigned to each seller Cj Cw is not more than the number
_∈_
of charging piles kj, namely we have |Aj| ≤ _kj. Thereby the_
budget balance is held in both TMC and EMC.
**Truthfulness: We select a winning buyer V50 ∈** Vw, a
losing buyer V86 /∈ Vw, a winning seller C1 ∈ Cw, and a
losing seller C3 /∈ Cw as the representatives to evaluate the
truthfulness of buyers and sellers in our TMC and EMC. Fig.
5 and Fig. 6 show the truthfulness of buyers and sellers in
TMC and EMC. Let us look at Fig. 5 in TMC first. For the
winning buyer V50, σ(50) = 1, it can get the maximum utility
_uˆ50 = 9.932 when giving the truthful bid b[1]50_ [=][ v]50[1] [= 0][.][905][.]
Here, we have I50 = {C1, C10} and its utility cannot be
improved when changing the bids to the seller in I50. If
the bid b[1]50 _[<][ 0][.][77][, its utility will decrease to][ 1][.][819][ since]_
the V50 will not be selected in H1 and then be assigned
to seller C10. Besides, by changing the bids to sellers that
are not in V (C ) or in V but not in I (C ) its utility
-----
cannot be improved as well. For the winning seller C1, it
can get the maximum utility ¯u1 when giving the truthful ask
_a1 = c1 = 0.434, which cannot be improved by changing its_
ask. For the losing buyer V86, the utility ˆu86 is impossible to
be more than zero when bidding untruthfully. Its utility will
be negative if increasing the bids to sellers in Vc (C3 and C6).
For the losing seller C3, it achieves zero utility when giving
the truthful ask a3 = c3 = 0.896, which will be negative if
decreasing its ask. Next, let us look at Fig. 6 in EMC. We
have the same observations in sub-figures shown as (b), (c),
and (d). For the winning buyer V50, σ(50) = 1, it has a little
different from that in TMC. If increasing the bids to other
sellers in Vc (C6 and C10), its utility will decrease, even be
negative. This is because the V50 will be assigned to C6 or
_C10 instead of C1 since total bids have been varied._
To evaluate the computational efficiency and system efficiency, we consider a smart area whose number of CSs m
ranges from 0 to 200. The parameters are sampled according
to the rules described in the simulation setup.
**Computational Efficiency: Fig. 7 shows the running time**
comparison between TMC and EMC. We default by n =
10 _m, thereby the time complexity O(nm log(nm)) can_
_·_
be considered as 10 _m[2]_ approximately. The trends shown
_·_
in Fig. 7 are in line with our expectations and they are
computationally efficient. Besides, we can observe that the
running time of EMC is slightly lower than that of TMC since
there are more entities eliminated in advance.
**System Efficiency: Here, the system efficiency can be**
characterized by the number of successful trades between
buyers and sellers, which is equal to the number of winning
buyers in Vw because each winning buyer will be assigned to a
winning seller and then begin to trade. Fig. 8 shows the system
efficiency comparison between TMC and EMC. The system
efficiency is not monotone since we sample the parameters
used in this simulation at each number of sellers independently.
Shown as Fig. 8, we can observe that the system efficiency
in EMC is apparently better than that in TMC, which implies
our proposed EMC is an effective approach to improve system
efficiency even though it does not guarantee the truthfulness of
buyers in some extreme cases. Next, the gap between TMC and
EMC increases gradually as the number of sellers increases.
This is because the buyer who bids higher can take over more
tentative sets of candidate sellers in Algorithm 3. However,
it can be assigned to only one of these tentative sets, which
causes a lot of waste and enlarge the gap.
_C. Further Discussion_
According to Lemma 7, we have known that truthfulness
is not held for buyers in some extreme cases. A buyer cannot
predict an untruthful bid to improve its utility in a deterministic
manner because it does not know the bidding strategies of
other buyers. For the buyers, it is very risky and difficult to
increase their utilities by changing their bids. Our simulation
result, shown as Fig. 6, also proves this point that EMC
satisfies the truthfulness to some extent. Shown as Fig. 7 and
Fig. 8, EMC has a lower running time and a much better
system efficiency than TMC. Therefore, we prefer to use our
EMC instead of TMC in practical applications
IX. CONCLUSION
In this paper, a charging scheduling system based on
blockchain technology and a constrained multi-item double
auction model has been designed and implemented. To achieve
privacy protection and scalability, we gave a lightweight charging scheduling framework based on asymmetric encryption
and DAG-based blockchain. To incentivize EVs and CSs to
participate in the market, we considered a constrained multiitem double auction model and designed two algorithms, TMC
and EMC, that attempt to assign EVs (buyers) in this area to be
charged in CSs (sellers). Both algorithms are feasible, which
ensures individual rationality, budget balance, truthfulness, and
computational efficiency. Here, EMC can get a better system
efficiency than TMC, but it weakens the truthfulness of buyers
to some extent. Finally, the results of numerical simulations
indicated that our model is robust and theoretical analysis is
correct.
ACKNOWLEDGMENT
This work was supported in part by the National Natural Science Foundation of China (NSFC) under Grant No.
62202055 and No. 62202016, the Start-up Fund from Beijing
Normal University under Grant No. 310432104, the Start-up
Fund from BNU-HKBU United International College under
Grant No. UICR0700018-22, the Project of Young Innovative
Talents of Guangdong Education Department under Grant No.
2022KQNCX102, and the National Science Foundation (NSF)
under Grant No. 1907472 and No. 1822985.
REFERENCES
[1] Z. Xiong, H. Xu, W. Li, and Z. Cai, “Multi-source adversarial sample attack on autonomous vehicles,” IEEE Transactions on Vehicular
_Technology, vol. 70, no. 3, pp. 2822–2835, 2021._
[2] Z. Wang, Q. Hu, R. Li, M. Xu, and Z. Xiong, “Incentive mechanism
design for joint resource allocation in blockchain-based federated learning,” IEEE Transactions on Parallel and Distributed Systems, vol. 34,
no. 5, pp. 1536–1547, 2023.
[3] R. Zhang, R. Xue, and L. Liu, “Security and privacy on blockchain,”
_ACM Computing Surveys (CSUR), vol. 52, no. 3, pp. 1–34, 2019._
[4] J. Guo, X. Ding, and W. Wu, “A blockchain-enabled ecosystem for
distributed electricity trading in smart city,” IEEE Internet of Things
_Journal, vol. 8, no. 3, pp. 2040–2050, 2020._
[5] ——, “An architecture for distributed energies trading in byzantinebased blockchains,” IEEE Transactions on Green Communications and
_Networking, vol. 6, no. 2, pp. 1216–1230, 2022._
[6] S. Popov, “The tangle,” cit. on, p. 131, 2016.
[7] D. Yang, X. Fang, and G. Xue, “Truthful auction for cooperative
communications,” in Proceedings of the Twelfth ACM International
_Symposium on Mobile Ad Hoc Networking and Computing, 2011, pp._
1–10.
[8] A.-L. Jin, W. Song, and W. Zhuang, “Auction-based resource allocation
for sharing cloudlets in mobile cloud computing,” IEEE Transactions
_on Emerging Topics in Computing, vol. 6, no. 1, pp. 45–57, 2015._
[9] A.-L. Jin, W. Song, P. Wang, D. Niyato, and P. Ju, “Auction mechanisms toward efficient resource sharing for cloudlets in mobile cloud
computing,” IEEE Transactions on Services Computing, vol. 9, no. 6,
pp. 895–909, 2015.
[10] J. Kang, R. Yu, X. Huang, S. Maharjan, Y. Zhang, and E. Hossain,
“Enabling localized peer-to-peer electricity trading among plug-in hybrid
electric vehicles using consortium blockchains,” IEEE Transactions on
_Industrial Informatics, vol. 13, no. 6, pp. 3154–3164, 2017._
[11] D. Wu, H. Zeng, C. Lu, and B. Boulet, “Two-stage energy management
for office buildings with workplace ev charging and renewable energy,”
_IEEE Transactions on Transportation Electrification, vol. 3, no. 1, pp._
225 237 2017
-----
[12] C. Liu, K. K. Chai, X. Zhang, E. T. Lau, and Y. Chen, “Adaptive
blockchain-based electric vehicle participation scheme in smart grid
platform,” IEEE Access, vol. 6, pp. 25 657–25 665, 2018.
[13] Z. Su, Y. Wang, Q. Xu, M. Fei, Y.-C. Tian, and N. Zhang, “A secure
charging scheme for electric vehicles with smart communities in energy
blockchain,” IEEE Internet of Things Journal, vol. 6, no. 3, pp. 4601–
4613, 2018.
[14] Z. Zhou, B. Wang, M. Dong, and K. Ota, “Secure and efficient
vehicle-to-grid energy trading in cyber physical systems: Integration of
blockchain and edge computing,” IEEE Transactions on Systems, Man,
_and Cybernetics: Systems, vol. 50, no. 1, pp. 43–57, 2019._
[15] S. Xia, F. Lin, Z. Chen, C. Tang, Y. Ma, and X. Yu, “A bayesian
game based vehicle-to-vehicle electricity trading scheme for blockchainenabled internet of vehicles,” IEEE Transactions on Vehicular Technol_ogy, vol. 69, no. 7, pp. 6856–6868, 2020._
[16] J. Guo, X. Ding, and W. Wu, “Reliable traffic monitoring mechanisms
based on blockchain in vehicular networks,” IEEE Transactions on
_Reliability, vol. 71, no. 3, pp. 1219–1229, 2021._
[17] J. Huang, L. Kong, G. Chen, M.-Y. Wu, X. Liu, and P. Zeng, “Towards
secure industrial iot: Blockchain system with credit-based consensus
mechanism,” IEEE Transactions on Industrial Informatics, vol. 15, no. 6,
pp. 3680–3689, 2019.
[18] V. Hassija, V. Chamola, S. Garg, N. G. K. Dara, G. Kaddoum, and
D. N. K. Jayakody, “A blockchain-based framework for lightweight
data sharing and energy trading in v2g network,” IEEE Transactions
_on Vehicular Technology, vol. 69, no. 6, pp. 5799–5812, 2020._
[19] D. Yang, G. Xue, X. Fang, and J. Tang, “Incentive mechanisms for
crowdsensing: Crowdsourcing with smartphones,” IEEE/ACM transac_tions on networking, vol. 24, no. 3, pp. 1732–1744, 2015._
[20] Y. Lin, Z. Cai, X. Wang, F. Hao, L. Wang, and A. M. V. V. Sai, “Multiround incentive mechanism for cold start-enabled mobile crowdsensing,”
_IEEE Transactions on Vehicular Technology, vol. 70, no. 1, pp. 993–_
1007, 2021.
[21] Y. Zhang and X. Zhang, “Incentive mechanism with task bundling for
mobile crowd sensing,” ACM Transactions on Sensor Networks, vol. 19,
no. 3, pp. 1–23, 2023.
[22] X. Ding, J. Guo, D. Li, and W. Wu, “An incentive mechanism for building a secure blockchain-based internet of things,” IEEE Transactions on
_Network Science and Engineering, vol. 8, no. 1, pp. 477–487, 2020._
[23] Z. Zhao, C. Feng, and A. L. Liu, “Comparisons of auction designs
through multiagent learning in peer-to-peer energy trading,” IEEE Trans_actions on Smart Grid, vol. 14, no. 1, pp. 593–605, 2022._
[24] D. An, Q. Yang, D. Li, and Z. Wu, “Distributed online incentive scheme
for energy trading in multi-microgrid systems,” IEEE Transactions on
_Automation Science and Engineering, 2023._
[25] W. Borjigin, K. Ota, and M. Dong, “In broker we trust: A double-auction
approach for resource allocation in nfv markets,” IEEE Transactions on
_Network and Service Management, vol. 15, no. 4, pp. 1322–1333, 2018._
[26] R. P. McAfee, “A dominant strategy double auction,” Journal of eco_nomic Theory, vol. 56, no. 2, pp. 434–450, 1992._
[27] D. C. Parkes, J. Kalagnanam, and M. Eso, “Achieving budget-balance
with vickrey-based payment schemes in exchanges,” in Proceedings of
_the 17th international joint conference on Artificial intelligence-Volume_
_2, 2001, pp. 1161–1168._
[28] S. Nakamoto, “Bitcoin: A peer-to-peer electronic cash system,” White
_Paper, 2008._
[29] G. Wood et al., “Ethereum: A secure decentralised generalised transaction ledger,” Ethereum project yellow paper, vol. 151, no. 2014, pp.
1–32, 2014.
[30] E. Androulaki, A. Barger, V. Bortnikov, C. Cachin, K. Christidis,
A. De Caro, D. Enyeart, C. Ferris, G. Laventman, Y. Manevich et al.,
“Hyperledger fabric: a distributed operating system for permissioned
blockchains,” in Proceedings of the thirteenth EuroSys conference, 2018,
pp. 1–15.
[31] E. K. Kogias, P. Jovanovic, N. Gailly, I. Khoffi, L. Gasser, and B. Ford,
“Enhancing bitcoin security and performance with strong consistency
via collective signing,” in 25th {usenix} security symposium ({usenix}
_security 16), 2016, pp. 279–296._
[32] V. Hassija, V. Chamola, D. N. G. Krishna, and M. Guizani, “A distributed
framework for energy trading between uavs and charging stations
for critical applications,” IEEE Transactions on Vehicular Technology,
vol. 69, no. 5, pp. 5391–5402, 2020.
[33] A. Churyumov, “Byteball: A decentralized system for storage and
transfer of value,” URL https://byteball.org/Byteball.pdf, 2016.
[34] C. LeMahieu, “Nano: A feeless distributed cryptocurrency network,”
_Nano [Online resource]. URL: https://nano.org/en/whitepaper (date of_
_24 03 2018) 2018_
[35] D. Johnson, A. Menezes, and S. Vanstone, “The elliptic curve digital signature algorithm (ecdsa),” International journal of information security,
vol. 1, no. 1, pp. 36–63, 2001.
**Jianxiong Guo received his Ph.D. degree from**
the Department of Computer Science, University of
Texas at Dallas, Richardson, TX, USA, in 2021,
and his B.E. degree from the School of Chemistry
and Chemical Engineering, South China University
of Technology, Guangzhou, China, in 2015. He is
currently an Assistant Professor with the Advanced
Institute of Natural Sciences, Beijing Normal University, and also with the Guangdong Key Lab of
AI and Multi-Modal Data Processing, BNU-HKBU
United International College, Zhuhai, China. He
is a member of IEEE/ACM/CCF. He has published more than 40 peerreviewed papers and been the reviewer for many famous international journals/conferences. His research interests include social networks, wireless
sensor networks, combinatorial optimization, and machine learning.
**Xingjian Ding received his B.E. degree in electronic**
information engineering from Sichuan University in
2012 and M.S. degree in software engineering from
Beijing Forestry University in 2017. He obtained
his Ph.D. degree from the School of Information,
Renmin University of China in 2021. He is currently
an assistant professor at the School of Software
Engineering, Beijing University of Technology. His
research interests include wireless rechargeable sensor networks, approximation algorithms design and
analysis, and blockchain.
**Weili Wu received the Ph.D. and M.S. degrees from**
the Department of Computer Science, University of
Minnesota, Minneapolis, MN, USA, in 2002 and
1998, respectively. She is currently a Full Professor
with the Department of Computer Science, The
University of Texas at Dallas, Richardson, TX, USA.
Her research mainly deals in the general research
area of data communication and data management.
Her research focuses on the design and analysis
of algorithms for optimization problems that occur
in wireless networking environments and various
database systems.
**Ding-Zhu Du received the M.S. degree from the**
Chinese Academy of Sciences, Beijing, China, in
1982, and the Ph.D. degree from the University
of California at Santa Barbara, Santa Barbara, CA,
USA, in 1985, under the supervision of Prof. R. V.
Book. Before settling at The University of Texas
at Dallas, Richardson, TX, USA, he was a Professor with the Department of Computer Science and
Engineering, University of Minnesota, Minneapolis,
MN, USA. He was with the Mathematical Sciences
Research Institute, Berkeley, CA, USA, for one year,
with the Department of Mathematics, Massachusetts Institute of Technology,
Cambridge, MA, USA, for one year, and with the Department of Computer
Science, Princeton University, Princeton, NJ, USA, for one and a half years.
Dr. Du is the Editor-in-Chief of the Journal of Combinatorial Optimization
and is also on the editorial boards for several other journals.
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/2010.01436, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "GREEN",
"url": "https://arxiv.org/pdf/2010.01436"
}
| 2,020
|
[
"JournalArticle"
] | true
| 2020-10-03T00:00:00
|
[
{
"paperId": "bf17cb050c156b2d5822e9da7dca66c167485039",
"title": "Distributed Online Incentive Scheme for Energy Trading in Multi-Microgrid Systems"
},
{
"paperId": "672d378cd798a03d6a90a79b4b402ff49fce5aa7",
"title": "Next-power: Next-generation framework for secure and sustainable energy trading in the metaverse"
},
{
"paperId": "83d2d11f73249d19ddb06df9acd0d8fbb476dfd0",
"title": "Incentive Mechanism with Task Bundling for Mobile Crowd Sensing"
},
{
"paperId": "f0ef374bd6095c16ffdea79de6ce3d00a2f42f15",
"title": "Comparisons of Auction Designs Through Multiagent Learning in Peer-to-Peer Energy Trading"
},
{
"paperId": "3bab76d6bfba8c821d109aa120d700d5a2580dad",
"title": "Understanding Characteristics and System Implications of DAG-Based Blockchain in IoT Environments"
},
{
"paperId": "4ad85567c7bb89e6f2074ed5308e833de49c63bc",
"title": "A DAG Blockchain-Enhanced User-Autonomy Spectrum Sharing Framework for 6G-Enabled IoT"
},
{
"paperId": "51fe4b279e6d4e7ae37aa56a6555271c64d1a326",
"title": "Incentive Mechanism Design for Joint Resource Allocation in Blockchain-Based Federated Learning"
},
{
"paperId": "86a7b1ae4bf592b2d2294e4bd05eda505639799a",
"title": "Exploiting Multi-Dimensional Task Diversity in Distributed Auctions for Mobile Crowdsensing"
},
{
"paperId": "26c136c5fb2961995847ef43b6a9f6bb560f0446",
"title": "Multi-Source Adversarial Sample Attack on Autonomous Vehicles"
},
{
"paperId": "0e9bb03f7c99807e9733f2357ffe4b10ca0455ea",
"title": "An Incentive Mechanism for Building a Secure Blockchain-Based Internet of Things"
},
{
"paperId": "29039ddf5725295af8cebe52b5709f5ee4dabb9c",
"title": "Multi-Round Incentive Mechanism for Cold Start-Enabled Mobile Crowdsensing"
},
{
"paperId": "63ee9b99e77dee8c5d497babdadf3fd569d85185",
"title": "Reliable Traffic Monitoring Mechanisms Based on Blockchain in Vehicular Networks"
},
{
"paperId": "44ff32b8bbfa0b4bdda40162afcbe12af6e5b1b5",
"title": "zkCrowd: A Hybrid Blockchain-Based Crowdsourcing Platform"
},
{
"paperId": "efcf2a07467ce55afbfc6942870a4376fbb29e95",
"title": "An Architecture for Distributed Energies Trading in Byzantine-Based Blockchains"
},
{
"paperId": "831ddcfd1ddfa0d5b9e34a625fab89d39d16f6a0",
"title": "A Bayesian Game Based Vehicle-to-Vehicle Electricity Trading Scheme for Blockchain-Enabled Internet of Vehicles"
},
{
"paperId": "d1eea8fbe1163e20576c7a43d0ca2cd1615e9da6",
"title": "A Blockchain-Enabled Ecosystem for Distributed Electricity Trading in Smart City"
},
{
"paperId": "5dee597647e62b79d07d34a3f31459456801ca28",
"title": "A Distributed Framework for Energy Trading Between UAVs and Charging Stations for Critical Applications"
},
{
"paperId": "0085516e1951f2e7c0f5fecb0fb9385820883e1a",
"title": "DEAL: Differentially Private Auction for Blockchain-Based Microgrids Energy Trading"
},
{
"paperId": "5603dc2ef70cde2de7af9140480f7756884ee5c5",
"title": "A Blockchain-Based Framework for Lightweight Data Sharing and Energy Trading in V2G Network"
},
{
"paperId": "1bd6c38338bbb44b6ab4ca6af36a0a20b5e1f06d",
"title": "Secure and Efficient Vehicle-to-Grid Energy Trading in Cyber Physical Systems: Integration of Blockchain and Edge Computing"
},
{
"paperId": "3edf84d91b475af359d490f0581a009591559664",
"title": "Double Auction Mechanisms For Dynamic Autonomous Electric Vehicles Energy Trading"
},
{
"paperId": "eeafc2434bbd5f63995c206eb3eedc1a01550145",
"title": "A Secure Charging Scheme for Electric Vehicles With Smart Communities in Energy Blockchain"
},
{
"paperId": "b269f69d4a8a9a7bdfff7780d4f7e7fd1ce554ec",
"title": "Coin Hopping Attack in Blockchain-Based IoT"
},
{
"paperId": "b4c3ae7667cc4c64ba6cf7114ab3be0b163312cf",
"title": "Security and Privacy on Blockchain"
},
{
"paperId": "9be319953ae3b40f7c9fe70f649c1ff60915e6f7",
"title": "Towards Secure Industrial IoT: Blockchain System With Credit-Based Consensus Mechanism"
},
{
"paperId": "dbd85a482b21038c2d792f1652f8c36976bf2690",
"title": "In Broker We Trust: A Double-Auction Approach for Resource Allocation in NFV Markets"
},
{
"paperId": "d05d8515953857cf3375211281838ec44fee669e",
"title": "Adaptive Blockchain-Based Electric Vehicle Participation Scheme in Smart Grid Platform"
},
{
"paperId": "88ad82e6f2264f75f7783232ba9185a2f931a5d1",
"title": "Facial Expression Analysis under Partial Occlusion"
},
{
"paperId": "1b4c39ba447e8098291e47fd5da32ca650f41181",
"title": "Hyperledger fabric: a distributed operating system for permissioned blockchains"
},
{
"paperId": "29ec6d070828fabc60982fb3251fe6acdd04eb19",
"title": "Energy Big Data Security Threats in IoT-Based Smart Grid Communications"
},
{
"paperId": "6ed916428e6596961f7a9ee769f5552a80af285f",
"title": "Distributed Auctions for Task Assignment and Scheduling in Mobile Crowdsensing Systems"
},
{
"paperId": "81bfea080e833fd0046b1e9b879a19429c1d08bf",
"title": "Enabling Localized Peer-to-Peer Electricity Trading Among Plug-in Hybrid Electric Vehicles Using Consortium Blockchains"
},
{
"paperId": "33709e0bb08464cdaf26b1ef68b7c388333a7156",
"title": "BlockChain: A Distributed Solution to Automotive Security and Privacy"
},
{
"paperId": "3b21b6f9e9e7b9477c4f8ab668a348bb25abf7c9",
"title": "Two-Stage Energy Management for Office Buildings With Workplace EV Charging and Renewable Energy"
},
{
"paperId": "f00e466cd4c74e9b50a622c4ef55f8fa0fc0f07c",
"title": "Auction Mechanisms Toward Efficient Resource Sharing for Cloudlets in Mobile Cloud Computing"
},
{
"paperId": "d8403984bff42604cce78876efb31c11321bfed8",
"title": "Incentive Mechanisms for Crowdsensing: Crowdsourcing With Smartphones"
},
{
"paperId": "efd99fe3b5b620d89aa03201199c45988c688670",
"title": "Enhancing Bitcoin Security and Performance with Strong Consistency via Collective Signing"
},
{
"paperId": "8bee02071879104528c87b659b9de05b2427d1cb",
"title": "Truthful auction for cooperative communications"
},
{
"paperId": "353d374013d9849c5534e7b586a583cd97fa275f",
"title": "Achieving Budget-Balance with Vickrey-Based Payment Schemes in Exchanges"
},
{
"paperId": "65ecd5eaec484201c701c4a1508f8d2132c125f0",
"title": "The Elliptic Curve Digital Signature Algorithm (ECDSA)"
},
{
"paperId": "36c724c2f381a74261b878cc78c77bb19bf90e96",
"title": "A dominant strategy double auction"
},
{
"paperId": null,
"title": "Blockchain-basedreverseauctionforv2v charging in smart grid environment"
},
{
"paperId": "e6c2cb33391083f4e60da9378079b453a5591891",
"title": "Auction-Based Resource Allocation for Sharing Cloudlets in Mobile Cloud Computing"
},
{
"paperId": "600c574adfbd0a6895934ec8d3dbfcb56fb2bd68",
"title": "Nano : A Feeless Distributed Cryptocurrency Network"
},
{
"paperId": null,
"title": "Currently, he is working toward the PhD degree in the School of Information, Renmin University of China, Beijing, China. His research interests include wireless rechargeable sensor networks algorithm"
},
{
"paperId": null,
"title": "Byteball: A decentralized system for storage and transfer of value"
},
{
"paperId": "43586b34b054b48891d478407d4e7435702653e0",
"title": "The Tangle"
},
{
"paperId": "64e428003bedbde11b7409999c836c848f735319",
"title": "Designing Truthful Spectrum Double Auctions with Local Markets"
},
{
"paperId": "3c50bb6cc3f5417c3325a36ee190e24f0dc87257",
"title": "ETHEREUM: A SECURE DECENTRALISED GENERALISED TRANSACTION LEDGER"
},
{
"paperId": "4e9ec92a90c5d571d2f1d496f8df01f0a8f38596",
"title": "Bitcoin: A Peer-to-Peer Electronic Cash System"
},
{
"paperId": null,
"title": "of California at Santa Barbara, Santa Barbara, CA, USA,"
},
{
"paperId": null,
"title": "Transactions: When CS C x receives the new transaction from the EV V i , it will check and sign this transaction by its private key as well. Now, the CS C x issues this new transaction"
},
{
"paperId": null,
"title": "A Double Auction for Charging Scheduling among Vehicles Using DAG-Blockchains"
},
{
"paperId": null,
"title": "— We model the assignment between EVs and CSs as a constrained multi-item double auction"
},
{
"paperId": null,
"title": "United International College, Zhuhai, China"
},
{
"paperId": null,
"title": "Received 28 August 2023; revised 15 February 2024; accepted 26"
},
{
"paperId": null,
"title": "sensor networks, combinatorial optimization, and machine learning"
},
{
"paperId": null,
"title": "2023. Trusted mobile edge computing: DAG blockchain-aided trust management and resource allocation"
},
{
"paperId": null,
"title": "Scheduling: The manager waits to receive request messages from EVs and status messages from CSs"
},
{
"paperId": null,
"title": "Confirm: If EV V i receives an order message from the manager, then it can be charged at the CS C x designated by the manager"
},
{
"paperId": null,
"title": "Order and Assignment"
},
{
"paperId": null,
"title": "Engineering, University of Minnesota, Minneapolis, MN, USA"
}
] | 25,517
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/0213f41fa2b6958b9142cc43454cfa1974de97dd
|
[
"Computer Science"
] | 0.913839
|
Databases in Cloud Computing: A Literature Review
|
0213f41fa2b6958b9142cc43454cfa1974de97dd
|
[
{
"authorId": "71456857",
"name": "Harrison John Bhatti"
},
{
"authorId": "2380227",
"name": "Babak Bashari Rad"
}
] |
{
"alternate_issns": null,
"alternate_names": null,
"alternate_urls": null,
"id": null,
"issn": null,
"name": null,
"type": null,
"url": null
}
|
— Information Technology industry has been using the traditional relational databases for about 40 years. However, in the most recent years, there was a substantial conversion in the IT industry in terms of commercial applications. Stand-alone applications have been replaced with electronic applications, committed servers with various appropriate servers and devoted storage with system storage. Lower fee, flexibility, the model of pay-as-you-go are the main reasons, which caused the distributed computing are turned into reality. This is one of the most significant revolutions in Information Technology, after the emergence of the Internet. Cloud databases, Big Table, Sherpa, and SimpleDB are getting to be more familiar to communities. They highlighted the obstacles of current social databases in terms of usability, flexibility, and provisioning. Cloud databases are essentially employed for information-escalated applications, such as storage and mining of huge data or commercial data. These applications are flexible and multipurpose in nature. Numerous value-based information administration applications, like banking, online reservation, e-trade and inventory administration, etc. are produced. Databases with the support of these types of applications have to include four important features: Atomicity, Consistency, Isolation, and Durability (ACID), although employing these databases is not simple for using in the cloud. The goal of this paper is to find out the advantages and disadvantages of databases widely employed in cloud systems and to review the challenges in developing cloud databases
|
Published Online April 2017 in MECS (http://www.mecs-press.org/)
DOI: 10.5815/ijitcs.2017.04.02
# Databases in Cloud Computing:
A Literature Review
## Harrison John Bhatti
School of Computing and Technology, Asia Pacific University of Technology and Innovation (APU), Technology Park
Malaysia (TPM), Bukit Jalil, Kuala Lumpur 57000 Malaysia
E-mail: harrisonjohn03@gmail.com
## Babak Bashari Rad
School of Computing and Technology, Asia Pacific University of Technology and Innovation (APU), Technology Park
Malaysia (TPM), Bukit Jalil, Kuala Lumpur 57000 Malaysia
E-mail: babak.basharirad@apu.edu.my
**_Abstract—Information Technology industry has been_**
using the traditional relational databases for about 40
years. However, in the most recent years, there was a
substantial conversion in the IT industry in terms of
commercial applications. Stand-alone applications have
been replaced with electronic applications, committed
servers with various appropriate servers and devoted
storage with system storage. Lower fee, flexibility, the
model of pay-as-you-go are the main reasons, which
caused the distributed computing are turned into reality.
This is one of the most significant revolutions in
Information Technology, after the emergence of the
Internet. Cloud databases, Big Table, Sherpa, and
SimpleDB are getting to be more familiar to communities.
They highlighted the obstacles of current social databases
in terms of usability, flexibility, and provisioning. Cloud
databases are essentially employed for informationescalated applications, such as storage and mining of
huge data or commercial data. These applications are
flexible and multipurpose in nature. Numerous valuebased information administration applications, like
banking, online reservation, e-trade and inventory
administration, etc. are produced. Databases with the
support of these types of applications have to include four
important features: Atomicity, Consistency, Isolation, and
Durability (ACID), although employing these databases
is not simple for using in the cloud. The goal of this paper
is to find out the advantages and disadvantages of
databases widely employed in cloud systems and to
review the challenges in developing cloud databases.
**_Index Terms—Cloud, Database, Cloud Computing,_**
Cloud Database, Cloud Service.
I. INTRODUCTION
All various branches of IT are obligated and committed
to provide and true enlisting, stockpiling, supporting the
workplaces and IT frameworks at the most lessened
achievable cost. According to [1], Enormous interest in
IT framework fills in as a prevention in its gathering,
especially for little scale affiliations. Down and out
affiliations hunt down alternatives, which can lessen their
capital endeavors incorporated into acquiring and keeping
up IT hardware and programming with the goal that they
can get greatest advantages of IT.
At this stage, Cloud databases are considered as a
smart answer for programmers, on the off chance that
they need to store the information of their applications in
a versatile and exceedingly accessible backend. These
administrations are alluded to as Database-as-a-Service
(DBaaS) [2].
A cloud-facilitated DBMS must have a few methods
for persistently putting away its database. One
methodology is to utilize a persistent stockpiling
administration gave inside of the cloud and got to over
the system by the DBMS. An illustration of this is
Amazon's Elastic Block Service (EBS), which gives
system available persevering stockpiling volumes that
can be connected to virtual machines [3].
This article investigates the preferences and
weaknesses of conveying database frameworks in the
cloud. We take a gander at how the run of the mill
features of available distributed computing influences the
decision of information administration applications to
transfer in the cloud. Due to the growing necessities of
today's commercial world for more investigation and
exploration, we can infer explanatorily and descriptive
information administration applications are more capable
of being employed in the cloud than value-based
information administration applications. We, therefore,
lay out an exploration motivation for huge size data
investigation in the cloud, to demonstrate that the
accessible frameworks are not suitable for cloud
organization, and to resist that there is a requisite for a
recently planned DBMS, designed especially for
distributed computing stages [4].
In this paper, cloud databases, cloud computing and the
databases that can be hosted and deployed in the cloud
have been discussed, respectively. Furthermore, the
advantages and disadvantages of the most widely used
database in cloud computing have been presented. This
paper is organized as follows. In the next section, cloud
database has been introduced, in brief. Then, the cloud
-----
computing and its features have been discussed. Next,
some popular databases used in cloud computing have
been reviewed, including advantages and disadvantages
of MySQL, and in the last section, the most important
challenges in the development of cloud databases have
been discussed. Finally, at the end of the article, a
summary, and list of references are given.
II. CLOUD DATABASE
The cloud database holds the information on distinctive
server farms situated in diverse areas. This makes the
cloud database structure not the same as the objective
database administration framework. Over a cloud
database, there are numerous hubs, intended for question
administrations, for server farms that are also corporate
farms and are situated in distinctive land areas. This
connection is required for the convenient and full access
to the database on the cloud administrations. Many
systems have been introduced to get the benefits of
databases over the cloud. The user can take its advantages
by means of a personal computer using the web, or by a
mobile device, which has capability of accessing the
cloud database using 3G/4G services. To understand the
infrastructure of the cloud databases, structure of cloud
database is demonstrated in Fig. 1 [5].
Fig.1. Structure of Cloud Database [5]
III. CLOUD COMPUTING
Distributed computing is a late idea and one of the
most recent PC industry trendy expressions. The idea is
gotten from the symbolism of the ''Internet cloud'', in
which the symbolism of a cloud is customarily ''used to
speak to the Internet or some extensive organized
environment''. The thought portrayed in the symbolism is
that customer information and applications are put away
and got to ''in the distance''. As being what is indicated,
one definition offered for distributed computing is the
''virtualization of assets that keeps up and oversees itself.
To improve the idea, distributed computing can be
basically characterized as the distribution and utilization
of assets and facilities of a system to finish the work with
no worry about possession or administration of the
system resources and assets. With distributed computing,
PC assets for finishing work and their information are no
more put away on one's PC, yet are facilitated somewhere
else to be made available in any area and whenever [6].
According to [7], Distributed computing is a
developing range of disseminated processing that
provides lots of benefits to the organization so that
companies can get their data by using a new technology
easy and faster. When an organization makes a contract
with the cloud service provider to save their programs
and data then cloud service provider make it possible so
that their clients can get full access anytime anywhere
with full benefits but still there are some things needed to
be considered like security and viewing data by others.
Distributed computing is not a solitary sort of framework,
but rather it incorporates a scope of basic advances and
setup choices. The qualities and weaknesses of the
distinctive cloud development, policies, structures,
administration models, and sending routines should be
considered by associations evaluating administrations to
satisfy their requirements.
_A. Features of Cloud Computing_
There are lots of features and facilities provided by
cloud computing, but here we will talk about some of
them. According to [7], a model for supporting valuable,
on interest framework access to a shared customizable
resources like servers, storage, and applications, which
are instantly supplied and available with simple
organization efforts or immaterial collaboration of
provider.
- _On-Demand Self-Service_
With distributed computing, affiliations can have on
-----
interest self-organization for handling capacities, for
instance, server time and framework stockpiling when
required, and through a single supplier [7].
- _Broad Network Access_
All services and facilities of clouds are accessible on
the network and will be available according to some
systematic mechanisms that improve the use by
heterogeneous thin or thick client platforms, like mobile
devices or portable workstations [8].
- _Pooling the Computer Resources_
The resources provided by the supplier are stored to
present appropriate services to different customers
through a multi-tenant model, with various real or virtual
resources allocated based on the requirements of
customers. While the range of the benefits like data
storage is provided, care of memory, network
transmission limits, and virtual machines is not needed to
be controlled by the user, it might be feasible for the
supporter of determining the nation, state, or server farm
that gives the cloud administrations [7].
- _Rapid Elasticity_
Cloud limits can be given to the endorser rapidly and
adaptable, allowing the supporter of either construct or
decrease organizations. The limits available consistently
appear, in every way, to be fast to the supporter and can
be gained in any sum at whatever point [7, 8].
- _Measured Service_
Cloud systems basically monitor, control and improve
the availability and performance of provided facilities and
services through a deliberate administration ability that is
fitting for the sort of administration offered. The use of
asset can be observed, monitored, and recorded to offer a
clear report to the supplier, as well as the users of the
services [7, 8].
- _Multi-Tenacity Services_
The cloud server can be informed for the prerequisites
for policy-driven administration, segmentation, isolation,
governance, service levels, and charging/payments for
various types of customers [9].
IV. DATABASES IN CLOUD COMPUTING ENVIRONMENT
Distributed computing innovation speaks to another
ideal model for facilitating programming applications.
This standard streamlines the prolonged procedures of
equipment provisioning, equipment acquirement, and
programming sending. In this manner, it reformed the
way computational assets and administrations are
popularized and conveyed to clients. These days,
distributed computing is becoming essentially. Cloud
suppliers progressively give new administrations and new
elements to their customers with proficient and practical
answers for their issues. Thus, the cloud has turned into
an appealing stage for the product designers and
endeavors to have their applications and frameworks.
Nonetheless, the administrations offered by distinctive
cloud suppliers are generally incongruent with one
another and do not bolster any institutionalized model or
interfaces. Along these lines, one of the real difficulties
for encouraging cloud appropriation is that of the cloud
interoperability and transportability [2].
On a fundamental level, cloud databases are at present
considered as an appealing answer for programming
designers on the off chance that they need to store the
information of their applications in an adaptable and
exceptionally accessible backend. These administrations
are alluded to as Database-as-a-Service (DBaaS). These
cloud-based information stockpiling administrations can
be arranged into two principle classifications: benefits
that backing conventional social databases (RDB) (e.g.,
Amazon RDS, Google SQL, Microsoft Azure), and
key/quality pair information stockpiling administrations
(e.g., Amazon Simple DB, Google Data Store), which are
otherwise called NoSQL Databases. In important, RDB
frameworks utilization organized inquiry dialect (SQL) as
an institutionalized interface to get the information in a
social database. On the other side, the NoSQL databases
stay unstandardized, so there is no brought together
information access approach. Consequently, every cloud
supplier has an alternate approach to overseeing and
access the database, which makes the information
convenience, and it is a testing assignment to accomplish
between these frameworks [2].
At the other compelling, applications can store their
information utilizing cloud-facilitated social database
administration frameworks (DBMS). Case in point,
customers of the base as an administration supplier, for
example, Amazon, can send DBMS in virtual machines
and utilize these to give database administration to their
applications. On the other hand, applications can utilize
administrations, for example, Amazon's RDS or
Microsoft SQL Azure in a comparative manner. This
methodology is most appropriate to applications that can
be upheld by a solitary DBMS case, or that can be shared
over various autonomous DBMS occurrences. High
accessibility is additionally an issue, as the DBMS speaks
to a solitary purpose of disappointment an issue normally
tended to utilize DBMS-level high accessibility systems.
In spite of these impediments, this methodology is
broadly utilized on the grounds that it puts the greater
part of the surely knew advantages of social DBMS, for
example, SQL question handling and exchange support,
in the administration of the application. This is the
methodology we concentrate on in this paper. A cloud
facilitated DBMS must have a few methods for diligently
putting away its database. One methodology is to utilize a
steady stockpiling administration give inside of the cloud
and got to over the system by the DBMS. A sample of
this is Amazon's Elastic Block Service (EBS), which
gives system open diligent stockpiling volumes that can
be joined to virtual machines [3].
Amazon provides deployment services to databases
like MS SQL Server, MySQL, and Oracle in its own
cloud which is EC2 [10].
-----
V. POPULAR DATABASES USED IN CLOUD COMPUTING
There are some most popular databases in cloud
computing. They are mentioned below:
- StromDB
- MySQL
- PostgreSQL
- Google Cloud SQL
- MongoLab.
_A. StromDB_
StromDB is a free and an open source appropriate
ongoing calculation framework. It is simple in StromDB
to dependably handle the infinite flow of information,
finishing for steady get ready what Hadoop achieved for
pack get ready. StromDB is very straightforward, it can
be utilized with any programming dialect, and is a
considerable measure of enjoyable to utilize. StromDB
has numerous utilization cases: constant examination,
online machine learning, nonstop processing, conveyed
RPC, ETL, and then some. StromDB is quick: a
benchmark timed it at more than a million tuples prepared
every second per hub. It is adaptable, deficiency tolerant
ensures your information will be prepared and is anything
but difficult to situated up and work. StromDB
coordinates with the queuing and database advancements
you as of now utilize. A Storm topology devours floods
of information and procedures those streams in selfassertively complex ways, repartitioning the streams
between every phase of the reckoning however required.
Read all the more in the instructional exercise [11].
_B. MySQL_
MySQL is an open-source social database
administration framework. It is possessed by Oracle
Corporation and can be utilized under either the GNU
General Public License or a standard business permit
acquired from Oracle. MySQL is a hearty, multi-strung,
value-based DBMS. It is profoundly versatile and can be
conveyed over numerous servers. Because it can be
utilized for nothing out of pocket, it holds a critical piece
of the pie inside of established researchers. While
frequently thought to be unseemly for spaces of high
security like budgetary organizations or certain territories
of the administration. MySQL has turned into the main
social database in numerous regions of the scholarly
world, including experimental research and instructing
understudies [12].
_C. PostgreSQL_
Cloud Database permits administration suppliers and
associations to offer versatile and profoundly adaptable
database-as-an administration (DBaaS) situations while
liberating DBAs and application designers from the rigors
of setting up and directing present day and vigorous
database situations. Postgres plus Cloud Database
rearranges the procedure of provisioning vigorous
Postgres arrangements while exploiting the advantages of
distributed computing. At the point when utilized with
Postgres Plus Advanced Server, Cloud Database
additionally gives an Oracle-perfect DBaaS, offering
sensational expense reserve funds and game changes [13].
Fig. 2 illustrate the PostgreSQL Performance in
Amazon.
Fig.2. PostgreSQL Performance in Amazon [14]
_D. Google Cloud SQL_
MySQL database has one more database, which can
easily be deployed in Google cloud known as “Google
Cloud SQL”. It has every one of the abilities and
usefulness of MySQL, with a couple of extra elements
and a couple of unsupported elements as recorded
underneath. Google Cloud SQL is anything but difficult
to utilize, doesn't require any product establishment or
support, and is perfect for little to medium-sized
applications [15].
-----
MySQL databases sent in the cloud without an object.
It is provided you by the Google Cloud Platform with
effective databases that run quick, don't come up short on
space and give your application the excess, it requires
dependable capacity [15].
_E. MongoLab_
MongoDB is an arranged open source JSON database
structure. Geir Magnusson and Dwight Merriman created
at 10gen. Instead of a complete quality store, it is planned
to be a bona fide article database. The data is stored in
JSON, like records with component developments. The
flexibility of key quality store and space is given. The
rich accommodation like records and part demand of
social databases are also provided. The flexibility level is
given too [1].
VI. WIDELY USED DATABASE IN CLOUD COMPUTING
(MYSQL)
MySQL is the best database framework being utilized
everywhere throughout the world, particularly when little
and medium size commercial enterprises are attempting
to cut expenses. Keeping in mind the end goal to meet the
level of administration requested by the clients, it is
discriminating that applications have the accessibility and
execution expected to pay little respect to the sort of use
or the work stack a framework has. For measuring the
execution in MySQL applications, discovery
methodology is the most widely recognized system
utilized for measuring the Transactions every Second [16].
_A. Advantages of MySQL Database in Cloud Computing_
There are some main advantages of MySQL database
in cloud computing [17]:
- _Availability_
It is very terrible to deal with a database going down
during high workload and sales times. Cloud-based
MySQL databases provide a guarantee to avoid this issue
using modern technology and accessible and distributed
resources.
- _Buy the database administration only_
Some cloud organizations just offer MySQL database
facilitating through a cloud-based facilitating record. As
of late organizations began offering databases as an
administration, permitting people to pay just for the
databases and not for a facilitating record that there is no
utilization for.
- _Easy to get outsource maintenance_
Innovation keeps on progressing, yet administration's
spending on IT office staff more often than scale in a
remarkable same manner. In case you're as of now overburden with system organization, sending parts of the
framework to the cloud permits you to offload upkeep
and redesign undertakings to the cloud supplier. You
cannot be totally uninvolved, however, each and every bit
makes a difference.
- _Versatility_
The versatility that originates from MySQL databases
cannot be coordinated by individual or devoted devices.
People would prefer not to ship in a bundle of database
servers for trivial needs, however, cloud-based MySQL
databases are ideal for such circumstance.
_B. Disadvantages of MySQL Database in Cloud_
_Computing_
Below is the discussion on limitations of both SQL and
MySQL [12].
- _Null Data_
Putting away deficient or obscure worldly information
in SQL is commonly finished with a NULL. DATE and
TIME information sorts, as portrayed by SQL-92, are
viewed as each to be made out of three different whole
numbers of different satisfactory extents. For instance,
DATE is the single information sort relegated to a table
quality that stores a date (and not a period). SQL takes
into consideration to win a big or bust nullability. That is,
the information, in general, can be invalid, yet parts of a
date can't.
- _Granularity_
Identified with putting away invalid information is
putting away information in different granularities. Since
zero is significant in every time field and in light of the
fact that MySQL additionally uses zero as an invalid
marker for every field in a period, putting away TIMEs or
DATETIMEs of different granularities in a solitary
segment which is impractical.
- _Overflow_
Within MySQL, a DATE can be categorized as one of
three essential extents: upheld, legitimate, and unlawful.
"Upheld" means acknowledged by the framework and
ensured to work. "Lawful" and "unlawful" are terms not
clearly characterized but rather which were extrapolated
from other phrasing utilized as a part of the instructional
booklet. "Legal" means perhaps acknowledged by the
framework however not ensured to work. "Illegal" means
not acknowledged by the framework.
- _Non-Gregorian Calendars_
MySQL utilizes the proleptic Gregorian schedule,
implying that all dates are settled around the Gregorian
datebook and that the Gregorian logbook is utilized to
speak to even those dates that occurred amid the time
when the Julian timetable was being used. The same
component can be found in play in the yearly dates of
Hanukkah. Hanukkah moves around the Gregorian
timetable on the grounds that it is taking into account the
Jewish datebook. Inside of the Jewish logbook, Hanukkah
holds an altered position, however, non-Jews have a
tendency to identify with Hanukkah regarding the
Gregorian schedule, which is the reason it seems to move
around from year to year.
-----
VII. CHALLENGES TO DEVELOP CLOUD DATABASE
Cloud DBMSs ought to bolster elements of Cloud
figuring and additional databases for more extensive
worthiness; it is a responsibility of Hercules. There are
some possible difficulties connected with cloud databases,
which are displayed in Fig. 3 [1].
Fig.3. Issues can occur during DB deployment [1]
- _Scalability_
The rapid growth of databases in size is a consequence
of involving large size multimedia data, which requires
novel scalable systems. Because users expect to easily
scale up and down the size of data in databases to ensure
requirements of their commercial aims, cloud systems
must provide scalable database services to meet the
expectations of their users.
This is the most important feature of cloud standard. It
prescribes the services that can be scaled-up or down
remarkably without accomplishing any impedance in the
association. It is a big challenge in the architecture of the
system to implement databases in the cloud to guarantee
that synchronous customers are supported and handled
and data can be improved.
- _Fault Tolerance and High Availability_
It is very vital to replicate information over wide
geographical locations to provide a high availability and
robustness of information, as well as high flexibility in
adaptation to internal failure.
The term availability of system can generally be
defined as the degree of accessibility and usability of
resources for individual users or staffs of organizations
[18]. This is one of the most important issues, which must
be considered by individuals or organizations before
starting to move to the cloud database.
If an interruption occurs due to a failure in cloud
service, it may affect the availability of databases,
temporarily or permanently, which may cause a serious
loss of data, partially or completely. Equipment failures,
Security deficiencies and attacks such as DOS are serious
threats to the availability of cloud database system. In
most cases, these types of failures are unpredictable and
can seriously influence the performance of organizations
or individuals’ activities, which may result in the
corruption of data or interruption of real-time services.
The performance of the majority of database applications
may seriously be affected due to unavailability or failure
of cloud service.
- _Integrity and Data Consistency_
In order to guarantee a high level of integrity of data, it
is vital to carefully control and monitor users of the
database, including the database administrator and
technical staffs, who legally permitted to access the
system [18].
Keeping the consistency of an exchange in a database
is also a very difficult task, even worse if it changes very
fast, particularly on account of value-based information.
Designers must resemble BASE (Basically Available,
Soft state, eventually consistent) features of database
precisely. They must be careful to ensure that there is no
risk of losing data integrity in their shift to cloud
databases.
- _Interface for Query_
Cloud Database is spread. Addressing passed on the
database is an imperative test that cloud planners face. A
passed on inquiry needs to get to particular focus
purposes of cloud database. There ought to be a
streamlined and sorted out solicitation interface for
investigating the database.
- _Privacy and Security of Database_
There are some security concerns which organization
-----
needs to consider, before transferring the traditional
database to the database on a cloud platform. These
security considerations are the main and significant
concern of the organizations, not the cloud service
provider, as the outcome will ultimately affect the
organization’s function. Specifically, if sensitive
information is stored on the local databases, during the
migration process it is important to promise users about
the security of cloud database. In particular, the
confidentiality and protection of data should be
guaranteed to users. It must be assured that the data will
not be illegally manipulated or stolen during the
procedure of transferring from the internal database to
cloud storage. To achieve this safe migration, a secure
procedure should be carefully designed and implemented
[18].
It is also essential to encrypt the data stored on the
outsourced databases hosted at cloud storage, in order to
achieve a high level of confidentiality.
Dangers are included in the storage of value-based
information on a host that is not adequately secured.
Significant information is encrypted before being stored
in the cloud to neutralize illegal access. The ability of
decryption of data in the cloud should be restricted for
different applications. It is a serious challenge to promise
the privacy and security of various databases on one
system.
- _Data Portability_
Information Portability is the capacity to execute
application prepared for a specific cloud supplier in
another cloud supplier's settings and systems.
Interoperability is the capability to provide some codes
that are enough adaptable to work with various cloud
suppliers, independent of their differences.
VIII. REVIEW OF RELATED WORKS
There are many researches on the cloud computing
database and its related issues, published recent years.
Some of them are discussed briefly in this section. In a
research paper by Vodomin and Androcec [19], the
authors present a practical prototype of a migration tool
for SQL databases, including MySQL, PostgreSQL and
Microsoft SQL Server. This research mainly contributed
in investigation of issues may happen in the process of
migration of databases into cloud storage. The
dissimilarities of storage models between commercial
clouds have also been discussed, which assist to identify
the potential issues may appear during the migration
between cloud storages.
Strauch, S., et al [20] introduced a methodology to
move applications to cloud. Their methodology considers
some significant aspects, such as differences in the
granularity of interactions and data confidentiality. It is
also necessary to allow the interaction of the application
with remote data sources. All these features have been
addressed in the proposed method. Furthermore, the
authors also developed a tool for decision support,
application refactoring and movement of data. This tool
aids the developers of applications to realize of the
suggested methodology. Both the proposed methodology
and the tool have been also evaluated by the authors
using a case study examination in partnership with an IT
enterprise.
In another research paper by Abourezq and Idrissi [21],
the researchers offered a benchmark of the main database
solutions presented by service providers as DBaaS
(DataBase as a Service). They reviewed the
characteristics of solutions and their adaptability to Big
Data applications.
In the paper written by Arora and Gupta [1], the state
of the art in the cloud databases and various architectures
have been reviewed and discussed. Furthermore, the
challenges in development of cloud databases, as well as
some of very common cloud databases have been
discussed and assessed. The main goal of their paper was
to review and discuss the recent trends, and to explore
and analyze the barriers and issues in development of
cloud databases technologies.
Alomari et al. in their paper published on 2014 [2]
focused on the challenges and issues of data transfer
between various cloud data storage. The authors proposed
a data model and an API for the modern generation of
NoSQL databases in cloud storage. The implementation
of their proposed framework involves three popular
NoSQL systems, including Google Datastore, Amazon
SimpleDB and MongoDB. The proposed framework was
established with high level of flexibility and could be
simply applied to other NoSQL systems. Moreover, the
framework include some tools to provide support for
adaptation, data transformation and exchange. The
authors also employed a case study to describe the
structure and implementation of the suggested framework.
In another study by Ferretti et al. [22], an alternative
architecture has been proposed which avoids intermediate
components, to achieve a level of availability and
scalability similar to unencrypted cloud database services.
Additionally, their proposed architecture ensure the
consistency of data in an environment where different
clients run SQL queries simultaneously, and the
configuration of the database can be changed.
Finally, In an article published by Shendeand Chapke
[23], the latest trends in cloud services provided for
database management systems have been discussed. The
benefits and drawbacks of database as a service have
been explored to allow users to make decision for using
database as a service. This article also discussed the
architecture of cloud based on database management
system.
IX. SUMMARY
This article introduced the basic knowledge and
concepts of cloud databases and explained some of their
important features. Organizations started to work on the
distributed computing for various aims and a pattern
begun by adopting distributed computing administrations
for an improved and faster accessibility of the data
instead of establishing a separate database server for each
-----
organization or company. Presently the cloud database
has advanced another measurement Database as a Service.
This service assists the organizations to exploit the
facilities provided by the suppliers, without any concern
about storage of the equipment and programming tools.
They get administrations from DBaaS supplier and take
advantages of the flexibility of a full-time available
database. There are also both favorable conditions and
inconveniences. However, the adoption the cloud
database has proved that the advantages are more than the
weaknesses. The cloud database services offer various
favorable features.
REFERENCES
[1] Arora, I. and A. Gupta, Cloud databases: a paradigm shift
_in databases. International J. of Computer Science Issues,_
2012. 9(4): p. 77-83.
[2] Alomari, E., A. Barnawi, and S. Sakr. _CDPort: a_
_framework of data portability in cloud platforms. in_
_Proceedings of the 16th International Conference on_
_Information Integration and Web-based Applications &_
_Services. 2014. ACM._
[3] Liu, R., A. Aboulnaga, and K. Salem. _Dax: a widely_
_distributed multitenant storage service for dbms hosting._
in _Proceedings of the VLDB Endowment. 2013. VLDB_
Endowment.
[4] Agrawal, D., S. Das, and A.E. Abbadi, Data management
_in the cloud: challenges and opportunities. Synthesis_
Lectures on Data Management, 2012. 4(6): p. 1-138.
[5] Al Shehri, W., _Cloud Database Database as a Service._
International Journal of Database Management Systems,
2013. 5(2): p. 1.
[6] Scale, M.-S.E., _Cloud computing and collaboration._
Library Hi Tech News, 2009. 26(9): p. 10-13.
[7] Radack, S., _Cloud computing: a review of features,_
_benefits, and risks, and recommendations for secure,_
_efficient implementations. National Institute of Standards_
and Technology, 2012.
[8] Puthal, D., B. Sahoo, S. Mishra, and S. Swain. _Cloud_
_computing features, issues, and challenges: a big picture._
in Computational Intelligence and Networks (CINE), 2015
_International Conference on. 2015. IEEE._
[9] Jula, A., E. Sundararajan, and Z. Othman, _Cloud_
_computing service composition: A systematic literature_
_review. Expert Systems with Applications, 2014. 41(8): p._
3809-3824.
[10] Aboulnaga, A., et al., _Deploying Database Appliances in_
_the Cloud. IEEE Data Eng. Bull., 2009. 32(1): p. 13-20._
[11] Marz, N., _Storm: distributed and fault-tolerant realtime_
_computation, in O'Reilly Strata Conference Making Data_
_Work. 2012, O'Reilly Media, Inc.: Santa Clara, California._
[12] Vicknair, C., D. Wilkins, and Y. Chen. _MySQL and the_
_trouble with temporal data. in_ _Proceedings of the 50th_
_Annual Southeast Regional Conference. 2012. ACM._
[13] Postgres Plus, _Cloud Database: Getting started Guide._
Retrieved 23rd November, 2012.
[14] Campbell, L., J. Edwards, and E. Calvo _RDBMS in the_
_Cloud: PostgreSQL on AWS. Amazon Web Services,_
2013.
[15] Krishnan, S. and J.L.U. Gonzalez, Google Cloud SQL, in
_Building Your Next Big Thing with Google Cloud_
_Platform. 2015, Springer. p. 159-183._
[16] Ahmed, M., M.M. Uddin, M.S. Azad, and S. Haseeb.
_MySQL performance analysis on a limited resource server:_
_Fedora vs. Ubuntu Linux. in_ _Proceedings of the 2010_
_Spring Simulation Multiconference. 2010. Society for_
Computer Simulation International.
[17] Summers, A. _Five advantages of running a SQL Server_
_database in a cloud environment or virtual machine. 2013._
[18] Sakhi, I., Database security in the cloud. 2012.
[19] Vodomin, G. and D. Androcec. _Problems during_
_Database Migration to the Cloud. in_ _Central European_
_Conference on Information and Intelligent Systems. 2015._
Faculty of Organization and Informatics Varazdin.
[20] Strauch, S., et al., Migrating enterprise applications to the
_cloud: methodology and evaluation. International Journal_
of Big Data Intelligence 5, 2014. 1(3): p. 127-140.
[21] Abourezq, M. and A. Idrissi, _Database-as-a-service for_
_big data: An overview. International Journal of Advanced_
Computer Science and Applications (IJACSA), 2016. 7(1).
[22] Ferretti, L., M. Colajanni, and M. Marchetti, _Supporting_
_security and consistency for cloud database, in_
_Cyberspace Safety and Security. 2012, Springer. p. 179-_
193.
[23] Shende, S.B. and P.P. Chapke, _Cloud Database_
_Management System (CDBMS). Compusoft, 2015. 4(1): p._
1462.
**Authors’ Profiles**
**Harrison** **John** **Bhatti** received his
Bachelors of Science in Computer Science
(BCS) degree in 2003 and M.Sc. of
Information Technology Management in the
field of Cloud Computing and Virtualization
in 2016 from Asia Pacific University of
Technology and Innovation (APU), Kuala
Lumpur in Collaboration with Staffordshire
University, UK. Harrison John is currently doing his second
Masters of Engineering in Industrial Management and
Innovation from University of Halmstad, Sweden. His core
research areas are Cloud Computing, Virtualization, Docker
Container and Strategic Planning and Innovation.
**Babak Bashari Rad received his B.Sc. of**
Computer Engineering in subfield of
Software in 1996 and M.Sc. of Computer
Engineering in field of Artificial
Intelligence and Robotics in 2001 from
University of Shiraz. He received his Ph.D.
in Computer Science, from University
Technology of Malaysia, in 2013. Dr. Babak is currently
Program Leader of Postgraduate Studies in School of
Computing and a Senior Lecturer in academic group of
Computer Science and Software Engineering (CSSE), Asia
Pacific University of Technology and Innovation (APU), Kuala
Lumpur. His main research interests cover a broad range of
various areas in Computer Science and Information Technology
including Information Security and Forensics, Malware
Detection, Machine Learning, Artificial Intelligence, Image
Processing, Cloud Computing, and other relevant fields.
-----
**How to cite this paper:** Harrison John Bhatti, Babak Bashari
Rad,"Databases in Cloud Computing: A Literature Review",
International Journal of Information Technology and Computer
Science(IJITCS), Vol.9, No.4, pp.9-17, 2017. DOI:
10.5815/ijitcs.2017.04.02
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.5815/IJITCS.2017.04.02?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.5815/IJITCS.2017.04.02, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "GOLD",
"url": "http://www.mecs-press.org/ijitcs/ijitcs-v9-n4/IJITCS-V9-N4-2.pdf"
}
| 2,017
|
[
"Review"
] | true
| 2017-04-08T00:00:00
|
[
{
"paperId": "10b13340aa6145b67d470c8989fa0339e7b1d075",
"title": "Problems during Database Migration to the Cloud"
},
{
"paperId": "e38f7af107e55ab01fffa6e730337ef18544b3bd",
"title": "Building Your Next Big Thing with Google Cloud Platform"
},
{
"paperId": "c7bcfbe38eae00e0d3c1d1615217cd084a6e5518",
"title": "Cloud Computing Features, Issues, and Challenges: A Big Picture"
},
{
"paperId": "74cf3121e2f901b501cfcc868baa8d110781225c",
"title": "Migrating enterprise applications to the cloud: methodology and evaluation"
},
{
"paperId": "1463029e078bda11a877ac533dde8a004cebbaca",
"title": "CDPort: A Framework of Data Portability in Cloud Platforms"
},
{
"paperId": "e72714c051aae9d4e32002d2b406a6a5f7d58d5f",
"title": "Cloud computing service composition: A systematic literature review"
},
{
"paperId": "7460d993bed5f8b624af3273fcd4873c952ceb7d",
"title": "Cloud Database Management System (CDBMS)"
},
{
"paperId": "9be96860856b2eb37d75111eb2bd8134e8034bd4",
"title": "Cloud Database Database as a Service"
},
{
"paperId": "b51b0c8de157121c8e00adcde03a1114b90a983e",
"title": "DAX: A Widely Distributed Multi-tenant Storage Service for DBMS Hosting"
},
{
"paperId": "f13fe25b56891d696341ed4720bca199d03c7cfb",
"title": "Data Management in the Cloud: Challenges and Opportunities"
},
{
"paperId": "55d611b4d79f39462f204a3e194457c2ef8b1017",
"title": "Supporting Security and Consistency for Cloud Database"
},
{
"paperId": "3111d9a0c472a6b9912b8990033fd45e99bdc732",
"title": "Cloud Computing: A Review of Features, Benefits, and Risks, and Recommendations for Secure, Efficient Implementations | NIST"
},
{
"paperId": "a66074da68167b4e2f1535b472c37f762db99d19",
"title": "MySQL and the trouble with temporal data"
},
{
"paperId": "da08c17d228e05e9a72174c906354155cdd79d8d",
"title": "MySQL performance analysis on a limited resource server: Fedora vs. Ubuntu Linux"
},
{
"paperId": "c6fa48334fb77707412634aaa0a04a7fc4716b4e",
"title": "Cloud computing and collaboration"
},
{
"paperId": "a0e7e17762d0297f9a4cb7c714aa1ee86962c512",
"title": "Database-as-a-Service for Big Data: An Overview"
},
{
"paperId": null,
"title": "RDBMS in the Cloud: PostgreSQL on AWS"
},
{
"paperId": null,
"title": "Five advantages of running a SQL Server database in a cloud environment or virtual machine"
},
{
"paperId": null,
"title": "Cloud databases: a paradigm shift in databases"
},
{
"paperId": "d0ba7b29e9dca93d39c85180afdbc42d860bd025",
"title": "Database security in the cloud"
},
{
"paperId": null,
"title": "Postgres Plus"
},
{
"paperId": null,
"title": "Storm: distributed and fault-tolerant realtime computation"
},
{
"paperId": null,
"title": "Society for Computer Simulation International"
},
{
"paperId": "12a911fa2c9babc1d91368a8fde757c4f8a282c1",
"title": "Deploying Database Appliances in the Cloud"
},
{
"paperId": "7fabc3a05d6246634ff81fe1c1b95e633d66492b",
"title": "Copyright"
},
{
"paperId": null,
"title": "Bhatti received his Bachelors of Science in Computer Science (BCS) degree in 2003 and M.Sc. of Information Technology Management in the field of Cloud Computing and Virtualization"
},
{
"paperId": null,
"title": "Innovation from University of Halmstad, Sweden"
}
] | 8,803
|
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/02143a479ab76c1a8793d3669a25e72ead827869
|
[
"Computer Science"
] | 0.837242
|
Symmetric-Key Based Proofs of Retrievability Supporting Public Verification
|
02143a479ab76c1a8793d3669a25e72ead827869
|
European Symposium on Research in Computer Security
|
[
{
"authorId": "10770425",
"name": "Chaowen Guan"
},
{
"authorId": "144222395",
"name": "K. Ren"
},
{
"authorId": "3283592",
"name": "Fangguo Zhang"
},
{
"authorId": "1682949",
"name": "F. Kerschbaum"
},
{
"authorId": "46380550",
"name": "Jia Yu"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"ESORICS",
"Eur Symp Res Comput Secur"
],
"alternate_urls": null,
"id": "0bddd5d7-2897-495a-a961-465abe6e04de",
"issn": null,
"name": "European Symposium on Research in Computer Security",
"type": "conference",
"url": "http://www.wikicfp.com/cfp/program?id=923"
}
| null |
# Symmetric-Key Based Proofs of Retrievability Supporting Public Verification
Chaowen Guan[1(][B][)], Kui Ren[1], Fangguo Zhang[1][,][2][,][3], Florian Kerschbaum[4],
and Jia Yu[1][,][5]
1 Department of Computer Science and Engineering,
University at Buffalo, Buffalo, USA
_{chaoweng,kuiren}@buffalo.edu, isszhfg@mail.sysu.edu.cn_
2 School of Information Science and Technology, Sun Yat-sen University,
Guangzhou, China
3 Guangdong Key Laboratory of Information Security Technology,
Guangzhou, China
4 SAP, Karlsruhe, Germany
florian.kerschbaum@sap.com
5 College of Information Engineering, Qingdao University, Qingdao, China
**Abstract. Proofs-of-Retrievability enables a client to store his data on**
a cloud server so that he executes an efficient auditing protocol to check
that the server possesses all of his data in the future. During an audit,
the server must maintain full knowledge of the client’s data to pass, even
though only a few blocks of the data need to be accessed. Since the first
work by Juels and Kaliski, many PoR schemes have been proposed and
some of them can support dynamic updates. However, all the existing
works that achieve public verifiability are built upon traditional publickey cryptosystems which imposes a relatively high computational burden
on low-power clients (e.g., mobile devices).
In this work we explore indistinguishability obfuscation for building a
Proof-of-Retrievability scheme that provides public verification while the
encryption is based on symmetric key primitives. The resulting scheme
offers light-weight storing and proving at the expense of longer verification. This could be useful in apations where outsourcing files is usually
done by low-power client and verifications can be done by well equipped
machines (e.g., a third party server). We also show that the proposed
scheme can support dynamic updates. At last, for better assessing our
proposed scheme, we give a performance analysis of our scheme and a
comparison with several other existing schemes which demonstrates that
our scheme achieves better performance on the data owner side and the
server side.
**Keywords: Cloud storage · Proofs of retrievability · Indistinguishabil-**
ity obfuscation
## 1 Introduction
Nowadays, storage outsourcing (e.g., Google Drive, Dropbox, etc.) is becoming
increasingly popular as one of the applications of cloud computing. It enables
_⃝c_ Springer International Publishing Switzerland 2015
-----
clients to access the outsourced data flexibly from any location. However, the
storage provider (i.e., server) is not necessarily trusted. This situation gives rise
to a need that a data owner (i.e., client) can efficiently verify that the server
indeed stores the entire data. More precisely, a client can run an efficient audit
protocol with the untrusted server where the server can pass the audit only
if it maintains knowledge of the client’s entire outsourced data. Formally, this
implies two guarantees that the client wants from the server: Authenticity and
_Retrievability. Authenticity ensures that the client can verify the correctness of_
the data fetched from the server. On the other hand, Retrievability provides
assurance that the client’s data on the server is intact and no data loss has
occurred. Apparently, the client should not need to download the entire data
from server to verify the data’s integrity, since this may be prohibitive in terms
of bandwidth and time. Also, it is undesirable for the server to read all of the
client’s outsourced data during an audit protocol.
One method that achieves the above is called Proofs of Retrievability (PoR)
which was initially defined and constructed by Juels and Kaliski [1]. Mainly, PoR
schemes can be categorized into two classes: privately verifiable ones and publicly verifiable ones. Note that privately verifiable PoR systems normally only
involve symmetric key primitives, which are cheap for the data owner in encrypting and uploading its files. However, in such systems the guarantees of the data’s
authenticity and retrievability largely depend on the data owners themselves due
to the fact that they need to regularly perform verifications (e.g., auditing) in
order to react as early as possible in case of a data loss. Nowadays, users create
and upload data everywhere using low power devices, such as mobile phones.
Obviously, such privately verifiable PoR system would inevitably impose expensive burdens on low power data owners in the long run. On the other hand, in
this scenario with low power users, it is reasonable to have a well equipped server
(trusted or semi-trusted) perform auditing on behalf of data owner which requires
publicly verifiable PoR systems. However, all of the existing PoR schemes that
achieve public verifiability are constructed based on traditional public key cryptography which implies more complex and expensive computations compared to
simple and symmetric key cryptographic primitives. (This observation can also
be spotted in outsourced computing schemes that support public verification
[34–36].) That means a PoR scheme using public key cryptographic primitives
incurs relatively expensive overheads on low-capability clients. One might want
to construct a public verifiable PoR scheme without relying on traditional public key cryptographic primitives. One cryptographic primitive that can help to
overcome this constraint is indistinguishability obfuscation (i ) which achieves
_O_
that obfuscations of any two distinct (equal-size) programs that implement the
same functionality are computationally indistinguishable from each other. i
_O_
has become so important since the recent breakthrough result of Garg et al. in
[2]. Garg et al. proposed the first candidate construction of an efficient indistinguishability obfuscator for general programs which are written as boolean
circuits. Subsequently, Sahai and Waters [3] showed the power of i as a cryp_O_
tographic primitive: they used i to construct denial encryption, public-key
_O_
encryption, and much more from pseudorandom functions. Most recently, by
-----
exploiting i, Ramchen et al. [4] built a fully secure signature scheme with fast
_O_
signing and Boneh et al. [5] proposed a multiparty key exchange protocol, an
efficient traitor tracing system and more.
**Our work. In this paper, we explore this new primitive, i**, for building PoR.
_O_
In particular, we modify Shacham and Waters’ privately verifiable PoR scheme
[6] and apply i to construct a publicly verifiable PoR scheme. Our results share
_O_
a similar property with Ramchen et al.’s signing scheme [4], that is, storing and
proving are fast at the expense of longer public verification. Such “imbalance”
could be useful in applications where outsourcing files is usually done by lowpower client and verifications can be done by well equipped machines (a semitrusted third party). Our contributions are summarized as follows:
1. We explore building proof-of-retrievability systems from obfuscation. The
resulting PoR scheme offers light-weight outsourcing, because it requires only
symmetric key operations for the data owner to upload files to the cloud
server. Likewise, the server also requires less workload during an auditing
compared to existing publicly verifiable PoR schemes.
2. We show that the proposed PoR scheme can support dynamic updates by
applying the Merkle hash tree technique. We first build a modified B+ tree
over the file blocks and the corresponding block verification messages σ. Then
we apply the Merkle hash tree to this tree for ensuring authenticity and
freshness.
3. Note that the current i construction candidate will incur a large amount of
_O_
overhead for generating obfuscation, but it is only a one-time cost during the
preprocessing stage of our system. Therefore its cost can be amortized over
plenty of future operations. Except for this one-time cost, we show that our
proposed scheme achieves good performance on the data owner side and the
cloud server side by analysis and comparisons with other recent existing PoR
schemes.
Indistinguishability obfuscation indeed provides attractive and interesting features, but the current i candidate construction offers impractical generation
_O_
and evaluation. Given the fact that the development of i is still in its nascent
_O_
stages, in Appendix, we discuss several possible future directions in works on
obfuscation in addition to those discussed in [2].
**1.1** **Related Work**
**Proof of Retrievability and Provable Data Possession. The first PoR**
scheme was defined and constructed by Juels and Kaliski [1], and the first Provable Data Possession (PDP) was concurrently defined by Ateniese et al. [7].
The main difference between PoR and PDP is the notion of security that they
achieve. Concretely, PoR provides stronger security guarantees than PDP does.
A successful PoR audit guarantees that the server maintains knowledge of all
of the client’s outsourced data, while a successful PDP audit only ensures that
-----
the server is retaining most of the data. That means, in a PDP system a server
that lost a small amount of data can still pass an audit with significant probability. Some PDP schemes [8] indeed provide full security. However, those schemes
requires the server to read the client’s entire data during an audit. If the data is
large, this becomes totally impractical. A detailed comparison can be found in
[9]. Since the introduction of PoR and PDP they have received much research
attention. On the one hand, subsequent works [6, 10–12] for static data focused
on the improvement of communication efficiency and exact security. On the
other hand, the works of [13–15] showed how to construct dynamic PDP scheme
supporting efficient updates. Although many efficient PoR schemes have been
proposed since the work of Juels et al., only a few of them supports efficient
dynamic update [16–18].
Observe that in publicly verifiable PoR systems, an external verifier (called
auditor) is able to perform an auditing protocol with the cloud server on behalf
of the data owner. However, public PoR systems do not provide any security
guarantees when the user and/or the external verifier are dishonest. To address
this problem Armknecht et al. recently introduced the notion of outsourced proofs
_of retrievability (OPoR) [19]. In particular, OPoR protects against the collusion_
of any two parties among the malicious auditor, malicious users and the malicious cloud server. Armknecht et al. proposed a concrete OPoR scheme, named
Fortress, which is mainly built upon the private PoR scheme in [6]. In order to
be secure in the OPoR security model, Fortress also employs a mechanism that
enables the user and the auditor to extract common pseudorandom bits using a
time-dependent source without any interaction.
**Indistinguishability Obfuscation. Program obfuscation aims to make com-**
puter programs “unintelligible” while preserving their functionality. The formal
study of obfuscation was started by Barak et al. [20] in 2001. In their work,
they first suggested a quite intuitive notion called virtual black-box obfuscation, for which they also showed impossibility. Motivated by this impossibility,
they proposed another important notion of obfuscation called indistinguishability
_obfuscation (i_ ), which asks that obfuscations of any two distinct (equal-size)
_O_
programs that implement the same functionalities are computationally indistinguishable from each other. A recent breakthrough result by Garg et al. [2] presented the first candidate construction of an efficient indistinguishability obfuscator for general programs that are written as boolean circuits. The proposed
construction was build on the multilinear map candidates [21, 22].
The works of Garg et al. [2] also showed how to apply indistinguishability
obfuscation to the construction of functional encryption schemes for general circuits. In subsequent work, Sahai and Waters [3] formally investigated what can
be built from indistinguishability obfuscation and showed the power of indistinguishability obfuscation as a cryptographic primitive. Since then, many new
applications of general-purpose obfuscation have been explored [24–28]. Most
recently, the works of Boneh et al. [5] and Ramchen et al. [4] re-explore the constructions of some existing cryptographic primitives through the lens of obfuscation, including broadcast encryption, traitor tracing and signing. Those proposed
-----
constructions indeed obtain some attractive features, although current obfuscation candidates incur prohibitive overheads. Precisely, Boneh et al.’s broadcast encryption achieves that ciphertext size is independent of the number of
users, and their traitor tracing system achieves full collusion resistance with
short ciphertexts, secret keys and public keys. On the other hand, Ramchen
et al. [4] proposed an imbalanced signing algorithm, which is ideally significantly
faster than comparable signatures that are not built upon obfuscation. Here
“imbalanced” means the signing is fast at the expense of longer verification.
## 2 Preliminaries
In this section we define proof-of-retrievability, indistinguishability obfuscation,
and variants of pseudorandom functions (PRFs) that we will make use of. All the
variants of PRFs that we consider will be constructed from one-way functions.
**2.1** **Proofs of Retrievability**
Below, we give the definition of publicly verifiable PoR scheme in a way similar
to that in [6]. A proof of retrievability scheme defines four algorithms, KeyGen,
Store, Prove and Verify, which are specified as follows:
(pk, sk) **KeyGen(1[λ]). On input the security parameter λ, this randomized**
_←_
algorithm generates a public-private keypair (pk, sk).
(M _[∗], t) ←Store(sk, M_ ). On input a secret key sk and a file M ∈{0, 1}[∗], this
algorithm processes M to produce M _[∗], which will be stored on the server,_
and a tag t. The tag t contains information associated with the file M _[∗]._
(0, 1) **Audit(Prove, Verify). The randomized proving and verifying algo-**
_←_
rithms together define an Audit-protocol for proving file retrievability. During
protocol execution, both algorithms take as input the public key pk and the
file tag t output by Store. Prove algorithm also takes as input the processed
file description M _[∗]_ that is output by Store, and Verify algorithm takes as
input public verification key V K. At the end of the protocol, Verify outputs
0 or 1, with 1 indicating that the file is being stored on the server. We denote
a run of two parties executing such protocol as:
_{0, 1} ←_ (Verify(pk, V K, t) ⇌ **Prove(pk, t, M** _[∗]))._
**Correctness. For all keypairs (pk, sk) output by KeyGen, for all files M**
_∈_
_{0, 1}[∗], and for all (M_ _[∗], t) output by Store(sk, M_ ), the verification algorithm
accepts when interacting with the valid prover:
(Verify(pk, V K, t) ⇌ **Prove(pk, t, M** _[∗])) = 1._
-----
**2.2** **Obfuscation Preliminaries**
We recall the definition of indistinguishability obfuscation from [2, 3].
**Definition 1. Indistinguishability Obfuscation (i** _). A uniform PPT machine_
_O_
_iO is called an indistinguishability obfuscator for a circuit class {Cλ}λ∈N if the_
_following conditions are satisfied:_
_– For all security parameters λ ∈_ N, for all C ∈Cλ, for all inputs x, we have
_that Pr[C_ _[′](x) = C(x) : C_ _[′]_ _i_ (λ, C)] = 1.
_←_ _O_
_– For any (not necessarily uniform) PPT distinguisher (Samp, D), there exists_
_a negligible function negl(_ ) such that the following holds: if for all security
_·_
_parameters λ ∈_ N, Pr[∀x, C0(x) = C1(x) : (C0; C1; τ ) ← _Samp(1[λ])] > 1 −_
_negl(λ), then we have_
_|Pr[D(τ, iO(λ, C0)) = 1 : (C0; C1; τ_ ) ← _Samp(1[λ])]−_
Pr[D(τ, iO(λ, C1)) = 1 : (C0; C1; τ ) ← _Samp(1[λ])]| ≤_ _negl(λ)._
**2.3** **Puncturable PRFs**
A pseudorandom function (PRF) is a function F : with K $ such
_K×M →Y_ _←K_
that the function F (K, ) is indistinguishable from random. A constrained PRF
_·_
[29] is a PRF F (K, ) that is able to evaluate at certain portions of the input
_·_
space and nowhere else. A puncturable PRF [3, 29] is a type of constrained PRF
that enables the evaluation at all bit strings of a certain length, except for any
polynomial-size set of inputs. Concretely, it is defined with two PPT algorithms
(EvalF, PunctureF ) such that the following two properties hold:
– Functionality Preserved under Puncturing. For every PPT algorithm
with input 1[λ] outputs a set S 0, 1, for all x 0, 1 _S, we have_
_A_ _⊆{_ _}[n]_ _∈{_ _}[n]\_
Pr[EvalF (K{S}, x) = F (K, x) : K _←K$_ _, K{S} ←_ PunctureF (K, S)] = 1
– Pseudorandom at Punctured Points. For every pair of PPT algorithms
(A1, A2) such that A1(1[λ]) outputs a set S ⊆{0, 1}[n] and a state σ, consider
an experiment where K _←K$_ _, K{S} ←_ PunctureF (K, S). It holds that
_|Pr[A2(σ, K{S}, S, F_ (K, S)) = 1)]−
Pr[A2(σ, K{S}, S, Um(λ)·|S|) = 1]| ≤ _negl(λ)_
## 3 Security Definitions
The security definitions of Authenticity and Retrievability in [17, 18] are essentially equivalent to the security definition of Soundness in [6]. Note that the
security definitions in [17, 18] are for dynamic PoR systems, while the one in
[6] considers only static PoR systems. The only difference between a static PoR
-----
scheme and a dynamic PoR scheme is that the latter one supports secure dynamic
updates, including modification, deletion and insertion. This affects the access
to oracles in the security game. Below we present the security definitions for
static PoR systems in the same way as [17, 18] and then point out how to obtain
the security definitions for dynamic PoR systems based on the static one.
**3.1** **Security Definitions on Static PoR**
**Authenticity. Authenticity requires that the client can always detect if any**
message sent by the server deviates from honest behavior. More precisely, consider the following game between a challenger, a malicious server and an
_C_ _S[�]_
honest server for the adaptive version of authenticity:
_S_
– The challenger initializes the environment and provides with public para_S[�]_
meters.
– The malicious sever _S[�] specifies a valid protocol sequence P = (op1, op2, · · ·,_
_oppoly(λ)) of polynomial size in the security parameter λ. The specified oper-_
ations opt can be either Store or Audit. C executes the protocol with both _S[�]_
and an honest server .
_S_
If at execution of any opj, the message sent by _S[�] differs from that of the honest_
server and does not output reject, the adversary wins and the game results
_S_ _C_ _S[�]_
in 1, else 0.
**Definition 2. A static PoR scheme is said to satisfy adaptive Authenticity, if**
_any polynomial-time adversary_ _wins the above security game with probability_
_S[�]_
_no more than negl(λ)._
**Retrievability. Retrievability guarantees that whenever a malicious server can**
pass the audit test with non-negligible probability, the server must know the
entire content of ; and moreover, can be recovered by repeatedly running
_M_ _M_
the Audit-protocol between the challenger and the server . More precisely,
_C_ _S[�]_
consider the following security game:
– The challenger initializes the environment and provides with public para_S[�]_
meters.
– The malicious server _S[�] specifies a protocol sequence P = (op1, op2, · · ·,_
_oppoly(λ)) of polynomial size in terms of the security parameter λ. The speci-_
fied operations opt can be either Store or Audit. Let M be the correct content
value.
– The challenger sequentially executes the respective protocols with . At the
_C_ _S[�]_
end of executing P, let stC and st �S [be the final configurations (states) of the]
challenger and the malicious server, respectively.
– The challenger now gets black-box rewinding access to the malicious server in
its final configuration st
�S [. Starting from the configurations (][st][C][, st][ �]S [), the chal-]
lenger runs the Audit-protocol repeatedly for a polynomial number of times
with the server and attempts to extract out the content value as .
_S[�]_ _M[′]_
-----
If the malicious server passes the Audit-protocol with non-negligible probability
_S[�]_
and the extracted content value =, then this game outputs 1, else 0.
_M[′]_ _̸_ _M_
**Definition 3. A static PoR scheme is said to satisfy Retrievability, if there exists**
_an efficient extractor_ _such that for any polynomial-time_ _, if_ _passes the_
_E_ _S[�]_ _S[�]_
_Audit-protocol with non-negligible probability, and then after executing the Audit-_
_protocol with_ _for a polynomial number of times, the extractor_ _outputs content_
_S[�]_ _E_
_value_ = _only with negligible probability._
_M[′]_ _̸_ _M_
The above says that the extractor will be able to extract out the correct
_E_
content value = if the malicious server can maintain a non-negligible
_M[′]_ _M_ _S[�]_
probability of passing the Audit-protocol. This means the server must retain full
knowledge of .
_M_
**3.2** **Security Definitions on Dynamic PoR**
The security definitions for dynamic PoR systems are the same as those for static
PoR systems, except that the oracles which the malicious server has access to
_S[�]_
are including Read, Write and Audit. Precisely, the security game for Authenticity
is the same as the for static PoR schemes, except that the malicious server
_S[�]_
can get access to Read, Write and Audit oracles. This means that the specified
operations opt by _S[�] in the protocol sequence P = (op1, op2, · · ·, oppoly(λ)) can_
be either Read, Write or Audit. Similarly, the security game for Retrievability is
the same as that for static PoR systems, except that the malicious server can
_S[�]_
get access to Read, Write and Audit oracles. Note that the winning condition for
both games remain unchanged.
## 4 Constructions
In this section we first give the construction of a static publicly verifiable PoR
system. Then we discuss how to extend this static PoR scheme to support efficient dynamic updates.
Before presenting our proposed constructions, we analyze a trivial construction of a publicly verifiable PoR scheme using i . Let n be the number of file
_O_
blocks, λ1 be the size of a file block (here assume every file block is equally
large), λ2 be the size of a block tag σ and I be the challenge index set requested
by the verifier. Since i can hide secret information, which is embedded into the
_O_
obfuscated program, from the users, one might construct a scheme as: (1) set
the tag for a file block mi as the output of a PRF F (k, mi) with secret key k;
(2) embed key k into the verification program and obfuscate it; (3) this verification program simply checks the tags for the challenged file blocks to see if
they are valid outputs of the PRF. Observe that this verification program takes
as inputs a challenge index set, the challenged file blocks and the corresponding file tags. Therefore, the circuit for this verification program will be of size
_O(poly(|I| · log n + |I| · λ1 + |I| · λ2)), where |I| is the size of index set I and_
_poly(x) is a polynomial in terms of x. Clearly, this method also costs much a lot_
of bandwidth due to the fact that it does not provide an aggregated proof.
While in our construction we modify the privately verifiable PoR scheme
in [6]. For consistency with the above analysis, assume that file blocks are not
-----
further divided into sectors. Then the verification program takes as input a
challenge index set I, an aggregation of the challenged file blocks μ and an
aggregated σ[′]. Consequently the circuit for the verification program will have
size O(poly(|I| · log n + λ1 + λ2)), which is much smaller than that in the trivial
construction. Clearly, the trivial construction will lead to a significantly larger
obfuscation of the verification program.
Similarly, we analyze the circuit’s size when a file block is further split into
_s sectors, as the scheme in [6] did. Let the size of a sector in a file block be λ3._
The circuit size in the trivial construction will remain unchanged, O(poly( _I_
_|_ _| ·_
log n + |I| · λ1 + |I| · λ2)). While the circuit in our construction will have size
_O(poly(|I| · log n + s · λ3 + λ3)) ≈_ _O(poly(|I| · log n + λ1 + λ3)), which is still_
much smaller than that in the trivial construction. As we can see, exploiting i
_O_
is not trivial although it is a powerful cryptographic primitive.
**4.1** **Static Publicly Verifiable PoR Scheme**
We modify Shacham and Waters’ privately verifiable PoR scheme in [6] and
combine it with i to give a publicly verifiable PoR scheme. Recall that in the
_O_
scheme in [6], a file F is processed using erasure code and then divided into n
blocks. Also note that each block is split into s sectors. This allows for a tradeoff
between storage overhead and communication overhead, as discussed in [6].
Before presenting the construction of the proposed static PoR scheme, we
give a brief discussion on how we apply indistinguishability obfuscation to the
PoR scheme in [6]. For doing that, we need to utilize a key technique introduced
in [3], named punctured programs. At a very high-level, the idea of this technique
is to modify a program (which is to be obfuscated) by surgically removing a key
element of the program, without which the adversary cannot win the security
game it must play, but in a way that does not change the functionality of the
program. Note that, in Shacham and Waters’ PoR scheme, for each file block, σi
is set as fprf (i) + [�]j[s]=1 _[α][j][m][ij][, where the secret key][ k][prf][ for PRF][ f][ is specific]_
for one certain file M . That means for different files, it uses different PRF key
_kprf_ ’s. As to make it a punctured PRF that we want in the obfuscated program,
we eliminate this binding between PRF key kprf and file M, and the same PRF
key kprf will be used in storing many different files. Thus, the PRF key kprf
will be randomly chosen in client KeyGen step, not in Store step. The security
will be maintained after this modification, due to the fact that it still provides
_σi with randomness without adversary getting the PRF key._
The second main change is related to the construction of a file tag t.
Note that, in Shacham and Waters’ scheme, t = n∥c∥MACkmac(n∥c), where
_c = Enckenc(kprf_ _∥α1∥· · · ∥αs). In our proposed scheme, the randomly selected_
elements α1, · · ·, αs will be removed. Instead, we use another PRF key fprf ′ to
generate s pseudorandom numbers, which will reduce the communication cost by
(s · ⌈log p⌉), where log p means each element αi ∈ Zp. As a consequence of these
two changes, the symmetric key encryption component c is no longer needed and
_σi will be made as fprf_ (i) + [�]j[s]=1 _[f][prf][ ′][(][j][)][ ·][ m][ij][.]_
-----
Let F1(k1, ·) be a puncturable PRF mapping ⌈log N _⌉-bit inputs to ⌈log Zp⌉. Here_
_N is a bound on the number of blocks in a file. Let F2(k2, ·) be a puncturable_
PRF mapping ⌈log s⌉-bit inputs to ⌈log Zp⌉. Let SSigssk(x) be the algorithm
generating a signature on x.
**KeyGen(). Randomly choose two PRF key k1** 1, k2 2 and a random
_∈K_ _∈K_
signing keypair (svk, ssk) _←R_ SKg. Set the secret key sk = (k1, k2, ssk). Let
the public key be svk along with the verification key VK which is an indistinguishability obfuscation of the program Check defined as below.
**Store(sk, M** ). Given file M and secret key sk = (k1, k2, ssk), proceed as follows:
1. apply the erasure code to M to obtain M _[′];_
2. split M _[′]_ into n blocks, and each block into s sectors to get {mij} for
1 _i_ _n, 1_ _j_ _s;_
_≤_ _≤_ _≤_ _≤_
3. set the file tag t = n∥SSigssk(n)
4. for each i, 1 ≤ _i ≤_ _n, compute σi = F1(k1, i) +_ [�]j[s]=1 _[F][2][(][k][2][, j][)][ ·][ m][ij][;]_
5. set as the outputs the processed file M _[′]_ = {mij}, 1 ≤ _i ≤_ _n, 1 ≤_ _j ≤_ _s,_
the corresponding file tag t and {σi}, 1 ≤ _i ≤_ _n._
**Verify(svk, V K, t). Given the tag t, parse t = n∥SSigssk(n) and use svk to verify**
the signature on t; if the signature is invalid, reject and halt. Otherwise, pick
a random l-element subset I from [1, n], and for each i _I, pick a random_
_∈_
element vi ∈ Zp. Send set Q = {(i, vi)} to the prover.
Parse the prover’s response to obtain μ1, · · ·, μs, σ ∈ Z[s]p[+1]. If parsing fails,
reject and halt. Otherwise, output VK(Q = {(i, vi)}i∈I _, μ1, · · ·, μs, σ)._
Check:
Inputs: Q = {(i, vi)}i∈I _, μ1, · · ·, μs, σ_
Constants: PRF keys k1, k2
**if σ =** [�](i,vi)∈Q _[v][i][ ·][ F][1][(][k][1][, i][) +][ �]j[s]=1_ _[F][2][(][k][2][, j][)][ ·][ μ][j][ then][ output 1]_
**else output**
_⊥_
**Prove(t, M** _[′]). Given the processed file M_ _[′], {σi}, 1 ≤_ _i ≤_ _n and an l-element_
set Q sent by the verifier, parse M _[′]_ = {mij}, 1 ≤ _i ≤_ _n, 1 ≤_ _j ≤_ _s and_
_Q = {(i, vi)}. Then compute_
� �
_μj =_ _vimij for 1 ≤_ _j ≤_ _s,_ and _σ =_ _viσi,_
(i,vi)∈Q (i,vi)
and send to the prove in response the values μ1, · · ·, μs and σ.
**4.2** **PoR Scheme Supporting Efficient Dynamic Updates**
A PoR scheme supporting dynamic updates means that it enables modification,
deletion and insertion over the stored files. Note that, in the static PoR scheme,
each σi associated with mij 1≤j≤s is also bound to a file block index i. If an
update is executed in this static PoR scheme, it requires to change every σi corresponding to the involved file blocks, and the cost could probably be expensive.
Let’s say the client needs to insert a file block Fi into position i. We can see
that this insertion manipulation requires to update the indices in σj’s for all
_i ≤_ _j ≤_ _n. On average, a single insertion incurs updates on n/2 σj’s._
-----
In order to offer efficient insertion, we need to disentangle σi from index i.
Concretely, F1(k1, ·) should be erased in the computing of σi, which leads to a
modified σi[′] [=][ �][s]j=1 _[F][2][(][k][2][, j][)][ ·][ m][ij][. However, this would make the scheme inse-]_
cure, because a malicious server can always forge, e.g., σi[′][/][2 =][ �][s]j=1 _[F][2][(][k][2][, j][)][ ·]_
(mij/2) for file block {mij/2}1≤j≤s with this σi[′][.]
Instead, we build σi as F1(k1, ri)+ [�][s]j=1 _[F][2][(][k][2][, j][)]_ _[·]_ _[m][ij][, where][ r][i][ is a random]_
element from Zp. Clearly, we can’t maintain the order of the stored file blocks
without associating σi with index i. To provide the guarantee that every upto-date file block is in the designated position, we use a modified B+ tree data
structure with standard Merkle hash tree technique.
Observe that, unlike Shacham and Waters’ scheme where the file is split into
_n blocks after being erasure encoded, the construction here assumes that each file_
block is encoded ‘locally’. (Cash et al.’s work [17] also started with this point.)
That is, instead of using an erasure code that takes the entire file as input, we
use a code that works on small blocks. More precisely, the client divides the file
_M into n blocks, i.e., M = (m1, m2, · · ·, mn), and then encodes each file block_
_mi individually into a corresponding codeword block ci = encode(mi). Next, the_
client performs the following PoR scheme to create σi for each ci. Auditing works
as before: The verifier randomly selects l indices from [1, n] and l random values,
and then challenges the server to respond with a proof that is computed with
those l random values and corresponding codewords specified by the l indices.
Note that, in this construction, each codeword ci will be further divided into s
sectors, (ci1, ci2, · · ·, cis) during the creation of σi. A more detailed discussion
about this and analysis of how to better define block size can be found in the
appendices in [6, 17].
Let F1(k1, ·) be a puncturable PRF mapping ⌈log N _⌉-bit inputs to ⌈log Zp⌉. Here_
_N is a bound on the number of blocks in a file. Let F2(k2, ·) be a puncturable_
PRF mapping ⌈log s⌉-bit inputs to ⌈log Zp⌉. Let Enck/Deck be a symmetric key
encryption/decryption algorithm, and SSigssk(x) be the algorithm generating a
signature on x.
**KeyGen(). Randomly choose puncturable PRF keys k1** 1 k2 2,
_∈K_ _∈K_
a symmetric encryption key kenc _enc and a random signing keypair_
_∈K_
(svk, ssk) _←R_ SKg. Set the secret key sk = (k1, k2, kenc, ssk). Let the public
key be svk along with the verification key VK which is an indistinguishability
obfuscation of the program CheckU defined as below.
**Store(sk, M** ). Given file M and secret key sk = (k1, k2, kenc, ssk), proceed as
follows:
1. split M _[′]_ into n blocks and apply the erasure code to each block mi to
obtain the codeword block m[′]i[, then divide each block][ m][′]i [into][ s][ sectors to]
get {m[′]ij[}][ for 1][ ≤] _[i][ ≤]_ _[n,][ 1][ ≤]_ _[j][ ≤]_ _[s][;]_
2. for each i, 1 ≤ _i ≤_ _n, choose a random element ri ∈_ Zp and compute
_σi = F1(k1, ri) +_ [�]j[s]=1 _[F][2][(][k][2][, j][)][ ·][ m]ij[′]_ [;]
3. set c = Enckenc(r1∥· · · ∥rn) and the file tag t = n∥c∥SSigssk(n∥c);
4. set as the outputs the processed file M _[′]_ = {m[′]ij[}][, 1][ ≤] _[i][ ≤]_ _[n,][ 1][ ≤]_ _[j][ ≤]_ _[s][,]_
the corresponding file tag t and {σi}, 1 ≤ _i ≤_ _n._
-----
**Verify(svk, V K, t). Given the file tag t, parse t = n∥c∥SSigssk(n∥c) and use**
_svk to verify the signature on t; if the signature is invalid, reject and halt._
Otherwise, pick a random l-element subset I from [1, n], and for each i _I,_
_∈_
pick a random element vi ∈ Zp. Sent set Q = {(i, vi)} to the prover.
Parse the prover’s response to obtain μ1, · · ·, μs, σ ∈ Z[s]p[+1]. If parsing fails,
reject and halt. Otherwise, output VK(Q = {(i, vi)}i∈I _, μ1, · · ·, μs, σ, t)._
CheckU:
Inputs: Q = {(i, vi)}i∈I _, μ1, · · ·, μs, σ, t_
Constants: PRF keys k1, k2, symmetric encryption key kenc
_n∥c∥SSigssk(n∥c) ←_ _t_
_r1, · · ·, rn ←_ _Deckenc_ (c)
**if σ =** [�](i,vi)∈Q _[v][i][ ·][ F][1][(][k][1][, r][i][) +][ �]j[s]=1_ _[F][2][(][k][2][, j][)][ ·][ μ][j][ then][ output 1]_
**else output**
_⊥_
**Prove(t, M** _[′]). Given the processed file M_ _[′], {σi}, 1 ≤_ _i ≤_ _n and an l-element_
set Q sent by the verifier, parse M _[′]_ = {m[′]ij[}][,][ 1][ ≤] _[i][ ≤]_ _[n,][ 1][ ≤]_ _[j][ ≤]_ _[s][ and]_
_Q = {(i, vi)}. Then compute_
� �
_μj =_ _vim[′]ij_ [for 1][ ≤] _[j][ ≤]_ _[s,]_ and _σ =_ _viσi,_
(i,vi)∈Q (i,vi)
and send to the prove in response the values μ1, · · ·, μs and σ.
**Modified B+ Merkle tree. In our construction, we organize the data files**
using a modified B+ tree, and then apply a standard Merkle Hash tree to provides guarantees of freshness and authenticity. In this modified B+ tree, each
node has at most three entries. Each entry in leaf node is data file’s σ and is
linked to its corresponding data file in the additional bottom level. The internal
nodes will no longer have index information. Before presenting the tree’s construction, we first define some notations. We denote an entry’s corresponding
computed σ by label( ), the rank of an entry (i.e., the number of file blocks that
_·_
can be reached from this entry) by rank( ), descendants of an entry by child( ),
_·_ _·_
left/right sibling of an entry by len( )/ren( ).
_·_ _·_
– entry w in leaf node: label(w) = σ, len(w) (if w is the leftmost entry, len(w) =
0) and ren(w) ((if w is the rightmost entry, ren(w) = 0);
– entry v in internal node and root node: rank(v), child(v) len(v) and ren(v),
where len(v) and ren(v) conform to the rules above.
An example is illustrated in Fig. 1a. Following the definitions above, entry v1
in root node R contains: (1) rank(v1) = 3, because w1, w2 and w3 can be reached
from v1; (2) child(v1) = w1∥w2∥w3; (3) len(v1) = 0; (4) ren(v1) = v2. Entry w2 in
leaf node W1 contains: (1) label(w2) = σ2; (2) len(w2) = w1; (3) ren(w2) = w3.
Note that the arrows connecting the entries in leaf nodes with F ’s means that
each entry is associated with its corresponding file block. Precisely, e.g., entry w1
is associated with the first data block F1 and label(w1) = σ1.
-----
**Fig. 1. An example of a modified B+ tree.**
To search for a σ and its corresponding file block, we need two additional
values of each entry, low( ) and high( ). low( ) gives the lowest-position data
_·_ _·_ _·_
block that can be reached from an entry, and high( ) defines the highest-position
_·_
data block that can be reached from an entry. Observe that these two values need
not be stored for every entry in the tree. We can compute them on the fly using
the ranks. For the current entry r, assume we know low(r) and high(r). Let
_child(r) = v1∥v2∥v3. Then low(vi)’s and high(vi)’s can be computed with entry’s_
_rank value in the following way: (1) low(v1) = low(r) and high(v1) = low(v1) +_
_rank(v1)_ _−_ 1; (2) low(v2) = high(v1)+1 and high(v2) = low(v2)+ _rank(v2)_ _−_ 1;
(3) low(v3) = high(v2) + 1 and high(v3) = high(r).
Using the entries’ rank values, we can reach the i-th data block (i.e., i-th
entry) in the leaf nodes. The search starts with entry v1 in root node. Clearly,
for the start entry of the tree, we have low(v1) = 1. On each entry v during the
search, if i [low(v), high(v)], we proceed the search along the pointer from v
_∈_
to its children; otherwise, check the next entry on v’s right side. We continue
until we reach the i-th data block. For instance, say we want to read the 6-th
data block in Fig. 1a. We start with entry v1, and the search proceeds as follows:
1. compute high(v1) = low(v1) + rank(v1) − 1 = 3;
2. i = 6 /∈ [low(v1), high(v1)], then check the next entry, v2;
3. compute low(v2) = high(v1) + 1 = 4, high(v2) = low(v2) + rank(v2) − 1 = 6;
4. i ∈ [low(v2), high(v2)], then follow the pointer leading to v2’s children;
5. get child(v2) = w4∥w5∥w6;
-----
6. now in leaf node, check each entry from left to right, and find w6 be the entry
connecting to the wanted data block.
Now it is only left to define the Merkle hash tree on this modified B+ tree.
Note that in our modified B+ tree, each node have at most 3 entries. Let upper
case letter denote node and lower case one denote entry. For each entry, the
hashing value is computed as follows:
– Case 0: w is an entry in a leaf node, compute f (w) = h(label(w)) = h(σ),
– Case 1: v is an entry in an internal node and it’s descendent is node V _[′],_
compute f (v) = h(rank(v)∥f (V _[′]))._
For each node (internal node or leaf node) consisting of entries v1, v2, v3 from left
to right, we define f (V ) = h(f (v1)∥f (v2)∥f (v3)). For instance, in Fig. 1.a, the
hashing value for the root node is f (R) = h(f (v1)∥f (v2)∥f (v3)), where f (vi) =
_h(rank(vi)∥f_ (Wi)) and f (Wi) = h(f (w(i−1)∗3+1)∥f (w(i−1)∗3+2)∥f (w(i−1)∗3+3)).
With this Merkle hash tree built over the modified B+ tree, the client keeps
track of the root digest. Every time after fetching a data block, the client fetches
its corresponding σ as well. Also the client receives the hashing values associated
with other entries in the same node along the path from root to the data block.
Then the client can verify the authenticity and freshness with the Merkle tree.
Say the client needs to verify the authenticity and freshness of block F3 in Fig. 1a,
where he/she possesses the root digest f (R). The path from root to F3 will be
(R → _W1). For verification, besides σ3, the client also receives f_ (w1), f (w2) in
node W1 and f (v2), f (v3) in node R.
**Update. The main manipulations are updating the data block and updating**
the Merkle tree. Note that the update affects only nodes along the path from a
wanted data block to root on the Merkle tree. Therefore, the running time for
updating the Merkle tree is O(logn). Also to update the Merkle tree, some hashing values along the path from a data block to root are needed from the server.
Clearly, the size of those values will be O(logn). Update operations include Modification, Deletion and Insertion. The update operations over our modified B+ tree
mostly conform to the procedures of standard B+ tree. A slight difference lies
in the Insertion operation when splitting node, due to the fact that our modified
B+ tree doesn’t have index information.
First, we discuss Modification and Deletion. To modify a data block, the client
simply computes the data block’s new corresponding σ and updates the Merkle
tree with this σ to obtain a new root digest. Then the client uploads the the new
data block and the new σ. After receiving this new σ, the server just needs to
update the Merkle tree along the path from the data block to root. To delete a
data block, the server simply deletes the unwanted data block by the client and
then updates the Merkle tree along the path from this data block to root.
Next, we give the details of Insertion. If the leaf node where the new data
block will be inserted is not full, the procedure is the same as Modification.
Otherwise, the leaf node needs to be split, and then the entry that leads to this
leaf node will also be split into two entries, with one entry leading to each leaf
node. Note that unlike operations on standard B+ tree, we don’t copy the index
-----
of the third entry (i.e., the index of the new generated node) to its parent’s node.
Instead, we simply create a new entry with a pointer leading to the node and
record the corresponding information as defined above. If the root node needs
to be divided, the depth of this Merkle tree will increment by 1. An example
of updating is shown as Fig. 1b and c. Say the client wants to insert a new file
block F10 in the 7-th position. First, it locates the position in the way mentioned
above. Note that we can locate the 6-th position or the 7-th position. Here we
choose to locate the 6-th position and insert a new entry w10 behind w6 in left
node W2 . (If choosing to locate the 7-th position, one should put the new entry
before w7.) Next, the information corresponding to this new file block F10 will
be written into entry w10 with a pointer pointing from w10 to F10, as shown in
Fig. 1b. Since it exceeds the maximum number of entries that a node can have,
this leaf node W2 needs to be split into two leaf nodes, W2[′] [and][ W][4] [with two non-]
empty entries in each node (this conforms to the rules of updating a B+ tree), as
shown in Fig. 1c. At the same time, a new entry v4 is created in the root node R
with a pointer leading v4 to leaf node W4. Similarly, this root node R is split into
two internal nodes, V1 and V1. Finally, a new root note R[′] is built, which has two
entries and two pointers leading to V1 and V2, respectively. Note that, now the
root node has entries r1 and r2, where r1 is the start entry of this tree, meaning
_low(r1) = 1. We also have rank(r1) = rank(V1) = rank(v1) + rank(v2) = 5 and_
_rank(r2) = 5._
**4.3** **Security Proofs**
**Theorem 1. The proposed static PoR scheme satisfied Authenticity as specified**
_in Sect. 3.1, assuming the existence of secure indistinguishability obfuscators,_
_existentially unforgeable signature schemes and secure puncturable PRFs._
**Theorem 2. The proposed static PoR scheme satisfies Retrievability as specified**
_in Sect. 3.1._
The detailed proof for Theorem 1 is given in the full version of this paper [23].
The proof for Theorem 2 will be identical to that in [6], because in our scheme, a
file is processed using erasure code before being divided into n blocks, the same
as that in [6] where the proof was divided into two parts, Sects. 4.2 and 4.3.
## 5 Analysis and Comparisons
In this section, we give an analysis of our proposed scheme and then compare it
with other two recently proposed schemes.
Our scheme requires the data owner to generate an obfuscated program during the preprocessing stage of the system. With the current obfuscator candidate,
it indeed costs the data owner a somewhat large amount of overhead, but this
is a one-time effort which can be amortized over plenty of operations in the
future. Thus, we focus on the analysis on the computation and communication
overheads incurred during writing and auditing operations rather than those in
-----
**Table 1. Comparison with existing dynamic PoRs.**
Scheme Write cost on Write Auditing cost Verifiability Dynamic
server bandwidth server read
Iris [16] _O(β)_ _O(β)_ _O(βλ[√]n)_ Private YES
Cash et al. [17] O(βλ(log n)[2]) _O(βλ(log n)[2])_ _O(βλ(log n)[2])_ Private YES
Shi et al. [18] _O(β log n) +_ _O(β) +_ _O(βλ log n)_ Public YES
_O(λ log n)_ _O(λ log n)_
This paper _O(β) +_ _O(β) +_ _O(βλ)_ Public YES
_O(λ log n)_ _O(λ log n)_
the preprocessing step. Like the private PoR system in [6] the data owner can
efficiently store files on the cloud server, and it takes the cloud server less overhead during an auditing protocol than in a public-key-based scheme. The cost
on the client device is mainly incurred by the operations over symmetric key
primitives, which are known to be much faster than public key cryptographic
primitives. The cost analysis on the server side is shown as Table 1.
In Table 1 showing a comparison with existing dynamic PoR schemes we let
_β be the block size in number of bits, λ be the security parameter and n be the_
number of blocks. We compare our scheme with the state-of-the-art scheme [18],
since a comparison between Shi et al.’s scheme and Cash et al.’s scheme is given
in [18]. Note that Shi et al.’s scheme needs amortized cost O(β log n) for writing
on the server side, due to the fact that an erasure-coding needs to be done on
the entire data file after Θ(n) updates, while our scheme uses an erasure code
that works on file blocks, instead of taking the entire file as inputs (more details
and discussions can be found in Sect. 4). That means, in our system modifying
a block does not require a change of the erasure codes of the entire file. Thus,
the cost for writing is only proportional to the block size being written. On the
other hand, during an auditing protocol, Shi et al.’s scheme incurs overhead
_O(βλ log n) on the server side, due to the features of the server-side storage_
layout. In their scheme, one single file will be stored as three parts, including
raw data part R, erasure-coded copy of the entire file C and hierarchical log
structure part H that stores the up-to-date file blocks in erasure-coded format.
Thus, during one auditing operation, Shi et al.’s scheme needs to check O(λ)
random blocks from C and O(λ) random blocks from each filled level in H.
While, in our scheme, the server performs every writing over the wanted block
directly, not storing the update block separately. Thus, our scheme only requires
_O(λ) random blocks of one file to check authenticity during auditing. (Note that_
this O(λ) usually would be Ω([√]nβ) if no pseudorandom permutation over the
locations of the file blocks is performed, because a small number proportional
to O(λ) might render the system insecure. Please refer to [17] for more details.)
Note that it is most likely that the auditing protocol is executed between a
well-equipped verification machine and the server, and the operations on server
|Scheme|Write cost on server|Write bandwidth|Auditing cost server read|Verifiability|Dynamic|
|---|---|---|---|---|---|
|Iris [16]|O(β)|O(β)|√ O(βλ n)|Private|YES|
|Cash et al. [17]|O(βλ(log n)2)|O(βλ(log n)2)|O(βλ(log n)2)|Private|YES|
|Shi et al. [18]|O(β log n) + O(λ log n)|O(β) + O(λ log n)|O(βλ log n)|Public|YES|
|This paper|O(β) + O(λ log n)|O(β) + O(λ log n)|O(βλ)|Public|YES|
-----
side only involve symmetric key primitives. Therefore, it will not have noticeable
effects on the system’s overall performance.
Clearly, the improvement in our work mainly results from i ’s power that
_O_
secret keys can be embedded into the obfuscated verification program without
secret keys being learnt by user. However, the current obfuscator candidate [2]
provides a construction running in impractical, albeit polynomial, time. (Note
that it is reasonable and useful that the obfuscated program is run on wellequipped machines.) Although i ’s generation and evaluation is not fast now
_O_
[30], studies on implementing practical obfuscation are developing fast [31]. It
is plausible that obfuscations with practical performance will be achieved in the
not too distant future. Note that the improvement on obfuscation will directly
lead to an improvement on our schemes.
## 6 Conclusions
In this paper, we explore indistinguishability obfuscation to construct a publicly
verifiable Proofs-of-Retrievability (PoR) scheme that is mainly built upon symmetric key cryptographic primitives. We also show how to modify the proposed
scheme to support dynamic updates using a combination of a modified B+ tree
and a standard Merkle hash tree. By analysis and comparisons with other existing schemes, we show that our scheme is efficient on the data owner side and the
cloud server side. Although it consumes a somewhat large amount of overheads
to generate an obfuscation, it is only a one-time effort during the preprocessing stage of the system. Therefore, this cost can be amortized over all of future
operations. Also note that the improvement on obfuscation will directly lead to
an improvement on our schemes.
**Acknowledgments. This work is supported in part by US National Science Founda-**
tion under grant CNS-1262277 and the National Natural Science Foundation of China
(Nos. 61379154 and U1135001).
## A Discussions and Future Directions Towards i O
As pointed out in [2], the current obfuscation constructions runs in impractical
polynomial-time, and it is an important objective to improve the efficiency for
_i_ usage in real life applications. Also Apon et al.’s showed the inefficiency in
_O_
_i_ ’s generation and evaluation in [30]. In this section, we give discussions on
_O_
three possible future directions in Obfuscation, in addition to those in [2].
**A.1** **Outsourced and Joint Generation of Indistinguishability**
**Obfuscation**
Image the scenario in our proposed publicly verifiable PoR system, where users
store their data on the same cloud server using the same PoR scheme but with
-----
different secret keys. One naive approach with i would be requiring each user
_O_
to generate his/her own individual obfuscated program for public verification.
This means that each user needs to afford the prohibitively expensive overhead
for i ’s generation on his/her own. Note that for the same PoR scheme, the
_O_
verification procedures are the same but with different user’s secret key. Also note
that each user “embeds” his/her own secret keys into the obfuscated verification
in a way that anyone else can’t learn anything about the embedded secret values.
Hence, we can have several users jointly and securely generate one obfuscated
verification program, where each user uses his/her own secret key as part of the
input to the generation. One promising way could be using Secure multiparty
computation. Observe that this generated obfuscated program has almost the
same computation as the one with only one user’s secret key embedded. The only
differences between this jointly generated obfuscation and the individual-usergenerated obfuscation are that (1) the jointly generated obfuscation is implanted
with more than one user’s secret key; (2) the jointly generated obfuscation needs
one more step to identify which user’s secret key it will use.
On the other hand, outsourced computing is useful in applications where
relatively low-power devices need to compute expensive and time-consuming
functions. Clearly, as for relatively low-power individual computers, the overhead
caused by the current i construction candidate is impractical. Thus, it would
_O_
be promising to find a specific way to efficiently outsource i ’s generation.
_O_
**A.2** **Reusability and Universality of Indistinguishability Obfuscation**
Reusability is related to i ’s joint generation to some extent. In the scenario
_O_
considered above, the jointly generated obfuscated program is embedded with a
group of users’ private key. This means that the same obfuscated program can
be used by verifiers on behalf of different users in this group.
Universality is relevant to an obfuscated program’s functionalities. More
Concretely, an universal i is supposed to support multiple functionalities. A
_O_
_straightforward example would be the obfucation-based functional encryption_
scheme in [2]. Recall that in their construction, the secret key skf for a function
_f is an obfuscated program. For this obfuscated program to become universal,_
_skf would need to be associated with more than one function. In this case,_
e.g., an universal obfuscated program skf can be associated with a class of
similar functions f = (f1, f2, · · ·, fk). This means that skf ’s holder can obtain
_f1(m), f2(m), · · ·, fk(m) from an encryption of m._
Recently, Hohenberger et al. [32] has shown that i can provide some other
_O_
cryptographic primitives with universality. They employed i to construct uni_O_
versal signature aggregators, which can aggregate across schemes in various algebraic settings (e.g., RSA, BLS). Prior to this universal signature aggregator, the
aggregation of signatures can only be built if all the signers use the same signing
algorithm and shared parameters. On the contrary, the universal signature aggregator enables the aggregation of the users’ signatures without requiring them to
execute the same signing behavior, which indicates a compressed authentication
overhead.
-----
**A.3** **Obfuscation for Specific Functions**
The current i construction candidate provides a way for obfuscating general cir_O_
cuits and runs in impractical polynomial-time. Note that an obfuscation designed
for some particular simple function with practical performance, such as computing two vectors’ inner product, can also be wanted. (like Wee’s work in STOC’05
[33]) This means that we want to obfuscate such simple functions in a practical
way that might be specific for those functions. Note that, for example, a practical obfuscated program computing the inner product of two vectors, where one
vector is an input to this program and the other one is embedded into the program without user learning its knowledge, could be useful in applications like
computational biometrics. Also, it is really likely that such a practical obfuscation for a specified function can be used as a building block to construct an
obfuscation supporting more complex functionalities by combining with other
existing practical cryptographic primitives.
## References
1. Juels, A., Kaliski, Jr., B.S.: PORs: Proofs of retrievability for large files. In: ACM
CCS, pp. 584–597 (2007)
2. Garg, S., Gentry, C., Halevi, S., Raykova, M., Sahai, A., Waters, B.: Candidate indistinguishability obfuscation and functional encryption for all circuits. In:
FOCS, pp. 40–49 (2013)
3. Sahai, A., Waters, B.: How to use indistinguishability obfuscation: deniable encryption, and more. In: STOC, pp. 475–484 (2014)
4. Ramchen, K., Waters, B.: Fully secure and fast signing from obfuscation. In: ACM
CCS, pp. 659–673 (2014)
5. Boneh, D., Zhandry, M.: Multiparty key exchange, efficient traitor tracing, and
more from indistinguishability obfuscation. In: Garay, J.A., Gennaro, R. (eds.)
CRYPTO 2014, Part I. LNCS, vol. 8616, pp. 480–499. Springer, Heidelberg (2014)
6. Shacham, H., Waters, B.: Compact proofs of retrievability. In: Pieprzyk, J. (ed.)
ASIACRYPT 2008. LNCS, vol. 5350, pp. 90–107. Springer, Heidelberg (2008)
7. Giuseppe, A., Randal, B., Reza, C., Herring, J., Kissner, L., Peterson, Z., Song, D.:
Provable data possession at untrusted stores. In: ACM CCS, pp. 598–609 (2007)
8. Benabbas, S., Gennaro, R., Vahlis, Y.: Verifiable delegation of computation over
large datasets. In: Rogaway, P. (ed.) CRYPTO 2011. LNCS, vol. 6841, pp. 111–131.
Springer, Heidelberg (2011)
9. K¨up¸c¨u, A.: Efficient cryptography for the next generation secure cloud: protocols,
proofs, and implementation. Lambert Academic Publishing, Saarbr¨ucken (2010)
10. Ateniese, G., Kamara, S., Katz, J.: Proofs of storage from homomorphic identification protocols. In: Matsui, M. (ed.) ASIACRYPT 2009. LNCS, vol. 5912, pp.
319–333. Springer, Heidelberg (2009)
11. Bowers, K.D., Juels, A., Oprea, A.: Proofs of retrievability: theory and implementation. In: The ACM Workshop on Cloud Computing Security, pp. 43–54 (2009)
12. Dodis, Y., Vadhan, S., Wichs, D.: Proofs of retrievability via hardness amplification. In: Reingold, O. (ed.) TCC 2009. LNCS, vol. 5444, pp. 109–127. Springer,
Heidelberg (2009)
13. Ateniese, G., Pietro, R.D., Mancini, L.V., Tsudik, G.: Scalable and efficient provable data possession. In: SecureComm 2008, pp. 9:1–9:10. ACM, New York (2008)
-----
14. Dynamic provable data possession. In: ACM CCS, pp. 213–222 (2009)
15. Wang, Q., Wang, C., Li, J., Ren, K., Lou, W.: Enabling public verifiability and
data dynamics for storage security in cloud computing. In: Backes, M., Ning, P.
(eds.) ESORICS 2009. LNCS, vol. 5789, pp. 355–370. Springer, Heidelberg (2009)
16. Stefanov, E., van Dijk, M., Juels, A., Oprea, A.: Iris: a scalable cloud file system
with efficient integrity checks. In: ACSAC, pp. 229–238 (2012)
17. Cash, D., K¨up¸c¨u, A., Wichs, D.: Dynamic proofs of retrievability via oblivious
RAM. In: Johansson, T., Nguyen, P.Q. (eds.) EUROCRYPT 2013. LNCS, vol.
7881, pp. 279–295. Springer, Heidelberg (2013)
18. Shi, E., Stefanov, E., Papamanthou, C.: Practical dynamic proofs of retrievability.
In: ACM CCS, pp. 325–336 (2013)
19. Armknecht, F., Bohli, J.M., Karame, G.O., Liu, Z., Reuter, C.A.: Outsourced
proofs of retrievability. In: ACM CCS, pp. 831–843 (2014)
20. Barak, B., Goldreich, O., Impagliazzo, R., Rudich, S., Sahai, A., Vadhan, S.: On
the (Im)possibility of obfuscating programs. In: Kilian, J. (ed.) CRYPTO 2001.
LNCS, vol. 2139, pp. 1–18. Springer, Heidelberg (2001)
21. Coron, J.-S., Lepoint, T., Tibouchi, M.: Practical multilinear maps over the integers. In: Canetti, R., Garay, J.A. (eds.) CRYPTO 2013, Part I. LNCS, vol. 8042,
pp. 476–493. Springer, Heidelberg (2013)
22. Garg, S., Gentry, C., Halevi, S.: Candidate multilinear maps from ideal lattices.
In: Johansson, T., Nguyen, P.Q. (eds.) EUROCRYPT 2013. LNCS, vol. 7881, pp.
1–17. Springer, Heidelberg (2013)
23. Guan, C., Ren, K., Zhang, F., Kerschbaum, F., Yu, J.: A symmetric-key based
[proofs of retrievability supporting public verification. full version. http://ubisec.](http://ubisec.cse.buffalo.edu/files/PoR_from_iO.pdf)
[cse.buffalo.edu/files/PoR from iO.pdf](http://ubisec.cse.buffalo.edu/files/PoR_from_iO.pdf)
24. Barak, B., Bitansky, N., Canetti, R., Kalai, Y.T., Paneth, O., Sahai, A.: Obfuscation for evasive functions. In: Lindell, Y. (ed.) TCC 2014. LNCS, vol. 8349, pp.
26–51. Springer, Heidelberg (2014)
25. Brakerski, Z., Rothblum, G.N.: Virtual black-box obfuscation for all circuits via
generic graded encoding. In: Lindell, Y. (ed.) TCC 2014. LNCS, vol. 8349, pp.
1–25. Springer, Heidelberg (2014)
26. Garg, S., Gentry, C., Halevi, S., Raykova, M.: Two-round secure MPC from indistinguishability obfuscation. In: Lindell, Y. (ed.) TCC 2014. LNCS, vol. 8349, pp.
74–94. Springer, Heidelberg (2014)
27. Goldwasser, S., Gordon, S.D., Goyal, V., Jain, A., Katz, J., Liu, F.-H., Sahai, A.,
Shi, E., Zhou, H.-S.: Multi-input functional encryption. In: Nguyen, P.Q., Oswald,
E. (eds.) EUROCRYPT 2014. LNCS, vol. 8441, pp. 578–602. Springer, Heidelberg
(2014)
28. Hohenberger, S., Sahai, A., Waters, B.: Replacing a random oracle: full domain
hash from indistinguishability obfuscation. In: Nguyen, P.Q., Oswald, E. (eds.)
EUROCRYPT 2014. LNCS, vol. 8441, pp. 201–220. Springer, Heidelberg (2014)
29. Boneh, D., Waters, B.: Constrained pseudorandom functions and their applications. In: Sako, K., Sarkar, P. (eds.) ASIACRYPT 2013, Part II. LNCS, vol. 8270,
pp. 280–300. Springer, Heidelberg (2013)
30. Apon, D., Huang, Y., Katz, J., Malozemoff, A.J.: Implementing cryptographic
program obfuscation. IACR Cryptol. ePrint Arch. 2014, 779 (2014)
31. Ananth, P., Gupta, D., Ishai, Y., Sahai, A.: Optimizing obfuscation: avoiding barrington’s theorem. In: ACM CCS, pp. 646–658 (2014)
32. Hohenberger, S., Koppula, V., Waters, B.: Universal signature aggregators. IACR
Cryptol. ePrint Arch. 2014, 745 (2014)
-----
33. Wee, H.: On obfuscating point functions. In: STOC, pp. 523–532 (2005)
34. Gennaro, R., Gentry, C., Parno, B.: Non-interactive verifiable computing: outsourcing computation to untrusted workers. In: Rabin, T. (ed.) CRYPTO 2010. LNCS,
vol. 6223, pp. 465–482. Springer, Heidelberg (2010)
35. Parno, B., Raykova, M., Vaikuntanathan, V.: How to delegate and verify in public:
verifiable computation from attribute-based encryption. In: Cramer, R. (ed.) TCC
2012. LNCS, vol. 7194, pp. 422–439. Springer, Heidelberg (2012)
36. Kerschbaum, F.: Outsourced private set intersection using homomorphic encryption. In: ASIACCS, pp. 85–86 (2012)
-----
|
{
"disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references', 'abstract'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1007/978-3-319-24174-6_11?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1007/978-3-319-24174-6_11, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "CLOSED",
"url": ""
}
| 2,015
|
[
"JournalArticle",
"Conference"
] | false
| 2015-09-21T00:00:00
|
[] | 17,829
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Physics",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Physics",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/0217a17ab73d370bbdbf3ecb95639bd0810b5690
|
[
"Computer Science",
"Physics"
] | 0.842535
|
Generation and Distribution of Quantum Oblivious Keys for Secure Multiparty Computation
|
0217a17ab73d370bbdbf3ecb95639bd0810b5690
|
Applied Sciences
|
[
{
"authorId": "145451615",
"name": "M. Lemus"
},
{
"authorId": "122259487",
"name": "Mariana F. Ramos"
},
{
"authorId": "2053153085",
"name": "P. Yadav"
},
{
"authorId": "144691583",
"name": "N. Silva"
},
{
"authorId": "2113134",
"name": "N. Muga"
},
{
"authorId": "151498130",
"name": "André Souto"
},
{
"authorId": "47144625",
"name": "N. Paunkovic"
},
{
"authorId": "144372606",
"name": "P. Mateus"
},
{
"authorId": "143888528",
"name": "A. Pinto"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"Appl Sci"
],
"alternate_urls": [
"http://www.mathem.pub.ro/apps/",
"https://www.mdpi.com/journal/applsci",
"http://nbn-resolving.de/urn/resolver.pl?urn=urn:nbn:ch:bel-217814"
],
"id": "136edf8d-0f88-4c2c-830f-461c6a9b842e",
"issn": "2076-3417",
"name": "Applied Sciences",
"type": "journal",
"url": "http://www.e-helvetica.nb.admin.ch/directAccess?callnumber=bel-217814"
}
|
The oblivious transfer primitive is sufficient to implement secure multiparty computation. However, secure multiparty computation based on public-key cryptography is limited by the security and efficiency of the oblivious transfer implementation. We present a method to generate and distribute oblivious keys by exchanging qubits and by performing commitments using classical hash functions. With the presented hybrid approach of quantum and classical, we obtain a practical and high-speed oblivious transfer protocol. We analyse the security and efficiency features of the technique and conclude that it presents advantages in both areas when compared to public-key based techniques.
|
# applied sciences
_Article_
## Generation and Distribution of Quantum Oblivious Keys for Secure Multiparty Computation
**Mariano Lemus** **[1,2], Mariana F. Ramos** **[3,4], Preeti Yadav** **[1,2], Nuno A. Silva** **[3,4]** **,**
**Nelson J. Muga** **[3,4]** **, André Souto** **[2,5,6,]*** **, Nikola Paunkovi´c** **[1,2]** **, Paulo Mateus** **[1,2]**
**and Armando N. Pinto** **[3,4]**
1 Departamento de Matemática, Instituto Superior Técnico, Av. Rovisco Pais 1, 1049-001 Lisboa, Portugal;
mariano.lemus@tecnico.ulisboa.pt (M.L.); pri8.phy@gmail.com (P.Y.);
npaunkov@math.tecnico.ulisboa.pt (N.P.); pmat@math.ist.utl.pt (P.M.)
2 Instituto de Telecomunicações, Av. Rovisco Pais 1, 1049-001 Lisboa, Portugal
3 Departamento de Eletrónica, Telecomunicações e Informática, Universidade de Aveiro, Campus
Universitário de Santiago, 3810-193 Aveiro, Portugal; marianaferreiraramos@live.ua.pt (M.F.R.);
nasilva@ua.pt (N.A.S.); muga@ua.pt (N.J.M.); anp@ua.pt (A.N.P.)
4 Instituto de Telecomunicações, Campus Universitário de Santiago, 3810-193 Aveiro, Portugal
5 Departamento de Informática, Faculdade de Ciências da Universidade de Lisboa, Campo Grande 016,
1749-016 Lisboa, Portugal
6 LASIGE, Faculdade de Ciências da Universidade de Lisboa, Campo Grande 016, 1749-016 Lisboa, Portugal
***** Correspondence: ansouto@fc.ul.pt
Received: 15 May 2020; Accepted: 10 June 2020; Published: 12 June 2020
**Featured Application: Private data mining.**
[����������](https://www.mdpi.com/2076-3417/10/12/4080?type=check_update&version=3)
**�������**
**Abstract: The oblivious transfer primitive is sufficient to implement secure multiparty computation.**
However, secure multiparty computation based on public-key cryptography is limited by the security
and efficiency of the oblivious transfer implementation. We present a method to generate and
distribute oblivious keys by exchanging qubits and by performing commitments using classical
hash functions. With the presented hybrid approach of quantum and classical, we obtain a practical
and high-speed oblivious transfer protocol. We analyse the security and efficiency features of the
technique and conclude that it presents advantages in both areas when compared to public-key
based techniques.
**Keywords: secure multiparty computation; oblivious transfer; quantum communications**
**1. Introduction**
In Secure Multiparty Computation (SMC), several agents compute a function that depends on
their own inputs, while maintaining them private [1]. Privacy is critical in the context of an information
society, where data is collected from multiple devices (smartphones, home appliances, computers,
street cameras, sensors, ...) and subjected to intensive analysis through data mining. This data
collection and exploration paradigm offers great opportunities, but it also raises serious concerns.
A technology able to protect the privacy of citizens, while simultaneously allowing to profit from
extensive data mining, is going to be of utmost importance. SMC has the potential to be that technology
if it can be made practical, secure and ubiquitous.
Current SMC protocols rely on the use of asymmetric cryptography algorithms [2], which are
considered significantly more computationally complex compared with symmetric cryptography
algorithms [3]. Besides being more computationally intensive, in its current standards, asymmetric
cryptography cannot be considered secure anymore due to the expected increase of computational
-----
_Appl. Sci. 2020, 10, 4080_ 2 of 11
power that a large-scale quantum computer will bring [4]. Identifying these shortcomings in efficiency
and security motivates the search for alternative techniques for implementing SMC without the need
of public key cryptography.
_1.1. Secure Multiparty Computation and Oblivious Transfer_
Consider a set of N agents and f (x1, x2, . . ., xN) = (y1, y2, . . ., yN) a multivariate function. For i ∈
_{1, . . ., N}, a SMC service (see Figure 1) receives the input xi from the i-th agent and outputs back the_
value yi in such a way that no additional information is revealed about the remaining xj, yj, for j ̸= i.
Additionally, this definition can be strengthened by requiring that for some number M < N of corrupt
agents working together, no information about the remaining agents gets revealed (secrecy). It can
also be imposed that if at most M[′] _< N agents do not compute the function correctly, the protocol_
identifies it and aborts (authenticity).
**Figure 1. In secure multiparty computation, N parties compute a function preserving the privacy of**
their own input. Each party only has access to their own input–output pair.
Some of the most promising approaches towards implementing SMC are based on oblivious
circuit evaluation techniques such as Yao’s garbled circuits for the two party case [5] and the GMW
or BMR protocols for the general case [2,6–8]. It has been shown that to achieve SMC it is enough to
implement the Oblivious Transfer (OT) primitive and, without additional assumptions, the security of
the resulting SMC depends only on that of the OT [9]. In the worst case, this requires each party to
perform one OT with every other party for each gate of the circuit being evaluated. This number can
be reduced by weakening the security or by increasing the amount of exchanged data [10]. Either way,
the OT cost of SMC represents a major bottleneck for its practical implementation. Finding fast and
secure OT protocols, hence, is a very relevant task in the context of implementing SMC.
Let Alice and Bob be two agents. A 1-out-of-2 OT service receives bits b0, b1 as input from Alice and
a bit c as input from Bob, then outputs bc to Bob. This is done in a way that Bob gets no information about
the other message, i.e., bc, and Alice gets no information about Bob’s choice, i.e., the value of c [11].
_1.2. State of the Art_
Classical OT implementations are based on the use of asymmetric keys, and suffer from two types
of problems. The first one is the efficiency: asymmetric cryptography relies on relatively complex key
generation, encryption, and decryption algorithms [12] (Chapter 1) and [13] (Chapter 6). This limits
achievable rates of OTs, and since implementations of SMC require a very large number of OTs [3,10],
this has hindered the development of SMC-based applications. The other serious drawback is that
-----
_Appl. Sci. 2020, 10, 4080_ 3 of 11
asymmetric cryptography, based on integer number factorization or discrete-logarithm problems, is
insecure in the presence of quantum computers, and therefore, it has to be progressively abandoned.
There are strong research efforts in order to find other hard problems that can support asymmetric
cryptography [4]. However, the security of these novel solutions is still not fully understood.
A possible way to circumvent this problem is by using quantum cryptography to improve
the efficiency and security of current techniques. Quantum solutions for secure key distribution,
Bit Commitment (BC) and OT have been already proposed [14]. The former was proved to be
unconditionally secure (assuming an authenticated channel) and realizable using current technology.
Although, it was shown to be impossible to achieve unconditionally secure quantum BC and OT [15–17],
one can impose restrictions on the power of adversaries in order to obtain practically secure versions
of these protocols [18,19]. These assumptions include physical limitations on the apparatuses, such as
noisy or bounded quantum memories [20–22]. For instance, quantum OT and BC protocols have been
developed and implemented (see [23–25]) under the noisy storage model. Nevertheless, solutions
based on hardware limitations may not last for long, because as quantum technology improves the rate
of secure OT instances will decrease. Other solutions include exploring relativistic scenarios using the
fact that no information can travel faster than light [26–28]. However, at the moment, these solutions
do not seem to be practical enough to allow the large dissemination of SMC.
In this work, we explore the resulting security and efficiency features of implementing oblivious
transfer using a well known quantum protocol [5] supported by using a cryptographic hash based
commitment scheme [29]. We call it a hybrid approach, since it mixes both classical and quantum
cryptography. We analyse the protocol stand alone security, as well as its composable security in the
random oracle model. Additionally, we study its computational complexity and compare it with the
complexity of alternative public key based protocols. Furthermore, we show that, while unconditional
information-theoretic security cannot be achieved, there is an advantage (both in terms of security
and efficiency) of using quantum resources in computationally secure protocols, and as such, they are
worth consideration for practical tasks in the near future.
This paper is organized as follows. In Section 2, we present a quantum protocol to produce OT
given access to a collision resistant hash function, define the concept of oblivious keys, and explain
how having pre-shared oblivious keys can significantly decrease the computational cost of OT during
SMC. The security and efficiency of the protocol is discussed in Section 3. Finally, in Section 4 we
summarize the main conclusions of this work.
**2. Methods**
_2.1. Generating the OTs_
In this section, we describe how to perform oblivious transfer by exchanging qubits. The protocol
_πQOT shown in Figure 2 is the well known quantum oblivious transfer protocol first introduced by_
Yao, which assumes access to secure commitments. The two logical qubit states 0 and 1 represent
_|_ _⟩_ _|_ _⟩_
_√_ _√_
the computational basis, and the states + = ( 0 + 1 )/ 2, = ( 0 1 )/ 2 represent the
_|_ _⟩_ _|_ _⟩_ _|_ _⟩_ _|−⟩_ _|_ _⟩−|_ _⟩_
Hadamard basis. We also define the states |(si, ai)⟩ for si, ai ∈{0, 1} according to the following rule:
(0, 0) = 0 (0, 1) = +
_|_ _⟩_ _|_ _⟩_ _|_ _⟩_ _|_ _⟩_
(1, 0) = 1 (1, 1) = .
_|_ _⟩_ _|_ _⟩_ _|_ _⟩_ _|−⟩_
Note that these states can be physically instantiated using, for instance, a polarization encoding
fiber optic quantum communication system, provided that a fast polarization encoding/decoding
process and an algorithm to control random polarization drifts in optical fibers are available [30,31].
-----
_Appl. Sci. 2020, 10, 4080_ 4 of 11
**Protocol πQOT**
**Parameters: Integers n, m < n.**
**Parties: The sender Alice and the receiver Bob.**
**Inputs: Alice gets two bits b0, b1 and Bob gets a bit c.**
_(Oblivious key distribution phase)_
1. Alice samples s, a ∈{0, 1}[n][+][m]. For each i ≤ _n + m she prepares the state |φi⟩_ = |(si, ai)⟩ and sends
_|φ⟩_ = |φ1φ2 . . . φn+m⟩ to Bob.
2. Bob samples ˜a ∈{0, 1}[n][+][m] and, for each i, measures |φi⟩ in the computational basis if ˜ai = 0, otherwise
measures it in the Hadamard basis. Then, he computes the string ˜s = ˜s1s˜2 . . . ˜sn+m, where ˜si = 0 if the
outcome of measuring |φi⟩ was 0 or +, and ˜si = 1 if it was 1 or −.
3. For each i, Bob commits (s˜i, ˜ai) to Alice.
4. Alice chooses randomly a set of indices T ⊂{1, . . ., n + m} of size m and sends T to Bob.
5. For each j ∈ _T, Bob opens the commitments associated to (s˜j, ˜aj)._
6. Alice checks if sj = ˜sj whenever aj = ˜aj for all j ∈ _T. If the test fails Alice aborts the protocol, otherwise she_
sends a[∗] = a|T to Bob and sets k = s|T.
7. Bob computes x = a[∗] _⊕_ _a˜|T and k[˜] = ˜s|T._
_(Oblivious transfer phase)_
8. Bob defines the two sets I0 = {i | xi = 0} and I1 = {i | xi = 1}. Then, he sends to Alice the ordered pair
(Ic, Ic⊕1).
9. Alice computes (e0, e1), where ei = bi �j∈Ic⊕i _[k]j[, and sends it to Bob.]_
10. Bob outputs b[˜]c = ec �j∈I0 _[k][˜]_ _j[.]_
**Figure 2. Quantum OT protocol based on secure commitments. The** [�] denotes the bit XOR of all the
elements in the family.
Intuitively, this protocol works because the computational and the Hadamard are conjugate bases.
Performing a measurement in the preparation basis of a state, given by ai, yields a deterministic
outcome, whereas measuring in the conjugate basis, given by ¯ai, results in a completely random
outcome. By preparing and measuring in random bases, as shown in steps 1 and 2, approximately
half of the measurement outcomes will be equal to the prepared states, and half of them will have
no correlation. As Alice sends the information of preparation bases to Bob in step 6, he gets to know
which of his bits are correlated with Alice’s. During steps 3 to 6, Bob commits the information of
his measurement basis and outcomes to Alice, who then chooses a random subset of them to test
for correlations. Passing this test (statistically) ensures that Bob measured his qubits as stated in the
protocol as opposed to performing a different (potentially joint) measurement. Such strategy may
extract additional information from Alice’s strings, but would fail to pass the specific correlation check
in step 6. At step 8, Bob separates his non-tested measurement outcomes in two groups: I0 where he
measured in the same basis as the preparation one, and I1, in which he measured in the different basis.
He then inputs his bit choice c by selecting the order in which he sends the two sets to Alice. During
step 9, Alice encrypts her first and second input bits with the preparation bits associated with the first
and second second sets sent by Bob respectively. This effectively hides Bob’s input bit because she is
ignorant about the measurements that were not opened by Bob (by the security of the commitment
scheme). Finally, Bob can decrypt only the bit encrypted with the preparation bits associated to I0.
In real implementations of the protocol one should consider imperfect sources, noisy channels,
and measurement errors. Thus, in step 6 Alice should perform parameter estimation for the statistics
of the measurements, and pass whenever the error parameter es below some previously fixed value.
Following this, Alice and Bob perform standard post-processing techniques of information reconciliation
and privacy amplification before continuing to step 7. These techniques indeed work even in the presence
of a dishonest Bob. As long as he has some minimal amount of uncertainty about Alice’s preparation
string s, an adequate privacy amplification scheme can be used to maximize Bob’s uncertainty of
one of Alice’s input bits. This comes at the cost of increasing the amount of qubits shared per OT [32].
-----
_Appl. Sci. 2020, 10, 4080_ 5 of 11
An example of these techniques applied in the context of the noisy storage model (where the commitment
based check is replaced by a time delay under noisy memories) can be found in [19].
_2.2. Oblivious Key Distribution_
In order to make the quantum implementation of OT more practical during SMC we introduce
the concept of oblivious keys. The protocol πQOT can be separated in two phases: the Oblivious Key
_Distribution phase which consists of steps 1 to 7 and forms the πOKD subprotocol, and the Oblivious_
Transfer phase which takes steps 8 to 10 and we denote as the πOK→OT subprotocol. Note that after
step 7 of πQOT the subsets I0, I1 have not been revealed to Alice, so she has no information yet on how
the correlated and uncorrelated bits between k and _k[˜] are distributed (recall that k and_ _k[˜] are the result_
of removing the tested bits from the strings s and ˜s respectively). On the other hand, after receiving
Alice’s preparation bases, Bob does know the distribution of correlated and uncorrelated bits between
_k and_ _k[˜], which is recorded in the string x (xi = 0 if ai = ˜ai, otherwise xi = 1). Note that until step 7 of_
the protocol all computation is independent of the input bits e0, e1, c. Furthermore, from step 8, only
the strings k, _k[˜], and x are needed to finish the protocol (in addition to the input bits). We call these three_
strings collectively an oblivious key, depicted in Figure 3. Formally, let Alice and Bob be two agents.
Oblivious Key Distribution (OKD) is a service that outputs to Alice the string k = k1k2 . . . kℓ and to
Bob the string _k[˜] =_ _k[˜]_ 1k[˜] 2 . . . k[˜] _ℓ_ together with the bit string x = x1x2 . . . xℓ, such that ki = _k[˜]_ _i whenever_
_xi = 0 and_ _k[˜]_ _i does not give any information about k whenever xi = 1. All of the strings are chosen at_
random for every invocation of the service. A pair (k, (k[˜], x)) distributed as above is what we call an
oblivious key pair. Alice, who knows k, is referred to as the sender, and Bob, who holds _k[˜] and x, is the_
receiver. In other words, when two parties share an oblivious key, the sender holds a string k, while
the receiver has only approximately half of the bits of k, but knows exactly which of those bits he has.
**Figure 3. Oblivious keys. Alice has the string k and Bob the string** _k[˜]. For each party, the boxes in the_
left and right represent the bits of their string associated to the indices i for which xi equals 0 (left box)
or 1 (right box). Alice knows the entire key, Bob only knows half of the key, but Alice does not know
which half Bob knows.
When two parties have previously shared an oblivious key pair, they can securely produce OT by
performing the steps πOK→OT of πQOT. This is significantly faster than current implementations of OT
without any previous shared resource and does not require quantum communication during SMC.
Note that the agents can perform, previously or concurrently, an OKD protocol to share a sufficiently
large oblivious key, which can be then partitioned and used to perform as many instances of OT as
needed for SMC.
Fortunately, it is possible to achieve fast oblivious key exchange if the parties have access to fast
and reliable quantum communications and classical commitments. In order to use this QOT protocol,
the commitment scheme must be instantiated. Consider the commitment protocol πCOMH shown in
Figure 4, first introduced by Halevi and Micali. It uses a combination of universal and cryptographic
hashing, the former to ensure statistical uniformity on the commitments, and the latter to hide the
-----
_Appl. Sci. 2020, 10, 4080_ 6 of 11
committed message. The motivation for the choice of this protocol for this task will become more
apparent during the following sections as we discuss the security and efficiency characteristics of
the composition of πQOT with πCOMH, henceforth referred as the πHOK (for Hybrid Oblivious Key)
protocol for OT.
The existence of a reduction from OT to commitments, while proven within quantum
cryptography through the πQOT protocol, is an open problem in classical cryptography. The existence
of commitment schemes such as πCOMH, which do not rely on asymmetric cryptography, provides
a way to obtain OT in the quantum setting while circumventing the disadvantages of asymmetric
cryptography.
**Protocol πCOMH**
**Parameters: Message length ˜n and security parameter k. A universal hash family F = { f : {0, 1}[ℓ]** _→{0, 1}[k]},_
with ℓ = 4k + 2˜n + 4. A collision resistant hash function H.
**Parties: The verifier Alice and the committer Bob.**
**Inputs: Bob gets a string ˜m of length ˜n.**
_(Commit phase)_
1. Bob samples r ∈{0, 1}[ℓ], computes y = H(r), and chooses f ∈ **F, such that f (r) = ˜m. Then, he sends ( f**, y)
to Alice.
_(Open phase)_
2. Bob sends r to Alice.
3. Alice checks that H(r) = y. If this test fails she aborts the protocol. Otherwise, she outputs f (r).
**Figure 4. Commitment protocol based on collision resistant hash functions.**
**3. Results and Discussion**
_3.1. Security_
In this section, we analyse the security of the proposed composition of protocols. The main result
is encapsulated in the following theorem.
**Theorem 1. The protocol πHOK is secure as long as the hash function is collision resistant. Moreover, if the**
_hash function models a Random Oracle, a simple modification of the protocol can make it universally-composable_
_secure._
**Proof. The security proof relies on several well-established results in cryptography. First, notice that**
the πHOK protocol is closely related to the standard Quantum OT protocol πQOT, which is proven
statistically secure in Yao’s original paper [33] and later universally composable in the quantum
composability framework [34]. The difference between the two is that πQOT uses ideal commitments,
as opposed to the hash-based commitments in πHOK. We start by showing that the protocol πHOK
is standalone secure. For this case, we only need to replace the ideal commitment of πQOT with
a standalone secure commitment protocol, such as the Halevi and Micali [29], which is depicted in
_πCOMH. Since the latter is secure whenever the hash function is collision resistant, we conclude that_
_πHOK is secure whenever the hash function is collision resistant._
Finally, we provide the simple modification of πHOK that makes it universally-composable secure
when the hash function models a Random Oracle. The modification is only required to improve upon
the commitment protocol, as Yao’s protocol with ideal commitments is universally-composable [34].
Indeed, we need to consider universally composable commitment scheme instead of πCOMH. This is
achieved by the HMQ construction [35] which, given a standalone secure commitment scheme and
a Random Oracle, outputs a universally-composable commitment scheme, which is perfectly hiding
and computationally binding, that is, secure as far as collisions cannot be found. So we just need to
replace πCOMH with the output of the HMQ construction, when πCOMH and H are given as inputs
and H models a Random Oracle.
-----
_Appl. Sci. 2020, 10, 4080_ 7 of 11
Regarding the above theorem we note that, for the composable security, the HMQ construction
mentioned in the proof formally requires access to a random oracle, which is an abstract object used
for studying security and cannot be realized in the real world. Hence, we leave it as an additional
security property, as hash functions are traditionally modelled as random oracles. Stand alone security
of the πHOK protocol does not require the hash function to be a random oracle.
The use of collision resistant hash functions is acceptable in the quantum setting, as it has
been shown that there exist functions for which a quantum computer does not have any significant
advantage in finding collisions when compared with a classical one [36]. One point to note about the
security of πOKD is that it is not susceptible to intercept now-decrypt later style of attacks. Bob can
attempt an attack in which he does not properly measure the qubits sent by Alice at step 2, and instead
waits until Alice reveals the test subset in step 4 to measure honestly only those qubits. For that he
must be able to control the openings of the commitment scheme such that Alice opens the values of
his measurement outcomes for those qubits. In order to do this, he must be able find collisions for H
before step 5. This means that attacking the protocol by finding collisions of the hash function is only
effective if it is done in real time, that is, between steps 3 and 5 of the protocol. This is in contrast to
asymmetric cryptography based OT, in which Bob can obtain both bits if he is able to overcome the
computational security at a later stage.
Finally, we point out that the OT extension algorithms that are used during SMC often rely only
on collision resistant hash functions [37] anyway. If those protocols are used to extend the base OTs
produced by πHOK, we can effectively speed up the OT rates without introducing any additional
computational complexity assumption.
_3.2. Efficiency_
Complexity-wise, the main problem with public-key based OT protocols is that they require
a public/private key generation, encryption, and decryption per transfer. In the case of RSA and
ElGamal based algorithms, this has complexity O(n[2.58]) (where N = 2[n] is the size of the group), using
Karatsuba multiplication and Berett reduction for Euclidian division [38]. Post-quantum protocols are
still ongoing optimization, but recent results show RLWE key genereration and encryption in time
_O(n[2]_ log(n)) [39].
To study the time complexity of the πHOK protocol, consider first the complexity of πCOMH.
It requires two calls of H and one call of the universal hash family F, ˜n bit comparisons (if using the
technique proposed in [29] to find the required f ), and one additional evaluation of f . Cryptographic
hash functions are designed so that their time complexity is linear on the size of the input, which in
this case is ℓ = 4k + 2n˜ + 4. To compute the universal hashing, the construction in [29] requires ˜nk
binary multiplications. Thus, the running time of πCOMH is linear on the security parameter k. On the
other hand, πQOT has two security parameters: n, associated to the size of the keys used to encrypt the
transferred bits, and m, associated to the security of the measurement test done by Alice. The protocol
requires n + m qubit preparations and measurements, n + m calls of the commitment scheme, and n
bit comparisons. This leads to an overall time complexity of O(k(n + m)) for the πHOK protocol, which
is linear in all of its security parameters.
In realistic scenarios, however, error correction and privacy amplification must be implemented
during the πOK→OT. For the former, LDPC codes [40] or the cascade algorithm [41] can be used, and the
latter can be done with universal hashing. For a given channel error parameter, these algorithms have
time complexity linear in the size of the input string, which in our case is n. Hence, πHOK stays efficient
when considering channel losses and preparation/measurement errors.
One of the major bottlenecks in the GMW protocol for SMC is the number of instances of OT
required (it is worth noting that GMW uses 1-out-of-4 OT, which can efficiently be obtained from two
instances of the 1-out-of-2 OT presented here [42]). A single Advanced Encryption Standard (AES)
circuit can be obtained with the order of 10[6] instances of OT. However, with current solutions, i.e.,
with computational implementations of OT based on asymmetric classical cryptography, one can
-----
_Appl. Sci. 2020, 10, 4080_ 8 of 11
generate 10[3] secure OTs per second in standard devices [43]. It is possible to use OT extension
_∼_
algorithms to increase its size up to rates of the order of 10[6] OT per second [3]. Several of such
techniques are based on symmetric cryptography primitives [43], such as hash functions, and could
also be used to extend the OTs generated by πHOK.
Due to the popularity of crypto-currencies, fast and efficient hashing machines have recently
become more accessible. Dedicated hashing devices are able to compute SHA-256 at rates of 10[12]
hashes per second (see Bitfury, Ebit, and WhatsMiner, for example). In addition, existent standard
Quantum Key Distribution (QKD) setups can be adapted to implement OKD, since both protocols
share the same requirements for the generation and measurement of photons. Notably, QKD setups
have already demonstrated secret key rates of the order of 10[6] bits per second [44–48]. It is also worth
mentioning that, as opposed to QKD, OKD is useful even in the case when Alice and Bob are at the
same location. This is because in standard key distribution the parties trust each other and, if at the
same location, they can just exchange hard drives with the shared key, whereas when sharing oblivious
keys, the parties do not trust each other and need a protocol that enforces security. Thus, for the cases
in which both parties being at the same location is not an inconvenience, the oblivious key rates can be
further raised, as the effects of channel noise are minimized.
Direct comparisons of OT generation speed between asymmetric cryptography techniques and
quantum techniques are difficult because the algorithms run on different hardware. Nevertheless, as
quantum technologies keep improving, the size and cost of devices capable of implementing quantum
protocols will decrease and their use can result in significant improvements of OT efficiency, in the
short-to-medium term future.
**4. Conclusions**
Motivated by the usefulness of SMC as a privacy-protecting data mining tool, and identifying
its OT cost as its main implementation challenge, we have proposed a potential solution for practical
implementation of OT as a subroutine SMC. The scheme consists on pre-sharing an oblivious key pair
and then using it to compute fast OT during the execution of the SMC protocol. We call this approach
hybrid because it uses resources traditionally associated with classical symmetric cryptography
(cryptographic hash functions), as well as quantum state communication and measurements on
conjugate observables, resources associated with quantum cryptography. The scheme is secure as far
as the chosen hash function is secure against quantum attacks. In addition, we showed that the overall
time complexity of πHOK is linear on all its security parameters, as opposed to the public-key based
alternatives, whose time complexities are at least quadratic on their respective parameters. Finally, by
comparing the state of current technology with the protocol requirements, we concluded that it has the
potential to surpass current asymmetric cryptography based techniques.
It was also noted that current experimental implementations of standard discrete-variable QKD
can be adapted to perform πHOK. The same post-processing techniques of error correction and privacy
amplification apply, however, fast hashing subroutines should be added for commitments during
the parameter estimation step. Future work includes designing an experimental setup, meeting the
implementation challenges, and experimentally testing the speed, correctness, and security of the
resulting oblivious key pairs. This includes computing oblivious key rate bounds for realistic scenarios
and comparing them with current alternative technologies. Real world key rate comparisons can help
us understand better the position of quantum technologies in the modern cryptographic landscape.
Regarding the use of quantum cryptography during the commitment phase; because of the
impossibility theorem for unconditionally secure commitments in the quantum setting [17], one must
always work with an additional assumption on top of needing quantum resources. The noisy
storage model provides an example in which the commitments are achieved by noisy quantum
memories [21,22,49]. The drawback of this particular assumption is the fact that advances in quantum
storage technology work against the performance of the protocol, which is not a desired feature.
The added cost of using quantum communication is a disadvantage. So far, to the knowledge of the
-----
_Appl. Sci. 2020, 10, 4080_ 9 of 11
authors, there are no additional practical quantum bit commitment protocols that provide advantages
in security or efficiency compared to classical ones once additional assumptions (such as random
oracles, common reference strings, computational hardness, etc.,) are introduced. Nevertheless, we are
optimistic that such protocols can be found in the future, perhaps by clever design, or by considering
a different a kind of assumption outside of the standard ones.
**Author Contributions: Conceptualization, P.M. and A.N.P.; methodology, M.F.R., N.A.S. and N.J.M.; validation,**
M.L., N.P., and A.S.; formal analysis, M.L., P.Y., N.P., A.S. and P.M.; investigation, M.L., M.F.R., N.A.S. and N.J.M.;
writing—original draft preparation, M.L., N.P., P.Y., M.F.R., N.A.S., N.J.M. and A.N.P.; writing—review and
editing, M.L. and P.M.; visualization, M.F.R., N.A.S., N.J.M. and A.N.P.; supervision, P.M. and A.N.P.; project
administration, P.M., A.S. and A.N.P.; funding acquisition, P.M., A.S. and A.N.P. All authors have read and agreed
to the published version of the manuscript.
**Funding:** This work is supported by the Fundação para a Ciência e a Tecnologia (FCT) through
national funds, by FEDER, COMPETE 2020, and by Regional Operational Program of Lisbon,
under UIDB/50008/2020, UIDP/50008/2020, UID/CEC/00408/2013, POCI-01-0145-FEDER-031826,
POCI-01-0247-FEDER-039728, PTDC/CCI-CIF/29877/2017, PD/BD/114334/2016, PD/BD/113648/2015, and
CEECIND/04594/2017/CP1393/CT0006A.
**Conflicts of Interest: The authors declare no conflict of interest. The funders had no role in the design of the**
study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, or in the decision to
publish the results.
**References**
1. Lindell, Y.; Pinkas, B. Secure Multiparty Computation for Privacy-Preserving Data Mining. J. Priv. Confid.
**[2009, 59–98. [CrossRef]](http://dx.doi.org/10.29012/jpc.v1i1.566)**
2. Laud, P.; Kamm, L. Applications of Secure Multiparty Computation; IOS Press: Amsterdam, The Netherlands,
2015; Volume 13.
3. Asharov, G.; Lindell, Y.; Schneider, T.; Zohner, M. More Efficient Oblivious Transfer Extensions. J. Cryptol.
**[2017, 30, 805–858. [CrossRef]](http://dx.doi.org/10.1007/s00145-016-9236-6)**
4. [Bernstein, D.J.; Lange, T. Post-quantum cryptography. Nature 2017, 549, 188. [CrossRef] [PubMed]](http://dx.doi.org/10.1038/nature23461)
5. Yao, A.C.C. How to generate and exchange secrets. In Proceedings of the 27th Annual Symposium on
Foundations of Computer Science (sfcs 1986), Toronto, ON, Canada, 27–29 October 1986; pp. 162–167.
[[CrossRef]](http://dx.doi.org/10.1109/SFCS.1986.25)
6. Goldreich, O.; Micali, S.; Wigderson, A. How to Play ANY Mental Game. In Proceedings of the Nineteenth
Annual ACM Symposium on Theory of Computing, New York, NY, USA, 25–27 May 1987; ACM: New York,
[NY, USA, 1987; pp. 218–229. [CrossRef]](http://dx.doi.org/10.1145/28395.28420)
7. Schneider, T.; Zohner, M., GMW vs. Yao? Efficient Secure Two-Party Computation with Low Depth Circuits.
In Financial Cryptography and Data Security, Proceedings of the 17th International Conference, FC 2013, Okinawa,
_Japan, 1–5 April 2013; Revised Selected Papers; Sadeghi, A.R., Ed.; Springer: Berlin/Heidelberg, Germany,_
2013; pp. 275–292.
8. Beaver, D.; Micali, S.; Rogaway, P. The round complexity of secure protocols. In Proceedings of the
Twenty-Second Annual ACM Symposium on Theory of Computing, Baltimore, MD, USA, 14–16 May 1990;
pp. 503–513.
9. Kilian, J. Founding Cryptography on Oblivious Transfer. In Proceedings of the Twentieth Annual ACM
Symposium on Theory of Computing, Chicago, IL, USA, 2–4 May 1988; ACM: New York, NY, USA, 1988;
pp. 20–31.
10. Harnik, D.; Ishai, Y.; Kushilevitz, E. How many oblivious transfers are needed for secure multiparty
computation? In Proceedings of the Annual International Cryptology Conference, Santa Barbara, CA, USA,
19–23 August 2007; pp. 284–302.
11. Rabin, M.O. How To Exchange Secrets; Technical Report TR-81; Aiken Computation Laboratory, Harvad
University, Cambridge, MA, USA, 1981.
12. Goldreich, O. Foundations of Cryptography, Volume I Basic Techniques; Cambridge University Press: Cambridge,
UK, 2001.
13. Paar, C.; Pelzl, J. Understanding Cryptography; Springer: Berlin/Heidelberg, Germany, 2010.
-----
_Appl. Sci. 2020, 10, 4080_ 10 of 11
14. Broadbent, A.; Schaffner, C. Quantum cryptography beyond quantum key distribution. Des. Codes Cryptogr.
**[2016, 78, 351–382. [CrossRef]](http://dx.doi.org/10.1007/s10623-015-0157-4)**
15. Shenoy-Hejamadi, A.; Pathak, A.; Radhakrishna, S. Quantum cryptography: Key distribution and beyond.
_[Quanta 2017, 6, 1–47. [CrossRef]](http://dx.doi.org/10.12743/quanta.v6i1.57)_
16. Lo, H.K.; Chau, H.F. Is Quantum Bit Commitment Really Possible? Phys. Rev. Lett. 1997, 78, 3410–3413.
[[CrossRef]](http://dx.doi.org/10.1103/PhysRevLett.78.3410)
17. Mayers, D. Unconditionally Secure Quantum Bit Commitment is Impossible. _Phys. Rev. Lett. 1997,_
_[78, 3414–3417. [CrossRef]](http://dx.doi.org/10.1103/PhysRevLett.78.3414)_
18. Wehner, S.; Schaffner, C.; Terhal, B.M. Cryptography from Noisy Storage. Phys. Rev. Lett. 2008, 100, 220502.
[[CrossRef]](http://dx.doi.org/10.1103/PhysRevLett.100.220502)
19. Wehner, S.; Curty, M.; Schaffner, C.; Lo, H.K. Implementation of two-party protocols in the noisy-storage
[model. Phys. Rev. A 2010, 81, 052336. [CrossRef]](http://dx.doi.org/10.1103/PhysRevA.81.052336)
20. Konig, R.; Wehner, S.; Wullschleger, J. Unconditional Security From Noisy Quantum Storage. IEEE Trans.
_[Inf. Theory 2012, 58, 1962–1984. [CrossRef]](http://dx.doi.org/10.1109/TIT.2011.2177772)_
21. Loura, R.; Almeida, Á.J.; André, P.; Pinto, A.; Mateus, P.; Paunkovi´c, N. Noise and measurement errors in
[a practical two-state quantum bit commitment protocol. Phys. Rev. A 2014, 89, 052336. [CrossRef]](http://dx.doi.org/10.1103/PhysRevA.89.052336)
22. Almeida, Á.J.; Stojanovic, A.D.; Paunkovi´c, N.; Loura, R.; Muga, N.J.; Silva, N.A.; Mateus, P.; André, P.S.;
Pinto, A.N. Implementation of a two-state quantum bit commitment protocol in optical fibers. J. Opt. 2015,
_[18, 015202. [CrossRef]](http://dx.doi.org/10.1088/2040-8978/18/1/015202)_
23. Erven, C.; Ng, N.; Gigov, N.; Laflamme, R.; Wehner, S.; Weihs, G. An experimental implementation of
[oblivious transfer in the noisy storage model. Nat. Commun. 2014, 5, 3418. [CrossRef] [PubMed]](http://dx.doi.org/10.1038/ncomms4418)
24. Furrer, F.; Gehring, T.; Schaffner, C.; Pacher, C.; Schnabel, R.; Wehner, S. Continuous-variable protocol for
[oblivious transfer in the noisy-storage model. Nat. Commun. 2018, 9, 1450. [CrossRef] [PubMed]](http://dx.doi.org/10.1038/s41467-018-03729-4)
25. Ng, N.H.Y.; Joshi, S.K.; Chen Ming, C.; Kurtsiefer, C.; Wehner, S. Experimental implementation of bit
[commitment in the noisy-storage model. Nat. Commun. 2012, 3, 1326. [CrossRef] [PubMed]](http://dx.doi.org/10.1038/ncomms2268)
26. Lunghi, T.; Kaniewski, J.; Bussières, F.; Houlmann, R.; Tomamichel, M.; Wehner, S.; Zbinden, H. Practical
[Relativistic Bit Commitment. Phys. Rev. Lett. 2015, 115, 030502. [CrossRef] [PubMed]](http://dx.doi.org/10.1103/PhysRevLett.115.030502)
27. Verbanis, E.; Martin, A.; Houlmann, R.; Boso, G.; Bussières, F.; Zbinden, H. 24-Hour Relativistic Bit
[Commitment. Phys. Rev. Lett. 2016, 117, 140506. [CrossRef]](http://dx.doi.org/10.1103/PhysRevLett.117.140506)
28. Pitalúa-García, D.; Kerenidis, I. Practical and unconditionally secure spacetime-constrained oblivious
[transfer. Phys. Rev. A 2018, 98, 032327. [CrossRef]](http://dx.doi.org/10.1103/PhysRevA.98.032327)
29. Halevi, S.; Micali, S. Practical and Provably-Secure Commitment Schemes from Collision-Free Hashing.
In Advances in Cryptology—CRYPTO ’96, Proceedings of the 16th Annual International Cryptology Conference Santa
_Barbara, California, CA, USA, 18–22 August 1996; Koblitz, N., Ed.; Springer: Berlin/Heidelberg, Germany,_
1996; pp. 201–215.
30. Pinto, A.N.; Ramos, M.F.; Silva, N.A.; Muga, N.J. Generation and Distribution of Oblivious Keys through
Quantum Communications. In Proceedings of the 2018 20th International Conference on Transparent Optical
[Networks (ICTON), Bucharest, Romania, 1–5 July 2018; pp. 1–3. [CrossRef]](http://dx.doi.org/10.1109/ICTON.2018.8473991)
31. Ramos, M.F.; Silva, N.A.; Muga, N.J.; Pinto, A.N. Reversal operator to compensate polarization random
[drifts in quantum communications. Opt. Express 2020, 28, 5035–5049. [CrossRef]](http://dx.doi.org/10.1364/OE.385196)
32. Lindell, Y.; Pinkas, B. An efficient protocol for secure two-party computation in the presence of malicious
adversaries. In Proceedings of the Annual International Conference on the Theory and Applications of
Cryptographic Techniques, Barcelona, Spain, 20–24 May 2007; pp. 52–78.
33. Yao, A.C.C. Security of Quantum Protocols Against Coherent Measurements. In Proceedings of the
Twenty-Seventh Annual ACM Symposium on Theory of Computing, Las Vegas, NV, USA, 29 May–1 June
1995; ACM: New York, NY, USA, 1995; pp. 67–75.
34. Unruh, D. Universally composable quantum multi-party computation. In Proceedings of the Annual
International Conference on the Theory and Applications of Cryptographic Techniques, French Riviera,
France, 30 May–3 June 2010; pp. 486–505.
35. Hofheinz, D.; Müller-Quade, J. Universally Composable Commitments Using Random Oracles. In Theory of
_Cryptography; Naor, M., Ed.; Springer: Berlin/Heidelberg, Germany, 2004; pp. 58–76._
36. Aaronson, S.; Shi, Y. Quantum Lower Bounds for the Collision and the Element Distinctness Problems.
_[J. ACM 2004, 51, 595–605. [CrossRef]](http://dx.doi.org/10.1145/1008731.1008735)_
-----
_Appl. Sci. 2020, 10, 4080_ 11 of 11
37. Asharov, G.; Lindell, Y.; Schneider, T.; Zohner, M. More Efficient Oblivious Transfer and Extensions for Faster
Secure Computation. In Proceedings of the 2013 ACM SIGSAC Conference on Computer & Communications
[Security, Berlin, Germany, 4–8 November 2013; ACM: New York, NY, USA, 2013; pp. 535–548. [CrossRef]](http://dx.doi.org/10.1145/2508859.2516738)
38. Menezes, A.J.; Katz, J.; Van Oorschot, P.C.; Vanstone, S.A. Handbook of Applied Cryptography; CRC Press, Inc.:
Boca Raton, FL, USA: 1996.
39. Ding, J.; Xie, X.; Lin, X. A Simple Provably Secure Key Exchange Scheme Based on the Learning with Errors
Problem. Iacr Cryptol. Eprint Arch. 2012, 2012, 688.
40. Martinez-Mateo, J.; Elkouss, D.; Martin, V. Key reconciliation for high performance quantum key distribution.
_[Sci. Rep. 2013, 3, 1576. [CrossRef] [PubMed]](http://dx.doi.org/10.1038/srep01576)_
41. Brassard, G.; Salvail, L. Secret-key reconciliation by public discussion. In Workshop on the Theory and
_Application of of Cryptographic Techniques; Springer: Berlin/Heidelberg, Germany, 1993; pp. 410–423._
42. [Naor, M.; Pinkas, B. Computationally secure oblivious transfer. J. Cryptol. 2005, 18, 1–35. [CrossRef]](http://dx.doi.org/10.1007/s00145-004-0102-6)
43. Chou, T.; Orlandi, C. The simplest protocol for oblivious transfer. In Proceedings of the International
Conference on Cryptology and Information Security in Latin America, Guadalajara, Mexico, 23–26 August
2015; pp. 40–58.
44. Comandar, L.; Fröhlich, B.; Lucamarini, M.; Patel, K.; Sharpe, A.; Dynes, J.; Yuan, Z.; Penty, R.; Shields, A.
Room temperature single-photon detectors for high bit rate quantum key distribution. Appl. Phys. Lett. 2014,
_[104, 021101. [CrossRef]](http://dx.doi.org/10.1063/1.4855515)_
45. Islam, N.T.; Lim, C.C.W.; Cahall, C.; Kim, J.; Gauthier, D.J. Provably secure and high-rate quantum key
[distribution with time-bin qudits. Sci. Adv. 2017, 3, e1701491. [CrossRef]](http://dx.doi.org/10.1126/sciadv.1701491)
46. Ko, H.; Choi, B.S.; Choe, J.S.; Kim, K.J.; Kim, J.H.; Youn, C.J. High-speed and high-performance
polarization-based quantum key distribution system without side channel effects caused by multiple lasers.
_[Photonics Res. 2018, 6, 214–219. [CrossRef]](http://dx.doi.org/10.1364/PRJ.6.000214)_
47. Wang, T.; Huang, P.; Zhou, Y.; Liu, W.; Ma, H.; Wang, S.; Zeng, G. High key rate continuous-variable
[quantum key distribution with a real local oscillator. Opt. Express 2018, 26, 2794–2806. [CrossRef]](http://dx.doi.org/10.1364/OE.26.002794)
48. Pirandola, S.; Andersen, U.; Banchi, L.; Berta, M.; Bunandar, D.; Colbeck, R.; Englund, D.; Gehring, T.;
Lupo, C.; Ottaviani, C.; et al. Advances in Quantum Cryptography. arXiv 2019, arXiv:1906.01645.
49. Loura, R.; Arsenovi´c, D.; Paunkovi´c, N.; Popovi´c, D.B.; Prvanovi´c, S. Security of two-state and four-state
[practical quantum bit-commitment protocols. Phys. Rev. A 2016, 94, 062335. [CrossRef]](http://dx.doi.org/10.1103/PhysRevA.94.062335)
_⃝c_ 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
[(CC BY) license (http://creativecommons.org/licenses/by/4.0/).](http://creativecommons.org/licenses/by/4.0/.)
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/1909.11701, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "GOLD",
"url": "https://www.mdpi.com/2076-3417/10/12/4080/pdf"
}
| 2,019
|
[
"JournalArticle"
] | true
| 2019-09-25T00:00:00
|
[
{
"paperId": "0e9c40922a9f9bad45bf71afb9d7dbc66ef48565",
"title": "Reversal operator to compensate polarization random drifts in quantum communications."
},
{
"paperId": "8ceda6f05d27ae88d8272f228bed78b4f0b3af13",
"title": "Advances in quantum cryptography"
},
{
"paperId": "83721103a6fd5535e943b1b575cf70862c2322a8",
"title": "Handbook of Applied Cryptography"
},
{
"paperId": "e3151b039935d2fba4933a1964514f2debb639f5",
"title": "Practical and unconditionally secure spacetime-constrained oblivious transfer"
},
{
"paperId": "db827e86f965bd12cb08e515c8e6940063102184",
"title": "Understanding Cryptography"
},
{
"paperId": "c98b82e5f2e1385e4f4ba9f9bdf23c09f60979fd",
"title": "Generation and Distribution of Oblivious Keys through Quantum Communications"
},
{
"paperId": "f2a2c3cee9d1e7ebc4835577938e0ba638eb7c04",
"title": "High key rate continuous-variable quantum key distribution with a real local oscillator."
},
{
"paperId": "e8815a601afc7d7a6246f45d98cc17214d79cc3b",
"title": "High-speed and high-performance polarization-based quantum key distribution system without side channel effects caused by multiple lasers"
},
{
"paperId": "7faf5e22615c238cae2cf77b125d0e6ca6a3945e",
"title": "Provably secure and high-rate quantum key distribution with time-bin qudits"
},
{
"paperId": "517943eea610e226c4aef8192cad506410ef1d9b",
"title": "Physical implementation of oblivious transfer using optical correlated randomness"
},
{
"paperId": "25b4a7b04cc7c1ff46f56d4f29a2dcbd7f444f8f",
"title": "Continuous-variable protocol for oblivious transfer in the noisy-storage model"
},
{
"paperId": "797f33b2208de43926407139d5bdc530f1aa3a09",
"title": "More Efficient Oblivious Transfer Extensions"
},
{
"paperId": "0feae0e89270536cdb45138ee0993aefe53e8905",
"title": "Quantum Cryptography: Key Distribution and Beyond"
},
{
"paperId": "cf0719b8d8c3c75c4212aa986600348f1fe0a3f2",
"title": "Securing the Internet of Things in a Quantum World"
},
{
"paperId": "168ec39c9eb2b4440bdf2736ee5101fafdc811c2",
"title": "Quantum computers ready to leap out of the lab in 2017"
},
{
"paperId": "5256995df636cecb7dee690e646ab2398cfbee82",
"title": "Security of two-state and four-state practical quantum bit-commitment protocols"
},
{
"paperId": "5210b5a0e98a7138710e2d7d88b641f6997a4dde",
"title": "24-Hour Relativistic Bit Commitment."
},
{
"paperId": "18f0f29bacb2441dd430271fc6ceaacca8ed68c4",
"title": "Computationally Binding Quantum Commitments"
},
{
"paperId": "98fad0706f87c3a07d69aab77f995b4dc1dc8bd0",
"title": "Quantum cryptography beyond quantum key distribution"
},
{
"paperId": "c52930652ef773a82572e8e7c88e4c6f384f16d8",
"title": "The Simplest Protocol for Oblivious Transfer"
},
{
"paperId": "737e72cfb698d1aa756b044bfa64c5001b90064b",
"title": "Practical Relativistic Bit Commitment."
},
{
"paperId": "ab34e3e90de34024a6727a0349d77a70118e8b2c",
"title": "Noise and measurement errors in a practical two-state quantum bit commitment protocol"
},
{
"paperId": "7470baf6080b7a835edf92f8467637ab3873c8a9",
"title": "Room temperature single-photon detectors for high bit rate quantum key distribution"
},
{
"paperId": "0166c8b5c6445043b94fc7b62d145d0c3c8b6483",
"title": "More efficient oblivious transfer and extensions for faster secure computation"
},
{
"paperId": "acb5ace0bbb9741bcc757cb09ec4faf6572823b8",
"title": "An experimental implementation of oblivious transfer in the noisy storage model"
},
{
"paperId": "9e003697fb9cc15aee9b97a8f42b88dbdf06eef3",
"title": "Key Reconciliation for High Performance Quantum Key Distribution"
},
{
"paperId": "9fa0ee74353fd008f2fbb1f6d724437678cbf9dd",
"title": "GMW vs. Yao? Efficient Secure Two-Party Computation with Low Depth Circuits"
},
{
"paperId": "206112b21f899101d090f78590f343d30b97fc33",
"title": "Experimental implementation of bit commitment in the noisy-storage model"
},
{
"paperId": "e4dc30ba02cd3c29735348ca02c5d49fc1df9eb7",
"title": "Implementation of two-party protocols in the noisy-storage model"
},
{
"paperId": "d531e2c1b879538d859a6d6a82d43e30e9d41c8e",
"title": "Universally Composable Quantum Multi-party Computation"
},
{
"paperId": "574d0df466d7e1244029365a44aa7b579cfd7fee",
"title": "Unconditional Security From Noisy Quantum Storage"
},
{
"paperId": "a6f644f6e739fa73ada11dc4c85b812b31f63d53",
"title": "Secure Multiparty Computation for Privacy-Preserving Data Mining"
},
{
"paperId": "99291ce0b97a31c786560241fea62604332afbf5",
"title": "Post-quantum cryptography"
},
{
"paperId": "253e994f744e9a0b9c4a305ebb246891ec0665e2",
"title": "Cryptography from noisy storage."
},
{
"paperId": "84ab7f80160291e5f1cda11d519a74864a8c6eda",
"title": "How Many Oblivious Transfers Are Needed for Secure Multiparty Computation?"
},
{
"paperId": "d4a8ecf10852322d5a162ba1a58687e9f5c16a19",
"title": "An Efficient Protocol for Secure Two-Party Computation in the Presence of Malicious Adversaries"
},
{
"paperId": "a93556bf7fee012764e36f70495cbbc43bdf4875",
"title": "Universally Composable Commitments Using Random Oracles"
},
{
"paperId": "4beae9ab6774cac4190ba8ceb998370317c65d09",
"title": "Quantum lower bounds for the collision and the element distinctness problems"
},
{
"paperId": "11032f14bf3fdb71476518922d3af4e6cd8b4af8",
"title": "Practical and Provably-Secure Commitment Schemes from Collision-Free Hashing"
},
{
"paperId": "9543136e66bf75eebc31adf98ad1a0ff99844fdf",
"title": "Unconditionally secure quantum bit commitment is impossible"
},
{
"paperId": "ba1c8fd36590fca91a67945b02d907da94098586",
"title": "Is Quantum Bit Commitment Really Possible?"
},
{
"paperId": "86bce9af17bae2cca7427b1f813cd046d3a7eb35",
"title": "Security of quantum protocols against coherent measurements"
},
{
"paperId": "cd81f9ed8a3adfecf6a69fdf5e1dbf92d7844dcc",
"title": "Secret-Key Reconciliation by Public Discussion"
},
{
"paperId": "e5302edfa2fa077525008333fcb56d9c2f3451ef",
"title": "The round complexity of secure protocols"
},
{
"paperId": "df2473061df11b76cebb7400c50246d0b354390c",
"title": "How to play ANY mental game"
},
{
"paperId": "29b0f06d18949fc7f3a38bb0022571aa15725dc7",
"title": "How to generate and exchange secrets"
},
{
"paperId": "aab2764de08142e8c256f6014d0cd1ff600517ad",
"title": "Founding Cryptography on Oblivious Transfer"
},
{
"paperId": "dc9701ffb5b2f1e345bf1d59065f9d2ee1afbedd",
"title": "Implementation of a two-state quantum bit commitment protocol in optical fibers"
},
{
"paperId": "d9082271933927a6b574b22e8e0bb8ad1b4903e2",
"title": "Applications of secure multiparty computation"
},
{
"paperId": "f104773bee7d9ac9e28f4591c8521008c48540e2",
"title": "A Simple Provably Secure Key Exchange Scheme Based on the Learning with Errors Problem"
},
{
"paperId": "de4b461d1f1cc7f7044c92b49c586a2463b28a8e",
"title": "Computationally Secure Oblivious Transfer"
},
{
"paperId": "e9fb87613db9138acc19682cbf109c3c37dad02b",
"title": "Founding crytpography on oblivious transfer"
}
] | 12,187
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Engineering",
"source": "s2-fos-model"
},
{
"category": "Physics",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/0219fc1452e1fe4236f49bfc7838f03a19ec7fba
|
[
"Computer Science"
] | 0.903706
|
5G-Compatible IF-Over-Fiber Transmission Using a Low-Cost SFP-Class Transceiver
|
0219fc1452e1fe4236f49bfc7838f03a19ec7fba
|
IEEE Access
|
[
{
"authorId": "104319582",
"name": "M. Fernandes"
},
{
"authorId": "1720762567",
"name": "Bruno T. Brandão"
},
{
"authorId": "1404994638",
"name": "A. Lorences-Riesgo"
},
{
"authorId": "2105730276",
"name": "P. Monteiro"
},
{
"authorId": "1394516775",
"name": "F. Guiomar"
}
] |
{
"alternate_issns": null,
"alternate_names": null,
"alternate_urls": [
"http://ieeexplore.ieee.org/servlet/opac?punumber=6287639"
],
"id": "2633f5b2-c15c-49fe-80f5-07523e770c26",
"issn": "2169-3536",
"name": "IEEE Access",
"type": "journal",
"url": "http://www.ieee.org/publications_standards/publications/ieee_access.html"
}
|
With the rise of 5G and beyond, the ever-increasing data-rates demanded by mobile access are severely challenging the capacity of optical fronthaul networks. Despite its high reliability and ease of deployment, legacy digital radio-over-fiber (RoF) technologies face an upcoming bandwidth bottleneck in the short term. This has motivated a renewed interest in the development of analog RoF alternatives, owing to their high spectral efficiency. However, unlike its digital counterpart, analog RoF transmission requires a highly linear transceiver to guarantee signal fidelity. Typical solutions exploited in recent research works tend to adopt the use of bulky benchtop components, such as directly modulated lasers (DML) and photodiodes. Although this provides a convenient and quick path for proof-of-concept demonstrations, there is still a considerable gap between lab developments and commercial deployment. Most importantly, a key question arises: can analog-RoF transceivers meet the 5G requirements while being competitive in terms of cost and footprint? Following this challenge, in this work we exploit the use of a low-cost commercial off-the-shelf (COTS) small form-factor pluggable (SFP) transceiver, originally designed for digital transmission at 1 Gbps, which is properly adapted towards analog RoF transmission. Bypassing the digital electronics circuitry of the SFP, while keeping the original transmitter optical sub-assembly (TOSA) and receiver optical sub-assembly (ROSA), we demonstrate that high-performance 5G-compatible transmission can be performed by reusing the key built-in components of current low-cost SFP-class transceivers. Particularly, we demonstrate error vector magnitude (EVM) performances compatible with 5G 64QAM transmission both at 100MHz and 400MHz. Furthermore, employing a memory polynomial model for digital pre-distortion of the transmitted signal, we achieve 256QAM-compatible performance at 100MHz bandwidth, after 20 km fronthaul transmission.
|
Received February 1, 2022, accepted February 23, 2022, date of publication February 25, 2022, date of current version March 9, 2022.
_Digital Object Identifier 10.1109/ACCESS.2022.3154784_
# 5G-Compatible IF-Over-Fiber Transmission Using a Low-Cost SFP-Class Transceiver
MARCO A. FERNANDES 1,2, (Member, IEEE), BRUNO T. BRANDÃO 1,2, (Member, IEEE),
ABEL LORENCES-RIESGO 1, PAULO P. MONTEIRO 1,2, (Senior Member, IEEE),
AND FERNANDO P. GUIOMAR 1,2, (Member, IEEE)
1Instituto de Telecomunicações, 3810-193 Aveiro, Portugal
2Department of Electronics, Telecommunications, and Informatics (DETI), University of Aveiro, 3810-193 Aveiro, Portugal
Corresponding author: Marco A. Fernandes (marcofernandes@av.it.pt)
This work was supported in part by the European Regional Development Fund (FEDER), through the Regional Operational Programme of
Centre (CENTRO 2020) of the Portugal 2020 framework; and in part by the Financial Support National Public [Fundação para a Ciência e
Tecnologia (FCT)] through Projects Optical Radio Convergence Infrastructure for Communications and Power Delivering (ORCIP)
(CENTRO-01-0145-FEDER-022141), Utilização de Tecnologias de Reflectometría no melhoramento do futuro Internet das Coisas e
Sistemas Ciber-Físicos (RETIOT) (Programa Operacional Competitividade e Internacionalização (POCI)-01-0145-FEDER-016432),
LANDmaRk (POCI-01-0145-FEDER-031527), and OptWire (PTDC/EEI-TEL/2697/2021). The work of Marco A. Fernandes and Bruno
T. Brandão was supported by the Ph.D. fellowships from FCT under Grant 2020.07521.BD and Grant 2021.05867.BD. The work of
Fernando P. Guiomar was supported by the ‘‘la Caixa’’ Foundation (ID 100010434), under Grant LCF/BQ/PR20/11770015.
**ABSTRACT With the rise of 5G and beyond, the ever-increasing data-rates demanded by mobile access**
are severely challenging the capacity of optical fronthaul networks. Despite its high reliability and ease
of deployment, legacy digital radio-over-fiber (RoF) technologies face an upcoming bandwidth bottleneck
in the short term. This has motivated a renewed interest in the development of analog RoF alternatives,
owing to their high spectral efficiency. However, unlike its digital counterpart, analog RoF transmission
requires a highly linear transceiver to guarantee signal fidelity. Typical solutions exploited in recent research
works tend to adopt the use of bulky benchtop components, such as directly modulated lasers (DML) and
photodiodes. Although this provides a convenient and quick path for proof-of-concept demonstrations, there
is still a considerable gap between lab developments and commercial deployment. Most importantly, a key
question arises: can analog-RoF transceivers meet the 5G requirements while being competitive in terms of
cost and footprint? Following this challenge, in this work we exploit the use of a low-cost commercial offthe-shelf (COTS) small form-factor pluggable (SFP) transceiver, originally designed for digital transmission
at 1 Gbps, which is properly adapted towards analog RoF transmission. Bypassing the digital electronics
circuitry of the SFP, while keeping the original transmitter optical sub-assembly (TOSA) and receiver
optical sub-assembly (ROSA), we demonstrate that high-performance 5G-compatible transmission can be
performed by reusing the key built-in components of current low-cost SFP-class transceivers. Particularly,
we demonstrate error vector magnitude (EVM) performances compatible with 5G 64QAM transmission both
at 100 MHz and 400 MHz. Furthermore, employing a memory polynomial model for digital pre-distortion
of the transmitted signal, we achieve 256QAM-compatible performance at 100 MHz bandwidth, after 20 km
fronthaul transmission.
**INDEX TERMS 5G, memory polynomial, analog radio-over-fiber.**
**I. INTRODUCTION**
The imminent rise of 5G and beyond radio communications,
together with the progressive adoption of the centralized
radio-access network (C-RAN) architecture [1], is bringing
new challenges for optical transceivers. Future transceivers
will have to cope with very tight requirements in terms of
The associate editor coordinating the review of this manuscript and
approving it for publication was Z. G. Zang .
bandwidth, latency and reliability [2]. Digital fronthauling
based on the common radio public interface (CPRI) [3] specification has been adopted as the de facto standard for 4G-LTE
signals, with its most recent version (eCPRI) dividing functionalities between the centralized unit and the distributed
unit, thereby achieving improved latency and capacity [17].
However, for next generation RANs, this fronthaul architecture must be able to provide data rates beyond of hundred
Gbps [4] due to the larger bandwidth of 5G signals and the use
-----
**FIGURE 1. Schematic of two RAN concepts: i) employing D-RoF transmission (above), and ii) another depicting an A-RoF architecture (below).**
of massive multiple input multiple output (MIMO) systems.
This together with the tight latency requirements defined
for 5G signals, has triggered research on whether analog
fronthaul can be used instead [5]. Whereas analog fronthaul
lacks the resilience of the digital counterpart, it does however
reduce the requirements on the transceiver bandwidth and
also minimizes the fronthaul latency [1].
The performance of analog radio-over-fiber (A-RoF)
transceivers is impaired by several effects, including nonlinearities in the transceiver [6] and fiber dispersion, whose
penalty can be enhanced by the laser chirp [7]. The presence of such undesired effects is enhanced when using lowcost transceivers, which are required for these applications.
To mitigate fiber dispersion, the use of an intermediate
frequency has been proposed for the mm-wave bands [8].
Mitigation of nonlinearities can be performed using
several digital pre-distortion (DPD) techniques, such as
memory polynomials [9], look-up-tables [10], or neural networks [11]. The complexity of these techniques should be
taken into account, and therefore lower complexity solutions such as look-up-tables or memory polynomial are
preferred.
Due to the widespread dissemination of digital communications, the main efforts from the industry rely on developing
low-cost packaged digital transceivers, while most analog
solutions remain based on costly discrete components, which
require further driving/adaptation so that they can be used
for research purposes. In this work, we address this scarcity
of analog RoF solutions, exploiting a workaround to obtain
low-cost analog transceivers through the adaptation of a commercial digital small form-factor pluggable (SFP) transceiver
to support the analog transmission of 5G signals. We address
key implementation issues such as the impact of crosstalk
between transmitter and receiver ports, and then we proceed
with experimentally testing the adapted low-cost transceiver
in an optical analog fronthaul. In order to maximize
the transceiver performance, we propose a low-complexity
DPD method based on a memory polynomial. We demonstrate that transmission over 20 km single-mode fiber (SMF)
does not impose any major impairment on the signal quality,
but nonlinear compensation is required to enable reception
of a 256QAM signal with EVM below the 3.5% established
limit [12].
In summary, the main novel contributions provided in this
work can be enumerated as follows,
i) proposal and demonstration of a simple adaption procedure over the electronic driving circuit of a low-cost
digital SFP transceiver, enabling its compatibility with
analog RoF transmission, while reusing the original
packaging and TOSA / ROSA components;
ii) detailed characterization of the adapted analog SFP
transceiver at the component level (S-parameters) and
system level (EVM performance in different 5G transmission scenarios);
iii) performance enhancement of the adapted analog SFP
transceiver using a low-complexity DPD model based
on a memory polynomial;
iv) experimental demonstration of the performance compatibility of the proposed analog SFP transceiver for the
transmission of 5G signals with 100 MHz and 400 MHz
with 64QAM modulated subcarriers, extended to
256QAM-compatible operation (100 MHz) enabled by
a memory polynomial DPD.
**II. ANALOG RADIO-OVER-FIBER FOR 5G ACCESS**
Figure 1 depicts the concept of a typical RAN employing
D-RoF transmission, and the alternative architecture with
analog transmission. The first difference observed consists
in the shift of equipment complexity from the distributed
unit (DU) to the centralized unit (CU). This is mainly due
-----
**FIGURE 2. Simple adaptation to convert a commercial digital SFP into an**
analog RoF transceiver. Red lines represent the shunt wires soldered into
the original PCB. (a) Modifications in the transmitter side and (b) in the
receiver side.
to the digital-to-analog converter (DAC) and analog-to-digital
converter (ADC) that were in the DU in the D-RoF scenario,
performing A/D conversion in the uplink and D/A conversion
in the downlink, which are relocated to the CU in the A-RoF
case, performing D/A conversion in the downlink and A/D
conversion in the uplink. This results in DUs that are simpler,
cheaper and with lower power consumption.
Other main drivers for the adoption of A-RoF systems
consist in their low-latency, ease of implementation and
high-spectral efficiency. Since in the A-RoF scenario, the
radio signals are directly transmitted, problems such as bandwidth multiplication are avoided. In typical D-RoF scenarios,
this problem cannot be avoided, resulting in enormous
data rates for 5G typical scenarios. In this scenario, considering the employement of CPRI (Option 8 / Split E),
the data rate can be calculated using the following
expression
:
data rate = M × Sr × N × 2 _Q[I]_ (1)
[×][ C][w][ ×][ C][,]
where M is the number of antennas per sector, Sr the sampling
rate, N is the number of bits per sample, 2(I _/Q) is a multipli-_
cation factor for in-phase (I) and quadrature-phase (Q) data,
_Cw is the factor of CPRI control word and C is a coding factor._
From the expression analysis it is easy to conclude that for
transmitting a given signal, the required fronthaul data rate
will be significantly expanded. For instance, considering the
case study of this paper, for the transmission of a 100 MHz
signal, if we consider M = 1, Sr = 1.5 × 100 MHz,
_N_ = 15, Cw = 16/15 and C = 66/64, the resulting
data rate for the D-RoF scenario will be roughly 2.5 Gb/s.
Considering the maximum bandwidth specified for 5G i.e.
16 aggregated components carriers (CCs) each with 400 MHz
(total aggregated bandwidth of 6.4 GHz), the required data
rate given by expression (1) with the aforementioned parameters is 158 Gbps, which would require the use of two highend 100GBASE-LR4 transceivers (or even ER4-class, if the
fronthaul is longer than 10 km), thus imposing a high cost
and power consumption at the E/O ends and also a rather
inefficient use of the optical spectrum (4 wavelengths per
transceiver). Instead, if A-RoF is considered, no bandwidth
expansion is imposed, and the 6.4 GHz radio signal can be
generated by an analog transceiver with a similar operating
bandwidth, thereby reducing the cost of the electronic components and also reducing the spectral occupancy of the optical signal. A similar calculation for more than 100 antenna
modules, a 4-sector device supporting 400 MHz baseband
channels would roughly require 10 Tbps, which is equivalent
to about 400 optical OOK-DSB-50GHz-grid channels [16].
These high-capacity examples clearly expose the critical
upscaling issues that are associated with digital fronthauling,
which is driving a renewed interest on the development of
A-RoF solutions for 5G and beyond.
Another tight requirement for 5G networks is ultra lowlatency communications, where this latency time can be
reduced to 1 ms in urgent and specific scenarios. This is
another advantage of A-RoF systems which have a significant
gain in latency-terms when compared to its digital counterpart. All these advantages are increasing the interest of the
scientific community in analog RANs, with multiple works
highlighting these advantages in field-trials applications and
real 5G networks [13]–[15]. Despite all the aforementioned
advantages of analog transmission it is difficult to encounter a
commercial off-the-shelf (COTS) analog optical transceiver.
In the next section, we propose a simple procedure to take
a COTS digital transceiver and convert it to perform analog
transmission.
**III. SFP ADAPTATION FOR ANALOG TRANSMISSION**
The transceivers under test are COTS SFP transceivers
designed for digital transmission at 1–10 Gb/s, which are
nowadays ubiquitous in fiber-optic access networks, and
whose cost is typically below 50 ¿. Let us start by unpacking the SFP transceiver and analyzing its key components.
In Figure 3.a, we can identify three key parts of the digital SFP transceiver: i) the transmitter optical sub-assembly
(TOSA), which performs electrical-to-optical (E/O) conversion using a directly-modulated laser, ii) the receiver optical
sub-assembly (ROSA), which performs optical-to-electrical
(O/E) conversion through an amplified photodiode, and iii)
the printed circuit board (PCB) that is responsible for electrically driving the TOSA and ROSA components. Since these
transceivers are designed for digital fiber communications,
a set of necessary adaptations are required to enable the
transmission of analog optical signals. Note, however, that
the key optical transmission components, i.e. the TOSA and
ROSA parts of the transceiver, are fundamentally responsible
for the E/O and O/E conversion regardless of the properties
of the transmitted/received signals, i.e. they are transparent to the type of transmission, and therefore can be kept
in their original form, thus benefiting from the low-cost
and small form-factor integration of these components. The
main adaptations to enable analog transmission with the
SFP transceiver are then required at the RF driving level in
the PCB.
-----
**FIGURE 3. Final prototype of the converted analog RoF transceiver.**
(a) Open package showing the simple shunt modifications applied to the
original SFP board and (b) analog RoF transceiver enclosed in the
original SFP package.
_A. ADAPTING THE ORIGINAL SFP ELECTRONICS FOR_
_ANALOG TRANSMISSION_
In order to preserve the original form-factor and pin-out of the
standard SFP transceiver, we reuse the original digital board
and perform a direct bypass of the digital electronics. Figure 2
shows a functional view of the modifications performed.
As evidenced in Figure 2, the modifications simply consist
in bypassing the digital part of the board (buffering, equalization, amplification, DC offset cancellation and amplitude
limitation) and directly wiring the input data pins to the output
ones. The final physical layout of the modified transceiver
can be observed in Figure 3, which evidences that these
small modifications enable to obtain a low-cost analog RoF
transceiver and keeping it in the original SFP package. Also
note that, in this work, we have used a SFP evaluation board
to provide access to the RF ports via SMA connectors.
**IV. CHARACTERIZATION OF THE ANALOG SFP**
Having successfully converted a digital transceiver into an
analog one, we proceeded with the characterization of these
transceivers. To this intent, firstly, S-parameter measurements
were performed to characterize the frequency response of the
analog transceivers. The E/O and O/E frequency response
of the transmitter and receiver are illustrated in Figure 4.
Note that the E/O frequency response of the TOSA is measured by using a calibrated photodiode at the receiver, while
the corresponding O/E frequency response of the ROSA is
obtained after de-embedding the frequency response of the
TOSA. From Figure 4a we see that the transmitter shows a
considerable conversion loss of 40 dB on average for frequencies up to 3 GHz and beyond 3 GHz the loss increases
gradually, exceeding 55 dB at 10 GHz. This shows the already
known poor conversion efficiency of direct modulated lasers.
For the receiver it can be seen that the O/E conversion gain
shows a decreasing trend from 45 dB near DC, down to 40 dB
at 7 GHz. After 7 GHz, the performance degrades severely.
When observing the combined response of the transmitter
and the receiver, (Figure 4b) we see that for frequencies
below 3 GHz there is a gain in the system, provided by the
receiver TIA. However, when increasing the frequency for
Final prototype of the converted analog RoF transceiver.
(a) Open package showing the simple shunt modifications applied to the
original SFP board and (b) analog RoF transceiver enclosed in the
original SFP package.
**FIGURE 4. (a) Transmitter and receiver E/O and O/E frequency responses**
respectively. (b) Combined |S21| of the transmitter and receiver.
**FIGURE 5. Electrical S11 and S22 responses of the transmitter and**
receiver respectively.
values above 7.5 GHz the system insertion loss becomes
too high to apply in a practical scenario. The results of
the impedance matching measurement at the transmitter and
receiver RF ports are shown in Figure 5. The input reflection
coefficient at the transmitter input RF port (|S11|) show a
bad impedance matching with the |S11| being mostly between
4 dB and 8 dB in the frequency range tested. In con− −
trast, the receiver results reveal two regions with good port
impedance matching, with |S22| equal or below −10 dB.
The first one is from DC up to 2.5 GHz and the second
one is from 5.5 GHz up to 6.5 GHz. The clear dip below
−30 dB in the |S11| of the transmitter can be attributed to a
resonance in the input circuit confined between 8 GHz and
9 GHz. Overall, this S-parameter characterization shows the
-----
**FIGURE 6. Diagram of the setup used to measure the transceiver**
crosstalk.
major impact of the transmitter conversion loss in the system.
It is worth noting that this simple digital-to-analog adaptation
procedure, basically consisting of four bypass wires, has been
performed with the main aim of providing a proof-of-concept
and easily reproducible demonstration. However, it should
be noted that improved impedance matching and conversion
efficiency might be achievable if the original SFP digital
electronics are entirely replaced by an analog driving board,
at the expense of a longer development time. Nevertheless,
the premise of this work is to assess the performance limits
of such a simple and fast digital-to-analog SFP conversion
procedure.
_A. CROSSTALK MEASUREMENTS_
Crosstalk between the transmitter and receiver can be a performance bottleneck in A-RoF systems. Addressing this subject, we advanced the characterization process by measuring
the crosstalk in our A-RoF transceiver. In order to quantify
the impact of this problem in further tests, we designed
a simple experiment consisting in measuring the received
RF power with the laser on and off. The setup used to measure
the crosstalk is depicted in Figure 6. If the transmitter and
receiver are disconnected, it is not expected to receive any
RF power, however, this power still has a significant value
due to the crosstalk.
We have defined crosstalk as the ratio between the power
received with the laser off and the one with laser on
:
_�P =_ _[P][Laser OFF]_ (2)
_PLaser ON_
Fixing the RF power to 0 dBm, we measured the value of
_�P sweeping the RF frequency between 1 GHz and 10 GHz._
These results are presented in Figure 7, which shows a clear
dependency between the crosstalk and the RF carrier frequency. For higher frequencies (above 8 GHz), the crosstalk
is so high, that the RF received power is the same, regardless
of whether the optical link is connected or disconnected.
Given the EVM-SNR relationship,
EVM (%) = 10[−] [SNR (dB)]20 × 100, (3)
if we assume all the crosstalk to be noise, we can measure the
impact that it will have on the EVM. Since we are aiming to
transmit 5G signals over an RF carrier of 3.5 GHz, the EVM
will always have a floor limit of 10%, which already does not
comply with the 3GPP requirements for modulations formats
of 256QAM and 64QAM [12]. These results render the joint
utilization of the transmitter and receiver in the transceiver
more challenging, degrading the overall transmission performance. To overcome this problem and to ensure that the
**FIGURE 7. Dependency of the crosstalk with the RF frequency with 0 dBm**
RF power at the transmitter.
transceiver crosstalk is avoided, in all of the remaining tests
we have decided to use TOSA and ROSA from different
SFPs packages. Although this choice implies some hardware
inefficiency (one TOSA/ROSA is unutilized per SFP pair),
it guarantees a crosstalk-free operation.
**V. EXPERIMENTAL SETUP**
In order to emulate a real analog optical fronthaul, we have
implemented the experimental setup shown in Figure 8.
The setup is composed of an Arbitrary Waveform Generator (AWG) responsible for the generation of the RF baseband
signal. This AWG has two differential output channels corresponding to the I and Q waveforms. This signal is then
up-converted by the IQ mixer to an RF carrier of 3.5 GHz,
corresponding to the standardized FR1 [2], before directly
modulating the analog optical transmitter. The optical signal
then travels through a given length of Single Mode Fiber
(SMF). Before the optical receiver, there is a Variable Optical
Attenuator (VOA), which enables the control of the optical
power at the photodiode. The optical receiver performs the
O/E conversion and this electrical signal is received in the
Vector Signal Analyzer (VSA). The VSA down-converts
the signal and demodulates it using standard compensation
algorithms.
In order to maximize the performance of the low-cost
transceivers, we propose the use of nonlinear DPD based on
the following memory polynomial model,
_Q−1_
�
_akq · y(n −_ _q) · |y(n −_ _q)|[k]_ _,_ (4)
_q=0_
_z(n)_
=
_K_ −1
�
_k=0_
where y is the transmitted signal without pre-distortion, z is
the pre-distorted signal, K is the nonlinear order and Q is the
memory depth. The optimization of the memory polynomial
coefficients, akq, follows the strategy described in Figure 8.
First, the signals are generated in the AWG without DPD. The
measured signals are then used to calculate the coefficients of
the memory polynomial by comparison with the transmitted
-----
**FIGURE 8. Experimental setup for extracting and applying a DPD model**
in a low-cost optical fronthaul.
waveform. After calculating the coefficients, the signal is
pre-distorted with the DPD model.
The experimental analysis is divided in the following
stages: i) back-to-back (B2B) performance assessment of
the TOSA/ROSA transceiver pair, ii) fronthaul performance
assessment, in which we consider a fiber link composed
of 20 km SMF, and iii) A-RoF performance enhancement.
**VI. EXPERIMENTAL 5G RESULTS**
_A. 5G FR1 TRANSMISSION_
1) OPTICAL B2B
The first test consisted in analyzing the performance of the
fronthaul in a simple scenario without optical fiber. The VOA
was used to set the optical power into the photodiode to
12 dBm. This test consisted in optimizing the RF power for
−
the 256QAM 5G signal. These optimizations were done for
various DPD models varying the values of K and Q described
in equation (4). The first implemented model consists only
of a linear compensation with one sample memory. After,
we tested a DPD model without memory (Q 1), while
=
increasing the nonlinear order of the model (K ) until obtaining the best performance. With the optimized value of K,
we increased the memory of the model until finding its best
value. The obtained results are plotted in Figure 9. As can
be seen, without DPD the best EVM was obtained for an RF
power of 0 dBm, and has a value of 4.3%. Note that this value
is above the 3.5% limit established by 3GPP for 256QAM
transmission. The linear compensation corresponding to the
blue dashed line in the figure presented similar results to
those obtained without DPD. With all the nonlinear models
implemented we obtained a considerable EVM reduction,
with the best DPD model (Q = 1, K = 5), yielding an
EVM of 3.2%. This model shows also to be better in terms of
RF power margin where the EVM is below 3.5%, providing
approximately a 2 dB tolerance for transmitted power detuning. Not shown here for brevity but increasing the nonlinear
order for values higher than 5 resulted in a progressive loss
of performance, likely triggered by a less accurate model
extraction of higher-order nonlinearities. It is interesting to
observe that the best absolute performance was obtained for
a higher power than the optimum without DPD (3 dBm),
**FIGURE 9. Measured EVM in B2B for the cases of no DPD, linear DPD and**
different nonlinear DPD based on memory polynomials, for a 100 MHz
signal with a 3.5 GHz carrier.
**FIGURE 10. Measured EVM after 20 km SMF for the cases of no DPD and**
different nonlinear DPD based on memory polynomials, for a 100 MHz
signal with a 3.5 GHz carrier.
which is a well-known advantage of nonlinear compensation
in general: by mitigating nonlinearities, higher powers can be
launched into the transmission system, thereby resulting in
an improved SNR and/or power-budget. It is worth noticing
that, all the considered DPD models in Figure 9 enable to
successfully transmit a 5G 256QAM signal with an EVM
below the 3GPP limit.
2) 20 km ANALYSIS
After studying the B2B performance, we added 20 km of
SMF in order to get a more realistic fronthaul scenario. With
this setup, there were two main goals. First, to verify if the
model obtained in B2B is accurate enough to achieve the
best transmission performance with a fiber fronthaul. This
would bring a great advantage in practical terms, enabling the
optimization of the DPD model for the optical transceivers in
a controlled laboratory environment without requiring individual optimization in different fronthaul networks. The other
main goal of this test, which is also inherently related with
the first, is to test whether a long fronthaul link composed of
20 km SMF would introduce memory effects in the system.
-----
With these goals in mind, we have continued the tests using
the best DPD model obtained in B2B (Q = 1, K = 5)
and with Q 2 and K 5. For the memoryless model
= =
(Q = 1, K = 5), we have measured the performance
applying the model obtained in B2B as well as extracting a
new model taking into account the 20 km in the system. It is
worth noticing that the insertion of the fiber link required to
increase the power in the photodiode to 10.5 dBm in order
−
to achieve the best performance. The obtained results are
shown in Figure 10. Without DPD the performance remains
similar to the B2B case, with minimum EVM of 4.3% when
the RF power is 0 dBm. Through the analysis of Figure 10,
it is possible to conclude that the model obtained in B2B is
still valid up to an RF power of 2 dBm, achieving a minimum
EVM of 3.3%. The benefits of extracting the model again
are visible for higher powers, enabling to achieve an EVM
of 3.2% for an RF power of 3 dBm. It is worth to notice that
the model obtained in B2B was still capable of presenting
a very good performance being only 0.1% worse than those
obtained with 20 km fiber link. The next step was to analyze
if the 20 km SMF introduced memory in the system, so we
extracted and applied a memory polynomial with Q 2 and
=
_K_ 5, which is signaled in the figure by a blue solid line. The
=
system remains memoryless since the performance degrades
relatively to the Q = 1, K = 5 scenario, yielding a minimum
EVM of 3.3% at the optimum RF power of 3 dBm. Once again
with all the DPD models, it was possible to obtain an EVM
below the 3GPP limit for 256QAM.
Considering equation (3) we can calculate the SNR gain
obtained with DPD. An EVM of 4.3% corresponds to an SNR
of 27.3 dB, whereas an EVM of 3.2% corresponds to an SNR
of 29.9 dB. Therefore, we may conclude that DPD effectively
provides an SNR improvement of approximately 3 dB, which
is a remarkable gain. These improvements are clear when
observing the spectrum of the signal with and without predistortion. In Figure 11 it is possible to observe the improvement on the signal spectrum at the power of 3 dBm caused by
the usage of DPD. Without DPD there is a clear nonlinear
phenomenon in the adjacent bands of the signal, which is
commonly designated as ‘‘spectral regrowth’’. It is possible to see that the applied DPD model compensate almost
completely for these bands. The signal spectra obtained at
the optimum power without DPD (0 dBm) is also shown in
Figure 11. From the results, we may conclude that DPD has
effectively compensated for the nonlinear distortions generated by the 3 dB increased power, thus resulting in an effective
gain of 3 dB in SNR, which very nearly matches the observed
gains in terms of EVM.
_B. 5G FR2 TRANSMISSION_
1) IF ANALYSIS
Since 3GPP has defined the mmWave range for FR2 transmission, to be able to transmit these signals with our adapted
transceivers, we need to convert them to an intermediate
frequency (IF), leading to a system configuration that is typi
**FIGURE 11. Measured spectra for the optical case without DPD (0 dBm),**
and for the optical case when performing DPD (3 dBm) before and after
applying the model.
**FIGURE 12. Measured EVM for different IF values with a 64QAM 400 MHz**
transmitted signal.
cally designated as IF-over-fiber (IFoF). To optimize the best
suited IF for our system, we started by sending a 400 MHz
64QAM signal using different IFs with an RF power of 0 dBm
and analysed the measured EVM. The results obtained from
this analysis are depicted in Figure 12. Despite not showing
a clear tendency, the results show an EVM minimum of
5.5% for an IF of 3.5 GHz. Besides maximizing the A-RoF
transceiver performance, this 3.5GHz IF choice also shows
the advantage of enabling an improved compatibility between
the FR1 and FR2 transmission modes.
2) PERFORMANCE ASSESSMENT
After optimizing the value of the transmitted IF we measured
the performance achievable in an FR2 scenario with our
setup. To this intent, we used the same 5G signal as before
(64QAM, 400 MHz) at the optimum IF of 3.5 GHz. With this
signal, we performed tests without fiber and with 20 km SMF.
Figure 13 shows the results obtained when sweeping the RF
power transmitted, without any DPD and with the best DPD
model found (Q = 2, K = 3), for each scenario. It is observable from the presented results that there is a slight increase
in the optimum RF power, which is related with having
-----
of real multiplications (NRMs) required to pre-distort each
sample of the transmitted signal. To proceed with this analysis, let us return to equation (4) and, since both akq and y(n)
are complex numbers, rewrite it as,
_z(n)_
=
_K_ −1 _Q−1_
� �
([akq,r _yr_ (n − _q) −_ _akq,iyi(n −_ _q)]_
_k=0_ _q=0_ � Real component (2 RMs)�� �
+j [akq,r _yi(n −_ _q) + akq,iyr_ (n − _q)])|y(n −_ _q)|[k]_ _,_
� �� �
Imaginary component (2 RMs)
**FIGURE 13. Measured EVM in an OB2B scenario and with 20 km SMF, for**
the cases of no DPD and best DPD model obtained (Q = 2, K = 3), for a
400 MHz signal with a 3.5 GHz carrier.
increased the signal bandwidth from 100 MHz to 400 MHz,
leading to a power spreading over the frequencies spectrum.
This bandwidth increase also leads to the enhancement of
filtering effects that introduce more memory into the system,
resulting in an optimum DPD model with a memory tap.
From the results, we observe that, for both scenarios, simply
optimizing the RF power driving the SFP is enough to achieve
64QAM transmission (where the 3GPP limit is 8% EVM).
Without any DPD we obtained an EVM of 5.2% and 6.3%,
in OB2B and with 20 km SMF, respectively. However, with
the best DPD models found, the performance was improved
to 4.8% in the OB2B scenario, and 5.5% with 20 km SMF.
_C. DPD COMPLEXITY ANALYSIS_
An underlying problem with introducing advanced techniques for nonlinear DPD is the increased complexity in these
systems. For this reason, we decided to perform a complexity
analysis of the proposed memory polynomial DPD method.
In order to quantify their complexity, we will use the number
- K = 1:
(5)
where we consider that the absolute value of y(n _q) can be_
−
computed as,[1]
�
|y(n − _q)| =_ _yr_ (n − _q)[2]_ + yi(n − _q)[2]_
� �� �
2 RMs
_._ (6)
Noting that the number of RMs grows linearly with the DPD
memory, Q, we can start by analyzing the model complexity
for the case of Q 1, with increasing polynomial order, K,
=
as shown in equations (8) to (11), as shown at the bottom of
the page.
Note that, when computing |y(n)|[k] for k > 1, we assume
that the value of _y(n)_ has already been previously com| |[k][−][1]
puted and stored in memory, and therefore there is only
one extra real multiplication needed to evaluate _y(n)_
| |[k] =
_y(n)_ _y(n)_ .
| |[k][−][1]| |
Finally, generalizing the above examples for any value of
_K and Q, we obtain that the following analytical expression_
1Note that, for simplicity, we neglect the complexity associated with
the [√](.) operation, as its hardware implementation might follow different
algorithms, namely resorting to the use of look-up tables. Nevertheless, it is
worth noting that for any memory polynomial of order K > 1 and memory
_Q, only Q square-root operations are actually required; i.e. once the value of_
|y(n − _q)| is first computed, it can be stored in memory for the subsequent_
evaluation of its |y(n − _q)|[k]_ products.
([a00,r _yr_ (n) − _a00,iyi(n)]_
� �� �
2 RMs
+j [a00,r _yi(n) + a00,iyr_ (n)]) |y(n)|[0]
� 2 RMs�� � ����=1
4 RMs (8)
−→
- K = 2:
- K = 3:
- K = 4:
[a10,r _yr_ (n) − _a10,iyi(n)]|y(n)|_ +j [a10,r _yi(n) + a10,iyr_ (n)]|y(n)|
� �� � � �� �
3 RMs 3 RMs
[a20,r _yr_ (n) − _a20,iyi(n)]|y(n)|[2]_ +j [a20,r _yi(n) + a20,iyr_ (n)]|y(n)|[2]
� �� � � �� �
3 RMs 3 RMs
[a30,r _yr_ (n) − _a30,iyi(n)]|y(n)|[3]_
� �� �
3 RMs
+j [a30,r _yi(n) + a30,iyr_ (n)]|y(n)|[3]
� �� �
3 RMs
+2 from eq. (6) +4 from eq. (8) 12 RMs (9)
−−−−−−−−−−−−−−−−−−→
+12 from eq. (9) +1 from |y(n)|[2]
19 RMs (10)
−−−−−−−−−−−−−−−−−−−→
+19 from eq. (10) +1 from |y(n)|[3]
26 RMs (11)
−−−−−−−−−−−−−−−−−−−→
-----
that fully describes the complexity (in number of RMs) of the
memory polynomial model,
[2] Technical Specification Group Services and System Aspects: Release 15
_Description, Standard 3GPP TR 21.915, 2019._
[3] Common Public Radio Interface (CPRI); Interface Specification, document CPRI Specification V7.0, 2015.
[4] Technical Specification Group Radio Access Network: Study on CU-DU
_Lower Layer Split for NR, Annex A: Fronthaul Bandwidth (Release 15),_
Standard 3GPP TR 38.816, 2017.
[5] C. Ranaweera, E. Wong, A. Nirmalathas, C. Jayasundara, and C. Lim, ‘‘5G
C-RAN with optical fronthaul: An analysis from a deployment perspective,’’ J. Lightw. Technol., vol. 36, no. 11, pp. 2059–2068, Jun. 1, 2018,
[doi: 10.1109/JLT.2017.2782822.](http://dx.doi.org/10.1109/JLT.2017.2782822)
[6] J. Wang, C. Liu, J. Zhang, M. Zhu, M. Xu, F. Lu, L. Cheng, and
G.-K. Chang, ‘‘Nonlinear inter-band subcarrier intermodulations of multiRAT OFDM wireless services in 5G heterogeneous mobile fronthaul networks,’’ J. Lightw. Technol., vol. 34, no. 17, pp. 4089–4103, Sep. 1, 2016,
[doi: 10.1109/JLT.2016.2584621.](http://dx.doi.org/10.1109/JLT.2016.2584621)
[7] B. G. Kim, S. H. Bae, H. Kim, and Y. C. Chung, ‘‘RoF-based mobile fronthaul networks implemented by using DML and EML for 5G wireless communication systems,’’ J. Lightw. Technol., vol. 36, no. 14, pp. 2874–2881,
[Jul. 15, 2018, doi: 10.1109/JLT.2018.2808294.](http://dx.doi.org/10.1109/JLT.2018.2808294)
[8] S.-H. Cho, H. Park, H. S. Chung, K. H. Doo, S. Lee, and J. H. Lee,
‘‘Cost-effective next generation mobile fronthaul architecture with multiIF carrier transmission scheme,’’ in Proc. Opt. Fiber Commun. Conf.,
Mar. 2014, pp. 1–3.
[9] J. Zhang, J. Wang, M. Xu, F. Lu, L. Chen, J. Yu, and G.-K. Chang,
‘‘Memory-polynomial digital pre-distortion for linearity improvement of
directly-modulated multi-IF-over-fiber LTE mobile fronthaul,’’ in Proc.
_Opt. Fiber Commun. Conf., Mar. 2016, pp. 1–3._
[10] X. N. Fernando and A. B. Sesay, ‘‘Look-up table based adaptive predistortion for dynamic range enhancement in a radio over fiber link,’’ in Proc.
_IEEE Pacific Rim Conf. Commun., Comput. Signal Process. (PACRIM)_
_[Conf., Aug. 1999, pp. 26–29, doi: 10.1109/PACRIM.1999.799469.](http://dx.doi.org/10.1109/PACRIM.1999.799469)_
[11] S. Liu, M. Xu, J. Wang, F. Lu, W. Zhang, H. Tian, and G.-K. Chang,
‘‘A multilevel artificial neural network nonlinear equalizer for millimeterwave mobile fronthaul systems,’’ J. Lightw. Technol., vol. 35, no. 20,
[pp. 4406–4417, Oct. 15, 2017, doi: 10.1109/JLT.2017.2717778.](http://dx.doi.org/10.1109/JLT.2017.2717778)
[12] User Equipment (UE) Radio Transmission and Reception; Part 1: Range
_1 Standalone (Release 15), Standard 3GPP TS 38.101-1, 2018._
[13] M. A. Fernandes, P. A. Loureiro, B. T. Brandao, A. Lorences-Riesgo,
F. P. Guiomar, and P. P. Monteiro, ‘‘Multi-carrier 5G-compliant DMLbased transmission enhanced by bit and power loading,’’ IEEE Pho_ton. Technol. Lett., vol. 32, no. 12, pp. 737–740, Jun. 15, 2020, doi:_
[10.1109/LPT.2020.2994045.](http://dx.doi.org/10.1109/LPT.2020.2994045)
[14] A. Mufutau, F. Guiomar, M. Fernandes, A. Lorences-Riesgo, A. Oliveira,
and P. Monteiro, ‘‘Demonstration of a hybrid optical fiber–wireless 5G
fronthaul coexisting with end-to-end 4G networks,’’ J. Opt. Commun.
_Netw., vol. 12, pp. 72–78, Mar. 2020._
[15] M. Alzenad, M. Z. Shakir, H. Yanikomeroglu, and M.-S. Alouini, ‘‘FSObased vertical backhaul/fronthaul framework for 5G+ wireless networks,’’
_IEEE Commun. Mag., vol. 56, no. 1, pp. 218–224, Jan. 2018._
[16] Z. Zakrzewski, ‘‘D-RoF and A-RoF interfaces in an all-optical fronthaul
of 5G mobile systems,’’ Appl. Sci., vol. 10, no. 4, p. 1212, Feb. 2020.
[17] Common Public Radio Interface Interface Specification, document eCPRI
Interface Specification V1.0, 2017.
MARCO A. FERNANDES (Member, IEEE)
received the M.Sc. degree in electronics and
telecommunications engineering from the University of Aveiro, in 2019. He is currently pursuing
the Ph.D. degree in MAP-tele doctorate program
from the University of Aveiro, the University of
Porto, and the University of Minho. During his
M.Sc. degree, he has worked with analog-radio
over fiber applied to 5G communications.
During his master’s, he worked with advanced
radio-over-fiber transmission providing 5G-solutions for Optical Radio Convergence Infrastructure for Communications and Power Delivering (ORCIPwww.orcip.pt) testbed. He is currently participates in multiple research
projects, mainly involved high-capacity free space optics (FSO) transmission
and machine learning applications. He has authored or coauthored more than
15 scientific publications in leading international journals and conferences.
He is an Optica Member. He received the Ph.D. Grant from FCT, in 2020.
In 2021, he was a finalist in OFC2021 Corning Student Award.
Nmul =
�
4Q, if K = 1,
(7)
(6 + 6(K − 1) + (K − 2))Q, if K ≥ 2.
Having obtained this relation, we can now calculate the
number of multiplications required to implement the memory polynomial DPD at the best complexity-vs-performance
tradeoff previously found in Figs. 9, 10 and 13, i.e. K 5,
=
_Q_ 1 and K 3, Q 2, yielding a complexity of
= = =
33 and 38 multiplications, respectively. Through this in-depth
complexity analysis, we can then conclude that the utilized
memory polynomial model is effectively a low-complexity
subsystem. As a baseline for comparison, the number of
multiplications required for its operation is lower than what
would be required for a standard linear filter with 10 taps (i.e.
4Q as in the upper branch of (8), for K 1).
=
**VII. CONCLUSION**
In this paper, we addressed one of the main challenges in
the upcoming next-generation RANs, namely, the bandwidth
bottleneck imposed by digital fronthauling in typical architectures. With the rise of 5G and the emergence of 6G specifications, it is required to search for alternative technologies that meet these unprecedented demands. Answering to
these requirements and responding to the scarcity of low-cost
analog optical transceivers, we have demonstrated a simple
procedure to take a low-cost COTS digital SPF transceiver,
and modify it to perform analog transmission. With the simple
modifications exposed in the paper, we were able to obtain
an SFP-packaged analog transceiver, capable of transmitting
100 MHz and 400 MHz 64QAM signals meeting the 3GPP
EVM requirements. Moreover, a memory-polynomial based
pre-distortion technique has been shown to partially counteract the limitations inherent to the simple digital-to-analog
adaption procedure, enabling to meet the EVM specifications
for transmitting a 100 MHz 256QAM signal over 20 km SMF.
Although the proposed digital-to-analog adaption of the
SFP transceiver is not deemed as a practical solution for the
marketization of analog RoF transceivers, the results presented in this work demonstrate that it is possible to design
high-performance analog RoF solutions using low-cost components that have found matured deployment in the low-end
digital optics market. Furthermore, the analog-adaption
methodology and digital pre-distortion technique demonstrated in this work might provide useful insights for the
research community, facilitating the access to low-cost RoF
solutions as an enabling technology to support the experimentation and prototyping of complex 5G and 6G optical access
architectures in laboratory environments.
**REFERENCES**
[1] I. A. Alimi, A. L. Teixeira, and P. P. Monteiro, ‘‘Toward an efficient C-RAN optical fronthaul for the future networks: A tutorial
on technologies, requirements, challenges, and solutions,’’ IEEE Com_mun. Surveys Tuts., vol. 20, no. 1, pp. 708–769, 1st Quart., 2018, doi:_
[10.1109/COMST.2017.2773462.](http://dx.doi.org/10.1109/COMST.2017.2773462)
-----
BRUNO T. BRANDÃO (Member, IEEE)
received the M.Sc. degree in electronics and
telecommunications engineering from the University of Aveiro, Portugal, in 2019. During his M.Sc.
degree, he has developed an analogue radio-overfibre link, based on low-cost optical transceivers,
for 4G and 5G communication support. Since
2019, he has been with the Telecommunications Ph.D. Program (MAP-Tele), a joint venture
between the Universities of Minho, Portugal, and
the Universities of Aveiro. In his Ph.D. studies, he is developing and implementing digital signal processing algorithms in reconfigurable hardware for
real-time coherent communication systems. Along with the Ph.D. studies,
he is working in the RETIOT Project under the M.Sc. fellowship in which he
is developing a distributed radio system for both radio communications and
coherent radar applications. He also designed an RF frontend to serve as an
interface between the remote user equipment and the radio-over-fibre link.
This work was done under the scope of ORCIP infrastructure (www.orcip.pt)
at the Instituto de Telecomunicações.
ABEL LORENCES-RIESGO received the Ph.D.
degree from the Chalmers University of Technology, in 2017. From 2017 to 2019, he worked as
a Postdoctoral Researcher at Telecomunicações–
Aveiro. In 2019, he joined the Optical Communication Technology Laboratory, Paris Research
Center, Huawei Technologies France, as a Senior
Engineer. He has authored or coauthored more
than 70 papers in leading international journals and
conferences. His main research interests include
fiber optic communications and digital signal processing.
PAULO P. MONTEIRO (Senior Member, IEEE) is
currently an Associate Professor at the University
of Aveiro and a Senior Researcher at the Instituto
de Telecomunicações, where he is also a Research
Coordinator of optical communication systems
(https://www.it.pt/Groups/Index/59). He successfully tutored over 14 Ph.D. students and 24 master’s students. He has participated in more than
26 research projects. He has authored/coauthored
more than 18 patent applications, over 115 articles
in journals, and 380 conference contributions. His main research interests
include optical communications and reflectometry systems.
FERNANDO P. GUIOMAR (Member, IEEE)
received the M.Sc. and Ph.D. degrees in electronics and telecommunications engineering from
the University of Aveiro, Portugal, in 2009 and
2015, respectively. Since 2017, he has been a
Senior Researcher at the Instituto de Telecomunicações, Aveiro, where his main research interests are focused within the area of fiber-based
and free-space optical communication systems,
including the development of digital signal processing algorithms, advanced modulation and coding, constellation shaping
and non-linear modeling and mitigation. He has authored or coauthored more
than 100 scientific publications in leading international journals and conferences. He is an OSA Member. In 2015, he received a Marie SkłodowskaCurie individual fellowship, jointly hosted by the Politecnico di Torino,
Italy, and CISCO Optical GmbH, Germany. In 2016, he has received the
Photonics21 Student Innovation Award, distinguishing industrial-oriented
research with high impact in Europe. In 2020, he was awarded a three year
Junior Leader Fellowship by the ‘‘la Caixa’’ Foundation.
BRUNO
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.1109/ACCESS.2022.3154784?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/ACCESS.2022.3154784, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "GOLD",
"url": "https://ieeexplore.ieee.org/ielx7/6287639/9668973/09721876.pdf"
}
| 2,022
|
[
"JournalArticle"
] | true
| null |
[
{
"paperId": "cabd0425553bb2e5fd6d61992a473c6f4db8f162",
"title": "Multi-Carrier 5G-Compliant DML-Based Transmission Enhanced by Bit and Power Loading"
},
{
"paperId": "22190aea929c806856b9467b9c911cdc36109b20",
"title": "D-RoF and A-RoF Interfaces in an All-Optical Fronthaul of 5G Mobile Systems"
},
{
"paperId": "938192a90615fac58aa9b8d0ac4fabeeb28395ac",
"title": "Demonstration of a hybrid optical fiber–wireless 5G fronthaul coexisting with end-to-end 4G networks"
},
{
"paperId": "95d73ad6062651c7d42e3a7f60efe43d604236c1",
"title": "RoF-Based Mobile Fronthaul Networks Implemented by Using DML and EML for 5G Wireless Communication Systems"
},
{
"paperId": "8b9edb606ee1c3f86ab39dbf823889bdd6487786",
"title": "5G C-RAN With Optical Fronthaul: An Analysis From a Deployment Perspective"
},
{
"paperId": "f6a3f3ef2a242e2c9d60bd7439fe1a2ef1f68ea5",
"title": "A Multilevel Artificial Neural Network Nonlinear Equalizer for Millimeter-Wave Mobile Fronthaul Systems"
},
{
"paperId": "3d9a5775b900725c2268cf1bf62ddfdb547aa548",
"title": "Nonlinear Inter-Band Subcarrier Intermodulations of Multi-RAT OFDM Wireless Services in 5G Heterogeneous Mobile Fronthaul Networks"
},
{
"paperId": "0cc8551881da255f2cfeebbbb973ec2d473c1b2d",
"title": "Memory-polynomial digital pre-distortion for linearity improvement of directly-modulated multi-IF-over-fiber LTE mobile fronthaul"
},
{
"paperId": "6978786aa1de5db57eeb90d1edeb03a96eccbbc0",
"title": "Cost-effective next generation mobile fronthaul architecture with multi-IF carrier transmission scheme"
},
{
"paperId": "4d91322b48291edb03f1fa6b26b2520062ed2d09",
"title": "Look-up table based adaptive predistortion for dynamic range enhancement in a radio over fiber link"
},
{
"paperId": "afbc1ba078afee5588dd2b0ae90153eeb0733692",
"title": "Toward an Efficient C-RAN Optical Fronthaul for the Future Networks: A Tutorial on Technologies, Requirements, Challenges, and Solutions"
},
{
"paperId": null,
"title": "User Equipment (UE) Radio Transmission and Reception"
},
{
"paperId": null,
"title": "FSObased vertical backhaul/fronthaul framework for 5G+ wireless networks"
},
{
"paperId": null,
"title": "Release 15 Description"
},
{
"paperId": null,
"title": "Technical Specification Group Radio Access Network: Study on CU-DU Lower Layer Split for NR, Annex A: Fronthaul Bandwidth (Release 15)"
},
{
"paperId": null,
"title": "Common Public Radio Interface Interface Specification, document eCPRI Interface Specification V1"
}
] | 12,275
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Economics",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/021a44734d384688e277b852910ac80cb2ab8c34
|
[
"Computer Science"
] | 0.870149
|
An Agent-Based Model Framework for Utility-Based Cryptoeconomies
|
021a44734d384688e277b852910ac80cb2ab8c34
|
arXiv.org
|
[
{
"authorId": "38081739",
"name": "Kiran Karra"
},
{
"authorId": "2225941210",
"name": "Tom Mellan"
},
{
"authorId": "2226149183",
"name": "Maria Silva"
},
{
"authorId": "1557361372",
"name": "Juan P. Madrigal-Cianci"
},
{
"authorId": "2225941452",
"name": "Axel Cubero Cortes"
},
{
"authorId": "1471432581",
"name": "Zixuan Zhang"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"ArXiv"
],
"alternate_urls": null,
"id": "1901e811-ee72-4b20-8f7e-de08cd395a10",
"issn": "2331-8422",
"name": "arXiv.org",
"type": null,
"url": "https://arxiv.org"
}
|
In this paper, we outline a framework for modeling utility-based blockchain-enabled economic systems using Agent Based Modeling (ABM). Our approach is to model the supply dynamics based on metrics of the cryptoeconomy. We then build autonomous agents that make decisions based on those metrics. Those decisions, in turn, impact the metrics in the next time-step, creating a closed loop that models the evolution of cryptoeconomies over time. We apply this framework as a case-study to Filecoin, a decentralized blockchain-based storage network. We perform several experiments that explore the effect of different strategies, capitalization, and external factors to agent rewards, that highlight the efficacy of our approach to modeling blockchain based cryptoeconomies.
|
SS 379 5980 (o e)
DOI 10.5195/LEDGER.20XX.XXX
RESEARCH ARTICLE
# An Agent-Based Model Framework for Utility-Based Cryptoeconomies
### Kiran Karra, Tom Mellan, Maria Silva, Juan P. Madrigal-Cianci, Axel Cubero Cortes, Zixuan Zhang [∗]
**Abstract. In this paper, we outline a framework for modeling utility-based blockchain-enabled**
economic systems using Agent Based Modeling (ABM). Our approach is to model the supply
dynamics based on metrics of the cryptoeconomy. We then build autonomous agents that
make decisions based on those metrics. Those decisions, in turn, impact the metrics in the
next time-step, creating a closed loop that models the evolution of cryptoeconomies over
time. We apply this framework as a case-study to Filecoin, a decentralized blockchain-based
storage network. We perform several experiments that explore the effect of different strategies,
capitalization, and external factors to agent rewards, that highlight the efficacy of our approach
to modeling blockchain based cryptoeconomies.
KEY WORDS
1. Agent-Based Modeling. 2. Cryptoeconomics. 3. Digital Twin.
### 1. Introduction
#### Cryptoeconomics is an interdisciplinary science that combines fields such as economics, cryp- tography, and computer science with the goal of designing and analyzing economic incentive structures for resource allocation in decentralized systems.[1] Accordingly, cryptoeconomic sys- tems are often used to create new forms of digital currency, utilities, and markets. Because
each system has its own goals and contexts in which it is applicable, cryptoeconomic incentive
structures usually need to be customized for each individual application. In addition, these systems typically show features associated with Complex Systems.[1] This means that the long-term
#### evolution of these systems cannot be easily inferred from local changes caused by individuals,
which makes the task of customizing cryptoeconomic systems to support a concrete application
more difficult.
#### Even though cryptoeconomics is a relatively young field,[2] some work has been done to address the complexities of designing and tuning decentralised economies. An exciting new
approach is to use Agent-based modeling (ABM).[3] ABM is a computational modeling technique
#### that has been used to study a wide variety of complex systems, including social systems,[4]
economic systems,[5,6] and biological systems.[7] In ABM, the system is modeled as a collection of agents, each of which has its own set of rules and behaviors. The agents interact with each
_∗_ Kiran Karra and Tom Mellan contributed equally to this work.
[All authors are research scientists at CryptoEconLab (https://cryptoeconlab.io)](https://cryptoeconlab.io)
-----
LEDGER VOL X (20XX) X X
other and with the environment, and the system’s behavior emerges from the interactions of the
individual agents.[3]
#### Within the cryptoeconomics space, ABM has the potential to support practitioners in three
main areas:
(1) Study the cryptoeconomics and robustness of the blockchain to agent behavior. As
#### an example, Struchkov et al.[8] used ABM to test how Decentralised Exchanges would
respond to stress market conditions and front-running, while Cocco et al.[9] used ABM to
analyse the mining incentives in the Bitcoin network;
(2) Explore the design space of blockchain networks. For instance, ABM has been applied to
compare different token designs and their impact on prediction markets;[10]
(3) Test new features and protocols. Following the fair launch allocation from Yearn Finance,
a group of researchers[11] used ABMs to examine the concentration of voting rights tokens
after the launch, under different trading modalities.
#### This paper explores how ABM can be adapted to a particular type of decentralised system, namely utility-based decentralised networks. These networks employ their own currency to
provide consumptive rights on the services or the product being offered by the network.[12] Thus,
#### these systems mediate a marketplace of providers of a specific good and users that want to
consume the good. Since the entire system depends on the good being traded, any tool attempting
to model such system needs to consider how changes in utility impact the system and its agents.
#### Therefore, we propose a framework for applying ABMs to utility-based cryptoeconomies. Our approach is complimentary to other methods in the literature[13] and builds on the work of Zhang, et. al[14] to enable multi-scale coupling between individual microeconomic preferences
and protocol specific supply dynamics. Rational users of cryptoeconomic systems will base their
decisions on some aspects of the network they are involved in, which in turn affects the network.
This natural feedback loop is well represented in our framework.
#### We apply the framework to Filecoin, a decentralised data storage network,[15] and conduct two experiments that uncover interesting aspects of the network. The first explores the agents’ reward trajectories under different lending rates, while the other examines how the current
cryptoeconomic mechanisms of Filecoin impact wealth distribution.
The rest of this paper is organized as follows. We begin in Section 2 by presenting the general
#### framework for applying ABM to utility-based systems. This framework is then applied to the Filecoin network, by first developing a mathematical model of Filecoin’s supply dynamics in
Section 3. In Section 4, we describe the ABM that leverages this model to simulate a closed loop
interaction of programmable agents within the Filecoin economy. Section 5 follows with the two
experiments that showcase the utility of our ABM framework to understand the Filecoin system.
We conclude in Section 6 by framing the results in the context of utility based token economies
and discussing future research paths.
### 2. ABM Framework for Utility Cryptoeconomies
#### Agent-based models (ABMs) are a tool for modeling complex systems[3] with a high degree of granularity. They consist of two primary components which interact with each other: a) the
environment, which models the system under study, and b) agents, which take actions that affect
the environment.
-----
LEDGER VOL X (20XX) X X
**Formal Definition**
#### We define the general framework of our deterministic, discrete-time ABM as follows. Let E
denote the set of all possible environmental variables. For a given time d ∈ N, let Ed ∈ E denote
#### the environment at time d, and define the set E ∋ Ed := E0 × ··· × Ed. In our specific setting,
_Ed,_ _Ed+1,... corresponds to the environments at day d, at day d +_ 1, etc. In addition, let A be
#### the abstract set of agents and let H denote the abstract set of actions. To each agent a ∈ A, there corresponds a given set of actions ha ∈ H . Furthermore, we define the update rules
_fd[H]+1_ [:][ A][ ×] _[E][d][ �→]_ _[H][,][ f][ E]d+1_ [:][ H][ ×] _[E][ �→]_ _[E][, and][ f][ A]d+1_ [:][ E][d][ �→] _[A][, as some abstract functions that]_
update the actions of each agent, the environment and agents respectively. Given a set of agents
_A0_ _A, at time d = 0, where each agent a0_ _A0 is equipped with a set of actions ha0_ _H and_
_⊂_ _∈_ _∈_
an initial environment E0, the ABM proceeds by iterating as follows:
_ha,d+1 = fd[S][(][a][d][,]_ _[E][d][)]_ _∀ad ∈_ _Ad,_ (1)
_Ed+1 = fd[E]_ [ą] _ha,d+1,_ _Ed_ (2)
_ad_ _∈Ad_
_Ad+1 = fd[A][(][E][d][)][.]_ (3)
**Environment**
The supply dynamics, defined as factors which affect the supply of tokens in the cryptoeconomy,
are modeled by the environment, E. Factors include the total supply of tokens, the rate at which
#### new tokens are mined, and the rate at which tokens are taken out of circulation through means
such as burning. This information is often used by investors and traders who want to understand
the potential value of a cryptoeconomy to then make decisions about potential investments into
that economy.
Three aspects of the environment must be defined:
(1) Network Performance Metrics - what are the key set of metrics that are to be modeled
and used as performance indicators when evaluating the results of the ABM simulation?
(2) Inputs - when actors take part in the cryptoeconomy, what is the subset of actions that
they can take that would affect the network metrics?
(3) Outputs - what are the outputs that flow back to miners, which in turn affect miners
decisions about their actions in the next time step? For example, in a typical blockchain,
rewards are issued to miners and because the rate and trajectory of rewards affects agent’s
behavior, this is a key output. Additional outputs, such as a subset of the overall network
metrics that miners can use to make rational decisions can also be included.
**Agents**
Miners in blockchains are mapped to agents, Ai. Miners take actions that correspond to the inputs
#### of the environment and make those decisions based on the outputs of the environment they are interacting with. The actions they take affect the network’s performance metrics, which then
affects the outputs that the agents are fed. Through this feedback loop, the dynamical nature of
the cryptoeconomy is modeled.
-----
LEDGER VOL X (20XX) X X
Fig. 1. Proposed framework for mapping cryptoeconomies to an ABM. Miners are represented
by agents and the environment is mapped to a supply dynamics model.
#### Fig. 1 shows a diagram of this proposed framework for mapping cryptoeconomies to an
ABM.
**Examples**
#### Three examples of cryptoeconomic projects where this framework can provide value include
Helium,[16] Ethereum,[17] and Filecoin.[15]
#### Helium is a decentralized wireless network where miners provide wireless coverage in
exchange for HNT tokens. An ABM utilizing the described framework can be used to understand,
#### for example, how the rate of token distribution and population density may affect expected network coverage. This requires a mathematical model of how tokens are minted (the supply
dynamics), given agent inputs (i.e. the wireless coverage they provide to the network). This can
then be used to design new incentive structures to ensure a more even coverage distribution.
Ethereum is another example where the described ABM framework can be applied. A specific
use case could be analyzing how user behavior (agents) could stress and effect the total circulating
#### supply of ETH tokens after the “Shapella” upgrade, which enabled easier unlocking of staked ETH. In this setting, one could model, e.g., the propensity of a participant to unlock, against more staking inflows due to the ability to unlock. Here, the agents actions (whether to stake or
unstake) have a direct effect on the supply dynamics and the outlined framework would enable
one quantify various scenarios that may play out.
We now discuss the application of the ABM framework to Filecoin in more detail.
### 3. Modeling Filecoin Supply Dynamics
#### Filecoin is a distributed storage network based on the blockchain, where miners, referred to as storage providers (SPs), provide storage capacity for the network and earn units of the Filecoin cryptocurrency (FIL) by periodically producing cryptographic proofs that certify they
are providing the promised storage capacity. In contrast to using Nakamoto-style proof of work
#### to maintain consensus on the chain, Filecoin uses proof of storage: a miner’s voting power — the probability that the network elects a miner to create a new block — is proportional to their
current quality-adjusted storage in use in relation to the rest of the network. The cryptoeconomics
#### of Filecoin are designed to incentivize storage providers to participate and grow the collective
utility of the data storage network.
The following subsections describe various aspects of the Filecoin supply dynamics.
-----
LEDGER VOL X (20XX) X X
**Circulating Supply**
Filecoin’s circulating supply Sd is modeled at a daily (d) level of aggregation and has four parts:
_Sd+1 = Md +Vd_ _−_ _Ld −_ _Bd_ _._ (4)
���� ����
inflow outflow
These correspond to minted block rewards Md, vested tokens Vd, locked tokens Ld, and burnt
tokens Bd.
**Power Onboarding and Renewals**
The dynamics of Md, Ld, and Bd depend on the amount of storage power onboarded and renewed
in the network.
#### In Filecoin, storage providers (SPs) participate by onboarding power onto the network by adding a sector for a committed duration. Power is measured in units of sectors, which can be either 32 GiB or 64 GiB in size. Each sector consists of a fraction of committed capacity (CC)
and verified deal data (FIL+).[18] An SPs can choose to renew CC sectors when they expire.
#### We model power in aggregate terms rather than at the sector level. This means that we model the network’s storage power to be split into two categories: 1) CC and 2) FIL+. This
approximation is valid for the granularity that we are seeking to achieve with our modeling.
Filecoin has two methods to measure the power of the network, the network’s raw byte power
#### (RBP) and the network’s quality adjusted power (QAP). Network RBP is a measure of the raw
storage capacity (in bytes) of the network — it does not distinguish between the kind of data that
#### is stored on the network. For example, empty or random data stored on the network is counted
the same as a widely used dataset when computing network RBP. A second measure of network
power is quality adjusted power (QAP). QAP is a derived measurement that captures the amount
of useful data being stored on the network. Considering the aggregated approximation discussed
above, we compute the quality adjusted power, Pd[QA] of the network on day d as
_Pd[QA]_ = (1 _−_ _γ)_ _·_ _Pd[RB]_ + 10 _·_ _γ ·_ _Pd[RB][,]_ (5)
#### where γ ∈ [0, 1] is the overall FIL+ rate of the network, and Pd[RB] is the raw byte power of the
network on day d. Eq. 5 reveals that FIL+ power is given a 10x multiplier when computing the
QA power of the network.[18]
An initial pledge collateral of FIL tokens is required in order to onboard or renew power, and
#### the specific amounts and time-windows are discussed below. In exchange for onboarding and renewing power onto the network, and continually submitting storage proofs to the chain, SPs
can receive block rewards in the form of FIL tokens.
**Rewards from minting**
#### Filecoin uses a hybrid minting model that has two components — simple minting and baseline
minting. The total number of tokens minted by day d is the sum of these two minting mechanisms:
_Md = Md[S]_ [+] _[M]d[B]_ (6)
-----
LEDGER VOL X (20XX) X X
Simple minting is defined by an exponential decay model
_Md[S]_ [=][ M]∞[S] _[·]_ [(][1] _[−]_ _[e][−][λ][ d][)][,]_ (7)
which decays at a rate of λ = [ln]6yrs[(][2][)] [, corresponding to a 6-year half-life.][ M]∞[S] [takes a value of 30%]
of the maximum possible minting supply of 1.1B tokens. Tokens emitted via simple minting are
independent of network power. This is similar to minting schemes present in other blockchains.[19]
#### The second component of minting in Filecoin is baseline minting, Md[B][. Baseline minting]
depends on network power and aims to align incentives with the growth of the network’s utility.
#### The minting function still follows an exponential decay, however, it now decays based on the
effective network time, θd. The equations describing this are:
_Md[B]_ [=][ M]∞[B] _[·]_ [(][1] _[−]_ _[e][−][λθ][d]_ [)]
� �
_θd =_ _g[1]_ [ln] _gRb0[∑]d_ + 1 (8)
_R[∑]d_ [=] ∑ min{bd, _Pd[RB][}]_
_d∈_ _D_
From these definitions, we can compute the cumulative baseline minting by day d from the
cumulative capped RBP of the network:
_−λ_ _gR[∑]d_ �
1 _e_ _g_ [ln][(] _b0_ [+][1][)]
_−_
_Md[B]_ [=][ M]∞[B] _[·]_
= M∞[B] _[·]_
�
1
_−_
=
(9)
� � _−gλ_
_gR[∑]d_ + 1
_b0_
In this expression M∞[B] [takes a value of 70% of the maximum possible supply of 1.1B tokens.][ R][∑]
#### is the cumulative capped network RBP; it is the sum of the point-wise minimum of network’s
RBP and the baseline storage function for each day:
_bd = b0 e_ _[gd]_ (10)
The baseline storage function serves as a target for the network to hit to maximize the baseline
minting rate. In this expression g = [log][(][2][)]
365 [, the baseline storage growth rate which corresponds to]
2 year doubling, and b0 = 2.88888888EiB is the initial baseline storage.
**Vesting**
Vesting supply, which can contribute 0.9B tokens, is modelled daily, summing across the set of
recipients R as
Vd = ∑ _Vr,d ._ (11)
_r∈_ _R_
Different recipients have different linear vesting schedules.
**Locked tokens**
Locked tokens in the network Ld are made up of storage collaterals Ld[S] [and vesting block rewards]
_Ld[R]_ [as]
-----
LEDGER VOL X (20XX) X X
_Ld = Ld[S]_ [+] _[L]d[R]_ (12)
The locked balance for vesting blocked rewards is modeled as
_Ld[R]_ [=][ 0][.][75] ∑ ∆ _Md−τ ·_ _rd[R]_ _[.]_ (13)
_τ≤d_
where ∆ _Md is daily minted rewards and rd[R]_ [is a vector specifying the release linear release over]
180days.
The locked storage collateral is modeled as having the following dynamics
_Ld[S]_ [=] ∑ ∆ _Ld[S]−τ_ _[·]_ _[r]d[S]_ _[.]_ (14)
_τ≤d_
#### where ∆ Ld[S] [is newly locked storage collateral tokens, and][ r]d[S] [is a vector specifying a release]
schedule.
Newly locked collateral tokens are given by
∆ _Ld[S]_ [=][ ∆] _[L]d[SP]_ [+] [∆] _[L]d[CP]_ _[.]_ (15)
where the ’storage pledge’ locked tokens are
∆ _Ld[SP]_ [=][ max] [(][20] _[·]_ [∆][M][d][,][ 0][)][,] (16)
and ’consensus pledge’ locked tokens are
(17)
∆ _Ld[CP]_ = max
_d_ _,_ 0
[0][.][3] _[·]�[S][d][ ·]_ [∆][P][QA]�
max _Pd[QA], bd_
where ∆Pd[QA] denotes new quality-adjusted power onboarded on day d.
**Burnt tokens**
Burnt tokens are modeled as consisting of termination fees B[T]d [and base fees from gas usage][ B]d[G]
as:
_Bd = B[T]d_ [+] _[B]d[G]_ (18)
#### where terminations are accounted for by aggregating agent decisions and gas fees as linearly
increasing as B[G]d [=][ β][ d][ at average rate][ β] [.]
### 4. ABM of Filecoin
Utilizing the framework developed in Section 2, we create an ABM of Filecoin. Storage providers
#### (SPs) are mapped to agents, and the environment consists of the supply dynamics described in Section 3. We divide the environment into three logical modules: a) Network State, b)
Forecasting, and c) the external environment. These components interact with each other in the
following manner.
-----
LEDGER VOL X (20XX) X X
Fig. 2. Summary of the Agent Based Model of the Filecoin network, with arrows indicating
direction of data flow.
#### Agents determine the amount of power they will onboard and renew onto the network for
day d. All agents decisions are aggregated and passed into the network state module. Using the
developed model of the supply dynamics, the network state is updated. By utilizing both historical
network metrics and network forecasting information, agents can make rational decisions. Finally,
#### an external environment simulates constraints that SPs are subject to in the real world, such as borrowing costs of pledge collateral. These components interact to create a closed-loop
simulation of the Filecoin economy.
Fig. 2 summarizes the components and dataflow of the Filecoin ABM.
**Agents**
Agents directly influence the outcomes of the simulation, since their actions are aggregated and
used as inputs to update the network state. We have developed three types of agents which use the
network and forecasting information in different ways to make decisions regarding onboarding
and renewing power:
(1) DCAAgent - This is the dollar cost averaging agent, and does not use any forecasting
information or historical network information to make decisions. The agent is configured
#### to onboard a constant amount of power per day, the percentage of that power which corresponds to verified deals, and the percentage of expiring power which should be renewed. This is a dollar-cost averaging strategy and can be useful in understanding
relative performance to more complex strategies.
(2) FoFRAgent - This agent utilizes the rewards/sector forecast provided by the network
#### to internally forecast the FIL-on-FIL returns (FoFR) of onboarding sectors for various sector durations, where FoFR = [rewards]
pledge [. This metric can additionally be generalized]
#### to introduce arbitrary cost structures. Because pledge (Eq. 17) is dependent upon the Network QAP and agents must make decisions for a given timestep before the overall Network QAP is aggregated for a day, it is approximated by using the previous day’s
pledge. If the estimated FoFR for any of the tested sector durations exceeds a configurable
#### threshold (which indirectly represents the risk profile of the agent), then the agent will
onboard a configured amount of power. It will also renew a configured amount of power
-----
LEDGER VOL X (20XX) X X
under the same condition.
(3) NPVAgent - This agent utilizes the rewards/sector forecast provided by the network
to compute the net present value (NPV) of onboarding power for various sector durations.
NPV is the present value of the expected rewards/sector less costs/sector. Present
#### value is computed using the continuous discounting formula. The agent discount rate, configured upon instantiation, is a proxy for the risk profile of the agent, with a higher
discount rate representing a higher risk aversion. The agent will onboard and renew power
at a sector duration which maximizes NPV, but will take no action for day d if NPV < 0
for all durations tested.
**Network State**
Network state consists of: a) network power, and b) the status of tokens.
The total network power is summed across all of the agents individual contributions for each
#### day, and the type of power onboarded (CC or FIL+) is also tracked. This enables the network
state to track both Pd[RB] and Pd[QA].
Token status consists of three parts: a) the amount mined, b) the amount locked due to pledge,
#### and c) the remaining that has been released into the circulating supply through vesting. The mined tokens are distributed to agents as rewards. The fraction of total rewards mined in a day
are distributed proportional to the fraction of total network QAP. Tokens are locked when agents
#### onboard power to provide a consensus pledge. This information is aggregated to compute the
overall network state of the number of tokens locked. Using Eq. 4, the circulating supply of the
network is computed.
**Forecasting**
#### To enable agents to make rational decisions, relevant forecasts are computed and provided to
agents, who can choose to utilize them. As previously mentioned, agents use rewards/sector
to make rational decisions. This quantifies the amount of rewards an agent can expect to receive
for onboarding a given sector. The metric is forecast by first utilizing the historical network RBP
#### and QAP to train a time-series forecasting model, M . M is used to forecast RBP and QAP trajectories until the end of the simulation, denoted by P[ˆ]d[RB] and P[ˆ]d[QA], respectively. P[ˆ]d[RB] is then
used to compute the expected minting rate, ˆmd, using Eqs. 6, 7, and 9. Finally, because rewards
are distributed proportional to the network’s QA power, _rewardssectorˆ_ [[][d][] =][ ˆ][m][d][/][ ˆ][P]d[QA].
There are currently two models implemented for RB and QA forecasting, linear extrapolation
#### and a variant of Markov-Chain Monte Carlo.[20,21] Agents may also elect to perform custom forecasting using network metrics, and this represents a competitive advantage that a certain
agent may have.
**External Environment**
Agents must borrow tokens in order to satisfy the consensus pledge (Eq. 17) needed to onboard
#### power. The borrowing rate is modeled as an external environment process that specifies the discount rate, Rd, at which agents can borrow tokens. Agents use this information to make rational decisions regarding onboarding and renewing power. The purpose of this is to model
-----
LEDGER VOL X (20XX) X X
Fig. 3. Validating our supply dynamics and ABM through backtesting. (a) The mined FIL of
the model to the historical data, and (b) The circulating supply computed by the model against
historical data.
realities that SPs have to face, when determining their strategy for being involved in the Filecoin
network. Additional real-world complexities can also be modeled here.
**Model Validation**
We begin by validating the model of Filecoin’s supply dynamics that was developed and described
#### in Section 3, using backtesting. Our approach is to instantiate one DCAAgent which onboards and renews the historical power that was onboarded onto the network for that day. This is in
contrast to the typical use-case of an agent, which is making daily decisions about whether and
#### how much power to onboard and renew. Then, the relevant statistics for circulating supply are
calculated from the start of simulation to the present date. This is then compared against actual
statistics from the Filecoin network retrieved from Spacescope.[22] Fig.3 shows the results of this
experiment: a) shows the minted tokens, and b) shows the circulating supply. For each of these
network statistics, the implemented model tracks the historical data with good accuracy. Slight
#### differences are observed, and these can be attributed to not modeling certain intricacies of the
Filecoin network, such as variable sector durations.
### 5. Experiments and Results
#### In this section, we describe some experiments that showcase the utility of ABM in modeling
blockchain networks, using the Filecoin network as a case study.
**Sensitivity of Rewards to External Discount Rates**
#### In this experiment, we explore how the cryptoeconomics of Filecoin and external factors such
as borrowing rates affect agent rewards. We instantiate two subpopulations of NPVAgents, one
subpopulation is configured to only onboard verified deals (which corresponds to FIL+ power),
while the second is configured to only onboard storage capacity (CC power). Both subpopulations
of agents are configured to have identical risk profiles by instantiating the agents with the same
#### discount rates. Fig. 4 shows agent rewards trajectory, with different colors indicating different
external discount rates.
#### Fig. 4(a) shows that irrespective of external discount rates, FIL+ agents are more profitable than CC agents. This is a direct consequence of the cryptoeconomic mechanism in place in
-----
LEDGER VOL X (20XX) X X
Fig. 4. Experiment exploring the sensitivity of returns to external discount rates with two
subpopulations of agents, FIL+ and CC. In (a) both exhibit the same risk profile, (b) the CC
agent has 2x the risk aversion of the FIL+ agent.
Filecoin to incentivize FIL+ data, through the 10x quality adjusted (QA) multiplier. Secondly, we
see the effect of external borrowing rates on agent profitability. As expected, higher rewards are
correlated with lower borrowing rates. However, the rewards trajectory does not change linearly
with the borrowing rate and starts to oscillate as borrowing rates increase. This is an example of
an interesting dynamic that emerges as a result of the agent based simulation.
#### We extend this experiment by altering one aspect of the previous experimental setup - that is, we increase the risk aversion of the CC agent to be two-times the risk aversion of the FIL+ agent. We then examine the agent rewards trajectory, shown in Fig. 4(b). Because the FIL+
agents have the same risk as before, their rewards trajectories are identical. However, we notice
#### that when the external discount rate is 30%, the risk-averse CC agent manages a more positive
rewards trajectory than the non risk-averse CC agent. The effect of this disappears as the external
borrowing rates decrease, however.
**Wealth Concentration**
#### In this experiment, we explore how the distribution of starting capital in the cryptoeconomic network affects the ability to get rewards from the network. Our experimental setup consists of five DCAAgents which are configured to represent different levels of capitalization. This is represented with a vector [a1, a2, a3, a4, a5], where the relative capitalization of ai is defined as
_ci = ai/_ ∑[5]i=1 _[a][i][.]_
In Filecoin, onboarding power requires, in addition to pledge collateral, sealing of sectors via
cryptographic proofs that require large computational resources. It is reasonable to assume that
agents with larger capitalization will have more hardware resources to perform this than agents
with smaller capitalization, thereby having a larger sealing throughput. To model this, we scale
#### how much power an agent is able to onboard and renew, per day, by its relative capitalization.
The mapping from capitalization to sealing throughput captures the idea of wealth concentration.
#### To compare and interpret the results, the overall power onboarded is kept constant across the
three experiments.
We test three distributions of initial capital:
(1) All agents have equal starting capital (20%) - this is considered the baseline and corre
-----
LEDGER VOL X (20XX) X X
Fig. 5. The trajectory of rewards for agents with various starting capitalizations, relative to the
baseline distribution, where all agents are equally capitalized (20%).
sponds to the the vector [1, 1, 1, 1, 1].
(2) One agent has 50% of the starting capital, and the remainder have 50/4 = 12.5% starting
capital each and corresponds to the vector [4, 1, 1, 1, 1].
(3) The agent capitalization follows the distribution: [33%, 27%, 20%, 13%, 7%] and corresponds to the configuration vector [5, 4, 3, 2, 1].
#### Fig. 5 shows the reward trajectories of each agent, relative to the baseline case where each
agent has 20% of the starting capital. We observe that relative to the max-capitalized agent, the
rewards trajectories of other agents are on a decreasing trend. This is a consequence of the fact
that both onboarding and renewals are a function of the agent capitalization.
### 6. Conclusion
#### In this paper, we have outlined a framework for applying ABM to modeling utility based blockchain economies, and validated our framework with Filecoin as a case study. Our experiments shed light on some interesting aspects of Filecoin, including agent reward trajectories when
#### taking into account external lending rates, and how the cryptoeconomic structure of Filecoin
distributes wealth.
The sensitivity experiments indicate that creating new, competitive lending markets with smart
#### contracts leveraging programmable platforms such as FVM[23] can enable network growth and
increase miner returns. The wealth concentration experiments indicate that starting capitalization
has a significant effect on total rewards in the future. By explicitly modeling this effect with the
supply dynamics, one can then design new incentive structures to either accentuate, maintain, or
perhaps reverse the trend based on the goals of the project. Insights such as these, enabled by the
ABM framework, can help designers and creators of cryptoeconomies to more efficiently achieve
their goals. This indicates that ABM can be a valuable tool for researchers to better understand
and design blockchain economies.
#### In the future, we plan to explore additional aspects of blockchain economies that are well
mapped to ABMs, such as the effect of information quality, availability, and lag on agent reward
#### trajectories, and related network science questions. Another potential research direction is to include uncertainty by considering a probabilistic ABM, while balancing the computational
constraints using methods such as Multi-level and Multi-Index Monte Carlo methods.[24–27]
-----
LEDGER VOL X (20XX) X X
### Author Contributions
KK developed the ABM codebase and helped devise experiments that were conducted with the
#### framework. TM developed the mathematical models of the Filecoin economy and steered the project. They both contributed equally to manuscript preparation. TM and MS implemented the initial mathematical models, which were then ported to the ABM framework. JC provided
formalism and consulting on ABM related topics. AC and ZZ helped putting the project in larger
context and helped with manuscript preparation.
### Notes and References
1 Voshmgir, S., Zargham, M., et al. “Foundations of cryptoeconomic systems.” Research Institute for Cryptoe_conomics, Vienna, Working Paper Series/Institute for Cryptoeconomics/Interdisciplinary Research 1._
2 Davidson, S., De Filippi, P., Potts, J. “Economics of Blockchain.” (2016) doi:10.2139/ssrn.2744751 URL
```
https://papers.ssrn.com/abstract=2744751.
```
3 Macal, C. M., North, M. J. “Agent-based modeling and simulation.” In Proceedings of the 2009 winter
_simulation conference (WSC) IEEE 86–98 (2009) ._
4 Terano, T. “A Perspective on Agent-Based Modeling in Social System Analysis.” In G. S. Metcalf,
K. Kijima, H. Deguchi (Eds.), Handbook of Systems Sciences Singapore: Springer 1–13 (2020)doi:
[10.1007/978-981-13-0370-8 5-1 URL https://doi.org/10.1007/978-981-13-0370-8_5-1.](https://doi.org/10.1007/978-981-13-0370-8_5-1)
5 Chen, S.-H., Chang, C.-L., Du, Y.-R. “Agent-based economic models and econometrics.” _The_ _Knowledge_ _Engineering_ _Review_ **27.2** 187–219 (2012) doi:10.1017/
S0269888912000136 publisher: Cambridge University Press URL `https://www.`
```
cambridge.org/core/journals/knowledge-engineering-review/article/abs/
agentbased-economic-models-and-econometrics/DF3E4987809567A9B277F83ED6E22E00.
```
6 Fagiolo, G., Guerini, M., Lamperti, F., Moneta, A., Roventini, A. “Validation of Agent-Based Models
in Economics and Finance.” In C. Beisbart, N. J. Saam (Eds.), Computer Simulation Validation: Fundamental
_Concepts, Methodological Frameworks, and Philosophical Perspectives Cham: Springer International Publish-_
ing Simulation Foundations, Methods and Applications 763–787 (2019)doi:10.1007/978-3-319-70766-2 31
[URL https://doi.org/10.1007/978-3-319-70766-2_31.](https://doi.org/10.1007/978-3-319-70766-2_31)
7 Eubank, S., et al. “Modelling disease outbreaks in realistic urban social networks.” Nature 429.6988 180–184
(2004).
8 Struchkov, I., Lukashin, A., Kuznetsov, B., Mikhalev, I., Mandrusova, Z. “Agent-Based Modeling of
Blockchain Decentralized Financial Protocols.” In 2021 29th Conference of Open Innovations Association
_(FRUCT) 337–343 (2021) doi:10.23919/FRUCT52173.2021.9435601 iSSN: 2305-7254._
9 Cocco, L., Tonelli, R., Marchesi, M. “An Agent Based Model to Analyze the Bitcoin Mining Activity and a
Comparison with the Gold Mining Industry.” Future Internet 11.1 8 (2019) doi:10.3390/fi11010008 number: 1
[Publisher: Multidisciplinary Digital Publishing Institute URL https://www.mdpi.com/1999-5903/11/1/](https://www.mdpi.com/1999-5903/11/1/8)
```
8.
```
10 Hulsemann, P., Tumasjan, A. “Walk this Way! Incentive Structures of Different Token Designs for¨
[Blockchain-Based Applications.” ICIS 2019 Proceedings URL https://aisel.aisnet.org/icis2019/](https://aisel.aisnet.org/icis2019/blockchain_fintech/blockchain_fintech/7)
```
blockchain_fintech/blockchain_fintech/7.
```
11 Fernandez, J. D., Barbereau, T., Papageorgiou, O. “Agent-based Model of Initial Token Allocations: Evaluating Wealth Concentration in Fair Launches.” (2022) 2208.10271.
12 Benedetti, H., Abarzua, L., Caceres Fuentes, C. “Utility Tokens.” (2021) doi:10.2139/ssrn.4088568 URL´
```
https://papers.ssrn.com/abstract=4088568.
```
[13 Akcin, O., Streit, R. P., Oommen, B., Vishwanath, S., Chinchali, S. https://eprint.iacr.org/2022/](https://eprint.iacr.org/2022/1492)
```
1492 “A Control Theoretic Approach to Infrastructure-Centric Blockchain Tokenomics.” (2022) Cryptology
```
[ePrint Archive, Paper 2022/1492 URL https://eprint.iacr.org/2022/1492.](https://eprint.iacr.org/2022/1492)
-----
LEDGER VOL X (20XX) X X
14 Zhang, Z., Zargham, M., Preciado, V. M. “On modeling blockchain-enabled economic networks as stochastic
dynamical systems.” Applied Network Science 5.1 1–24 (2020).
15 Benet, J., Greco, N. “Filecoin: A decentralized storage network.” Protoc. Labs 1–36.
16 Haleem, A., Allen, A., Thompson, A., Nijdam, M., Garg, R. “A decentralized wireless network.” Helium
_Netw 3–7._
17 Buterin, V., et al. “A next-generation smart contract and decentralized application platform.” white paper
**3.37 2–1 (2014).**
[18 Labs, P. apr 28, 2023 “Filecoin Spec.” (2018) URL https://spec.filecoin.io.](https://spec.filecoin.io)
19 Nakamoto, S. “Bitcoin: A peer-to-peer electronic cash system.” Decentralized business review 21260.
20 Neal, R. M., et al. “MCMC using Hamiltonian dynamics.” Handbook of markov chain monte carlo 2.11 2
(2011).
21 Hoffman, M. D., Gelman, A., et al. “The No-U-Turn sampler: adaptively setting path lengths in Hamiltonian
Monte Carlo.” J. Mach. Learn. Res. 15.1 1593–1623 (2014).
[22 Labs, S. apr 28, 2023 “Spacescope.” (2018) URL https://spacescope.io.](https://spacescope.io)
[23 Accessed: 2022-06-09 “Introducing the Filecoin Virtual Machine.” https://filecoin.io/blog/posts/](https://filecoin.io/blog/posts/introducing-the-filecoin-virtual-machine/)
```
introducing-the-filecoin-virtual-machine/.
```
24 Giles, M. B. “Multilevel monte carlo methods.” Acta numerica 24 259–328 (2015).
25 Madrigal-Cianci, J. P., Kristensen, J. “Time-efficient Decentralized Exchange of Everlasting Options with
Exotic Payoff Functions.” In 2022 IEEE International Conference on Blockchain (Blockchain) 427–434 (2022)
doi:10.1109/Blockchain55522.2022.00066.
26 Madrigal-Cianci, J. P., Nobile, F., Tempone, R. “Analysis of a class of multilevel Markov chain Monte Carlo
algorithms based on independent Metropolis–Hastings.” SIAM/ASA Journal on Uncertainty Quantification 11.1
91–138 (2023).
27 Qian, E., Peherstorfer, B., O’Malley, D., Vesselinov, V. V., Willcox, K. “Multifidelity Monte Carlo estimation
of variance and sensitivity indices.” SIAM/ASA Journal on Uncertainty Quantification 6.2 683–706 (2018).
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/2307.15200, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "CLOSED",
"url": "https://arxiv.org/pdf/2307.15200"
}
| 2,023
|
[
"JournalArticle"
] | true
| 2023-07-27T00:00:00
|
[
{
"paperId": "79d0267c870ed7ef88625c0571808220523a82e0",
"title": "Analysis of a Class of Multilevel Markov Chain Monte Carlo Algorithms Based on Independent Metropolis-Hastings"
},
{
"paperId": "301bc3c7c0665488cbb31294acce73218a00e15a",
"title": "A Control Theoretic Approach to Infrastructure-Centric Blockchain Tokenomics"
},
{
"paperId": "ad414c6ad72556d4cdd5930de3fec4a5a52a1ed2",
"title": "Agent-based Model of Initial Token Allocations: Evaluating Wealth Concentration in Fair Launches"
},
{
"paperId": "e629c12ea77ca1ac79e6bc109e97c41da674307c",
"title": "Time-efficient Decentralized Exchange of Everlasting Options with Exotic Payoff Functions"
},
{
"paperId": "90c80f9286abbc460bf5130adcf72868e80fe02b",
"title": "Agent-Based Modeling of Blockchain Decentralized Financial Protocols"
},
{
"paperId": "1cb9f168dbf66cff10891f9ee29a31bfa86ef32f",
"title": "On modeling blockchain-enabled economic networks as stochastic dynamical systems"
},
{
"paperId": "8ea45c6b86d702db129a32d2c50caf37c7b86cd3",
"title": "Foundations of Cryptoeconomic Systems"
},
{
"paperId": "70a881f0c11bc5c17654269aa8df8c14a36eed28",
"title": "Validation of Agent-Based Models in Economics and Finance"
},
{
"paperId": "6749c78e9f12012a9fb58194438a24bad641dfd7",
"title": "An Agent Based Model to Analyze the Bitcoin Mining Activity and a Comparison with the Gold Mining Industry"
},
{
"paperId": "f0ae6eda680130ae4e0701febf6886c65f8b6b3f",
"title": "Multifidelity Monte Carlo Estimation of Variance and Sensitivity Indices"
},
{
"paperId": "13fca50483e298e931e1b804e95089c26a773cd9",
"title": "Economics of Blockchain"
},
{
"paperId": "d16a7375c98f888333ebe253062a767c784b9dfb",
"title": "Agent-based economic models and econometrics"
},
{
"paperId": "e1103d528d874a9e8e84ca443fe3fd5c1ff9eb9e",
"title": "The No-U-turn sampler: adaptively setting path lengths in Hamiltonian Monte Carlo"
},
{
"paperId": "3b0821ff22fdffc95b0caae1f9660773eb54dc52",
"title": "Handbook of Markov Chain Monte Carlo"
},
{
"paperId": "b716f735fb9b229c8227495e596e2d52f2981c33",
"title": "MCMC Using Hamiltonian Dynamics"
},
{
"paperId": "d9f25ae6216f47d8351705cab828219362d7099f",
"title": "Agent-based modeling and simulation"
},
{
"paperId": "1413dbfbbae1b59d52656db3dc48a4ee278e082f",
"title": "Modelling disease outbreaks in realistic urban social networks"
},
{
"paperId": "d512f19f7cdf5994df388d8bc5c5a828c311b870",
"title": "Multilevel Monte Carlo Methods"
},
{
"paperId": "9103b4aa754ab20959a30473c7e9a54fe7d51248",
"title": "Utility Tokens"
},
{
"paperId": "bc18522debf8b54c660ec21d404538e609333d56",
"title": "A Perspective on Agent-Based Modeling in Social System Analysis"
},
{
"paperId": "a93cf4648862778128c10dfa7f48da1347b4cb82",
"title": "Computer Simulation Validation - Fundamental Concepts, Methodological Frameworks, and Philosophical Perspectives"
},
{
"paperId": "5f6de94b4b150f200e1e3d4f3bf8390720b08680",
"title": "Walk this Way! Incentive Structures of Different Token Designs for Blockchain-Based Applications"
},
{
"paperId": null,
"title": "Spacescope"
},
{
"paperId": "0dbb8a54ca5066b82fa086bbf5db4c54b947719a",
"title": "A NEXT GENERATION SMART CONTRACT & DECENTRALIZED APPLICATION PLATFORM"
},
{
"paperId": "4e9ec92a90c5d571d2f1d496f8df01f0a8f38596",
"title": "Bitcoin: A Peer-to-Peer Electronic Cash System"
},
{
"paperId": null,
"title": "“ Filecoin : A decentralized storage network"
},
{
"paperId": null,
"title": "Introducing the Filecoin Virtual Machine"
},
{
"paperId": null,
"title": "Filecoin Spec"
},
{
"paperId": null,
"title": "Study the cryptoeconomics and robustness of the blockchain to agent behavior"
}
] | 10,067
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Business",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/021bb3f039ede8bfe80eb9749a32b3e9c8bc6b29
|
[
"Computer Science"
] | 0.897764
|
Flexible Integration of Blockchain with Business Process Automation: A Federated Architecture
|
021bb3f039ede8bfe80eb9749a32b3e9c8bc6b29
|
CAiSE Forum
|
[
{
"authorId": "50696201",
"name": "M. Adams"
},
{
"authorId": "2914969",
"name": "S. Suriadi"
},
{
"authorId": "2116451209",
"name": "Akhil Kumar"
},
{
"authorId": "143613819",
"name": "A. Hofstede"
}
] |
{
"alternate_issns": null,
"alternate_names": null,
"alternate_urls": null,
"id": null,
"issn": null,
"name": null,
"type": null,
"url": null
}
| null |
-----
# Flexible Integration of Blockchain with Business Process Automation: A Federated Architecture
Michael Adams[1], Suriadi Suriadi[1], Akhil Kumar[2], and
Arthur H. M. ter Hofstede[1]
1 Queensland University of Technology, Brisbane, Australia
_{mj.adams, s.suriadi, a.terhofstede}@qut.edu.au_
2 Smeal College of Business, Penn State University, University Park, USA
```
akhil@psu.edu
```
**Abstract. Blockchain technology enables various business transactions**
to be performed in an immutable and transparent manner. Within the
business process management community, blockchain technology has been
positioned as a way to better support the execution of inter-organisational
business processes, where the entities involved may not completely trust
each other. However, the architectures proposed thus far in the literature
for blockchain-enabled business process management can be described as
“heavy-weight”, since they promote the blockchain platform as the monolithic focal point of all business logic and process operations. We propose
an alternative: a federated and flexible architecture that leverages the capabilities of blockchain, but without overloading the functionalities of the
blockchain platform with those already extant in Business Process Management Systems (BPMSs). We illustrate its benefits, and demonstrate
its feasibility, through the implementation of a prototype.
**Keywords: blockchain; process flexibility; business process automation;**
business process management systems.
## 1 Introduction
A blockchain is a tamper-proof, replicated and distributed ledger [10] to which
multiple parties can append transactional records in such a way that modification is prevented, in a technically-enforceable manner. Blockchain technology
effectively guarantees that transactions, once recorded, become immutable [17],
facilitating the execution of transactions across multiple, potentially untrusted
parties without the need for a trusted intermediary. Naturally, blockchain opens
up new opportunities to support the execution of cross-organisational business
processes (i.e. those processes that necessitate interactions involving multiple
discrete players) typically seen in many domains, such as supply chain management and manufacturing.
In recent years, the Business Process Management (BPM) community has
investigated ways to exploit blockchain for secure, cross-organisational process
execution (see [6–8, 19] for some initial approaches). In this paper, we specifically focus on the alternative architectural designs that integrate blockchain
-----
2 M. Adams et al.
technologies with business process management systems (BPMS) to support process executions involving multiple, independent parties. We call such a system a
_blockchain-integrated BPMS._
A prominent architecture proposed for blockchain-integrated BPMS transforms a business process, expressed as one or more process models, into smart
contracts (programmable transactions) that are then executed entirely upon a
blockchain platform [6,19]. That is, all business rules, branching logic, instance
data, resource allocation, access authorisations and process state management is
deployed to and handled by the blockchain platform. Thus, the focal point of this
architecture resides in the blockchain and the different parties involved in the
business process must interact directly with this blockchain, both during process
design time and runtime executions. We shall refer to such an architecture as
_blockchain-centric._
While blockchain-centric architectures may be appealing for some business
process applications, and under certain threat assumptions and/or risk scenarios, it is not a universal solution. It is a heavy-weight architecture, with a rigidity that may not be necessary, or even desirable, in many other business process applications, for example where interactions between multiple parties are
loosely-coupled and/or may involve asynchronous-type interactions. A heavyweight architecture also overloads a blockchain system with a host of supporting
compilers, components and mechanisms required to wholly accommodate business process design and execution within a distributed ledger. In effect, this tight
integration necessitates a duplication of the capabilities that already exist within
core execution engines of BPMSs.
Hence, we propose an alternative federated architecture, that is more decentralised, cooperative, and flexible, simpler to realise and better suited to meet
the needs of a wide variety of cross-organisational settings. The proposed architecture allows component parts to each perform their fit-for-purpose capabilities
in a federated whole, rather than overloading components with functionalities
that are better performed by others. This architecture is flexible in that it is
not tied to any particular type of blockchain platform or BPMS. It supports
the minimisation of append operations to a blockchain, which are known to be
resource-intensive [18,21], and does not require the creation and propagation of
multiple smart contracts per execution instance.
This paper is structured as follows: a background discussion and related
work are presented in Section 2, Section 3 establishes the need for a federated
blockchain-integrated BPMS, and Section 4 describes the proposed architecture,
while Section 5 illustrates its implementation. A comparative discussion of the
federated and blockchain-centric approaches is presented in Section 6, followed
by the conclusion of the paper.
## 2 Background and Related Work
The key advantages of blockchain, besides immutabiity, include: visibility (all
authorised participants can view the transactions); validation (transactions are
-----
Flexible Integration of Blockchain with Process Automation 3
endorsed by peers through a designated consensus mechanism prior to being
written to the chain); and resilience (a replicated ledger means there is no single
point-of-failure).
A blockchain system can be permissioned (exercise membership control) or
_permissionless (publicly accessible). For example, Ethereum[1]_ [3] is (by default)
a permissionless blockchain platform, where any peer can join to read or submit transactions at any time. Moreover, there is no central entity to manage
membership, although private and permissioned blockchains can also be configured. Permissioned blockchain systems are designed to better address concerns
around transaction security, privacy and scalability [1]. Hyperledger Fabric[2] [4]
is an example of a permissioned blockchain framework. Another key aspect of
blockchain technology is the provision of so-called smart contracts [5,15], i.e. executable scripts that reside on the blockchain and automate the steps and rules
corresponding to the business logic of the bespoke transactional operations.
In recent research efforts towards integrating blockchain technology with
BPM [6,10,19], the authors propose an architecture that tightly integrates business process execution with blockchain by encapsulating the entire business process logic into smart contracts. In this approach, a translator component takes
a process specification as input and generates a set of corresponding smart contracts per process instance. In addition, a choreography monitor uses smart
contracts to control a collaborative business process. A prototype has been developed for the Ethereum platform [19].
Architectural design issues of blockchain based systems with an eye towards
quality and performance attributes are addressed in [20] in the form of a taxonomy and flowchart. Other performance issues that have been addressed are
availability [18] and latency [21]. Methods for optimising execution of business
processes on an Ethereum blockchain by improving data structures and runtime
components are discussed in [6] and demonstrated in a prototype called Caterpillar [8]. Approaches for implementing collaborative, data-aware business processes on blockchain using the business artifact paradigm are discussed in [2,7],
focussing on a new business collaboration language.
Sturm et al. [14] develop a generic approach to control-flow management
within the blockchain by having one contract that handles choice and parallel
structures. However, the control-flow capabilities are limited and data management is not discussed. There are also a plethora of approaches to interorganisational process management that use platforms and environments other
than blockchain, for example [9,11,13].
All these related approaches have helped to locate our work in context. However, our approach is different in that we believe that the essential functionality
of a BPM system should not be migrated to the blockchain. Instead, we explore
a lean approach (along the lines of [14]) wherein the BPM system can interface with the blockchain as a repository of reliable data and for executing key
contractual terms through smart contracts.
1 https://www.ethereum.org/
2 https://www.hyperledger.org/projects/fabric
-----
4 M. Adams et al.
## 3 Towards a Federated Blockchain-integrated BPMS
Consider the pharmaceutical use case scenario shown in Figure 1. In this crossorganisational process, a Pharmacy places an order for medical supplies with
its Distributor, who in turn requests the production of the pharmaceuticals
by the Manufacturer. Once the pharmaceuticals are manufactured, they are
delivered to the Distributor who then sends them to the Pharmacy.
Using different types of blockchain with any pair of distributor and pharmacy
(at different process instances)
|Blockchain Type B (with Manufacturer 1)|Col2|Col3|Col4|
|---|---|---|---|
||Blockchain Type B (with Manufacturer 1)|||
|||Blockchain Type B (with Manufacturer 1)||
|||||
Using different types of blockchain with any pair of
distributor and manufacturer
(at different process instances)
**Fig. 1. Pharmaceutical Supply Process - Multiple Ledgers**
When this process is executed, there is a potential for conflict across different
parties. For example, if the Distributor fails to deliver the ordered pharmaceuticals on time, the Distributor may blame the Manufacturer for being late with
production, or the Distributor may dispute the date and time when it received
the original order. Therefore, the use of blockchain in recording the process transactions can be beneficial. Moreover, each organisation can exercise full control
over their own private business process, and share information of only selected
activities that involve cross-organisational interactions, as shown in Figure 1.
There are many desirable features of this approach. Firstly, the parties in the
process do not need to agree on a common inter-organisational process. They
may even be on different blockchain platforms so long as they are compatible.
Secondly, the lower transparency requirement will increase the willingness of the
-----
Flexible Integration of Blockchain with Process Automation 5
parties to cooperate with each other. Thirdly, there is more scalability in such an
arrangement since in general a pharmacy will deal with multiple distributors, and
a distributor, in turn, with multiple manufacturers. Thus, this use case calls for
a more flexible, decentralised, loosely-coupled and distributed approach based on
platform heterogeneity, for both two-party and multi-party interactions, which
minimises the need for interactions with the blockchain platform.
**Towards a Federated Approach We propose a federated, blockchain inte-**
grated BPMS architecture to address the issues identified above. Such an architecture should provide the following properties:
– Separation of Concerns: A clear separation of capabilities should be maintained between business logic operations and distributed transactional execution records, with the aim of minimising the performance hit on blockchain
operations and maximising the fit-for-purpose capabilities of the BPMS and
blockchain platforms.
– Platform Heterogeneity: The architecture should allow the use of more
than one compatible blockchain platform within and across a composite set
_of interacting process instances._
– Compartmentalisation of Interactions: A requirement that all interactions between any two participating parties need to be transparent to all
parties involved should not be imposed. A blockchain-centric architecture
may perhaps support this through the use of, for example, separate permissioned channels, but this should not be seen as a necessary realisation, and it
still imposes the requirement that they share the same blockchain platform.
– Single-party Interaction: The architecture should not assume that all interactions between a business process and a blockchain involve multi-party
communication. Hence, it should support simple single-party interaction between an organisation’s business process and its corresponding blockchain.
## 4 Conceptual Architecture
In our federated approach, each organisation hosts a discrete BPMS that encapsulates a service or middleware component through which it will delegate
designated tasks, designed to perform a required inter-organisational activity,
within a process execution instance. The service will then interact with a properly configured blockchain network.
Each participating service in an inter-organisational process is granted authorisation to a discrete permissioned channel (or other authenticating, secure
pipeline) on a blockchain network. A channel is a private overlay that partitions
a blockchain network to provide data isolation and confidentiality [1]. Whenever a new block is written, an event notification is generated by the blockchain
platform and then relayed to the BPMS through the service. The service will
by default listen for events as they occur, but it may also be configured to
periodically request the event history from past blocks, to accommodate those
-----
6 M. Adams et al.
**Fig. 2. Conceptual internal architecture**
deployments where connection to the blockchain network is not always available.
The service will take one of three actions for each received event notification,
depending on how the service has been configured for each event: (1) release
a task that has been waiting for the event to occur; (2) launch a new process
instance, using the event as a trigger; or (3) ignore the event.
Hence, the only information exchanged between organisations is that required
for work to be handed over and performed within each organisation (e.g. purchase order, invoice, contract, application). The state of a process instance can
be inferred from the history of data associated with it on the blockchain, for
example an order has been placed, a shipment was sent, a payment was made,
etc. This eliminates the need for sharing additional information about exact process state on the blockchain, or any process definitions, business logic and rules,
organisational data, or resource allocations that should remain private to their
respective organisations.
A transaction (such as placing a purchase order) submitted by one organisation to the blockchain will, within a short period, be written to a block on the
blockchain after it is validated by other peer nodes on the network using a validation algorithm, and ordered along with other transactions into a block structure.
The creation of a new block will trigger an event notification which may be used
by another organisation to complete a task in one of its own processes or to
commence a new process instance (see Section 5 for more details).
An internal architecture of the proposed approach is given in Figure 2. The
BPMS of an organisation will delegate the execution of certain tasks to the
blockchain service (middleware component) using the appropriate API along
with the requisite data. The subcomponents of the middleware are:
– Smart Contract Invoker: Performs smart contract calls on the blockchain
to either query the current instance data that has been written to the chain,
-----
Flexible Integration of Blockchain with Process Automation 7
or requests the creation of a new transaction to store data to be shared with
another organisation.
– Event Listener: Listens and responds to events generated by the blockchain
network each time a new transaction is created. An event may trigger the
completion of a waiting task, or the launch of a new process instance via a
call to the BPM engine’s API.
– Task Cache: Stores tasks that are waiting for some event to occur on the
blockchain, that is some specific data to be made available from another organisation (e.g. order received, invoice produced, etc). When the designated
event occurs, that task can be further processed and/or completed, allowing
its parent process instance to continue.
– Authority Certificate Store: Stores the private and public keys authorising the service to access the channel to read from and submit to the ledger
on behalf of its owner organisation. Each call of a smart contract must be
accompanied by the appropriate certificates.
## 5 Implementation
A prototype service that implements the conceptual architecture described in
Section 4 has been realised in the YAWL business process management environment [16]. YAWL was selected because it is robust, fully open-source, and
offers a service-oriented architecture, allowing an interactive blockchain service
to be implemented independent of existing components. However, the generic
federated architecture is not limited to the YAWL environment, but rather is
applicable to any BPMS that supports the addition of service-oriented or middleware components for interacting with external networks and applications.
Importantly, absolutely no changes were required to be made to the YAWL environment itself to enable support for communication and interaction with a
blockchain network. The YAWL Blockchain Service, and its source code, can be
freely downloaded from the YAWL repository[3].
For this prototype implementation, a Hyperledger Fabric blockchain network was chosen, because it is open-source, can be deployed freely, does not require crypto-currency payments for its operations and supports a permissioned
network natively. Again, the architecture is not limited by this choice; other
blockchain platforms may be used.
The Blockchain Service has been developed as a YAWL custom service and so
may have tasks assigned to it at design time via the process editor. At runtime,
the engine delegates all such assigned tasks to the service for action, passing task
input data and metadata to it via a specific process engine API. Communication
between the YAWL Blockchain Service and the Hyperledger Fabric network is
handled via the Java software development kit (SDK) for Hyperledger[4]. Each
organisation maintains its own discrete YAWL environment and Blockchain Service.
3 https://github.com/yawlfoundation/yawl
4 https://github.com/hyperledger/fabric-sdk-java
-----
8 M. Adams et al.
**5.1** **Event Handling**
The architecture leverages the event generating capabilities of the blockchain
platform to provide change-of-state announcements in the end-to-end process,
in particular notifying parties in cross-organisational processes of actions that
have been taken by others. Events of interest can be used to release a task that
has been waiting for an action to occur, or can signify the triggering of a new
process instance within an organisation.
For a task that has been designated to wait for an action, a dedicated data
structure, which specifies the event to wait for and the values that uniquely identify the event as related to the current end-to-end process instance, is included
as an input parameter of the task. On being delegated the task at runtime, the
Blockchain Service stores details of the task, and the specifics of the WAIT data
values, and compares each incoming event with those parameters. When a match
occurs, the task is updated with the values attached to the event, and released,
i.e. returned to the core BPMS engine, allowing the process instance to continue.
The other type of event of interest to the Blockchain Service signifies the
triggering of a new case instance. For example, if a pharmacy raises a purchase
order and submits it to a blockchain, the event produced by the blockchain when
the order transaction is written to a block can be captured by the Blockchain
Service of a pharmaceuticals distributor and used to trigger the creation of a new
process to fulfil that order (see Section 5.2 below for more details). Events can
be defined as process triggering events via a dynamically loaded configuration
file, or via an input data variable for a task, or by using an administration tool.
**5.2** **Illustrative Example**
An example execution of typical interactions among the three organisations in
the supply chain scenario of Section 3 is illustrated in Figure 3. The processes
have been somewhat simplified for clarity in the discussion below, and are depicted in the YAWL language.
There are three interacting organisations: a Pharmacy that places orders
for the supply of pharmaceuticals, a Distributor that fulfils those orders, and a
_Manufacturer that fabricates and supplies pharmaceuticals for distribution. The_
Pharmacy interacts only with the Distributor, similarly the Manufacturer interacts only with the Distributor, and consequently the Distributor interacts with
both. To ensure data isolation and confidentiality, two channels are created, one
_Pharmacy_ _Distributor (called chPharmDist), the other Manufacturer_
_←→_ _←→_
_Distributor (chManuDist). Importantly, all internal processes remains private_
to each organisation, only the transactional data necessary to collaborate with
another organisation is shared via the blockchain.
While this scenario concerns a specific pharmacy-distributor-manufacturer,
more generally a distributor would deal with a number of different pharmacies
and manufacturers, and vice versa, all of which would potentially participate
as peers within the blockchain platform and may play a role in the validation
-----
Flexible Integration of Blockchain with Process Automation 9
**Fig. 3. Inter-organisational process interactions – supply chain example**
consensus that occurs when a transaction is submitted to the chain. Further, it
is of course also possible to have a single channel for all three parties if desired.
To illustrate a complete sequence of interactions, with reference to Figure 3
and the numeric labelling within it:
1. A composite process instance begins with the pharmacy process, when a new
order is generated and then sent, i.e. submitted and subsequently written to a
new block on the chain via a task delegated to its YAWL Blockchain Service.
2. Since the permissioned channel chPharmDist is shared by the Pharmacy
and the Distributor, the Distributor’s Blockchain Service detects the new
_BlockEvent and interprets it as a trigger to launch a new instance of its_
internal ‘supply’ process. The transaction data sent with the event (i.e. the
purchase order) is used as the originating data for the new instance.
3. The Distributor adds the order to a batch, then at a designated time submits
the batch order to the Manufacturer via submission to the blockchain via
the shared chManuDist channel shared by those two organisations.
4. The Manufacturer’s service receives the write BlockEvent, which triggers a
new instance of its own ‘manufacture’ process, using the transaction data in
the event (i.e. the batch order) as originating data.
5. Once the Manufacturer ships the order, the process archives the order details
on the blockchain.
6. This BlockEvent triggers the release of the waiting Receive and Verify task
in the Distributor’s process, allowing that process to continue.
7. Later, an invoice is produced by the Distributor and submitted to the blockchain via the chPharmDist channel.
8. The subsequent writing of the invoice to the chain causes a BlockEvent that
triggers the release of the waiting Receive Invoice task in the Pharmacy’s
process.
9. Eventually, the Pharmacy pays the invoice by submitting the payment transfer details to the chain.
-----
10 M. Adams et al.
10. The payment causes a BlockEvent that triggers the release of the waiting
_Receive Payment task in the Distributor’s process._
Significantly, this example illustrates that secure inter-organisational process
automation can be achieved using a federated architecture, and that the approach affords several concrete advantages when compared to the more heavyweight, blockchain-centric architectures:
– Efforts to combine the three processes into one overarching, monolithic, endto-end process model are no longer required, negating the need for a great
deal of collaboration between all parties, and the translation of the result
into a set of factory smart contracts. The architecture also avoids the need
for the creation, verification and storage of a new set of smart contracts for
every instance of the inter-organisational process.
– Because all business logic, branching rules, resources allocations, etc. are
handled by the BPMS, the smart contracts here are not overloaded with
procedural code, resulting in much simpler, faster to process transactions.
In this example, the smart contracts define data structures for order, invoice
and payment, and a trivial invoke function that either submits a transaction
or performs a query over existing blocks. The data structures are used to
(de)serialise JSON strings passed to/from the BPMS into block data.
– There is no need to ‘centralise’ the process on the blockchain. Each organisation retains autonomy of its own processes, and the foci of operations are
retained within the processes of each organisation’s BPMS.
– There is no requirement for the creation and maintenance of the “intricate
set of components” [19], prerequisite to the heavyweight architecture. Only
a standard BPMS environment, a simple middleware service and vanilla
blockchain network are needed.
– Unlike many blockchain-centric architectures, there is no requirement for
a central ‘mediator’ process to choreograph the interactions between each
organisation’s processes.
– There are no limitations placed on the types of process patterns supported.
Any pattern supported by the process language used by the process execution
environment (i.e. the BPMS) can be used in this approach, including those
more complex patterns that are difficult, if not impossible, to transform
into a smart contract, since all process executions are contained within the
BPMS, rather than on the blockchain.
– An unimpeachable audit trail is stored on the blockchain(s) and can be extrapolated for all inter-process activity instances between organisation pairs.
## 6 Discussion and Conclusion
Many blockchain-centric approaches use a blockchain monolithically as an entire
execution platform for business processes. Thus, a potentially large volume of
data, including process definitions in the form of smart contracts, business rule
definitions, datasets representing the work of a process instance, as well as its
-----
Flexible Integration of Blockchain with Process Automation 11
constantly updating state information, is written to, read from, and executed on
the blockchain. Depending on the smart contracts and business rules executed,
such data could contain potentially-confidential internal data of an organisation,
thus inadvertently and unnecessarily exposing private data to external parties.
As per our case example (Figure 3), it will require 20 large, custom contracts to be created (one for each task) in such blockchain-centric architectures
versus 7 short, generic contracts that merely write important transactions to
the blockchain in our approach. Additionally, each custom contract will require
considerable effort for verification, and the blind trust of each organisation that
the translation tool generates error-free smart contracts. Each update of a smart
contract requires that each peer must compile, instantiate and validate it before
it is committed to the blockchain, thus consuming resources and adding to the
overhead of the blockchain’s performance.
It is clear that blockchain is much more expensive as a medium for processing and storage than traditional media. Hence, it should be used as sparingly as
possible by minimising both the size of smart contracts and the amount of data
stored, while maintaining trust by means of a reliable audit trail. Extraneous
processing and data should go to traditional platforms that offer better performance, flexibility and technology heterogeneity, and less visibility across parties.
We are not convinced that it is necessary to reinvent the functionality of a BPM
engine, which includes complex control flow management, data management and
resource allocation, within a blockchain platform.
As we have demonstrated in our implementation, it is less work to integrate
blockchain into an application with our federated approach, when compared
to the more heavyweight blockchain-centric architectures. To fully transfer all
the features of an industrial strength BPM system onto a blockchain platform
could amount to a very long, risky and expensive undertaking, especially when
considering the non-trivial processes in real-world scenarios. Our prototype illustrates the advantages of dedicating the existing capabilities of BPMS for process
automation, and those of blockchain as an immutable, distributed ledger, to automate secure, cross-organisational process interactions without the overheads
necessitated by the heavyweight, blockchain-centric approach.
We believe our proposal aligns better with the underlying philosophy of
blockchain technology based on distributed autonomous organisations (DAOs)
[12]. We have presented a conceptual architecture and an implementation that
demonstrates the feasibility of the approach. The comparisons presented here
are mostly qualitative; a more thorough empirical comparison through experiments and quantitative data is needed and it will form our future work. More
work is also needed to optimise the distribution of on-chain and off-chain data,
and to validate the applicability of the federated approach with different types
of scenarios and use cases.
## References
1. Androulaki, E., Barger, A., Bortnikov, V., Cachin, C., et al.: Hyperledger Fabric: a
distributed operating system for permissioned blockchains. In: Proceedings of the
-----
12 M. Adams et al.
Thirteenth EuroSys Conference. p. 30. ACM (2018)
2. Astigarraga, T., Chen, X., Chen, Y., Gu, J., et al.: Empowering business-level
blockchain users with a rules framework for smart contracts. In: 16th International
Conference on Service-Oriented Computing. pp. 111–128. Springer (2018)
3. Buterin, V.: Ethereum: A next-generation smart contract and decentralized application platform (2014), https://github.com/ethereum/wiki/wiki/White-Paper
4. Cachin, C.: Architecture of the Hyperledger Blockchain Fabric. In: Distributed
Cryptocurrencies and Consensus Ledgers. vol. 310, p. 4 (2016)
5. Christidis, K., Devetsikiotis, M.: Blockchains and smart contracts for the internet
of things. IEEE Access 4, 2292–2303 (2016)
6. Garc´ıa-Ba˜nuelos, L., Ponomarev, A., Dumas, M., Weber, I.: Optimized execution
of business processes on blockchain. In: BPM. pp. 130–146. Springer, Cham (2017)
7. Hull, R., Batra, V.S., Chen, Y.M., Deutsch, A., et al.: Towards a shared ledger business collaboration language based on data-aware processes. In: 14th International
Conference on Service-Oriented Computing. pp. 18–36. Springer (2016)
8. L´opez-Pintado, O., Garc´ıa-Ba˜nuelos, L., Dumas, M., Weber, I., Ponomarev, A.:
Caterpillar: a business process execution engine on the ethereum blockchain. Software: Practice and Experience 49(7), 1162–1193 (2019)
9. Mendling, J., Hafner, M.: From inter-organizational workflows to process execution:
Generating BPEL from WS-CDL. In: CoopIS’05. pp. 506–515. Springer (2005)
10. Mendling, J., Weber, I., van der Aalst, W., vom Brocke, J., et al.: Blockchains for
business process management - challenges and opportunities. ACM Transactions
on Management Information Systems 9(1), 4:1–4:16 (Feb 2018)
11. Narendra, N.C., Norta, A., Mahunnah, M., Ma, L., Maggi, F.M.: Sound conflict
management and resolution for virtual-enterprise collaborations. Service Oriented
Computing and Applications 10(3), 233–251 (Sep 2016)
12. Norta, A.: Creation of smart-contracting collaborations for decentralized autonomous organizations. In: International Conference on Business Informatics Research. pp. 3–17. Springer (2015)
13. Norta, A., Grefen, P., Narendra, N.C.: A reference architecture for managing dynamic inter-organizational business processes. Data & Knowledge Engineering 91,
52–89 (2014)
14. Sturm, C., Szalanczi, J., Sch¨onig, S., Jablonski, S.: A lean architecture for
blockchain based decentralized process execution. In: International Conference on
Business Process Management Workshops. pp. 361–373. Springer (2018)
15. Szabo, N.: Formalizing and securing relationships on public networks. First Monday
**2(9) (1997)**
16. ter Hofstede, A., van der Aalst, W., Adams, M., Russell, N. (eds.): Modern Business
Process Automation: YAWL and Its Support Environment. Springer (2010)
17. Underwood, S.: Blockchain beyond bitcoin. Communications of the ACM 59(11),
15–17 (2016)
18. Weber, I., Gramoli, V., Ponomarev, A., Staples, M., et al.: On availability for
blockchain-based systems. In: 2017 IEEE 36th Symposium on Reliable Distributed
Systems (SRDS). pp. 64–73 (Sept 2017)
19. Weber, I., Xu, X., Riveret, R., Governatori, G., et al.: Untrusted business process
monitoring and execution using blockchain. In: BPM. pp. 329–347. Springer (2016)
20. Xu, X., Weber, I., Staples, M., Zhu, L., et al.: A taxonomy of blockchain-based
systems for architecture design. In: ICSA. pp. 243–252. IEEE (April 2017)
21. Yasaweerasinghelage, R., Staples, M., Weber, I.: Predicting latency of blockchainbased systems using architectural modelling and simulation. In: ICSA. pp. 253–256.
IEEE (April 2017)
-----
|
{
"disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references', 'abstract'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1007/978-3-030-58135-0_1?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1007/978-3-030-58135-0_1, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "CLOSED",
"url": ""
}
| 2,020
|
[
"JournalArticle"
] | false
| null |
[] | 7,468
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Medicine",
"source": "external"
},
{
"category": "Engineering",
"source": "s2-fos-model"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Business",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/021c89fecb5deff33c4bd7ab7f22800ad64ad22d
|
[
"Computer Science",
"Medicine"
] | 0.883362
|
IoT Data Qualification for a Logistic Chain Traceability Smart Contract
|
021c89fecb5deff33c4bd7ab7f22800ad64ad22d
|
Italian National Conference on Sensors
|
[
{
"authorId": "145893525",
"name": "Mohamed Ahmed"
},
{
"authorId": "3269191",
"name": "C. Taconet"
},
{
"authorId": "103678321",
"name": "Mohamed Ould"
},
{
"authorId": "1776634",
"name": "S. Chabridon"
},
{
"authorId": "1695916",
"name": "A. Bouzeghoub"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"SENSORS",
"IEEE Sens",
"Ital National Conf Sens",
"IEEE Sensors",
"Sensors"
],
"alternate_urls": [
"http://nbn-resolving.de/urn/resolver.pl?urn=urn:nbn:ch:bel-142001",
"http://www.mdpi.com/journal/sensors",
"https://www.mdpi.com/journal/sensors"
],
"id": "3dbf084c-ef47-4b74-9919-047b40704538",
"issn": "1424-8220",
"name": "Italian National Conference on Sensors",
"type": "conference",
"url": "http://www.e-helvetica.nb.admin.ch/directAccess?callnumber=bel-142001"
}
|
In the logistic chain domain, the traceability of shipments in their entire delivery process from the shipper to the consignee involves many stakeholders. From the traceability data, contractual decisions may be taken such as incident detection, validation of the delivery or billing. The stakeholders require transparency in the whole process. The combination of the Internet of Things (IoT) and the blockchain paradigms helps in the development of automated and trusted systems. In this context, ensuring the quality of the IoT data is an absolute requirement for the adoption of those technologies. In this article, we propose an approach to assess the data quality (DQ) of IoT data sources using a logistic traceability smart contract developed on top of a blockchain. We select the quality dimensions relevant to our context, namely accuracy, completeness, consistency and currentness, with a proposition of their corresponding measurement methods. We also propose a data quality model specific to the logistic chain domain and a distributed traceability architecture. The evaluation of the proposal shows the capacity of the proposed method to assess the IoT data quality and ensure the user agreement on the data qualification rules. The proposed solution opens new opportunities in the development of automated logistic traceability systems.
|
# sensors
_Article_
## IoT Data Qualification for a Logistic Chain Traceability Smart Contract
**Mohamed Ahmed** **[1,2,]*, Chantal Taconet** **[2,]*** **, Mohamed Ould** **[1], Sophie Chabridon** **[2]** **and Amel Bouzeghoub** **[2]**
1 ALIS International, 4 Rue du Meunier, 95724 Roissy-en-France, France; Mohamed.Ould@alis-intl.com
2 Samovar, Télécom SudParis, Institut Polytechnique de Paris, 9 rue Charles Fourier,
91011 Evry-Courcouronnes CEDEX, France; Sophie.Chabridon@telecom-sudparis.eu (S.C.);
Amel.Bouzeghoub@telecom-sudparis.eu (A.B.)
***** Correspondence: Mohamed.Ahmed@alis-intl.com (M.A.); Chantal.Taconet@telecom-sudparis.eu (C.T.)
[����������](https://www.mdpi.com/article/10.3390/s21062239?type=check_update&version=2)
**�������**
**Citation: Ahmed, M.; Taconet, C.;**
Ould, M.; Chabridon, S.; Bouzeghoub,
A. IoT Data Qualification for a
Logistic Chain Traceability Smart
Contract. Sensors 2021, 21, 2239.
[https://doi.org/10.3390/s21062239](https://doi.org/10.3390/s21062239)
Academic Editor: Muhamed
Turkanovi´c
Received: 29 January 2021
Accepted: 19 March 2021
Published: 23 March 2021
**Publisher’s Note: MDPI stays neutral**
with regard to jurisdictional claims in
published maps and institutional affil
iations.
**Copyright: © 2021 by the authors.**
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
[Attribution (CC BY) license (https://](https://creativecommons.org/licenses/by/4.0/)
[creativecommons.org/licenses/by/](https://creativecommons.org/licenses/by/4.0/)
4.0/).
**Abstract: In the logistic chain domain, the traceability of shipments in their entire delivery process**
from the shipper to the consignee involves many stakeholders. From the traceability data, contractual decisions may be taken such as incident detection, validation of the delivery or billing. The
stakeholders require transparency in the whole process. The combination of the Internet of Things
(IoT) and the blockchain paradigms helps in the development of automated and trusted systems.
In this context, ensuring the quality of the IoT data is an absolute requirement for the adoption of
those technologies. In this article, we propose an approach to assess the data quality (DQ) of IoT
data sources using a logistic traceability smart contract developed on top of a blockchain. We select
the quality dimensions relevant to our context, namely accuracy, completeness, consistency and
currentness, with a proposition of their corresponding measurement methods. We also propose a
data quality model specific to the logistic chain domain and a distributed traceability architecture.
The evaluation of the proposal shows the capacity of the proposed method to assess the IoT data
quality and ensure the user agreement on the data qualification rules. The proposed solution opens
new opportunities in the development of automated logistic traceability systems.
**Keywords: IoT; data quality; smart contract; traceability; logistic; sensor; blockchain; supply chain**
**1. Introduction**
In the logistic chain domain, multiple stakeholders need to exchange data about
_shipments transiting from the shipper to the consignee. The data exchange purpose is to give_
visibility to all the stakeholders about the shipments progress in the logistic chain and trace
the path as well as the transport conditions throughout the entire chain.
We refer to the data collected during shipments transit as traceability data, the system
in charge of collecting, saving and sharing those data as traceability system and the whole
process of data collection and processing as the traceability process.
Traditional traceability systems handle traceability data in a central system hosted by
one of the stakeholders, which constitutes a risk on the availability of the traceability data
(single point of failure). The lack of transparency in the qualification process could also
be a source of dispute on the correct application of data handling and qualification rules
agreed by all the traceability process stakeholders.
The advent of blockchain technology and smart contracts help develop new traceability systems. Such systems allow stakeholders to achieve the secure and transparent sharing
of traceability data, using the blockchain secured and distributed ledger. In addition, smart
contracts allow stakeholders to share data handling and decision-making rules, in order to
ensure that the same agreed rules are applied by all the stakeholders.
Increasingly, IoT devices are used to automatically collect field data. Those data are
used both for traceability purpose and to take automatic decisions, such as the creation
of shipment incidents, when one or more of the negotiated shipment transport conditions
-----
_Sensors 2021, 21, 2239_ 2 of 25
are not respected. As a result, the human intervention is limited in the process, as well
as process error probability. To automate the traceability process decision making, new
traceability system architectures have been proposed in the literature combining smart
contracts and IoT (see, e.g., [1–4]).
However, in the existing smart contracts and IoT traceability systems literature, many
of the provided architectures propose to integrate the IoT data directly into the smart
contract (see, e.g., [1–4]). This could lead to unsound decisions taken by the smart contract
based on erroneous data collected and sent directly to the smart contract by the IoT data
sources. To overcome this issue, we propose to introduce an IoT data qualification process
in smart-contracts and IoT-based traceability architectures.
We proposed in [5] to enhance traceability architectures using blockchain, smart
contracts and IoT, combined with a lightweight IoT data qualification process. However,
in this previous work, the proposed qualification process covered only outlier measure
detection, which is only one facet of data quality. Furthermore, we did not compute the
qualification at different levels such as the measure and the sensor level. Moreover, the
IoT data qualification process was centralized at one stakeholder’s site, and there was
no guarantee for the other stakeholders on the correct execution of the agreed IoT data
qualification rules. In addition, the stakeholder in charge of the IoT data qualification
represented a single point of failure of the architecture on the IoT data qualification part.
To overcome the above limitations, the main contributions of this article are threefold:
_(i) The literature review of IoT data qualification highlights that the data quality of a system_
is assessed by means of several dimensions. Considering the logistic chain properties,
the first contribution is to identify the most relevant IoT data qualification dimensions
and provide measurement methods for each of them. (ii) To help the stakeholders to get
an end-to-end visibility of the data quality and to identify the quality issues causes, the
second contribution aims at measuring the data quality at four levels: IoT data events,
IoT data sources, shipments and IoT data sources-shipments associations. (iii) To ensure
the stakeholders agreement on the traceability data, the data qualification rules, and the
decisions taken based on the data, such as the creation of incidents, the third contribution
consists in integrating the data qualification measurement methods in a traceability smart
contract.
The rest of the article is organized as follows. In Section 2, we present the logistic
domain context and its requirements through an example use case. Section 3 highlights the
main research questions addressed in the paper together with their motivations. Section 4
studies the works related to the IoT data quality and the use of the blockchain to assess
this quality. In Section 5, we present the use of the selected IoT data quality dimensions to
measure the data quality. Section 6 presents the architecture of the proposed traceability
solution. The evaluation of our proposed IoT data quality assessment approach is presented
in Section 7. Finally, we conclude in Section 8 and present some future works.
**2. Medical Equipment Cold Chain Use Case**
In this section, we present a business-to-business logistic chain emblematic example.
Because of its specific constraints, the medical equipment cold chain is handled by specific
transport means. We chose this use case for two reasons: (1) the requirement for transport
monitoring; and (2) we worked with an ALIS customer specialized in the production of
medical equipment and we were able to discuss with this customer about their traceability
needs for this specific cold chain context.
Some of the equipment, such as perishable medical diagnostic kits used in blood tests,
needs to be transported under strict conditions with a temperature between a minimum
of +2 and a maximum of +8 _[◦]C. The non-compliance with this temperature interval may_
render the medical diagnostic kits unusable. The stakeholders should be notified of any
temperature non-compliance.
At least three traceability stakeholders are involved in the traceability system of this
medical equipment cold chain: a shipper (at the origin of the transport request), a carrier (in
-----
_Sensors 2021, 21, 2239_ 3 of 25
charge of transport operation) and a consignee (the recipient of the transported equipment).
In our context, we use the term shipment to designate any object entrusted to the carrier by
the shipper, in order to be forwarded to the consignee.
The transport condition data are collected through IoT data sources declared by the
stakeholders and every IoT data source has its own data communication interval. Among
the declared IoT data sources used in our scenario, there is a connected object equipped
with sensors that accompanies the shipments and that is assigned by the shipper.
For visibility and transparency purpose, the stakeholders need to securely share all
the traceability data created manually or collected automatically by the IoT data sources.
The stakeholders need also to be sure that the traceability data processing conforms to the
rules agreed between the stakeholders.
The shipper is responsible for the shipment creation in the traceability system, with all
the data required by the carrier for the good execution of the transport operation, such as
the origin, destination, transport temperature thresholds and IoT data reception interval. In
this scenario, we focus on the management of incidents that could be detected automatically
by the traceability system, based on the data sent by the IoT data sources, such as the
non-compliance with the negotiated transport temperature interval.
The data received from the IoT data sources will be used to automatically create
incidents in the traceability process if necessary. Hence, these data should not be integrated
directly into the system. A data qualification process is required to ensure that the IoT
data quality is good enough to ensure the proper incidents detection. For this purpose, the
stakeholders should have the ability to set the required thresholds for the IoT data quality.
Thus, the data that do not meet the quality thresholds requirements could not be used in
the traceability process.
The data qualification process has many advantages: it not only provides a quality
degree to each shipment related IoT event and a performance measure of its associated data
source but also helps the users to choose the most trustworthy data source and facilitates
the detection of damaged ones in order to repair or replace them.
**3. Research Questions and Motivations**
Based on the above-mentioned use case, we can highlight six main research questions
addressed in this article and their motivations: (1) How accurate are the data? In other
words, do the data reflect the reality of the shipment transport operation? Measuring data
accuracy avoids the use of unreliable data. (2) Are the data complete? Indeed, the existence
of gaps in the collected data may affect the shipment traceability. (3) Are the data consistent?
The consistency issue arises when the collected data assigned to a shipment comes from
several sources with possibly discrepancies leading to incidents. In this case, an agreement
could be defined to tolerate a minimum deviation between the data, for example, a gap
of 0.5 _[◦]C in the temperature may be considered as acceptable. (4) Are the data timely_
valid? That is, are the data compliant with the receiving window agreed between the
stakeholders? The non-respect of this interval may significantly affect the stakeholder’s
visibility and the required transparency of ongoing transport operations.
Each above question reflects a facet (dimension) of the quality process that this paper
addresses and thus the main contribution of this paper is to propose quality measures
for each dimension identified as relevant in our context namely: accuracy, completeness,
consistency and currentness. These quality dimensions are defined in the next section.
In addition, to the above quality dimensions questions, there is a concern about quality
granularity. (5) How can the system provide different levels of quality: data events, IoT data
sources and per shipment performances? This high precision quality monitoring facilitates
the identification at the right time of the data sources that need to be repaired or removed.
Finally, there is a question concerning transparency. (6) How can the data and the
data quality measurement rules be shared securely among the stakeholders to ensure their
agreement on the correct application of these rules? To address this issue, we propose to im
-----
_Sensors 2021, 21, 2239_ 4 of 25
plement the above quality measures into a smart contract, in order to ensure the agreement
of all the stakeholders on the correct application of the proposed quality measures.
**4. Related Works**
Data quality is not a recent research topic. The first data quality studies concerned
databases. Many data quality aspects have been considered such as the accuracy, consistency and reliability to improve the quality of data inputs into databases and handle
databases incompatibility and time critical delivery data [6].
With the advent of the IoT as new data sources, the existing data quality studied
aspects needed to be extended to the specificities of those new data sources. The data
collected from IoT data sources need to be controlled even more due to the limited capacity
of these sources to ensure the security and the quality of their data. The “Never trust user
input” should evolve to “Never trust things input”, as stated by Karkouch et al. [7].
Moreover, the emergence of blockchain opens new opportunities for systems that
involve multiple stakeholders. The logistic chain domain, which involves multiple stakeholders, provides relevant use cases for this technology [8], especially for traceability
purpose [9]. The blockchain promotes the development of smart logistics [10], using smart
contracts.
Before providing a literature review, it is important first to define some terms used in
the domain of data quality and their meaning in the logistic context.
_4.1. Data Quality Definitions_
Data quality dimensions are attributes representing a single aspect of the data quality,
as stated by Richard Y. Wang [11]. In this work, we consider the following data quality
dimensions: accuracy, completeness, consistency and currentness.
The accuracy, as stated by ISO [12], refers to: “the degree to which data has attributes
that correctly represent the true value of the intended attribute of a concept or event in a
specific context of use”. In our context, it is difficult to know if a received measurement
reflects the real shipment situation, especially when the shipment transport operation is
ongoing. However, we can define an accuracy measurement method based on the received
measure and the measure source specifications.
The completeness, according to ISO [12], corresponds to “the degree to which subject
data associated with an entity has values for all expected attributes and related entity
instances in a specific context of use”. In our context, the completeness depicts the fact that
all the expected events have been received by a data source or a shipment according to the
update interval agreed by all the stakeholders.
The consistency, according to ISO [12], refers to “The degree to which data has attributes
that are free from contradiction and are coherent with other data in a specific context of
use”. It is also referred to as concordance in some works [13]. In our context, the consistency
dimension corresponds to the degree of coherence between IoT data events sent by different
IoT data sources and related to the same shipment.
The currentness was defined by ISO [12] as: “The degree to which data has attributes
that are of the right age in a specific context of use”. It is also referred to as timeliness,
currency, freshness, delay or contemporaneous, in some works [13,14]. In our context, an
event is considered of the right age when it is received at the expected time according to
the update interval agreed by the stakeholders and defined in the smart contract.
_4.2. Related Works Study Criteria_
The combination of the blockchain smart contracts and the IoT helps in the development of trusted [15] and automated systems. However, the IoT data quality is a hindrance
to the development and adoption of this new generation of systems.
In this article, we present the works related to the IoT data quality issue according
to three criteria: (C1) the quality dimensions; (C2) the quality levels; and (C3) the use of
blockchain smart contracts for data quality management.
-----
_Sensors 2021, 21, 2239_ 5 of 25
4.2.1. Quality Dimensions (C1)
The IoT data quality issue has been addressed using data quality dimensions. For this
purpose, the traditional data quality dimensions [11] have been used and adapted to the
IoT context needs [13].
The definition of the IoT quality dimensions and their corresponding calculation
methods facilitates their usage and application in the target IoT based systems. Due to
the lack of works on IoT data using quality dimensions in the logistic chain context, we
selected some representative related works from other domains.
Many of the existing works show the interest of using those quality dimensions for
IoT data quality handling. In each work, the authors selected the dimensions relevant to
their domain and defined the corresponding measurement methods for the selected quality
dimensions.
Li et al. [16] defined and measured the currency, availability and validity metrics in a
pervasive environment (IoT context) and the problem of data expiration (data no longer
usable). It is worth noting that, in the traceability context, the data do not expire. It is
important to get all the data for traceability purpose even though the data received late
will have a poor currentness quality index.
Sicari et al. [17] proposed a quality-aware and secured architecture handling: accuracy,
currentness, completeness and other quality dimensions. A framework for determining the
quality of heterogeneous information sources was proposed by Kuemper et al. [18], using
the dimensions of accuracy and consistency.
To ensure a real-time data allocation and data quality in multiple partitions collection
and storage, Kolomvatsos [19] proposed a real time data pre-processing mechanism, using
Fuzzy Logic and handling the accuracy dimension.
In the domain of Ambient Assisted Living (AAL) systems, Kara et al. [20] proposed
a quality evaluation model. Their approach is based on the definition and execution of
quality metrics and the use of fuzzy logic to evaluate the metrics and decide of the data
quality level. In the same precedent domain, Erazo-Garzon et al. [21] defined, measured
and evaluated the quality of data collected from an intelligent pillbox, using seven data
quality dimensions, among them the accuracy, the completeness, the currentness and the
confidentiality.
All the above works use some of or all our required IoT quality dimensions. However, their measurements methods do not meet our needs of dimensions definition and
measurement at different levels: data event, data source and shipment.
4.2.2. Quality Levels (C2)
In the logistic chain context, the stakeholders need to be provided with a full quality
visibility at different levels of the manipulated objects. This is our second criterion (C2).
It is helpful for the data quality management and simplifies the investigation in case of
discrepancy between the stakeholders IoT data sources. Some works proposed data quality
models to handle this issue.
A generic data quality metamodel for data stream management was proposed by
Karkouch et al. [22]; in the evaluation of their work, the authors used the accuracy and
completeness dimensions. There is also the work of Fagúndez et al. [23] on a data quality
model to assess sensors data quality in the health domain, using the dimensions of accuracy,
completeness, freshness and consistency.
The above cited models do not meet our context needs. On the one hand, the data
sources in our context are reused and affected by different shipments in different transport
operations. On the other hand, to meet the criterion (C2) in our proposition, we provide
the stakeholders with a full visibility of the data quality at different object levels, using an
adequate quality model.
-----
_Sensors 2021, 21, 2239_ 6 of 25
4.2.3. Blockchain Smart Contracts for Data Quality Management (C3)
Traceability data need to be shared securely among the stakeholders in order to
ensure their agreement on the data quality and the correct application of the agreed data
calculation methods. This is our third criterion (C3). The following representative works
from the literature propose IoT-blockchain based architectures to handle this issue.
In the domain of crowdsensing platforms, there are many recent works, proposing to
use the blockchain in order to improve the quality of the collected IoT data, such as the
works of Gu et al. [24], Nguyen and Ali [25], Wei et al. [26], Cheng et al. [27], Huang et
al. [28], Zou et al. [29] and Javaid et al. [30]. Their propositions are based essentially on
users’ reviews, reputation and reward mechanisms to incentivize the users to improve the
quality of their provided data. Those mechanisms are not applicable in the logistic chain
context, in which the stakeholders are known and responsible of their provided data.
Casado-Vara [31] proposed an IoT data quality framework based on the use of a
blockchain, in the context of smart homes. The proposed solution is limited to the accuracy
dimension and does not involve multiple stakeholders, each having its own data sources
as is in our context.
In the context of a fish farm, Hang et al. [32] proposed a blockchain based architecture
to ensure agriculture data integrity. Their proposed fish farm architecture includes an
outlier filter, that removes measurements beyond the expected values. This outlier filter is
implemented outside the blockchain, using a Kalman filter algorithm.
Leal et al. [14] proposed a framework for end-to-end traceability and data integrity, in
the domain of pharmaceutical manufacturing. They addressed the problem of temporal and
multi-source variability using probability distribution methods. In our logistic traceability
context, we do not need to estimate sensor measurement data, so we should just report
these data as they are sent by sensors. If some data are missed or out of the expected ranges,
this results in a quality incident on which the involved stakeholders need to agree.
In our proposition, we implement the data quality measurement methods in a blockchain
smart contract in order to ensure a secured sharing and agreement of all the stakeholders
on the correct application of the measurement method and the resulting data quality.
_4.3. Summary of the Related Works Study_
To enhance and secure the IoT data quality in the logistic chain, we propose in this
article a data quality assessment architecture using accuracy, completeness, consistency and
currentness dimensions (C1) in a blockchain smart contract for logistic chain traceability.
The proposed architecture provides the logistic chain stakeholders with data quality visibility at different levels (C2) and guarantees the user agreement on the correct quality rules
application (C3). Besides, the proposed architecture does not only increase the integrated
data quality, but also the stakeholder’s trust and adherence to the resulting automatic
decisions.
Table 1 summarizes the selected related works and how they meet the studied three
criteria.
**Table 1. Related works comparison summary.**
**C3 (Use of Blockchain Smart Contracts**
**C1 (Quality Dimension)** **C2 (Quality Levels)**
**for Data Quality Management)**
IoT
Li et al. [16] Currentness and others Data N/A
Sicari et al. [17] accuracy, currentness, completeness and others Data and stream window N/A
Kuemper et al. [18] Accuracy and consistency Data and data source N/A
Kolomvatsos et al. [19] Accuracy Data N/A
Kara et al. [20] Accuracy, completeness and others Data N/A
Accuracy, completeness, consistency (lack of
Erazo-Garzom et al. [21] Data and data source N/A
measurement method), currentness and others.
IoT Data Quality models
Karkouch et al. [22] Accuracy and completeness (in the evaluation) Data and stream window N/A
Accuracy, completeness, freshness and
Fagúndez et al. [23] Data and stream window N/A
consistency
-----
_Sensors 2021, 21, 2239_ 7 of 25
**Table 1. Cont.**
**C3 (Use of Blockchain Smart Contracts**
**C1 (Quality Dimension)** **C2 (Quality Levels)**
**for Data Quality Management)**
Blockchain and IoT Crowdsensing platforms
Gu et al. [24], Nguyen and Ali [25],
Wei et al. [26], Cheng et al. [27],
Huang et al. [28], Zou et al. [29]
and Javaid et al. [30]
Blockchain, IoT and data qualification
Data quality ensured through reviews,
N/A N/A reputations and rewards mechanisms
implemented in blockchain smart contracts
Accuracy qualified outside the blockchain
Casado-Vara et al. [31] Accuracy and consistency N/A
smart contract and consistency inside it
Outlier’s filtering outside the blockchain
Hang et al. [32] Accuracy (outliers filtering) N/A
smart contract
Accuracy, consistency (multi-source variability) Data qualification outside and inside the
Leal et al. [14] N/A
and currentness (contemporaneous) blockchain smart contract
Accuracy, completeness, consistency and
Our proposition
currentness
Data, data source, shipment
and shipment data source
relationship (equivalent to
Stream window)
Data qualified using quality dimensions
implemented in a blockchain smart
contract
**5. Data Qualification Using Data Quality Dimensions**
In this work, the data qualification refers to the definition of data quality measurement
methods and the application of those methods on every data received and handled by the
smart contract.
We focus on the qualification of traceability IoT data. Because these data are automatically collected and used by the smart contract for the detection of incidents, their
qualification is essential for building reliable and automated traceability system.
Thanks to a data quality study adapted to the logistic chain domain, we identified:
_(i) relevant IoT data quality dimensions; and (ii) their respective measurement methods._
The IoT data quality model purpose is to be implemented in the traceability smart
contract, in order to assess the shipment data quality and consequently improve the incidents
creation process. Among the quality models proposed in the literature, the model by
Karkouch et al. [22] is the closest to our above needs, and we decided to implement and
extend this model for the logistic chain domain.
As depicted in Figure 1, we added the Shipment entity to collect the data quality
at the shipment level with its own IoTQualityDimension. Furthermore, for capturing the
data quality during the association of the IoTDataSource and the Shipment, we added the
_Assignment entity which reflects this temporary relationship._
Furthermore, we highlight in Figure 1 all the model entities and attributes added
for the quality assessment purpose. The main entity of this model is Shipment which
has its own IoTQualityDimension and its own IoTDataSource affected to it through the
_Assignment entity. It is worth noting that the IoTQualityDimension has a weight attribute_
defining the importance of the dimension according to the stakeholders needs.
In our context, we need to distinguish different application levels of each dimension,
for quality visibility at every object level. The quality index resulting from a dimension
application is calculated differently for each dimension related entity in the schema. In
some cases (detailed in the next sections), an IoTQualityDimension is not defined for
some entities of the schema. For example, the completeness dimension is not defined for
_IoTDataEvent and IoTMeasure; it is used only for entities with an update time interval_
constraint such as IoTDataSource and Shipment.
Moreover, we introduce in this model a qualityCon f idenceIndex, in order to provide
users with an overview of the data quality for the main objects manipulated in our traceability system, which are the IoTDataSource, Shipment, Assignment and IoTDataEvent.
-----
_Sensors 2021, 21, 2239_ 8 of 25
**Assignment**
- id: Long
- startAssiTime: Timestamp
- endAssiTime: Timestamp 0..*
- confidenceIndex: Long
- dimensionsQualityIndex: Map<String, Double>
- lastReceivedEvtTimestamp: Long
0..*
1..*
assigned
Shipment **IoTDataSource**
- id: Long - id: String
- pickupTimestamp: Timestamp - name: String related
- deliveryTimestamp: Timestamp - owner: String
- origin: String - startTimestamp 0..*
- destination: String - qualityConfidenceIndex: Double
- mode: String - dimensionsQualityIndex: Map<String, Double>
- carrier: String - measureIntervalInSeconds: Integer
- dataQualityIndexThreshold: Double related
**- qualityConfidenceIndex: Double**
- dataQualityIndexDimensionsThresholds: Map<String, Double>
**- dimensionsQualityIndex: Map<String, Double>**
**- globalDataQualityThreshold: Double** 1..1
**- dataQualityIndexThreshold: Double** 1..1 0..1 0..1 0..* 0..*
**- dataQualityIndexDimensionsThresholds: Map<String, Double>**
**IoTDataEvent**
related related
0..* 1..1 0..1 0..* 0..* - id: Long
- srcId: String
**DataQualityIncident** - timestamp: Timestamp
has shipmentIncident has shipmentCondition
- receptionTimestamp: Timestamp
- id: Long - qualityConfidenceIndex: Double
1..1 0..* - label: String - dimensionsQualityIndex: Map<String, Double>
- creationTime: Timestamp - dataQualityIndexThreshold: Double
ShipmentIncident ShipmentCondition - closingTime: Timestamp- stakholder: String 0..* - dataQualityIndexDimensionsThresholds: Map<String, Double>
- id: Long - id: Long **IoTMeasure**
- label: String - code: String
- creationTime: Timestamp - label: String
- closingTime: Timestamp - min: Long - code: String
- stakholders: List<String> - max: Long - qualityConfidenceIndex: Double
- stakholders: List<String> 0..* 0..* - dimensionsQualityIndex: Map<String, Double> 1..*
**_IoTQualityDimension_**
# code: String; 1..*
# name: String
**IoTMeasureValue**
# weight: Integer
# timeToleranceThreshold: Integer
- code: String
- value: Double
+ caculateDimensionConfidenceIndex(shipment:Shipment): Double - minValue: Double
+ caculateDimensionConfidenceIndex(ioTDataSource:IoTDataSource): Double - maxValue: Double
+ caculateDimensionConfidenceIndex(ioTDataEvent:IoTDataEvent): Double - precision: Double
+ caculateDimensionConfidenceIndex(ioTMeasure:IoTMeasure): Double
**IoTQualityAccuracy** **IoTQualityCompleteness** **IoTQualityConsistency** **IoTQualityCurrentness**
**Figure 1. IoT Data Quality Entity class diagram.**
The calculation of the quality index takes into account the weight W of dimensions
fixed by the users for IoTDataSource and Shipment. We calculate this quality index
for the IoTDataSource and Assignment as an average of their n IoTDataEvents and m
_IoTQualityDimensions:_
|Col1|Col2|Col3|Col4|
|---|---|---|---|
|IoTQualityCompleteness||IoTQualityConsistency||
_qualityCon f idenceIndex =_
_m_ _n_
_j∑=1(Wj ∗_ _i∑=1_ _dimensionQualityIndexjIoTDataEventi )_
(1)
_m_
∑ _Wj_
_j=1_
For the Shipment quality index calculation, we use the quality indexes of its related
_Assignment objects. Regarding the IoTDataEvent, we use the average quality of its related_
-----
_Sensors 2021, 21, 2239_ 9 of 25
_IoTMeasures. The methods used to calculate the IoTQualityDimensions are detailed in the_
next sections.
The quality thresholds are set by the stakeholders to define the minimum accepted
quality index. Values that do not respect this quality will be stored for traceability purpose
but will not be used for dynamic incident detection.
To monitor the compliance of the received data according to both the quality threshold
and the Shipment transport conditions defined in the smart contract, we added in the model,
respectively, the entities: DataQualityIncident and ShipmentIncident. DataQualityIncident
results from a non-compliance with the agreed quality thresholds and ShipmentIncident
results from a non-compliance with the agreed business transport conditions. For example,
consider a IoTDataSource with an interval of possible values from 0 to 50 _[◦]C, and monitor-_
ing a Shipment with business transport conditions of 2–8 _[◦]C. If this IoTDataSource sends_
a temperature value of 100[◦]C, this value is considered as non-compliant with the quality
thresholds and generates a DataQualityIncident. However, if the sent value is 20 _[◦]C, it_
is considered as non-compliant with the business transport conditions and generates a
_ShipmentIncident._
In the next subsections, we detail how the dimensions are used to calculate the quality
indexes for the different object levels.
_5.1. Accuracy_
The accuracy measurement method is based on the IoTDataSource specifications
(sensor measure precision value and sensor minimum and maximum measurable values).
Using this method, we can ensure that the received measurement is a possible normal
value that can be sent by the concerned IoTDataSource.
Therefore, the received measurement could be used by the traceability smart contract,
for example to create an incident, if the received measurement is out of the ranges fixed by
the shipper for this specific measurement.
In the following subsection, we detail the accuracy calculation method depending on
the object level.
Accuracy Levels
We identify five accuracy levels: the IoTMeasureValue accuracy AccMsrVal,
the IoTMeasure accuracy AccMsr, the IoTDataEvent accuracy AccEvt, the IoTDataSource
accuracy AccSrc and the Shipment accuracy AccShp.
The IoTMeasureValue accuracy as indicated by its name is related to only one value
of the IoTMeasure. It is used to indicate if a value of the IoTMeasure is in the range of relevant and acceptable values of this specific IoTMeasureValue, based on the IoTDataSource
specifications. For example, consider an IoTMeasureValue m, with precision p, and FThmin
and FThmax are, respectively, the minimum and the maximum possible values given by the
_IoTDataSource manufacturer._
We calculate the IoTMeasureValue accuracy AccMsrVal using the following formula:
1 If (m − _p) ≥_ _FThmin and (m + p) ≤_ _FThmax_
_m−FThp_ _min_ if (m − _p) < FThmin and m ≥_ _FThmin_
_FThmaxp_ _−m_ if (m + p) > FThmax and m ≤ _FThmax_
0 _otherwise_
_AccMsrVal =_
(2)
The IoTMeasure is composed of n IoTMeasureValues, and consequently we calculate
the IoTMeasure accuracy AccMsr as an average of all its IoTMeasureValues accuracies:
_AccMsr =_
_n_
∑ _AccMsrVali_
_i=1_
(3)
_n_
-----
_Sensors 2021, 21, 2239_ 10 of 25
The IoTDataEvent accuracy AccEvt corresponds to an overview of the accuracies
of all its related IoTMeasures. This is useful in our context where the IoTDataEvent is
considered as a coherent set of IoTMeasures. If this is not the case, the accuracy calculated at
the IoTMeasure level can directly be used, and the IoTDataEvent accuracy can be ignored.
However, for an IoTDataEvent with n related IoTMeasures, the IoTDataEvent accuracy
corresponds to the average of all the IoTDataEvent related IoTMeasures accuracies:
_AccEvt =_
_n_
∑ _AccMsri_
_i=1_
(4)
_n_
The IoTDataSource accuracy AccSrc gives an overview of all the IoTDataSource related IoTDataEvents accuracies, it is related to the historic of IoTDataEvents received from
the IoTDataSource. In our context, we consider that it is important to take in consideration
this historic of IoTDataEvents in the calculation of IoTDataSource accuracy, because it
indicates the reliability of the IoTDataSource since it has been deployed and used in our
traceability system.
If the users are interested only in the IoTDataSource IoTMeasures accuracies, the
accuracy calculated at the IoTMeasure level could be reused at the IoTDataSource level
in order to give them an IoTDataSource accuracy per IoTMeasure. The accuracy of an
_IoTDataSource corresponds to the accuracy average of all its related IoTDataEvents:_
_AccSrc =_
_n_
∑ _AccEvti_
_i=1_
(5)
_n_
Finally, the Shipment level accuracy emphasizes all the Shipment related IoTDataSource
accuracies for the specific time period in which the IoTDataSource is assigned to the
_Shipment. Every Shipment is considered as an independent transport operation that should_
have its own accuracy value.
For a Shipment with n Assignments to IoTDataSources, the accuracy AccShp corresponds to the average of all the Shipment-IoTDataSource Assignments. For each Assignment
accuracy AccAssigni, the number of IoTDataEvents nEvtAssign to be considered in the accuracy calculation, corresponds to the number of IoTDataEvents sent by the IoTDataSource
for this specific Shipment Assignment relationship:
_nEvtAssign_
∑ _AccEvtj_
_j=1_
(6)
_nEvtAssign_
_AccShp =_
_5.2. Completeness_
_n_
∑ _AccAssigni_
_i=1_
such as AccAssigni =
_n_
The completeness measurement method calculates the gap in the data reception for
a specific object. It concerns the levels of the IoTDataSource, the Assignment and the
_Shipment._
5.2.1. Completeness Levels
At the IoTDataSource level, the completeness is calculated based on the source
_startTimestamp, the source measure interval I, the number of received IoTDataEvents n_
from the IoTDataSource and the reception timestamp of the last IoTDataEvent lastTimestamp,
related to the IoTDataSource:
_ComSrc =_
�
1 If n ≥ _[lastTimestamp][−]I[startTimestamp]_
(7)
_n∗I_ otherwise
_lastTimestamp−startTimestamp_
-----
_Sensors 2021, 21, 2239_ 11 of 25
The Assignment completeness ComAssign means that all the expected IoTDataEvents
of the assigned IoTDataSource Src, have been received by the Shipment during the IoTData
_Source and Shipment association time period enshrined in the smart contract._
Consequently, for the Shipment, the IoTDataEvent frequency is at least one IoTData
_Event per IoT update time interval I defined in the smart contract. The ComAssign highlights_
for the stakeholders the capacity of each IoTDataSource to send all the expected data during
its association with a Shipment. This helps the stakeholders to decide on the reusability
of the IoTDataSource for further Shipments in the case of a good completeness value or,
otherwise, to take over the IoTDataSource in order to identify the completeness source
problem.
The ComAssign evolves during the Shipment and the IoTDataSource association time
period, and it is recalculated for every new IoTDataEvent reception at the timestamp
_evtTimestamp, based on the current number of received IoTDataEvents n, the Shipment up-_
date interval _I,_ the _IoTDataSource-Shipment_ _Assignment_ _startAssignTime_ and
_endAssignTime timestamps._
1 If n ≥ _[evtTimestamp][−]I[startAssignTime]_
and evtTimestamp ]startAssignTime, endAssignTime]
_∈_
Or
_n ≥_ _[endAssignTime][−]I[startAssignTime]_ (8)
and evtTimestamp > endAssignTime
_endAssignTimen−∗startAssignTimeI_ If evtTimestamp > endAssignTime
0 _otherwise_
At the Shipment level, the completeness ComShp gives an idea of the completeness
trend of all the Shipment related IoTDataSources. It is calculated as a ComAssign average of
the nAssign IoTDataSources assigned to the Shipment:
_ComAssign =_
_ComShp =_
_nAssign_
∑ _ComAssigni_
_i=1_
(9)
_nAssign_
5.2.2. Completeness Incidents
The completeness problem reflects the missing IoTDataEvents. Many reasons could
be at the origin of missing IoTDataEvents: network errors, synchronization problems or
device malfunctions [33]. If it is not handled, missing data seriously affect the reliability of
the data collected through the IoTDataSource.
We propose to generate a completeness incident, if the completeness index of the
object fall below the completeness threshold fixed by the stakeholders. The update-missing
incident created by the smart contract will also remain there in order to trace the history of
data quality problems related to the event IoTDataSource.
_5.3. Consistency_
It is important to calculate the coherence degree between IoTDataEvents and to alert
the stakeholders in the case of incoherence detection. The stakeholders should take a
corrective action, such as identifying and removing failing IoTDataSource, adapting new
threshold values, etc.
The main IoTDataSource in this work is the shipper shipment connected object. However, other IoTDataSources could be added by any of the Shipment transport stakeholders.
When two or more IoTDataSources assigned to the Shipment monitor the same transport
conditions, we calculate the consistency of those IoTDataSources, by comparing their
_IoTMeasures._
-----
_Sensors 2021, 21, 2239_ 12 of 25
The IoTMeasures comparison takes into account two tolerance thresholds: the time
tolerance threshold Ttth and the measure tolerance threshold Mtth. Those two thresholds should be defined at the Shipment creation for every IoTDataSource assigned to the
_Shipment, of course through a mutual agreement between the stakeholders in charge of_
those IoTDataSources.
Consistency Levels
The consistency dimension concerns the levels of IoTDataEvent and Shipment. When
an IoTDataEvent Evti is received from IoTDataSource Srci at a timestamp Rti, and contains a list Msri of IoTMeasures, we check if there are other IoTDataEvents related to
the Shipment and sent by other IoTDataSources, verifying that for each IoTDataEvent
_Evtj, received from IoTDataSource Srcj at the timestamp Rtj and containing a list Msrj of_
_IoTMeasures:_
Srci ̸= Srcj
_|Rti −_ _Rtj| ≤_ _Ttth_ (10)
Msri ∩ _Msrj ̸= ∅_
where IoTMeasures are compared using their codes (see Figure 1).
If there is only one IoTDataSource for the Shipment, or there are no IoTDataEvents
verifying the above conditions, then there is no consistency calculation to do. Otherwise,
the IoTDataEvent consistency is calculated using the following method:
1 _∀m ∈_ _Msri ∩_ _Msrj, |Valmi −_ _Valmj_ _| ≤_ _Mtth_ _Valmi is the value of m in Msri, and_
_Valmj is the value of m in Msrj_
_NbConNbEvtEvti_ _NbConEvti is the number of events_
concordant with Evti, and NbEvt
is the total number of events
verifying the above consistency
conditions
_ConEvti =_
(11)
The Shipment consistency ConShp gives an overview of the Shipment data consistency
between all the IoTDataSources related to the Shipment and monitoring the same transport
conditions. It is calculated as an average of the Shipment related Assignments consistency:
_ConAssign._
_nEvtAssign_
∑ _ConEvtj_
_j=1_
(12)
_nEvtAssign_
_ConShp =_
_5.4. Currentness_
_n_
∑ _ConAssigni_
_i=1_
such as ConAssigni =
_n_
In the logistic traceability context, the currentness dimension may not be critical.
Indeed, the most important is to detect incidents, even though the data are received late.
However, currentness may reveal incidents concerning data acquisition. Thus, the stakeholders define the Shipment currentness threshold according to the use case.
5.4.1. Currentness Levels
We consider the following currentness levels: IoTDataEvent, IoTDataSource and
_Shipment._
For an IoTDataEvent Evti, the currentness CurEvti is calculated based on the previous
_IoTDataEvent reception timestamp ti−1, the update interval defined in the smart contract_
-----
_Sensors 2021, 21, 2239_ 13 of 25
_I, the expected next IoTDataEvent timestamp ti+1 which is equal to ti−1 + 2 ∗_ _I and the_
current IoTDataEvent reception timestamp ti.
_CurEvti =_
�
1 − _[|][(][t][i][−][1][+]I_ _[I][)][−][t][i][|]_ If ti ∈]ti−1, ti+1[ (13)
0 _otherwise_
For the Shipments, the interval I is a shipper requirement that should be met through
the sending of an IoTDataEvent to the smart contract, every time that this interval has
elapsed. Consequently, the currentness indicates not only the quality of the data but also
the meet degree of one of the more important shipper requirements defined in the smart
contract, the Shipment update interval I.
Furthermore, the CurEvti at the IoTDataSource level is calculated using the same
above method, but it is worth noting that the IoTDataSource has its own update interval
that could be different from the Shipment update interval.
Regarding the IoTDataSource, the currentness corresponds to the degree to which the
_IoTDataSource has met the update interval time requirement, in the entire history of its_
related IoTDataEvents, including the last received IoTDataEvent.
The currentness dimension helps the users in the choice of the IoTDataSources to be
assigned to the Shipment, users will always choose the IoTDataSource with the highest
currentness among the available IoTDataSources. The IoTDataSource currentness CurSrc
is calculated as the average of all the IoTDataSource related IoTDataEvents:
_CurSrc =_
_n_
∑ _CurEvti_
_i=1_
(14)
_n_
From the Shipment perspective, the currentness indicates the degree to which the
_shipper update time interval requirement has been met for the Shipment by all its related_
_IoTDataSources, during the Shipment-IoTDataSource association time period. To measure_
the currentness performance of the Shipment-IoTDataSource association, the currentness
calculated for this association CurAssign is saved in the Assignment object.
The CurAssign is useful when the Shipment stakeholders need to investigate a low
_Shipment currentness, as it helps to identify the Shipment related IoTDataSource(s) respon-_
sible(s) of the low currentness value. The Shipment currentness CurShp corresponds to the
average CurAssigni of all its n related Assignment objects. The CurAssign is calculated as a
_CurEvtj average of the nEvtAssign IoTDataEvents received from the IoTDataSource for the_
_Shipment, during their Assignment association:_
_nEvtAssign_
∑ _CurEvtj_
_j=1_
(15)
_nEvtAssign_
_CurShp =_
_n_
∑ _CurAssigni_
_i=1_
such as CurAssigni =
_n_
5.4.2. Currentness Incidents
There are two currentness control points, the reception of the IoTDataEvent by the
stakeholder IS (shipper IS, carrier IS or consignee IS) and the reception of the IoTDataEvent
by the smart contract. In case of non-reception of the IoTDataEvent by the stakeholder IS,
this leads to a missing update on the smart contract side.
The IoTDataSource is configured to send an IoTDataEvent every n seconds. If this
interval has elapsed and no new IoTDataEvent has been received from the IoTDataSource,
the situation is considered as a missing update problem.
The missing update is not critical if the IoT update interval Isc of the smart contract
is larger than n seconds, because the smart contract generally does not wait for a new
_IoTDataEvent as long as this update interval does not expire._
In contrast, if the update interval is equal to n seconds, the stakeholder IS notifies
the smart contract in the case of missing data. Once notified, the smart contract assigns
-----
_Sensors 2021, 21, 2239_ 14 of 25
a missing update related incident to the IoTDataSource owner. The origins of this kind
of incidents are multiple; for example, the IoTDataSource is not able to connect to the IoT
network, the IoTDataSource has internal problem or an IoT cloud data platform problem.
**6. The Distributed Architecture of the Traceability System**
This section presents the architecture of the proposed traceability solution and its
main components: the blockchain smart contract and the IoT data sources.
To respond to the identified criteria for secured traceability data, data qualification
and transparency, we propose a distributed, secured and trusted architecture, based on the
use of blockchain smart contracts, as depicted in Figure 2. The main components of this
architecture are a smart contract shared by all the stakeholders and IoT data sources. The
arrows in this figure indicate data transmission directions.
Shipper Smart contract Consignee
Node Node
Carrier Node
Shipper IS Consignee IS
Carrier IS
Shipper Consignee
Carrier
Delivered
shipments
To be transported
shipments On going
transport
shipments
Shipment Connected Object
Shipment Connected Object
Shipment Connected Object
Shipper Consignee
Factories Factories
IoT Cloud
Data Platform
Shipper (IoTCDP) Carrier Consignee
Warehouses Warehouses Warehouses
LPWAN
Gateway
**Figure 2. Distributed architecture of the traceability system.**
For the smart contract component, we chose to work on a Hyperledger Fabric
blockchain [34]. It is a permissioned blockchain that presents many advantages in comparison to the other blockchains, among them: a node architecture based on the notion of
organization to establish a trust model more adapted to the enterprise context, the support
of the Go, Javascript and Java languages for writing smart contracts and a parameterized
consensus protocol [5].
The smart contract is installed on the top of a blockchain involving all the stakeholders.
This smart contract holds all the rules about the collected traceability data management,
-----
_Sensors 2021, 21, 2239_ 15 of 25
the incident management and the IoT data qualification. Those rules have been validated
by all the traceability system stakeholders.
IoT data sources are assigned to each shipment. They are responsible of the field data
collection about the shipment transport conditions.
It is worth noting that the three stakeholders depicted in Figure 2 are given as examples.
As many stakeholders as needed may be added to this architecture. The addition of
stakeholders is enabled by the underlying Hyperledger Fabric- based architecture [34].
In addition, the maximum number of stakeholders in the context of logistic chain is limited.
For example, in our use case, this number is of the order of tens stakeholders.
The stakeholders to be added to this architecture are those who need to participate
in the traceability process. They are added before the creation of any shipment transport
operation in which they will be involved.
_6.1. The Smart Contract_
Smart contracts are “trusted distributed applications” [34]. They are secured by the
underlying blockchain and the peers consensus mechanism. In the transport traceability
context, we need a distributed and secured application to share traceability data among all
the traceability process stakeholders and ensure their agreement on the shared data quality
and the incidents created based on this data.
We proposed in [5] to implement a lightweight IoT data qualification application and
a traceability smart contract handling all the shipment transport operation process. The
implemented smart contract allowed the stakeholders to define all the transport conditions
terms, update the transport status and transport related milestones status, integrating
IoT data about the shipment transport operation progress and creating both manual or
automatic transport related incidents.
The contractual constraints, negotiated between stakeholders, are enshrined in the
smart contract, and should be respected by all the stakeholders. Any gap between those
constraints and the data provided by a stakeholder results in a non-compliance incident
created automatically by the smart contract. The contractual constraints are communicated
to the smart contract by the shipper system at the shipment creation time.
In this article, we extend the traceability proposal presented in [5] to overcome two
important limitations. (i) The IoT data qualification is centralized at the shipper side, and
there is a lack of guarantees for the other stakeholders on the good execution of the agreed
IoT data quality calculation rules. (ii) The lightweight IoT data qualification module is
limited to data outlier’s detection.
Therefore, we propose in this work to enhance the IoT data application through the
implementation of the quality model presented in Section 5, into the traceability smart
contract. This allows ensuring the stakeholders agreement on the correct application of the
data qualification rules. The data qualification module is also improved by the integration
of the accuracy, completeness, consistency and currentness dimensions.
_IoTDataEvents that do not conform to the defined IoT quality model constraints_
generate DataQualityIncident visible by all the stakeholders. They are not discarded but
saved in the blockchain for audit purpose.
As examples of decisions taken automatically by the smart contract based on the
received events, Table 2 shows some temperature events values received by the smart
contract, their Quality Indexes (QI) and their corresponding decisions. For these examples,
we consider multiple IoT data sources with manufacturer temperature specifications interval of [0 _[◦]C, 50_ _[◦]C]. Those data sources are assigned to a shipment with a temperature_
conditions transport interval of [2 _[◦]C, 8_ _[◦]C]. The quality dimensions’ weights are set to 4_
for accuracy, 4 for consistency and 1 for Currentness. If the event QI is below the quality
index threshold (0.7) a DataQualityIncident is generated for the event. Consequently, the
event QI is calculated as follows:
_EventQI =_ [4][ ∗] [(][AccuracyQI][) +][ 4][ ∗] [(][ConsistencyQI][) +][ 1][ ∗] [(][CurrentnessQI][)] (16)
9
-----
_Sensors 2021, 21, 2239_ 16 of 25
where 9 is the sum of dimensions’ weights (4 + 4 + 1).
**Table 2. Smart contract decisions examples.**
**Received Temperature**
**Accuracy QI** **Consistency QI** **Currentness QI** **Event QI** **Smart Contract Decision**
**Events Values**
10 _[◦]C_ 1 1 1 1 Create a ShipmentIncident
_−20_ _[◦]C_ 0 1 1 0.55 Create an accuracy DataQualityIncident
Create an accuracy DataQualityIncident, and
Many times −20 _[◦]C from the_ 0 1 1 0.55 finally a completeness DataQualityIncident at
same source
the Source-Shipment Assignment level
Create a consistency and accuracy
_DataQualityIncident from Event1 and a_
_ShipmentIncident from Event2_
Event1 of −20 _[◦]C from a source 1_
and Event2 of 10 _[◦]C from a_
source 2
0 for Event1 0.33 for Event1
and 1 for 0.5 for both 1 for both and 0.77 for
Event2 Event2
10 _[◦]C received late from_ 1 1 0 0.88 Create a ShipmentIncident
one source
_6.2. The IoT Data Sources_
In the proposed traceability architecture, the IoT data could be received from many
IoT data sources. Each stakeholder could decide to assign an IoT data source that it owns
to a shipment in which it has a stakeholder role, at any time during the shipment progress in
the logistic chain. The only condition to do so is that the IoT data source and the shipment
have already been created in the smart contract.
The assignment of an IoT data source to a shipment is for a limited period. Every data
source assigned to a shipment sends IoT data about the shipment transport conditions at a
fixed time interval defined in the shipment smart contract instance.
If a data-related incident is detected by the smart contract, it is automatically affected
to the IoT data source owner declared in the smart contract. The smart contract has
a detailed description of the IoT data source specifications collected at the data source
creation in the smart contract. This is a requirement for the correct application of the data
quality measures.
The shipper in our context has a principal IoT data source which is the shipment
connected object accompanying the shipment. The role of this object is to collect data about
the shipment transport conditions, throughout the transport operation.
To send the collected data to the shipper IS (Information System), the connected object
uses an LPWAN (Low Power Aera Network) network Gateway, which transmits the
received messages to the IoT Cloud Data Platform (IoTCDP) before their reception in the
_shipper IS._
The shipper IS sends the received messages to the shipper node including the connected
object id of the messages. This connected object id is used by the smart contract to link the
received IoT messages to the right shipment in the smart contract. In this context, the data
are pushed by the IoT object. The pull/push of data from/to the connected object is out
of the scope of our work. The shipment connected object collects data about the shipment
pickup, transport and delivery conditions.
Each stakeholder could declare other IoT data sources, such as IoT data sources related
to factories, warehouses, transport vehicles, etc. In general, every data source that can
collect and send automatically measurements about the shipments could be declared by
the stakeholder as an IoT data source. Moreover, all IoT data sources, except the shipment
connected object, help to collect data about the shipment conditions in a specific segment of
the transport operation. Only the shipment connected object that accompanies the shipment
continues to collect data about the shipment transport conditions during the whole transport
operation.
**7. Evaluation**
The objectives of this section are: (i) to evaluate the proposed quality measures; (ii) to
evaluate the impact of the IoT data quality module on the number of created incidents; and
-----
_Sensors 2021, 21, 2239_ 17 of 25
_(iii) to evaluate the impact of the IoT data quality module on the IoT data event insertion_
time in the blockchain.
We evaluated our proposed quality measures to measure their pertinence and performance. We also monitored the number of quality incidents created to highlight the impact
of the quality module. The number of shipment incidents was also monitored to emphasize
the impact of the quality module on the business decisions.
The IoT data event insertion time in the blockchain was also measured in our tests,
firstly with the quality module activated and then with the quality module inactivated, in
order to evaluate the impact of our proposed quality module on the data event insertion
time and ensure the final users that this time is acceptable while ensuring the quality.
_7.1. Smart Contract Architecture_
For the implementation purpose, we used the same architecture used in our previous
work on the traceability using smart contracts and IoT [5]. It is an architecture based
on the use of Hyperledger Fabric as the blockchain implementation, with three peers
(stakeholders): a shipper, a carrier and a consignee. On the top of this blockchain, we
implemented our traceability and IoT data qualification smart contract.
The smart contract used in this evaluation was developed on the top of a Hyperledger
Fabric blockchain, using the Fabric Java Framework. We used in this evaluation a Virtual
Machine (VM) with the characteristics depicted in Table 3.
**Table 3. Test VM characteristics.**
**Characteristic** **Details**
OS Ubuntu 18.04.4 desktop amd64
CPU 4 CPU Intel(R) Core™i7-8565U
RAM 8 G
Virtual Disk 50 G
Furthermore, we set the Hyperledger Fabric block creation timeout to 1 s and the
maximum number of transactions per block to 15. This means that, after the reception
of a new transaction, the system will trigger the block creation either after a time wait of
1 s or after a total number of 15 new transactions is reached. In addition, we used in this
evaluation the Raft consensus algorithm, with a unique ordering service node [5].
In the existing traceability smart contract [5], we added many new methods such
as createDataSource and assignDataSource. The createDataSource method inserts the data
source given as input in the blockchain. The assignDataSouce method assigns an existing
IoT data source to an existing shipment, using their IDs. Based on the quality measures
proposed in this article, we updated the addIoTEvent method with the following new
functionalities: (i) calculate the event quality measures; and (ii) update the IoT data source
and the quality measures of the shipments related to this IoT data source.
_7.2. Evaluation Experimental Choices_
Due to a lack of real data to evaluate the proposed architecture in our use case, we
chose to simulate our use case data with a well-known dataset in the IoT domain. The
Intel Berkeley dataset is a collection of sensor data, collected by Intel research team in the
Intel Berkeley Research lab, between 28 February and 5 April 2004 [35]. An example of the
dataset content is depicted in Table 4.
-----
_Sensors 2021, 21, 2239_ 18 of 25
**Table 4. Samples of the Intel Berkeley dataset.**
**Date** **Time** **Event ID** **Sensor ID** **Temperature** **Humidity** **Light** **Voltage**
12 March 2004 16:29:04.084098 39302 1 21.8308 43.5855 165.6 2.53812
14 March 2004 15:45:11.669786 44974 2 26.9464 41.814 264.96 2.54901
19 March 2004 19:01:21.094445 59766 3 21.9092 45.1103 39.56 2.44412
. . . ... ... . . . . . . . . . . . . . . .
To adapt this dataset to our context, we considered every sensor as an IoT data source.
This gives us 54 data sources to be handled. For the shipments, we used every 24 h of sensor
data collection as a shipment, which results in 2052 shipments (54 sensors multiplied by 38,
the number of data collection days), for the whole dataset.
Furthermore, we considered only the temperature measures in this evaluation because
it is the main measure for our use case, but the module could be used to handle any other
measure type.
We began the evaluation phase by defining the user’s quality thresholds requirements
for all the data sources and shipments. We used the same threshold for the data sources,
the shipments and the four quality dimensions. We made a series of tests by varying the
defined threshold, going from 0 (no quality constraints) to 1 (strict quality), to show the
impact of those thresholds on the number of created quality and shipments incidents.
In Table 5, we establish a classification of data quality indexes for our dimensions
and objects. This classification helps in the presentation and the analysis of the evaluation
results.
**Table 5. Quality indexes and thresholds classification.**
**Data Quality Index and Threshold Interval** **Label** **Code**
[0, 0.5) Poor quality P
[0.5, 0.7) Low quality L
[0.7, 0.9) Good quality G
[0.9, 1] High quality H
We chose the following weights for the quality dimensions based on their importance
for the use case in the context of the medical equipment cold chain: a weight of 4 for the
accuracy, the completeness, and the consistency, which are the most important for our
users, and a weight of 1 for the currentness, which is not as critical as the other dimensions,
as explained in Section 5.
For the shipment incidents, we chose an accepted temperature interval of 20 to 25 _[◦]C_
based on the work of Hui et al. [36]. Beyond this temperature interval, if the received event
quality is compliant with the shipment quality threshold, this event results in a shipment
incident created for all the shipments that have an active assignment relationship with the
event data source.
There was no information in the dataset about the sensor’s precision value. Consequently, we chose to set this value to 0.5 _[◦]C, which is a recurrent value in the temperature_
sensors.
In the following evaluation results, we did not take into account the sensor 5 from
which we did not see any event. We also ignored some other events with the sensor IDs 55,
56 and 58, because in the dataset reference the number of sensors was only 54, and events
coming from the same sensor with the same event number (113,474 events in the dataset).
There were also 355 events in the dataset that we could not parse correctly due to their
data presentation errors and 526 incomplete lines, from which we could not get all the
event required data. This results in a total of 2,199,327 events integrated correctly in our
quality tests, from a total of 2,313,682 events present in the dataset.
We used the event timestamp in the dataset as an event reception timestamp in this
evaluation. Moreover, we used this timestamp to order and identify the events, for shipment
-----
_Sensors 2021, 21, 2239_ 19 of 25
incident creation and closing purpose. The results of this choice were 10,299 duplicated
events, because they had the same timestamp as previously received events from the same
sensor.
Furthermore, we use the quality threshold to define the stakeholder’s requirement for
the quality indexes of events to be integrated in the data source or sent to shipments. All the
events with a quality index below the defined quality threshold value results in a quality
incident and are not used to create shipment incidents in case of non-compliance with the
agreed transport conditions. If the quality incident is detected by the data source, it will
not send the event to its related shipments.
_7.3. Results Concerning the Accuracy, Completeness and Currentness Dimensions_
Firstly, regarding the accuracy, the sensors used to collect the Intel Berkeley dataset, a
valid temperature value should be in the range of 0–50 _[◦]C according to [37], otherwise we_
consider this temperature as inaccurate.
Regarding the completeness, we used the following parameters: the update interval
of 31 s, the maximum timestamp among the already integrated events timestamps, the
start IoT data source and the shipment start timestamp. We set the IoT data source start
timestamp at 28 February 2004 at 00:00:00 am, and, for the shipment, the start timestamp is
the shipment date and the start time set at 00:00:00 am and the end at 11:59:59 pm.
Concerning the currentness, we used the measure interval of 31 s given for the dataset.
We used this same update interval for the data sources and the shipments. In our tests, we
did not consider the difference that could exist between the event reception timestamp
and the event production timestamp. This difference could affect the test and need to be
addressed in future works.
Table 6 shows the classification of quality results obtained for the sensors (data
sources), regarding the different quality dimensions defined in this work and using multiple quality threshold values. Those results show that in the 53 retained sensors: 42 have a
good accuracy, 29 have a poor completeness and 29 have a lower currentness.
**Table 6. Sources quality evaluation results.**
**Quality Threshold** **Accuracy** **Completeness** **Currentness** **Quality Index**
0, 0.5, 0.7, 0.9 and 1 0P 1L 43G 9H 29P 22L 2G 0H 1P 29L 20G 3H 0P 38L 15G 0H
Regarding the global sensor quality index, most sensors (38) have a low-quality index.
If the quality threshold is set to a good quality value (e.g., 0.7), only 15 sensors are usable,
and, in the case of threshold of high quality (e.g., 0.9), there is no usable sensor in this
dataset.
Thanks to the quality module, all the events with a quality incident problem are not
integrated into the shipments assigned to the event data source, and this keeps the shipment
events quality at the level fixed and agreed by all the stakeholders. For example, in the
case of Sensor 45, when we set the quality threshold at 1, 9% of the events received from
this sensor have not been integrated into the source related shipments, due to their quality
problems.
In Table 7, we can clearly see the impact of the threshold choice on the percentage
of quality incidents. This percentage represents the events that do not respect the agreed
quality thresholds. The events are filtered at the data source level according to the selected
quality threshold value.
-----
_Sensors 2021, 21, 2239_ 20 of 25
**Table 7. Quality and shipments incidents results according to the quality threshold.**
**Shipments Quality** **Percentage of Quality** **Percentage of Shipments**
**Threshold** **Incidents** **Incidents**
0 0 0.21
0.5 25 0.4
0.7, 0.9 and 1 21 0.3
Consequently, the percentage of quality incidents drops from around 25% of the total
received events for a threshold at 0.5 to around 21% when the quality threshold was greater
or equal to 0.7. The percentage of shipment incidents evolution is not linear due to the
_shipments number evolution depending on the selected quality threshold, as depicted in_
Table 8.
**Table 8. Shipments events number evolution.**
**Shipments Quality** **Number of Shipments** **Number of Shipments with**
**Threshold** **Without Any Event** **at Least One Event**
0 421 1631
0.5, 0.7, 0.9 and 1 821 1231
Regarding the shipments quality results, it is important to note that there were 421
_shipments for which we did not receive any event, no matter what the quality threshold_
value was. This number increases to 821 shipments, when we set the quality threshold at
0.5, 0.7, 0.9 or 1, as depicted in Table 8. Consequently, we did not consider those shipments
in the following shipment quality results, because all our quality dimension calculations are
based on the events values and timestamps.
Table 9 shows that the percentage of shipments with a high accuracy level increase as
the shipments quality thresholds increases, and this is the same for the currentness. The
percentage of events with a poor completeness index increases due to events blocked by
the quality threshold at the data source level.
**Table 9. Shipments quality evaluation results.**
**Quality Threshold** **Accuracy (in %)** **Completeness (in %)** **Currentness (in %)** **Quality Index (in %)**
0 26P 1L 1G 72H 48P 29L 19G 4H 18P 30L 40G 13H 27P 16L 44G 13H
0.5, 0.7, 0.9 and 1 0P 0L 0G 100H 64P 19L 17G 1H 12P 32L 41G 15H 2P 47L 42G 10H
The shipment quality index also is improved by the quality threshold increase; for
example, we went from 27% of poor data quality shipments when the quality threshold was
at 0 to only 2%, when the quality threshold was up to 0.5.
_7.4. Results Concerning the Consistency Dimension_
For the consistency evaluation, we selected four groups of sensors placed in proximity
zones, as depicted in Figure 3: {1, 2, 3}, {11, 12, 13}, {15, 16, 17} and {49, 50, 51}. For each
group, we linked each sensor to all its related sensors shipments in the same sensors group.
-----
_Sensors 2021, 21, 2239_ 21 of 25
**Figure 3. Intel Berkeley sensors arrangement diagram.**
The total number of shipments related to the selected groups was 456 (12 sensors
multiplied by 38 data collection days). There were 84 shipments related to those groups, for
which we did not receive any event from the sensors, whatever the quality threshold value.
This number increases to 171 shipments when we set the quality threshold at 0.5, 0.7, 0.9 or
1, due to the events quality filtering at the data source level.
Furthermore, we set in this evaluation the tolerance time interval to 31 s and the
consistency tolerance temperature to 0.5 _[◦]C. This means that two events are considered as_
eligible to the consistency test only when their timestamps difference is lower than 31 a,
and they are considered as concordant if their reported temperatures difference is lower
than 0.5 _[◦]C._
Table 10 summarizes the consistency evaluation results for the selected sensors groups.
The group {1, 2, 3} has at least 76% of its shipments with a high consistency index. Those
results show that the events reported by the group {1, 2, 3} were more concordant than
those reported by the other groups.
**Table 10. Shipments consistency evaluation results.**
**Quality Threshold** **Sensors Group** **Consistency (in %)**
0 {1, 2, 3} 0P 3L 21G 76H
{11, 12, 13} 0P 0L 73G 27H
{15, 16, 17} 0P 0L 74G 26H
{49, 50, 51} 0P 0L 68G 32H
0.5, 0.7, 0.9 and 1 {1, 2, 3} 0P 0L 15G 85H
{11, 12, 13} 0P 0L 62G 38H
{15, 16, 17} 0P 0L 88G 12H
{49, 50, 51} 0P 0L 81G 19H
The consistency results for the selected groups were generally good to high, except
for 3% of shipments related to the group {1, 2, 3}, when the quality threshold was at 0. This
shows the impact of the quality threshold on the consistency quality results.
_7.5. Impact of the IoT Data Quality Module on the IoT Data Event Insertion_
For the smart contract IoT data quality evaluation, and due to our blockchain architecture response time (around 1 s per operation), we selected a sample of 3000 events from
the dataset. This sample corresponds to the first 1000 events received from the Sensors 1–3
on 28 February 2004.
-----
_Sensors 2021, 21, 2239_ 22 of 25
The average response time of the addIoTEvent using the 3000 events data sample was
around 1.7 s, with an average standard deviation of 0.174 s. When we disabled the quality
module, with the same data sample, the average response time of this method drops to
around 1.6 s, with an average standard deviation of 0.158 s.
This result shows that our quality module adds only around 0.1 s to the event integration time. The additional quality module cost is acceptable regarding the data quality
improvement brought by this module.
_7.6. Related Works Discussion_
As shown in Section 4, the works of Casado-Vara et al. [31], Hang et al. [32] and
Leal et al. [14] are the closest to our work.
Casado-Vara et al. [31] proposed a vote method to address the accuracy and the
consistency problems. Their vote method is based on the game theory to find a cooperative
temperature among all the used temperature sensors. It is not applicable in our context,
because we have different data sources owned by different stakeholders, and we need to
report all the data sent by those data sources for audit purpose.
In the case of discrepancy between the stakeholder’s data sources related to the same
_shipment, we need to trace this discrepancy, and, if it goes below the fixed quality threshold,_
a corresponding quality incident is created by the smart contract. However, the vote
method in [31] could be used in the very specific case of many shipments with similar
data sources, the same shipper, the same carrier and from which we want to have a global
measure trend.
Hang et al. [32] proposed a Hyperledger Fabric based architecture. This blockchain
implementation choice is perfectly adapted to our B2B use case, and we used the same in
our proposed architecture. However, they only addressed the accuracy problem (outlier
filtering) using the Kalman filter.
Besides, the standard version of Kalman filter did not meet our needs, because the
outlier interval limits are not fixed and evolve according to the received data. This could be
problematic when the Kalman filter goes in fail mode, as stated by Berman [38]. The usage
of an assisted version of the Kalman filter needs to be explored in future work.
Leal et al. [14] proposed using an Ethereum traceability-based architecture. Their
Ethereum choice is justified by the solution monetization goal. However, in our use case,
we chose to work with Hyperledger Fabric which does not need any cryptocurrency
management and has an organization architecture more adapted to our B2B logistic chain
context, in terms of data access levels management.
In addition, Leal et al. [14] addressed the accuracy, consistency and currentness
problems using probability distribution methods, but they did not provide further details
about their application and evaluation of those methods.
Furthermore, the authors of [14] proposed to filter the data inside and outside the
blockchain, which is a good idea, and we already have in our architecture the inside
blockchain data filtering. Besides, we need to explore the adding of a data filtering first
level outside the blockchain, in future works.
The outside blockchain filtering needs to be done carefully, because it should not
prevent the blockchain from getting the required traceability data; although, in some cases
these data will be outliers, they need to be traced for further audit purposes.
_7.7. Conclusions on the Evaluation_
This evaluation section demonstrates the pertinence of the proposed IoT data quality
module and the impact of this module on the data to be used in the traceability smart
contract. The entire data qualification process is executed in a secured and distributed
application on which users agree on every datum to be included, on its qualification process
and decisions to be taken based on this datum.
It is worth noting that the choice of the quality thresholds has a huge impact on the
data filtering process set at the data source level. The events with a quality index below
-----
_Sensors 2021, 21, 2239_ 23 of 25
the defined quality threshold will never be sent to the shipment. This leads directly to data
loss at the shipment level. For this reason, stakeholders may prefer selecting a good quality
threshold ([0.7,0.9]), rather than a high one ([0.9,1]).
Although the proposed architecture evaluation shows encouraging results, this architecture still needs to be tested in a real-life scenario with more data and stakeholders to get
more information about its real performances.
**8. Conclusions and Future Works**
In this article, we propose a distributed architecture and a smart contract to enhance
the IoT data quality in the context of logistic traceability. The proposed architecture uses a
model of IoT data quality with four main data quality dimensions: accuracy, currentness,
completeness and consistency.
We also propose an approach for the calculation of the selected data quality dimensions. The dimensions calculation results are used in our traceability smart contract to set
and control the data quality of events, data sources, shipments and shipments data sources
associations.
The proposed architecture ensures the stakeholders agreement on the data quality
calculation and application rules, and consequently their trust in the decisions taken
automatically by the traceability smart contract. We evaluated our proposed IoT data
quality assessment architecture based on an online available dataset, and the results show
the relevancy of this architecture.
This work could be extended by evaluating the scalability of the proposition when
adding more stakeholders. The approach used to calculate the quality dimensions could
be combined with algorithms, such as DBSCAN [39] or an assisted version of the Kalman
filter [40], to improve the quality index calculation.
The blockchain data charge could be alleviated by adding in this architecture a first
level of data filtering on each stakeholder side. The IoT data sources’ security and interoperability also need to be addressed. Finally, the architecture evaluation needs to be done in
a real-life scenario to ensure its performance in the context of logistic chain traceability.
**Author Contributions: Conceptualization, M.A. and C.T.; methodology, M.A., C.T., S.C., and A.B.;**
software, M.A.; validation, C.T., M.O., S.C. and A.B.; resources, M.A., C.T., S.C. and A.B.; data
curation, M.A.; writing—original draft preparation, M.A.; witing—review and editing, M.A., C.T.,
M.O., S.C., and A.B.; supervision, C.T., M.O., S.C. and A.B.; and project administration, C.T. and M.O.
All authors have read and agreed to the published version of the manuscript.
**Funding: This research was funded by ALIS.**
**Data Availability Statement: Publicly available datasets were analyzed in this study. This data can**
[be found here: http://db.csail.mit.edu/labdata/labdata.html.](http://db.csail.mit.edu/labdata/labdata.html)
**Conflicts of Interest: The authors declare no conflict of interest.**
**References**
1. Hasan, H.; AlHadhrami, E.; AlDhaheri, A.; Salah, K.; Jayaraman, R. Smart contract-based approach for efficient shipment
[management. Comput. Ind. Eng. 2019, 136, 149–159. [CrossRef]](http://doi.org/10.1016/j.cie.2019.07.022)
2. Bumblauskas, D.; Mann, A.; Dugan, B.; Rittmer, J. A blockchain use case in food distribution: Do you know where your food has
[been? Int. J. Inf. Manag. 2020, 52, 102008. [CrossRef]](http://dx.doi.org/10.1016/j.ijinfomgt.2019.09.004)
3. Casino, F.; Kanakaris, V.; Dasaklis, T.K.; Moschuris, S.; Rachaniotis, N.P. Modeling food supply chain traceability based on
blockchain technology. In Proceedings of the 9th IFAC Conference on Manufacturing Modelling, Management and Control MIM
[2019, Berlin, Germany, 28–30 August 2019. [CrossRef]](http://dx.doi.org/10.1016/j.ifacol.2019.11.620.)
4. Wen, Q.; Gao, Y.; Chen, Z.; Wu, D. A Blockchain-based Data Sharing Scheme in The Supply Chain by IIoT. In Proceedings of the
2019 IEEE International Conference on Industrial Cyber Physical Systems (ICPS), Taipei, Taiwan, 6–9 May 2019; pp. 695–700.
[[CrossRef]](http://dx.doi.org/10.1109/ICPHYS.2019.8780161)
5. Ahmed, M.; Taconet, C.; Ould, M.; Chabridon, S.; Bouzeghoub, A. Enhancing B2B supply chain traceability using smart contracts
and IoT. In Proceedings of the Hamburg International Conference of Logistics (HICL) 2020, Online: hosted by the Hamburg
[University of Technology, 23–25 September 2020; Volume 29, pp. 559–589. [CrossRef]](http://dx.doi.org/10.15480/882.3110)
-----
_Sensors 2021, 21, 2239_ 24 of 25
6. Lee, Y.W.; Strong, D.M.; Kahn, B.K.; Wang, R.Y. AIMQ: A methodology for information quality assessment. Inf. Manag. 2002,
_[40, 133–146. [CrossRef]](http://dx.doi.org/10.1016/S0378-7206(02)00043-5)_
7. Karkouch, A.; Mousannif, H.; Al Moatassime, H.; Noël, T. Data quality in internet of things: A state-of-the-art survey. J. Netw.
_[Comput. Appl. 2016, 73, 57–81. [CrossRef]](http://dx.doi.org/10.1016/j.jnca.2016.08.002)_
8. Suciu, G.; N˘adrag, C.; Istrate, C.; Vulpe, A.; Ditu, M.; Subea, O. Comparative Analysis of Distributed Ledger Technologies. In
[Proceedings of the 2018 Global Wireless Summit (GWS), Chiang Rai, Thailand, 25–28 November 2018; pp. 370–373. [CrossRef]](http://dx.doi.org/10.1109/GWS.2018.8686563)
9. Pournader, M.; Shi, Y.; Seuring, S.; Koh, S.L. Blockchain applications in supply chains, transport and logistics: A systematic
[review of the literature. Int. J. Prod. Res. 2020, 58, 2063–2081. [CrossRef]](http://dx.doi.org/10.1080/00207543.2019.1650976)
10. Issaoui, Y.; Khiat, A.; Bahnasse, A.; Ouajji, H. Smart logistics: Study of the application of blockchain technology. In Proceedings of
the 9th International Conference on Current and Future Trends of Information and Communication Technologies in Healthcare
[(ICTH-2019), Coimbra, Portugal, 4–7 November 2019. [CrossRef]](http://dx.doi.org/10.1016/j.procs.2019.09.467)
11. Richard, Y.; Wang, D.M.S. Beyond Accuracy: What Data Quality Means to Data Consumers. J. Manag. Inf. Syst. 1996, 12, 5–33.
12. [ISO 25012: Quality of Data Product. Available online: https://iso25000.com/index.php/en/iso-25000-standards/iso-25012](https://iso25000.com/index.php/en/iso-25000-standards/iso-25012)
(accessed on 26 December 2020).
13. [Liu, C.; Nitschke, P.; Williams, S.; Zowghi, D. Data quality and the Internet of Things. Computing 2019. [CrossRef]](http://dx.doi.org/10.1007/s00607-019-00746-z)
14. Leal, F.; Chis, A.E.; Caton, S.; González–Vélez, H.; García–Gómez, J.M.; Durá, M.; Sánchez–García, A.; Sáez, C.; Karageorgos, A.;
Gerogiannis, V.C.; et al. Smart Pharmaceutical Manufacturing: Ensuring End-to-End Traceability and Data Integrity in Medicine
[Production. Big Data Res. 2021, 24, 100172. [CrossRef]](http://dx.doi.org/10.1016/j.bdr.2020.100172)
15. Byabazaire, J.; O’Hare, G.; Delaney, D. Data Quality and Trust: Review of Challenges and Opportunities for Data Sharing in IoT.
_[Electronics 2020, 9, 2083. [CrossRef]](http://dx.doi.org/10.3390/electronics9122083)_
16. Li, F.; Nastic, S.; Dustdar, S. Data Quality Observation in Pervasive Environments. In Proceedings of the 2012 IEEE 15th
International Conference on Computational Science and Engineering, Paphos, Cyprus, 5–7 December 2012; pp. 602–609.
17. Sicari, S.; Rizzardi, A.; Miorandi, D.; Cappiello, C.; Coen-Porisini, A. A secure and quality-aware prototypical architecture for the
[Internet of Things. Inf. Syst. 2016, 58, 43–55. [CrossRef]](http://dx.doi.org/10.1016/j.is.2016.02.003)
18. Kuemper, D.; Iggena, T.; Toenjes, R.; Pulvermueller, E. Valid.IoT: A Framework for Sensor Data Quality Analysis and Interpolation.
In Proceedings of the 9th ACM Multimedia Systems Conference, MMSys ’18, Amsterdam, The Netherlands, 12–15 June 2018;
[Association for Computing Machinery: New York, NY, USA, 2018; pp. 294–303. [CrossRef]](http://dx.doi.org/10.1145/3204949.3204972)
19. Kolomvatsos, K. A distributed, proactive intelligent scheme for securing quality in large scale data processing. Computing 2019,
_[101, 1687–1710. [CrossRef]](http://dx.doi.org/10.1007/s00607-018-0683-9)_
20. Kara, M.; Lamouchi, O.; Ramdane-Cherif, A. A Quality Model for the Evaluation AAL Systems. In Proceedings of the 7th
International Conference on Current and Future Trends of Information and Communication Technologies in Healthcare (ICTH[2017), Lund, Sweden, 18–20 September 2017. [CrossRef]](http://dx.doi.org/10.1016/j.procs.2017.08.354)
21. Erazo-Garzon, L.; Erraez, J.; Illescas-Peña, L.; Cedillo, P. A Data Quality Model for AAL Systems. In Information and Communication
_Technologies of Ecuador (TIC.EC); Fosenca, C.E., Rodríguez Morales, G., Orellana Cordero, M., Botto-Tobar, M., Crespo Martínez, E.,_
Patiño León, A., Eds.; Springer International Publishing: Cham, Switzerland, 2020; pp. 137–152.
22. Karkouch, A.; Mousannif, H.; Al Moatassime, H.; Noel, T. A model-driven architecture-based data quality management
framework for the internet of Things. In Proceedings of the 2016 IEEE 2nd International Conference on Cloud Computing
Technologies and Applications (CloudTech), Marrakech, Morocco, 24–26 May 2016; pp. 252–259.
23. Fagúndez, S.; Fleitas, J.; Marotta, A. Data Stream Quality Evaluation for the Generation of Alarms in the Health Domain. J. Intell.
_[Syst. 2015, 24, 361–369. [CrossRef]](http://dx.doi.org/10.1515/jisys-2014-0166)_
24. Gu, X.; Peng, J.; Yu, W.; Cheng, Y.; Jiang, F.; Zhang, X.; Huang, Z.; Cai, L. Using blockchain to enhance the security of fog-assisted
crowdsensing systems. In Proceedings of the 2019 IEEE 28th International Symposium on Industrial Electronics (ISIE), Vancouver,
[BC, Canada, 12–14 June 2019; pp. 1859–1864. [CrossRef]](http://dx.doi.org/10.1109/ISIE.2019.8781332)
25. Nguyen, D.; Ali, M.I. Enabling On-Demand Decentralized IoT Collectability Marketplace using Blockchain and Crowdsensing.
[In Proceedings of the 2019 Global IoT Summit (GIoTS), Aarhus, Denmark, 17–21 June 2019; pp. 1–6. [CrossRef]](http://dx.doi.org/10.1109/GIOTS.2019.8766346)
26. [Wei, L.; Wu, J.; Long, C. A Blockchain-Based Hybrid Incentive Model for Crowdsensing. Electronics 2020, 9, 215. [CrossRef]](http://dx.doi.org/10.3390/electronics9020215)
27. Cheng, J.; Long, H.; Tang, X.; Li, J.; Chen, M.; Xiong, N. A Reputation Incentive Mechanism of Crowd Sensing System Based
on Blockchain. In Proceedings of the 6th International Conference on Artificial Intelligence and Security (ICAIS 2020), Hohhot,
[China, 17–20 July 2020; Springer: Singapore, 2020; pp. 695–706. [CrossRef]](http://dx.doi.org/10.1007/978-981-15-8086-4_65)
28. Huang, J.; Kong, L.; Dai, H.; Ding, W.; Cheng, L.; Chen, G.; Jin, X.; Zeng, P. Blockchain-Based Mobile Crowd Sensing in Industrial
[Systems. IEEE Trans. Ind. Inform. 2020, 16, 6553–6563. [CrossRef]](http://dx.doi.org/10.1109/TII.2019.2963728)
29. Zou, S.; Xi, J.; Wang, H.; Xu, G. CrowdBLPS: A Blockchain-Based Location-Privacy-Preserving Mobile Crowdsensing System.
_[IEEE Trans. Ind. Inform. 2020, 16, 4206–4218. [CrossRef]](http://dx.doi.org/10.1109/TII.2019.2957791)_
30. Javaid, A.; Zahid, M.; Ali, I.; Khan, R.; Noshad, Z.; Javaid, N. Reputation System for IoT Data Monetization using Blockchain.
In Proceedings of the 14th International Conference on Broad-Band Wireless Computing, Communication and Applications
(BWCCA), Antwerp, Belgium, 7–9 November 2019; Lecture Notes in Networks and Systems Series; Springer: Cham, Switzerland,
[2019; Volulme 97, pp. 173–184. [CrossRef]](http://dx.doi.org/10.1007/978-3-030-33506-9_16)
-----
_Sensors 2021, 21, 2239_ 25 of 25
31. Casado-Vara, R.; de la Prieta, F.; Prieto, J.; Corchado, J.M. Blockchain Framework for IoT Data Quality via Edge Computing. In
Proceedings of the BlockSys’18: 1st Workshop on Blockchain-Enabled Networked Sensor Systems, Shenzhen, China, 4 November
[2018; Association for Computing Machinery: New York, NY, USA, 2018; pp. 19–24. [CrossRef]](http://dx.doi.org/10.1145/3282278.3282282)
32. Hang, L.; Ullah, I.; Kim, D.H. A secure fish farm platform based on blockchain for agriculture data integrity. Comput. Electron.
_[Agric. 2020, 170, 105251. [CrossRef]](http://dx.doi.org/10.1016/j.compag.2020.105251)_
33. Mary, I.P.S.; Arockiam, L. Imputing the missing data in IoT based on the spatial and temporal correlation. In Proceedings of the
2017 IEEE International Conference on Current Trends in Advanced Computing (ICCTAC), Bangalore, India, 2–3 March 2017;
pp. 1–4.
34. Androulaki, E.; Barger, A.; Bortnikov, V.; Cachin, C.; Christidis, K.; De Caro, A.; Enyeart, D.; Ferris, C.; Laventman, G.; Manevich,
Y.; et al. Hyperledger Fabric: A Distributed Operating System for Permissioned Blockchains. In Proceedings of the Thirteenth
[EuroSys Conference—EuroSys ’18, Porto, Portugal, 23–26 April 2018; pp. 1–15. [CrossRef]](http://dx.doi.org/10.1145/3190508.3190538)
35. [Madden, S. Intel Lab Data. Available online: http://db.csail.mit.edu/labdata/labdata.html (accessed on 15 December 2020).](http://db.csail.mit.edu/labdata/labdata.html)
36. Hui, Z.; Fred, B.; Ed, A.; Yongchao, Z.; Darryl, D.; Xiang, Z.; Maohui, L. Reducing building over-cooling by adjusting HVAC
supply airflow setpoints and providing personal comfort systems. In Proceedings of the 15th Conference of the International
Society of Indoor Air Quality & Climate (ISIAQ), Philadelphia, PA, USA, 22–27 July 2018.
37. [MPR/MIB User’s Manual. Available online: http://www-db.ics.uci.edu/pages/research/quasar/MPR-MIB%20Series%20](http://www-db.ics.uci.edu/pages/research/quasar/MPR-MIB%20Series%20User%20Manual%207430-0021-06_A.pdf)
[User%20Manual%207430-0021-06_A.pdf (accessed on 6 January 2021).](http://www-db.ics.uci.edu/pages/research/quasar/MPR-MIB%20Series%20User%20Manual%207430-0021-06_A.pdf)
38. Berman, Z. Outliers rejection in Kalman filtering—Some new observations. In Proceedings of the 2014 IEEE/ION Position,
[Location and Navigation Symposium—PLANS 2014, Monterey, CA, USA, 5–8 May 2014; pp. 1008–1013. [CrossRef]](http://dx.doi.org/10.1109/PLANS.2014.6851466)
39. Ester, M.; Kriegel, H.P.; Sander, J.; Xu, X. A Density-Based Algorithm for Discovering Clusters in Large Spatial Databases with
Noise. In Proceedings of the Second International Conference on Knowledge Discovery and Data Mining KDD’96, Portland, OR,
USA, 2–4 August 1996; Simoudis, E., Han, J., Fayyad, U., Eds.; AAAI Press: Portland, OR, USA, 1996; pp. 226–231.
40. [Kalman, R.E. A New Approach to Linear Filtering and Prediction Problems. J. Basic Eng. 1960, 82, 35–45. [CrossRef]](http://dx.doi.org/10.1115/1.3662552)
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://pmc.ncbi.nlm.nih.gov/articles/PMC8005206, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "GOLD",
"url": "https://www.mdpi.com/1424-8220/21/6/2239/pdf?version=1616563883"
}
| 2,021
|
[
"JournalArticle"
] | true
| 2021-03-01T00:00:00
|
[
{
"paperId": "194575920b0e5c8241f3328e86fb1d1961098295",
"title": "Smart Pharmaceutical Manufacturing: Ensuring End-to-End Traceability and Data Integrity in Medicine Production"
},
{
"paperId": "02de8b8ad8da33ae3c31cd497f604e0cd36df874",
"title": "Data Quality and Trust: Review of Challenges and Opportunities for Data Sharing in IoT"
},
{
"paperId": "b44cbf45a571973780b638712f1bad9f13885e1b",
"title": "Enhancing B2B supply chain traceability using smart contracts and IoT"
},
{
"paperId": "b9344230b8a0f59e2604c04b5939466a3c6314c8",
"title": "A blockchain use case in food distribution: Do you know where your food has been?"
},
{
"paperId": "52086111ab99095fa5f6dda8832ea51b8b2e500c",
"title": "CrowdBLPS: A Blockchain-Based Location-Privacy-Preserving Mobile Crowdsensing System"
},
{
"paperId": "17cfa4c5d53735dba4a4ed4239a9b4e2ddb0b57d",
"title": "A secure fish farm platform based on blockchain for agriculture data integrity"
},
{
"paperId": "6f9d0b9bbec43810bff5d1aa5ce611e4ec74836e",
"title": "A Blockchain-Based Hybrid Incentive Model for Crowdsensing"
},
{
"paperId": "145709e8c6b5a84c1b2164a4510a13a511099de4",
"title": "Blockchain-Based Mobile Crowd Sensing in Industrial Systems"
},
{
"paperId": "03eab3fe6012f30649c189e72f9c71e1d3778ccc",
"title": "A Data Quality Model for AAL Systems"
},
{
"paperId": "bdfd9dbc612de159e42efe1688b3694c0a8e2365",
"title": "Reputation System for IoT Data Monetization Using Blockchain"
},
{
"paperId": "6eec7163206b1ce96142c997347d600892a6eaab",
"title": "Smart contract-based approach for efficient shipment management"
},
{
"paperId": "b5680cb1a98e30d061cc6ba1bda8b5f2b6ddea71",
"title": "Blockchain applications in supply chains, transport and logistics: a systematic review of the literature"
},
{
"paperId": "515193e0624c4b6583e48c098e16551cb7a12e0c",
"title": "Data quality and the Internet of Things"
},
{
"paperId": "5f3628418c7feb81a410bb01d256d12b162d2d02",
"title": "Enabling On-Demand Decentralized IoT Collectability Marketplace using Blockchain and Crowdsensing"
},
{
"paperId": "a1eee1ef44305cbf9cced3dd8bd20260cf930f07",
"title": "Using blockchain to enhance the security of fog-assisted crowdsensing systems"
},
{
"paperId": "750e7b8fca5c29ee9597a66927b237a94e02a8dc",
"title": "A Blockchain-based Data Sharing Scheme in The Supply Chain by IIoT"
},
{
"paperId": "365b86d2757ac753e8f8fa2d4a4d03649ae92299",
"title": "A distributed, proactive intelligent scheme for securing quality in large scale data processing"
},
{
"paperId": "5b8e9b3546d1785341003fbc6146ad1882c86130",
"title": "Blockchain framework for IoT data quality via edge computing"
},
{
"paperId": "cbdc83b5b1d096fcf362e63b595443d7286d856c",
"title": "Comparative Analysis of Distributed Ledger Technologies"
},
{
"paperId": "c053f1163d56efec3207b69820fc2c3e0590b3ab",
"title": "Reducing building over-cooling by adjusting HVAC supply airflow setpoints and providing personal comfort systems"
},
{
"paperId": "a111f604876b3f291578bf29b865cb34c74e501f",
"title": "Valid.IoT: a framework for sensor data quality analysis and interpolation"
},
{
"paperId": "1b4c39ba447e8098291e47fd5da32ca650f41181",
"title": "Hyperledger fabric: a distributed operating system for permissioned blockchains"
},
{
"paperId": "2e2a40965a593370e3ef681649053dcf56c0031a",
"title": "Imputing the missing data in IoT based on the spatial and temporal correlation"
},
{
"paperId": "823f01704db0b8ad0ef5106a50925504be91cb4b",
"title": "Data quality in internet of things: A state-of-the-art survey"
},
{
"paperId": "2facd1ce6754b44a660e90ea48110462210dd63e",
"title": "A secure and quality-aware prototypical architecture for the Internet of Things"
},
{
"paperId": "adc43198003787ff3cf0425dfe645424f714b0c6",
"title": "A model-driven architecture-based data quality management framework for the internet of Things"
},
{
"paperId": "1eefedddf29f648d0bf1bb0e21044218cc564b4a",
"title": "Data Stream Quality Evaluation for the Generation of Alarms in the Health Domain"
},
{
"paperId": "e1f9bd356faf6dde103d061a72f0f7a338f86127",
"title": "Outliers rejection in Kalman filtering — Some new observations"
},
{
"paperId": "3c967f91408413b7d8ab1d7ac1f85612474dc7f8",
"title": "Data Quality Observation in Pervasive Environments"
},
{
"paperId": "4e82d1fe3066ad65970434c7b1d0a10c3443e94d",
"title": "AIMQ: a methodology for information quality assessment"
},
{
"paperId": "5c8fe9a0412a078e30eb7e5eeb0068655b673e86",
"title": "A Density-Based Algorithm for Discovering Clusters in Large Spatial Databases with Noise"
},
{
"paperId": "b057cc625984119d48846dbf08f30b565f8c263d",
"title": "Beyond Accuracy: What Data Quality Means to Data Consumers"
},
{
"paperId": null,
"title": "Quality of Data Product"
},
{
"paperId": "da87e8e8699fe05ee2c954b49e1b93fdeaed28e9",
"title": "A Reputation Incentive Mechanism of Crowd Sensing System Based on Blockchain"
},
{
"paperId": "7d05797d0b7c96d1a7b43e50775c15b788a64c3e",
"title": "Modeling food supply chain traceability based on blockchain technology"
},
{
"paperId": "7de2516008bb3d6dc8e4e4cc251d3bd837907078",
"title": "Smart logistics: Study of the application of blockchain technology"
},
{
"paperId": "4261315e89f886c44e7a3f9b25aca7e4768cb798",
"title": "A Quality Model for the Evaluation AAL Systems"
},
{
"paperId": "255a77422b1da74da05d1714b7875356187385bd",
"title": "A New Approach to Linear Filtering and Prediction Problems"
},
{
"paperId": null,
"title": "MPR / MIB User ’ s Manual"
},
{
"paperId": null,
"title": "Intel Lab Data"
}
] | 23,393
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/021d12b8bd39c09f273ee6e30fc70916bf3e52ca
|
[
"Computer Science"
] | 0.855219
|
A Distributed Path Query Engine for Temporal Property Graphs
|
021d12b8bd39c09f273ee6e30fc70916bf3e52ca
|
IEEE/ACM International Symposium on Cluster, Cloud and Internet Computing
|
[
{
"authorId": "1491748089",
"name": "S. Ramesh"
},
{
"authorId": "1491747900",
"name": "Animesh Baranawal"
},
{
"authorId": "1761220",
"name": "Yogesh L. Simmhan"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"Clust Comput Grid",
"CCGRID",
"IEEE/ACM Int Symp Clust Cloud Grid Comput",
"IEEE/ACM International Symposium Cluster, Cloud and Grid Computing",
"Cluster Computing and the Grid",
"IEEE/ACM Int Symp Clust Cloud Internet Comput"
],
"alternate_urls": null,
"id": "57f970eb-366a-4bfa-aa06-2ff70d834806",
"issn": null,
"name": "IEEE/ACM International Symposium on Cluster, Cloud and Internet Computing",
"type": "conference",
"url": "http://www.buyya.com/ccgrid/"
}
|
Property graphs are a common form of linked data, with path queries used to traverse and explore them for enterprise transactions and mining. Temporal property graphs are a recent variant where time is a first-class entity to be queried over, and their properties and structure vary over time. These are seen in social, telecom and transit networks. However, current graph databases and query engines have limited support for temporal relations among graph entities, no support for time-varying entities and/or do not scale on distributed resources. We address this gap by extending a linear path query model over property graphs to include intuitive temporal predicates that operate over temporal graphs. We design a distributed execution model for these temporal path queries using the interval-centric computing model, and develop a novel cost model to select an efficient execution plan from several. We perform detailed experiments of our $\mathcal{G}ranite$ distributed query engine using temporal property graphs as large as 52M vertices, 218M edges and 118M properties, and an 800-query workload, derived from the LDBC benchmark. We offer sub-second query latencies in most cases, which is 149×-1140× faster compared to industry-leading Neo4J shared- memory graph database and the JanusGraph/Spark distributed graph query engine. Further, our cost model selects a query plan that is within 10% of the optimal execution time in 90% of the cases. We also scale well, and complete 100% of the queries for all graphs, compared to only 32-92% by baseline systems.
|
## A Distributed Path Query Engine for Temporal Property Graphs *
#### Shriram Ramesh, Animesh Baranawal and Yogesh Simmhan
_Department of Computational and Data Sciences,_
_Indian Institute of Science, Bangalore 560012, India_
_Email: {shriramr, animeshb, simmhan}@iisc.ac.in_
**Abstract**
Property graphs are a common form of linked data, with path
queries used to traverse and explore them for enterprise transactions
and mining. Temporal property graphs are a recent variant where time
is a first-class entity to be queried over, and their properties and structure vary over time. These are seen in social, telecom, transit and
epidemic networks. However, current graph databases and query engines have limited support for temporal relations among graph entities, no support for time-varying entities and/or do not scale on distributed resources. We address this gap by extending a linear path
query model over property graphs to include intuitive temporal pred_icates and aggregation operators over temporal graphs. We design a_
_distributed execution model for these temporal path queries using the_
interval-centric computing model, and develop a novel cost model to
select an efficient execution plan from several. We perform detailed
experiments of our _ranite distributed query engine using both static_
_G_
and dynamic temporal property graphs as large as 52M vertices, 218M
edges and 325M properties, and a 1600-query workload, derived from
the LDBC benchmark. We often offer sub-second query latencies on a
commodity cluster, which is 149 –1140 faster compared to industry_×_ _×_
leading Neo4J shared-memory graph database and the JanusGraph/Spark distributed graph query engine. _ranite also completes 100% of_
_G_
the queries for all graphs, compared to only 32–92% workload completion by the baseline systems. Further, our cost model selects a query
plan that is within 10% of the optimal execution time in 90% of the
cases. Despite the irregular nature of graph processing, we exhibit a
weak-scaling efficiency of 60% on 8 nodes and 40% on 16 nodes,
_≥_ _≥_
for most query workloads.
*An extended version of the paper that appears in IEEE/ACM International Symposium on Cluster, Cloud and Internet Computing (CCGrid), 2020.
[doi:10.1109/CCGrid49817.2020.00-43](https://dx.doi.org/10.1109/CCGrid49817.2020.00-43)
1
-----
### 1 Introduction
Graphs are a natural model to represent and analyze linked data in various domains. Property graphs allow vertices and edges to have associated
_key–value pair properties, besides the graph structure. This forms a rich_
information schema and has been used to capture knowledge graphs (concepts, relations) [1], social networks (person, forum, message) [2], epidemic
networks (subject, infected status, location) [3], and financial and retail
transactions (person, product, purchase) [4].
_Path queries are a common class of queries over property graphs [5,_
6]. Here, the user defines a sequence of predicates over vertices and edges
that should match along a path in the graph. E.g., in the property graph
for a community of users in Figure 1, the vertices are labeled with their
_IDs, their colors indicate their type – blue for Person and orange for a_
_Post, and they have a set of properties listed as Name:Value. The edges_
are relationships, with types such as Follows, Likes and Created. We can
define an example 3-hop path query “[EQ1] Find a person (vertex type) who
_lives in the country ‘UK’ (vertex property) and follows (edge type) a person_
_who follows another person who is tagged with the label ‘Hiking’ (vertex_
_property)”. This query would match Cleo_ _Alice_ _Bob, if we ignore the time_
_→_ _→_
intervals. Path queries are used to identify concept pathways in knowledge
graphs, find friends in social networks, fake news detection, and suggest
products in retail websites [5, 6, 7]. They also need to be performed rapidly,
within 1 sec, as part of interactive requests from websites or exploratory
_≈_
queries by analysts.
While graph databases are designed for transactional read and write
workloads, we consider graphs that are updated infrequently but queried
often. For these workloads, graph query engines load and retain property
graphs in-memory to service requests with low latency, without the need
for locking or consistency protocols [8, 9]. They may also create indexes
to accelerate these searches [10, 11]. Property graphs can be large, with
10[5]–10[8] vertices and edges, and 10’s of properties on each vertex or edge.
This can exceed the memory on a single machine, often dominated by the
properties. This necessitates the use of distributed systems to scale to large
graphs [12, 13].
**Challenges** Time is an increasingly common graph feature in a variety of
domains [3, 14, 15, 16]. However, existing property graph data models fail to
consider it as a first-class entity. Here, we distinguish between graphs with a
_time interval or a lifespan associated with their entities (properties, vertices,_
edges), and those where the entities themselves change over time and the
history is available. We call the former static temporal graphs and the latter
_dynamic temporal graphs. Yet another class is streaming graphs, where the_
topology and properties change in real-time, and queries are performed on
2
-----
|Col1|[5,100) Tag: Hiking|
|---|---|
**Likes**
**Follows**
**Created**
**Don**
**Cleo**
**Pic**
**Post**
Figure 1: Sample Temporal Property Graph of a Community of Users
this evolving structure [17, 18]; that is outside the scope of this article.
E.g., in the temporal graph in Figure 1, the lifespan, [start, end), is
indicated on the vertices, edges and properties. The start time is inclusive
while the end time is exclusive. Other than the properties of Cleo, the
remaining entities of the graph form a static temporal graph as they are each
valid only for a single time range. But the value of the Country property of
_Cleo changes over time, making it a dynamic temporal graph._
This gap is reflected not just in the data model but also in the queries
_supported. We make a distinction between time-independent (TI) and time-_
_dependent (TD) queries, both being defined on a temporal graph [19]. TI_
queries are those which can be answered by examining the graph at a single
point in time (a snapshot), e.g. EQ1 executed on the temporal graph. In
contrast, TD queries capture temporal relations between the entities across
consecutive time intervals, e.g., “[EQ2] Find people tagged with ‘Hiking’
_who liked a post tagged as ‘Vacation’, before the post was liked by a per-_
_son named ‘Don’ ”,_ and “[EQ3] Find people who started to follow an_other person, after the latter stops following ‘Don’ ”._ Treating time as
just another property fails to express temporal relations such as ensuring time-ordering among the entities on the path. While EQ2 and EQ3
should match the paths Bob _PicPost_ _Don and Alice_ _Bob_ _Don, respec-_
_→_ _→_ _→_ _→_
tively, such queries are hard, if not impossible, to express in current graph
databases. This problem is exacerbated for path queries over dynamic tem_poral graphs. E.g., the query EQ1 over the dynamic temporal graph should_
_not match Cleo_ _Alice_ _Bob since at the time Cleo was living in ‘UK’, she_
_→_ _→_
was not following Alice.
While platforms which execute snapshot at a time [19,20] can be adapted
to support TI queries over temporal graphs, TD queries cannot be expressed
meaningfully. Even those that support TD algorithms enforce strict temporal ordering [21], requiring that the time intervals along the path should be
increasing or decreasing, but not both; this limits query expressivity. These
motivate the need to support intuitive temporal predicates to concisely ex
3
**Alice**
**Bob**
-----
press such temporal relations, and flexible platforms to execute them. Lastly,
the scalability of existing graph systems is also limited, with few property
graph query engines that operate on distributed memory systems with low
latency [8, 22], let alone on temporal property graphs.
We make the following specific contributions in this article:
We propose a temporal property graph model, and intuitive temporal
_•_
_predicates and aggregation operators for path queries on them (§3)._
We design a distributed execution model for these queries using the
_•_
interval-centric computing model (§4).
We develop a novel cost model that uses graph statistics to select the
_•_
best from multiple execution plans (§5).
We conduct a detailed evaluation of the performance and scalability
_•_
of _ranite for 8 temporal graphs and up to 1600 queries, derived from_
_G_
the LDBC benchmark. We compare this against three configurations
of Neo4J, and JanusGraph which uses Apache Spark (§6).
We discuss related work in Section 2 and our conclusions in Section 7.
A prior version of this work appeared as a conference paper [23]. This
article substantially extends this. Specifically, it introduces the temporal
aggregation operator to the query model (Section 3.3) and implements it
within the execution model; offers details, illustrations and complexity metrics for our query model, distributed execution model and query optimizations (Sections 3, 4 and 5); and provides a rigorous empirical evaluation,
including two additional large dynamic temporal graphs, aggregation query
workloads, weak scaling experiments, and results on the component times of
query execution, besides more detailed analysis for the cost model benefits
and baseline platform comparisons (Section 6).
### 2 Related Work
#### 2.1 Distributed and Temporal Graph Processing
There are several distributed graph processing platforms for running graph
algorithms on commodity clusters and clouds [24]. These typically offer programming abstractions like Google Pregel’s vertex-centric computing model [20]
and its component-centric variants [25, 26] to design algorithms such as
Breadth First Search, centrality scores and mining [27]. These execute using a Bulk Synchronous Parallel (BSP) model, and scale to large graphs and
applications that explore the entire graph. They offer high throughput batch
_processing that take_ (mins)– (hours). We instead focus on exploratory
_O_ _O_
and transactional path queries that are to be processed in (secs). This
_O_
4
-----
requires careful use of distributed graph platforms and optimizations for fast
responses.
There are also parallel graph platforms for HPC clusters and accelerators [28]. These optimize the memory and communication access to scale
to graphs with billions of entities on thousands of cores [29]. They focus on
high-throughput graph algorithms and queries over static graphs [30]. We
instead target commodity hardware and cloud VMs with 10’s of nodes and
100’s of total cores, and are more accessible. We also address queries over
temporal property graphs.
A few distributed abstractions and platforms support designing of temporal algorithms and their batch execution [19, 31, 32]. Most are limited to
executing TI algorithms, snapshot at a time, and are unable to seamlessly
model TD queries. Our prior work Graphite offers an interval-centric computing model (ICM) to represent TI and TD algorithms, but limits it to
time-respecting algorithms [21]. We use it as the base framework for our
proposed distributed path query engine, while relaxing the time-ordering,
including indexing and proposing different query execution plans for lowlatency response. There are also some platforms that support incremental
computing over streaming graph updates [33, 34]. We rather focus on materialized property graphs with temporal lifespans on their vertices, edges
and properties that have already been collected in the past. In future, we
will also consider incremental query processing over such streaming graphs.
#### 2.2 Property and Temporal Graph Querying
Query models over property graphs and associated query engines are popular for semantic graphs [30, 35, 36]. Languages like SPARQL offer a highly
flexible declarative syntax, but are costly to execute in practice for large
graphs [37, 38]. Others support a narrower set of declarative query primitives, such as finding paths, reachability and patterns over property graphs,
but manage to scale to large graphs using a distributed execution model [39,
40]. However, none of these support time as a first-class entity, during query
specification or execution.
There has been limited work on querying and indexing over specific temporal features of property graph. Semertzidis, et al. [41] propose a model for
finding the top-k graph patterns which exist for the longest period of time
over a series of graph snapshots. They offer several indexing techniques to
minimize the snapshot search space, and perform a brute-force pattern mining on the restricted set. This multi-snapshot approach limits the pattern
to one that fully exists at a single time-point and recurs across time, rather
than spans time-points. It is also limited to a single-machine execution,
which limits scaling.
_TimeReach [42] supports conjunctive and disjunctive reachability queries_
on a series of temporal graph snapshots. It builds an index from strongly
5
-----
connected components (SCC) for each snapshot, condenses them across
time, and use this to traverse between vertices in different SCCs within
a single hop. It assumes that the graph has few SCCs that do not change
much over time. They also require the path to be reachable within a single
snapshot rather than allow path segments to connect across time. Likewise,
_TopChain [43] supports temporal reachability query using an index label-_
ing scheme. It unrolls the temporal graph into a static graph, with time
expanded as additional edges, finds the chain-cover over it, and stores the
top-k reachable chains from each vertex as labels. It uses this to answer
time-respecting reachability, earliest arrival path and fastest path queries.
Paths can span time intervals. However, they do not support any predicates
over the properties. Neither of these support distributed execution.
There is also literature on approximate querying over graphs. Arrow [44]
examines reachability queries on both non-temporal and temporal graphs
using random walks. These are performed from both the source and the
sink vertices, and an intersection of the two vertex sets gives the result.
They use approximation by bounding the walk length based on the diameter
of the graph and a tunable parameter which balances accuracy and query
latency. Iyer, et al. [45] consider approximate pattern mining on large nontemporal graphs. They use statistical techniques to sample the graph edges
and estimate the number of occurrences of a specific pattern in the graph.
However, their approach cannot enumerate the actual vertices and edges
forming the pattern.
_ChronoGraph [46] supports temporal traversal queries over interval prop-_
erty graphs, and is the closest to our work. They implement this by extending the Gremlin property graph query language with temporal properties.
They propose optimizations to the Gremlin traversal operators, and parallelization and lazy traversals within a single machine, which are executed
by the TinkerGraph engine. However, they do not support novel temporal
operators such as the edge-temporal relationship that we introduce. They
also do not use indexes or query planning to make the execution plan more
efficient. Their optimizations are tightly-coupled to the execution engine,
which does not support distributed execution.
Lastly, there are several open-source and proprietary graph database
systems [8, 47, 48] which provide general-purpose property graph storage
and querying capabilities while allowing transactional access to graph data.
However, these systems do not have first class support for time-varying
graphs and query models that can leverage the temporal dimension. This
leads to temporal queries written in their native language which are neither
intuitive in expressing temporal notion nor efficient during execution due to
lack of time-aware query optimizer and execution engine.
In summary, these various platforms lack one or more of the following
capabilities we offer: modeling time as a first-class graph and query concept;
enabling temporal path queries that span time and match temporal relations
6
-----
across entities; and distributed execution on commodity clusters that scales
to large graphs using a query optimizer that leverages the graph’s structure,
temporal features, and property values.
### 3 Temporal Graph and Query Models
#### 3.1 Temporal Concepts
The temporal property graph concepts used in this paper are drawn from
our earlier work [21]. Time is a linearly ordered discrete domain Ωwhose
range is the set of non-negative whole numbers. Each instant in this domain
is called a time-point and an atomic increment in time is called a time-unit.
A time interval is given by τ = [ts, te) where ts, te ∈ Ωwhich indicates an
interval starting from and including ts and extending to but excluding te.
_Interval relations [49] are Boolean comparators between intervals; fully be-_
_fore relation is denoted by_, starts before relation by, fully after relation
_≪_ _≺_
by, starts after relation by, during relation by, equals relation by =,
_≫_ _≻_ _⊂_
_during or equals relation by_ and overlaps relation by .
_⊆_ _⊓_
#### 3.2 Temporal Property Graph Model
We formally define a temporal property graph as a directed graph G =
(V, E, PV, PE). V is a set of typed vertices where each vertex ⟨vid, σ, τ _⟩∈_ _V_
is a tuple with a unique vertex ID, vid, a vertex type (or schema) σ, and
the lifespan of existence of the vertex given by the interval, τ = [ts, te). E
is a set of directed typed edges, with ⟨eid, σ, vidi, vidj, τ _⟩∈_ _E. Here, eid_
is a unique ID of the edge, σ its type, vidi and vidj are its source and
sink vertices respectively, and τ = [ts, te) is its lifespan. We have a schema
function : σ _K, that maps a given vertex or edge type σ to the set of_
_S_ _→_
_property keys (or names) it can have. PV is a set of vertex property values,_
where each ⟨vid, κ, val, τp⟩∈ _PV represents a value val for the key κ ∈_ _K_
for the vertex vid, with the value valid for the interval τp ⊆ _τ_ . A similar
definition applies for edge property values ⟨eid, κ, val, τp⟩∈ _PE._
Further, the graph G must meet the uniqueness constraint of vertices
and edges, i.e., a vertex or an edge with a given ID exist at most once and
for a single continuous duration; referential integrity constraints, where the
lifespan of an edge must be contained within the lifespan of its incident
vertices; and constant edge association, which enforces that the vertices incident on an edge remain the same during the edge’s lifespan. These are
defined in [50].
A static temporal property graph is a restricted version of the temporal
property graph such that τp = τ for the vertex and edge properties, i.e.,
each property key has a static value that is valid for the entire vertex or
edge lifespan, formally stated as:
7
-----
_∀⟨vid, κ, val, τp⟩∈_ _PV, ⟨vid, σ, τ_ _⟩∈_ _V =⇒_ _τp = τ_ and _∀⟨eid, κ, val, τp⟩∈_
_PE, ⟨eid, σ, vidi, vidj, τ_ _⟩∈_ _E =⇒_ _τp = τ Temporal property graphs with-_
out this restriction are called dynamic temporal property graphs, and allow
keys for a vertex or an edge to have different values for non-overlapping time
intervals, i.e., τp ⊆ _τ_ . E.g., Figure 1 is a dynamic temporal property graph
as Cleo’s property values change over time, but omitting Cleo makes it a
static temporal property graph.
#### 3.3 Temporal Path Query
An n-hop linear chain path query matches a path with n vertex predicates
and n 1 edge predicates. The syntax rules for this query model and its
_−_
predicates are given below, and illustrated for the example queries from
earlier in Table 1.
_path_ ::= _ve-fragment_ _ve-int-fragment_ - _v-predicate_
_⟨_ _⟩_ _⟨_ _⟩⟨_ _⟩_ _⟨_ _⟩_
| _ve-fragment_ _ve-int-fragment_ - _v-predicate_ _aggregate_
_⟨_ _⟩⟨_ _⟩_ _⟨_ _⟩⊕⟨_ _⟩_
_ve-fragment_ ::= _v-predicate_ _e-predicate_
_⟨_ _⟩_ _⟨_ _⟩⊢⟨_ _⟩_
_ve-int-fragment_ ::= _ve-fragment_ _v-predicate_ _etr-clause_ _e-predicate_
_⟨_ _⟩_ _⟨_ _⟩| ⟨_ _⟩⟨_ _⟩⊢⟨_ _⟩_
_v-predicate_ ::= _predicate_
_⟨_ _⟩_ _⟨_ _⟩_
_e-predicate_ ::= _predicate_ _direction_
_⟨_ _⟩_ _⟨_ _⟩⟨_ _⟩_
_direction_ ::=
_⟨_ _⟩_ _→| ←| ↔_
_predicate_ ::= ⋆ _bool-predicate_ _prop-clause_ _time-clause_
_⟨_ _⟩_ _| ⟨_ _⟩| ⟨_ _⟩| ⟨_ _⟩|_
_time-clause_ AND _bool-predicate_
_⟨_ _⟩_ _⟨_ _⟩_
_bool-predicate_ ::= _prop-clause_ _prop-clause_ OR _bool-predicate_
_⟨_ _⟩_ _⟨_ _⟩| ⟨_ _⟩_ _⟨_ _⟩_
| _prop-clause_ AND _bool-predicate_
_⟨_ _⟩_ _⟨_ _⟩_
_prop-clause_ ::= ve-key _prop-compare_ value
_⟨_ _⟩_ _⟨_ _⟩_
_time-clause_ ::= ve-lifespan _time-compare_ interval
_⟨_ _⟩_ _⟨_ _⟩_
_etr-clause_ ::= el-lifespan _time-compare_ er-lifespan
_⟨_ _⟩_ _⟨_ _⟩_
_prop-compare_ ::= ‘==’ ‘!=’
_⟨_ _⟩_ _|_ _| ∋_
_time-compare_ ::=
_⟨_ _⟩_ _≺| ≪| ≻| ≫| ⊓| ̸ ⊓_
_aggregate_ ::= _aggregate-op_ [ v-key | ⋆ ]
_⟨_ _⟩_ _⟨_ _⟩_
_aggregate-op_ ::= count min max
_⟨_ _⟩_ _|_ _|_
As we can see, the property and time clauses are the atomic elements of
the predicate and allow one to compare in/equality and containment between
a property value and the given value, and a more flexible set of comparisons between a vertex/edge/property lifespan and a given interval (time_compare). These temporal clauses allow a wide variety of comparison within_
8
-----
the context of a single vertex or edge, and their properties. These clauses
can be combined using Boolean AND and OR operators. Edge predicates can
have an optional direction. The wildcard ⋆ matches all vertices or edges at
a hop.
A novel and powerful temporal operator we introduce is edge time re_lationship (ETR). Unlike the time clause, this etr-clause allows comparison_
across edge lifespans. Specifically, it is defined on an intermediate vertex in
the path (ve-int-fragment), and allows us to compare the lifespans of its left
(el-lifespan) and right (er-lifespan) edges in the path. The motivation for
this operator comes from social network mining [6] and to identify flow and
frauds in transactions networks [4]. E.g., the queries EQ2 and EQ3 from
Section 1 can be concisely captured using this.
We also support a novel temporal aggregate operator to group the resultset from the path query. The paths are grouped on the first vertex in the
resulting temporal paths, and computes a specific aggregation on a property
at the last vertex of the path. The grouping is time-aware; specifically, it
is based on the duration of the first vertex in the result path. E.g., if the
result-set for a query contains i = 1..m paths of length n each, v1[i] _[−]_ _[e]1[i]_ _[−]_
_v2[i]_ 2 _n[, and the first vertex][ v]1[i]_ [in a result matches the query for the]
_[−]_ _[e][i]_ _[−]_ _[...]_ _[−]_ _[v][i]_
time period τi = [t[i]s[, t][i]e[), then we perform a “group by” of the result paths]
by the temporal vertex {v1[i] _[.id,][ [][t][i]s[, t][i]e[)][}][. For all the paths][ j][ in a group, we]_
perform an aggregation operation ⊕ on vn[j] _[.prop][, where][ prop][ is a property on]_
the last vertex that is selected by the user and may be omitted for a count
aggregation. We return the aggregated result {v1[i] _[.vid,][ [][t][i]s[, t][i]e[)][,][ ⊕]_ _n[.prop][)][}]_
_j_ [(][v][j]
for each unique temporal vertex group [1]. This can help answer queries such
as “[EQ4] Count the number of persons followed by a person ‘Bob’ during
_his existence in the network”. The answer to this for Figure 1 varies across_
time, taking value 1 during [10, 30) [50, 100) and 0 during [5, 10) [30, 50).
_∪_ _∪_
Our _ranite implementation supports count, min and max operations for_
_G_
, while others can also easily be added.
_⊕_
### 4 Distributed Query Engine
#### 4.1 Relaxed Interval Centric Computing
The high-level architecture of our distributed query engine, _ranite, is_
_G_
shown in Figure 2a. Our query engine uses a distributed in-memory iterative
execution model that extends and relaxes the Interval-centric Computing
_Model (ICM) [21]. ICM adds a temporal dimension to Pregel’s vertex-centric_
iterative computing model [20], and allows users to define their computation
from the perspective of a single interval-vertex, i.e., the state and properties
1The valid duration for the first vertex can be disjoint, in which case each maximal
contiguous interval for that vertex vid forms a separate temporal group.
9
-----
Table 1: Query Syntax Examples
**Example Query** **Query Syntax**
**EQ1 Find a person who lives** **Type == Person AND Country == UK**
_⊢_
in ‘UK’ and follows a per- **Type == Follows**
_→_
son who follows another person **Type == Person** **Type == Follows**
_⊢_ _→_
who is tagged with ‘Hiking’ **Type == Person AND Tag** _Hiking_
_∋_
**Type == Person AND Tag** _Hiking_ **Type**
**EQ2** Find people tagged _∋_ _⊢_
== Likes
with ‘Hiking’ who liked a post _→_
**Type** == _Post_ AND **Tag** _Vacation_
tagged as ‘Vacation’ before the _∋_
el-lifespan er-lifespan
post was liked by a person _≺_
**Type == Likes**
named ‘Don’ _⊢_ _←_
**Type == Person AND Name == Don**
**Type == Person** **Type == Follows**
**EQ3 Find people who started** _⊢_ _→_
**Type** == _Person_ el-lifespan
to follow another person, after _≫_
er-lifespan **Type == Follows**
they stopped following ‘Don’ _⊢_ _→_
**Type == Person AND Name == Don**
**EQ4** Count the number of
**Type == Person AND Name == Bob** **Type**
persons followed by a person _⊢_
== Follows
‘Bob’ during his existence in _→_
**Type == Person** count [⋆]
the network. _⊕_
for a certain interval of a vertex’s lifespan. In each iteration (superstep)
of an ICM application, a user-defined compute function is called on each
active interval-vertex, which operates on its prior state and on messages it
receives from its neighbors, for that interval, and updates the current state.
A TimeWarp function aligns the lifespans of the input messages to the lifespans of the partitioned interval states for an interval vertex. So each call to
compute executes on the temporally intersecting messages and states for a
vertex. Then, a user-defined scatter function is called on the out-edges of
that interval-vertex, which allows them to send temporal messages containing, say, the updated vertex state to its neighboring vertices. The message
lifespan is usually the intersection of the state and the edge lifespans.
Messages are delivered in bulk at a barrier after the scatter phase,
and the compute phase for the next iteration starts after that. Vertices
receiving a message whose interval overlaps with its lifespan are activated
for the overlapping period. This repeats across supersteps until no messages
are generated after a superstep. The execution of compute and scatter
functions are each data-parallel within a superstep, and their invocation on
different interval vertices and edges can be done by concurrent threads.
We design _ranite using the compute and scatter primitives offered_
_G_
the Graphite implementation of ICM over Apache Giraph, as illustrated in
Figure 2b. However, ICM enforces time-respecting behavior, i.e., the intervals between the messages and the interval-vertex state have to overlap for
compute to be called on the messages; intervals between the states updated
10
|Example Query|Query Syntax|
|---|---|
|Example Query|Query Syntax|
|---|---|
|EQ1 Find a person who lives in ‘UK’ and follows a per- son who follows another person who is tagged with ‘Hiking’|Type == Person AND Country == UK ⊢ Type == Follows → Type == Person Type == Follows ⊢ → Type == Person AND Tag Hiking ∋|
|EQ2 Find people tagged with ‘Hiking’ who liked a post tagged as ‘Vacation’ before the post was liked by a person named ‘Don’|Type == Person AND Tag Hiking Type ∋ ⊢ == Likes → Type == Post AND Tag Vacation ∋ el-lifespan er-lifespan ≺ Type == Likes ⊢ ← Type == Person AND Name == Don|
|EQ3 Find people who started to follow another person, after they stopped following ‘Don’|Type == Person Type == Follows ⊢ → Type == Person el-lifespan ≫ er-lifespan Type == Follows ⊢ → Type == Person AND Name == Don|
|EQ4 Count the number of persons followed by a person ‘Bob’ during his existence in the network.|Type == Person AND Name == Bob Type ⊢ == Follows → Type == Person count [⋆] ⊕|
-----
**Stats** **Master .Optimizer**
**Receive & broadcast query to worker**
- **Select query plan using cost model**
- **Coordinate worker exec. for query**
**plan ◊** **Return result set to client**
**Init/**
**Compute**
**(Resume)**
**vertex**
**predicate**
**evaluation.**
**V**
**V**
Controls Messaging
|Granite|Master .Optimizer Stats Receive & broadcast query to worker ◊Select query plan using cost model ◊Coordinate worker exec. for query plan ◊Return result set to client|Col3|Init/ Scatter Compute Evaluate edge (Resume) predicate, temporal vertex edge relation. predicate Send partial results evaluation. message to sink vertex.|
|---|---|---|---|
|ICM on Interval Property Graph Graphite VCM on Graph Apache Giraph Query Worker Worker Worker Worker Master||ICM on Interval Property Graph Graphite||
|||VCM on Graph Apache Giraph||
|Giraph Worker Partition Compute Graphite Interval Compute Evaluate Vertex Granite P Ar ce td ivic ea vte e r𝜋 ti𝑖ceo sn init / compute V V V scatter|Col2|Col3|Col4|Col5|Col6|Col7|Col8|
|---|---|---|---|---|---|---|---|
|||||V V V||||
||scatter|||||||
**Scatter**
**Evaluate edge**
**predicate, temporal**
**edge relation.**
**Send partial results**
**message to sink vertex.**
**Partition Compute**
Evaluate Edge
_MatchedPredicate verticesഥ𝜋𝑖_ on
(b) Iterative query execution
across ICM supersteps
**V**
HDFS
(a) Architecture of Granite
**Graphite**
**Interval**
**Compute**
**Granite**
**init /**
**compute**
**scatter**
Figure 2: Architecture and ICM execution model of _ranite_
_G_
by the compute and the edge lifespans have to overlap for scatter to be
called; and scatter sends messages only on edges whose lifespan overlaps
with the updated states.
But the temporal path queries do not need to meet these requirements,
e.g., a query may need to navigate from a vertex to an adjacent vertex
that occurs after it. The TimeWarp operator of ICM enforces this timerespecting behavior. So we relax ICM to allow non-time respecting behavior between compute, scatter and messages to meet the execution
requirements of our path queries, while leveraging its other interval-centric
features.
#### 4.2 Distributed Execution Model
In our execution model, each vertex predicate for a path query and the
succeeding edge predicate, if any, are evaluated in a single ICM superstep.
Specifically, the vertex predicates are evaluated in the compute function
and the edge predicates in the scatter function. We use a specialized
logic called init for evaluating the first vertex predicate in a query. This
is shown in Figs. 2b and 3a.
A Master receives the path query from the client, and broadcasts it to
all Workers to start the first superstep (Figure 2a). Each Worker operates
over a set of graph partitions with one thread per partition, and each thread
calls the compute and scatter functions on every active vertex in its
partition. The init logic is called on all vertices in the first superstep.
It resets the vertex state for this new query and evaluates the first vertex
predicate of the query. If the vertex matches, its state is updated with a
_matched flag and scatter is invoked for each of its incident in or out edges,_
as defined in the query. Scatter evaluates the next edge predicate, and if
it matches, sends the partial path result and the evaluated path length to
the destination vertex as a message. If a match fails, this path traversal is
pruned.
11
-----
In the next iteration, our compute logic is called for vertices receiving
a message. This evaluates the next vertex predicate in the path and if it
matches, it puts all the partial path results from the input messages in the
vertex state, and scatter is called on each incident edge. If the edge
matches the next edge predicate, the current vertex and edge are appended
to each prior partial result and sent to the destination vertex. This repeats
for as many supersteps as the path length. In the last superstep, the vertices
receiving matching paths in their messages send it to the Master to return
to the client.
Figure 3a (Plan 1) illustrates this for a sample path query with vertex
and edge predicates, V 1 _E1_ _V 2_ _E2_ _V 3. In superstep 1, init is called_
_−_ _−_ _−_ _−_
on all vertices to evaluate the vertex predicate V 1, and for the ones that
match, scatter is called to evaluate the edge predicate E1. Those edges
that match send a message to their remote vertex in superstep 2, where all
vertices that receive a message invoke their compute logic to evaluate the
vertex predicate V 2 of the second hop. This is (optionally) preceded by
the TimeWarp operator on all the messages received by an interval vertex.
Vertices that match V 2 call scatter on their edges to match the predicate
_E2, and send messages if they too match. In the last superstep, vertices_
that receive messages evaluate the predicate V 3, and if there is a match,
return that result path to the user. Each vertex in the last superstep may
return multiple matching paths based on the messages received, and different
vertices may return result paths to the Master.
Scatter also evaluates the edge temporal relationship. Here, the scatter
of the preceding edge passes its lifespan in the result message, and this is
compared against the current edge’s lifespan by the next scatter to decide
on a match. In the case of temporal aggregate queries, the result set is constructed in the last superstep, similar to the non-aggregate queries. Then,
the first vertex in each result path, its associated lifespan, and the count or
the property value of the last vertex to be aggregated are extracted and sent
to the Master. The Master temporally groups the values for each distinct
temporal vertex using the TimeWarp operator, and applies the aggregation
operator on the values in each group.
For static temporal graphs, we do not use any interval-centric features of
ICM, such as TimeWarp, and the entire lifespan of the vertex is treated as a
single interval-vertex for execution, and likewise for edges. However, we do
use the property graph model and state management APIs offered by the
interval-vertex.
For dynamic temporal graphs with time varying properties, we leverage
the interval-centric features of ICM. Specifically, we enable TimeWarp of
message intervals with the vertex properties’ lifespans so that compute is
called on an interval vertex with messages temporally aligned and grouped
against the property intervals. Scatter is called only for edges whose
lifespans overlap with the matching interval-vertex, and its scope is limited
12
-----
_Query_
**E1** **E2**
**V1** **V2** **V3**
|SS|V1|Col3|E1|V2|E2|V3|
|---|---|---|---|---|---|---|
|1|I||S||S||
|2||||CJ|||
|SS|V1|Col3|E1|V2|E2|V3|
|---|---|---|---|---|---|---|
|1|I||S||||
|2||||C|S||
|3||||||C|
**V2**
**V3**
(a) Query execution phases across different supersteps
(SS) for two plans of the input query
**Tag ϶ Vacation**
**Likes** **Likes**
**Tag ϶** **Name**
**V1** **V2** **V3**
**Hiking** **== Don**
**Sub** **Sub**
**V1** **V2** **Split** **V2** **V3**
**query 1** **query 2**
**SS 1** **SS 2** **SS 2** **SS 1**
**Join**
**Partial** **Person**
**results** **Post**
```
el-lifespan ≺ er-lifespan
```
(b) Splitting of query and Joining of results for EQ2,
similar to Plan 2 in Figure 3a
Figure 3: Query execution plans in _ranite_
_G_
to the period of overlap. The compute or scatter functions only access
messages and properties that are relevant to their current interval, and both
can be called multiple times, for different intervals, on the same vertex and
edge.
#### 4.3 Distributed Query Execution Plans
Queries can be evaluated by splitting them into smaller path query segments
that are independently evaluated left-to-right, and the results then combined. Each vertex predicate in the path query is a potential split point.
E.g., a query V 1 _E1_ _V 2_ _E2_ _V 3 can be split at V 2 into the segments:_
_−_ _−_ _−_ _−_
_V 1_ _E1_ _V 2 and V 3_ _E2_ _V 2; execution proceeds inwards, from the outer_
_−_ _−_ _−_ _−_
predicates (V 1 and V 3) to the split vertex (V 2) which joins the results. This
is illustrated in Figs. 3a (Plan 2) and 3b. A trivial split at the last vertex
predicate V 3 is the default execution of the query from left-to-right, shown
in Figure 3a (Plan 1), while an alternative split at the first vertex predicate
_V 1 evaluates this from right-to-left as V 3_ _E2_ _V 2_ _E1_ _V 1._
_−_ _−_ _−_ _−_
13
**Person**
**Post**
**V2**
**V2**
**V2**
**V1**
-----
Each split point and plan can be beneficial based on how many vertices and edges match the predicates in the graph. Intuitively, a good plan
should evaluate the most discriminating predicate first (low selectivity, few
vertex/edges match) to reduce the solution space early. A cost model, discussed in Section 5, attempts to select the best split point.
We modify our _ranite logic to handle the execution of two path seg-_
_G_
ments concurrently. E.g., for a split point V 2, in the first superstep, we
evaluate predicates V 1 _E1 and V 3_ _E2 in the same compute/init and_
_−_ _−_
scatter logic, while in the second superstep we evaluate predicate V 2, as
shown in Figure 3a (Plan 2). In the final superstep when results from both
the segments are available, we do a nested loop join to get the cross-product
of the results. This can be extended to multiple split points in the future.
#### 4.4 System Optimizations
**4.4.1** **Type-based Graph Partitioning**
Giraph by default does a hash-partitioning of the vertices of the graph by
their vertex IDs onto workers. But we use knowledge of the entity schema
types to create graph partitions hosting only a single vertex type. This helps
us eliminate the evaluation of all vertices in a partition if its type does not
match the vertex type specified in that hop of the query. This filtering is
done before the compute is called, at the partitionCompute of Giraph.
We first group vertices by type to form a typed partition each, e.g., Type A
and Type B, as illustrated in Figure 4a. But these can have skewed sizes, and
there may be too few types (hence partitions) to fully exploit the parallelism
available on the workers and their threads. So we further perform a secondlevel topological partitioning of each typed partition into p sub-partitions
using METIS [51]. This only considers the edges between vertices of the
same type, i.e., within each typed partition, and uses the edge lifespan as
their weight. This second-level partitioning can also reduce the network
messaging cost between vertices of the same type. The sub-partitions from
each typed partition are then distributed in a round-robin manner among
all the workers. So if there are w workers, t types and p sub-partition per
type, each worker is expected to have _[t][×][p]_ sub-partitions, with _p_
_w_ _w_ [of each]
type. Since each superstep typically evaluates a query predicate for a single
vertex type, this ensures load balancing of the typed sub-partitions across
all workers during a superstep execution.
In our experiments using the 100k:A-S graph, described in Section 6.1, we
observe that using type-based partitioning at the first level instead of hash
partitioning improves the average execution time for our query workloads
by 5.8 . When we combine this with METIS partitioning in the second
_×_
level, we see a further improvement of 32%. All our results we later report
use this optimization.
14
-----
G
D
B
E
**1**
**4**
**5**
**7**
**7**
**9**
**8**
**5**
**3**
**2**
**3**
**Type B**
**4**
**9**
**D**
**D**
**H**
**A**
**F**
**Input Graph** _Type-based_ **Sub-partitions**
_Partitioning_ _METIS_
(a) Two-level load-balanced partitioning of Input Graph to Workers, by type and then by
topology
H
F
**6**
**1**
**4**
**1**
**7**
**H**
**E** _Input Graph_ **B** **D** **F** **H**
C **C** **C**
**B** **E** **F** **H**
✓ _Predicate Match_ _Predicate Mismatch_ **C** **E** **F** **H**
(b) Message tree propagation during
query evaluation
A
**3**
**9**
**2**
**5**
**8**
**H**
**B**
**F**
**C**
**F**
Figure 4: Examples of system optimizations in _ranite_
_G_
**4.4.2** **Message Optimization**
Path results can have a lot of overlaps. But each partial result path is
separately maintained and sent in messages during query execution. This
redundancy leads to large message sizes and more memory. Instead, we
construct a result tree, where vertices/edges that match at a previous hop
are higher up in the tree and subsequent vertex/edge matches are its descendants. E.g., assuming a full binary tree expansion for a path query
with h hops and n = 2[h][−][1] matching paths, this reduces the result size from
(h _n) to_ (2n 1). When execution completes, a traversal of this result
_O_ _×_ _O_ _−_
tree gives the expanded result paths.
This is illustrated in Figure 4b. Here, vertices A, B and C match the
vertex and edge predicates in the first hop and send their partial result to
their neighbors. D receives the messages from A and B and evaluates itself
for the second-hop predicate. But this execution is not unique to A or B, but
rather shared across them. If D matches, rather than send a message with
two sub-paths, A _D and B_ _D, we instead send a sub-tree,_ _A, B_ _D in the_
_−_ _−_ _{_ _}−_
message, to its neighbors. Similarly, E which receives messages fro B and C
and matches for the second predicate sends a sub-tree _B, C_ _E message._
_{_ _} −_
_F receives two sub-trees as messages, evaluates itself for the third predicate_
that matches, and sends a larger sub-tree, _A, B_ _D_ _,_ _B, C_ _E_ _F_, to
_{{_ _}−_ _}_ _{{_ _}−_ _}−_
its neighbor H. G is not a match and prunes its traversal, with no messages
sent. H matches the last predicate successfully, and sends the final resulttree with H as the root to the Master, which unrolls the tree to return the
paths from H to every leaf as individual results to the client.
**4.4.3** **Memory Optimizations**
In our graph data model, all property keys and values, excluding time intervals, are strings. In Java, string objects are memory-heavy. Since keys will
often repeat for different vertices in the same JVM, we map every property
key to a byte, and rewrite the query at the Master based on this mapping.
15
C
**8**
**6**
**H**
**B**
-----
Further, for property values that repeat, such as country, we use interning
in Java that replaces individual string objects with shared string objects.
This works as the graph is read-only. Besides reducing the base memory
usage for the graph by 5%, it also allows predicate comparisons based on
_≈_
pointer equivalence.
### 5 Query Planning and Optimization
A given path query can be executed using different distributed execution
plans, each having a different execution time. The goal of the cost model
is to quickly estimate the expected execution time of these plans and pick
the optimal plan for execution. Rather than absolute accuracy of the query
execution time, what matters is its ability to distinguish poor plans with
high execution times from good plans with low execution times.
We propose an analytical cost model that uses statistics about the temporal property graph, combined with estimates about the time spent in
different stages of the distributed execution plan, to estimate the execution
time for the different plans of a given query. We first enumerate the possible
plans, contributed by each split point in the path query. The graph statistics are then used to predict the number of vertices and edges that will be
active at each superstep of query execution, and the number of vertices that
will match the predicates in this superstep and activate the next hop of the
query (superstep). Based on the number of active and matched vertices and
edges, our cost model will estimate the runtime for each superstep of the
plan. Adding these up across supersteps returns the estimated execution
time for a plan. Next, we discuss the statistics that we maintain, and the
models to predict the vertex and edge counts, and the execution time.
#### 5.1 Graph Statistics
We maintain statistics about the temporal property graph to help estimate
the vertices and edges matching a specific query predicate. Typically, relational databases maintain statistics on the frequency of tuples matching
different value ranges, for a given column (property). A unique challenge for
us is that the property values can be time variant. Hence, for each property
key present in the vertex and edge types, we maintain a 2D histogram, where
the Y axis indicates the different value ranges for the property and the X
axis the different time ranges. Each entry in the histogram has a count of
vertices or edges that fall within that value range for that time range.
E.g., Figure 5a(top) shows such a histogram for the Country property.
Its Y axis lists different country values appearing in the vertices of the
property graph, such as India, UK and US. The X axis divides the lifespan
of the graph into time intervals, say, [0, 50) in steps of 10. The cell values
indicate the number of vertices that have these property values for those time
16
-----
_10_ _20_ _30 40 50_
**_Time_**
India **9** **10** **12** **9** **14**
UK **4** **6** **9** **5** **14**
USA **2** **4** **8** **10** **12**
**Tiling**
_10_ _20_ _30 40 50_
**_Time_**
India **10**
**14**
UK
**5** **8**
USA **12**
(a) 2-D Histogram of Statistics
|9|10|12|9|14|Col6|
|---|---|---|---|---|---|
|4 6 2 4||9 5 14 8 10 12||||
|Col1|1|0|Col4|Col5|Col6|
|---|---|---|---|---|---|
||5||8|14||
|||||12||
**India : 10**
**India/UK : 14**
**USA : 12**
(b) Interval Tree for Statistics
**UK/USA : 8**
Figure 5: Query planning
Table 2: Vertex and edge count estimates per superstep and execu_tion time calculated by the model for two execution plans, for query_
_EQ2 on 100k:F-S graph_
**Plan** **SS** _ai_ _fi_ _mi_ _ai_ _fi_ _mi_ _Ti (ms)_
1 1 100k 3.7 10[−][2] 3.7k 6.2M 35M 1.3M 531
_×_
2 1.3M 7.7 10[−][4] 1k – – – 132
_×_
2 1 51M 7.7 10[−][4] 39k 273k 88M 67k 4147
_×_
2 67k 3.7 10[−][2] 2.5k – – – 35
_×_
intervals, in the entire graph. Here, 9 vertices have the Country property
value as India during time interval [0, 10) and 10 vertices have it during
[10, 20), and similarly for other countries and time intervals.
Formally, for a given property key κ, we define a histogram function
_Hκ : (val, τ_ ) →⟨f, δin, δout⟩, that returns an estimate of the frequency f of
vertices or edges which have the property value val during a time interval
_τ_, and the average in and out degrees δ of the matching vertices, which are
maintained for a vertex property.
The granularity of the value and time ranges has an impact on the size of
the statistics maintained and the accuracy of the estimated frequencies. We
make several optimizations in this regard. We use Dynamic Programming
(DP) to coarsen the ranges of the histogram along both axes to form a
_hierarchical tiling [52]._ This ensures that the frequency variance among
the individual value–time pairs in each tile is no more than a threshold.
For example, in Figure 5a(bottom), the frequencies 9, 10, 12 and 9 for the
property value India during the interval [10, 40) are close to each other
and hence tiled, i.e., aggregated and replaced by their average value 10.
Similarly, India and UK have the same frequency 14 for the interval [40, 50)
and are tiled. This reduces the number of entries that are maintained in
the histogram, i.e., the space complexity, while bounding its impact on the
accuracy of the statistics.
For important properties like vertex and edge types, out-degree and in
17
|Plan|SS|a f m i i i|a f m i i i|T (ms) i|
|---|---|---|---|---|
|1 1 100k 3.7 10−2 3.7k 6.2M 35M 1.3M 531 × 2 1.3M 7.7 10−4 1k – – – 132 ×|1|100k 3.7 10−2 3.7k ×|6.2M 35M 1.3M|531|
|---|---|---|---|---|
|Plan 1|SS 1 2|ai fi mi 100k 3.7 × 10−2 3.7k 1.3M 7.7 × 10−4 1k 4|ai fi mi 6.2M 35M 1.3M – – –|Ti (ms) 531 132|
|---|---|---|---|---|
|2|1|51M 7.7 10−4 39k ×|273k 88M 67k|4147|
||2|67k 3.7 10−2 2.5k ×|– – –|35|
**UK/USA : 5**
-----
degree, we pre-coarsen the time steps into, say, weeks, and for other properties into, say, months to reduce the size of the histogram – the actual
coarsening factor is decided based on how often the properties change in
the graph. For properties with 1000’s of enumerated values (e.g., Tag in
Figure 1), we sort them based on their frequency, cluster them into similar
frequencies, and perform tiling on these clusters. We retain a map between
property values and clusters for these, which is used to rewrite the input
query to replace the property values with these cluster IDs instead.
We use an interval tree to maintain each histogram, with each tile inserted into this tree based on its time range. The nodes of the tree will have
a set of tiles (property value ranges and their frequencies) that fall within
its time interval. The invariant for all the nodes in the tree is such that the
interval of a parent node will be after the left child (i.e., start time and/or
end time of parent’s interval is after the left child’s interval), and before
the right child. E.g., the interval tree in Figure 5b is constructed from the
2D histogram in Figure 5a(bottom). Every tile in the histogram becomes
a node or part of a node in the interval tree. We insert a tile in the right
subtree if its interval is greater than the parent node’s interval, in the left
sub-tree if it is lesser, and in the parent if it overlaps with it. To perform
a lookup, we check if the lookup interval is greater than or less than the
parent interval and prune the search space accordingly, similar to a binary
search tree. Calling the function performs a lookup in this interval tree,
_H_
and matches within the set of property ranges.
The time complexity to construct each interval tree includes the time
to aggregate the statistics from the graph, taking (n _k), where n is the_
_O_ _·_
number of vertices in the graph and k the average number of property names
or keys per vertex type. For each property key, the time taken is dominated
by the tiling step that uses DP, and takes (p[3]t[3]), where p is the number
_O_
of (clustered) values for the property key, and t the number of (coarsened)
time units they span [52]. The cost of building the interval tree is (m _t),_
_O_ _·_
where m is the number of tiles in the coarsened histogram. The lookup time
is (p _t) in the worst case; for a balanced tree the expected lookup time is_
_O_ _·_
(log m + k), where k is the number of intersecting intervals in the tree.
_O_
The raw size of the statistics for the graphs used in our experiments
ranges from 4200–5600 kB for about 13–15 property keys.
#### 5.2 Estimating the Active and Matching Vertex and Edge Counts
A query plan contains either one or two path query segments. The query
predicates on each vertex and its edges in the segment are evaluated in a
single superstep. If two path segments are present, their results are joined
at the split point. Aggregation operators, if any, are also evaluated in the
last superstep. For each segment, we estimate a count of active and match
18
-----
ing vertices and edges in each superstep, given by the recurrence relation
discussed next.
Let P = [π1, π1, ..., πn] denote the sequence of n vertex predicates, π, and
_n_ 1 edge predicates, π, for a given path query segment. Each predicate
_−_
_π has a set of property clauses CP (π) = {⟨κ, val⟩} and a temporal clause_
_CT (π) = ⟨lifespan, τ_ _⟩, where κ is a property key, val is a value to compare_
its value against, and τ is the interval to compare that vertex/edge/property’s lifespan against; and similarly for π. These clauses themselves can be
combined using AND and OR Boolean operators, as described in the query
syntax earlier.
Let σi (σi) denote the type of the vertex (edge) enforced by a clause of
predicate πi (πi). Let Vσ (Eσ) denote the set of vertices (edges) of that
type; if the vertex (edge) type is not specified in the predicate, these sets
degenerate to all vertices (edges) in the graph.
As shown in Figure 2b, each superstep is decomposed into 2 stages: calling init or compute on the active vertices to find the vertices matching
the vertex predicate, and calling scatter on the active edges (i.e., in or
out edges of the matching vertices) to identify the edges matching the edge
predicates. These in turn help identify the active vertices for the next superstep of execution. Initially, all vertices of the graph are active, but if a
type is specified in the starting vertex predicate, we can use the type-based
partitioning to limit the active vertices to the ones having that vertex type.
Let ai and mi denote the number of active and matched vertices, respectively, for vertex predicate πi with type σi, and ai and mi denote the
_number of active and matched edges, respectively, for the edge predicate πi_
with type σi. These can be recursively defined as:
_ai_ =
�
_|Vσ|_, if i = 1 (1)
min(mi−1, |Vσ|), otherwise
�
_⟨fi, δin[i]_ _[, δ]out[i]_ _[⟩]_ = _Hκ(val, τ_ )
_⟨κ,val⟩∈CP (πi)_
_⟨lifespan,τ_ _⟩∈CT (πi)_
_mi_ = _ai ×_ _|V[f]σ[i]_ _|_ (2)
_ai_ = _m[σ]i_ _in_ [+][ δ]out[i] [)] (3)
_[×][ (][δ][i]_
�
_⟨fi, −, −⟩_ = _Hκ(val, τ_ )
_⟨κ,val⟩∈CP (πi)_
_⟨lifespan,τ_ _⟩∈CT (πi)_
_fi_
_mi_ = _ai ×_ _|Vσ| × (δ[¯]in[σ]_ [+ ¯][δ]out[σ] [)] (4)
In Equation 1, we set the active vertex count in the first superstep to
be equal to the number of vertices of type σ. This reflects the localization
19
-----
of the search space in the init function to only vertices in the partitions
matching that vertex type. For subsequent supersteps, the active vertex
search space is upper-bounded by |Vσ| but is usually expected to be the
number of matching edges in the previous superstep [2], which would send a
message to activate these vertices and call its compute function.
Next, in Equation 2, we use the graph statistics to find the fraction of
vertices _fi_
_|Vσ|_ [that match the vertex predicate][ π][i][ (also called][ selectivity][) and]
multiply this with the number of active vertices to estimate the matched vertices. This is the expected matched output count from init or compute.
We use to find the selectivity by iterating through all clauses of a predi_H_
cate πi, get their frequency, average in degree and average out degree of the
vertex matches for each along with any temporal clause, and then aggregate
( ) these frequencies. The aggregation between adjacent clauses can be ei_⊗_
ther AND or OR, and based on this, we apply the following aggregation logic
for the frequencies and degrees.
�
_f_ = (f 1, f 2) =
�
_δ_ = (⟨fi, δi⟩, ...) =
�
min(f 1, f 2), if = AND
_⊗_ (5)
max(f 1, f 2), if = OR
_⊗_
�
_i_ _[f][i][ ×][ δ][i]_
(6)
�
_i_ _[f][i]_
Equation 5 returns the smaller of the frequencies while performing an AND,
and the larger of the two with an OR; the former can be an over-estimate
while the latter an under-estimate if the two properties are not statistically
independent. Equation 6 finds the weighted average of the degrees of the
vertices matching the predicates. Once the frequencies of the clauses are
aggregated, we divide it by the number of vertices of this vertex type to get
the selectivity for the vertex predicate.
Then, in Equation 3, we identify the number of edges for which scatter
will be triggered by multiplying the matched vertices with the sum of the
in and out degrees for the matching vertices δ. Lastly, in Equation 4 we
estimate the number of edges matched by the edge predicate πi. Here, we
get the edge selectivity using the frequency of edge matches returned by
the graph statistics, and normalized by the number of preceding vertices of
type σ, times the average of the in and out degrees of vertices of this type,
_δ. The edge selectivity is multiplied by the active edge count to get the_
matched edges that is expected from the scatter call. These edges will
send messages to their destination vertices, and this will feed into the active
vertex count in superstep i + 1.
E.g., Table 2 shows the cost model and statistics in action for query
_EQ2 on graph 100k:F-S that is described later in Section 6.1. It reports_
the counts for the active and matched vertex and edge counts (a, m) using
2This is in the worst case, if vertices and edges that match in the preceding hop activate
mutually exclusive vertices in the next hop.
20
-----
Equations 1–4, and the frequency of the vertices and edges (f ) as returned
by the histogram, for each superstep of two different query plans. We see
that a1 is higher for Plan 2 than Plan 1 since the plans start at different
vertex types during the init phase, and this will lead to different execution
times for this phase (ι, discussed later). The frequency f1 in Plan 1 is equal
to f2 in Plan, 2 and likewise for f2 of Plan 1 and f1 of Plan 2. This is
expected since the predicate evaluated in superstep 1 of Plan 1 is same as
that of superstep 2 in Plan 2. The cost model also estimates the messages
sent m1 to be 1.3M and 67k for the two plans. Since we assume that the
property values are independent, the selectivities remain constant. a2 = m1
for both plans since we assume that each message from a superstep is sent
to a unique vertex in the next superstep. While the compute calls for Plan
2 is higher, and the scatter calls and messages for Plan 1 is higher. The
execution time model discussed next helps decide which of these plans has
a lower estimated latency.
The clauses for time can also have comparators like,, etc. and prop_≻_ _≺_
erty clauses can have ! =. These are supported by the histogram and cost
model. E.g., we get the frequency for a operator by summing the frequen_≺_
cies for all values smaller than the given value, and for ! = by subtracting
from the total frequency the frequency of values that equal the given value.
All time-variant statistics are maintained in the histogram, while invariants
such as the count of vertex and edges of each type are maintained as part
of global statistics for the graph.
#### 5.3 Execution Time Estimate
Given the estimates of the active/matched vertices/edges in each superstep, we incorporate them into execution time models for the different
stages within a superstep to predict the overall execution time. We use
micro-benchmarks to fit a linear regression model for the execution times,
_,_ _,_ _,_ _, and_, used below. These are unique to a cluster deployment
_I_ _M_ _S_ _CC_ _IC_
of _ranite, and can be reused across graphs and queries._
_G_
As shown in Figure 2b, the init function is called on the active vertices
_a1 in the first superstep, and generates m1 outputs that affect the states_
of the interval vertex. Its execution time estimate is given by the function
_ι = I(a0, m0). For subsequent supersteps i, the compute function is called_
similarly on the active vertices, ai, to generate the matched vertices mi. This
has a slightly different execution logic since it has to process an estimated
_mi−1 input messages from the previous superstep and does not have to_
initialize data structures, unlike init. Its execution time estimate is, ci =
_M(ai, mi, mi−1). In a superstep i, scatter is called on the active edges and_
generates matched edges, with an estimated time of si = S(ai, mi). Besides
these, there are per-superstep platform overheads: for iterating over vertices
matching a given type, cci = CC(|Vσ|) in the partitionCompute phase, and a
21
-----
Table 3: Cost model coefficients for linear regression fit for each execution
phase, as used in our experiments
**Init (** **)** **Compute (** **)** **Scatter (** **)**
_I_ _C_ _S_
_a0_ _m0_ cons. _ai_ _mi_ _mi−1_ cons. _ai_ _mi_ cons.
9.4e-5 -3.1e-5 3.83 7.2e-5 3.3e-5 1.8e-5 1.63 7.9e-5 0 -3.81
**Interval** **Partition**
**Compute** **Compute**
**(** **)** **(** **)**
_IC_ _CC_
_ai_ cons. _Vσ_ cons.
-5.1e-6 8.6e-2 -8.0e-6 28.7
base overhead of ici = IC(ai) per active vertex for Graphite.
Given these, the total estimated execution time of the cost model for a
query path segment with n hops is:
|Init ( ) I|Compute ( ) C|Scatter ( ) S|
|---|---|---|
|a m cons. 0 0|a m m cons. i i i−1|a m cons. i i|
|Interval Compute ( ) IC|Partition Compute ( ) CC|
|---|---|
|a cons. i|V cons. σ|
|-5.1e-6 8.6e-2|-8.0e-6 28.7|
|---|---|
_T = (ι + s1 + cc1 + ic1) +_
_n_
�
_ci + si + cci + ici_
_i=2_
In practice, these functions are determined by fitting simple linear regression models over query micro-benchmarks performed on the cluster on
which the platform will be deployed. This is done once, and the functions
are common for different graphs and query workloads on that cluster. E.g.,
Table 3 shows the coefficients for the linear equations that we fit for these
functions, for the experiment setup in Section 6.2. Also, Table 2 shows the
estimated execution time Ti in each superstep i for the two execution plans,
using these coefficients. Plan 1 takes lesser time than Plan 2 due to the
latter taking 7.8 longer in superstep 1. This is caused by a high init
_×_
execution time, ι, since it has to evaluate 51M vertices (a1) compared to
only 100k in Plan 1. Since the total time is dominated by the init time,
our cost model will choose Plan 1 for executing of this query.
We exclude the time to perform join and aggregation (for aggregate
queries) from the cost model equation. This is based on our observation
that this time is negligible (e.g., 20–30 ms in our experiments) compared to
the overall execution time of a query (1000 ms) in most cases. In contrast,
the execution time for init and the three compute functions together take
about 900 ms. Further, the join and aggregate costs are proportional to the
result set size. Even with a large result set size, there would inevitably
be a large number of intermediate compute calls, and so the relative time
taken by join and aggregate will remain low. Avoiding their inclusion helps
keep the model concise, with only the most significant costs included. The
time taken to find the optimal split point for a query using the approach
described in this section is 2–9ms.
22
-----
Figure 6: Modified LDBC Temporal Property Graph schema used in the
evaluation
### 6 Results
#### 6.1 Workload
We use the social network benchmark from the Linked Data Benchmark
Council (LDBC) [53] for our evaluation of _ranite._ It is a community_G_
standard workload with realistic transactional path queries over a social
network property graph. There are two parts to this benchmark, a social
network graph generator and a suite of benchmark queries.
**Property Graph Datasets** The graph generator S3G2 [54] models a
social network as a large correlated directed property graph with diverse
distributions. Vertices and edges have a schema type and a set of properties
for each type. Vertex types include person, message, comment, university,
_country, etc., while edge types are follows, likes, isLocatedIn, etc. The graph_
is generated for a given number of persons in the network, and a given degree distribution of the person–follows–person edge: Altmann (A), Discrete
_Weibull (DW), Facebook (F) or Zipf (Z)._
We make two changes to the LDBC property graph generator. One,
we denormalize the schema to embed some vertex types such as country,
_company, university and tag directly as properties inside person, forum,_
_post and comment vertices._ This simplifies the data model. Two, while
LDBC vertices are assigned a creation timestamp that can fall within a
3-year period, we include an end time of to form a time interval. We
_∞_
also add lifespans to the edges incident on vertices based on their referential
integrity constraints, and replace time-related properties like join date and
_post date with the built-in lifespan property instead. The vertex and edge_
lifespans are also inherited by their properties. Figure 6 shows this modified
graph schema.
23
-----
Table 4: Characteristics of graphs used in the experiments
_Frequent Vertex Types_
**Graph** **V** **E** **Persons Posts Comments Forums**
_|_ _|_ _|_ _|_
_Static Temporal Graphs_
**10k:DW-S** 5.5M 20.8M 8.9k 1.1M 4.3M 82k
**100k:Z-S** 12.1M 23.9M 89.9k 7.4M 2.3M 815k
**100k:A-S** 25.4M 78.2M 89.9k 8.7M 15.7M 816k
**100k:F-S** 52.1M 217.6M 100k 12.6M 38.3M 996k
_Dynamic Temporal Graphs_
**10k:DW-D** 6.6M 29.3M 10k 1.4M 5.1M 100k
**100k:Z-D** 15.2M 37.1M 100k 9.3M 4.8M 995k
**100k:A-D** 32.0M 112.2M 100k 10.8M 20.1M 995k
**100k:F-D** 52.0M 216.5M 100k 12.6M 38.2M 995k
_Frequent Edge Types_ **Unrolled**
**Graph** **hasMember[∗]** **hasCreator[†]** **Properties[#]**
_Static Temporal Graphs_
**10k:DW-S** 3.3M 4.3M 35M
**100k:Z-S** 1.5M 2.3M 60M
**100k:A-S** 12.7M 15.8M 157M
**100k:F-S** 52.2M 38.4 M 325M
_Dynamic Temporal Graphs_
**10k:DW-D** 7.2M 5.1M 30M
**100k:Z-D** 3.2M 4.8M 57M
**100k:A-D** 25.6M 20.1M 132M
**100k:F-D** 51.8M 38.3M 222M
_∗_ _forum_hasMember_person_ _† comment_hasCreator_person_
# Unrolls multi-valued properties into individual ones
However, this is still only a static temporal property graph. To address
this, we introduce temporal variability into the properties, worksAt, country
and hasInterest of the person vertex. For worksAt, we generate a new property every year using the LDBC distribution; the country is correlated with
_worksAt, and hence updated as well. We update the hasInterest property_
based on the list of tags for a forum that a person joins, at different time
points.
Table 4 shows the vertex and edge counts, the number of vertices of
each type and the total number of property values, for graphs we generate
with 10[4] (10k) or 10[5] (100k) persons, with different distributions (DW, Z,
A, F), and with static (S, top) and dynamic (D, bottom) properties. As we
see, the Comments type dominates the number of vertices, with up to 400
comments per person over a 3 year period, followed by about 100 Posts per
person. The most frequent edge types are forum_hasMember_person and
_comment_hasCreator_person, while each person Follows 10.2 other friends_
on average. Properties such as hasInterest for person and hasTag for com
24
|Graph|V | ||E | ||Frequent Vertex Types Persons Posts Comments Forums|
|---|---|---|---|
|10k:DW-S 100k:Z-S 100k:A-S 100k:F-S|5.5M 12.1M 25.4M 52.1M|20.8M 23.9M 78.2M 217.6M|8.9k 89.9k 89.9k 100k|1.1M 7.4M 8.7M 12.6M|4.3M 2.3M 15.7M 38.3M|82k 815k 816k 996k|
|---|---|---|---|---|---|---|
|10k:DW-D 100k:Z-D 100k:A-D 100k:F-D|6.6M 15.2M 32.0M 52.0M|Col3|29.3M 37.1M 112.2M 216.5M|10k 100k 100k 100k|1.4M 9.3M 10.8M 12.6M|5.1M 4.8M 20.1M 38.2M|Col8|100k 995k 995k 995k|
|---|---|---|---|---|---|---|---|---|
|Graph||Frequent Edge Types hasMember∗ hasCreator†|||||Unrolled Properties#||
|10k:DW-S 100k:Z-S 100k:A-S 100k:F-S|3.3M 1.5M 12.7M 52.2M|4.3M 2.3M 15.8M 38.4 M|35M 60M 157M 325M|
|---|---|---|---|
|10k:DW-D 100k:Z-D 100k:A-D 100k:F-D|7.2M 3.2M 25.6M 51.8M|5.1M 4.8M 20.1M 38.3M|30M 57M 132M 222M|
|---|---|---|---|
-----
10[8]
10[7]
10[6]
10[5]
10[4]
10[3]
10[2]
10[1]
10[0]
|Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|
|---|---|---|---|---|---|---|---|---|
||||||||||
||7|.5K|||||||
||||806||815|3 500|.1K||
|2|6|||4|||||
Q1 Q2 Q3 Q4 Q5 Q6 Q7 Q8
Query Type
10[8]
10[7]
10[6]
10[5]
10[4]
10[3]
10[2]
10[1]
10[0]
|Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|
|---|---|---|---|---|---|---|---|---|
||||||||||
||||||||||
|||606|16||103|73|265||
|1|2|||1|||||
|Q1 Q||2 Q3 Q4 Q5 Q6 Q7 Q8 Query Type|||||||
(a) 10k:DW
(b) 100k:Z
10[8]
10[7]
10[6]
10[5]
10[4]
10[3]
10[2]
10[1]
10[0]
|Col1|Col2|5.1K|324|Col5|673|6|96 81|Col9|
|---|---|---|---|---|---|---|---|---|
|7|7|||2|||||
|1|42 3|0|3|7K 2|2. 1|5. 5K|2K 4.|7K|
|---|---|---|---|---|---|---|---|---|
||||||||||
Q1 Q2 Q3 Q4 Q5 Q6 Q7 Q8
Query Type
5.2K4.7K
2.5K
142
30 21
Q1 Q2 Q3 Q4 Q5 Q6 Q7 Q8
Query Type
4.3M
10[8]
10[7]
10[6]
10[5]
10[4]
10[3]
10[2]
10[1]
10[0]
(c) 100k:A
(d) 100k:F
Figure 7: Box and whiskers distribution plot of the result set count for the
100 instances of each non-aggregate query type. Q1–Q7 are reported on the
static graphs while Q8 is on the dynamic graphs. The median result set
count is labeled.
ment take up the most space since they are multi-valued, with an average
of 23 interests per person and 1.22 tags per comment.
**Query Workload** We select a subset of query templates provided in the
LDBC query workload [53] that conform to a linear path query, and adapt
them for our temporal graphs. Table 5 describes the query templates. These
are either from the Business Intelligence (BI) or the Interactive Workload
_(IW). We also include two additional query templates Q5 and Q6 to fully_
exercise our query model. Also, query template Q8 depends on worksAt
which is a dynamic property and so it is only evaluated for dynamic temporal
graphs.
Each template has some parameterized property or time value. We generate 100 query instances for each template by randomly selecting a value for
the parameters, evaluating the query on the temporal graph, and ensuring
that there is at least 1 valid result set in most cases. Query instances are
generated for both the static and dynamic graphs. In addition to these non
25
-----
Table 5: Description of query workload used in the experiments
**Property** **Time** **Has ER** **Description of path to**
**Query LDBC ID Hops** **Predi-** **Predi-** **Predi-** **find (Parameterized property**
**cates** **cates** **cate?** _values are underlined)_
Two messages with different
**Q1** BI/Q9 3 4 1 Yes belong to the same forum
time ordering between the messages
A person with a given
**Q2** BI/Q10 2 6 1 No _message with the same_
given date.
A person from a given
**Q3** BI/Q16 3 6 1 Yes commented or liked a
_person from another given_
Mutual friendships between three
_persons, but with a time-respecting_
**Q4** BI/Q17 4 3 2 Yes
order in which they befriend each
other.
A person posts a message
given tag to a forum
**Q5** – 5 7 3 Yes time offset, they post another
_message to the same_
different tag.
A person with a specific
**Q6** – 5 7 1 Yes replies to a post after another
_person replies to it._
A person posts a message
outside their home country
**Q7** BI/Q23 4 5 3 Yes befriends another person
_person then posts another_
from outside their home
Two persons working in different
**Q8** IW/Q11 3 3 1 Yes _companies have a common_
at a time-point.
_aggregate queries, we also create another workload that includes a count_
temporal aggregate operator to these query templates, i.e., it will group the
results of the original query by the first vertex and its time intervals, and
return the count for each vertex-interval. This helps evaluate the performance of aggregate queries. For brevity, we limit these aggregate queries to
the two largest graphs, 100k:A and 100k:F. Figures 7a, 7b, 7c and 7d show
the distribution of the result set count for the (non-aggregate) queries on
the different graphs in our workload.
These illustrate the expressivity of our query model, and ability to intuitively extend it to the time domain. The query length varies between
2 and 5 hops, allowing us to evaluate the cost model and _ranite perfor-_
_G_
26
|Query|LDBC ID|Hops|Propert Predi- cates|y Time Predi- cates|Has ER Predi- cate?|Description of path to find (Parameterized property values are underlined)|
|---|---|---|---|---|---|---|
|Col1|Col2|Col3|cates|cates|cate?|values are underlined)|
|---|---|---|---|---|---|---|
|Q1|BI/Q9|3|4|1|Yes|Two messages with different tags belong to the same forum, with a time ordering between the messages|
|Q2|BI/Q10|2|6|1|No|A person with a given tag creates a message with the same tag after a given date.|
|Q3|BI/Q16|3|6|1|Yes|A person from a given country has commented or liked a post before a person from another given country.|
|Q4|BI/Q17|4|3|2|Yes|Mutual friendships between three persons, but with a time-respecting order in which they befriend each other.|
|Q5|–|5|7|3|Yes|A person posts a message with a given tag to a forum and, after a time offset, they post another message to the same forum with a different tag.|
|Q6|–|5|7|1|Yes|A person with a specific gender replies to a post after another person replies to it.|
|Q7|BI/Q23|4|5|3|Yes|A person posts a message from outside their home country, then befriends another person, and that person then posts another message from outside their home country.|
|Q8|IW/Q11|3|3|1|Yes|Two persons working in different companies have a common friend at a time-point.|
-----
mance for different lengths. All the vertex types appear as predicates in our
workload. The queries filter on both single-valued properties like country
and lastName, and multi-valued properties like hasInterest and hasTag. All
edge types except forum_hasModerator_person are used in the workload.
7 out of the 8 query types have ETR predicate and all the queries have at
least 1 time predicate. They are diverse with respect to result sizes too, as
shown in Figure 7a, 7b, 7c and 7d, and the result counts span several orders
of magnitude, from 10[0]–10[4].
In our experiments, each query is given an execution budget of 600 secs,
after which it is terminated and marked as failed. The average execution
times are only reported on the successful queries. We verify the correctness
of all queries on _ranite and baseline platforms. For the performance eval-_
_G_
uations, the queries only return the count of the result sets for timeliness.
#### 6.2 Experiment Setup
Our commodity cluster has 18 compute nodes, each with one Intel Xeon E52620 v4 CPU with 8 cores (16 HT) @ 2.10GHz, 64 GB RAM and 1 Gbps
Ethernet, running CentOS v7. For some shared-memory experiments on
other baseline graph platforms, we also use a “big memory” head node with
2 similar CPUs and 512 GB RAM. _ranite is implemented over our in-house_
_G_
Graphite v1.0 ICM platform [21], Apache Giraph v1.3.0, Hadoop v3.1.1 and
Java v8. By default, our distributed experiments use 8 compute nodes
in this cluster, run one _ranite Worker JVM per machine with 8 threads_
_G_
per Worker, and have 50 GB RAM available to the JVM. The graphs are
initially loaded into _ranite from JSON files stored in HDFS, with their_
_G_
pre-computed cost model statistics, and the query workloads run on this
distributed in-memory copy of the graph.
#### 6.3 Baseline Graph Platforms
We use the widely-used Neo4J Community Edition v3.2.3 [47] as a baseline graph database to compare against. This is a single-machine, singlethreaded platform. We use three variants of this. One specifies the workload
queries using the community standard Gremlin query language (N4J-Gr, in
our plots), and the other uses Neo4J’s native Cypher language (N4J-Cy).
Both these variants run on a single compute node with 50 GB heap size. A
third variant uses Cypher as well, but is allocated 8 50 = 400 GB of heap
_×_
space on the head node (N4J-Cy-M ). As graph platforms are often memory
bound, this configuration matches the total distributed memory available to
our _ranite setup by default. We build indexes on all properties in Neo4J._
_G_
There are few open source distributed graph engines available. Janus_Graph [8], a fork from Titan, is popular, and uses Apache Spark v2.4.0 as a_
distributed backend engine to run Gremlin queries (Spark, in our plots). It
27
-----
uses Apache Cassandra v2.2.10 to store and access the input graph. Spark
runs on 8 compute nodes with 1 Worker each and 50 GB heap memory
per Worker. Cassandra is deployed on 8 additional compute nodes. This is
based on the recommended configuration for JanusGraph on Cassandra [3].
Spark initially loads the graph from Cassandra into its distributed memory present on its 8 compute nodes. This load time is not considered as
part of the query execution time. So effectively, only the 8 Spark nodes
are used during query execution. For all baselines, we follow the standard
performance tuning guidelines provided in their documentation [4 5].
Since these platforms do not natively support temporal queries over
_dynamic temporal graphs, we transform the graphs into a static tempo-_
ral graph using techniques described by Wu, et al. and used earlier by
Graphite [21, 43]. This static property graph converts the time-intervals on
vertices and edges of the original interval graph into an expanded set of
vertices and edges that are valid for just a single discrete time point. This
lets us adapt the query to operate on the static graph, albeit a bloated one.
Also, temporal aggregation is not feasible internally on these platforms. So
we perform the final aggregation at the client side for queries with an aggregate operator. JanusGraph/Spark is unable to load these two large graphs
in-memory, and hence was not evaluated for the aggregate queries. The
results from all platforms for all queries are verified to be identical.
#### 6.4 Effectiveness of Cost Model
We first evaluate the effectiveness of _ranite’s cost model in identifying the_
_G_
optimal split point for the distributed query execution. For each query type
(template), we execute its 100 query instances using all their possible query
_plans, i.e., every possible split point is considered for each query. From the_
execution time of all plans for a query, we pick the smallest as its optimal
_plan. We compare this against the plan selected by our cost model, and_
report the % of excess execution time that our model-selected plan takes
above the optimal plan. This is the effective time penalty when we select a
sub-optimal plan.
Figure 8a shows a violin plot of the the distribution of this % excess
time over optimal, for the different fixed split points 1–4 executed for the
100 queries of type Q4 (non-aggregate) on graph 100k:A-S, compared to the
plan selected by our cost model (CM) – lower this value, closer to optimal
the performance. We see that the execution time varies widely across the
plans, with some taking 8 longer than optimal. Also, some split points like
_×_
2 and 3 are in general better than the others, but among them, neither is
consistently better. In contrast, our cost model plan has a low mean excess
3https://docs.janusgraph.org/storage-backend/cassandra/
4https://neo4j.com/docs/operations-manual/3.2/performance/
5https://docs.janusgraph.org/advanced-topics/hadoop/
28
-----
(a) Distribution of queries that exceed
the optimal plan’s time by a % (Y axis),
for each fixed plan and for the cost
model, for 100k:A-S graph on Q4 type
queries
10[1]
10[0]
|Col1|Col2|Col3|Q1|Col5|Col6|Col7|Col8|Col9|Col10|Col11|Col12|Col13|Col14|Col15|Col16|
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
||||Q2 Q3|||||||||||||
||||Q4|||||||||||||
|||||||||||||||||
||||Q5 Q6 Q7|||||||||||||
|||||||||||||||||
|||||||||||||||||
|||||||||||||||||
|||||||||||||||||
|||||||||||||||||
|||||||||||||||||
|||||||||||||||||
|||||||||||||||||
Best Second Third Fourth
Split Point
(c) Ratio of estimated average execution
cost of the other plans relative to the optimal plan, for all query types of 100k:A_S graph_
(b) Actual vs. Cost model estimated execution time for all query instances of
_100k:A-S graph, with correlation coeffi-_
cient of ρ = 0.87
Optimal 2nd Best Rest
100
80
60
40
20
0
(d) Cost Model Accuracy. % of times
the optimal plan, 2[nd] best plan and
other plans were selected by our model
for all graphs
Figure 8: Effectiveness of cost model in picking the best plan for nonaggregate queries
time of 2.9%, relative to 12.2% and 6.9% excess time taken by these other
split points. Also, it is not possible to a priori find a single fixed split point
which is generally better than the rest, without running the queries using
all split points. These motivate the need for an automated analytical cost
model for query plan selection.
We analyze the accuracy of the cost model for 100k:A, the second largest
graph, in more detail, and its impact on the execution cost. First, Figure 8b
shows a scatter plot between the actual and the model-estimated execution
29
-----
Table 6: % excess time spent over the Optimal plan by the Cost Model selected plan, for
different query percentiles of each query type
(a) 100k:A-S
_%le_ _Q1 Q2 Q3 Q4 Q5 Q6 Q7_
75 1.8 0 2.2 0 0 0 0
90 6.8 0 12.6 0 0 0 0
95 8.5 0 24.6 0 56 0 0
99 17.6 0 47.1 0 123 0 195
(c) 100k:A-S (Temporal Aggregate)
_%le Q1 Q2 Q3 Q4 Q5 Q6 Q7_
75 0 0 0 0 0 0 0
90 5.7 0 20 0 19 0 0
95 6.3 0 24 0 24 0 0
99 8.3 0 30 0 52 0 0
Optimum 2nd Best Rest
80
60
40
20
0
Q1 Q2 Q3 Q4 Q5 Q6 Q7
Query Type
(a) 100k:A-S
Optimum 2nd Best Rest
80
60
40
20
0
Q1 Q2 Q3 Q4 Q5 Q6 Q7
Query Type
(c) 100k:A-S, Temporal Agg.
(b) 100k:A-D
_%le Q1 Q2 Q3 Q4 Q5 Q6 Q7 Q8_
75 0 0 0 0 0 0 0
90 0 0 42 0 7.1 0 0 59
95 2.4 0 124 66 8.8 0 0 112
99 3.6 0 198 191 12 0 0 277
(d) 100k:A-D (Temporal Aggregate)
_%le Q1 Q2 Q3 Q4 Q5 Q6 Q7 Q8_
75 0 0 0 0 0 0 0 0
90 4 0 6 0 28 0 0 57
95 12 28 21 132 39 0 0 175
99 18 145 84 166 57 0 0 643
Optimum 2nd Best Rest
100
80
60
40
20
0
Q1 Q2 Q3 Q4 Q5 Q6 Q7 Q8
Query Type
(b) 100k:A-D
Optimum 2nd Best Rest
100
80
60
40
20
0
Q1 Q2 Q3 Q4 Q5 Q6 Q7 Q8
Query Type
(d) 100k:A-D, Temporal Agg.
|%le|Q1|Q2|Q3|Q4|Q5|Q6|Q7|
|---|---|---|---|---|---|---|---|
|75|1.8|0|2.2|0|0|0|0|
|90|6.8|0|12.6|0|0|0|0|
|95|8.5|0|24.6|0|56|0|0|
|99|17.6|0|47.1|0|123|0|195|
|%le|Q1|Q2|Q3|Q4|Q5|Q6|Q7|Q8|
|---|---|---|---|---|---|---|---|---|
|75|0|0|0|0|0|0|0|0|
|90|0|0|42|0|7.1|0|0|59|
|95|2.4|0|124|66|8.8|0|0|112|
|99|3.6|0|198|191|12|0|0|277|
|%le|Q1|Q2|Q3|Q4|Q5|Q6|Q7|
|---|---|---|---|---|---|---|---|
|75|0|0|0|0|0|0|0|
|90|5.7|0|20|0|19|0|0|
|95|6.3|0|24|0|24|0|0|
|99|8.3|0|30|0|52|0|0|
|%le|Q1|Q2|Q3|Q4|Q5|Q6|Q7|Q8|
|---|---|---|---|---|---|---|---|---|
|75|0|0|0|0|0|0|0|0|
|90|4|0|6|0|28|0|0|57|
|95|12|28|21|132|39|0|0|175|
|99|18|145|84|166|57|0|0|643|
Figure 9: Cost Model Accuracy. % of times the optimal plan, the second
best plan and the other plans were selected by our model
30
-----
time for the 100k:A-S static graph; the plot has 2500 points. Overall, we
_≈_
see a high correlation coefficient of ρ = 0.87. There is an over-estimation
for Q7 (maroon) due to an inaccurate estimation of the number of matching
edges in the second hop, and under-estimates for Q1 to Q5. But Q6 (purple)
shows a high correlation of ρ = 0.94.
Given these execution time inaccuracies of the model, we examine its
effect on: (1) picking the optimal execution plan, and (2) on the latency
penalty when it does not pick the optimal plan. Figures 9 show the fraction
of times the cost model selects the optimal plan, the second best plan, and
the rest of the plans, for the static and dynamic variants of 100k:A, and for
non-aggregate and aggregate queries. We also have corresponding data in
Tables 6 which report for different query types (columns), and for different
percentiles of their queries (rows), what is the % excess execution time over
the optimal spent by the plan chosen by the cost model.
For the non-aggregate queries, the best or the second best plan were
selected over 97% of the time across all queries, as seen in Figures 9a and 9b.
For queries Q2, Q4, Q6 and Q7, the optimal plan was chosen 99% of the
time. In Q2, this is due to a short query length of 2 that reduces the
cumulative errors in the model, as well as a high difference in cost between
the best and the second best plans. This is seen in Figure 8c, which gives the
ratio of the 2[nd], 3[rd] and 4[th] best plan relative to the optimal. For Q2, the
best plan evaluates the person vertices first, which are 500 fewer than the
_×_
_message vertices evaluated first by the other plan. As a result, the optimal_
execution time is 10 smaller than the other and the model easily selects
_×_
the former plan. Similarly, Q6 also exhibits a high difference in cost between
the optimal plan and the remaining three. But the top two plans for queries
_Q4 and Q7 have a similar cost. For Q4, starting at either ends causes a high_
fan-out and hence the plans that start at the two intermediate hops have a
lower, but similar, cost. In such cases, as Figures 9a and 9b show, we may
occasionally select the second best plan.
However, the consequence of choosing the second best plan on the actual
execution latency is low when the top-2 plans have a similar model cost. In
fact, for 100k:A-S, we see from Table 6a that the execution time of the
model-selected plan is within 2% of the optimal execution time for the 75[th]
percentile query, within a query type, and within 13% for the 90[th] percentile
query. Its only at the 95[th] percentile query that we see higher penalties of
8–56% for 3 of the 7 query types. Even for the dynamic graph 100k:A-D,
6 of the 8 query types have negligible time penalties at the 90[th] percentile
query in Table 6b, while two, Q3 and Q8, have higher penalties of 42–59%.
The sub-optimal behavior happens when the execution model predicts a
similar cost for the top-2 plans but selects the actual second-best, and the
observed runtime for the second-best is much worse than the best. E.g.,
for the 100k:A-S graph, the difference in actual execution cost between the
optimum and second best plans for query Q3 is 18%. This causes the model
31
-----
to select the second best plan 28% of the time, and causes 5% of the
_≈_ _≈_
queries to take 25% or longer to execute than the optimal plan.
We see similar trends for the temporal aggregate queries as well, in Figures 9c and 9d, and Tables 6c and 6d. The models predict the same costs
for these aggregate queries since it ignores the aggregate operation and join
costs due to their negligible overheads. Despite that, these queries perform
on par or better than the equivalent queries without the aggregation step.
In fact, this is broadly applicable to all the graphs, as observed in Figure 8d.
It reports that across all queries and graphs evaluated, our cost model picks
the best (optimal) or the second best plan over 95% of the time.
In summary, the cost model is accurate when the query is of shorter
length, and accurate enough to distinguish between the similar good plans
and the rest when certain predicate have high cardinalities. So we predominantly pick a plan that is optimal, or has an execution time that is
close to the optimal plan. Thus, while our cost model is not perfect, it is
accurate enough to discriminate between the better and the worse plans,
and consequently reduce the actual query execution time.
#### 6.5 Comparison with Baselines
Figures 10 show the average execution time on _ranite and the baseline_
_G_
platforms (Y axis, log scale) for the different non-aggregate query types (X
axis) for the static temporal graphs, and Figures 11 for the dynamic temporal
_graphs. Only queries that complete in the 600 sec time budget are plotted._
As Table 7 shows, Janus/Spark did not run (DNR) for several larger graphs
due to resource limits when loading the graph in-memory from Cassandra.
32–79% of queries did not finish (DNF) within the time budget on Neo4J
for 100k:F-S, the largest graph. _ranite completes all queries on all graphs,_
_G_
often within 1 sec. For the largest graph 100k:F-S, _ranite uses 16 nodes_
_G_
to ensure that the graph fits in distributed memory.
The bar plots show that _ranite is much faster than the baselines, across_
_G_
_all graphs and all query types, except for Q5 on the smallest graph, 10k:DW-_
_S. On average, we are 149_ faster than N4J-Cy-M, 192 faster than N4J_×_ _×_
Cy, 154 faster than N4J-Gr and 1140 faster than Spark. Other than
_×_ _×_
the largest graph, _ranite completes on an average within 500 ms for all_
_G_
static graphs and most query types, and on an average within 1000 ms for
_100k:F-S and all the dynamic graphs._
Focusing on specific query types for the largest static temporal graph,
_100k:F-S, Q2 takes the least time for_ _ranite due to its short path length_
_G_
of 2. The left-to-right execution by the baseline platforms is the optimal
query plan, but we are still able to out-perform them due to the parallelism
provided by partitioning. _ranite takes_ 5 secs for Q3 due to the huge
_G_ _≈_
number of results, 5.9M on average. But this query does not even com_≈_
plete for N4J-Cy and Spark. _ranite’s tree-based result structure is more_
_G_
32
-----
10[7]
10[5]
10[7]
10[5]
10[3]
10[1]
10[3]
10[1]
10[7]
10[5]
10[7]
10[5]
10[3]
10[1]
|Spark N4J-Cy-M|N4J-Gr N4J-Cy Granite|
|---|---|
|||
|||
|Q1 Q2 Q3 Q Query (a) 10k:D|4 Q5 Q6 Q7 Type W-S|
|Spark N4J-Cy-M|N4J-Gr N4J-Cy Granite|
|||
|R R R R|R R R F|
|DN DN DN DN|DN DN DN DN|
|7|Col2|Col3|
|---|---|---|
||Spark N4J-Cy-M|N4J-Gr N4J-Cy Granite|
||||
||||
|Q1 Q2 Q3 Q Quer (b) 100k:|||
||Spark N4J-Cy-M||
||||
||R R R F R F||
||DN DN DN DN DN DN||
Q1 Q2 Q3 Q4 Q5 Q6 Q7
Query Type
Q1 Q2 Q3 Q4 Q5 Q6 Q7
Query Type
10[3]
10[1]
(c) 100k:A-S
(d) 100k:F-S
Figure 10: Comparison of average execution time of _ranite with baseline_
_G_
systems for non-aggregate query types, on Static Temporal Graphs
compact, reducing memory and communication costs. Q4 for this graph is
also 89–112 better in _ranite than the baselines, with large result sizes_
_×_ _G_
of 72k on average. Here, there is a rapid fan-out of matching vertices
_≈_
followed by a fan-in as they fail to match downstream predicates, leading to
high costs. _Q7 queries are able to complete only in_ _ranite and not on the_
_G_
baseline platforms. This query has an optimal split point of 1 or 2 which
is not adopted by the baselines. In fact, baselines use the worst possible
left-to-right plan, which we see is 4 slower than the optimal for _ranite._
_×_ _G_
_ranite is also consistently better for the dynamic graphs. Similar to_
_G_
the static graphs, the only time that our average query time is slower than
a baseline is for Q5 on 10k:DW-D. Here, the default left-to-right execution
is near-optimal, and the query has a low traversal fan-out and < 10 results.
So the baselines are in an ideal configuration while _ranite has overheads_
_G_
for distributed execution.
_Neo4J using Cypher, on the single compute node (N4J-Cy) and the big_
memory node (N4J-Cy-M), are the next best to _ranite. The large memory_
_G_
variant gives similar performance as the regular memory one for the smaller
graphs, but for larger graphs like 100k:A and 100k:F, it out-performs. For
33
-----
10[7]
10[5]
10[7]
10[5]
10[3]
10[1]
10[3]
10[1]
Q1 Q2 Q3 Q4 Q5 Q6 Q7 Q8
Query Type
(b) 100k:Z-D
10[7]
10[5]
10[7]
10[5]
10[3]
10[1]
|Col1|Col2|N4J- Gra|Col4|Gr nite|Col6|N|Col8|4J-Cy|Col10|Col11|Col12|N4J-|Col14|Cy-M|Col16|
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
|||||||||||||||||
|||||||||||||||||
|Q|1 Q (||2 Q a) 1||3 Q Que 0k:||4 Q ry T DW||5 Q ype -D||6 Q||7 Q||8|
|||N G|4J- ra|Gr nit|e||N|4J|-Cy|||N|4J-|Cy|-M|
|||||||||||||||||
|||||F|F|||||F|F|||F|F|
|||||DN|DN|||||DN|DN|||DN|DN|
|N4J-Cy N4J-|Col2|Cy-M|Col4|
|---|---|---|---|
|||||
|||||
|Q4 Q5 Q6 Q uery Type 00k:Z-D|7 Q||8|
|N4J-Cy N G|4J- ra|Cy nit|-M e|
|||||
|FF F FF|F|FF||
|DDNN DN DDNN|DN|DDNN||
Q1 Q2 Q3 Q4 Q5 Q6 Q7 Q8
Query Type
N4J-Gr N4J-Cy N4J-Cy-M
Granite
Q1 Q2 Q3 Q4 Q5 Q6 Q7 Q8
Query Type
10[3]
10[1]
(c) 100k:A-D
(d) 100k:F-D
Figure 11: Comparison of average execution time of _ranite with baseline_
_G_
systems for non-aggregate query types, on Dynamic Temporal Graphs
the latter graph, N4J-Cy could not finish several query types. Though Neo4J
uses indexes to help filter the vertices for the first hop, query processing for
later hops involves a breadth first traversal and pruning of paths based on the
predicates. There are also complex joins between consecutive edges along the
path to apply the temporal edge relation. These affect their execution times.
Gremlin and Cypher variants of Neo4J are comparable in performance, with
no strong performance skew either way. Interestingly, the Gremlin variant of
Neo4J is able to run most query workloads for all graph, albeit with slower
performance.
The Janus/Spark distributed baseline takes the most time for all these
queries. This is despite omitting its initial graph RDD creation time (
_≈_
80 secs). _ranite persists the graph in-memory across queries. Despite us-_
_G_
ing distributed machines, Spark is unable to load large graphs in memory
and often fails to complete execution within the time budget. A similar
challenge was seen even for alternative engines like, Hadoop, used by JanusGraph and Spark was the best of the lot.
In the bar plots, we also show a black bar for the single-machine baselines, which is marked at the 1/8[th] execution time-point – this shows the
34
-----
10[7]
10[5]
10[7]
10[5]
10[3]
10[1]
Q1 Q2 Q3 Q4 Q5 Q6 Q7
Q1 Q2 Q3 Q4 Q5 Q6 Q7
10[3]
10[1]
10[7]
10[5]
10[7]
10[5]
10[3]
10[1]
|N4J-Gr Granite|N4J-Cy|N4J|-Cy-M|Col5|
|---|---|---|---|---|
||||||
|||||F|
|||||DN|
|Q1 Q2 Q Q (a) 1|3 Q4 Q5 uery Type 00k:A-S|Q6||Q7|
|N4J-Gr Granite|N4J-Cy|N4J-|C|y-M|
||||||
||||||
|F||F||F|
|DN||DN||DN|
|N G|4J-Gr ranite|N|4J-Cy|N4J|-Cy-M|
|---|---|---|---|---|---|
|||||||
||||||FFF|
||||||DDDNNN|
|Q1 Q|2 Q Q (b) 1|3 Q uery 00k:|4 Q5 Type F-S|Q6 Q7||
|N4 Gr|J-Gr anite|N|4J-Cy|N4J-|Cy-M|
|||||||
|||||||
||FF|FF|F|FFF|FF|
||DDNN|DDNN|DN|DDDNNN|DDNN|
Q1 Q2 Q3 Q4 Q5 Q6 Q7 Q8
Query Type
Q1 Q2 Q3 Q4 Q5 Q6 Q7 Q8
Query Type
10[3]
10[1]
(c) 100k:A-D
(d) 100k:F-D
Figure 12: Comparison of average execution time of _ranite with baseline_
_G_
systems for Temporal Aggregate query types
theoretical time that would be taken by these platforms if they had perfect
parallel scaling on 8 machines, though they do not support parallel execution. As we see, _ranite is often able to complete its execution within_
_G_
that mark, showing that our distributed engine shows scaling performance
comparable or better than highly optimized single-machine platforms, even
if they had ideal scaling.
Lastly, we compare the performance of temporal aggregate queries for the
two largest static and dynamic graphs, 100k:A and 100k:F. Their execution
times on the different platforms are shown in Figure 12. For the static
graphs, we observe from Figures 12a and 12b that _ranite is much faster_
_G_
than all the baselines for most query types. On average, we are 165 faster
_×_
than N4J-Cy-M, 175 faster than N4J-Cy and 95 faster than N4J-Gr. This
_×_ _×_
is 10 faster even when compared with the perfect scaling extrapolation for
_×_
the baselines.
These temporal aggregate queries are slower compared to their nonaggregate equivalents. Specifically, for 100k:A-S, _ranite takes 64% (_
_G_ _≈_
315 ms) more on average while the baseline platforms on average are 56%
(N4J-Cy), 42% (N4J-Gr) and 78% (N4J-Cy-M) slower, which translates to
35
-----
Table 7: % of queries that complete within 600 seconds for different platforms on the temporal graphs
**Graph** **Spark** **N4J-Gr** **N4J-Cy** **N4J-Cy-M** _ranite_
_G_
_Static Graphs, Non-aggregate queries_
10k:DW 100 99 99 80 100
100k:Z 93 90 100 100 100
100k:A DNR 100 90 98 100
100k:F DNR 66 21 68 100
_Dynamic Graphs, Temporal aggregate queries_
100k:A DNR 98 65 99 100
100k:F DNR 60 68 65 100
**Graph** **Spark** **N4J-Gr** **N4J-Cy** **N4J-Cy-M** _ranite_
_G_
_Dynamic Graphs, Non-aggregate queries_
10k:DW DNR 96 98 96 100
100k:Z DNR 100 90 98 100
100k:A DNR 97 46 46 100
100k:F DNR 30 20 75 100
_Dynamic Graphs, Temporal aggregate queries_
100k:A DNR 95 46 99 100
100k:F DNR 36 19 78 100
24–53 secs longer, per query. The baselines’ time increase considerably
_≈_
due to the additional overhead of sending the entire result set back to client
to perform the temporal aggregation, as opposed to just sending the total
number of results for the non-aggregate queries. Since _ranite does this_
_G_
natively in a distributed manner, we mitigate this cost.
_ranite completes all these queries when executed using the plan se-_
_G_
lected by the cost model (Table 7). The baseline platforms are only able to
complete, on average, 79% (N4J-Gr), 67% (N4J-Cy) and 82% (N4J-Cy-M)
of the queries on the static graphs, and this is worse for the dynamic graphs,
ranging from 33%–89%.
Also, as Figures 12a and 12b show, we take under 1 sec to run all queries
on 100k:A-S except Q7, and within 2.1 secs for all queries on the largest
graph, 100k:F-S, except Q3 – query Q3 takes longer due to the large result
count of 4.3M (Figure 7d). For dynamic graphs, we take under 3 secs
_≈_
for all queries on 100k:A-D except Q4 and Q7, and within 9.2 secs for all
queries on the largest graph, 100k:F-D, except Q3 (Figures 12c and 12d).
None of the baseline platforms could finish query type Q7 for 100k:F-S or
_100k:F-D. This query starts and ends with the Post vertex type, which has_
a high cardinality. Also, these queries on the baseline platforms need to
accumulate all the results for client-side aggregation. Both of these lead to
memory-pressure for the larger graphs.
36
|Graph|Spark N4J-Gr N4J-Cy N4J-Cy-M ranite G|
|---|---|
|Graph|Spark N4J-Gr N4J-Cy N4J-Cy-M Granite|
|---|---|
||Static Graphs, Non-aggregate queries|
|10k:DW|100 99 99 80 100|
|100k:Z|93 90 100 100 100|
|100k:A|DNR 100 90 98 100|
|100k:F|DNR 66 21 68 100|
|10k:DW 100k:Z 100k:A 100k:F|Static Graphs, Non-aggregate queries 100 99 99 80 100 93 90 100 100 100 DNR 100 90 98 100 DNR 66 21 68 100|
|---|---|
||Dynamic Graphs, Temporal aggregate queries|
|100k:A|DNR 98 65 99 100|
|100k:F DNR 60 68 65 100||
|Graph|Spark N4J-Gr N4J-Cy N4J-Cy-M ranite G|
|Graph|Spark N4J-Gr N4J-Cy N4J-Cy-M Granite|
|---|---|
||Dynamic Graphs, Non-aggregate queries|
|10k:DW|DNR 96 98 96 100|
|100k:Z|DNR 100 90 98 100|
|100k:A|DNR 97 46 46 100|
|100k:F|DNR 30 20 75 100|
|10k:DW 100k:Z 100k:A 100k:F|Dynamic Graphs, Non-aggregate queries DNR 96 98 96 100 DNR 100 90 98 100 DNR 97 46 46 100 DNR 30 20 75 100|
|---|---|
||Dynamic Graphs, Temporal aggregate queries|
|100k:A|DNR 95 46 99 100|
|100k:F|DNR 36 19 78 100|
-----
#### 6.6 Components of Execution Time
Next, we briefly examine where the time is spent in distributed execution.
As an exemplar, Figure 13 shows a stacked bar plot of the time taken by
_Q7 in different supersteps, and within different workers in a superstep, for_
the 100k:A-S graph. The stacks represent the time taken by the init/com_pute, scatter, and join phases of_ _ranite, the interval compute parent phase_
_G_
of Graphite (ICM), the partition compute grand-parent phase of Giraph
(VCM), and other residual time such as barrier synchronization and JVM
garbage collection (GC), in each superstep. These times are averaged across
all 100 instances of the query type. For deterministic execution, we select a
fixed split point for the execution plan that is optimal for a majority of the
queries, which, for Q7 is at the third vertex in the path.
For Q7, the first superstep time is dominated by the init logic as the
predicate operates on the Post vertex type, which has 8.7M vertices. Its
_scatter time is minimal as only 71k out edges match out of 250k and are_
used to send messages. The overheads of interval compute are small, but
_partition compute takes longer at 140 ms. In the latter, the Giraph logic_
which we extend selects the active partitions based on the vertex type of
the query predicate (Post, in the case of Q7), iterates through its active
vertices, invokes interval compute on each with the incoming messages, and
clears the message queue. The other time is non-trivial at 145 ms. This is
caused by GC triggering due to memory pressure, and taking 110 ms, with
the rest going to the superstep barrier.
In superstep 2, the compute time is negligible at 1.5 ms as only 3.4k
_Person vertices are active across both branches of the query plan, but scatter_
takes 247 ms since 2.83M edges are processed along one branch of the plan
– the Person vertex has a high out-edge degree – out of which 31k satisfy
the predicate. About 100 ms is taken by partition and interval computes, for
selecting and iterating over the relevant active vertices, and for performing
TimeWarp and state initialization, while there is a GC overhead of 64 ms
in other. In the last superstep, there is a small time taken for compute and
to join the results.
Interestingly, the time taken by each phase is similar across the different
workers in a superstep for this query. This indicates that the partitioning
manages to balance the load for this query type. However, for other queries
like Q4 (not shown for brevity), we observe that in some supersteps, scatter
takes 79% longer for the slowest worker compared to the fastest due to a
skew in the number of edges activated per worker. Also, queries like Q4 take
less time for the first superstep but a larger time in superstep 2 due to a
high fanout, going from 36k edges processed in the first step to 1.48M edges
in the second step. In others like Q3, the first superstep is dominated by
scatter since the initial vertex type Person has only 89k vertices with 770 of
them matching, but these cause 950k edges to be processed of which 122k
37
-----
Figure 13: Stacked bar plot of component execution times
in each superstep, averaged over all queries of query type
_Q7 on 100k:A-S graph. Header labels indicate average_
component time across Workers in a superstep.
6
2 4 8 16
5 100
4 80
3 60
2 40
1 20
0 0
Q1 Q2 Q3 Q4 Q5 Q6 Q7
Query Type
Figure 14: Relative execution time (left axis,
bar) and Scaling efficiency%= _[t][2]_
_tw_ [% (right axis, cir-]
cle) for Worker counts w = 4, 8, 16, relative to
_{_ _}_
_w = 2 for Weak Scaling runs with (w_ 6.25k):F-S
_×_
graphs
match and trigger messaging.
In summary, the different supersteps have high variability in execution
times and there is also variability in the time taken by each phase. Despite
that, the cost model is able to discriminate and select near-optimal plans.
The load is mostly balanced across workers in a superstep, though this
depends on the query type. Much of the time is spent directly in processing
the query using compute and scatter, with some additional overheads for the
other phases.
#### 6.7 Weak Scaling
We evaluate the weak scaling capabilities of _ranite using the static Facebook-_
_G_
distribution graphs. We use 4 different system resource sizes – 2, 4, 8 and 16
Workers, with 1 compute node per Worker, and the graph sizes increase pro
38
-----
portional to the Worker count – 12.5k:F-S, 25k:F-S, 50k:F-S and 100k:F-S.
This attempts to keep the workload per Worker constant across the scaling configurations, with the per-Worker vertex and edge counts remaining
within 18% and 23% of their mean, respectively. 100k:F-S is partitioned
_±_ _±_
into 512 partitions (128 per vertex type), and other graphs into 256 partitions (64 per vertex type). This ensures that we have enough partitions
for the compute threads to process them in parallel across all Workers. We
generate and use a 100 query workload for each query type, for each graph.
The left Y axis of Figure 14 (bars) shows for each query type, the average
relative execution time when using w = 4, 8, 16 Workers, compared to w = 2
Workers. The right Y axis (circles) shows the scaling efficiency = _t2_
_tw_ [%,]
i.e., time taken on 2 Workers vs. time taken on w Workers. With perfect
weak scaling, the relative time should be constant and efficiency 100%. The
asymmetric nature of graph data structure makes it rare to get ideal weak
scaling. However, we do see that query types Q1, Q5, Q6 and Q7 offer
60% scaling efficiency on up to 8 Workers, and all queries but Q3 and
_≥_
_Q4 have_ 40% efficiency on up to 16 Workers. Q3 and Q4 are unable to
_≥_
fully exploit the additional resources due to stragglers among their threads,
which are often 10 slower due to uneven load. These two queries also have
_×_
the largest result cardinality, which causes more messages to be sent over
the network as the number of machines increase. As a result, they have poor
scaling efficiency.
### 7 Conclusions
In this article, we have motivated the need for querying over large temporal
property graphs and the lack of such platforms. We have proposed an intuitive temporal path query model to express a wide variety of requirements
over such graphs, and designed the _ranite distributed engine to implement_
_G_
these at scale over the Graphite ICM platform. Our novel analytical cost
model uses concise information about the graph to allow accurate selection of a distributed query execution plan from several choices. These are
validated through rigorous experiments on 8 temporal graphs with a 1600query workload, derived from the LDBC benchmark. _ranite out-performs_
_G_
the baseline graph platforms and gives < 1 sec latency for most queries.
As future work, we plan to explore out of core execution models to scale
beyond distributed memory, indexing techniques to accelerate performance,
more generalized temporal tree and reachability query models, and compare
performance with other research prototypes and metrics from literature.
Designing incremental query execution strategies over streaming propertygraph updates is also a related and under-explored challenge. The _ranite_
_G_
platform is also finding relevance in analyzing epidemiological networks that
form temporal property graphs constructed from, say, digital contact tracing
39
-----
for the COVID-19 pandemic. This may motivate the need for further query
operators.
### Acknowledgements
The first author of this work was supported by the Maersk CDS M.Tech.
Fellowship, and the last author was supported by the Swarna Jayanti Fellowship from DST, India. We thank Ravishankar Joshi from BITS-Pilani,
Goa for his assistance with the experiments.
### References
[1] T. Mitchell et al., “Never-ending learning,” Communications of the
_ACM, vol. 61, no. 5, 2018._
[2] M. Cha, H. Haddadi, F. Benevenuto, and K. P. Gummadi, “Measuring user influence in twitter: The million follower fallacy,” in AAAI
_International conference on weblogs and social media (ICWSM), 2010._
[3] S. Liu, S. Poccia, K. S. Candan, G. Chowell, and M. L. Sapino, “epidms:
data management and analytics for decision-making from epidemic
spread simulation ensembles,” The Journal of infectious diseases, vol.
214, pp. S427–S432, 2016.
[4] B. Haslhofer, R. Karl, and E. Filtz, “O bitcoin where art thou? insight
into large-scale transaction graphs.” in International Workshop on Se_mantic Change & Evolving Semantics (SuCCESS), 2016._
[5] K. Shu, A. Sliva, S. Wang, J. Tang, and H. Liu, “Fake news detection on
social media: A data mining perspective,” ACM SIGKDD Explorations
_Newsletter, vol. 19, no. 1, 2017._
[6] W. Fan, “Graph pattern matching revised for social network analysis,”
in International Conference on Database Theory (ICDT), 2012.
[7] Z. Huang, W. Chung, and H. Chen, “A graph model for e-commerce
recommender systems,” Journal of the American Society for informa_tion science and technology, vol. 55, no. 3, 2004._
[[8] Sharp, Austin et al., “JanusGraph,” 2020, https://janusgraph.org/.](https://janusgraph.org/)
[9] V. G. Castellana, A. Morari, J. Weaver, A. Tumeo, D. Haglin, O. Villa,
and J. Feo, “In-memory graph databases for web-scale data,” Computer,
vol. 48, no. 3, 2015.
40
-----
[10] Y. Yan, C. Wang, A. Zhou, W. Qian, L. Ma, and Y. Pan, “Efficient
indices using graph partitioning in rdf triple stores,” in 2009 IEEE
_25th International Conference on Data Engineering._ IEEE, 2009, pp.
1263–1266.
[11] T. Milo and D. Suciu, “Index structures for path expressions,” in Inter_national Conference on Database Theory. Springer, 1999, pp. 277–295._
[12] M. Junghanns, M. Kießling, N. Teichmann, K. Gómez, A. Petermann, and E. Rahm, “Declarative and distributed graph analytics with
gradoop,” Proceedings of the VLDB Endowment, vol. 11, no. 12, 2018.
[13] B. Shao, H. Wang, and Y. Li, “Trinity: A distributed graph engine on
a memory cloud,” in Proceedings of the 2013 ACM SIGMOD Interna_tional Conference on Management of Data, 2013, pp. 505–516._
[14] D. Greene, D. Doyle, and P. Cunningham, “Tracking the evolution of
communities in dynamic social networks,” in 2010 International con_ference on advances in social networks analysis and mining._ IEEE,
2010.
[15] B. George and S. Shekhar, “Time-aggregated graphs for modeling spatio-temporal networks,” in Journal on Data Semantics XI.
Springer, 2008.
[16] L. Zhao, G.-J. Wang, M. Wang, W. Bao, W. Li, and H. E. Stanley,
“Stock market as temporal network,” Physica A: Statistical Mechanics
_and its Applications, vol. 506, 2018._
[17] O. Walavalkar, A. Joshi, T. Finin, Y. Yesha et al., “Streaming knowledge bases,” in Fourth International Workshop on Scalable Semantic
_Web knowledge Base Systems, 2008._
[18] D. Ediger, J. Riedy, D. A. Bader, and H. Meyerhenke, “Tracking structure of streaming social networks,” in 2011 IEEE International Sympo_sium on Parallel and Distributed Processing Workshops and Phd Forum,_
2011.
[19] M. Then, T. Kersten, S. Günnemann, A. Kemper, and T. Neumann,
“Automatic algorithm transformation for efficient multi-snapshot analytics on temporal graphs,” Proceedings of the VLDB Endowment,
vol. 10, no. 8, 2017.
[20] G. Malewicz et al., “Pregel: a system for large-scale graph processing,”
in SIGMOD, 2010.
[21] S. Gandhi and Y. Simmhan, “An interval-centric model for distributed
computing over temporal graphs,” in IEEE International Conference
_on Data Engineering (ICDE), 2020._
41
-----
[22] K. Zeng, J. Yang, H. Wang, B. Shao, and Z. Wang, “A distributed
graph engine for web scale rdf data,” The VLDB Journal, vol. 6, no. 4,
2013.
[23] S. Ramesh, A. Baranawal, and Y. Simmhan, “A distributed path query
engine for temporal property graphs,” in IEEE/ACM International
_Symposium on Cluster, Cloud and Internet Computing (CCGrid), 2020._
[24] Y. Guo, M. Biczak, A. L. Varbanescu, A. Iosup, C. Martella, and T. L.
Willke, “How well do graph-processing platforms perform? an empirical
performance evaluation and analysis,” in IEEE IPDPS, 2014.
[25] Y. Simmhan, A. Kumbhare, C. Wickramaarachchi, S. Nagarkar,
S. Ravi, C. Raghavendra, and V. Prasanna, “GoFFish: A sub-graph
centric framework for large-scale graph analytics,” in EuroPar, 2014.
[26] J. E. Gonzalez, R. S. Xin, A. Dave, D. Crankshaw, M. J. Franklin,
and I. Stoica, “Graphx: Graph processing in a distributed dataflow
framework,” in OSDI, 2014.
[27] H. Chen, M. Liu, Y. Zhao, X. Yan, D. Yan, and J. Cheng, “G-miner: An
efficient task-oriented graph mining system,” in ACM EuroSys, 2018.
[28] D. Gregor and A. Lumsdaine, “The parallel bgl: A generic library
for distributed graph computations,” Parallel Object-Oriented Scien_tific Computing (POOSC), 2005._
[29] H.-V. Dang, R. Dathathri, G. Gill, A. Brooks, N. Dryden, A. Lenharth,
L. Hoang, K. Pingali, and M. Snir, “A lightweight communication runtime for distributed graph analytics,” in IEEE IPDPS, 2018.
[30] J. Shi, Y. Yao, R. Chen, H. Chen, and F. Li, “Fast and concurrent
RDF queries with rdma-based distributed graph exploration,” in 12th
_{_ _}_
_USENIX_ _Symposium on Operating Systems Design and Implementa-_
_{_ _}_
_tion (_ _OSDI_ _16), 2016, pp. 317–332._
_{_ _}_
[31] Y. Simmhan et al., “Distributed programming over time-series graphs,”
in IEEE IPDPS, 2015.
[32] W. Han et al., “Chronos: a graph engine for temporal graph analysis,”
in ACM EuroSys, 2014.
[33] T. A. Zakian, L. A. Capelli, and Z. Hu, “Incrementalization of vertexcentric programs,” in IEEE IPDPS, 2019.
[34] R. Cheng, J. Hong, A. Kyrola, Y. Miao, X. Weng, M. Wu, F. Yang,
L. Zhou, F. Zhao, and E. Chen, “Kineograph: taking the pulse of a
fast-changing and connected world,” in ACM EuroSys, 2012.
42
-----
[35] D. Chavarría-Miranda, V. G. Castellana, A. Morari, D. Haglin, and
J. Feo, “Graql: A query language for high-performance attributed graph
databases,” in IEEE IPDPS Workshops, 2016.
[36] J. Zhou, G. V. Bochmann, and Z. Shi, “Distributed query processing in
an ad-hoc semantic web data sharing system,” in IEEE IPDPS Work_shops, 2013._
[37] “SPARQL Query Language for RDF,” 2008. [Online]. Available:
[http://www.w3.org/TR/rdf-sparql-query/](http://www.w3.org/TR/rdf-sparql-query/)
[38] J. Huang, D. J. Abadi, and K. Ren, “Scalable sparql querying of large
rdf graphs,” VLDB Endowment, vol. 4, no. 11, 2011.
[39] N. Jamadagni and Y. Simmhan, “GoDB: From batch processing to
distributed querying over property graphs,” in IEEE/ACM CCGrid,
2016.
[40] M. Sarwat, S. Elnikety, Y. He, and M. F. Mokbel, “Horton+: A distributed system for processing declarative reachability queries over partitioned graphs,” PVLDB, vol. 6, no. 14, 2013.
[41] K. Semertzidis and E. Pitoura, “Top-k durable graph pattern queries
on temporal graphs,” IEEE TKDE, vol. 31, no. 1, 2018.
[42] K. Semertzidis, E. Pitoura, and K. Lillis, “Timereach: Historical reachability queries on evolving graphs,” in EDBT, 2015.
[43] H. Wu, Y. Huang, J. Cheng, J. Li, and Y. Ke, “Reachability and timebased path queries in temporal graphs,” in ICDE, 2016.
[44] N. Sengupta, A. Bagchi, M. Ramanath, and S. Bedathur, “Arrow: Approximating reachability using random walks over web-scale graphs,”
in 2019 IEEE 35th International Conference on Data Engineering
_(ICDE), 2019._
[45] A. P. Iyer, Z. Liu, X. Jin, S. Venkataraman, V. Braverman, and I. Stoica, “ ASAP : Fast, approximate graph pattern mining at scale,” in
_{_ _}_
_13th_ _USENIX_ _Symposium on Operating Systems Design and Imple-_
_{_ _}_
_mentation (_ _OSDI_ _18), 2018._
_{_ _}_
[46] J. Byun, S. Woo, and D. Kim, “Chronograph: Enabling temporal graph
traversals for efficient information diffusion analysis over time,” IEEE
_TKDE, vol. 32, no. 3, 2019._
[[47] “Neo4j graph platform,” 2020. [Online]. Available: https://neo4j.com/](https://neo4j.com/)
[48] “Orientdb graph database,” 2020. [Online]. Available: [https:](https://orientdb.com/graph-database/)
[//orientdb.com/graph-database/](https://orientdb.com/graph-database/)
43
-----
[49] J. Allen, “Maintaining knowledge about temporal intervals,” Commu_nications of the ACM, vol. 26, no. 11, 1983._
[50] V. Z. Moffitt and J. Stoyanovich, “Temporal graph algebra,” in ACM
_International Symposium on Database Prog. Languages (DBPL), 2017._
[51] G. Karypis and V. Kumar, “A fast and high quality multilevel scheme
for partitioning irregular graphs,” SIAM Journal on Scientific Comput_ing, vol. 20, no. 1, 1998._
[52] S. Muthukrishnan, V. Poosala, and T. Suel, “On rectangular partitionings in two dimensions: Algorithms, complexity and applications,” in
_International Conference on Database Theory (ICDT), 1999._
[53] “The LDBC social network benchmark (version 0.3.2),” Linked Data
Benchmark Council, Tech. Rep., 2019.
[54] M.-D. Pham, P. Boncz, and O. Erling, “S3g2: A scalable structurecorrelated social graph generator,” in Selected Topics in Performance
_Evaluation and Benchmarking, 2013._
44
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/2002.03274, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "GREEN",
"url": "https://arxiv.org/pdf/2002.03274"
}
| 2,020
|
[
"JournalArticle"
] | true
| 2020-02-09T00:00:00
|
[
{
"paperId": "b97084d8e61f062ba63338528701c0c1ec7f3da4",
"title": "An Interval-centric Model for Distributed Computing over Temporal Graphs"
},
{
"paperId": "69b47761baff2b3d25fc73f5cc90c3e3518d2b4f",
"title": "ChronoGraph: Enabling Temporal Graph Traversals for Efficient Information Diffusion Analysis over Time"
},
{
"paperId": "605bad962c56a74d4fe5e42b36271c2e67becaa3",
"title": "Incrementalization of Vertex-Centric Programs"
},
{
"paperId": "2cec2e0c71d24744da6035b369f5c96084787c04",
"title": "A Lightweight Communication Runtime for Distributed Graph Analytics"
},
{
"paperId": "24a441897a0fd3b781052cb4ced38532c031acd8",
"title": "G-Miner: an efficient task-oriented graph mining system"
},
{
"paperId": "3f92a52c6823d7be1a7672be56333b3379a778b0",
"title": "Temporal graph algebra"
},
{
"paperId": "cb40a5e6d4fc0290452345791bb91040aed76961",
"title": "Fake News Detection on Social Media: A Data Mining Perspective"
},
{
"paperId": "3ff2f80eee5256aeebab1dced07defc73a41bd70",
"title": "GraQL: A Query Language for High-Performance Attributed Graph Databases"
},
{
"paperId": "975668821b26eeae79265135be73337ad7e3be7a",
"title": "GoDB: From Batch Processing to Distributed Querying over Property Graphs"
},
{
"paperId": "1c948041fd224135d780a552ce78be1503ee129d",
"title": "Reachability and time-based path queries in temporal graphs"
},
{
"paperId": "0ba86604228b555475496e200f31878df3aabd6e",
"title": "Never-Ending Learning"
},
{
"paperId": "d6adbafb07c7a5590516404208fe8776f1363eb3",
"title": "GraphX: Graph Processing in a Distributed Dataflow Framework"
},
{
"paperId": "8b48dde8978253066a0b19ab10e3525a2ebbdad4",
"title": "Distributed Programming over Time-Series Graphs"
},
{
"paperId": "92d797407719ecf859431caf943de13ca7c9a38b",
"title": "How Well Do Graph-Processing Platforms Perform? An Empirical Performance Evaluation and Analysis"
},
{
"paperId": "6f7cd29a3dfdcb2f6880a022e13054542020c5ce",
"title": "Chronos: a graph engine for temporal graph analysis"
},
{
"paperId": "05370a6cc820ffe5393fcc948d7d600b5949a217",
"title": "GoFFish: A Sub-graph Centric Framework for Large-Scale Graph Analytics"
},
{
"paperId": "ad88f7b783a742ca2fe7b1f0cb54b7bbb40e61ce",
"title": "Horton+: A Distributed System for Processing Declarative Reachability Queries over Partitioned Graphs"
},
{
"paperId": "688ef3247342fb1a2292503125bb3b1ccde6a1ea",
"title": "Distributed Query Processing in an Ad-hoc Semantic Web Data Sharing System"
},
{
"paperId": "f0b21ae25f918818d032d9b6b326f334b3510caa",
"title": "S3G2: A Scalable Structure-Correlated Social Graph Generator"
},
{
"paperId": "282bc59faefb734137d2ea978cb1eb5699e67c7c",
"title": "Kineograph: taking the pulse of a fast-changing and connected world"
},
{
"paperId": "ead04453dc387f849472e1edf4f9ba5bed6571f7",
"title": "Graph pattern matching revised for social network analysis"
},
{
"paperId": "2d867297dfe0d3ce2ed5b1d0f2dff88cac46ee94",
"title": "Pregel: a system for large-scale graph processing"
},
{
"paperId": "1ad8410d0ded269af4a0116d8b38842a7549f0ae",
"title": "Measuring User Influence in Twitter: The Million Follower Fallacy"
},
{
"paperId": "8232d17f69b491d92d944af0fb93b5a10632497c",
"title": "Efficient Indices Using Graph Partitioning in RDF Triple Stores"
},
{
"paperId": "4468768ed9f544a8ddeccb49af3857dfcf5df359",
"title": "Time-Aggregated Graphs for Modeling Spatio-temporal Networks"
},
{
"paperId": "7e2de7f6458ad745bb64daef2e5145ffaf2ad647",
"title": "On Rectangular Partitionings in Two Dimensions: Algorithms, Complexity, and Applications"
},
{
"paperId": "df86d2a8c217776786bac9019d8b20029e4c0dd5",
"title": "A Fast and High Quality Multilevel Scheme for Partitioning Irregular Graphs"
},
{
"paperId": "f4d642d674aa63aafc11562c2557cb4772946147",
"title": "Maintaining knowledge about temporal intervals"
},
{
"paperId": "1d5d5a2b448fa93a131d44edba8bc5483c23ef3e",
"title": "Top-<inline-formula><tex-math notation=\"LaTeX\">$k$</tex-math><alternatives> <inline-graphic xlink:href=\"semertzidis-ieq1-2823754.gif\"/></alternatives></inline-formula> Durable Graph Pattern Queries on Temporal Graphs"
},
{
"paperId": null,
"title": "The LDBC social network benchmark (version 0.3.2)"
},
{
"paperId": "96b95da0ab88de23641014abff2a5c0b5fec00c9",
"title": "O Bitcoin Where Art Thou? Insight into Large-Scale Transaction Graphs"
},
{
"paperId": "865dc9c42de72ae1581e63d52936e6ed5f41d919",
"title": "TimeReach: Historical Reachability Queries on Evolving Graphs"
},
{
"paperId": "d7f449c199ce86d3b8039899caabb31b54ced7f2",
"title": "The Parallel BGL : A Generic Library for Distributed Graph Computations"
},
{
"paperId": null,
"title": "A distributed path query engine for temporal property graphs"
},
{
"paperId": null,
"title": "JanusGraph"
}
] | 34,427
|
en
|
[
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Environmental Science",
"source": "s2-fos-model"
},
{
"category": "Engineering",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/021d3d91a15cd3a247255a205d8e4228e04609a0
|
[] | 0.917905
|
Improving Collaborative Intrusion Detection System Using Blockchain and Pluggable Authentication Modules for Sustainable Smart City
|
021d3d91a15cd3a247255a205d8e4228e04609a0
|
Sustainability
|
[
{
"authorId": "2152962051",
"name": "R. Gupta"
},
{
"authorId": "2202681571",
"name": "Vedant Chawla"
},
{
"authorId": "144821177",
"name": "R. K. Pateriya"
},
{
"authorId": "2099551831",
"name": "P. Shukla"
},
{
"authorId": "2213062",
"name": "S. Mahfoudh"
},
{
"authorId": "2195025023",
"name": "Syed Bilal Hussain Shah"
}
] |
{
"alternate_issns": null,
"alternate_names": null,
"alternate_urls": [
"http://mdpi.com/journal/sustainability",
"http://nbn-resolving.de/urn/resolver.pl?urn=urn:nbn:ch:bel-172127"
],
"id": "8775599f-4f9a-45f0-900e-7f4de68e6843",
"issn": "2071-1050",
"name": "Sustainability",
"type": "journal",
"url": "http://www.e-helvetica.nb.admin.ch/directAccess?callnumber=bel-172127"
}
|
The threat of cyber-attacks is ever increasing in today’s society. There is a clear need for better and more effective defensive tools. Intrusion detection can be defined as the detection of anomalous behavior either in the host or in the network. An intrusion detection system can be used to identify the anomalous behavior of the system. The two major tasks of intrusion detection are to monitor data and raise an alert to the system administrators when an intrusion takes place. The current intrusion detection system is incapable of tackling sophisticated attacks which take place on the entire network containing large number of nodes while maintaining a low number of login attempts on each node in the system. A collaborative intrusion detection system (CIDS) was designed to remove the inefficiency of the current intrusion detection system which failed to detect coordinated distributed attacks. The main problem in the CIDS is the concept of trust. Hosts in the network need to trust the data sent by other peers in the network. To bring in the concept of trust and implement the proof-of-concept, blockchain was used. Pluggable authentication modules (PAM) were also used to track login activity securely before an intruder could modify the login activity. To implement blockchain, an Ethereum-based private blockchain was used.
|
## sustainability
_Article_
# Improving Collaborative Intrusion Detection System Using Blockchain and Pluggable Authentication Modules for Sustainable Smart City
**Rajeev Kumar Gupta** **[1]** **, Vedant Chawla** **[2], Rajesh Kumar Pateriya** **[2], Piyush Kumar Shukla** **[3],**
**Saoucene Mahfoudh** **[4]** **and Syed Bilal Hussain Shah** **[4,]***
1 Computer Science and Engineering Department, Pandit Deendayal Energy University,
Gandhinagar 382007, India
2 Computer Science and Engineering Department, Maulana Azad National Institute of Technology,
Bhopal 462003, India
3 Computer Science & Engineering Department, University Institute of Technology, Rajiv Gandhi Proudyogiki
Vishwavidyalaya (Technological University of Madhya Pradesh), Bhopal 462033, India
4 School of Engineering, Computing and Informatics, Dar Al-Hekma University, Jeddah 22246, Saudi Arabia
***** Correspondence: sshah@dah.edu.sa
**Citation: Gupta, R.K.; Chawla, V.;**
Pateriya, R.K.; Shukla, P.K.;
Mahfoudh, S.; Shah, S.B.H.
Improving Collaborative Intrusion
Detection System Using Blockchain
and Pluggable Authentication
Modules for Sustainable Smart City.
_[Sustainability 2023, 15, 2133. https://](https://doi.org/10.3390/su15032133)_
[doi.org/10.3390/su15032133](https://doi.org/10.3390/su15032133)
Academic Editors: Dhananjay Singh,
Paulo J. Sequeira Gonçalves, Pradeep
Kumar Singh, Pradip Sharma and
Pao-Ann Hsiung
Received: 21 November 2022
Revised: 27 December 2022
Accepted: 16 January 2023
Published: 23 January 2023
**Copyright:** © 2023 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
[Attribution (CC BY) license (https://](https://creativecommons.org/licenses/by/4.0/)
[creativecommons.org/licenses/by/](https://creativecommons.org/licenses/by/4.0/)
4.0/).
**Abstract: The threat of cyber-attacks is ever increasing in today’s society. There is a clear need for**
better and more effective defensive tools. Intrusion detection can be defined as the detection of
anomalous behavior either in the host or in the network. An intrusion detection system can be used
to identify the anomalous behavior of the system. The two major tasks of intrusion detection are
to monitor data and raise an alert to the system administrators when an intrusion takes place. The
current intrusion detection system is incapable of tackling sophisticated attacks which take place
on the entire network containing large number of nodes while maintaining a low number of login
attempts on each node in the system. A collaborative intrusion detection system (CIDS) was designed
to remove the inefficiency of the current intrusion detection system which failed to detect coordinated
distributed attacks. The main problem in the CIDS is the concept of trust. Hosts in the network need
to trust the data sent by other peers in the network. To bring in the concept of trust and implement
the proof-of-concept, blockchain was used. Pluggable authentication modules (PAM) were also used
to track login activity securely before an intruder could modify the login activity. To implement
blockchain, an Ethereum-based private blockchain was used.
**Keywords: sustainable smart city; intrusion detection system; collaborative intrusion detection**
system; authentication; blockchain
**1. Introduction**
According to the report of the Indian Computer Emergency Response Team (2021),
more than 26,100 websites were victims of cyber-attacks in India in the year 2020 alone.
This clearly indicates the need for better and more effective defensive tools. The role of
blockchain in smart, sustainable cities is vital because it helps to foster the kind of trust
necessary for smart cities. Blockchain should serve as the cornerstone for the development
of a smart city and is a crucial assurance for the proper design and execution of the
management strategy and planning scheme. A smart city naturally combines smart energy,
smart transportation, smart government, and other services under the same umbrella.
Decentralization and the availability of clear data place strict constraints on the big data
service platform. Finding the problematic node among the hundreds of millions of nodes
in a network is a time-consuming operation if a network encounters a problem or is the
target of an attack. Most current Internet-of-Things networks are centralized. A huge
server or centralized cloud is connected to hundreds of millions of nodes, which causes
-----
_Sustainability 2023, 15, 2133_ 2 of 14
bottlenecks in the price and computer storage capacity. Blockchain-distributed technology
can guarantee that even if one or more nodes are hacked, the total network data remains
trustworthy and secure. Distributed computing makes use of point-to-point computing
to handle the hundreds of billions of transactions that the Internet-of-Things generates.
This significantly lowers the cost of computing and storage by utilizing the computing and
storage capabilities of a large number of idle devices deployed in unused locations.
In order to increase the degree of security for secure transmission and safe storage,
additional protection mechanisms need to be implemented due to the privacy of numerous
people involved. Blockchain has shown to be secure, dependable, and suitable for this
purpose. The disaster recovery system cannot be enhanced due to the high expense of
creating a data center and data storage. Therefore, a key issue at hand is how to lower
storage costs while enhancing disaster recovery capabilities. Blockchain, which connects
distributed and centralized services, can successfully stop an attack on the vital network
infrastructure. The main objective of intrusion detection is to observe anomalous behavior
either in a network or in a host. In the current scenario, the current IDS is not sophisticated
enough to detect the wide variety of threats. Collaborative intrusion detection has some of
the capabilities to at least detect some of those threats and send them for further processing.
Based on the deployed location, IDSs can be categorized as a host-based intrusion detection
system (HIDS) and network-based intrusion detection system (NIDS). A HIDS monitors the
characteristics of a particular node and the system events in a node for malicious activities,
whereas a NIDS monitors the network by placing packet sniffers in the network at various
points. These packet sniffers pick up the data and send the data to analysis units who
compare the present state of the system with that of an anomaly.
Based on the approach of the detection, an IDS can again be classified into two types:
signature-based IDSs and anomaly-based IDSs. Signature-based detection detects an attack
by comparing stored signatures with the observed system or network events for possible
occurrences. A signature (also known as a rule) is a pattern that describes a known attack or
exploit. Anomaly-based detection works by detecting large deviations between its pre-built
normal profile and the observed events, and hence detects suspicious activity. A normal
profile is frequently generated by observing the features of ordinary activity over time, and
it might represent the regular behavior of users, network connections, and programs [1]. If
an abnormal circumstance is discovered, an alert may be triggered. The main disadvantage
of IDS is that it cannot detect sophisticated attacks which take place on the entire network
of nodes cumulatively as they monitor only a single node or a single network. For example,
if we have a series of stand-alone IDSs, they are incapable of detecting a distributed attack
which takes place across multiple hosts in a network. This is because they do not have the
ability to co-relate the events which take place. To address this weakness, the concept of
CIDSs was introduced.
CIDSs were introduced to address the weakness of IDSs which can be seen during
distributed attacks. CIDSs generally consists of several monitor units and analysis units.
The monitor units jot down the information and send it to the analysis units, which process
the information and make decisions based on it. Based on the architectural differences,
CIDSs again can be classified into three categories as shown in Figure 1, namely: centralized,
decentralized, and distributed [2]. A centralized CIDS is the most basic version and the
simplest one. However, it is prone to a single point of failure (SPoF) and performance
bottleneck in cases of network overload. In distributed CIDSs, the SPoF disadvantage is
somewhat removed but it still has disadvantages. In this, information is lost at each level
of the hierarchy and hence is somewhat unreliable. In a decentralized CIDS, each node
behaves as both monitor and analysis units. It is a P2P architecture which facilitates data
sharing, correlation, and aggregation of data across all nodes. However, CIDS also has
some disadvantages. The network cost incurred is very high as all nodes are in constant
communication with each other. Furthermore, the idea of trust is very important among
these nodes. To remove the trust issue among the nodes, the concept of blockchain was
-----
###, g y g
_Sustainabilityall nodes are in constant communication with each other. Furthermore, the idea of trust 2023, 15, 2133_ 3 of 14
### is very important among these nodes. To remove the trust issue among the nodes, the concept of blockchain was introduced. This problem, along with CIDS, is discussed in a introduced. This problem, along with CIDS, is discussed in a detailed manner further in detailed manner further in the later stages. Figure 1 illustrates the architecture of CIDS.the later stages. Figure 1 illustrates the architecture of CIDS.
#### Figure 1. Overview of a CIDS architecture. Figure 1. Overview of a CIDS architecture.
The main contributions of this paper are:
### The main contributions of this paper are: Proposed system will be able to detect coordinated distributed attacks.
Hosts in the network need to trust the data sent by other peers in the network. To
### Proposed system will be able to detect coordinated distributed attacks.bring in the concept of trust and implement the proof-of-concept, blockchain was used. Hosts in the network need to trust the data sent by other peers in the network. To Pluggable authentication modules (PAM) were also used to track login activity securely
before an intruder could modify the login activity.
### bring in the concept of trust and implement the proof-of-concept, blockchain was used.
To implement blockchain, an Ethereum-based private blockchain was used
### Pluggable authentication modules (PAM) were also used to track login activity se-This paper is organized as follows: Section 2 discusses the basics of blockchain along
curely before an intruder could modify the login activity.with the different components of blockchain, Section 3 discusses different existing intrusion
To implement blockchain, an Ethereum-based private blockchain was useddetection systems, and Section 4 explains the proposed improved collaborative IDS which
uses blockchain and pluggable authentication modules. Section 5 discusses the result
### This paper is organized as follows: Section 2 discusses the basics of blockchain along analysis and Section 6 summarizes the entire work and gives direction for future work.
with the different components of blockchain, Section 3 discusses different existing intru
**2. Blockchain**
### sion detection systems, and Section 4 explains the proposed improved collaborative IDS
Blockchain can be defined as a distributed peer-to-peer network of blocks. Each block
### which uses blockchain and pluggable authentication modules. Section 5 discusses the re
is linked to the previous block using a cryptographic hash. Blockchain technology has been
### sult analysis and Section 6 summarizes the entire work and gives direction for future applied to several fields such as healthcare, education, energy, etc. There are three types work. of blockchain ledgers which are currently in use: public, consortium, and private. Public
blockchains (such as Ethereum) are accessible to anyone with internet access and anyone
can read the blockchain and maintain the blockchain ledger, i.e., there is no membership
### 2. Blockchain mechanism in place. The consortium blockchains (such as the Hyperledger Fabric) are
maintained by an established body which grants access to others and has a pre-defined
### Blockchain can be defined as a distributed peer-to-peer network of blocks. Each block
consortium of peers maintaining the chain. Private blockchains are maintained by one
### is linked to the previous block using a cryptographic hash. Blockchain technology has entity that provides access to others and there is no consensus process. been applied to several fields such as healthcare, education, energy, etc. There are three
_2.1. Block Structure_
### types of blockchain ledgers which are currently in use: public, consortium, and private.
The most basic definition of blockchain is that it is a chain of blocks with each block
### Public blockchains (such as Ethereum) are accessible to anyone with internet access and connected to the one before it with the help of a mathematical relationship. The block in anyone can read the blockchain and maintain the blockchain ledger, i.e., there is no mem-itself is a container of data. The main premise underlying blockchain is that each block
contains a unique self-identifying hash that ensures the chain’s integrity. The hash of the
### bership mechanism in place. The consortium blockchains (such as the Hyperledger Fabric)
block index, data, timestamp, and, of course, the hash of the previous block hash, make
### are maintained by an established body which grants access to others and has a pre-defined up this self-identifying hash. It also contains a record of the transactions, called a ledger, consortium of peers maintaining the chain. Private blockchains are maintained by one which took place during the time of blockchain production. As each block references the entity that provides access to others and there is no consensus process
-----
### up i e i e i yi g a I a o o ai a e o o e a a io, a e a e
_Sustainability 2023, 15, 2133_ which took place during the time of blockchain production. As each block reference4 of 14
### one before it, there is a record of all transactions that took place prior to the current bl generation. The Figure 2 shows the structure of the block chain generation.
one before it, there is a record of all transactions that took place prior to the current block’s
generation. The Figure 2 shows the structure of the block chain generation.
##### Figure 2. Figure 2.Structure of a blockchain [3]. Structure of a blockchain [3].
_2.2. Consensus_
### 2.2. Consensus
Consensus algorithms allow the participants to reach an agreement about the state
### Consensus algorithms allow the participants to reach an agreement about the of the network without the presence of a central authority. Any blockchain model is only
of the network without the presence of a central authority. Any blockchain model isas effective as its consensus model. There are two major consensus algorithms in the
blockchain world which are the proof-of-work and the proof-of-stake. The proof-of-work
### as effective as its consensus model. There are two major consensus algorithms in the b
algorithm is implemented by Bitcoin, whereas the proof-of-stake algorithm is implemented
### chain world which are the proof-of-work and the proof-of-stake. The proof-of-work by Ethereum and is currently in deployment. Proof-of-work is founded on the premise rithm is implemented by Bitcoin, whereas the proof-of-stake algorithm is implementethat a participant establishes its identity by demonstrating that it worked. In the case of Ethereum and is currently in deployment. Proof-of-work is founded on the premiseBitcoin, each participant’s purpose is to find a hash value that is less than a number set by
the network as the difficulty level. This is an example of a computational puzzle where
### a participant establishes its identity by demonstrating that it worked. In the case of Bit
a brute-force, guess-and-check method is the most effective way to solve it. This process,
### each participant’s purpose is to find a hash value that is less than a number set bknown as mining, ensures that no single player has an edge in creating the next block. network as the difficulty level. This is an example of a computational puzzle whAs a result, miners are not required to provide any authentication or a-priori knowledge.
The chances of a block being modified successfully diminishes exponentially with the size
### brute-force, guess-and-check method is the most effective way to solve it. This pro
of the blockchain. Proof-of-work, on the other hand, is subject to the 51 percent attack,
### known as mining, ensures that no single player has an edge in creating the next bloc
in which a coalition with more than half of the possible mining power can insert blocks
### a result, miners are not required to provide any authentication or a-priori knowledgeinto the blockchain. To counter this, Ethereum built a new consensus algorithm called chances of a block being modified successfully diminishes exponentially with the siproof-of-stake. Proof-of-stake relies on a group of validators with a financial stake in the
network voting and proposing the next block in turn. The method chooses validators for
### the blockchain. Proof-of-work, on the other hand, is subject to the 51 percent attac
block production in a pseudo-random manner, preventing advance knowledge of when a
### which a coalition with more than half of the possible mining power can insert blocksspecific participant would create a block. The quantity of cryptocurrency, or stake, that a the blockchain. To counter this, Ethereum built a new consensus algorithm called pparticipant has determines his or her chances of being chosen as a validator. While there of-stake. Proof-of-stake relies on a group of validators with a financial stake in the netare several drawbacks to this method of implementation, it does address the 51 percent
attack problem which the proof-of-work had and is currently being developed by Ethereum.
### voting and proposing the next block in turn. The method chooses validators for b
Table 1 illustrates the fields of block structure in blockchains and their uses.
### production in a pseudo-random manner, preventing advance knowledge of when a cific participant would create a block. The quantity of cryptocurrency, or stake, that a ticipant has determines his or her chances of being chosen as a validator. While ther several drawbacks to this method of implementation, it does address the 51 percent a problem which the proof-of-work had and is currently being developed by Ethereum ble 1 illustrates the fields of block structure in blockchains and their uses.
-----
_Sustainability 2023, 15, 2133_ 5 of 14
**Table 1. Fields in a blockchain [4].**
**Name** **Description**
Version It refers to the protocol’s identifying rules.
Timestamp
Nodes can use timestamps to properly change the mining difficulty for each block
creation period. Timestamps allow the net- work to calculate how long it takes to
extract blocks during a certain time period and alter the mining difficulty
parameter accordingly.
Previous Hash Used to link the previous block with the current block in the chain.
Target It defines the difficulty level of the consensus algorithm.
Nonce is an abbreviation for “number only used once” which is added to a hashed
Nonce block in a blockchain that when rehashed, fits the difficulty level limitations. In
order to receive cryptocurrency, blockchain miners must solve for a nonce.
A Merkle Root is the hash of all the hashes of all the legitimate transactions that
Merkle Root
make up a block.
Hashing transaction occurs by Merkel tree, where each node is related with its
Hash parent node; therefore, if the transaction is modified, then it will affect all hash
trees from the leaf node to the Merkle Root, respectively.
**3. Literature Survey**
Kanth et al. [5] implemented a collaborative intrusion detection system capable of
recording login activity via a private blockchain-based ledger and hence it is immutable. In
initial stages the authors were successfully capable of proving that blockchain-based CIDSs
were a viable method to detect doorknob-rattling attacks and hence can prevent any act of
an intruder trying to modify the activity records. The author [6] uses CPU utilization as a
metric to accurately determine whether an intrusion is taking place or not.
Golomb et al. [7] introduces CIoTA, which is a blockchain-based solution for collaborative anomaly detection across a large number of IoT devices. While staying resilient
to adversarial attacks, CIoTA constantly trains an anomaly detection model. CIoTA can
also distinguish between uncommon benign events and malevolent activity by harnessing
the knowledge of the crowd. One downside of CIoTA is that each IoT model/firmware
must have its own chain published. As a result, CIoTA is best suited to large industrial
settings and smart cities in its current state. We intend to develop CIoTA in the future
to support a variety of frameworks and increase its detection capability, for example, by
investigating API flows rather than lower-level control flows. Ide et al. [8] presented a
system (CollabDict) based on blockchain and the Gaussian mixture learning algorithm
for collaborative anomaly detection. The major challenges which the author faced here
were building the consensus, validating the data, and security of the data. However, the
performance of the CollabDict is better than the fuses multitask learning algorithm.
Kumari et al. [9] primarily examine the issue of harmful behaviors occurring in blockchain
networks, and then attempt to remedy the problem using the clustering protocol. As a result,
the authors keep a check on each node’s behavior pattern. Had the authors tried to perform
manually for each node, it would have been practically impossible to do so for all the nodes.
The K-means clustering approach was utilized to perform the clustering. However, with that
algorithm, considerable improvisation was required. As a result, an adapted version of the
k-means method was used. This in turn made the blockchain safer against any unlawful or
unusual activity. However, the major disadvantage is that the authors used the mean value
for each cluster, thus an inaccurate cluster head could be selected.
Dey [10] employs game theory and supervised machine learning techniques to identify
anomalous player behavior in a blockchain network. The author provides the probability
for each attack based on the value of each transaction; however, the implementation was
still in its early phases and hence required a lot of improvements in the defense mechanism.
Signorini et al. [11] proposed BAD (blockchain anomaly detection). BAD, in particular,
enables the detection of abnormal transactions and the prevention of their propagation.
-----
_Sustainability 2023, 15, 2133_ 6 of 14
While forks can occur naturally in the blockchain life cycle owing to network delays, they
can also be generated purposefully by attackers and used to commit fraud. Malicious
acts are dispersed throughout the chain. By gathering data, BAD enables the avoidance
of repeated attacks and builds a tamper-proof threat database that is distributed (thus
preventing any single point of failure), trusted (the majority of the network collects and
verifies any behavioral data), and private.
Kanth et al. [5] implemented a collaborative intrusion detection system capable of
recording login activity via a private blockchain-based ledger and hence it is immutable. In
initial stages the authors were successfully capable of proving that blockchain-based CIDSs
were a viable method to detect doorknob-rattling attacks and hence can prevent any act of
an intruder trying to modify the activity records. The author also uses CPU utilization as a
metric to accurately determine whether an intrusion is taking place or not.
Steichen et al. [12] discusses security issues regarding private or consortium blockchains.
In this paper, the authors discussed how an attacker can target individual nodes since the
number of nodes undertaking blockchain-related tasks is generally restricted. As a result,
ChainGuard, which is built as an SDN module and identifies and intercepts excessively
large flows at the network level, was proposed in this study. ChainGuard’s implementation
specifics were discussed and trials were carried out. The tests conducted by the author indicate
that ChainGuard can effectively resist DoS and DDoS assaults while allowing a restricted
number of packets to cross the SDN network, and hence permits communication between
benign blockchain nodes to continue in the case of an attack. Zhu et al. [13] discussed a
novel approach to achieving the controllable blockchain CBDM, which is used to obtain
storage efficiency in the cloud computing network and to reduce the risk of attacks which are
malicious in the blockchain. Though not tested in a real environment, it provides huge scope
for development of the prototype.
Hu et al. [14] discussed the multi-microgrid system which it creates a collaborative
intrusion detection (CID) paradigm based on blockchain technology. It stores the CID
goal in a blockchain and uses a consensus mechanism to create a multi-microgrid system
correlation model. It also reduces the false-negative rate and considerably improves the
DPoS consensus algorithm by continuously using multiple patterns. However, the major
drawback here is that this method does not provide a higher level of true-positive rates
and is also limited to fewer types of attacks. N. Alexopoulos et al. [15] uses blockchain
technology to improve CIDSs and also provides a combined architecture based on the CIDS
and blockchain. This paper proposes a model which considerably reduces the overhead
and volume of the blockchain. However, the author has provided a prototype, or a higher
view, and the model was not tested in the real environment.
Li et al. [16] focused on signature-based collection in their study and proposed a
CBSigIDS, a general framework for a collaborative blockchained, signature-based IDS
that used blockchains to help gradually share and construct a trusted signature database,
inspired by previous blockchain applications. It improved the effectiveness of signaturebased IDSs. However, a major drawback is that it was prone to advanced attacks, and the
need for verification and updates in the blockchain resulted in the diminishing performance
of the overall network.
In recent years, many other IDSs have been proposed [17–21]. These IDSs can be
used in any domain to identify intrusions or abnormalities, which can then lead to the
development of a secure solution for smart cities. In smart cities, everything is connected
to the internet so a smart IDS can play a significant role to provide the security for this
created network. Aloqaily et al. [22] proposed an IDS for securing transportation. This IDS
will help in vehicular service management to secure the network from attacks and ensure
the quality-of-service availability. Elrawy et al. [23] discussed the role of IDSs and the IoT
in the smart environment. This article first discusses various existing works that have
contributed to the smart environment using IoT sensors, and then discusses the existing
IDSs used to provide security in an IoT context. Elsaeidy et al. [24] introduced a smart IDS
to prevent distributed denial of service (DDoS) attacks in smart cities. This article used
-----
_Sustainability 2023, 15, 2133_ 7 of 14
the restricted Boltzmann machines (RBMs) technique to design the IDS. Saba et al. [25]
proposed an ensemble-based IDS for smart city hospitals.
The current IDS is not sophisticated enough to detect distributed, parallel attacks that
take place throughout the nodes in the network instead of a single node. In the case of
CIDSs, the ability to correlate events is crucial. The events occurring across all nodes in the
network must be aggregated for further processing and raising of alerts. The concept of
trust is crucial among the nodes. While the discussed approaches have their advantages,
there was clearly a lack of scalable architecture in the case of the CIDS. The main aim
of this paper is to demonstrate an approach through which a scalable architecture could
be developed, and trust could be established among the nodes in the CIDS architecture.
Blockchain was proposed to solve the trust issue in the case of the CIDS.
**4. Proposed Methodology and Implementation**
In the case of the doorknob-rattling scenario, there is a clear need for a CIDS as
demonstrated by Alexopolous et al. [15]. The doorknob-rattling scenario can be further
explained. In this case, suppose 50 stand-alone nodes in the network are using IDSs and
tracking login attempts. The threshold for an individual machine could be set to 4. Instead
of making 4 incorrect attempts on each node, the attacker would make a series of 2 incorrect
attempts until he successfully logs on to a particular node. To utilize this, the attacker
_Sustainability 2023, 15, x FOR PEER REVIEW uses the common list of user-ids and passwords available. If the system is using an IDS,8 of 17_
these activities would go unnoticed. However, in the case of a CIDS, these activities would
clearly be noticed, and a clear spike would be seen as in Figure 3.
**Figure 3. Figure 3.Doorknob-rattling attack [5]. Doorknob-rattling attack [5].**
Table 2 indicates the parameters necessary for the building of a CIDS system. While
Table 2 indicates the parameters necessary for the building of a CIDS system. While
these are necessary, some of them are complementary to each other. That is, while satisfying
these are necessary, some of them are complementary to each other. That is, while satis
one of the requirements, there would be a high chance of violating the other. For example,
fying one of the requirements, there would be a high chance of violating the other. For
accountability means disclosing some of the information about the node in the system
example, accountability means disclosing some of the information about the node in the
while that would clearly defy the rules of privacy. Hence, there are clear trade-offs between
system while that would clearly defy the rules of privacy. Hence, there are clear trade-offs
one and another.
between one and another.
**Table 2. Requirements of a CIDS system [26].**
Accountability Nodes must be responsible for the actions taken by them.
Integrity Data cannot be manipulated once entered into the system.
Resilience The system should be free from SPoF.
-----
_Sustainability 2023, 15, 2133_ 8 of 14
**Table 2. Requirements of a CIDS system [26].**
Accountability Nodes must be responsible for the actions taken by them.
Integrity Data cannot be manipulated once entered into the system.
Resilience The system should be free from SPoF.
Consensus Nodes in the system must trust the data sent by other nodes.
Scalability The system must be able to scale as the number of nodes increases.
Overhead The overhead cost must be minimized to achieve scalability.
Privacy Privacy must be a concern for the participants in the system.
The main challenge here was to develop a CIDS framework which would decrease
the overhead costs and would be scalable in cases where the size of the network increases.
Blockchain was used to implement the proof-of-concept described earlier. A private
Ethereum-based ledger was used in our case. In order to log the successful attempts made,
pluggable authentication modules were used. The pluggable authentication modules were
used in order to securely log the attempts made in the system so that these attempts
could be transferred to the blockchain as environment variables. In order to use this, the
pam_exec.so was used to run a shell script login_success.sh and the pam_exec file passed
the login information to the shell script to the shell script as environment variables. These
variables were then sent to the log files which were stored automatically. In order to send
data to blockchain, cron utility software was used in Linux which scheduled the transfer of
data from the log files directly to the blockchain by running a python script at a continuous
interval of 5 minutes. If the login was successful, data from the logs would immediately be
sent to the blockchain. This is because a scenario was imagined where the attacker would
gain access to these log files and could tamper or remove them in order to remove the proof
of his presence in the node. In order to simulate an attack, continuous login attempts were
made from different machines using SSH (secure shell) to check whether the attempts were
being logged or not. To make continuous attacks from another host, cronjobs were again
used to call shell script files at an interval of 5 min. This shell script would make a series of
wrong attempts trying to log in as different users in the target machine. This is undertaken
to reduce the overhead cost incurred every time a transaction is made on the blockchain.
In order to make the system more secure, another parameter was used. This was
measuring the CPU utilization of the target machine. There may be a case where an
attacker, after gaining access to the system, would try to run malicious programs. In order
to detect this, CPU utilization was also logged and stored in log files. To measure CPU
utilization, the command used was top | head -3 | tail -1. This command would give
the CPU utilization at that given instant. Our system was designed in such a way that
if the CPU utilization would exceed a given threshold, this would be logged on the log
files. After being logged, these files would be sent to the blockchain at an interval of every
5 min. The spike in the gas cost which can be seen through the Ganache UI would warn
the system administrator of the possible attacks which might be taking place in the node.
**5. Results Discussion**
_5.1. CPU Utilization_
A simple experiment was performed to confirm that our CIDS setup would be able to
precisely record CPU utilization information. The system would record CPU utilization
every minute and record the result. If the CPU utilization exceeded a particular threshold,
for example, 50%, it would get recorded in a log file named cpu.log. To spike the CPU
utilization, another program (prime) was running in the background. The purpose of prime
was to spike the CPU utilization, utilizing 10 threads so that the results could be logged
onto the cpu.log file. Based on the usage of the current system, a threshold was set to 50%
to trigger the sending of data to the log files. This threshold could be modified according to
-----
utilization, another program (prime) was running in the background. The purpose of
_Sustainability 2023, 15, 2133_ prime was to spike the CPU utilization, utilizing 10 threads so that the results could be 9 of 14
logged onto the cpu.log file. Based on the usage of the current system, a threshold was set
to 50% to trigger the sending of data to the log files. This threshold could be modified
the usage of the system and based on user needs. Figureaccording to the usage of the system and based on user needs. Figure 4 illustrates the 4 illustrates the cpu.log recording
all CPU utilization.cpu.log recording all CPU utilization.
**Figure 4. Snapshot from cpu.log recording all CPU utilization.**
**Figure 4. Snapshot from cpu.log recording all CPU utilization.**
_5.2. Login Attempts5.2. Login Attempts_
After the CIDS was set up, the main objective was to capture different authenticationAfter the CIDS was set up, the main objective was to capture different authentication
requests, including the login attempts. Figurerequests, including the login attempts. Figure 5 shows the output of the log file when a 5 shows the output of the log file when a
_Sustainability 2023, 15, x FOR PEER REVIEW_ 10 of 17
person tries to become a super-user. This is determined by the $PAM_RUSER field becauseperson tries to become a super-user. This is determined by the $PAM_RUSER field beboth sudo and su can change the context of the user.cause both sudo and su can change the context of the user.
**Figure 5. Output of auth.log when someone tries to become a super-user.**
**Figure 5. Output of auth.log when someone tries to become a super-user.**
##### Table 3 illustrates the results when the user tries to become a super-user using the Table 3 illustrates the results when the user tries to become a super-user using the
sudo and su commands. sudo and su commands.
**Table 3. The attempts when the user tries to become a super-user using the sudo and su commands.**
**Table 3. The attempts when the user tries to become a super-user using the sudo and su commands.**
**$PAM_USER** **$PAM_TYPE** **$PAM_SERVICE** **$PAM_RUSER** **Date** **Success/Failure**
##### $PAM_USE $PAM_TYP $PAM_SER- $PAM_RUS
Tue 27 MayDate **Success/Failure**
vedant auth **R** **E sudo** **VICE vedant** **ER** Success
01:56:00
##### Tue May 27 vedant auth sudo vedant Tue 27 May Success
vedant auth su vedant 01:59:0001:56:00 Success
##### Tue May 27 vedant auth su vedant Success
Table 4 shows the use of an external host by making use of the ssh command to log01:59:00
onto the CIDS remotely. This example makes use of the $PAM_RHOST field, containingTable 4 shows the use of an external host by making use of the ssh command to log
##### onto the CIDS remotely. This example makes use of the $PAM_RHOST field, containing information about requesting hosts. In the current example, an external agent (192.168.87.4) information about requesting hosts. In the current example, an external agent successfully logged into the CIDS node as ‘vedant’ via the ssh vedant@192.168.87.3 command.
This is a crucial use case as the doorknob-rattling attack typically involves remote users
##### (192.168.87.4) successfully logged into the CIDS node as ‘vedant’ via the ssh
attempting to penetrate the target network [27].
##### vedant@192.168.87.3 command. This is a crucial use case as the doorknob-rattling attack typically involves remote users attempting to penetrate the target network [27].
**Table 4. Login attempt evidence.**
##### Suc- $PAM SER
-----
_Sustainability 2023, 15, 2133_ 10 of 14
**Table 4. Login attempt evidence.**
**$PAM_USER** **$PAM_TYPE** **$PAM_SERVICE** **$PAM_RHOST** **Date** **Success/Failure**
Wed 27 May
vedant auth sshd 192.168.87.3 Success
01:56:23
Using the current method to record the external login attempts evidenced by Table 4,
a simulated doorknob-rattling attack was tried against one of the machines in the network.
During the test, the intruder tried using different user accounts on a single machine. The
FOR PEER REVIEW output from the log files is shown in Figure 6. The attacking machine tried to penetrate11 of 17
the machine thrice using a secure shell (SSH) into each of the user accounts. Each of these
attempts was recorded and sent to the blockchain as transactions [28–30].
_Sustainability 2023, 15, x FOR PEER REVIEW_ 12 of 17
#### Figure 6. Doorknob-rattling attack in a ledger. Figure 6. Doorknob-rattling attack in a ledger.
**2023, 15, x FOR PEER REVIEW**
Each transaction had a varying gas cost based on the number of attempts the intruder
made to penetrate the machine. All records from the log files were pushed onto theA brief summary of the doorknob-rattling attack events in the case of a single attacker
##### is shown in Table 5. The transaction which took place in Ganache can be seen in Figure 7. blockchain either at the end of a specified time interval (in the case where all the attempts
during the interval were failed login attempts) or were sent immediately to the blockchain
**Table 5. if there were any successful logins [Summary of doorknob-rattling attack.31].**
A brief summary of the doorknob-rattling attack events in the case of a single attacker
is shown in Table 5. The transaction which took place in Ganache can be seen in FigureNumber of Login At- 7.
##### User IP Address of Request
tempts
**Table 5.user1 Summary of doorknob-rattling attack.192.168.87.3** 1
##### user2 User IP Address of Request192.168.87.3 Number of Login Attempts2
user6 user1 192.168.87.3 192.168.87.3 1 3
user7 192.168.87.3 2
user2 192.168.87.3 2
##### There was a total of eight login attempts over four different user accounts. Figure 7
user6 192.168.87.3 3
##### shows the transaction which was submitted to the blockchain and the transaction hash
user7 192.168.87.3 2
##### and all the other details.
**Figure 7. Figure 7.Transactional data which were sent to the blockchain. Transactional data which were sent to the blockchain.**
##### Figure 8 shows a list of all the blocks which were mined during the entire process. It
also shows the gas cost incurred during the entire attack
-----
_Sustainability 2023, 15, 2133_ 11 of 14
, 15, x FOR PEER REVIEW There was a total of eight login attempts over four different user accounts. Figure13 of 17 7
shows the transaction which was submitted to the blockchain and the transaction hash and
all the other details.
Figure 8 shows a list of all the blocks which were mined during the entire process. It
also shows the gas cost incurred during the entire attack.
##### Figure 8. List of all the blocks mined and the respective gas costs.
**Figure 8. List of all the blocks mined and the respective gas costs.**
#### The timestamps of these transactions show that the attacks were permanently rec-The timestamps of these transactions show that the attacks were permanently recorded
in the CIDS distributed ledger. The given sequence of events and protection of related data
#### orded in the CIDS distributed ledger. The given sequence of events and protection of re
shows that the nodes can be protected, and the system administrator would be made aware
#### lated data shows that the nodes can be protected, and the system administrator would be
of the possible intrusion immediately. This also proves that Ethereum and Ganache work
#### made aware of the possible intrusion immediately. This also proves that Ethereum and smoothly with Linux and the integration between them to achieve data ingest for intrusion Ganache work smoothly with Linux and the integration between them to achieve data detection is successful. ingest for intrusion detection is successful.
_5.3. Detecting an Anomaly: Thwarting a Doorknob-Rattling Attack_
The main aim of our CIDS architecture was to record data that could be used to detect
#### 5.3. Detecting an Anomaly: Thwarting a Doorknob-Rattling Attack anomalies. This is because data collection at the end is not enough. It needs to be processed
The main aim of our CIDS architecture was to record data that could be used to detect and the analysis should be performed to find potential threats. Subroutines were created
which were scheduled to run at specific time intervals using the cron utility software. This
#### anomalies. This is because data collection at the end is not enough. It needs to be processed
made the traffic steady over a period of time. The cron software utility in Linux was used
#### and the analysis should be performed to find potential threats. Subroutines were created
to schedule the bash script anomaly.sh. The anomaly.sh is used to ensure that 20 to 30 login
#### which were scheduled to run at specific time intervals using the cron utility software. This attempts were made from random users on the virtual machine at an interval of every five made the traffic steady over a period of time. The cron software utility in Linux was used minutes. This was continued for several iterations. All these attacks were logged into our
target machine and data were being sent to the blockchain. The main idea behind this
#### to schedule the bash script anomaly.sh. The anomaly.sh is used to ensure that 20 to 30
approach was that an increase in the number of transactions would mean higher gas cost.
#### login attempts were made from random users on the virtual machine at an interval of
The gas cost for each transaction is further used to analyze whether an attack has actually
#### every five minutes. This was continued for several iterations. All these attacks were taken place or not. Table 6 shows the transactions which took place during the attack and logged into our target machine and data were being sent to the blockchain. The main idea the total gas cost incurred. There were three instances where the number of attacks crossed behind this approach was that an increase in the number of transactions would mean the threshold. The time instances were 10:40, 11:20, and 11:45.
The transactions were then plotted onto a graph with the number of transactions on
#### higher gas cost. The gas cost for each transaction is further used to analyze whether an
the primary axis and the number of attempts on the secondary axis. The peaks show that
#### attack has actually taken place or not. Table 6 shows the transactions which took place the number of transactions crossed the threshold as shown in Figure 9. Figure 10 shows the during the attack and the total gas cost incurred. There were three instances where the bar chart of gas costs in various time frames. Although the inability to detect all the attacks number of attacks crossed the threshold. The time instances were 10:40, 11:20, and 11:45. was problematic, this statistical method did correctly identify that there was an anomaly,
which would lead a system administrator to further investigate.
##### Table 6. Transactions and their respective gas costs in an interval of five minutes.
#### Time Number of Transactions Total Gas Cost
-----
_Sustainability 2023, 15, 2133_ 12 of 14
**Table 6. Transactions and their respective gas costs in an interval of five minutes.**
**Time** **Number of Transactions** **Total Gas Cost**
10:20 1 22,280
10:25 1 22,280
10:40 18 44,184
10:45 1 22,280
_Sustainability 2023, 15, x FOR PEER REVIEW_ 14 of 17
10:50 4 26,152
11:00 4 23,576
11:15 2 23,576
11:45 1 22,280
11:20 8 26,152
11:50 4 26,152
11:25 1 22,280
11:55 1 22,280
11:30 1 22,280
The transactions were then plotted onto a graph with the number of transactions on
the primary axis and the number of attempts on the secondary axis. The peaks show that 11:35 26 54,488
the number of transactions crossed the threshold as shown in Figure 9. Figure 10 shows 11:45 1 22,280
the bar chart of gas costs in various time frames. Although the inability to detect all the 11:50 4 26,152
attacks was problematic, this statistical method did correctly identify that there was an
11:55 1 22,280
anomaly, which would lead a system administrator to further investigate.
60,000 30
50,000 25
40,000 20
30,000 15
20,000 10
10,000 5
0 0
1 2 3 4 5 6 7 8 9 10 11 12 13 14
_Sustainability 2023, 15, x FOR PEER REVIEW_ 15 of 17
**Figure 9.** Graph showing number of attempts and corresponding gas costs at an interval of five
**Figure 9. Graph showing number of attempts and corresponding gas costs at an interval of five minutes.**
minutes.
60,000
50,000
40,000
30,000
20,000
10,000
0
TIME
**Figure 10. Bar chart of gas costs in various time frames.**
**Figure 10. Bar chart of gas costs in various time frames.**
**6. Conclusions and Future Work**
The sharing of information is extremely crucial between the nodes in a CIDS system
-----
_Sustainability 2023, 15, 2133_ 13 of 14
**6. Conclusions and Future Work**
The sharing of information is extremely crucial between the nodes in a CIDS system
in order to prevent the system from attacks as a whole. Information sharing is extremely
important in a scenario where distributed attacks are taking place increasingly. CIDSs,
along with blockchain, appears to be highly suitable for the ingesting of data, especially
in the case of building a smart sustainable city. This paper showed that commercial and
open-source blockchain technologies may be used to create an information-sharing system
that records both doorknob-rattling attacks using pluggable authentication modules and
CPU utilization data as blockchain transactions. This also proves that a blockchain system
can also be used as a logging mechanism for multiple machines and hence can be used
to aggregate data which could be later processed for intrusion detection. This research
provides positive indications that blockchain technology could be used on a large scale for
solving the intrusion detection problem and building a CIDS at a very large scale.
The most significant contribution made in this paper is that it provides an end-to-end
proof-of-concept for CIDS. It also showed at an initial level that attacks or intrusions can be
detected using blockchain as a backbone of the CIDS framework. However, there is a need
to consider the cost of setting up such a system and how sound it is. The proof-of-concept,
which was discussed in the literature, was not implemented at an end-to-end level.
The main aim of this paper was to build an IDS which could be potentially used
to detect system abnormalities and intrusions. There are several avenues which are left
to explore in this paper for additional work. The main aim going further would be to
create a large-scale system which could detect anomalies, block them, and trigger alerts to
the system administrator. Further research is also required to see how the overhead cost
of running the blockchain client would be taken care of. Currently, Ganache (a private
blockchain running at a particular node) is used for testing and carrying out transactions in
the blockchain. Public or other test nets could be used to carry out system tests.
**Author Contributions: Methodology, V.C.; Formal analysis, S.M.; Investigation, R.K.P.; Resources,**
R.K.G.; Writing—original draft, S.B.H.S.; Writing—review & editing, P.K.S. All authors have read and
agreed to the published version of the manuscript.
**Funding: This research received no external funding.**
**Data Availability Statement: The data that support the findings of this study are available on request**
from the corresponding author.
**Conflicts of Interest: The authors declare no conflict of interest.**
**References**
1. [Jose, S. A Survey on Anomaly Based Host Intrusion Detection System. J. Phys. Conf. Ser. 2018, 1000, 1–11. [CrossRef]](http://doi.org/10.1088/1742-6596/1000/1/012049)
2. Li, W. Surveying Trust-Based Collaborative Intrusion Detection: State-of-the-Art, Challenges and Future Directions. IEEE
_[Commun. Surveys Tuts 2021, 280–305. [CrossRef]](http://doi.org/10.1109/COMST.2021.3139052)_
3. [“What Is Hashing? Step-by-Step Guide-Under Hood of Blockchain. August. 2017. Available online: https://blockgeeks.com/](https://blockgeeks.com/guides/what-is-hashing/)
[guides/what-is-hashing/ (accessed on 2 July 2022).](https://blockgeeks.com/guides/what-is-hashing/)
4. Salam, A.-E.; Mohammed, A.; Yousef, S.; Selvakumar, M.; Iznan, H. Intrusion Detection Systems Using Blockchain Technology:
[A Review, Issues and Challenges. Comput. Syst. Sci. Eng. 2021, 40, 87–112. [CrossRef]](http://doi.org/10.32604/csse.2022.017941)
5. Kanth, V.; Mcabee, A.; Tummala, M.; Mceachen, J. Collaborative Intrusion Detection leveraging Blockchain and Pluggable
Authentication Modules. In Proceedings of the 53rd Hawaii International Conference on System Sciences, Maui, HI, USA, 7–10
January 2020.
6. Dreger, H.; Feldmann, A.; Paxson, V.; Sommer, R. Predicting the Resource Consumption of Network Intrusion Detection Systems.
[International Workshop on Recent Advances in Intrusion Detection. 2008. Available online: https://link.springer.com/chapter/](https://link.springer.com/chapter/10.1007/978-3-540-87403-4_8)
[10.1007/978-3-540-87403-4_8 (accessed on 2 July 2022).](https://link.springer.com/chapter/10.1007/978-3-540-87403-4_8)
7. Golomb, T.; Mirsky, Y.; Elovici, Y. CIoTA: Collaborative IoT anomaly detection via Blockchain. arXiv 2018, arXiv:1803.03807.
8. Idé, T. Collaborative Anomaly Detection on Blockchain from Noisy Sensor Data. In Proceedings of the 2018 IEEE International
Conference on Data Mining Workshops (ICDMW), Singapore, 17–20 November 2018; pp. 120–127.
9. Kumari, R.; Catherine, M. Anomaly detection in Blockchain using clustering protocol. Int. J. Pure Appl. Math. 2018, 118, 391–396.
10. Dey, S. Securing majority-attack in blockchain using machine learning and algorithmic game theory: A proof of work. In Proceedings
of the 2018 10th Computer Science and Electronic Engineering (CEEC), Colchester, UK, 19–21 September 2018; pp. 7–10.
-----
_Sustainability 2023, 15, 2133_ 14 of 14
11. Signorini, M.; Pontecorvi, M.; Kanoun, W.; Di-Pietro, R. ADvISE: Anomaly Detection tool for Blockchain Systems. In Proceedings
of the 2018 IEEE World Congress on Services (SERVICES), San Francisco, CA, USA, 2–7 July 2018; pp. 65–66.
12. Steichen, M.; Homme, S.; State, R. ChainGuard—A firewall for blockchain applications using SDN with OpenFlow. In Proceedings
of the 2017 Principles, Systems and Applications of IP Telecommunications (IPTComm), Chicago, IL, USA, 25–28 September 2017;
pp. 1–8.
13. Zhu, L.; Wu, Y.; Gai, K.; Choo, K.R. Controllable and trustworthy blockchain-based cloud data management. Future Gener. Comput.
_[Syst. 2019, 91, 527–535. [CrossRef]](http://doi.org/10.1016/j.future.2018.09.019)_
14. Hu, B.; Zhou, C.; Tian, Y.C.; Qin, Y.; Junping, X. A collaborative intrusion detection approach using Blockchain for multimicrogrid
[systems. IEEE Trans. Syst. Man Cybern. Syst. 2019, 49, 1–11. [CrossRef]](http://doi.org/10.1109/TSMC.2019.2911548)
15. Alexopoulos, N.; Vasilomanolakis, E.; Ivánkó, N.R.; Mühlhäuser, M. Towards Blockchain-based collaborative intrusion detection
systems. In Proceedings of the Critical Information Infrastructures Security 12th International Conference, CRITIS 2017, Lucca,
Italy, 8–13 October 2017; D’Agostino, G., Scala, A., Eds.; Springer: Cham, Switzerland, 2018.
16. Vasilomanolakis, E.; Karuppayah, S.; Mühlhäuser, M.; Fischer, M. Taxonomy and survey of collaborative intrusion detection.
_[ACM Comput. Surv. 2015, 47, 1–33. [CrossRef]](http://doi.org/10.1145/2716260)_
17. Liang, C.; Shanmugam, B.; Azam, S.; Karim, A.; Islam, A.; Zamani, M.; Kavianpour, S.; Idris, N.B. Intrusion Detection System for
[the Internet of Things Based on Blockchain and Multi-Agent Systems. Electronics 2020, 9, 1120. [CrossRef]](http://doi.org/10.3390/electronics9071120)
18. Ghaleb, F.; Saeed, F.; Al-Sarem, M.; Ali Saleh Al-rimy, B.; Boulila, W.; Eljialy, A.E.M.; Aloufi, K.; Alazab, M. Misbehavior-Aware
On-Demand Collaborative Intrusion Detection System Using Distributed Ensemble Learning for VANET. Electronics 2020, 9, 1411.
[[CrossRef]](http://doi.org/10.3390/electronics9091411)
19. Radoglou-Grammatikis, P.I.; Sarigiannidis, P.G.; Efstathopoulos, G.; Panaousis, E.A. A Novel Multivariate Intrusion Detection
[System for Smart Grid. Sensors 2020, 20, 5305. [CrossRef] [PubMed]](http://doi.org/10.3390/s20185305)
20. Iwendi, C.; Anajemba, J.H.; Biamba, C.; Ngabo, D. Security of Things Intrusion Detection System for Smart Healthcare. Electronics
**[2021, 10, 1375. [CrossRef]](http://doi.org/10.3390/electronics10121375)**
21. Kotecha, K.; Verma, R.; Rao, P.V.; Prasad, P.; Mishra, V.K.; Badal, T.; Jain, D.; Garg, D.; Sharma, S. Enhanced Network Intrusion
[Detection System. Sensors 2021, 21, 7835. [CrossRef] [PubMed]](http://doi.org/10.3390/s21237835)
22. Aloqaily, M.; Otoum, S.; Al Ridhawi, I.; Jararweh, Y. An intrusion detection system for connected vehicles in smart cities. Ad Hoc
_[Netw. 2019, 90, 101842. [CrossRef]](http://doi.org/10.1016/j.adhoc.2019.02.001)_
23. Elrawy, M.F.; Awad, A.I.; Hamed, H.F. Intrusion detection systems for IoT-based smart environments: A survey. J. Cloud Comput.
**[2018, 7, 1–20. [CrossRef]](http://doi.org/10.1186/s13677-018-0123-6)**
24. Elsaeidy, A.; Munasinghe, K.S.; Sharma, D.; Jamalipour, A. Intrusion detection in smart cities using Restricted Boltzmann
[Machines. J. Netw. Comput. Appl. 2019, 135, 76–83. [CrossRef]](http://doi.org/10.1016/j.jnca.2019.02.026)
25. Saba, T. Intrusion Detection in Smart City Hospitals using Ensemble Classifiers. In Proceedings of the 13th International
[Conference on Developments in eSystems Engineering (DeSE), Liverpool, UK, 14–17 December 2020. [CrossRef]](http://doi.org/10.1109/DeSE51703.2020.9450247)
26. Zhu, B.; Joseph, A.; Sastry, S. A taxonomy of cyber attacks on scada systems. In Proceedings of the 2011 International Conference
on Internet of Things and 4th International Conference on Cyber, Physical and Social Computing, Washington, DC, USA, 19–22
October 2011; pp. 380–388.
27. [Debar, H.; Dacier, M.; Wespi, A. Towards a taxonomy of intrusion-detection systems. Comput. Netw. 1999, 31, 805–822. [CrossRef]](http://doi.org/10.1016/S1389-1286(98)00017-6)
28. Proffitt, T. How Can You Build and Leverage SNORT IDS Metrics to Reduce Risk? SANS Institute. 2013. Available online:
[https://www.sans.org/reading-room/whitepapers/tools/paper/34350 (accessed on 2 July 2022).](https://www.sans.org/reading-room/whitepapers/tools/paper/34350)
29. [Hu, J. Host-Based Anomaly Intrusion Detection; Springer: Berlin/Heidelberg, Germany, 2010; pp. 235–255. [CrossRef]](http://doi.org/10.1007/978-3-642-04117-4_13)
30. Khan, A.R.; Kashif, M.; Jhaveri, R.H.; Raut, R.; Saba, T.; Bahaj, S.A. Deep Learning for Intrusion Detection and Security of Internet
[of Things (IoT): Current Analysis, Challenges, and Possible Solutions. Secur. Commun. Netw. 2022, 2022, 1–13. [CrossRef]](http://doi.org/10.1155/2022/4016073)
31. Parwani, D.; Dutta, A.; Shukla, P.K.; Tahiliyani, M. Various Techniques of DDoS Attacks Detection & Prevention at Cloud:
A Survey. J. Comput. Sci. Technol. 2015, 8, 110–120.
**Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual**
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3390/su15032133?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/su15032133, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "GOLD",
"url": "https://www.mdpi.com/2071-1050/15/3/2133/pdf?version=1674464426"
}
| 2,023
|
[] | true
| 2023-01-23T00:00:00
|
[
{
"paperId": "1d3d953363a6a3070c3a87c940af432edba24e18",
"title": "Deep Learning for Intrusion Detection and Security of Internet of Things (IoT): Current Analysis, Challenges, and Possible Solutions"
},
{
"paperId": "39b7dbf60eaace5503793ab39d2b48e25101e816",
"title": "Enhanced Network Intrusion Detection System"
},
{
"paperId": "c1cf420379a410014a5307e46ef8307b3cecfb1c",
"title": "Security of Things Intrusion Detection System for Smart Healthcare"
},
{
"paperId": "ae82780a0deb576c647393a20392867e3657173a",
"title": "Misbehavior-Aware On-Demand Collaborative Intrusion Detection System Using Distributed Ensemble Learning for VANET"
},
{
"paperId": "a7fb6fd0076b3026d9bb37e1d52e8fb6b3a68853",
"title": "ARIES: A Novel Multivariate Intrusion Detection System for Smart Grid"
},
{
"paperId": "e50d6a9f5ea1be9eb5c8bfd55213c193e553a195",
"title": "Intrusion Detection System for the Internet of Things Based on Blockchain and Multi-Agent Systems"
},
{
"paperId": "94554103c275907d052fc87598c032ca041a74c9",
"title": "An intrusion detection system for connected vehicles in smart cities"
},
{
"paperId": "2ef192b8fb19fd6b2366de6ee6869e49eb41bd08",
"title": "Intrusion detection in smart cities using Restricted Boltzmann Machines"
},
{
"paperId": "809d947cf0f222ecb58e42a8be69c0dc00042880",
"title": "A Collaborative Intrusion Detection Approach Using Blockchain for Multimicrogrid Systems"
},
{
"paperId": "d9b5194f3f959eda2e95df6a340254f52ced46f4",
"title": "Controllable and trustworthy blockchain-based cloud data management"
},
{
"paperId": "a7eea1d6dfbcee9a434e9ab6c9a7f67f1506d3db",
"title": "Intrusion detection systems for IoT-based smart environments: a survey"
},
{
"paperId": "6d7b1e32a3c8bb8cf4f51f4c0561146c437bc766",
"title": "A Survey on Anomaly Based Host Intrusion Detection System"
},
{
"paperId": "00be303628cc63a79cf6ffd89402a07525dab4fe",
"title": "CIoTA: Collaborative IoT Anomaly Detection via Blockchain"
},
{
"paperId": "e11d5a4edec55f5d5dc8ea25621ecbf89e9bccb7",
"title": "Taxonomy and Survey of Collaborative Intrusion Detection"
},
{
"paperId": "62fa81980219842a6c5945cc2d3f5b142240e3eb",
"title": "Towards a taxonomy of intrusion-detection systems"
},
{
"paperId": "2899af503b92d4feea21468549013fc63ad72618",
"title": "Intrusion Detection Systems Using Blockchain Technology: A Review, Issues and Challenges"
},
{
"paperId": "13a12faf033612507e6364faeb2bc6aa6958d7d5",
"title": "Anomaly Detection in Block chain Using Clustering Protocol"
},
{
"paperId": null,
"title": "Various Techniques of DDoS Attacks Detection & Prevention at Cloud: A Survey"
}
] | 14,329
|
en
|
[
{
"category": "Economics",
"source": "s2-fos-model"
},
{
"category": "Business",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/021ecd84d3d3c6ddb6874023f60808dc635bc2c2
|
[] | 0.885968
|
Is Bitcoin an emerging market? A market efficiency perspective
|
021ecd84d3d3c6ddb6874023f60808dc635bc2c2
|
Central European Economic Journal
|
[
{
"authorId": "2237636921",
"name": "Mateusz Skwarek"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"Central Eur Econ J"
],
"alternate_urls": [
"https://content.sciendo.com/view/journals/ceje/ceej-overview.xml"
],
"id": "40fcf287-61ca-4a3d-9606-874810723e41",
"issn": "2543-6821",
"name": "Central European Economic Journal",
"type": null,
"url": "https://content.sciendo.com/view/journals/ceej/ceej-overview.xml"
}
|
Abstract Despite recent studies focused on comparing the dynamics of market efficiency between Bitcoin and other traditional assets, there is a lack of knowledge about whether Bitcoin and emerging markets efficiency behave similarly. This paper aims to compare the market efficiency dynamics between Bitcoin and the emerging stock markets. In particular, this study indicates whether the dynamics of Bitcoin market efficiency mimic those of emerging stock markets. Thus, the paper's contribution emerges from the combination of Bitcoin and emerging markets in the field of dynamics of market efficiency. The dynamics of market efficiency are measured using the Hurst exponent in the rolling window. The study uses daily data for the MSCI Emerging Markets Index and the Bitcoin market over the period 2011–2022. Our results show that there is at most a moderate correlation between the dynamics of Bitcoin and emerging stock markets’ efficiency over the entire study period. The strongest correlations occur mainly in periods of high economic policy uncertainty in the largest Bitcoin mining countries. Therefore, the association between Bitcoin market efficiency and emerging stock markets’ efficiency may strengthen with an increase in economic policy uncertainty. These findings may be useful for investors and portfolio managers in constructing better investment strategies.
|
**ISSN:** 2543-6821 (online)
[Journal homepage: http://ceej.wne.uw.edu.pl](http://ceej.wne.uw.edu.pl)
## **Mateusz Skwarek**
# **Is Bitcoin an emerging market?
** **A market efficiency perspective**
**To cite this article**
Skwarek, M. (2023). Is Bitcoin an emerging market? A market efficiency
perspective. Central European Economic Journal, 10(57), 219-236.
**DOI:** 10.2478/ceej-2023-0013
[To link to this article: https://doi.org/10.2478/ceej-2023-0013](https://doi.org/10.1515/ceej-2018-0003)
-----
##### **Mateusz Skwarek
**
Poznań University of Economics and Business, Institute of Accounting and Finance
Management, al. Niepodległości 10, 61-875 Poznań, Poland,
corresponding author: Mateusz.Skwarek@phd.ue.poznan.pl
### **Is Bitcoin an emer in market? A market efficienc ers ective** **g g y p p**
**Abstract**
Despite recent studies focused on comparing the dynamics of market efficiency between Bitcoin and other traditio-
nal assets, there is a lack of knowledge about whether Bitcoin and emerging markets efficiency behave similarly. This
paper aims to compare the market efficiency dynamics between Bitcoin and the emerging stock markets. In particular,
this study indicates whether the dynamics of Bitcoin market efficiency mimic those of emerging stock markets. Thus,
the paper’s contribution emerges from the combination of Bitcoin and emerging markets in the field of dynamics of
market efficiency. The dynamics of market efficiency are measured using the Hurst exponent in the rolling window.
The study uses daily data for the MSCI Emerging Markets Index and the Bitcoin market over the period 2011–2022.
Our results show that there is at most a moderate correlation between the dynamics of Bitcoin and emerging stock
markets’ efficiency over the entire study period. The strongest correlations occur mainly in periods of high economic
policy uncertainty in the largest Bitcoin mining countries. Therefore, the association between Bitcoin market effi-
ciency and emerging stock markets’ efficiency may strengthen with an increase in economic policy uncertainty. These
findings may be useful for investors and portfolio managers in constructing better investment strategies.
**Keywords**
bitcoin | market efficiency | emerging stock markets | long-range dependence | Hurst exponent
**JEL Codes**
G11, G14, G15
#### **1. Introduction**
Bitcoin is the largest (when it comes to capitalisation)
and the most researched cryptocurrency in the context
of informational efficiency (Urquhart, 2016; Bariviera,
2017; Kristoufek, 2018; Kumar & Zargar, 2019; Tran
& Leirvik, 2020; Noda, 2021). The majority of these
studies have confirmed that the Bitcoin market is the
least inefficient among cryptocurrencies. Thus, Bitcoin
seems to be the most mature market and representative
cryptocurrency (in terms of researchers’ and investors’
attention). However, similarly to other cryptocurrency
markets, many previous studies also indicate that the
Bitcoin market is still inefficient (e.g. Kosc, Sakowski
& Ślepaczuk, 2019).
The market is efficient when investors are not able
to earn abnormal returns based on their past values
(Fama, 1970). In other words, the market prices include
all information. However, changes in market conditions
and behavioural biases may make that market efficiency
dynamic. For example, loss aversion may affect
investor decision-making under business uncertainty
(Kahneman & Tversky, 1979). The Adaptive Markets
Hypothesis combines these components and assumes
that investors learn from their mistakes. After some
time (change in market conditions), investors adapt
to this new environment and then the market may be
very close to efficiency. But market conditions vary
over time, leading to behavioural biases of investors
(e.g. overconfidence, overreaction) and, in effect, the
dynamics of market efficiency can be observed (Lo,
2004). For example, Lim, Brooks and Kim (2008)
find the dynamics of stock market efficiency during
different market conditions. So it seems that changes
in economic uncertainty related to different market
conditions affect market efficiency.
So far, there is no comprehensive answer to the
question of whether the Bitcoin market efficiency is
developing better than that of emerging markets. This
paper aims to fill this gap by comparing the dynamics
of market efficiency in the Bitcoin market with that of
the emerging stock markets. This may help investors
to allocate capital more efficiently. In particular, the
survey of the relationship between emerging markets
-----
CEEJ **• 10** (57) **•** 2023 **•** pp. 219-236 **•** ISSN 2543-6821 **•** DOI: 10.2478/ceej-2023-0013 **221**
and Bitcoin over time could indicate whether a
portfolio’s chance of obtaining a given return is
greater by including both markets. On the other hand,
this analysis from a market efficiency perspective
may show in which sub-periods there is a delay in the
price’s reaction to information and what the size of
the price’s deviation from the random walk process
is. Thus, it could be used by investors in constructing
better investment strategies.
The latest research on cryptocurrency market
efficiency has focused on the resilience of this Bitcoin
market efficiency to global shocks such as the Covid-19
pandemic. Phiri (2022) shows that the pandemic has
affected the dynamics of Bitcoin market efficiency. A
similar view is developed by Fernandes et al. (2022), who
conclude that the response of Bitcoin market efficiency
to Covid-19 is different from other markets. This
evidence is supported by others (Wang et al., 2021; DinizMaganini, Diniz & Rasheed, 2021; Mensi et al. 2022)
who also compare the dynamics of market efficiency
in Bitcoin with developed markets and traditional
investments. The above-mentioned studies do not
include emerging stocks in this context. However, Lim et
al. (2008) confirm that some emerging markets exhibit
higher market inefficiency in times of financial crisis.
Baur and McDermott (2010) note that large emerging
markets react differently to economic shock (compared to
developed stock markets). Thus, the dynamics of market
efficiency of both the emerging and Bitcoin markets
seem to be affected by unexpected events in different
ways than developed stock markets. Because of this, a
study on the resilience of Bitcoin and emerging markets’
efficiency to economic shocks is needed. Therefore,
this paper joins a discussion on the comparison of the
dynamics of Bitcoin market efficiency with the dynamics
of market efficiency in other traditional markets.
Existing studies cover the relationship between
Bitcoin and emerging markets in other aspects than
dynamics of market efficiency (Carrick, 2016; Bouri
et al., 2017; Shahzad et al., 2019; Mizerka, StróżyńskaSzajek & Mizerka, 2020; Bouri et al., 2020).
Specifically, the studied areas include the dependence
between Bitcoin and emerging markets returns, the
co-movement of markets at different time horizons,
the predictability of asset returns from stock market
returns. The majority of these studies suggest that
this association is weak and may be time-varying.
The reaction of the correlation between the markets
and economic shocks may affect the losses of their
investors during these events (Baur & McDermott,
2010). In particular, they notice the reactions of
emerging markets, their difference from developed
markets, during extreme events. However, there is a
lack of empirical analysis of the relationship between
the dynamics of Bitcoin market efficiency and the
dynamics of emerging markets’ efficiency over time.
Therefore, the studies of the association between
markets during economic shocks should be deepened.
Several studies indicate some similarities between
Bitcoin and emerging markets from a market
efficiency perspective. Urquhart (2016), Bariviera
(2017), and Takaishi and Adachi (2020) find that the
Bitcoin market has become more efficient over time.
This trend in market efficiency is also documented in
the case of some emerging stock markets by Cajueiro
and Tabak (2004), Sukpitak and Hengpunya (2016), and
Hkiri et al. (2021). In this context, it can be concluded
that some emerging markets and Bitcoin become
more efficient in the years 2015–2016. However, in
the Chinese and Bitcoin markets during the years
2014–2015, bubble-like price dynamics could be
observed. According to Kristoufek (2018), the Bitcoin
market is efficient only after price bubbles, that is,
during low Bitcoin price dynamics. Some studies (e.g.
Lim et al., 2008; Hull & McGroarty, 2014) also report
that emerging markets exhibit higher inefficiency in
certain periods. Motivated by various results, this
paper verifies whether both emerging and Bitcoin
markets will become more efficient over time.
The research gap consists of several strands.
Firstly, it concerns the comparison of Bitcoin with
emerging stock markets’ efficiency. In particular, the
relative dynamics of market efficiency in both markets
have not been analysed in times of different turmoils.
Furthermore, so far, the relationship between Bitcoin
and emerging markets’ efficiency have not been
linked to the high economic policy uncertainty events
in its largest mining countries during this period.
Besides, to the best of the author’s knowledge, the
rolling correlation between the dynamics of market
efficiency of Bitcoin and emerging markets has not
been measured. It may help to compare the relative
chance of profitable investment strategies in different
markets at some time horizon. Thus the following
research question was asked: What is the relationship
between Bitcoin and emerging stock markets from the
market efficiency perspective?
The main purpose of this article is to compare the
dynamics of market efficiency between Bitcoin and
the emerging stock markets. For this aim, the weak
form of market efficiency is analysed over time by
applying the Hurst exponent in the rolling window.
-----
CEEJ **• 10** (57) **•** 2023 **•** pp. 219-236 **•** ISSN 2543-6821 **•** DOI: 10.2478/ceej-2023-0013 **222**
Specifically, the dataset consists of daily closing prices
of the MSCI Emerging Markets Index and the Bitcoin
market from the period 2011 to 2022. Finally, the
correlation coefficient is measured between two-time
series of Hurst exponents on a rolling window. In
effect, the dynamic relationship between the Bitcoin
market efficiency and the emerging stock markets’
efficiency is shown. Thus, it is indicated whether
Bitcoin and emerging stock markets show some
similarities in their efficiency.
Findings show that there is a moderate correlation
between the market efficiency of Bitcoin and emerging
stock markets at some periods (the strong value of
the correlation is not confirmed in the additional
tests). This is in contrast to previous research on the
association between Bitcoin and emerging market
returns. Bitcoin market efficiency and emerging
markets’ efficiency report the most common
fluctuations in periods of large economic policy
uncertainty. Specifically, the jumps in correlation
values occurred in the periods: the threat of a spillover
of the euro area crisis in 2012, the threat of the US
debt crisis at the end of 2013, Russia’s aggression
against Ukraine on February 2014, China’s economic
downturn in 2015, the USA – China trade tensions in
the years 2018–2019, Covid-19 in 2020 and the Russia
– Ukraine war in 2022.
Thus, the results can be assigned to the events
related to the high economic policy uncertainty of
the largest Bitcoin mining countries (e.g. China, the
USA). For example, the uncertainty related to the
US presidential election results in 2016 (Trump’s
election), the USA – China trade policy tensions in the
years 2018–2019 and the threat of Covid-19 may be
associated with the jumps in the values of correlation.
The identified economic shocks extend the conclusions
of the existing research on the dynamics of Bitcoin
market efficiency because recent studies mainly verify
the importance of Covid-19 for the dynamics of
market efficiency as a global crisis. It can be supposed
that the economic shocks of the largest Bitcoin mining
countries also would have an impact on the dynamics
of Bitcoin market efficiency and its relationship
with the dynamics of emerging markets’ efficiency.
Therefore, investors should pay attention to the role of
high economic policy uncertainty of these countries
in the profitability of their portfolio diversification,
which includes Bitcoin and emerging markets.
The contribution of this study is at least threefold.
Firstly, this paper adds to the previous literature by
comparing the dynamics of market efficiency in Bitcoin
and emerging stock markets. In the existing research,
there is no clear evidence of the ‘emerging’ nature
of the cryptocurrency market efficiency. However,
researchers refer to cryptocurrencies as an emerging
market (Alvarez-Ramirez, Rodriguez & Ibarra-Valdez,
2018; Khuntia & Pattanayak, 2018; Kumar & Zargar,
2019). Inappropriate classification of the Bitcoin market
may cause investors to treat it as less risky than it is.
Therefore, this study contributes to the possibility of
a better allocation, as it shows the actual and relative
level of both Bitcoin market maturity and the degree of
predictability of the returns time series of the studied
markets. Thus, it is important to verify that the emerging
market is the proper category, in the case of Bitcoin.
Secondly, the study extends a discussion on the
resilience of market efficiency to economic shocks.
The more resilience of market efficiency of one asset
from another could be a potential attribute of the
safe haven (Wang et al., 2021). In other words, the
safe haven could be identified by the observation
of negative predictability from the stock market
to the (safe haven) asset or by the fact that losses
from one investment are compensated by gains
from another (Shahzad et al., 2019). This is the first
study to compare the market efficiency dynamics
of emerging stock markets and Bitcoin in different
periods of economic shocks. Recent studies focus
on the relationship between Bitcoin and emerging
economies in the context of portfolio diversification
opportunities (Bouri et al., 2017; Shahzad et al., 2019;
Mizerka et al., 2020; Bouri et al., 2020). The low (or
negative) correlation between Bitcoin and other assets
may indicate benefits from portfolio diversification,
especially during periods of market stress that may
be characterised by a different herd behaviour of
investors (because of different perceptions of the
impact of a given shock on markets). However, there
is still a research gap in this phenomenon from a
market efficiency perspective. The findings show
whether the chance to obtain profitable strategies
based on historical quotations in one market may be
higher than in another. If both markets are included
in one portfolio, the low (or negative) correlation
between the degree of the predictability of returns
in these markets may indicate potential benefits of
a safe haven from the investment strategies based
on the market performance; that is, economic shock
effects on both markets differ in terms of degree
and/or nature of dependence (momentum/meanreversion) in return time series in a given time. Thus,
in terms of practical contribution, this study may
help investors in developing better diversification
-----
CEEJ **• 10** (57) **•** 2023 **•** pp. 219-236 **•** ISSN 2543-6821 **•** DOI: 10.2478/ceej-2023-0013 **223**
investment strategies. For example, the largest positive
correlations between Bitcoin and emerging markets’
efficiency in a period of market stress confirm that
investment strategies based on the historical returns
obtained in these markets should rather assume a
reduction of the share of these investment assets in
the portfolio during some unexpected events.
Thirdly, this paper contributes to the literature
on the dynamics of Bitcoin market efficiency and its
potential factors. Despite many recent studies on the
dynamics of Bitcoin market efficiency, there is no
comprehensive evidence on whether the dynamics of
Bitcoin market efficiency are related to uncertainty
(Wang et al., 2021; Diniz-Maganini et al., 2021; Mensi
et al. 2022; Phiri, 2022; Fernandes et al., 2022; Mnif,
Mouakhar & Jarboui, 2023). In addition, so far it has
not been verified in the context of emerging stock
markets. This research shows that high economic
uncertainty potentially affects the changes in both
Bitcoin and emerging stock markets’ efficiency. In
effect, investors should take the economic policy
uncertainty of the largest emerging countries in
Bitcoin mining into consideration . Thus, the novelty
of this paper arises from the combination of the
dynamics of Bitcoin and emerging markets’ efficiency
and uncertainty. This research makes a theoretical
contribution by explaining the co-movements in the
markets’ efficiency dynamics of different investment
assets by the sub-optimal investor reaction to the
high economic policy uncertainty events concerning
the largest countries in terms of the capital flows
between these markets. In other words, the reason
for the increase in the correlation between markets’
efficiency may be that investors are more subject to
the representativeness heuristic in times of high
economic uncertainty events. Thus, this study deepens
understanding of the Adaptive Markets Hypothesis.
The structure of the article is as follows: The
first section is the introduction. The second part
presents the literature background. Next, the data
and methodology applied in this paper are described.
The third section reports the results. The fourth part
of the article consists of additional analyses. The last
sections are the discussion and the conclusion.
#### **2. Literature review**
Bitcoin is the most popular cryptocurrency and the
largest in terms of market capitalisation, which accounts
for about 43% of the cryptocurrency market share
(January 4, 2023). The purpose of the creation of Bitcoin
is to be used as a payment system. In fact, some users
treat Bitcoin as an alternative currency or even a store of
value (Polasik et al., 2015). However, most participants
in the cryptocurrency market perceive it as a speculative
investment (Hileman & Rauchs, 2017, p. 24).
One of the most discussed issues in the context
of Bitcoin is its market efficiency. This popular topic
has been studied for many years in the stock markets
since Fama (1970) formulated the efficient market
hypothesis. According to Fama (1970), the efficient
market hypothesis (EMH) means it is impossible to use
past prices to predict future prices (weak form). Thus,
it refers to informational efficiency (Czekaj, Woś &
Żarnowski, 2001, 30) which is an important global
problem, because the growth in market efficiency may
lead to a better allocation of capital (both from the
global and individual investors’ perspectives).
The majority of early studies on Bitcoin market
efficiency report that its price behaviour is nonrandom and characterised by dynamics. Urquhart
(2016) and Bariviera (2017) applying the Hurst
exponent showed that the Bitcoin market was
inefficient in the years 2010–2016/2017 and that lately,
there was a trend toward an efficient market. Other
researchers (Aggarwal, 2019; Bouri et al., 2019; Jiang,
Nie & Ruan, 2018; Kumar & Zargar, 2019; Takaishi &
Adachi, 2020) also confirmed that the inefficiency of
the Bitcoin market varies over time. Several of them
found a long memory of Bitcoin returns, which signals
a positive autocorrelation (e.g. Alvarez-Ramirez et al.,
2018). Thus, the above-mentioned evidence suggested
that Bitcoin may become more efficient over time.
However, it seems to be still an inefficient market
with the presence of long memory.
Similar results can be found in the context of
emerging markets (Cajueiro & Tabak, 2004; Sukpitak
& Hengpunya, 2016; Hkiri et al., 2021). These studies
confirmed that emerging markets have become
more efficient over time. Hull and McGroarty
(2014), however, noticed that the emerging markets’
efficiency was time-varying and characterised by a
long-memory process most of the time. Therefore, the
following research question was addressed: What is
the relationship between Bitcoin and emerging stock
markets from the market efficiency perspective?
Despite recent papers mainly contradicting
Bitcoin market efficiency, many of them also focus
on the factors of market efficiency (e.g. Brauneis &
Mestel, 2018; Wei, 2018; Köchling, Müller & Posch,
-----
CEEJ **• 10** (57) **•** 2023 **•** pp. 219-236 **•** ISSN 2543-6821 **•** DOI: 10.2478/ceej-2023-0013 **224**
2019; Khuntia & Pattanayak, 2020; Takaishi & Adachi,
2020; Noda, 2021; Phiri, 2022) or the relationship
between the dynamics of market efficiency of Bitcoin
and traditional financial assets (e.g. Al-Yahyaee, Mensi
& Yoon, 2018; Plastun et al., 2019; Diniz-Maganini et
al., 2021; Wang et al., 2021; Mensi et al., 2022) or other
cryptocurrencies (e.g. Caporale, Gil-Alana & Plastun,
2018; Wei, 2018; Borowski & Matusewicz, 2019; Aslan
& Sensoy, 2020; Noda, 2021; Assaf et al., 2022). In this
context, several researchers (Brauneis & Mestel, 2018;
Wei, 2018; Noda, 2021) found that Bitcoin was the
least inefficient compared to other cryptocurrencies.
Therefore, taking that this is the most studied,
least inefficient, and largest cryptocurrency into
consideration, it seems to be the most suitable
representative of the cryptocurrency market with
which to examine the dynamics of market efficiency.
Recent studies compare Bitcoin market efficiency
to the efficiency of traditional assets such as gold,
currencies, bonds, stock markets, and commodities
(Al-Yahyaee et al., 2018; Plastun et al., 2019; Wang et
al., 2021; Diniz-Maganini et al., 2021; Mensi et al., 2022;
Chowdhury et al., 2023). Most of them indicate that
the size of the Bitcoin market inefficiency is different
from other markets. However, several studies confirm
that these markets exhibit some similarities when it
comes to market price reactions to economic shocks
such as Covid-19. Mensi et al. (2022) and Wang et al.
(2021) documented that the inefficiency of Bitcoin and
other studied markets of traditional financial assets
increased during the time of Covid-19. Lim et al. (2008)
also support these findings in the case of the reaction of
emerging market efficiency to a financial shock. Thus,
it can be expected that in times of market turmoil, the
correlation between the market efficiency of traditional
emerging markets and Bitcoin strengthens and the
market efficiency deteriorates in both cases.
Another interesting conclusion can be drawn in
the context of the market efficiency resistance to high
economic policy uncertainty. For example, Wang et
al. (2021), Diniz-Maganini et al. (2021), and Mensi et
al. (2022) observed that during Covid-19 the increase
in the Bitcoin market inefficiency was smaller than
for the other studied markets, which could be an
attribute of a safe haven (Wang et al., 2021). A similar
conclusion was developed by Fernandes et al. (2022),
who stated that the dynamics of cryptocurrency
market efficiency is robust to unpredictable shocks
such as Covid-19. In contrast to them, Phiri (2022)
obtained findings that contradict the resistance of
the dynamics of Bitcoin market efficiency to shocks.
Recently, Rufino (2023) confirmed that Bitcoin market
efficiency deteriorated during the pandemic period.
This is supported by Mnif et al. (2023), who also
reported that during unexpected events such as the
Russia-Ukraine war, the Bitcoin market inefficiency
increases. However, Chowdhury et al. (2023) noticed
that during the Covid-19 period, the market efficiency
of the S&P 500 changed more than did that of Bitcoin.
Thus, the majority of results provide evidence that
supports the greater resilience of Bitcoin market
efficiency to economic shocks compared to the
markets of traditional investment assets. There is still
no consistency, however, concerning whether Bitcoin
market efficiency is robust to unexpected events.
So far, the studied factors of cryptocurrency market
efficiency include liquidity (Brauneis & Mestel, 2018;
Wei, 2018; Köchling et al. 2019; Takaishi & Adachi,
2020; Noda, 2021), halving (Phiri, 2022), market
capitalisation (Brauneis & Mestel, 2018), or trading
volume (Khuntia & Pattanayak, 2020). Although
the latest research indicates that in the case of a
speculative bubble or global crisis, there is an increased
comovement of Bitcoin and most cryptocurrencies
(Assaf et al., 2022) or traditional investments efficiency
(Wang et al., 2021; Mensi et al., 2022). This has not
been verified yet for the market efficiency of Bitcoin
and emerging markets. Moreover, some researchers
(Czarnecki, Grech & Pamuła, 2008; Hkiri et al., 2021)
have noticed that the behaviour of the Hurst exponent
of the developing stock markets may be related to
the financial or political crisis. Therefore, it can be
assumed that, similar to other traditional markets,
the dynamics of emerging markets’ efficiency may
be more related to the dynamics of Bitcoin market
efficiency during times of high economic policy
uncertainty events. This is supported by large capital
flows between emerging markets and Bitcoin in the
context of cryptocurrency mining (Statista, 2022).
Only Plastun et al. (2019) examined emerging
markets’ efficiency and Bitcoin market efficiency.
Specifically, they compared the market efficiency
of two emerging markets (Russia and Ukraine) to
Bitcoin market efficiency. Thus, they did not take
China into account, which is the largest emerging
market (the share of China in the MSCI Emerging
Markets index was about 30% in late 2022) and one
of the most important countries in the case of Bitcoin
mining. Therefore, further studies which will include
the largest emerging markets in this field are needed.
Plastun et al. (2019) also concluded that these
markets exhibited different degrees of persistence in
-----
CEEJ **• 10** (57) **•** 2023 **•** pp. 219-236 **•** ISSN 2543-6821 **•** DOI: 10.2478/ceej-2023-0013 **225**
returns for days of the week in the years 2014–2018,
which is contrary to the efficient market hypothesis.
Moreover, the majority of other studies suggested that
the association between emerging markets and Bitcoin
is weak and may be time-varying (Carrick, 2016;
Shahzad et al., 2019; Bouri et al., 2020). Along these
lines, it can be expected that the relationship between
emerging markets and Bitcoin from a perspective of
dynamics of market efficiency is weak for the whole
period. On the other hand, Plastun et al. (2019) used
the Hurst exponent for emerging markets in crosssection on different days of the week. This static
approach does not include the dynamics of correlation
and volatility clustering, changes in the underlying
process which drives Bitcoin prices (Aggarwal, 2019).
Thus, the results obtained by Plastun et al. (2019)
should be verified by applying a different approach,
such as dynamic correlation (e.g. sliding window).
Besides, this evidence suggests that the relationship
between emerging markets and Bitcoin efficiency may
be time-varying.
To sum up, most studies of cryptocurrencies’
market efficiency have been conducted on the Bitcoin
market. This cryptocurrency has the largest share of
the market and is the least inefficient. Furthermore,
previous research looking at Bitcoin from a
market efficiency perspective have focused on its
relationship with other cryptocurrencies, uncertainty,
or traditional assets such as gold, currencies,
commodities, and developed countries’ stocks. These
studies don’t take the large capital flows between
emerging countries and Bitcoin into account. As an
effect, no evidence includes the largest emerging stock
markets from this perspective. However, the majority
of them suggest that the association between emerging
markets and Bitcoin is weak and may be time-varying
(Carrick, 2016; Shahzad et al., 2019; Bouri et al., 2020).
In particular, some studies conducted from a market
efficiency perspective find separately that both Bitcoin
(Urquhart, 2016; Bariviera, 2017; Takaishi & Adachi,
2020) and the emerging markets (Cajueiro & Tabak,
2004; Sukpitak & Hengpunya, 2016; Hkiri et al., 2021)
have become more efficient over time. Thus, the
research question concerns the relationship between
Bitcoin and emerging stock markets’ efficiency.
#### **3. Data and methodology**
In line with Urquhart (2016), Bariviera (2017),
Kristoufek (2018), and Jiang et al. (2018), logarithmic
returns are calculated to provide time series for analysis
of market efficiency. To verify the market efficiency, the
Hurst exponent is adopted. It is a measure of long-range
dependence. Following Urquhart (2016) and Bariviera
(2017), the Hurst exponent is calculated using the
rescaled range analysis (R/S). According to Kristoufek
(2010), this method can be represented as an analysis of
the rescaled range of a time series for different scales
of a given length. In effect, there is a dependence on a
distraction (range - *R* ) from different lengths of scale ( *i* ).
Briefly, this relation is presented below:
*R* / *S* = *a* ∙ *i* *[H]* (1)
where *H* is the Hurst exponent, *S* is the standard
deviation of the sums of departures of returns from
the average in a given period, *R* (range) is a difference
between the maximum and minimum of the sums of
deviations from the average in each subinterval of ‘ *i* ’
length, ‘ *a* ’ is a constant.
When the above relationship imitates a linear
trend in a double-logarithmic scale, there is a random
walk of the time series. So, if the Hurst exponent
equals 0.5, the market is efficient. The value of the
Hurst exponent of more than 0.5 means that the time
series is long memory persistent. On the other hand,
when the value of the Hurst exponent is less than 0.5
it can be interpreted as a mean-reversion property of
a time series.
As pointed out by Kristoufek (2010), the standard
deviations for the rescaled range analysis are smaller
compared to the detrended fluctuation analysis (DFA)
which is a very popular alternative in this case.
However, he states that in general, the results of both
methods are quite similar. Furthermore, Kristoufek
(2010) recommends applying a minimum scale of 16
observations and a minimum length of time series
equal to 512 data points in the case of R/S. He argues
that too-small scales can lead to an incorrect value
of the standard deviation (bias), which is used to
rescale the ranges during the estimation of the Hurst
exponent. However, too-large scales may cause the
impact of extreme values to be underestimated. Thus,
the minimum scale of 16 and the length of 512 for the
time series are used. In effect, to show the dynamics
of market efficiency, we calculate the Hurst exponent
over a rolling window of 512 data points (a fixed size)
with one-day step. This is comparable to the two-year
window exploited by Bariviera (2017).
Similar to Polanco-Martínez (2019), to present
a dynamic relationship between two variables –
-----
CEEJ **• 10** (57) **•** 2023 **•** pp. 219-236 **•** ISSN 2543-6821 **•** DOI: 10.2478/ceej-2023-0013 **226**
the Hurst exponents of the Bitcoin and the MSCI
Emerging Markets Index – the rolling window
correlation coefficients are estimated. Specifically, the
Spearman rank correlation with p-values is exploited,
because this is more robust to non-linear relationships
of the analysed data series. The reasons for applying
the rolling window correlation are the presence of
volatility clustering (Bariviera, 2017) and structural
breaks (Jiang et al., 2018) in financial time series
which could signal nonlinear patterns in the dynamics
of market efficiency. The correlations are based on
the series of Hurst exponents. For example, it means
that the dynamic correlation coefficient for the first
of January 2021, refers to the behaviour of market
returns in the previous period of two years plus the
window size for the rolling correlation (251 or 126
data points).
Finally, the robustness of the results is verified
in several ways. As proposed by Polanco-Martínez
(2019), the dynamic correlation is applied for different
window sizes. On the one hand, the small length
of the time series which is used to compute the
correlation could influence the significance of results.
On the other hand, the use of larger window sizes
may mean that the one correlation value includes the
impact of several ‘unpredictable’ events (which are
rare), so it will be difficult to isolate the importance of
one event for the studied association. Therefore, only
two window sizes are used: 126 and 251 data points.
Besides, Kendall’s correlation is applied to verify
whether the results are robust.
Similar to Borowski and Matusewicz (2020),
detrended fluctuation analysis (DFA) is also adopted
to provide additional estimates for market efficiency.
In contrast to the rescaled range analysis, DFA exploits
the squared fluctuations function that is a measure of
variability (instead of the range). Additionally, the
overlapping rolling window of 512 observations with
a minimum scale of 16 data points is used. As an effect,
the dynamic Spearman correlation is applied to the
Hurst exponents based on the DFA method.
The dataset consists of daily closing prices of
Bitcoin and the MSCI Emerging Markets Index in the
period September 13, 2011, through August 11, 2022.
This period is limited by the availability of quotations
from the Bitstamp exchange. Another reason for the
length of this period is that it includes the largest
changes in economic policy uncertainty, which allow
us to study the price’s reaction to different levels of
information uncertainty. Similar to Bariviera et al.
(2017) and Takaishi and Adachi (2020), the Bitcoin
data from Bitstamp (the world’s longest-standing
cryptocurrency exchange) are used and collected
through the website: http://api.bitcoincharts.com/
v1/csv/. In this case, for each common business day,
the day’s closing price is exploited (according to Aslan
and Sensoy (2020)). The MSCI Emerging Markets
Index prices are sourced from the *Wall Street Journal*
[(https://www.wsj.com/market-data/quotes/index/](https://www.wsj.com/market-data/quotes/index/XX/891800/historical-prices)
[XX/891800/historical-prices).](https://www.wsj.com/market-data/quotes/index/XX/891800/historical-prices) Information about
economic policy uncertainty events for China, Russia,
and USA is downloaded from www.policyuncertainty.
com (except for China’s downturn in 2015).
Table 1 presents estimates of basic descriptive
statistics for Bitcoin and the MSCI Emerging Markets
Index. It can be noticed that Bitcoin reports a higher
maximum daily return of 48% and the largest decrease
of 66%, compared to the MSCI Emerging Markets
Index. Besides, both return series are left skewed and
leptokurtic. However, the left tail of the distribution
in Bitcoin returns is much longer (-1.0666) than for
the MSCI Emerging index (−0.5098). Results of the
ADF test imply that returns for Bitcoin and the MSCI
Emerging index are stationary. These findings suggest
that Bitcoin returns are more volatile, and their
distribution is more non-normal than in the case of
the emerging stock markets.
**Table 1.** Descriptive statistics for the logarithmic return
series of Bitcoin (BTC) and the MSCI Emerging Markets
Index from 13 September 2011 to 11 August 2022
**BTC** **MSCI Emerging Markets**
Mean 0.0029 1.78E-05
Median 0.0026 4.55E-04
Maximum 0.4848 5.58E-02
Minimum -0.6639 -6.94E-02
Std. Dev. 0.0558 0.0101
Skewness -1.0666 -0.5098
Kurtosis 23.3798 8.0321
ADF -11.764*** -14.166***
Observations 2835 2835
Note: *** means a 1% significance level. Source: Own
calculations
-----
CEEJ **• 10** (57) **•** 2023 **•** pp. 219-236 **•** ISSN 2543-6821 **•** DOI: 10.2478/ceej-2023-0013 **227**
#### **4. Results**
Figure 1 shows the time series of the Hurst exponents
for both studied markets over time. The red and blue
lines denote the Hurst exponents for Bitcoin and
MSCI Emerging Markets Index, respectively. As seen
in Figure 1, at most times there is a time-varying long
memory of Bitcoin and the MSCI Emerging Markets
Index, because both markets obtained the most values
of the Hurst exponent more than 0.5. In particular,
most of the largest deviations of the Hurst exponent
from 0.5 are reached for Bitcoin. Thus, generally, the
Bitcoin market seems to be more inefficient compared
to the emerging stock markets. A similar conclusion
can be drawn based on the results of another study
(Plastun et al., 2019) in the context of a comparison of
the market efficiency between two emerging markets.
Secondly, in the first half of the study period, the
time series of the Hurst exponents shows a decreasing
trend towards the value of 0.5 (an efficient market) for
both the MSCI Emerging Markets Index and Bitcoin.
This trend can be assigned to the announcements
of Bitcoin regulations, and recommendations of
supervisory authorities, which frequently occurred
in the years 2013–2018. In particular, it concerns
mainly the largest Bitcoin mining countries, which
are China and the USA (Statista, 2022). These
‘regulatory’ events might ensure better access (for
investors) to information that had been undefined
(uncertain) before it appeared. In effect, a decrease
in the Bitcoin market inefficiency can be observed,
which is consistent with Urquhart (2016), Bariviera
(2017), and Takaishi and Adachi (2020). During
this period, there is also an improvement in market
efficiency for the MSCI Emerging Markets Index. It
confirms the findings of studies on market efficiency
for some emerging stock markets (Cajueiro & Tabak,
2004; Sukpitak & Hengpunya, 2016; Hkiri et al., 2021).
In Figure 1, one can see that the behaviour of the
Hurst exponent at some subperiods seems to be very
close for the emerging markets and Bitcoin, especially
during the pandemic era from 2020–2022. To be more
precise, at the beginning of the pandemic period, a
meaningful increase in market inefficiency can be
observed for both studied markets. However, the initial
reaction of the Hurst exponent to the economic shock
(pandemic) is less for Bitcoin compared to the traditional
assets, which are emerging stock markets. These above
findings for the Bitcoin market are in line with Wang
et al. (2021), Assaf et al. (2022), Mensi et al. (2022), and
Chowdhury et al. (2023). In effect, it can be expected
**Figure 1.** Hurst exponents of daily returns for Bitcoin and
the MSCI Emerging Markets Index
Note: The date denotes the endpoints of the sliding
windows. The red and blue lines mean Hurst exponents for
Bitcoin and MSCI Emerging Markets Index, respectively.
The dashed line denotes an efficient market – the value of
the Hurst exponent is 0.5. Source: Own work
that the strength of the relationship between Bitcoin
and emerging stock markets may be time-varying and
related to some global economic shocks. Therefore,
a dynamic correlation between the dynamics of the
market efficiency of Bitcoin and the MSCI Emerging
Markets Index is presented in Figure 2.
Time-varying correlation (Figure 2) shows that
in the context of the dynamics of market efficiency,
the relationship between Bitcoin and MSCI Emerging
Markets is quite strong in some periods. The
maximum Spearman coefficients are 0.83 (Panel A)
and 0.81 (Panel B), which indicate a strong correlation.
In particular, it can be observed that the significant
(p-values less than 10%) and the largest correlations
occur mainly in several periods, e.g. mid-2014, at the
end of 2014, 2015, at the end of 2016, in early 2017,
from 2018– 2019, 2020, in early 2021, at the end of
2021, and in early 2022.
The above periods can be assigned to the events
of high economic policy uncertainty in the largest
Bitcoin mining countries. Specifically, China and
the USA were the largest Bitcoin mining countries
in the last few years (Statista, 2022). However, until
2015 most Bitcoin mining industries were located in
Europe and the USA (Tovanich, Soulié & Isenberg,
2021). Apart from that, China accounts for almost a
third of the MSCI Emerging Markets Index. Another
large emerging economy is Russia. In the study period,
there are several economic shocks concerning these
countries: the threat of a spillover of the euro area
crisis in 2012, at the end of 2013 the threat of the US
-----
CEEJ **• 10** (57) **•** 2023 **•** pp. 219-236 **•** ISSN 2543-6821 **•** DOI: 10.2478/ceej-2023-0013 **228**
**Figure 2.** Dynamic correlation between Bitcoin and MSCI Emerging Markets Index using R/S at different lengths of the
rolling window: 251 (Panel A), 126 (Panel B)
Note: Black and grey lines indicate correlation coefficients and p-values, respectively. The horizontal red line means
p-values at 10%. Rolling window sizes are 251 (Panel A) and 126 Hurst exponents (Panel B). The date corresponds to the
endpoints of the sliding windows for the correlation coefficient. Source: Own calculations
debt crisis, Russia’s aggression against Ukraine on
February 2014, China’s economic downturn in 2015,
the USA – China trade tensions in the years 2018–2019
(also related to the election of Donald Trump at the end
of 2016), Covid-19 in 2020 and the Russia – Ukraine
war in 2022.
In particular, the turbulence in the eurozone
may cause investors to reduce the share of emerging
economies in their investment portfolios. At the end
of 2013, there was the threat of a debt crisis in the
USA and China. China was America’s largest foreign
creditor in 2013. At that time, the US and Europe also
had the largest share of Bitcoin mining. If Congress
had not passed an increase in the national debt limit
by October 17, foreign payments could be stopped.
As a result, there was a partial shutdown of the US
government for 16 days, because Congress could not
agree on a budget. Next, there was Russia’s aggression
against Ukraine on February 2014. Another shock was
related to uncertainty about the fact that Donald Trump
won the election in late 2016. During the presidential
campaign, he spoke about his future policy against the
existing trade agreements with China. As a result, in
January 2018, the USA set tariffs on China. This trade
conflict intensified through 2019. The increase in
the correlation value in the years 2020/2021 could be
linked to the appearance of uncertainty related to the
Covid-19 pandemic. In 2022, the threat of the Russia
Ukraine war could have affected investors in terms of
increasing fear and herd behaviour.
The common feature of the above-mentioned
economic shocks is their unpredictability. Because of
a lack of certain information about these events (e.g.
the threat of a debt crisis, the Covid-19 vaccine, trade
policy between the USA and China, and war), they
could not be included in market prices by the rational
expectations of investors. Furthermore, investors could
over- or underestimate the importance of these shocks
for the economy due to the presence of a high level of
fear. Thus, the irrationality of Bitcoin investors could
arise from an increase in economic policy uncertainty.
Besides, the specific features of cryptocurrencies also
may affect investors’ behaviour. The computing power
which concentrates on this ‘cryptocurrency system’ is
not related to one geographic territory. So, to estimate
the distributed policy uncertainty of Bitcoin, investors
may use heuristics based on the information from its
largest mining countries.
Thus, it seems that events related to the high
economic policy uncertainty in the largest emerging
economies (e.g. China) have a meaningful impact on
the comovement in dynamics of market efficiency of
Bitcoin and the emerging stock markets. Specifically,
in times of the highest economic policy uncertainty,
the correlation value seems to strengthen. Despite
-----
CEEJ **• 10** (57) **•** 2023 **•** pp. 219-236 **•** ISSN 2543-6821 **•** DOI: 10.2478/ceej-2023-0013 **229**
the association between Bitcoin and emerging stock
markets efficiency is quite strong in some periods, a
sign of the correlation changes. This may be because
in the case of economic shock, in the short term,
behavioural factors may be the main determinants of
the efficiency of both markets. However, in the long
run, the market efficiency of emerging stocks may be
determined more by fundamentals, in contrast to the
cryptocurrency market efficiency which may still be
mainly the effect of behavioural factors. Furthermore,
the negative values of the correlation could signal the
feature of the safe haven in some periods for Bitcoin.
This is in line with Wang et al. (2021), Maganini,
Diniz and Rasheed (2021), and Mensi et al. (2022). The
results are robust in the context of adopting different
sizes of the rolling window. Generally, using 251
(Panel A) and 126 (Panel B) observations as a length
of the sliding window, the significant and largest
correlations are obtained mainly in similar periods.
However, the values of the correlation based on the
‘126’ data points are more volatile due to the shorter
time series used for its calculation.
#### **5. Additional analyses**
To analyse the relationship between Bitcoin and
emerging stock market efficiency more deeply,
additional tests were carried out. One of them is the
calculation of the correlations on the first differences of
the Hurst exponents. In this case, the Hurst exponents
are also based on the rescaled range analysis (R/S).
The results are presented in Figure 3.
Time-varying correlation (Figure 3) shows that
in the context of the dynamics of market efficiency,
the relationship between Bitcoin and MSCI Emerging
Markets in the pandemic period is one of the strongest
compared to the whole period. However, the values of
the association between Bitcoin and emerging markets
suggest a weak or lack of statistical correlation in the
analysed period. The strength of this association is
different from the correlation based on the values
(Figure 2). However, it could be expected, because the
transformation of values to the first differences may
result in the loss of some information.
In Figure 3, it can be observed that there are four
local maxima of the correlation. Therefore, three
main subperiods with different trends in the study
association can be distinguished: the years 2014–2017,
**Figure 3.** Dynamic correlation on the first differences of Hurst exponents between Bitcoin and MSCI Emerging Markets
Index
Note: Black and grey lines indicate correlation coefficients and p-values respectively. The horizontal red line means
p-values at 10%. Rolling window sizes are 251 (Panel A) and 126 Hurst exponents (Panel B). The date corresponds to the
endpoints of the sliding windows for the correlation coefficient. Source: Own calculation
-----
CEEJ **• 10** (57) **•** 2023 **•** pp. 219-236 **•** ISSN 2543-6821 **•** DOI: 10.2478/ceej-2023-0013 **230**
**Figure 4.** Dynamic Kendall correlation between Bitcoin and MSCI Emerging Markets Index using R/S at different lengths
of the rolling window: 251 (Panel A), 126 (Panel B)
Note: Black and grey lines indicate correlation coefficients and p-values, respectively. The horizontal red line means
p-values at 10%. Rolling window sizes are 251 (Panel A) and 126 Hurst exponents (Panel B). The date corresponds to the
endpoints of the sliding windows for the correlation coefficient. Source: Own work
late 2017–2020, and from the end of 2020. The
significant (p-values less than 10%) and the largest
correlations occur mainly in three smaller periods
– 2014–early 2015, in the years 2017–early 2018,
and 2020–early 2021. These periods are covered by
previous findings (Figure 2).
Figure 4 presents the dynamic correlation using
Kendall’s τ instead of the Spearman correlation.
Notice that the results are identical to those reached
by the Spearman correlation. In particular, the
dynamics of the correlation coefficient are similar to
that observed in the case of the Spearman correlation.
The subperiods with the largest values of the study
association are the same as before (Figure 2). However,
the maximum value of the correlation coefficient
(0.61) is smaller compared to that of 0.83, noted for
the Spearman method using 251 data points as a
length of the rolling window. This is also supported
by the results obtained using the sliding window of
126 observations.
Figure 5 reflects estimates of the Hurst exponent
using DFA (detrended fluctuation analysis). In Figure
5, some similarities can be seen in the dynamics of
market efficiency using both methods – rescaled
range analysis and detrended fluctuation analysis.
Specifically, the significant (p-values less than 10%)
and the largest correlations between Bitcoin and
MSCI Emerging Markets from a market efficiency
perspective can be observed mainly in similar
subperiods – end of 2014/early 2015, at the end of
2015, at the end of 2016/early 2017, 2018, at the end of
2018 and 2019, at the beginning of 2020, and the end
of 2020/the beginning of 2021, at the end of 2021, and
in early 2022. Except for the very rare cases (periods:
2016/2017 and 2018/2019), the signs or values of
the correlation are very similar for both methods.
However, DFA reached on average lower maximum
correlation coefficients compared to R/S while the
studied relationship may be considered moderate in
most periods of market stress for both methods. The
reason for this can be that only DFA (contrary to R/S)
uses a polynomial fit detrending in subperiods. It may
be more resistant to the non-stationarity of time series
compared to R/S (Kristoufek, 2010).
To show more precisely the periods in which
the correlations are the strongest, the Spearman
correlations based on Hurst exponents by both
DFA and R/S methods, and the First differences are
presented together in Figure 6.
In some periods Figure 6 presents similar dynamics
of the Spearman correlation coefficients based on the
Hurst exponents of both methods (DFA and R/S). In
-----
CEEJ **• 10** (57) **•** 2023 **•** pp. 219-236 **•** ISSN 2543-6821 **•** DOI: 10.2478/ceej-2023-0013 **231**
**Figure 5.** Dynamic correlation between Bitcoin and MSCI Emerging Markets Index using DFA at different lengths of the
rolling window: 251 (Panel A), 126 (Panel B)
Note: Black and grey lines indicate correlation coefficients and p-values, respectively. The horizontal red line means
p-values at 10%. Rolling window sizes are 251 (Panel A) and 126 Hurst exponents (Panel B). The date corresponds to the
endpoints of the sliding windows for the correlation coefficient. Source: Own calculations
**Figure 6.** Dynamic Spearman correlation between Bitcoin and MSCI Emerging Markets Index based on Hurst exponents
using R/S, DFA and First differences of Hurst exponents in the rolling window of 251 observations
Note: Blue and red lines indicate correlation coefficients based on Hurst exponents using R/S and DFA, respectively. The
green line means correlation coefficients based on the first differences of Hurst exponents using R/S. The grey colour
indicates the range of the correlation values (minimum, maximum) relative to the time point (x-axis). The correlation
coefficients located in the area between two horizontal black dashed lines are statistically insignificant (p-values less
than 10%). The date corresponds to the endpoints of the sliding windows for the correlation coefficient. Source: Own
calculation
-----
CEEJ **• 10** (57) **•** 2023 **•** pp. 219-236 **•** ISSN 2543-6821 **•** DOI: 10.2478/ceej-2023-0013 **232**
particular, the most similar dynamics and correlation
values can be observed during the pandemic. Thus, the
link between global economic shocks and the dynamics
of correlation seems to be the most confirmed. On the
other hand, the correlation values in some cases differ
in the sign of the correlation or its value (especially
in the years 2018–2019). However, dynamics of
both correlations are very similar in the period of
local economic shocks (e.g. from mid-2018 to 2019).
Therefore, it cannot be unequivocally stated whether
“local” economic shocks contribute to the largest
correlation values between Bitcoin and the emerging
stock markets’ efficiency. But, in general, the results
confirm that the dynamics of the correlation between
the emerging stock and Bitcoin markets efficiency
behave similarly during times of high uncertainty
related to the economic turmoils in the countries with
the largest Bitcoin mining or global economic shocks.
This is also supported by the separate phases of the
dynamic correlation trend for the first differences in
market efficiency. The correlation of changes in the
dynamics of market efficiency strengthens with the
accumulation of uncertainty related to economic shocks.
#### **6. Discussion**
Generally, the results show that in times of unexpected
events, the correlation between emerging stock and
Bitcoin markets efficiency is the strongest, although
the strength of this association can be considered
moderate. However, only at some subperiods is there
a negative sign of the correlation, which may indicate
Bitcoin’s potential to be a safe haven in the context
of market efficiency. This is in line with the view
presented by others (Wang et al., 2021; Maganini,
Diniz, & Rasheed, 2021; Mensi et al., 2022). Besides,
this paper confirms the findings of previous studies,
uncovering in separate research that the market
efficiency in the emerging markets and Bitcoin is
time-varying and characterised by the long-memory
process most of the time, for example as in Bariviera
(2017) and Hull and McGroarty (2014).
Furthermore, our results indicate that the market
efficiency dynamics of Bitcoin and emerging stock
markets are different. In particular, the findings
obtained by detrended fluctuation analysis show at
most a moderate association between the dynamics
of Bitcoin market efficiency and emerging markets
efficiency. This confirms the results reached by
Plastun et al. (2019) for Russia and Ukraine. Thus,
future studies should treat Bitcoin rather as a specific
investment instead of an emerging market from the
perspective of the dynamics of market efficiency. This
is contrary to the nomenclature presented by others
(Alvarez-Ramirez et al., 2018; Kumar & Zargar, 2019).
The findings suggest that the main events
related to economic policy uncertainty may affect
the dynamics of market efficiency of Bitcoin and
emerging stock markets. These economic shocks
mainly concern the largest Bitcoin mining countries
and their major trading partners, and global economic
threats such as Covid-19. Thus, investors should track
the economic policy uncertainty of the largest Bitcoin
mining ‘geographic territory’. Furthermore, different
reactions of market efficiency in these markets to
some economic shocks imply the potential to benefit
from a diversification strategy using Bitcoin and the
emerging markets in one investment portfolio during
economic turmoil. Thus, the result may have an impact
on the more efficient allocation of capital. Besides, our
findings indicate that regulators of the Bitcoin market
and the emerging markets should be cautious about
the impact of their economic policy transparency on
the reaction of these market investors.
The research has shed light on the dependence
between the dynamics of market efficiency of Bitcoin
and emerging markets in the context of high economic
policy uncertainty in major Bitcoin mining countries.
Future studies should deepen this issue. Furthermore,
because Bitcoin is more inefficient than emerging
markets most of the time, its dynamics may be more
dependent on behavioural factors (these factors, such
as investor emotions, make investment decisions more
difficult). This could be verified by future studies. Our
results suggest that the largest emerging countries
with a meaningful share in Bitcoin mining could
play an essential role in this market during economic
shocks. Therefore, a study of the dynamic relationship
between Bitcoin market efficiency and emerging stock
markets efficiency across countries is needed. Finally,
the results showing a negative correlation are only
true for some identified economic shocks, so it is still
uncertain whether Bitcoin can be treated as a safe
haven from a market efficiency perspective.
#### **7. Conclusion**
This paper attempts to compare the dynamics of
market efficiency between Bitcoin and the emerging
stock markets in the years 2011–2022. It clarifies
-----
CEEJ **• 10** (57) **•** 2023 **•** pp. 219-236 **•** ISSN 2543-6821 **•** DOI: 10.2478/ceej-2023-0013 **233**
whether Bitcoin can be treated as an emerging market
or whether its market efficiency is more resistant to
economic shocks than emerging markets’ efficiency.
To this end, the Hurst exponent is exploited as a
measure of market efficiency. The Hurst exponent is
calculated using the rescaled range analysis. Besides,
the sliding window is applied to show the dynamics
of market efficiency. Thirdly, the rolling window
correlation is utilised to show how the association
between studied variables varies over time.
The contribution of this article to the literature is
at least threefold. Firstly, it concerns drivers of Bitcoin
market efficiency. So far, there is a lack of knowledge
of whether the dynamics of Bitcoin market efficiency
are linked to economic policy uncertainty. Our results
provide new insights into this issue, suggesting that
future studies should focus on a dependence between
the dynamics of Bitcoin market efficiency and
economic policy uncertainty in the largest Bitcoin
mining countries. Secondly, this paper adds to the
literature by the verification of the ‘emerging’ nature
of the cryptocurrency market efficiency. The findings
report that the dynamics of Bitcoin and emerging stock
markets’ efficiency are different. Besides, this research
presents different relative reactions of the Bitcoin and
emerging stock markets’ efficiency to economic shock
(e.g. the pandemic).
This study has several limitations. Firstly, Bitcoin’s
weekend prices are excluded from our sample,
because the stock market exchanges are closed at the
weekend. Secondly, it is difficult to precisely identify
concrete shocks in economic policy uncertainty as an
interpretation of the values of correlation because the
length of the ‘overlapping’ sliding windows (for the
correlation and Hurst exponent calculation) covers
from 2.5 to 3 years. Another limitation is the utilisation
of the aggregate index of the largest emerging stock
markets. In this case, the impact of economic policy
uncertainty on a particular stock market may be
different. Probably, the shocks will be more apparent
in the smallest emerging markets, because of lower
policy stability compared to the largest economies.
#### **References**
Aggarwal, D. (2019). Do Bitcoins follow a random
walk model? *Research in Economics,* 73, 15–22. https://
doi.org/10.1016/j.rie.2019.01.002
Alvarez-Ramirez, J., Rodriguez, E., & IbarraValdez, C. (2018). Long-range correlations and
asymmetry in the Bitcoin market. *Physica A: Statistical*
*Mechanics and Its Applications*, 492, 948–955. https://
doi.org/10.1016/j.physa.2017.11.025
Al-Yahyaee, K. H., Mensi, W., & Yoon, S. (2018).
Efficiency, multifractality, and the long-memory
property of the Bitcoin market: A comparative analysis
with stock, currency, and gold markets. *Finance*
*Research Letters*, 27, 228–234. https://doi.org/10.1016/j.
frl.2018.03.017
Aslan, A., & Sensoy, A. (2020). Intraday
efficiency-frequency nexus in the cryptocurrency
markets. *Finance Research Letters*, 35(C). https://doi.
org/10.1016/j.frl.2019.09.013
Assaf, A., Bhandari, A., Charif, H., & Demir, E.
(2022). Multivariate long memory structure in the
cryptocurrency market: The impact of COVID-19.
*International Review of Financial Analysis*, 82(C). https://
doi.org/10.1016/j.irfa.2022.102132
Bariviera, A. F. (2017). The inefficiency of Bitcoin
revisited: A dynamic approach. *Economics Letters,* 161,
1–4. https://doi.org/10.1016/j.econlet.2017.09.013
Bariviera, A. F., Basgall, M. J., Hasperué, W.,
& Naiouf, M. (2017). Some stylized facts of the
Bitcoin market. *Physica A: Statistical Mechanics and its*
*Applications*, 484(C), 82–90. https://doi.org/10.1016/j.
physa.2017.04.159
Baur, D., & McDermott, T. (2010). Is gold a safe
haven? International evidence. *Journal of Banking &*
*Finance*, 34(8), 1886–1898. https://doi.org/10.1016/j.
jbankfin.2009.12.008
Borowski, K., & Matusewicz, M. (2019). The
day-of-the-week effect on the example of 82
cryptocurrencies. *Przedsiębiorstwo i Finanse*, *3* (26),
31–50. Retrieved from http://cejsh.icm.edu.pl/cejsh/
element/bwmeta1.element.ojs-issn-2084-1361-year2019-issue-3-article-df281222-63ea-385f-a265
72dc5dc83783
Borowski, K., & Matusewicz, M. (2020).
Calculating Hurst Exponent with the Use of the
Siroky Method in Developed and Emerging Markets.
*Finanse I Prawo Finansowe*, 3(27), 25–61. https://doi.
org/10.18778/2391-6478.3.27.02
Bouri, E., Molnár, P., Azzi, G., Roubaud, D.,
& Hagfors, L. I. (2017). On the hedge and safe
haven properties of Bitcoin: Is it really more than a
diversifier?. *Finance Research Letters*, 20(C), 192–198.
[https://doi.org/10.1016/j.frl.2016.09.025](https://doi.org/10.1016/j.frl.2016.09.025)
-----
CEEJ **• 10** (57) **•** 2023 **•** pp. 219-236 **•** ISSN 2543-6821 **•** DOI: 10.2478/ceej-2023-0013 **234**
Bouri, E., Gil-Alana, L. A., Gupta, R., & Roubaud,
D. (2019). Modelling long memory volatility in the
Bitcoin market: Evidence of persistence and structural
breaks. *International Journal of Finance & Economics*, 24,
412–26. https://doi.org/10.1002/ijfe.1670
Bouri, E., Shahzad, S. J. H., Roubaud, D.,
Kristoufek, L., & Lucey, B. (2020). Bitcoin, gold, and
commodities as safe havens for stocks: New insight
through wavelet analysis. *The Quarterly Review*
*of Economics and Finance*, 77, 156–164. https://doi.
org/10.1016/j.qref.2020.03.004
Brauneis, A., & Mestel, R. (2018). Price discovery
of cryptocurrencies: Bitcoin and beyond. *Economics*
*Letters*, 165, 58–61. https://doi.org/10.1016/j.
econlet.2018.02.001
Cajueiro, D., & Tabak, B. (2004). The Hurst
exponent over time: testing the assertion that emerging
markets are becoming more efficient. *Physica A:*
*Statistical Mechanics and its Applications*, 336(3), 521–537.
https://doi.org/10.1016/j.physa.2003.12.031
Caporale, G. M., Gil-Alana, L., & Plastun, A.
(2018). Persistence in the cryptocurrency market.
*Research in International Business and Finance*, 46(C),
141–148. https://doi.org/10.1016/j.ribaf.2018.01.002
Carrick, J. (2016). Bitcoin as a Complement to
Emerging Market Currencies. *Emerging Markets*
*Finance and Trade*, 52(10), 2321–2334. https://doi.org/
10.1080/1540496X.2016.1193002
Chowdhury, M. A. F., Abdullah, M., Alam, M.,
Abedin, M. Z., & Shi, B. (2023). NFTs, DeFi, and
other assets efficiency and volatility dynamics: An
asymmetric multifractality analysis. *International*
*Review of Financial Analysis*, 87(C). https://doi.
org/10.1016/j.irfa.2023.102642
Czarnecki, Ł., Grech, D., & Pamuła, G. (2008).
Comparison study of global and local approaches
describing critical phenomena on the Polish stock
exchange market. *Physica A-statistical Mechanics and Its*
*Applications*, 387, 6801–6811. https://doi.org/10.1016/j.
physa.2008.08.019
Czekaj, J., Woś, M., & Żarnowski, J. (2001).
*Efektywność giełdowego rynku akcji w Polsce. Z*
*perspektywy dziesięciolecia* . Warszawa: Wydawnictwo
Naukowe PWN
Diniz-Maganini, N., Diniz, E. H., & Rasheed,
A. A. (2021). Bitcoin’s price efficiency and safe
haven properties during the COVID-19 pandemic:
A comparison. *Research in International Business*
*and Finance*, 58, 101472. https://doi.org/10.1016/j.
ribaf.2021.101472
Fama, E. F. (1970). Efficient Capital Markets:
A Review of Theory and Empirical Work. *The*
*Journal of Finance*, 25(2), 383–417. https://doi.
org/10.2307/2325486
Fernandes, L. H. S., Bouri, E., Silva, J. W.L., Bejan
L., & de Araujo, F. H. A. (2022). The resilience of
cryptocurrency market efficiency to COVID-19 shock.
*Physica A: Statistical Mechanics and its Applications*, 607.
https://doi.org/10.1016/j.physa.2022.128218
Hileman, G., & Rauchs, M. (2017). Global
cryptocurrency benchmarking study. *Cambridge Centre*
*for Alternative Finance* . Retrieved from https://www.
jbs.cam.ac.uk/wp-content/uploads/2020/08/2017-0420-global-cryptocurrency-benchmarking-study.pdf
Hkiri, B., Bejaoui, A., Gharib, C., & Al Nemer H. A.
(2021). Revisiting efficiency in MENA stock markets
during political shocks: evidence from a multi-step
approach. *Heliyon*, 7(9). https://doi.org/10.1016/j.
heliyon.2021.e08028
Hull, M., & McGroarty, F. (2014). Do emerging
markets become more efficient as they develop?
Long memory persistence in equity indices. *Emerging*
*Markets Review*, 18(C), 45–61. https://doi.org/10.1016/j.
ememar.2013.11.001
Jiang, Y., Nie, H., & Ruan, W. (2018). Time-Varying
Long-Term Memory in Bitcoin Market. *Finance*
*Research Letters*, 25, 280–284. https://doi.org/10.1016/j.
frl.2017.12.009
Kahneman, D., & Tversky, A. (1979).
Prospect Theory: An Analysis of Decision under
Risk. *Econometrica*, 47(2), 263–291. https://doi.
org/10.2307/1914185
Khuntia, S., & Pattanayak, J. (2018). Adaptive
market hypothesis and evolving predictability of
Bitcoin. *Economics Letters*, 167, 26–28. https://doi.
org/10.1016/j.econlet.2018.03.005
Khuntia, S., & Pattanayak, J. (2020). Adaptive Long
Memory in Volatility of Intra-day Bitcoin Returns and
the Impact of Trading Volume. *Finance Research Letters*,
32, 101077. https://doi.org/10.1016/j.frl.2018.12.025
Kosc, K., Sakowski, P., & Ślepaczuk, R.
(2019). Momentum and contrarian effects on the
cryptocurrency market. *Physica A: Statistical Mechanics*
*and its Applications* [, 523, 691–701. https://doi.](https://doi.org/10.1016/j.physa.2019.02.057)
[org/10.1016/j.physa.2019.02.057](https://doi.org/10.1016/j.physa.2019.02.057)
-----
CEEJ **• 10** (57) **•** 2023 **•** pp. 219-236 **•** ISSN 2543-6821 **•** DOI: 10.2478/ceej-2023-0013 **235**
Kristoufek, L. (2010). Rescaled Range Analysis
and Detrended Fluctuation Analysis: Finite
Sample Properties and Confidence Intervals.
*Czech Economic Review*, Charles University Prague,
Faculty of Social Sciences, Institute of Economic
Studies, 4(3), 315–329. Retrieved from https://
www.researchgate.net/profile/LadislavKristoufek/
publication/227360892_Rescaled_Range_Analysis_
and_Detrended_Fluctuation_Analysis_Finite_
Sample_Properties_and_Confidence_Intervals/
links/0fcfd50ddb6e3bdcf5000000/Rescaled-RangeAnalysis-and-Detrended-Fluctuation-AnalysisFinite-Sample-Properties-and-Confidence-Intervals.
pdf
Kristoufek, L. (2018). On the Bitcoin market
inefficiency and its Evolution. *Physica A: Statistical*
*Mechanics and its Applications*, 503, 257–262. https://doi.
org/10.1016/j.physa.2018.02.161
Kumar, D., & Zargar, F. N. (2019). Informational
inefficiency of Bitcoin: A study based on highfrequency data. *Research in International Business*
*and Finance*, 47, 344–353. https://doi.org/10.1016/j.
ribaf.2018.08.008
Köchling, G., Müller, J., & Posch, P. N. (2019).
Price delay and market frictions in cryptocurrency
markets. *Economics Letters*, 174(C), 39–41. https://doi.
org/10.1016/j.econlet.2018.10.025
Lim, K. P., Brooks, R. D., & Kim, J. H. (2008).
Financial crisis and stock market efficiency: Empirical
evidence from Asian countries. *International Review*
*of Financial Analysis*, 17(3), 571–591. https://doi.
org/10.1016/j.irfa.2007.03.001
Lo, A. W. (2004). The Adaptive Markets
Hypothesis: Market Efficiency from an Evolutionary
Perspective. *The Journal of Portfolio Management*, 30(5),
15–29. Retrieved from https://www.researchgate.
net/publication/228183756_The_Adaptive_
Markets_Hypothesis_Market_Efficiency_from_an_
Evolutionary_Perspective
Mensi, W., Sensoy, A., Vo, X. V., & Kang,
S. H. (2022). Pricing efficiency and asymmetric
multifractality of major asset classes before and
during COVID-19 crisis. *The North American Journal of*
*Economics and Finance*, 62(C). https://doi.org/10.1016/j.
najef.2022.101773
Mizerka, J., Stróżyńska-Szajek, A., & Mizerka, P.
(2020). The role of Bitcoin on developed and emerging
markets—on the basis of a Bitcoin users graph analysis.
*Finance Research Letters*, 35. https://doi.org/10.1016/j.
frl.2020.101489
Mnif, E., Mouakhar, K., & Jarboui, A. (2023).
Energy-conserving cryptocurrency response during
the COVID-19 pandemic and amid the Russia–Ukraine
conflict. *Journal of Risk Finance*, 24(2), 169–185. https://
doi.org/10.1108/JRF-06-2022-0161
Noda, A. (2021). On the evolution of
cryptocurrency market efficiency. *Applied Economics*
*Letters*, *28* (6), 433–439. https://doi.org/10.1080/13504
851.2020.1758617
Phiri, A. (2022). Can wavelets produce a clearer
picture of weak-form market efficiency in Bitcoin?
*Eurasian Economic Review.* *Eurasia Business and Economics*
*Society*, 12(3), 373–386. https://doi.org/10.1007/
s40822-022-00214-8
Plastun, A., Kozmenko, S., Plastun V., & Filatova,
H. (2019). Market anomalies and data persistence:
The case of the day-of-the-week effect. *Journal of*
*International Studies*, 12(3), 122–130. https://doi.
org/10.14254/2071-8330.2019/12-3/10
Polanco-Martínez, J. M. (2019). Dynamic
relationship analysis between NAFTA stock markets
using nonlinear, nonparametric, non-stationary
methods. *Nonlinear Dynamics,* 97, 369–389. https://doi.
org/10.1007/s11071-019-04974-y
Polasik, M., Piotrowska, A. I., Wisniewski,
T. P., Kotkowski, R., & Lightfoot, G. (2015). Price
Fluctuations and the Use of Bitcoin: An Empirical
Inquiry. *International Journal of Electronic Commerce*,
20(1), 9–49. https://doi.org/10.1080/10864415.2016.1
061413
Rufino, C. C. (2023). On the Volatility and
Market Inefficiency of Bitcoin During the COVID-19
Pandemic. *DLSU Business & Economics Review*, 32(2),
23–32. Retrieved from https://www.dlsu.edu.ph/
wp-content/uploads/2023/04/2rufino-040323.pdf
Shahzad, S. J. H., Bouri, E., Roubaud, D.,
Kristoufek, L., & Lucey, B. (2019). Is Bitcoin a better
safe-haven investment than gold and commodities?
*International Review of Financial Analysis*, 63, 322–330.
https://doi.org/10.1016/j.irfa.2019.01.002
Statista. (2022, January 12). *Bitcoin mining by*
*country* . Retrieved from https://www.statista.com/
statistics/1200477/bitcoin-mining-by-country/
Sukpitak, J., & Hengpunya, V. (2016). Efficiency of
Thai stock markets: Detrended fluctuation analysis.
*Physica A: Statistical Mechanics and its Applications*, 458(C),
[204–209. https://doi.org/10.1016/j.physa.2016.03.076](https://doi.org/10.1016/j.physa.2016.03.076)
-----
CEEJ **• 10** (57) **•** 2023 **•** pp. 219-236 **•** ISSN 2543-6821 **•** DOI: 10.2478/ceej-2023-0013 **236**
Takaishi, T., & Adachi, T. (2020). Market
Efficiency, Liquidity, and Multifractality of Bitcoin:
A Dynamic Study. *Asia-Pacific Financial Markets*, 27,
145–154. https://doi.org/10.1007/s10690-019-09286-0
Tovanich, N., Soulié, N., & Isenberg, P. (2021,
April). *Visual analytics of bitcoin mining pool evolution:*
*on the road toward stability?* 3rd International
Workshop on Blockchains and Smart Contracts
held in conjunction with the 11th IFIP International
Conference on New Technologies, Mobility and
Security, France, Paris, 1–5. https://doi.org/10.1109/
NTMS49979.2021.9432675
Tran, V. L., & Leirvik, T. (2020). Efficiency in the
markets of crypto-currencies. *Finance Research Letters*,
35. https://doi.org/10.1016/j.frl.2019.101382
Urquhart, A. (2016). The inefficiency of Bitcoin.
*Economics Letters*, 148, 80–82. https://doi.org/10.1016/j.
econlet.2016.09.019
Wang, J., & Wang, X. (2021). COVID-19 and
financial market efficiency: Evidence from an
entropy-based analysis. *Finance Research Letters*, 42(C).
https://doi.org/10.1016/j.frl.2020.101888
Wei, W. C. (2018). Liquidity and market efficiency
in cryptocurrencies. *Economics Letters*, 168, 21–24.
[https://doi.org/10.1016/j.econlet.2018.04.003](https://doi.org/10.1016/j.econlet.2018.04.003)
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.2478/ceej-2023-0013?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.2478/ceej-2023-0013, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "HYBRID",
"url": "https://sciendo.com/pdf/10.2478/ceej-2023-0013"
}
| 2,023
|
[
"JournalArticle"
] | true
| 2023-01-01T00:00:00
|
[
{
"paperId": "b016c9aa7731ee26cf581cfd6aaa1ac6e9a42f30",
"title": "NFTs, DeFi, and other assets efficiency and volatility dynamics: An asymmetric Multifractality analysis"
},
{
"paperId": "3f789cecc80f6d377bf8c849602ea45bb98ef3f3",
"title": "The resilience of cryptocurrency market efficiency to COVID-19 shock"
},
{
"paperId": "f6053e9032dafa604440b955e338d046decb2d85",
"title": "Can wavelets produce a clearer picture of weak-form market efficiency in Bitcoin?"
},
{
"paperId": "9895af7968af5a01dd2f02750a4bb27495483856",
"title": "Pricing efficiency and asymmetric multifractality of major asset classes before and during COVID-19 crisis"
},
{
"paperId": "9e87f3cd6b92cff97ee9b118902f230567e62308",
"title": "Multivariate long memory structure in the cryptocurrency markets: The impact of COVID-19"
},
{
"paperId": "b9090fafd634482b7d9b78d922ce6ad3881a6887",
"title": "Revisiting efficiency in MENA stock markets during political shocks: evidence from a multi-step approach"
},
{
"paperId": "63f68191f7b0249915c047dac6992f6bcff9cea8",
"title": "Bitcoin’s price efficiency and safe haven properties during the COVID-19 pandemic: A comparison"
},
{
"paperId": "c3ea681c2d760f8dfa46b2602e9081cdbb53f5e0",
"title": "COVID-19 and financial market efficiency: Evidence from an entropy-based analysis"
},
{
"paperId": "416f1b4a9615ae04b74a34327a867236049d4b16",
"title": "Calculating Hurst Exponent with the Use of the Siroky Method in Developed and Emerging Markets"
},
{
"paperId": "ebb567a45ecff5f4f13de1197a157cee840a2c7a",
"title": "Bitcoin, gold, and commodities as safe havens for stocks: New insight through wavelet analysis"
},
{
"paperId": "e549e886ab9c666acd3f0a6f996307168268d8de",
"title": "The role of Bitcoin on developed and emerging markets – on the basis of a Bitcoin users graph analysis"
},
{
"paperId": "22379d0695ab678330424360dea3711ea3a799d1",
"title": "Intraday efficiency-frequency nexus in the cryptocurrency markets"
},
{
"paperId": "003451d154fa358df6809d1c0e228ffbf3212476",
"title": "Efficiency in the Markets of Crypto-Currencies"
},
{
"paperId": "d18e968c64d66e5908d32dc1b9ea723d9275c418",
"title": "Market anomalies and data persistence: The case of the day-of-the-week effect"
},
{
"paperId": "58d9ebc0b7bffca6d28d78cba0007b64f7b28cbd",
"title": "Dynamic relationship analysis between NAFTA stock markets using nonlinear, nonparametric, non-stationary methods"
},
{
"paperId": "9a149d283607e973cbc375a50e33528bcf42df42",
"title": "Momentum and contrarian effects on the cryptocurrency market"
},
{
"paperId": "421f93e54b5fcf7e6c92966a7c2e82ab8d1fa4b0",
"title": "Is Bitcoin a better safe-haven investment than gold and commodities?"
},
{
"paperId": "33d746db9f1ef2e80fa8a7c828fa09d75bb87a93",
"title": "On the evolution of cryptocurrency market efficiency"
},
{
"paperId": "2ec77e703c24221bb74bf8317cd1214837658f15",
"title": "Do bitcoins follow a random walk model?"
},
{
"paperId": "6c5efe78379aafa0843c3f28ba878189a91c5cfa",
"title": "Market Efficiency, Liquidity, and Multifractality of Bitcoin: A Dynamic Study"
},
{
"paperId": "0c1a4fc61eb82a94a9c1568f7995d837355638f8",
"title": "Informational inefficiency of Bitcoin: A study based on high-frequency data"
},
{
"paperId": "217895bb50976672fd9004ee9e3b22ce6f0e1e31",
"title": "Efficiency, multifractality, and the long-memory property of the Bitcoin market: A comparative analysis with stock, currency, and gold markets"
},
{
"paperId": "9fc88fcfae6b3187c97cb0c332f5093500423245",
"title": "Price Delay and Market Frictions in Cryptocurrency Markets"
},
{
"paperId": "6ea1864c4299ddee3189c2bdfeb924b283603402",
"title": "Modelling long memory volatility in the Bitcoin market: Evidence of persistence and structural breaks"
},
{
"paperId": "b61d0259f2c02c945f2946679f0347e36285fc70",
"title": "On Bitcoin markets (in)efficiency and its evolution"
},
{
"paperId": "408687691e3fc29950250da71d4be91697471704",
"title": "Liquidity and market efficiency in cryptocurrencies"
},
{
"paperId": "0a7dc72b8e07e2622b47ecfedd9693ccfd0ba3dc",
"title": "Adaptive market hypothesis and evolving predictability of bitcoin"
},
{
"paperId": "1d0ff715d8d15d48b7c577fef1bb6fd496b5667a",
"title": "Price discovery of cryptocurrencies: Bitcoin and beyond"
},
{
"paperId": "747785ad0cf34f4f8abe99438b47474afa8a49f2",
"title": "Long-range correlations and asymmetry in the Bitcoin market"
},
{
"paperId": "569fd1797fe6e29ac0876d1e4d4e08ef8bf9b0c9",
"title": "Time-varying long-term memory in Bitcoin market"
},
{
"paperId": "3dae2f3c7ca5e9c60aa41df4e6dfba35f9586d83",
"title": "Persistence in the Cryptocurrency Market"
},
{
"paperId": "719b219f9b5e2de92b2014b11289fd9d388fb046",
"title": "The Inefficiency of Bitcoin Revisited: A Dynamic Approach"
},
{
"paperId": "572271b7f4d35459aaa4d8d84c3b26c9f2380765",
"title": "Some Stylized Facts of the Bitcoin Market"
},
{
"paperId": "26d69459477acc088945ceeb331e6fb6bd8af07a",
"title": "2017 Global Cryptocurrency Benchmarking Study"
},
{
"paperId": "becc76a130cd37d9441ff38ff9dc60c5218bff38",
"title": "On the Hedge and Safe Haven Properties of Bitcoin: Is it Really More than a Diversifier?"
},
{
"paperId": "f69c079991496fdfcda933ba9744eacc6007b926",
"title": "Efficiency of Thai stock markets: Detrended fluctuation analysis"
},
{
"paperId": "99d1e84f089518c9f74a9b4487a95caa794fde9c",
"title": "The Inefficiency of Bitcoin"
},
{
"paperId": "685a0a1a54cb732c466cbae58ea174211297ac04",
"title": "Price Fluctuations and the Use of Bitcoin: An Empirical Inquiry"
},
{
"paperId": "d561eb0a7d8858139c0ed390b35493d929bd946a",
"title": "Do emerging markets become more efficient as they develop? Long memory persistence in equity indices"
},
{
"paperId": "2c33761ce13d10605ce507ec30861e770c60f840",
"title": "Comparison study of global and local approaches describing critical phenomena on the Polish stock exchange market"
},
{
"paperId": "b376a646c744af2f340a3ee809e54987151c041a",
"title": "Financial crisis and stock market efficiency: Empirical evidence from Asian countries"
},
{
"paperId": "eebc4f8e7b5437bd12cf80bb28be2c6a23247b5b",
"title": "The Hurst exponent over time: testing the assertion that emerging markets are becoming more efficient"
},
{
"paperId": "26b8dfda9be1da51ffe0830381773bb6e68455ad",
"title": "Efficient Capital Markets: A Review of Theory and Empirical Work"
},
{
"paperId": "d38668a1eba0be85c75f9ea0c3cd1c892bf3c422",
"title": "Adaptive long memory in volatility of intra-day bitcoin returns and the impact of trading volume"
},
{
"paperId": null,
"title": "Ambiguity Aversion and Comparative Ignorance"
},
{
"paperId": "51c36fc0431bd2c99b6e36792a8da8e58a2e0137",
"title": "Institute for International Integration Studies Is Gold a Safe Haven? International Evidence Is Gold a Safe Haven? International Evidence Is Gold a Safe Haven? International Evidence"
}
] | 18,678
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Physics",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Physics",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/0223db3a9042664f4529dc235ad9f7a662dc15af
|
[
"Computer Science",
"Physics"
] | 0.898775
|
Discussion of Quantum Consensus Algorithms
|
0223db3a9042664f4529dc235ad9f7a662dc15af
|
arXiv.org
|
[
{
"authorId": "2108025805",
"name": "Lifu Zhang"
},
{
"authorId": "2051458279",
"name": "Samuel Fulton"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"ArXiv"
],
"alternate_urls": null,
"id": "1901e811-ee72-4b20-8f7e-de08cd395a10",
"issn": "2331-8422",
"name": "arXiv.org",
"type": null,
"url": "https://arxiv.org"
}
|
Leader election is a crucial process in many areas such as cloud computing, distributed systems, task orchestration, and blockchain. Oftentimes, in a distributed system, the network needs to choose a leader, which would be responsible for synchronization between different processors, data storage, information distribution, and more.
|
Boston University Department of Physics
## Discussion of Qantum Consensus Algorithms
#### Samuel Fulton, Lifu Zhang
**Abstract Leader election is a crucial process in many areas such as cloud computing, distributed systems, task orchestra-**
tion, and blockchain. Ofentimes, in a distributed system, the network needs to choose a leader, which would be responsible
for synchronization between diferent processors, data storage, information distribution, and more. In the case where the
network is anonymous, no classical algorithms could solve the problem exactly. However, in the seting of quantum computers, this problem is readily solved. In this paper, we analyze the quantum consensus algorithm developed by Seiichiro Tani.
We look at the inner workings of the algorithm and develop a circuit representation of the key steps. We review Mochon’s
fault tolerant leader election algorithm. We then implement a simple leader election algorithm on a quantum computer.
### Introduction
In this paper, we will introduce distributed systems and consensus algorithms. We then zoom in
and look at a specific class of distributed systems
known as anonymous or symmetric distributed
systems. From there, we introduce the classical approach to breaking symmetry and electing
a leader. We show that there is no classical deterministic algorithm for leader election among an
anonymous distributed system. We then introduce
two quantum algorithms developed by Seiichiro
Tani in [Tani et al. (2012)] for anonymous leader
election. The quantum algorithms for anonymous leader election are more than computational
speedups; they are deterministic solutions to a classically non-deterministic problem. We implement
the second quantum leader election algorithm in
Qiskit. Lastly, we analyze work by Mochon and Kitaev, [Mochon (2007)], on developing fault-tolerant
leader election algorithms.
### Distributed Systems
We will start with a brief definition of consensus
and distributed systems. We define a distributed
system to be a collection of communicating processors all working towards a common goal. Let’s consider a toy example. Imagine a distributed system
of computers atempting to factor a large number
_𝑁_ . Once one processor “believes” that it has factored the number, it proposes its prime factors, 𝑃𝑖,
of N. The other processors in the distributed system check if [�]𝑖 _[𝑃]𝑖_ [=][ 𝑁] [. If the majority of the pro-]
cessors agree that [�]𝑖 _[𝑃]𝑖_ [=][ 𝑁] [then][ 𝑃]𝑖 [is accepted]
as the factors of 𝑁 . All of the processors can now
move on to factoring a diferent prime number 𝑀,
all the while remembering that the prime factors
of 𝑁 have been decided. Other examples where
consensus algorithms come into play include leader
election, blockchain, load balancing, clock synchronization, and more.
Three things characterize a distributed system:
Agreement, Validity, and Termination. As the name
suggests agreement means that all non-faulty processors must agree on the same value. In the case
of factoring 𝑁, all non-faulty processors must either agree that 𝑃𝑖 are the factors of N or all agree
that 𝑃𝑖 are not the factors of 𝑁 . Validity is the assertion that under non-Byzantine conditions, the distributed system will never return an incorrect result. In the example of the processors atempting
the factor 𝑁, this would mean that given suficient
time all non-faulty processors would find the same
prime factors of 𝑁 . Termination is the assertion
that given enough time the processors are guaranteed to complete the task.
What do we mean when by non-faulty processors? There are two types of faulty processors.
The first is a processor that experienced a crash.
This processor stops responding to other processors in the distributed system. This is a common
occurrence. However, since distributed systems
only need the majority of processors to work, crash
failures are readily dealt with. The second type
of faulty processor experiences Byzantine failure.
Byzantine failure occurs when a processor malfunctions in a way such that it sends incorrect data
1
-----
Boston University Department of Physics
to the distributed system. An example of a Byzantine processor is a hacked processor. For consensus problems having Byzantine failures is the worst
scenario. Fault-tolerant consensus algorithms address crash and Byzantine failure. The most widely
used consensus algorithms in distributed and cloud
computing systems are Paxos and its variants such
as Raf. These algorithms are used for leader election and typically tolerate non-Byzantine failures.
### Qantum Consensus
Mazzarella categorizes quantum consensus into
four classes in [Mazzarella et al. (2015)]. _𝜎-_
expectation consensus, reduced state consensus, symmetric state consensus, and single 𝜎measurement consensus. Consider a quantum network consist of three qubits, and three observables
of the form.
_𝜎_ [1] = 𝜎[𝑧] ⊗ _𝐼_ ⊗ _𝐼,_
_𝜎_ [2] = 𝐼 ⊗ _𝜎[𝑧]_ ⊗ _𝐼,_
_𝜎_ [3] = 𝐼 ⊗ _𝐼_ ⊗ _𝜎[𝑧]_ _._
The system is in consensus concerning the expectation of 𝜎𝑧 if
_𝑇𝑟_ (𝜌𝜎 [1]) = 𝑇𝑟 (𝜌𝜎 [2]) = 𝑇𝑟 (𝜌𝜎 [3]).
Noted that Qantum consensus is achieved by a
quantum network rather than traditional computational resources. They are similar counterparts,
though we must take account of probabilistic outcomes due to the stochastic nature of quantum
mechanics. We will show how quantum entanglements can ofer an advantage in terms of reaching
an agreement in a distributed seting.
### Leader Election
As described in [Bro], ”leader election is the simple
idea of giving one thing (a process, host, thread, object, or human) in a distributed system some special powers. Those special powers could include
the ability to assign work, the ability to modify a
piece of data, or even the responsibility of handling
all requests in the system.” Leader election is extremely useful for improving eficiency. A leader
can ofen bypass consensus algorithms and simply inform the system about changes that will be
made. Leaders can help with consistency because
they can see all of the changes that have been made
to the system. By acting as a central data cache,
a leader can improve consistency across the entire
system.
Figure 1:
A single leader does introduce some drawbacks.
Namely, a single leader is a critical point for failure. If the leader crashes the entire distributed system may halt. Furthermore, if a single leader experiences Byzantine failure, the entire system may
waste time following incorrect protocols. However,
many of these drawbacks are mitigated through
the use of consensus algorithms. Ofentimes, the
improved eficiency of leader election out ways any
drawbacks. In the next section, we will explore how
leaders can be fairly elected.
### Leader Election Algorithm I
One interesting consensus problem is the anonymous leader election. Anonymous leader election
is used in the case where we have a collection of
identical processors and wish to designate a leader.
Figure 1 depicts an anonymous distributed system.
In the system, processors do not have unique identification and all run the same protocol. Thus, the
system is symmetric under all permutations. The
symmetry of the system prevents non-probabilistic
leader election. The classical approach outlined by
Seiichiro Tani in [Tani et al. (2005)] is to install a
coin flip in each processor. Each processor flips
a coin if heads it is eligible for leader election. If
the coin is tails, it is a follower. If multiple processors get heads, then the protocol is repeated with
the eligible candidates. Note that this process is
non-deterministic and has an expected run time of
_𝑂_ (log(𝑛)), where 𝑛 is the number of processors.
Let’s compare this to a simplistic quantum leader
election algorithm, which was proposed in [Tani
et al. (2012)]. We will start with a two-processor
scenario. For this algorithm, we have two processors 𝐴 and 𝐵. We generate the quantum state
_𝑊_ = √[1] (|01⟩+ |10⟩).
2
2
-----
Boston University Department of Physics
#### |R1⟩ • • X • • X |R1⟩
|R2⟩ • • X • • X |R2⟩
|0⟩1 |S1⟩ |0⟩2 |Sn⟩
Figure 2:
We send the first qubit to processor 𝐴 and the second qubit to processor 𝐵. If processor 𝐴 measures
|1⟩, we know processor 𝐵 must measure 0 and vice
versa. Whichever processor measures |1⟩ becomes
the leader. For 𝑛 processors we generate the state
_𝑊𝑛_ = √[1] (|10...0⟩+ ... + |0...01⟩),
_𝑛_
or
(|01⟩+ |10⟩) |00⟩ _._
If the system collapses to the (|01⟩+|10⟩) |00⟩ state,
processor 𝐴 measures its |𝑅1⟩ state and processor
_𝐵_ measures |𝑅2⟩ state. Whichever processor measures |1⟩ is the leader. If the system collapses into
the state (|00⟩+ |11⟩) |11⟩, both processors apply
the unitary operation
|• • X • • X | • • X • • X | S | S ||Col2|Col3|Col4|X|Col6|
|---|---|---|---|---|---|
|•|•||•|||
|•|•||•|||
|||||||
|||||||
_𝑈_ = [1]
√
2
�
1 −𝑖
−𝑖 1
�
= [1]
√
_𝑛_
∑︁𝑛−1
_𝑘=0_
2𝑘 � _._
���
Each processor receives a qubit. Whichever processor measures |1⟩ is the leader. This algorithm terminates afer one run, which is in stark contrast to
the non-deterministic classical leader election.
### Algorithm II
There is a more robust approach to quantum leader
election proposed in [Tani et al. (2012)]. Consider
the case with processors 𝐴 and 𝐵. Each processor
prepares the state
|𝑅1⟩ = |𝑅2⟩ = √[1] (|0⟩+ |1⟩).
2
We send |𝑅1⟩ and |𝑅2⟩ through the following circuit,
where _𝑋_ is the Pauli matrix _𝑋_ corresponding to a bit
flip, and the remaining two gates are control not
gates.
The output of the circuit is the state
|𝑅1 𝑅2 𝑆1 𝑆2⟩ = (|00⟩+ |11⟩) |11⟩
+ (|01⟩+ |10⟩) |00⟩ _._
In words, 𝑆1 and 𝑆2 and both |1⟩ if |𝑅1⟩ and |𝑅2⟩ are
equal. Processor 𝐴 has the state |𝑅1 𝑆1⟩ and processor 𝐵 has the |𝑅2 𝑆2⟩ qubit. We note that |𝑆1⟩ and
|𝑆2⟩ are entangled. Afer each processor measures
its 𝑆 state, the system will collapse to either
00 11 11
(| ⟩+ | ⟩) | ⟩
to their 𝑆 states.
_𝑈𝐴_ ⊗ _𝑈𝐵_ (|00⟩+ |11⟩)
= 𝑈𝐴 |0⟩⊗ _𝑈𝐵_ |0⟩+ 𝑈𝐴 |1⟩⊗ _𝑈𝐵_ |1⟩ _,_
= (|0⟩− _𝑖_ |1⟩) ⊗(|0⟩− _𝑖_ |1⟩)
+ (|0⟩− _𝑖_ |1⟩) ⊗(|0⟩− _𝑖_ |1⟩),
= |00⟩− _𝑖_ |01⟩+ 𝑖[2] |11⟩
+ 𝑖[2] |00⟩− _𝑖_ |01⟩− _𝑖_ |10⟩+ |11⟩ _,_
= −𝑖 (|01⟩+ |10⟩).
Just like in the first case, processor 𝐴 measures
its |𝑅1⟩ state, and processor 𝐵 measures |𝑅2⟩ state.
Whichever processor measures |1⟩ is the leader.
This algorithm is readily extended into the case
with 𝑛 nodes. Before we start we need one definition. We say a string 𝑥 = 𝑥1𝑥2...𝑥𝑛 is consistent if
all substrings 𝑥𝑖 are equal. Each processors starts
by generating the state
_𝑅𝑖_ = √[1] (|0⟩+ |1⟩).
2
The state of the system is
Each processor stores the consistency of the system in the qubit |𝑆𝑖 ⟩. The circuit for the process is
shown in Fig 3
The global system becomes
|𝑅1...𝑅𝑛𝑆1...𝑆𝑛⟩ = (��0⊗𝑛 � + ��1⊗𝑛 �) ��1⊗𝑛 �
+ 2∑︁[𝑛]−2 |𝑖⟩ ��0⊗𝑛 � _._
_𝑖=1_
Each processor now measures its |𝑆⟩ state. The system collapse to either
(��0⊗𝑛 � + ��1⊗𝑛 �) ��1⊗𝑛 �
or
2∑︁[𝑛]−2 |𝑖⟩ ��0⊗𝑛 � _._
_𝑖=1_
� 1
_𝑅𝑖_ = √
_𝑖_ 2[𝑛]
2∑︁[𝑛]−1
|𝑖⟩ _._
_𝑖=0_
3
-----
Boston University Department of Physics
_|R1⟩_ _•_ ... _•_ _X_ _•_ ... _•_ _X_ _|R1⟩_
_|R2⟩_ _•_ ... _•_ _X_ _•_ ... _•_ _X_ _|R2⟩_
...
_|Rn⟩_ _•_ ... _•_ _X_ _•_ ... _•_ _X_ _|Rn⟩_
_|0⟩1_ _|S1⟩_
...
_|0⟩n_ _|Sn⟩_
Figure 3:
If the system collapses to
2∑︁[𝑛]−2 |𝑖⟩ ��0⊗𝑛 �
_𝑖=1_
then the system 𝑅1…𝑅𝑛 is inconsistent. Since the
system is inconsistent at least one 𝑅𝑖 = |0⟩ and at
�
least one ��𝑅 _𝑗_ = |1⟩. Any processor that measures
its |𝑅𝑖 ⟩ = |1⟩ is a leader candidate. Any processor
that measures its |𝑅𝑖 ⟩ = |0⟩ is a follower. There
are now at most (𝑛 − 1) leader candidate, for which
the process is repeated. In the event that the system collapsed to the (��0⊗𝑛 � + ��1⊗𝑛 �) ��1⊗𝑛 � state, we
need to apply a unitary such that the symmetry is
broken. If the number of states 𝑛 is even then we
apply the unitary
_𝑈_ . Instead, for each processor we need an additional register |𝑇𝑖 ⟩ initialized to |0⟩. Set _𝑇𝑖_ = 𝑅𝑖 ⊕ _𝑇𝑖_ .
Then apply 𝑉𝑛 to 𝑅𝑖 ⊗ _𝑇𝑖_ . We define √𝑅1𝑛+1𝑉𝑛 as the
matrix
|• ... • X • ... • X | • ... • X • ... • X | . . . • ... • X • ... • X | |S . . . |S|Col2|X|Col4|
|---|---|---|---|
|•||||
|• . . .||||
|•||||
|. . .||||
|||||
�𝑛
�𝑛
1/√2 0 √𝑅𝑛 _𝑒[𝑖]_ _[𝜋]𝑛_ /√2
1/√2 0 −[√]𝑅𝑛𝑒[−][𝑖] _[𝜋]𝑛_ _𝑒[−][𝑖]_ _[𝜋]𝑛_ /√2
����� √𝑅𝑛 0 −𝑖𝑒√2[−]𝑅[𝑖𝜋𝐼𝑛]22𝑛𝑛 −[√]𝑅𝑛
√
� 0 _𝑅𝑛_ + 1 0 0
�����
�
_,_
where 𝑅𝑛 and 𝐼𝑛 are the real and imaginary
parts of 𝑒[𝑖] _[𝜋]𝑛_, respectively. This matrix is well de
fined since 0 < |𝑅𝑛 | < 1. With some calculations 𝑉𝑛 is shown to be unitary. Similar to the
case where 𝑛 is even, for symmetry to be preserved
the system must be in one of the following states
|00⟩ [⊗][𝑘] _, |01⟩_ [⊗][𝑘] _, |10⟩_ [⊗][𝑘] _, |11⟩_ [⊗][𝑘] . However, afer each
processor applies 𝑉𝑛, the probability of the system
being in any one of these states is
�
_,_
_𝑈_ = [1]
√
2
�
1 _𝑒[−][𝑖𝜋]_ [/][𝑛]
−𝑒[𝑖𝜋] [/][𝑛] 1
to each |𝑅𝑖 ⟩, which we will show breaks the symmetry. For symmetry to be preserved the system
|𝑅1...𝑅𝑛⟩ must be in the state |0⟩ [⊗][𝑛] or |1⟩ [⊗][𝑛] . Afer
each processor applies 𝑈 to |𝑅𝑖 ⟩, the probability of
being in either of these states is
Prob(|00⟩ [⊗][𝑛] ) = √[1] �� √ 1
2 2𝑅𝑛 + 2
� _𝑒[𝑖]_ _[𝜋]𝑛_ �𝑛�
+ √ _,_
2𝑅𝑛 + 2
= 0,
Prob(|01⟩ [⊗][𝑛] ) = √[1] �� √ 1
2 2𝑅𝑛 + 2
� −𝑒[𝑖] _[𝜋]𝑛_ �𝑛�
+ √ _,_
2𝑅𝑛 + 2
= 0,
�𝑛�
_,_
�𝑛�
_,_
Prob(|0⟩ [⊗][𝑛] ) = √[1]
2
�� 1
√
2
�𝑛 � _𝑒𝑖_ _[𝜋]𝑛_
+ √
2
and
= 0,
Prob(|1⟩ [⊗][𝑛] ) = √[1]
2
= 0.
�� 1
√
2
�𝑛 � −𝑒𝑖 _[𝜋]𝑛_
+ √
2
Thus, afer applying 𝑈 the probability of being in
a symmetric state is zero. Afer applying 𝑈, if a
processor measures its |𝑅𝑖 ⟩ to be |1⟩, it is a leader
candidate. Since the system is in an asymmetric
state, at least one processor will lose eligibility, and
at least one processor will remain eligible. If the
number of states 𝑛 is odd we cannot simply apply
,
Prob(|10⟩ [⊗][𝑛] ) = √[1] �� − √ 1 �𝑛
2 2𝑅𝑛 + 2
� 1 �𝑛�
+ √ _,_
2𝑅𝑛 + 2
= 0,
Prob(|11⟩ [⊗][𝑛] ) = 0.
Thus, the symmetry is broken. Each processor
now measures |𝑅𝑖 _𝑇𝑖_ ⟩, and the processors with the
largest value of |𝑅𝑖𝑇𝑖 ⟩ are candidates for the next
round. Again, since symmetry was broken, at least
one processor will lose eligibility, and at least one
processor will remain eligible.
4
-----
Boston University Department of Physics
### Qantum Consensus and Non- Bias Leader Election
Up to this point, we have been assuming no faulty
processors. Maor Ganz considers the case with a
group of 𝑛 processors who do not trust each other
and want to elect a leader. In his paper, [Ganz
(2017)], Ganz considers an algorithm that gives an
honest processor at least _𝑛[1]_ [−] _[𝜖]_ [probability of win-]
ning. Using classical consensus, this problem was
shown to be impossible by Mochon in [Mochon
(2007)]. However, using quantum consensus Mochon showed that in certain cases one can formulate an algorithm with arbitrarily small 𝜖.
This algorithm is based on a series of quantum
coin flips in tournament style. In other words, processors are paired and a single quantum coin flip is
used to eliminate a processor from each pair. The
main dificulty is in creating fault-tolerant coinflipping.
There are two types of bias coin flipping: strong
and weak coin-flipping. A strong coin-flipping protocol with bias _𝜖_ is a protocol in which neither party
is capable of forcing the probability of any given
flip to be greater than 1/2 + 𝜖. In weak coin flipping, both parties, Alice and Bob, have a predetermined desired coin outcome. For example, a 1
can be thought of as Alice winning and a 0 can be
thought of as Bob winning. In weak coin flipping,
neither player can shif the probability of the coin
flip towards their desired outcome with probability greater than 1/2 + 𝜖. In the classical scenario,
weak and strong coin flipping are essentially equivalent. However, as Mochon states ”in the quantum
world the two are very diferent,” [Mochon (2007)].
For this paper, we will only be concerned with weak
coin-flipping.
Machon’s algorithm, or rather his proof of the
existence of an algorithm for weak coin flipping
with arbitrarily small bias is a significant result in
quantum algorithms. However, as stated by Ganz
”the result has not been peer-reviewed, its novel
techniques (and in particular Kitaev’s point game
formalism) have not been applied anywhere else,
and an explicit protocol is missing,” [Ganz (2017)].
With that said, the basic setup for weak coin flipping is quite similar to the setup for algorithms II.
Figure 4 illustrates the process. Alice starts with
�
state ��𝜓𝐴,0 on space 𝐴 and Bob starts with state
�
��𝜓𝐵,0 on state 𝐵. On every odd round Alice applies
a unitary 𝑈𝐴,𝑖 and projection 𝐸𝐴,𝑖 to space 𝐴 ⊗ _𝑀,_
and on every even round Bob applies a unitary 𝑈𝑏,𝑖
and projection 𝐸𝐵,𝑖 to space 𝑀 ⊗ _𝐵. The basic idea is_
that by applying specific unitaries and projections,
an honest player can decrease any bias to arbitrar
ily small values.
Figure 4: retrieved from [Aharonov et al. (2014)]
Mochon’s paper proving this result is 80 pages,
and we do not have the time to go into detail.
However, this is an impressive result in quantum
information and demonstrates some of the beauty
of the field.
### Implementation
By using several existing quantum sofware packages, we were able to simulate Qantum Leader
Election Algorithm II. We used the packages listed
below.
- Qiskit is an open-source SDK for working
with quantum computers at the level of
pulses, circuits, and application modules.
- ProjectQ is an open-source sofware framework for quantum computing. It provides
tools for implementing and running quantum algorithms using either classical hardware.
- SimulaQron allows distributed simulation of
the nodes in a quantum internet network.
5
-----
Boston University Department of Physics
The source of implementation for simulating
Qantum Leader Election Algorithm can be found
at `https://github.com/lifuzhang1108/`
```
quantum-consensus.
```
**Pseudo code:**
1. Prepare one-qubit quantum registers
_𝑅1,…,𝑅6, 𝑇1, ...,𝑇6 and 𝑆6, ...,𝑆6._
2. For each processor, do the following:
3. If 𝑠𝑡𝑎𝑡𝑢𝑠 = “eligible,”
√
Set |𝑅𝑖 ⟩ = (|0⟩+ |1⟩)/ 2
Set |𝑆⟩ = |0⟩;
4. Apply circuit in figure 3.
5. Measure |𝑆⟩.
6. If |𝑆⟩ = |0⟩, measure |𝑅𝑖 ⟩.
If |𝑅𝑖 ⟩ = |1⟩, 𝑠𝑡𝑎𝑡𝑢𝑠 = “eligible.”
If |𝑅𝑖 ⟩ = |0⟩, 𝑠𝑡𝑎𝑡𝑢𝑠 = “ineligible.”
7. If |𝑆⟩ = |1⟩
and there are an even number of eligible
processors,
apply unitary 𝑈 to |𝑅𝑖 ⟩,
measure |𝑅𝑖 ⟩.
If |𝑅𝑖 ⟩ = |1⟩, 𝑠𝑡𝑎𝑡𝑢𝑠 = “eligible,”
If |𝑅𝑖 ⟩ = |0⟩, 𝑠𝑡𝑎𝑡𝑢𝑠 = “ineligible,”
8. If |𝑆⟩ = |1⟩
and there are an odd number of eligible
processors,
initialize 𝑇𝑖 and apply unitary 𝑉𝑛 to |𝑅𝑖𝑇𝑖 ⟩.
Measure all |𝑅𝑖𝑇𝑖 ⟩ If _𝑅𝑖𝑇𝑖_ =
_𝑚𝑎𝑥_ (𝑅1𝑇1, .., 𝑅𝑛𝑇𝑛), 𝑠𝑡𝑎𝑡𝑢𝑠 = “eligible,”
If 𝑅𝑖𝑇𝑖 _< 𝑚𝑎𝑥_ (𝑅1𝑇1, .., 𝑅𝑛𝑇𝑛), 𝑠𝑡𝑎𝑡𝑢𝑠 = “ineligible,”
9. Output status.
### Summary
Qantum computing provides tools for achieving
consensus in a distributed system. It is shown by
Tani that the classically non-deterministic anonymous leader election problem is can be solved deterministically using quantum computers. As a
proof of concept demonstration, we implemented
the quantum algorithms in Qiskit and simulated
quantum network using SimulaQron, the algorithm can successfully elect a single leader among
anonymous parties. Mochon demonstrated that
quantum consensus algorithms can be used to
fairly elect a leader even under Byzantine conditions. While these algorithms have not been used
in practice, they ofer excellent insight into both information theory and quantum mechanics.
### References
Leader election in distributed systems.
htps://d1.awsstatic.com.
Aharonov, Dorit, Andr´e Chailloux, Maor Ganz, Iordanis Kerenidis, and Lo¨ıck Magnin. 2014. A simpler proof of existence of quantum weak coin
flipping with arbitrarily small bias. SIAM Jour_nal on Computing, 45._
Ganz, Maor. 2017. Qantum leader election. Qan_tum Information Processing, 16:1–17._
Mazzarella, Luca, Alain Sarlete, and Francesco
Ticozzi. 2015. Consensus for quantum networks:
Symmetry from gossip interactions. IEEE Trans_actions on Automatic Control, 60(1):158–172._
Mochon, Carlos. 2007. Qantum weak coin flipping
with arbitrarily small bias.
Tani, Seiichiro, Hirotada Kobayashi, and Keiji Matsumoto. 2005. Qantum leader election via exact
amplitude amplification.
Tani, Seiichiro, Hirotada Kobayashi, and Keiji Matsumoto. 2012. Exact quantum algorithms for the
leader election problem. _ACM Trans. Comput._
_Theory, 4:1:1–1:24._
6
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/2206.04710, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "CLOSED",
"url": "http://arxiv.org/pdf/2206.04710"
}
| 2,022
|
[
"JournalArticle"
] | true
| 2022-06-09T00:00:00
|
[
{
"paperId": "c7824a6369e33dde3738f204797f965b37185751",
"title": "A Simpler Proof of the Existence of Quantum Weak Coin Flipping with Arbitrarily Small Bias"
},
{
"paperId": "68074c123229f29fa742e354a2c0a8b6af8057aa",
"title": "Consensus for Quantum Networks: Symmetry From Gossip Interactions"
},
{
"paperId": "c5d78f1656b6f42642173788fc431a65caad06d3",
"title": "Quantum weak coin flipping with arbitrarily small bias"
},
{
"paperId": "338dc1c9be94c6c817ce987f7ba2a8919f143241",
"title": "Exact Quantum Algorithms for the Leader Election Problem"
},
{
"paperId": "b8acfea8a2279ff72e4116a4e73ebd5cba0f5dc2",
"title": "Quantum Leader Election via Exact Amplitude Amplification"
}
] | 8,190
|
en
|
[
{
"category": "Medicine",
"source": "external"
},
{
"category": "Medicine",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/0224fec689cae36bb85c4d63ae3c5a4b060abcd3
|
[
"Medicine"
] | 0.922305
|
Infection-related hospitalization following ureteroscopic stone treatment: results from a surgical collaborative
|
0224fec689cae36bb85c4d63ae3c5a4b060abcd3
|
BMC Urology
|
[
{
"authorId": "39726081",
"name": "A. Cole"
},
{
"authorId": "49745989",
"name": "J. Telang"
},
{
"authorId": "2110993326",
"name": "Tae-Kyung Kim"
},
{
"authorId": "118276505",
"name": "K. Swarna"
},
{
"authorId": "152322873",
"name": "J. Qi"
},
{
"authorId": "3808902",
"name": "C. Dauw"
},
{
"authorId": "9831716",
"name": "B. Seifman"
},
{
"authorId": "6599584",
"name": "M. Abdelhady"
},
{
"authorId": "152413146",
"name": "W. Roberts"
},
{
"authorId": "31678075",
"name": "J. Hollingsworth"
},
{
"authorId": "145918894",
"name": "K. Ghani"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"BMC Urol"
],
"alternate_urls": [
"http://www.pubmedcentral.nih.gov/tocrender.fcgi?journal=67",
"https://bmcurol.biomedcentral.com/"
],
"id": "2369a370-b3f9-4eec-9f80-5e9540ea27e2",
"issn": "1471-2490",
"name": "BMC Urology",
"type": "journal",
"url": "http://www.biomedcentral.com/bmcurol/"
}
|
Background Unplanned hospitalization following ureteroscopy (URS) for urinary stone disease is associated with patient morbidity and increased healthcare costs. To this effect, AUA guidelines recommend at least a urinalysis in patients prior to URS. We examined risk factors for infection-related hospitalization following URS for urinary stones in a surgical collaborative. Methods Reducing Operative Complications from Kidney Stones (ROCKS) is a quality improvement (QI) initiative from the Michigan Urological Surgery Improvement Collaborative (MUSIC) consisting of academic and community practices in the State of Michigan. Trained abstractors prospectively record standardized data elements from the health record in a web-based registry including patient characteristics, surgical details and complications. Using the ROCKS registry, we identified all patients undergoing primary URS for urinary stones between June 2016 and October 2017, and determined the proportion hospitalized within 30 days with an infection-related complication. These patients underwent chart review to obtain clinical data related to the hospitalization. Multivariable logistic regression analysis was performed to determine risk factors for hospitalization. Results 1817 URS procedures from 11 practices were analyzed. 43 (2.4%) patients were hospitalized with an infection-related complication, and the mortality rate was 0.2%. Median time to admission and length of stay was 4 and 3 days, respectively. Nine (20.9%) patients did not have a pre-procedure urinalysis or urine culture, which was not different in the non-hospitalized cohort (20.5%). In hospitalized patients, pathogens included gram-negative (61.5%), gram-positive (19.2%), yeast (15.4%), and mixed (3.8%) organisms. Significant factors associated with infection-related hospitalization included higher Charlson comorbidity index, history of recurrent UTI, stone size, intra-operative complication, and procedures where fragments were left in-situ. Conclusions One in 40 patients are hospitalized with an infection-related complication following URS. Awareness of risk factors may allow for individualized counselling and management to reduce these events. Approximately 20% of patients did not have a pre-operative urine analysis or culture, and these findings demonstrate the need for further study to improve urine testing and compliance
|
p g
## RESEARCH ARTICLE
## Open Access
# Infection‑related hospitalization following ureteroscopic stone treatment: results from a surgical collaborative
### Adam Cole[1*], Jaya Telang[1], Tae‑Kyung Kim[1], Kavya Swarna[1], Ji Qi[1], Casey Dauw[1], Brian Seifman[2], Mazen Abdelhady[3], William Roberts[1], John Hollingsworth[1] and Khurshid R. Ghani[1] on behalf of for the Michigan Urological Surgery Improvement Collaborative
**Abstract**
**Background: Unplanned hospitalization following ureteroscopy (URS) for urinary stone disease is associated with**
patient morbidity and increased healthcare costs. To this effect, AUA guidelines recommend at least a urinalysis in
patients prior to URS. We examined risk factors for infection-related hospitalization following URS for urinary stones in
a surgical collaborative.
**Methods: Reducing Operative Complications from Kidney Stones (ROCKS) is a quality improvement (QI) initiative**
from the Michigan Urological Surgery Improvement Collaborative (MUSIC) consisting of academic and community
practices in the State of Michigan. Trained abstractors prospectively record standardized data elements from the
health record in a web-based registry including patient characteristics, surgical details and complications. Using the
ROCKS registry, we identified all patients undergoing primary URS for urinary stones between June 2016 and Octo‑
ber 2017, and determined the proportion hospitalized within 30 days with an infection-related complication. These
patients underwent chart review to obtain clinical data related to the hospitalization. Multivariable logistic regression
analysis was performed to determine risk factors for hospitalization.
**Results: 1817 URS procedures from 11 practices were analyzed. 43 (2.4%) patients were hospitalized with an infec‑**
tion-related complication, and the mortality rate was 0.2%. Median time to admission and length of stay was 4 and
3 days, respectively. Nine (20.9%) patients did not have a pre-procedure urinalysis or urine culture, which was not
different in the non-hospitalized cohort (20.5%). In hospitalized patients, pathogens included gram-negative (61.5%),
gram-positive (19.2%), yeast (15.4%), and mixed (3.8%) organisms. Significant factors associated with infection-related
hospitalization included higher Charlson comorbidity index, history of recurrent UTI, stone size, intra-operative com‑
plication, and procedures where fragments were left in-situ.
**Conclusions: One in 40 patients are hospitalized with an infection-related complication following URS. Awareness**
of risk factors may allow for individualized counselling and management to reduce these events. Approximately 20%
of patients did not have a pre-operative urine analysis or culture, and these findings demonstrate the need for further
study to improve urine testing and compliance
**Keywords: Ureteroscopy, Infection, Outcomes, Quality improvement, Urolithiasis**
*Correspondence: aicole@med.umich.edu
1 Department of Urology, University of Michigan, Ann Arbor, MI 48103,
USA
Full list of author information is available at the end of the article
**Background**
Ureteroscopy (URS) is now the most common treatment modality for treating upper urinary tract stones in
North America [1, 2]. Due to technological advances and
© The Author(s). 2020 Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which
permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the
original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or
other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line
to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory
regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this
[licence, visit http://creativecommons.org/licenses/by/4.0/. The Creative Commons Public Domain Dedication waiver (http://creativeco](http://creativecommons.org/licenses/by/4.0/)
[mmons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data.](http://creativecommons.org/publicdomain/zero/1.0/)
-----
widespread availability of equipment, URS is often performed in the outpatient setting [3]. Despite this, morbidity, especially infection-related complications, may
occur in up to 5–18% of patients [4–8]. These often result
in hospital admission and can have a significant impact
on patients, providers, and payers [3, 9–11]. A hospital
admission for sepsis can cost approximately $20,000 [12].
Therefore, efforts to mitigate infection-related complications following URS would be beneficial in reducing
healthcare expenditures.
Prior studies investigating infection-related complications after URS have provided some insights on
risk factors, which include stone, patient, and operative characteristics. However, most are single institution
series from tertiary referral or academic medical centers
[4–10], which may limit generalizability of the results to
the wider swathe of urologic patients commonly treated
by diverse practitioners in community or multi-specialty
group practices.
In the state of Michigan, we have developed a quality improvement (QI) initiative and a clinical registry—
Reducing Operative Complications from Kidney Stones
(ROCKS)—to better understand processes of care, outcomes, and quality indicators for patients undergoing
URS for urinary stones. A strength of this registry is its
diversity of patients and practicing urologists. In our
drive to improve outcomes for URS, we sought to better understand risk factors for infection-related hospitalization using data from this surgical collaborative. We
also sought to assess care in relation to guideline-based
practice. We hypothesize that there are modifiable factors which lead to infection related morbidity. Identifying
high-risk patients may allow for individualized counseling, and development of QI interventions that reduce
adverse events, and the associated patient morbidity and
healthcare costs.
**Methods**
**Data source**
The Michigan Urological Surgery Improvement Collaborative (MUSIC) was established in 2011 in partnership
with Blue Cross Blue Shield of Michigan. The ROCKS
QI initiative within MUSIC comprises diverse community and academic urology practices in the state of
Michigan and started in 2016. For patients with urinary
stones undergoing URS, trained abstractors prospectively record standardized data elements in a web-based
registry including patient and stone characteristics, surgical details and complications. Patient data are entered
into the registry 60 days after a URS procedure, and data
entry is guided by standard variable definitions and collaborative-wide operating procedures. To ensure data
quality, the coordinating center performs on-site data
audits on a semi-annual basis.
**Patient selection and outcomes**
We identified all patients undergoing URS for primary
treatment of urinary stones between June 2016 and
October 2017. During this period, ROCKS consisted of
11 practices. To be included in the ROCKS registry, a
patient had to be at least 18 years of age and undergone
unilateral URS for urinary stones. Patients who underwent bilateral URS, had an ipsilateral nephrostomy at
the time of URS, or underwent URS after percutaneous
renal surgery were ineligible. We identified all patients
who were discharged after surgery and then subsequently
hospitalized (at any institution) within 30 days of their
procedure. An infection-related hospitalization was
determined by chart review, based on the presence of
SIRS criteria with or without bacteriuria. Patients admitted for other indications (pain, hematuria, etc.) were classified as a non-infectious hospitalization. Stone-free rate
(SFR) was defined as absence of any fragment on X-ray,
CT or ultrasound reports obtained within 60 days. Chart
review was performed on all patients with infectionrelated hospitalizations, including urine culture pathogen
data, length of stay, and timing from surgery.
**Statistical analyses**
We generated descriptive summary statistics of all
patients in the analytic sample. Chi-square tests and student’s t tests were performed for categorical and continuous variables, respectively, to compare demographic and
operative factors between the two groups. Significant
pre-operative and operative variables were then used as
covariates in a multivariable analysis to determine which
factors were associated with higher odds of an infectionrelated hospitalization. Multivariable analysis was performed using a logistic regression model. The odds ratios
and 95% confidence intervals were reported. Significant
variables with less than 10 events were not included in
the multivariable final model. All analyses were performed with SAS 9.4 (SAS Institute, Cary, NC) at a 5%
significance level.
**Results**
A total of 1817 URS procedures in 1737 patients from
11 practices were analyzed. In total, 80 (4.4%) patients
were hospitalized within 30 days of their URS. Fortythree (2.4%) patients were hospitalized with an infectionrelated complication (Fig. 1). Median time from surgery
to admission was 4 days (range 0–30) and median length
of stay was 3 days (range 1–33) for the patients admitted with an infection-related complication. The majority
of admissions (74.4%) occurred within 7 days of surgery,
-----
and more than half of patients (55.8%) were admitted for
longer than 2 days. One (2.3%) patient had a prior ureteroscopy within 1 month of the index surgery. Of patients
with a positive urine culture during hospitalization
(n = 26), isolated pathogens included 16 (61.5%) gramnegative, 5 (19.2%) gram-positive, 4 (15.4%) yeast, and 1
(3.8%) with gram-positive and -negative cocci. Only 9 of
these 26 patients (34.6%) had positive urinalysis (defined
by positive nitrite) or urine culture prior to surgery.
Three patients died during their hospitalization (mortality rate 0.2%).
Pre-operative, intra-operative, and post-operative
characteristics among infection-related hospitalized,
and non-hospitalized patients, are provided in Table 1.
Significant factors for hospitalization with an infectionrelated complication on bivariate analysis were public
insurance status, older age, higher Charlson Comorbidity
Index (CCI), history of recurrent UTI (registry variable
based on clinic note by physician indicating history of
prior UTIs), spinal cord injury, urinary diversion, intraoperative complication, and larger stone size. Patients
that were hospitalized were less likely to be on pre-operative alpha-blockers. There was no statistical difference
in the proportion of patients who had an indwelling ureteral stent prior to URS. Of those hospitalized with an
infection-related complication, 9 (20.9%) did not have a
pre-procedure urinalysis or urine culture, compared to
355 (20.5%) in the non-hospitalized group (p = 0.95). 12
(27.9%) patients who were hospitalized with an infectious
indication had a positive urinalysis or urine culture prior
to surgery, compared to 261 (15.4%) in the non-hospitalized patients (p = 0.08). None of the 12 patients with
abnormal pre-operative urine studies who were hospitalized were treated with antibiotics prior to surgery.
Patients who were hospitalized for infectious reasons
were more likely to have an intra-operative complication (7.0%). Complications included inability to complete
procedure due to bleeding or perforation. There was no
difference in the rate of ureteral stent placement, ureteral dilation, or use of ureteral access sheath at the time
of surgery between the infection-related hospitalization
group and the non-admitted group. Those hospitalized
with an infection-related complication were more likely
to have lithotripsy with fragments left in situ at the conclusion of the operation.
On multivariable analysis (Table 2), significant risks
factors associated with hospitalization for infectionrelated causes included higher CCI, history of recurrent
UTI, increasing stone size, history of intra-operative
complication, and lithotripsy with fragments left in-situ.
The strongest risk factors were the presence of an intraoperative complication (OR 3.7) and history of recurrent
UTI (OR 3.74).
**Discussion**
We found that in 11 diverse urology practices across the
state of Michigan, 1 in 40 patients were hospitalized with
an infection-related complication following URS for urinary stones. During admission the most commonly identified organisms were gram-negative, however a small
proportion of patients had yeast identified. Risk factors
for an infection-related admission were higher Charlson
comorbidity index, history of recurrent UTI, larger stone
size, intra-operative complication and cases where lithotripsy was performed with fragments left in-situ. Overall,
20% of all patients did not have a documented urinalysis
or urine culture prior to URS. Collectively, these findings
represent an opportunity for the development of QI initiatives to decrease the risk of infection and sepsis after
URS, as well as better adherence to American Urological
Association (AUA) guidelines.
Previous investigators have examined risk factors for
infectious complications following URS. Zhong et al.
examined 250 patients that underwent URS for stone
treatment, and found an 8.1% incidence of systemic
inflammatory response syndrome (SIRS) following the
procedure. Risk-factors included stone size, smaller
caliber ureteral access sheath, higher irrigation flow
rate, and presence of struvite calculi [8]. Other studies
have also identified female gender [5, 6, 13], history of
-----
**Table 1 Patient characteristics for patients undergoing ureteroscopy for urinary stones in MUSIC ROCKS stratified**
**by post-operative course**
**Risk factor** **Infection-related hospitalization** **Non-hospitalized (n = 1737)** **_P value_**
**(n = 43)**
_Pre-operative characteristics_
Public insurance 25 (58.1%) 669 (39.7%) 0.01
Mean age (SD) 60.1 (15.8) 54.4 (15.5) 0.02
Male gender 19 (44.2%) 867 (50.1%) 0.44
BMI > 30 24 (58.5%) 788 (46.8%) 0.13
CCI ≥ 1 26 (60.5%) 495 (28.5%) < 0.01
CCI ≥ 2 14 (32.6%) 241 (13.9%) < 0.01
Presence of hydronephrosis on pre-operative imaging 27 (67.5%) 1048 (66.7%) 0.92
Largest stone size (mm), mean (SD) 10.1 (6.5) 7.8 (5.4) < 0.01
Solitary kidney 2 (4.7%) 25 (1.5%) 0.13
Horseshoe kidney 1 (2.3%) 6 (0.4%) 0.16
History of recurrent UTI 9 (20.9%) 88 (5.1%) < 0.01
Urinary diversion 2 (4.7%) 7 (0.4%) 0.02
Spinal cord injury 2 (4.7%) 3 (0.2%) < 0.01
Anti-platelet therapy 13 (30.2%) 345 (20.4%) 0.12
Pre-operative urinalysis/urine culture not performed 9 (20.9%) 355 (20.5%) 0.95
Positive pre-operative urinalysis/urine culture 12 (27.9%) 266 (15.4%) 0.08
Positive pre-operative urinalysis/urine culture treated 0 (0%) 59 (22.2%) 0.07
Urgent/emergent surgery 1 (2.3%) 146 (8.4%) 0.15
Peri-operative antibiotic use 38 (95%) 1513 (96.9%) 0.49
Alpha-blocker therapy prior to URS 11 (26.2%) 755 (45.1%) 0.01
Pre-stenting (ureteral stent in place) 20 (46.5%) 649 (37.5%) 0.23
_Stone location_
Renal 17 (45.9%) 502 (31.0%) 0.07
Ureter 14 (37.8%) 906 (56.0%)
Both 6 (16.2%) 211 (13.0%)
_Intra-operative characteristics_
Intra-operative complication 3 (7.0%) 33 (1.9%) < 0.01
Complication: bleeding 2 (4.7%) 14 (0.8)
Complication: perforation 0 (0.0%) 4 (0.2%)
Complication: other 1 (2.3%) 15 (0.9%)
Ureteral dilation 6 (13.9%) 340 (19.7%) 0.44
Ureteral access sheath use 18 (41.9%) 626 (36.6%) 0.49
Lithotripsy with fragments left in-situ 30 (69.8%) 716 (42.3%) < 0.01
Stenting during URS 31 (72.1%) 1248 (72.1%) 0.99
_Post-operative characteristics_
Discharged with antibiotics 15 (36.6%) 638 (39.2%) 0.74
Discharged with antibiotics and stent placed 9 (20.9%) 509 (29.3%) 0.23
Discharged with alpha-blocker 27 (65.8%) 911 (55.9%) 0.21
Stone free rate 19 (57.6%) 579 (77.5%) < 0.01
obstructive pyelonephritis [5, 6], positive pre-operative
urine culture [5, 6], and prolonged ureteral stent dwell
time [5] as risk factors for SIRS/sepsis, with rates of
SIRS/sepsis from 0.30–8% [5–8, 14]. We also found similar risk factors for hospitalization related to infectious
complications, including higher Charlson comorbidity
index, history of recurrent UTI, intra-operative complication, and stone size. Interestingly, female gender
and pre-operative ureteral stenting were not risk factors
in this analysis. Female gender has been a reported risk
factor in some series [5, 6, 13], however in other studies
this was not a risk factor [7, 14] suggesting differences in
-----
**Table 2 Multi-variable logistic regression demonstrating**
**association between patient characteristics and risk**
**of infection-related hospitalization**
**Risk factor** **OR** **CI** **_P value_**
Age 1.01 0.98–1.03 0.95
Comorbidity (CCI 0 vs. 1) 3.12 1.37–7.14 < 0.01
Comorbidity (CCI 0 vs. 2) 2.72 1.16–6.37 0.02
Stone size 1.04 1.01–1.07 0.02
History of recurrent UTI 3.74 1.55–9.00 < 0.01
Insurance (public vs. private) 1.57 0.75–3.25 0.23
Alpha-blocker prior to URS 0.51 0.24–1.06 0.07
Complete fragment removal 0.32 0.16–0.65 < 0.01
Intra-operative complication 3.70 1.22–11.25 0.02
study design. Perhaps a prospective study would be helpful. Additionally, public insurance was associated with an
increased risk of an infectious-hospitalization on univariate analysis. However, this association was not seen in
our multi-variable model, suggesting that the association
of insurance and infection may be due to other factors.
Awareness of risk factors can allow for an individualized approach to pre-operative antibiotic selection,
adoption of intra-operative technical factors such as
considering a ureteral access sheath or limiting the irrigation flow rate, and post-operative antibiotic therapy in
patients at risk for developing sepsis. Since there was a
strong relationship between an intra-operative complication and subsequent hospitalization, patients who suffer
this event could be considered for prolonged observation
in the recovery room or even admission and observation.
Likewise, patients with a history of recurrent UTI should
be considered for pre-operative urine culture (not urinalysis) and be managed with culture-directed pre-operative
antibiotics. While almost all patients received peri-operative antibiotics, more patients in the hospitalized group
had an abnormal urine study prior to surgery, and none
of these patients were treated with antibiotics. There
are a very small number of patients in both groups with
untreated positive urine cultures prior to ureteroscopy.
This represents a focus for subsequent quality improvement initiatives with the goal to improve pre-operative
testing and follow-up.
We found that in patients with a positive urine culture
during hospitalization, only 34.6% had a positive UA or
urine culture prior to surgery. It is possible this discordance lies in our definition of a positive urinalysis (nitrite
positivity), which can be altered by medications such
as pyridium. Additionally, any positive pre-op culture,
regardless of organism or colony count, is deemed positive. These represent limitations of our study, however,
previous studies have also reported discordance between
pre-operative, intra-operative and post-operative urine
cultures in patients undergoing stone surgery. Paonessa
et al. examined pre-operative urine cultures and intraoperative stone cultures in patients undergoing percutaneous nephrolithotomy and found that 9.7% of patients with
negative pre-operative urine cultures had positive stone
cultures. In patients with both positive pre-operative
urine and intra-operative stone cultures, the organisms
differed in 13.3%, representing an overall discordance in
almost a quarter of cases [15]. Marien et al. also demonstrated 27% discordant voided and upper tract urine cultures after decompression for obstructing stones [16].
The AUA Guidelines on Surgical Management of Stones
advises clinicians to obtain a urinalysis prior to URS, and
in patients with clinical or laboratory signs of infection,
a urine culture should be obtained [17]. EAU Guidelines
state a urine culture or urinary microscopy are mandatory before treatment [18]. In our cohort, approximately
20% of patients who were admitted with an infectious
complication did not have a urinalysis or urine culture
prior to surgery. This aspect of care, where patients are
not managed in accordance with current guidelines represents an area for improvement. Interestingly, this rate
was similar in the group of non-hospitalized patients. It
would appear that obtaining a pre-operative urinalysis or
urine culture did not alter the risk of hospitalization for
an infection-related reason. One major limitation of our
work is that we do not differentiate between urine culture
or urinalysis in our registry. Additionally, the pre-operative screening requirements vary by center due to institutional protocols, work-flow, staffing, and resources.
Some institutions require a urine culture within 30 days
of surgery, while others use urinalysis with reflex culture.
Despite these difference, pre-operative urine studies were
not obtained in approximately 20% of all patients, likely
for a variety of reasons: urine studies may not have been
ordered, urine studies were ordered but not performed
by the patient, or they were performed at outside institutions but not available. Our findings warrant further
investigation to address these quality of care gaps.
Lithotripsy with fragments left in-situ was associated
with an increased risk of infection-related hospitalization in our cohort. This variable is determined by review
of the operative notes by data abstractors based on key
phrases, such as “all fragments were removed,” or “all
remaining fragments were 1 mm of less.” The database
does not detect specific stone treatment technique, and
it is difficult to ascertain if this indicates dusting technique, or a hybrid technique of basketing and dusting.
It is possible that our results are confounded by patients
with large stone burden. Also, while patients in the nonhospitalized group were more likely to be on pre-operative alpha-blockers, this was not significantly associated
-----
with a lower risk of an infection-related hospitalization
on multi-variate analysis. In a recent study, 1 week of
pre-operative alpha-blocker therapy was associated with
lower overall complications after URS [19]. The same
mechanism by which alpha-blockers are prescribed to
facilitate ureteral stone passage—inhibition of the alpha
receptors in the distal ureter and reduced ureteral muscle tone and peristalsis—has been proposed to facilitate
ureteroscopy and instrumentation of the ureter [20]. This
is an area of interest that will be the subject of future
investigation.
Our work has several limitations. First, our patients are
located in a single state and it is possible these results are
not applicable to all patients in the United States or outside the country. Our registry does not collect information on pre-operative ureteral stent dwell-time, technical
intraoperative details such as size of access sheath, irrigation rate, surgical time, and other factors that may place
patients at higher risk for developing an infectious complication. In addition, stone cultures are not captured by
registry. Additionally, it is possible we are underreporting
the number of events based on our study design. From
the registry we were able to identify all patients that were
hospitalized within 30 days. We then made a determination if the hospitalization was due to infectious or noninfectious etiologies (pain, hematuria, etc.) based on
chart review. Along with the admission notes, SIRS criteria were used to determine if the admission was due to
infectious-indication, as culture data was not available or
obtained after administration of antibiotics. Therefore it
is possible that some patients were admitted with infectious-indications without SIRS criteria. Finally, the small
number of hospitalization events may alter the fit of our
multi-variable model.
Our findings do have several implications. URS is
among the most commonly performed urologic surgeries, and unplanned healthcare encounters following URS
are not uncommon. We demonstrate suboptimal statewide compliance with guidelines regarding pre-operative urine screening. Efforts should be taken to comply
with best practice statements, and this will be the subject of future QI initiatives in MUSIC. In particular we
are considering collecting information on which specific
pre-operative urine study was performed to determine
whether urinalysis is insufficient as a screening tool to
mitigate the risk of sepsis after URS.
**Conclusion**
We found that nearly 1 in 40 patients are hospitalized with
an infection-related complication following URS for urinary stones in diverse practices in Michigan. Awareness of
risk factors may allow for individualized counselling and
management to reduce these events. Approximately 20% of
patients did not have a pre-operative urine analysis or culture, and these findings demonstrate the need for further
study to improve urine testing and compliance.
**Abbreviations**
URS: Ureteroscopy; AUA: American Urological Association; ROCKS: Reducing
Operative Complications from Kidney Stones; MUSIC: Michigan Urological
Surgery Improvement Collaborative; QI: Quality improvement; SFR: Stone-free
rate; UTI: Urinary tract infection; CCI: Charlson Comorbidity Index; OR: Odds
ratio; SIRS: Systemic inflammatory response syndrome; UA: Urinalysis; EAU:
European Association of Urology.
**Acknowledgements**
We would like to thank the significant contribution of the clinical champions,
urologists and data abstractors in each participating MUSIC ROCKS practice.
In addition, we would like to acknowledge the support provided by the Value
Partnerships program at BCBSM.
**Authors’ contributions**
Made substantial contributions to the conception: AC, KG. Made substantial
contributions to design of the work: AC, JT, TK, KS, JQ, Made substantial con‑
tributions to the acquisition and analysis of the data: AC, JT, TM, KS, JQ. Made
substantial contributions to interpretation of the data: AC, CD, WR, JH, KG.
Made substantial contributions to drafting and revising the manuscript: AC,
CD, BS, MA, WR, JH, KG. All authors read and approved the final manuscript.
**Funding**
Michigan Urological Surgery Improvement Collaborative (MUSIC) is funded
by Blue Cross Blue Shield of Michigan (BCBSM). Blue Cross Blue Shield of
Michigan did not have a role in the design and conduct of the study; col‑
lection, management, analysis, and interpretation of the data; preparation,
review, or approval of the manuscript; and decision to submit the manuscript
for publication.
**Availability of data and materials**
The datasets generated and/or analyzed during the current study are not
publicly available, and are managed by the MUSIC urology coordinating
center. MUSIC urology was founded with the guiding principle to improve the
urologic care across the entire state. We do not compare institutions within
our registry for the purposes of maintaining confidentiality. As such, our data
are internally maintained and not publically available.
**Ethics approval and consent to participate**
The MUSIC registry was issued a Notice of Determination of “Not Regulated”
Status by the University of Michigan Institutional Review Board (IRBMED),
ID: HUM00054438. This registry did not fit the definition of human subjects
research requiring IRB approval because the program is focused on quality
improvement versus research and the human subjects themselves.
**Consent for publication**
Not applicable.
**Competing interests**
The authors declare that they have no competing interests.
**Author details**
1 Department of Urology, University of Michigan, Ann Arbor, MI 48103, USA.
2 Michigan Institute of Urology, West Bloomfield, MI 48322, USA. 3 Detroit
Medical Center, Department of Urology, Detroit, MI 48201, USA.
Received: 4 April 2020 Accepted: 15 September 2020
-----
**References**
1. Oberlin DT, Flum AS, Bachrach L, Matulewicz RS, Flury SC. Contempo‑
rary surgical trends in the management of upper tract calculi. J Urol.
2015;193:880–4.
2. Ordon M, et al. The surgical management of kidney stone disease: a
population based time series analysis. J Urol. 2014;192:1450–6.
3. San Juan J, Hou H, Ghani KR, Dupree JM, Hollingsworth JM. Variation in
spending around surgical episodes of urinary stone disease: findings
from Michigan. J Urol. 2018;199:1277–82.
4. Mitsuzuka K, Nakano O, Takahashi N, Satoh M. Identification of factors
associated with postoperative febrile urinary tract infection after ureter‑
oscopy for urinary stones. Urolithiasis. 2016;44:257–62.
5. Nevo A, Mano R, Baniel J, Lifshitz DA. Ureteric stent dwelling time: a risk
factor for post-ureteroscopy sepsis. BJU Int. 2017;120:117–22.
6. Uchida Y, Takazawa R, Kitayama S, Tsujii T. Predictive risk factors for
systemic inflammatory response syndrome following ureteroscopic laser
lithotripsy. Urolithiasis. 2018;46:375–81.
7. Blackmur JP, et al. Analysis of factors’ association with risk of postoperative
urosepsis in patients undergoing ureteroscopy for treatment of stone
disease. J Endourol. 2016;30:963–9.
8. Zhong W, Leto G, Wang L, Zeng G. Systemic inflammatory response
syndrome after flexible ureteroscopic lithotripsy: a study of risk factors. J
Endourol. 2015;29:25–8.
9. Bloom J, Matthews G, Phillips J. Factors influencing readmission after
elective ureteroscopy. J Urol. 2016;195:1487–91.
10. Du K, et al. Unplanned 30-day encounters after ureterorenoscopy for
urolithiasis. J Endourol. 2018;32:1100–7.
11. Scales CD, et al. The impact of unplanned postprocedure visits in the
management of patients with urinary stones. Surgery. 2014;155:769–75.
12. Arefian H, et al. Hospital-related cost of sepsis: a systematic review. J
Infect. 2017;74:107–17.
13. Martov A, et al. Postoperative infection rates in patients with a nega‑
tive baseline urine culture undergoing ureteroscopic stone removal: a
matched case–control analysis on antibiotic prophylaxis from the CROES
URS global study. J Endourol. 2015;29:171–80.
14. Somani BK, et al. Complications associated with ureterorenoscopy
(URS) related to treatment of urolithiasis: the Clinical Research Office of
Endourological Society URS Global study. World J Urol. 2017;35:675–81.
15. Paonessa JE, Gnessin E, Bhojani N, Williams JC, Lingeman JE. Preopera‑
tive bladder urine culture as a predictor of intraoperative stone culture
results: clinical implications and relationship to stone composition. J Urol.
2016;196:769–74.
16. Marien T, Mass AY, Shah O. Antimicrobial resistance patterns in cases of
obstructive pyelonephritis secondary to stones. Urology. 2015;85:64–8.
17. Assimos D, et al. Surgical management of stones: American uro‑
logical association/endourological society guideline, PART I. J Urol.
2016;196:1153–60.
18. Türk C, et al. EAU guidelines on interventional treatment for urolithiasis.
Eur Urol. 2016;69:475–82.
19. Sokhal A, et al. Do pre-operative alpha blockers facilitate ureteroscope
insertion at the vesico-ureteric junction? An answer from a prostective
case–controlled study. Eur Med J. 2017;2:82–6.
20. Alsaikhan B, Koziarz A, Lee J, Pace K. Preoperative alpha-blockers for
ureteroscopy for ureteral stones: a systematic review and meta-analysis of
randomized controlled trials. J Endourol. 2020;34:33–41.
21. Margel D, et al. Clinical implication of routine stone culture in percutane‑
ous nephrolithotomy—a prospective study. J Urol. 2006;67:26–9.
**Publisher’s Note**
Springer Nature remains neutral with regard to jurisdictional claims in pub‑
lished maps and institutional affiliations.
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://pmc.ncbi.nlm.nih.gov/articles/PMC7607640, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "GOLD",
"url": "https://bmcurol.biomedcentral.com/track/pdf/10.1186/s12894-020-00720-4"
}
| 2,020
|
[
"JournalArticle",
"Review"
] | true
| 2020-11-03T00:00:00
|
[
{
"paperId": "33e1185d60ea7e6205382bf0e0c34e34dca5de28",
"title": "Preoperative alpha-blockers for ureteroscopy for ureteric stones: a systematic review and meta-analysis of randomized controlled trials."
},
{
"paperId": "f8050ba31a9d825e6e17d63a1b6ac10a7b250a95",
"title": "Unplanned 30-Day Encounters After Ureterorenoscopy for Urolithiasis."
},
{
"paperId": "1b9e2f9b4d5de4087e800f2f1ee1d5ef2d3a9b89",
"title": "Predictive risk factors for systemic inflammatory response syndrome following ureteroscopic laser lithotripsy"
},
{
"paperId": "631a8efe123b5637c3841d7e1ae5e223434d3178",
"title": "Variation in Spending around Surgical Episodes of Urinary Stone Disease: Findings from Michigan"
},
{
"paperId": "c7f2fd44f52df94e1b90305ece9d88dfed9a9bf8",
"title": "Ureteric stent dwelling time: a risk factor for post‐ureteroscopy sepsis"
},
{
"paperId": "b7f638363c21a56620d5374cbfe46960a9ecd4a0",
"title": "Hospital-related cost of sepsis: A systematic review."
},
{
"paperId": "0258eee3a5a0168630716ee1540b96ad808f9f51",
"title": "Surgical Management of Stones: American Urological Association/Endourological Society Guideline, PART I."
},
{
"paperId": "520b02ef54beb6ae44329e0db05e6bebc34fa1dc",
"title": "Analysis of Factors' Association with Risk of Postoperative Urosepsis in Patients Undergoing Ureteroscopy for Treatment of Stone Disease."
},
{
"paperId": "32da42ebda1955581f19f294e7ee8835d332effe",
"title": "Preoperative Bladder Urine Culture as a Predictor of Intraoperative Stone Culture Results: Clinical Implications and Relationship to Stone Composition."
},
{
"paperId": "705d5bfd5b6366e6b6a1e1830af88ff435a7bdfd",
"title": "Complications associated with ureterorenoscopy (URS) related to treatment of urolithiasis: the Clinical Research Office of Endourological Society URS Global study"
},
{
"paperId": "da70f9ccc04c43df38f1d8a7df39ae3407f4de46",
"title": "Identification of factors associated with postoperative febrile urinary tract infection after ureteroscopy for urinary stones"
},
{
"paperId": "b2763dec136065e5800f64abafc705897deb57fa",
"title": "Factors Influencing Readmission after Elective Ureteroscopy."
},
{
"paperId": "01ccb68973043d4faf61f48bcc3b6eb5db2cc968",
"title": "EAU Guidelines on Interventional Treatment for Urolithiasis."
},
{
"paperId": "1f1ea544f14135d571edfd655be402062715374e",
"title": "Postoperative infection rates in patients with a negative baseline urine culture undergoing ureteroscopic stone removal: a matched case-control analysis on antibiotic prophylaxis from the CROES URS global study."
},
{
"paperId": "8b38d4835f39e6e2d44f53a76ddcf6b1636c8e2f",
"title": "Systemic inflammatory response syndrome after flexible ureteroscopic lithotripsy: a study of risk factors."
},
{
"paperId": "7824449747113e519bff0b5279132ddebbb0ad58",
"title": "The surgical management of kidney stone disease: a population based time series analysis."
},
{
"paperId": "3f44eb409ffa0ee009d0227cdc3dfbb372eb7753",
"title": "The impact of unplanned postprocedure visits in the management of patients with urinary stones."
},
{
"paperId": "ada76768377758f114cf28a7ef936acfbfb591b0",
"title": "Contemporary surgical trends in the management of upper tract calculi."
},
{
"paperId": null,
"title": "Do pre‐operative alpha blockers facilitate ureteroscope insertion at the vesico‐ureteric junction? An answer from a prostective case–controlled study"
},
{
"paperId": "10dce973d02844eaefed790886946d4a38ed714f",
"title": "Antimicrobial resistance patterns in cases of obstructive pyelonephritis secondary to stones."
},
{
"paperId": "f3ea80f42e37f5d01aad7a805975c7e562a7ecff",
"title": "Clinical implication of routine stone culture in percutaneous nephrolithotomy--a prospective study."
},
{
"paperId": "34b9635d7779e219e9d60e0d3d33919ca9bc123c",
"title": "Publisher's Note"
},
{
"paperId": null,
"title": "Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations"
}
] | 7,884
|
en
|
[
{
"category": "Computer Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/0227d8aca57de41a14d396350e12f28d4aabe480
|
[] | 0.902386
|
Towards Improving Privacy and Security of Identity Management Systems Using Blockchain Technology: A Systematic Review
|
0227d8aca57de41a14d396350e12f28d4aabe480
|
Applied Sciences
|
[
{
"authorId": "2193919526",
"name": "H. Alanzi"
},
{
"authorId": "1404419908",
"name": "M. Alkhatib"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"Appl Sci"
],
"alternate_urls": [
"http://www.mathem.pub.ro/apps/",
"https://www.mdpi.com/journal/applsci",
"http://nbn-resolving.de/urn/resolver.pl?urn=urn:nbn:ch:bel-217814"
],
"id": "136edf8d-0f88-4c2c-830f-461c6a9b842e",
"issn": "2076-3417",
"name": "Applied Sciences",
"type": "journal",
"url": "http://www.e-helvetica.nb.admin.ch/directAccess?callnumber=bel-217814"
}
|
An identity management system (IDMS) manages and organizes identities and credentials information exchanged between users, identity providers (IDPs), and service providers (SPs) to ensure confidentiality and enhance privacy of users’ personal data. Traditional or centralized IDMS rely on a third party to store a user’s personal information, authenticate the user, and organize the entire process. This clearly constitutes threats to the privacy of the user, in addition to other issues, such as single point of failure (SPOF), user tracking, and data availability issues. Blockchain technology has many useful features that can contribute to solving traditional IDMS issues, such as decentralization, immutability, and anonymity. Blockchain represents an attractive solution for many issues related to traditional IDMS, including privacy, third-party control, data leakage, and SPOF, supported by Distributed Ledger Technology (DLT) security features and powerful smart contracts technology. The current study presents a systematic literature review and analysis for recently proposed solutions that adopt the traditional centralized approach, as well as solutions based on blockchain technology. The study also aims to provide a deep understanding of proposed IDMS solutions and best practices, and highlight the research gaps and open issues related to IDMSs and users’ privacy. In particular, the current research focuses on analyzing the blockchain-based solutions and illustrating their strengths and weaknesses, as well as highlighting the promising blockchain technology framework that can be utilized to enhance privacy and solve security issues in a centralized IDMS. Such a study is an important step towards developing efficient solutions that address the pressing needs in the field.
|
## applied sciences
_Systematic Review_
### Towards Improving Privacy and Security of Identity Management Systems Using Blockchain Technology: A Systematic Review
**Haifa Alanzi * and Mohammad Alkhatib ***
Department of Computer Science, College of Computer and Information Sciences, Imam Mohammad Ibn Saud
Islamic University, Riyadh 11564, Saudi Arabia
*** Correspondence: haifaalanzi1995@gmail.com (H.A.); mohkhatib83@gmail.com (M.A.)**
**Citation: Alanzi, H.; Alkhatib, M.**
Towards Improving Privacy and
Security of Identity Management
Systems Using Blockchain
Technology: A Systematic Review.
_[Appl. Sci. 2022, 12, 12415. https://](https://doi.org/10.3390/app122312415)_
[doi.org/10.3390/app122312415](https://doi.org/10.3390/app122312415)
Academic Editor: Gianluca Lax
Received: 6 October 2022
Accepted: 26 November 2022
Published: 4 December 2022
**Publisher’s Note: MDPI stays neutral**
with regard to jurisdictional claims in
published maps and institutional affil
iations.
**Copyright:** © 2022 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
[Attribution (CC BY) license (https://](https://creativecommons.org/licenses/by/4.0/)
[creativecommons.org/licenses/by/](https://creativecommons.org/licenses/by/4.0/)
4.0/).
**Abstract: An identity management system (IDMS) manages and organizes identities and credentials**
information exchanged between users, identity providers (IDPs), and service providers (SPs) to
ensure confidentiality and enhance privacy of users’ personal data. Traditional or centralized IDMS
rely on a third party to store a user’s personal information, authenticate the user, and organize
the entire process. This clearly constitutes threats to the privacy of the user, in addition to other
issues, such as single point of failure (SPOF), user tracking, and data availability issues. Blockchain
technology has many useful features that can contribute to solving traditional IDMS issues, such
as decentralization, immutability, and anonymity. Blockchain represents an attractive solution for
many issues related to traditional IDMS, including privacy, third-party control, data leakage, and
SPOF, supported by Distributed Ledger Technology (DLT) security features and powerful smart
contracts technology. The current study presents a systematic literature review and analysis for
recently proposed solutions that adopt the traditional centralized approach, as well as solutions
based on blockchain technology. The study also aims to provide a deep understanding of proposed
IDMS solutions and best practices, and highlight the research gaps and open issues related to IDMSs
and users’ privacy. In particular, the current research focuses on analyzing the blockchain-based
solutions and illustrating their strengths and weaknesses, as well as highlighting the promising
blockchain technology framework that can be utilized to enhance privacy and solve security issues in
a centralized IDMS. Such a study is an important step towards developing efficient solutions that
address the pressing needs in the field.
**Keywords: identity management; blockchain; distributed ledger technology; self-sovereign identity;**
privacy
**1. Introduction**
Today, digital identities are essential for users on the internet to obtain services from
electronic service providers (SPs). Digital identity represents the user’s personality in the
digital world and carries their necessary data that allows the identity holder to access
various resources on the internet provided by SPs [1]. Managing and protecting the user’s
identity, as well as related transactions and data, are critical tasks that need to be considered.
The IDMS is an organizational process that aims to achieve these tasks and makes it easy
for authorized users to access required services through their digital identity credentials. In
addition, IDMS seeks to provide necessary security services, such as privacy, confidentiality,
and availability, to counter recently emerged cyberattacks and threats. There are three
general basic parties in IDMS: the identity provider (IDP), the SP (or relying party RP), and
the user [2]. The digital identity of the user is created by the IDP, as they are responsible for
creating the digital identity and certifying it for the SP; the user needs to obtain a service
from the SP, which provides the necessary authentication for the user. The SP provides
the user with various resources after verifying their identity through the IDP. An IDMS
becomes essential for modern applications and e-transactions to organize and manage
-----
_Appl. Sci. 2022, 12, 12415_ 2 of 20
identity information and credentials between the involved parties; the user, the SP, and
the IDP. Furthermore, the IDMS is required to control the process of user authorization
and support the role-based access system. The IDMS can be realized using centralized and
decentralized approaches.
A centralized IDM approach is the process of controlling and managing user identities
and their relations using other central parties: an IDP and an SP. It is based on two primary
operations, authentication and authorization, to provide an identity verification process and
to increase access control (AC) security. However, a centralized IDMS suffers from potential
risks that threaten users’ privacy and decrease system transparency because of its reliance
on centralization in controlling and managing users’ data. The major risks associated with
a centralized IDMS include issues related to user privacy, such as user behavior monitoring,
and third-party control, in addition to issues relevant to the availability of data, such as the
single point of failure (SPOF) [3].
The decentralized blockchain infrastructure is one of the most important proposed
solutions to solve the centralized IDMS issue approaches, as a result of its powerful security
features and promising technologies. The blockchain has multiple features that contribute to
improving the problems of the current central systems, such as the features of distribution,
peer-to-peer (P2P), immutability, and others. Two important concepts were launched in
2013 that served to transform IDMs from centralization to decentralization, Ethereum, and
the smart contract. In smart contracts, transactions between parties can be conducted and
tasks can be performed without the involvement of a third party, since it is a self-executing
program that runs whenever the conditions are met. There are many features of blockchain
technology that can enhance user privacy. Decentralization is the most important. In
addition, avoiding complete dependence on a central authority reduces the risk of a SPOF.
By using the blockchain, the user is protected from relying on third parties, and therefore,
the possibility of tracking and studying their behavior is eliminated. However, despite the
blockchain’s many advantages, it still faces some challenges, such as its scalability.
A study comparing the various solutions offered by this technology is important as
the blockchain offers multiple features that can help solve the problems associated with centralizing identity management. Several issues have been addressed in the current systems
in which blockchain technology has been applied, as well as addressing research that has
compared and uncovered the most suitable method of centralized identity management
using various types of blockchain.
This research presents a systematic literature review of recent studies that have proposed blockchain-based solutions for centralized IDMSs across different domains. The aim
of this study is to explore blockchain privacy and security solutions, study and compare
those solutions, and analyze the results to highlight the current research gaps and best practices. These efforts seek to develop efficient blockchain-based solutions for IDMSs which
represent an essential need for the current internet-based applications and businesses.
The remaining sections of this paper are as follows. Section 2: Background; Section 3:
Literature Review; Section 4: Method; Section 5: Result and Discussion; and finally, the
conclusion is outlined in Section 6.
**2. Background**
_2.1. Overview of IDMSs_
Digital identities are needed to identify users when they request access to digital
resources. To manage these digital identities, in addition to related information and
credentials, an efficient IDMS is required. There are many identity management models
that have been created and categorized based on the use of identity and the need for a
cross-domain, such as an isolated user identity model, a federated identity model, and a
user-centric model [1].
-----
_Appl. Sci. 2022, 12, 12415_ 3 of 20
2.1.1. The Isolated User Identity Model (SILO) or Centralized Model
IDMSs have undergone multiple stages of development. First, there was the Isolated
User Identity (SILO) model, which is the cornerstone and the most simple model most
widely used [4]. It is based on identity management between only two parties, the IDP and
the user. The IDP in this system plays the role of the SP, as it allows the user to create a
digital identity to obtain services provided in a specific field, which means that the user
needs to create several digital identities to obtain services in multiple domains [5]. This is
perhaps a major defect in this model owing to the difficulty of managing multiple identities
by the user, in addition to full dependence on the IDP, which may cause a violation of user
privacy, such as user movements tracking.
2.1.2. Federated Identity Model
Another IDM was then created, which is Federated IDMS [1]. It differs from the
previous system as it is based on three parties instead of two: the IDP, the SP, and the
user [6]. The IDP here is the responsible party for user identity creation, authentication,
and necessary credentials. In this model, the user depends on the IDP to issue credentials
related to their identity and authenticate them to the SP. Therefore, there must be an element
of trust between the IDP and the SP (Circle of Trust principle), which means that for every
IDP in the system, there is a group of trusted SPs that the user can obtain services from [2,4].
Full dependence on the IDP, in addition to being fully informed of all user behaviors
and relationships, is a threat to user privacy and may lead to the SPOF. These are serious
problems in the centralized identity management approach that depends on a central party
to provide the required identity creation and authentication services; the IDP.
2.1.3. User-Centric Model
This model is also referred to as the Open Trust Model, as all parties in the system
are required to trust each other [1]. In this model, the user can select the attributes and
credentials to be sent, in addition to the ability of choosing the IDP. It is very similar to the
federated model, and it also has the same privacy concerns. The second law of identity
(justifiable parties) is not satisfied in this model and the sharing policy with SP can be
defined by the user, but it is still under the control of the IDP [5]. User privacy is violated
in this model because of the IDP control.
2.1.4. Self-Sovereign Identity Model (SSI)
The abovementioned IDM model requires full dependence on a third party, the IDP,
to manage and control the identity, in addition to providing the credentials necessary for
authentication. This represents a clear threat to the user’s privacy, as all user behavior and
movements are exposed to the IDP. To raise the level of user privacy in the field of digital
identities, and to find a solution to the problems associated with the user’s dependence on
the IDP (problems related to the centralized approaches), a model based on the principle of
decentralization has appeared in the field of IDM. The adoption of a decentralized IDM
approach has been instigated by many researchers to find solutions regarding the privacy
and SPOF problems in the previous centralized models. The Self-sovereign Identity model
(SSI) is an emerging decentralized IDMS that provides the user with the ability to control
their identity, as well as its related data and transactions [7]. Unlike the three previously
mentioned models of online identity, centralized, federated, and user-centric, SSI provides
all three of the basic requirements, security, control, and portability. Therefore, the user
is both the controller and the manager of the identity, and there are no external central
control parties; reducing the hacking risk. During hacking, when the IDP obtains the
data of all users who trust it, the attacker needs to individually hack each user one by
one, which necessitates higher costs, more time, and more effort. To develop an efficient
decentralized IDM system capable of addressing problems related to privacy, SPOF, and
other security issues, an appropriate infrastructure must be made available. Distributed
Ledger Technology (DLT), also called blockchain, has been proposed by numerous research
-----
_Appl. Sci. 2022, 12, 12415_ 4 of 20
studies as an infrastructure by which to develop an IDM system and find effective solutions
to the issues of security, privacy, and SPOF, as well as to give users the freedom to manage
and exchange their data privately without the presence of or observation by controlling
parties [5].
_2.2. Blockchain_
Blockchain was invented in 2008 by an unknown entity who went under the pseudonym
Satoshi Nakamoto [8]. Blockchain technology is a technology that is built on several technologies, which include: blockchain data structure, public key infrastructure PKI, distributed ledger technology DLT, and a consensus mechanism [9]. Blockchain technology
has many characteristics that have contributed to its widespread adoption and significance
today, the most important of them being the decentralization feature. Using decentralization correctly is one of the most important steps towards solving the SPOF problem, which
poses one of the biggest challenges to centralized systems. There is also a significant impact
factor in the field of data protection associated with blockchain technology, since the data
stored cannot be deleted or modified once it has been stored on the blockchain [10,11].
Blockchain is one of the most important decentralized technologies. It has been widely
spread in the recent years and has been used in many domains, such as IOT [12–16]; supply
chain [17–20]; AC and Identity Management in [21–26], cloud IDM in [27], ad-hoc network
(VANET) in [28–30], healthcare in [31–33], internet of connected vehicles in [34,35], and
even for the undirected graph authentication, as discussed in [36]. Blockchain is a type of
DLT which makes it very difficult to modify or hack any data and transactions stored on
the blockchain platform through a secure and tamper-proof way [5]. The main components
of blockchain technology are:
- A block: A block of data which has a 32-bit randomly generated number (nonce) and
cryptographic hash, which is like a fingerprint of the block data. The first block of the
chain is called the Genesis Block, and it does not contain a previous hash, because it
is the original and the first block on the chain, and thus it is the only block with this
feature [37].
- Miners: The blockchain technology requires miners to solve complex math algorithms
to generate the cryptographic hash from the random nonce for each block created.
- Nodes: The nodes can be any electronic device holding all of the blockchain transactions copies.
- Chain: Group of blocks.
- Consensus protocol: Operations implementation rules.
The blockchain distributes the data blocks over multiple nodes on the internet [2].
Therefore, it is working to publish and transmit data in the form of multiple blocks linked
together. Each of the blocks contains the hash of the previous block, and that is why it is
called a chain of blocks (blockchain) because all the blocks are cryptographically linked to
each other through the hash, so if anyone tries to tamper with one of the blocks, the hash
of the block will no longer match up and the chain of blocks will be invalid, which is an
immutable ledger feature. Blockchain features such as decentralization, immutability, and
individual control of data, help to solve the most important issues of centralized IDMs by
giving the user full control of their data to increase privacy by limiting third-party control,
which is the main shortcoming of centralized IDM systems. The security and transparency
features avoid the central authority issue while no single entity owns the data. Another
important feature is that the blocks on a blockchain cannot be modified, and that is a very
important feature in the field of security as it has a major role in reducing attacks [38].
A Distributed P2P Network is one blockchain feature where each device in the network
is connected to all the other devices in the same network, and each device has a copy of
the blockchain. Therefore, with each new block created in the chain, a copy of the block
will be sent to all the peers under a cryptographic role. This is a very important security
feature where any system errors or tampering of any block will be detected because the
blockchain constantly checks all its peers to make sure that there are no issues. If any of the
-----
will be sent to all the peers under a cryptographic role. This is a very important security
_Appl. Sci. 2022, 12, 12415_ feature where any system errors or tampering of any block will be detected because the 5 of 20
blockchain constantly checks all its peers to make sure that there are no issues. If any of
the peers has a tampered block, the majority of the peers will compare the block and re
peers has a tampered block, the majority of the peers will compare the block and replaceplace the tampered block with the original one. As a result of this feature, it is difficult to
the tampered block with the original one. As a result of this feature, it is difficult to hackhack the block, since the hacker would have to tamper with more than 50% of the blocks
the block, since the hacker would have to tamper with more than 50% of the blocks at theat the same time in order to succeed [39,40]. In addition to the security features provided
same time in order to succeed [by blockchain technology, it eliminates the need for a third party to process transactions, 39,40]. In addition to the security features provided by
blockchain technology, it eliminates the need for a third party to process transactions, andand hence, supports decentralization via the use of smart contracts technology. A smart
hence, supports decentralization via the use of smart contracts technology. A smart contractcontract is a conditional transaction process in the blockchain that occurs when the condiis a conditional transaction process in the blockchain that occurs when the condition is mettion is met (a self-executed program). Smart contracts provide many advantages, such as
(a self-executed program). Smart contracts provide many advantages, such as increasingincreasing performance, saving time, and, most importantly, increasing privacy comperformance, saving time, and, most importantly, increasing privacy compared to otherpared to other traditional methods [41]. Smart contracts are run on many blockchain plattraditional methods [forms such as Hyperledger Fabric, Waves, Ethereum, and NEO. 41]. Smart contracts are run on many blockchain platforms such as
Hyperledger Fabric, Waves, Ethereum, and NEO.
Many IDM solutions have been designed without using DLT. As a result, there have
Many IDM solutions have been designed without using DLT. As a result, there have
been some issues related to central authority or third-party control, as in [3,6,42–44]. On
been some issues related to central authority or third-party control, as in [3,6,42–44]. On
the other hand, some research attempts have proposed solutions based on blockchain
the other hand, some research attempts have proposed solutions based on blockchain
technology. However, proposed blockchain-based IDM systems have certain issues re
technology. However, proposed blockchain-based IDM systems have certain issues related
lated to centralization when a private blockchain is used [13]; these pertain to private BC,
to centralization when a private blockchain is used [13]; these pertain to private BC, central
central authority in [45], data availability in [46], and key management issues in
authority in [45], data availability in [46], and key management issues in [47].There are
[47].There are many challenges in the field of user privacy in central identity manage
many challenges in the field of user privacy in central identity management, such as relying
ment, such as relying on the third party to create, verify, and authenticate the identity and
on the third party to create, verify, and authenticate the identity and its attributes, in
its attributes, in addition to the increased risk of user tracking, because the user needs the
addition to the increased risk of user tracking, because the user needs the third party every
third party every time they want to obtain a service from the service provider. The SPOF
time they want to obtain a service from the service provider. The SPOF is also one of the
is also one of the most important challenges facing central identity management.
most important challenges facing central identity management.
Integrating blockchain with identity management has many promising features that
Integrating blockchain with identity management has many promising features that
may help in solving and improving the system quality and user privacy. Decentralization,
may help in solving and improving the system quality and user privacy. Decentralization,
transparency, and immutability are among the most important characteristics that sup
transparency, and immutability are among the most important characteristics that support
this improvement, but there are also challenges that still need to be addressed, such asport this improvement, but there are also challenges that still need to be addressed, such
scalability of the blockchain system.as scalability of the blockchain system.
**3. Literature Review3. Literature Review**
The current paper aims to present a comprehensive discussion and review for both
traditional IDM systems that adopt the centralized approach, and the blockchain-based
IDMSs that rely on the decentralized DLT to improve privacy and achieve self-sovereign
identity concepts.
_3.1. Traditional IDMSs_
In [3], a study concerning Digital Identity and IDM Technologies, the author illustrated
a variety of technologies used in the field of IDM. Among the several competing standards
in the IDM field, the security assertion markup language (SAML) was the only applicable
-----
_Appl. Sci. 2022, 12, 12415_ 6 of 20
choice, as it had a high level of acceptance at that time. This is because it was part of the
solution to the problem of single sign-on. Later, another technology emerged and received
some attention in the community, called the WS-Federation. As users need to have multiple
identities for different service providers, the multiple identities used can cause a degree of
inconvenience to the user in terms of managing them. The author concluded that both are
similar in functionality but had different names: IDP and the service provider in SAML;
security token service and relying party in WS-Federation.
Microsoft CardSpace is a claim-based IDM system proposed by Microsoft to satisfy
the seven laws of identity. It gives the user the right to control their digital identities and
choose the card after they have completed the SP policy through the identity selector. The
identity selector is the intermediary between the user, the IDP, and the SP, as they retrieve
the security policy after the user picks the card and completes the user authentication with
the IDP on behalf of the user, and then forwards the security token to the SP to log the
user in after they have received it from the IDP. The system guarantees the integrity of
security tokens through an xml-signature and preserves the confidentiality of the IDP and
SP security policies by making transactions over an SSL/TLS channel. However, this model
violates user privacy, as it requires presenting the user credentials to the identity selector.
Another drawback for this model is that the user must carry out the authentication step
every time before a token is issued [42].
Another research study, this time conducted by the Liberty Alliance project, was a
single sign-on federated IDMS proposed in 2001. The project proposed several frameworks: the identity federation framework (ID-FF), the identity web services framework
(ID-WSF), the identity service interface specification (ID-SIS), the Liberty identity assurance
framework (LIAF), and the identity governance framework (IGF). The authentication and
authorization frameworks were separated in the system. The user in the Liberty Alliance
system was monitored by the IDP, as they knew who all the services providers were
accessed by the user, which violated user privacy [6].
In [48], researchers introduced Shibboleth, which is a Federated IDMS, and its single
sign-on framework, but it does not support single sign-off. The proposed system tries to
increase user privacy by using a short-term, random ID to maintain anonymity. Unlike
the previous project, the authentication and authorization frameworks can be combined.
In Shibboleth, IDP discovery is performed by the SP using the WAYF technique, which
can increase the risks to the user by connecting with a fake IDP, redirecting them via a
malicious SP. This also increased the risk of stolen credentials.
The OpenID system is an open-source IDMS, released in 2005. It supports SSO and
uses the concept of a global identifier to enable the user to contact any OpenID-enabled SP.
The system does not use any proof of rightful possession, which makes it vulnerable to the
risk of credential theft. In addition, it may create other risks such as directing the user to a
fake IDP via a malicious SP, and the risk of a man-in-the-middle (MITM) attack [44].
Reference [43] suggested two proposed solutions in the implementation layer to
improve the level of authentication with the user in a claim-based IDMS. A proof-ofauthenticity method and challenge-response method appeared as suggested solutions to
solve the problem of the malicious IDP, which may cause considerable damage to the SP
and the user. The authors suggested a proof-of-authenticity method as the first solution,
which uses an additional authentication layer through creating a random secret value by
the SP, and then sends it to the user (known only to the user and the SP) after each complete
authentication. The challenge-response method is the second proposed solution where the
user has to accept a challenge sent by the SP, and they must respond with the expected
result computed by using a private signature key or shared secret key between the SP and
the user. Both proposed solutions had a positive impact on solving the problem studied by
the authors, where, in addition to enhancing the user authentication, they also increased
the level of privacy in the claim-based IDM system.
The previously reviewed studies had many features that improve the quality and
performance of the system, but they also had many challenges that violate user privacy,
-----
_Appl. Sci. 2022, 12, 12415_ 7 of 20
such as data disclosure [42], user monitoring and increasing the risk of credentials being
stolen [6], a man-in-the-middle attack, and fake parties [44]. These were in addition to the
SPOF, which is one of the main issues associated with the centralized IDM approach.
The next section presents state-of-the-art studies that adopted a decentralized approach
for IDMS using the blockchain technology.
_3.2. Blockchain-Based IDMSs_
In [13], the authors presented a new IDM approach based on a private blockchain,
which aims to provide an efficient and simple protocol that meets all the needs of Internet
of Things (IOT) organizations. Researchers implemented a Hyper-ledger Fabric for the
smart homes model and wrote the chain codes using Golang language. The main functions
of the IDM systems are split into three phases to allow simultaneous execution: identity
registration, identity verification, and identity revocation; the three phases employed smart
contracts to interact with the blockchain. The author discussed how this approach would
enhance IOT entities communications by including a consortium membership service and
identity management protocol. The author chose to use a private blockchain in the model
to achieve more security and better scalability; however, in terms of characteristics, it was
more like centralization than decentralization, and that increased the risk of SPOF and
central authority issues.
The authors in [49] developed a decentralized IDM system prototype using the Hyperledger Indy blockchain as a proof-of-concept in the public transportation sector, based
on self-sovereign identity principles. The proposed system can reduce the need for using
multiple travel cards for the people who travel frequently and who use several modes of
transportation within multiple jurisdictions. The system aims to give the users full identity
control by creating a direct identity layer based on the principles of decentralization using
a blockchain-based IDM system to provide a Single European Transport for users. The
proposed system will provide the ability to create many decentralized identifiers for any
person, in addition to creating a key pair for each user so they can securely share the data.
In [45], researchers proposed a blockchain-based decentralized IDM system for the
public sector in South Korea by providing a mobile application by which to create electronic
identity cards, issued and managed by a national central authority. The user stores their
driver licenses on their device and verifies their identity through the app by using a onetime QR code. The client server in the system is developed by using Hyper-ledger Fabric
V1.0 to increase the privacy level. Amazon web service (AWS) is used in the system to
provide a faster process and increase efficiency. Data for any identity in the system is linked
to a central government agency in South Korea to complete the identification process.
User data is stored in a database in the form of keys and values paired on a hash map,
in addition to the chain code. The developer also used a modern user interface to make
users feel more comfortable using the system. The application is very effective in using
blockchain, but it appears to be centralized, even with blockchain, as the national central
authority is the data manager, and license requirement in the verification process might
be a disadvantage because such an application is not appropriate for many e-commerce
systems or for obtaining online services as there will be licenses or other types of formal
document involvement.
Authors of [46] used a smart contract to design a cross-domain self-sovereign identity
management system. The system contains three types of smart contracts; each one built to
perform a specific function. The services smart contract SSC is the first contract and the
basis contract in the system which controls the publishing of a user identity contract, and it
is created and published when the SP joins the system. The second is the identity smart
contract ISC, which is requested by the user from the SP after they have been identified and
verified, and their address is recorded in the SSC. The ISC is controlled by the user after
it is published. The Recovery Smart Contract (RSC) is also created at the same time. The
RSC is automatically created for each ISC to give the user the ability to recover their lost
password from a list of friends. The system, as proposed by the designer, performs better
-----
_Appl. Sci. 2022, 12, 12415_ 8 of 20
compared to three other systems using the same concept, but it also has a limitation in that
it uses the address of the ISC as a universal unique identifier UUID, which is not readable
by users, and, as the system stores the full attributes information in the user device, that
will decrease the availability of information when the user is offline.
In the study presented in [47], a hybrid methodology was proposed as a part of the
Impilo project for data management in healthcare by combining a central database and
decentralized infrastructure “blockchain”. The new approach tries to create ownership and
management of data on the patient side to increase security of electronic health records
and keep it shareable at the same time. Patient information is stored on a central database
during the validation process, and the transaction is stored on the blockchain. The system
operation begins by logging into the Impilo app and storing the registration information
in a new file, and then communicating with the DB to store the medical information.
The blockchain will generate a new hash, communicate with both sides, and then store
the transaction details on the chain if the verification process is correctly completed. In
this approach, the decryption key of medical information in a database is the user login
password; so, if an attacker knows the user login password, they will have access to all the
user information, and this decreases the security of the database.
In [50], researchers proposed a framework to solve the centralized problem of access
control and its related privacy and ethical issues, and to give users full control of their
IOT devices. The proposed framework is based on two main concepts: a blockchain and a
machine learning algorithm. The researchers addressed two problems in IoT environment
access control: centralized access control (AC) and security policy management. The
proposed framework distributes the security policy (a set of guidelines and security rules)
in the blockchain by using a smart contract instead of storing it in a server, as in a traditional
AC, and improves it by using an online learning mechanism of machine learning algorithms
to solve the problem of a non-contextual security policy. An online learning machine type
is used to detect any AC rules which do not satisfy the security policy, or which may lead
to any security threat.
Authors in [36] used the private Ethereum network to design a cryptographic authentication scheme. The authors developed a smart contract and published it on a private chain,
and then evaluated the scheme’s functions by using web3j and a proof of security model.
The research introduced a transitively closed undirected graph authentication (TCUGA)
scheme to update the certificates by the signatory with no re-signing process needed by
using a trapdoor hash function and allowing the administrator to prove the certificate
relationships “even when they are not in the same equivalence class” after they are received
from the signatory.
A permissioned blockchain-based IDM user authentication scheme was introduced
in [33] to solve key management and authentication issues in e-health systems by using
a key distributed mechanism of personal biometrics. The proposed system contains four
main members: the founder, the user (U), the registration center (RC), and the medical
server (MS), in addition to the smart contract that provides access control functions. It has
two major mathematical problems: the computational Diffie-Hellman problem (CDHP)
and the discrete logarithm problem (DLP). The proposed scheme is provided with a mutual
authentication equation and achieves anonymity by making the user’s identity hidden. The
designer tested the proposed system and guaranteed the security requirements by using
the Scyther tool, which is an automatic verification tool for security protocols.
An attempt to solve traditional banking issues by developing a blockchain-based IDM
and access control (BIMAC) framework was presented in [51]. The researchers used an
MVC (Model-View-Controller) structure for this purpose. The implemented framework
improved user experience by creating a user login to many bank accounts without the
need to remember all their accounts and passwords. The prototype applied the concept of
self-sovereign identity in the open banking field and provided an efficient authentication
framework.
-----
_Appl. Sci. 2022, 12, 12415_ 9 of 20
In [28], the authors tried to solve the problem of traffic disruption caused by malicious
vehicles through incorrect information propagation. As a way to maintain privacy, they
suggested using a blockchain-based authentication scheme and asymmetric key encryption
to secure vehicle communication. Additionally, elliptic curve cryptography was used to
increase transactions pseudonymity. According to the study of [34], it has been found
that when cooperating with unauthorized vehicles, it is possible to steal information,
compromise privacy, and exploit a variety of threats in terms of security. The authors
proposed a blockchain-based Internet of Vehicles (IoV) protocol that was developed on
the Ethereum platform, to improve the privacy of vehicle data and relationships with
the help of blockchain technology. However, too much IoV information stored in the
blockchain will affect the system’s scalability. [35] In addition, the paper discussed the
increased difficulty of managing certificates for vehicular communications, along with the
cost of anonymizing vehicle identities. This study proposes a blockchain-based pseudonym
management solution which has the ability to reuse existing pseudonyms in order to
simplify pseudonym management. Additionally, in [30], the authors attempted to enhance
vehicle privacy and trust relationships. As a result of the use of blockchain technology by
these authors, they proposed a blockchain-based anonymous reputation system (BARS),
which is based on a reputation evaluation algorithm.
A proof-of-concept IoT identity management system for a business case scenario was
implemented by the authors in [12], to ensure the integrity of the data provenance records
in the organization-networked IOT resources using blockchain and smart contracts. Solidity
language is used to code the proposed blockchain model and it is deployed in Kaleido.
The authors of [21] proposed a Hyperledger fabric blockchain system to enhance
Modbus, one of the Industrial Internet of Things IIoT protocols that faces many security
challenges, such as SPOFs. On-chain authentication and authorization are supported by
the designed decentralized identity system. By providing both security and scalability for
Modbus connections, it can be used in a system with more than one organization.
Self-sovereign Identity, blockchain, and Inter Planetary File technologies were used
by [17] to improve food supply chains. By using SSI concepts, the study proposed a way to
manage certifications throughout the supply chain. A certificate is issued by a certifying
body and stored in IPFS, with only some key information being stored on the chain; verifiers
need this information to verify whether a certificate is valid in the chain. To improve supply
chain security, the authors in [18] also implemented a Hyperledger Fabric framework
to ensure each registered device in the supply chain is tracked and to improve system
security. Furthermore, reference [19] proposed a supply chain traceability system, though
this proposed system tracks and validates both sides of the transaction. Additionally,
reference [20] used a permissioned blockchain network in order to take advantage of
smart contract features and to increase supply chain management security. The proposed
framework provides the user with control over the data and increases identity protection
by using cryptographic proof.
In recent years, telehealth has become a necessity, especially since the COVID-19 pandemic started. In [31], the authors addressed the problem of trusting e-health application
service providers and not knowing whether they comply with regulations to ensure privacy
and security. Blockchain technology was used to provide authentication and identification
processes to users and service providers across a variety of health domains. A smart
contract was implemented in the proposed system using Ethereum.
In edge computing, the privacy and security of user data are two of the most important
factors that need to be considered. As discussed in [22], the authors used smart contracts as
a means of presenting the Access Management System by using blockchain technology.
In order to improve the Internet of Things HIoT privacy, the authors in [15] proposed
verifiable anonymous identity management systems (VAIM), through which they improved
blockchain identity management and enhanced the unlinkability of the system by using
zero-knowledge proof (ZKP) algorithms.
-----
_Appl. Sci. 2022, 12, 12415_ 10 of 20
User privacy has been affected by third-party dependencies in identity management
systems in a variety of fields, including the Internet of Things. An SPOF is also one of
the most important issues resulting from third-party control. Using Hyperledger Fabric,
the authors in [13] implemented a smart-home-based scenario architecture to improve the
quality and efficiency of home sensors and to enhance IoT centralization issues. A proposed
architecture would divide the functions of the system into three main parts: registration,
authorization, and revocation. The authors tried to improve the scalability of the system by
splitting the functions.
The authors in [32] attempted to solve the problem of electronic health records information being exposed, which poses a threat to the privacy of the users and those whose
records are accessed. The authors implemented a proof of concept through the use of
Hyperledger Fabric’s permissioned blockchain technology to ensure anonymity for the
EHR data and to enhance privacy for patients.
Using a DNS-like approach, the authors of [23] proposed a DNS-IDMs architecture
that is implemented on Ethereum’s permissioned ledger. In order to enhance the privacy
of the user, users and service providers would be able to create identity attribute claims
and verify them using the services of real-world identity attribute benefactors. By using
blockchain transactions, users can also control and manage their identities.
There are many security challenges associated with large-scale IoT systems due to
centralization concepts, such as unauthorized access requests to IoT-enabled devices, which
are an issue of access control. To make the system more flexible and adaptable, reference [14]
implemented a private blockchain POC prototype using Ethereum and smart contracts.
BlendCAC was the name of the framework proposed by the authors.
An Ethereum-based IDM cloud protocol was proposed by [27], an improved version of
CIDM (Consolidated Identity Management). The proposed protocol attempts to solve the
third-party reliance problem in traditional identity management systems. Smart contracts
were used in the proposed system to increase data transmission privacy and to enhance
system flexibility.
The authors of [24] provided a method that allows users to sign transactions using a
different Ethereum identity in order to enhance user untraceability by granting the user the
right to delete their data and allow them to discard their identity afterward. The proposed
method represents identity through web3js-based implementation and data erasing can be
requested by the user or an end of service.
It was proposed in [25] that attribute trust could be enhanced by using an Attribute
Trust-enhancing Identity Broker (ATIB) architecture in order to enhance the aggregation
of system attributes by following the ten SSI principles. As part of the proposed proof
of concept, the service providers role would be enhanced with the help of the protocol
manager, which is the main component in the proposed architecture that will be able to
support the implementation of many identities and access protocols to the system.
In [26], the authors proposed a method of integrating distributed identity provider
technology (OLYMPUS) with blockchain technology while utilizing smart contract technology as a means of evolution of distributed identity provider technology. It was proposed
that the proposed architecture will improve system security and enhance the privacy of
users.
As a result of a combination of a cryptographic authentication scheme and blockchain
technology, reference [36] proposed a transitively closed undirected graph authentication
scheme (TCUGA). The proposed scheme manages vertices and edges, and it can prove the
absence of any edge between two vertices.
A permissioned blockchain was used with attribute-based access control (ABAC) and
an identity-based signature (IBS) in order to improve the security of an Internet of Things
system [16]. In this paper, a cross-domain blockchain-based IoT access control system
was proposed to address some of the challenges related to IoT systems, such as SPOFs,
information leaks, and Distributed Denial of Service (DDoS).
-----
_Appl. Sci. 2022, 12, 12415_ 11 of 20
By adopting an existing technology, the authors in [33] enhanced E-health identity
authentication and solved some major security issues, including reply attack and an MITM
attack. In order to provide a secure mutual authentication and key distribution system, the
proposed authentication scheme is implemented in permissioned blockchains.
A fine-grained AC scheme was proposed in [29] to enhance Vehicular Ad Hoc Network
(VANET) data sharing. In order to increase data sharing security and decrease SPOFs, a
combination of blockchain technology, IPFS, and ciphertext-based attribute encryption
(CP-ABE) is proposed. A smart contract is also used in the proposed scheme in order to
increase the scalability of the systems.
In [52], a private blockchain was used to help the agricultural sector and farmers in
India to ensure that their communication with their customers can take place directly with
them without any intervention from third parties in the process. The proposed model was
built on Hyperledger Fabric to enable direct communication between the farmer and the
customer at the same time.
**4. Method**
To achieve the study’s key aim of exploring the use of a public blockchain platform to
integrate the principle of decentralization with IDMS, we conducted a systematic review following the Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA)
guidelines, which help in analyzing the steps of the systematic review by identifying specific and clear research questions, and following a specific methodology to obtain answers
through the use of a sample of research papers that are determined by of exclusion and
inclusion criteria [53]. For this purpose, we selected previous studies that use blockchain
technology on IDMs. Further elaboration on research and selection strategy explanations is
given below:
_4.1. Research Need Identification_
An objective of this systematic literature review is to examine how blockchain-based
systems can be used to enhance privacy, as well as improve a system by eliminating or
reducing centralization issues in trading systems, such as SPOF risks, central authority
issues, and third-party control risks.
_4.2. Research Questions_
Q1: What are the current issues that threaten user privacy and security in centralized
IDMSs?
Q2: Will decentralizing identity management by using distributed ledger technology solve
user privacy problems, and if so, why?
Q3: What are the blockchain-based technologies that may be utilized to enhance user
privacy?
Q4: What is the most efficient blockchain-based development platform for IDMSs?
_4.3. Information Source and Database_
We selected multiple databases for the information sources, as shown in Table 1. The
literature review was limited to research studies published between 2018 and 2022.
**Table 1. Information Source.**
**Database** **Website**
IEEE Xplore Digital Library [https://www.ieee.org](https://www.ieee.org)
MDPI [https://www.mdpi.com](https://www.mdpi.com)
_4.4. Research String_
The research strings are described in Table 2.
-----
_Appl. Sci. 2022, 12, 12415_ 12 of 20
**Table 2. Research String.**
**Open** **After Deleting**
**Database** **Keywords** **NO**
**Access** **Duplicate**
**After**
**Reading**
**Paper**
IEEE
MDPI
“Identity management
319 38 38 14
systems AND blockchain”
“Identity management
101 11 2 0
system AND smart contract”
“Ethereum AND identity
41 5 4 0
management system”
“Identity management
26 26 26 11
systems AND blockchain”
“Identity management
7 7 1 0
system AND smart contract “
“Ethereum AND identity
2 2 0 0
management system”
_4.5. Criteria Selection_
The study only included research written in the English language from 2018 until the
present day. In addition, surveys papers or systematic review papers were not considered.
Instead, papers that proposed systems were considered, as shown in Table 3.
_4.6. Inclusion and Exclusion Criteria_
We followed the PRISMA flow diagram in the study selection process, as shown in
Figure 1, and by following the inclusion and exclusion criteria of the current systematic
review described in Table 3, the authors extracted approximately 496 studies relevant to
blockchain-based IDM systems. Following the two main inclusion criteria, only 71 papers
fulfilled the research aims. After downloading and reading the abstracts, 46 more papers
were excluded during screening. Only 26 research articles were assessed and recognized
_Appl. Sci. 2022, 12, x FOR PEER REVIEW_ 13 of 21
against the research criteria. The current systematic followed the PRISMA standards for
data extraction and selection, as shown in Figure 1.
**Figure 1.Figure 1. PRISMA flow diagram of the study selection process.PRISMA flow diagram of the study selection process.**
-----
_Appl. Sci. 2022, 12, 12415_ 13 of 20
**Table 3. Criteria selection.**
**Inclusion Criteria** **Exclusion Criteria**
Written English Studies written in other languages
Studies From 2018 until now Studies before 2018
Original research paper Survey, systematic review papers
Proposed solution implemented Proposed solutions not implemented
**5. Results and Discussion**
In this section, the research review will be discussed, and the results are presented in
detail. The results are presented in multiple sub-sections according to the field to which
they belong.
_5.1. Study Characteristics_
The current systematic review focused on developing blockchain-based solutions for
privacy and security issues in IDMSs. To highlight important characteristics of the reviewed
studies, we designed Tables A1 and A2 for the two databases considered in this study.
Each table contains Title, Author with Year, Type, Publisher, the use of BC, and the use of
SC. Due to the role blockchain types play in solving existing research problems, the tables
indicate which type of blockchain was used in each research, in addition to the possibility
of using smart contracts.
_5.2. Discussion and Result_
In this section, we present the information collected from the research papers after the
systematic review. In Section 5.2.1, we review the domains in which blockchain technology
was adopted to enhance privacy and security of IDM, and then Section 5.2.2 discusses the
blockchain types and technologies that were applied to address different issues related to
privacy and security in order to highlight the best practices and efficient solutions, as well
as to provide an understanding of the potential solutions that can be offered by blockchain
technologies. Section 5.2.3 discusses the research and issues addressed via using smart
contracts technology, as it represents a cornerstone and powerful blockchain technology
that can effectively contribute to developing efficient solutions for problems relevant to the
privacy issue. Finally, in Section 5.2.4, the research questions are answered in detail.
5.2.1. Domain
The current systematic review surveyed the previously proposed IDMS solutions
that adopted a decentralized approach. Previous literature has illustrated that the use of
blockchain technology improved the security and privacy of IDMSs in many domains, such
as IOT [12–16]; supply chain [17–20]; AC and Identity Management in [21–26], cloud IDM
in [27], ad-hoc network (VANET) in [28–30], healthcare in [31–33], internet of connected
vehicles in [34,35], and even for the undirected graph authentication, as discussed in [36].
5.2.2. Issues and Blockchain Type
This section sheds light on the different blockchain types adopted in previous research
and the security issues addressed by each type. This assists in understanding the potential
solutions that can be addressed by particular blockchain types or technology.
The majority of the reviewed studies adopted access control and IDM to find solutions
for system issues by using the Ethereum blockchain type. In [27], the IDMS adopted by
cloud users relies too much on third-party services. Studies published in [24] and [18]
suffered from third-party issues, especially trackability, and both used Ethereum in their
solutions. In [20], authors used Ethereum-based IDM Protocol as a solution for the U.S.
beef cattle supply chain. By utilizing Ethereum blockchain, the authors in [19] provided
a solution for identifying the root cause of system problems. An Ethereum-based food
-----
_Appl. Sci. 2022, 12, 12415_ 14 of 20
supply chain system was proposed in [17]. Other studies have also used the Ethereum
blockchain type to improve their systems, such as [14,23,29,31,34,36].
Other types of blockchain have also been used in some of the studies reviewed. A
permissioned blockchain was used in [15] as a solution for the same third-party issue in a
different domain. Trust relationships between SPs, users, and IDPs in ABC systems have
many privacy concerns, and the authors in [26] tried to improve this by using Hyperledger
technology. The later blockchain type was used by [16] to solve three main issues: (1) single
failure point; (2) privacy information leak; (3) Distributed Denial of Service (DDoS) attack
of the delegate node. In addition, [30] preserved a vehicle’s identity privacy by using
blockchain to prevent fake message distribution. Communication and computational
overheads in healthcare systems were discussed by [33], using a permissioned blockchain
to improve them. The reviewed studies proposed solutions to enhance and improve
centralized systems by using blockchain technology in a different way, but there are still
open issues that need to be addressed and enhanced, such as enhancing the scalability of
blockchain-based IDMS platforms, system usability, and privacy enhancement.
5.2.3. Smart Contract
Smart contracts are a very important concept in the field of blockchains. They provide
many important features to enhance system functionality and to increase the speed of
operations. In the current review, only seven research papers did not use smart contracts
in their proposed solutions: [15,21,25,28,30,32,35]. On the other hand, 18 research papers
adopted smart contracts to provide more efficient solutions for the privacy problems in
IDMS: [12–14,16–20,22–24,26,27,29,31,33,34,36].
The analysis of statistics related to the previous research shows that there has been an
increase in the number of publications over recent years that adopted blockchain technology
in the field of IDMS, as depicted in Figure 2. In terms of the blockchain type, the analysis
results presented in Figure 3 show that Ethereum has been more frequently used than the
other types of blockchain. There are several reasons for this. The smart contract is one of
the most important components of an Ethereum system’s development and improvement.
_Appl. Sci. 2022, 12, x FOR PEER REVIEW_ 15 of 2
The Solidity Language is another important reason, along with the fact that Ethereum is
involved in several applications, the most important of which is the DApp.
12
10
8
6
4
2
0
#### YE A R A N D N UM B E R O F P U BLICA TION S
2018 2019 2020 2021 2022
**Figure 2.Figure 2. Years and number of publications.Years and number of publications.**
There has been a significant increase in identity control on the proposed blockchainbased systems because of the third-party limitations caused by the decentralization feature.
The system is powerful and operates faster when it is using smart contracts as they are
# BC Types
self-executed codes, but there is some uncertainty about the security of the stored data. As
The system is powerful and operates faster when it is using smart contracts as they are
# BC Types
self-executed codes, but there is some uncertainty about the security of the stored data. As
-----
_Appl. Sci. 2022, 12, 12415_ 2 15 of 20
0
2018 2019 2020 2021 2022
a result, there have been many research papers on identity management systems that are
trying to reduce the different risks and to mitigate cyberattacks encountered in this field.Figure 2. Years and number of publications.
# BC Types
**24%**
**36%**
**40%**
Blockchain Ethereum Hyperledger Fabric
**Figure 3.Figure 3. Blockchain Types.Blockchain Types.**
# BC Types
**24%**
**36%**
**40%**
Blockchain Ethereum Hyperledger Fabric
It can be seen from the research articles shown in Tables A1 and A2 that blockchain
#### There has been a significant increase in identity control on the proposed blockchain
technology, the underlying technology for decentralized IDMSs, has been proposed as an
#### based systems because of the third-party limitations caused by the decentralization fea
effective solution for privacy and security issues in a variety of fields, such as IOT, supply
#### ture. The system is powerful and operates faster when it is using smart contracts as they
chains, ad-hoc networks, cloud IDM, healthcare, internet of connected vehicles, and access
#### are self-executed codes, but there is some uncertainty about the security of the stored data
control. Previous research has illustrated that blockchain is a powerful technology and has
many features that may effectively contribute to enhancing user privacy and increasing theAs a result, there have been many research papers on identity management systems tha
level of self-control over personal data in the field of IDM and relevant applications.are trying to reduce the different risks and to mitigate cyberattacks encountered in thi
#### field.
5.2.4. Research Questions and Answers
#### It can be seen from the research articles shown in Tables A1 and A2 that blockchain
Q1: What are the current issues that threaten user privacy and security in centralizedtechnology, the underlying technology for decentralized IDMSs, has been proposed as an
IDMSs?
#### effective solution for privacy and security issues in a variety of fields, such as IOT, supply chains, ad-hoc networks, cloud IDM, healthcare, internet of connected vehicles, and accesCentral identity management systems suffer from certain privacy issues, as discussed
in the previous section. One of the most important problems is centralization, since it reliescontrol. Previous research has illustrated that blockchain is a powerful technology and
upon one central party, which results in the high risk of an SPOF. Third-party control is
#### has many features that may effectively contribute to enhancing user privacy and increas
considered one of the most important threats in centralized systems, since the user is under
#### ing the level of self-control over personal data in the field of IDM and relevant applica
the control of a third party, which can compromise their privacy, such as monitoring their
#### tions.
movements and studying their behavior.
Q2: Will decentralizing identity management by using distributed-ledger technology solve
#### 5.2.4. Research Questions and Answers
user privacy problems, and if so, why?
#### Q1: What are the current issues that threaten user privacy and security in centralized ID
Decentralization of identity management by using distributed ledger technology
#### MSs?
addresses the problem of a SPOF because copies of the system are distributed over multiple
peers. As the peers constantly compare and verify the validity of the copies, when oneCentral identity management systems suffer from certain privacy issues, as discussed
fails, the rest discover the error and recopy the system in the correct chain. Furthermore,in the previous section. One of the most important problems is centralization, since it relie
technology provides the smart contract, which plays a major role in limiting the control ofupon one central party, which results in the high risk of an SPOF. Third-party control i
third parties, as tasks are assigned to the smart contract, and the tasks are automatically
executed without the intervention of any third parties.
Q3: What are the blockchain-based technologies that may be utilized to enhance user
privacy?
Using the smart contract as an intermediary to carry out tasks between the parties
enhances the privacy of the parties, since, for example, users can send tokens through the
smart contract to a service provider, whose tokens have attributes certified by third parties.
Since a smart contract acts as an intermediary, third parties and service providers cannot
Blockchain Ethereum Hyperledger Fabric
-----
_Appl. Sci. 2022, 12, 12415_ 16 of 20
track user relations or actions. Additionally, the user can control how much data is shown
in each token created for a service provider through a smart contract. The smart contract
can also be used to track all the viewers of the token data by recording their addresses and
the time they viewed it. So, yes, this technology enhances user privacy.
Q4: What is the most efficient blockchain-based development platform for IDMSs?
As a result of the research, most of the applications used the public blockchain
(Ethereum) because it is open source and has smart contract technology. Furthermore,
Ethereum works with a special currency called Ether, and has a special programming
language called Solidity.
**6. Conclusions**
In the domain of IDM, the adaptation of distributed ledger technology has attracted
attention due to its ability to enhance user privacy and address issues, such as the SPOF
and third-party control. The current work reviewed recent research papers in the area of
identity management systems; both traditional and those which have adopted blockchain
technology. Many articles covering IDM and blockchain technologies were reviewed in
this research. Many reviewed research attempts to provide the user with increased identity
control by trying to solve third-party control issues, address the SPOF, and avoid fake
message distribution. Furthermore, the review of previous research about IDMS showed
that there are still open issues relating to user privacy in the traditional centralized IDMSs,
including third-party control and user movement monitoring or tracking, in addition to
the problem of the SPOF. This prompted the need to search for an efficient solution to
enhance user privacy in IDMSs and avoid other problems associated with the decentralized
approach. Decentralized IDM by using blockchain has many advantages, including solving
the problem of third-party control by giving each user full control of their private information and activities, improving performance, and saving time by using smart contracts
and other blockchain features. In addition, the use of blockchain-based IDMS can avoid
the SPOF and ensure that data and services are available to legitimate parties once needed.
However, blockchain-based solutions that use a private type have some weaknesses related
to privacy, and they inherit certain problems from the centralized approach. In addition,
the use of weak authentication methods is a significant issue that needs to be addressed in
recently proposed block-chain-based IDMSs.
The systematic literature review presented in this paper discussed and analyzed
the recent solutions and current challenges in the field of IDM, while concentrating on
the contributions made by using blockchain technology. This aims to provide a better
understanding of the role and significance of adopting blockchain technologies in the field
of IDM and the advances that can be achieved using this powerful technology. Moreover,
the current review attempts to identify the research gaps and open issues, and motivate
future research works that may utilize the promising features of blockchain in improving
user privacy and addressing other challenges in the field of IDM.
As part of our future work, we intend to implement a system prototype for a decentralized identity management system utilizing the Ethereum blockchain to solve the problems
identified in this research and assess its advantages and disadvantages.
**Author Contributions: Writing—original draft preparation, H.A.; writing—review and editing, H.A.,**
M.A.; supervision, M.A. All authors have read and agreed to the published version of the manuscript.
**Funding: The authors extend their appreciation to the Deanship of Scientific Research at IMSIU for**
founding and supporting this work through the Graduate Student Research Support Program.
**Institutional Review Board Statement: Not applicable.**
**Informed Consent Statement: Not applicable.**
**Data Availability Statement: The literature review was limited to research studies published in IEEE**
[and MDPI databases. https://www.ieee.org. whttps://ww.mdpi.com.](https://www.ieee.org)
-----
_Appl. Sci. 2022, 12, 12415_ 17 of 20
**Acknowledgments: Authors acknowledge the support from Imam Mohammad ibn Saud Islamic**
University (IMISIU) for this research. The authors extend their appreciation to the Deanship of
Scientific Research at IMSIU for founding and supporting this work through the Graduate Student
Research Support Program.
**Conflicts of Interest: The authors declare no conflict of interes.**
**Appendix A Included Studies**
**Table A1. MDPI Included studies.**
**Smart**
**Study NO** **Title** **Authors** **Year** **Type** **Publisher** **BC Used and Filed**
**Contract**
[28]
EBAS: An Efficient
Blockchain-based
Authentication Scheme for
Secure Communication in
Vehicular Ad Hoc Network
Developing an IoT Identity
[12] Management System Using
Blockchain
[21]
Modbus Access Control
System Based on SSI over
Hyperledger Fabric
Blockchain
Blockchain and Self Sovereign
[17] Identity to Support Quality in
the Food Supply Chain
Blockchain, r secure
Xia Feng et al. 2022 article MDPI communication in no
VANET
Sitalakshmi
2022 article MDPI Blockchain, IOT Yes
Venkatraman et al.
Santiago Hyperledger fabric
Figueroa-Lorenzo 2021 article MDPI blockchain, Modbus no
et al. access control.
Luisanna Ethereum Blockchain,
2021 article MDPI yes
Cocco et al. Food Supply chain
Ethereum
Ibrahim Tariq
2021 article MDPI consortium yes
Javed et al.
blockchain, e health
[31]
Health-ID: A
Blockchain-Based
Decentralized Identity
Management for Remote
Healthcare
Blockchain-Enabled Access
Blockchain, edge
[22] Management System for Edge Yong Zhu et al 2021 article MDPI yes
computing
Computing
[34]
ABlockchain-based
Authentiaction Protocol For
Cooperative Vehicular Ad
Hoc Network
Ethereum blockchain,
A. F. M. Suaib
2021 article MDPI internet of Vehicles yes
Akhter et al.
(IoV)
Alightweight Blockchain
[13] based IOT Identity
Managemnt Approach
Aprivacy-preserving
[32] Healthcare Framework Using
Hyperledger Fabric
consortium
blockchain-based
identity management,
IoT(implement by
Hyperledger Fabric)
Hyperledger Fabric’s
permissioned
blockchain
framework,
healthcare
private Ethereum
network
(permissioned
Ethereum ledger)
Mohammed
2021 article MDPI
Amine Bouras et al.
Charalampos
2020 article MDPI
Stamatellis et al.
Jamila Alsayed
2019 article MDPI
Kassem et al.
yes
no
yes
[23]
[14]
DNS-IDM: A Blockchain
Identity Management System
to Secure Personal Data
Sharing in A Network
BlendCAC: ASmart
Contract-Enabled
Decentralized
Capability-Based Access
Control Mechanism For
The IOT
private Ethereum
Ronghua Xu
2018 article MDPI blockchain, AC in IoT yes
et al.
devices.
-----
_Appl. Sci. 2022, 12, 12415_ 18 of 20
**Table A2. IEEE Included studies.**
**Smart**
**Study NO** **Title** **Authors** **Year** **Type** **Publisher** **BC Used and Filed**
**Contract**
EIDM: A Ethereum-Based Cloud
shangping Ethereum blockchain,
[27] User Identity Management 2019 article IEEE yes
wang et al. cloud IDM
Protocol
Burnable Pseudo-Identity: A iván gutiérrez
Ethereum, Anonymous
[24] Non-binding Anonymous agüero 2021 article IEEE yes
Identity
Identity Method for Ethereum et al.
Pseudonym Management
Through Blockchain:
shihan bao Blockchain, internet of
[35] Cost-efficient Privacy 2019 article IEEE no
et al. connected vehicles.
Preservation on Intelligent
Transportation Systems
VAIM: Verifiable Anonymous permissioned
Identity Management for gyeongjin ra blockchain, the human
[15] 2021 article IEEE no
Human-centric Security and et al. internet of things
Privacy in the Internet of Things (HIoT)
ATIB: Design and Evaluation of
an Architecture for Brokered
Blockchain,
Self-Sovereign Identity andreas grüner
[25] 2021 article IEEE IDM(attributes no
Integration and Trust-Enhancing et al.
aggregations.)
Attribute Aggregation for the
Service Provider
A Trusted Approach for
Decentralized and rafael torres Hyperledger fabric,
[26] 2021 article IEEE yes
Privacy-Preserving Identity moreno et al. IDM
Management
A New Transitively Closed
Undirected Graph Authentication Ethereum, undirected
[36] chao lin1 et al. 2018 article IEEE yes
Scheme for Blockchain-based graph.
Identity Management Systems
Blockchain-Based IoT Access Hyperledger fabric
shuang sun
[16] Control System: Towards Security, 2021 article IEEE permissioned yes
et al.
Lightweight, and Cross-Domain blockchain, IOT AC.
A Permissioned Blockchain-based
permissioned
Identity Management and User xinyin xiang
[33] 2020 article IEEE blockchain, e-health yes
Authentication Scheme for et al.
systems
E-Health Systems
FADB: A Fine-grained Access Ethereum, Vehicular
[29] Control Scheme for VANET Data hui li et al. 2020 article IEEE Ad Hoc Network yes
Based on Blockchain (VANET)
Hyperledger fabric
A Blockchain-based Framework pinchen cui permissioned
[18] 2019 article IEEE yes
for Supply Chain Provenance et al. blockchain, Supply
Chain
Smart Contract-based Product
shangping Ethereum, Supply
[19] Traceability System in the Supply 2019 article IEEE yes
wang et al. Chain
Chain Scenario
Blockchain, vehicular
A Privacy-Preserving Trust Model zhaojun lu
[30] 2018 article IEEE ad hoc networks no
Based on Blockchain for VANETs et al.
(VANETs)
permissioned
A Permissioned Distributed
tanvir ferdousi blockchain network,
[20] Ledger for the US Beef Cattle 2020 article IEEE yes
et al. Ethereum Supply
Supply Chain
Chain
**References**
1. L’Amrani, H.; Berroukech, B.; Ajhoun, R.; El Idrissi, Y. Identity Management Systems: Laws of Identity for Models[′] Evaluation. In
Proceedings of the 2016 4th IEEE International Colloquium on Information Science and Technology (CiSt), Tangier, Morocco,
24–26 October 2016.
2. Liu, Y.; He, D.; Obaidat, M.; Kumar, N.; Khan, M.; Choo, K. Blockchain-based identity management systems: A review. J. Netw.
_[Comput. Appl. 2020, 166, 102731. [CrossRef]](http://doi.org/10.1016/j.jnca.2020.102731)_
3. Agudo, I. Digital Identity and Identity Management Technologies. Serb. Publ. InfoReview Joins UPENET Netw. CEPIS Soc. J. Mag.
**2010, 6.**
4. Jøsang, A.; AlZomai, M.; Suriadi, S. Usability and Privacy in Identity Management Architectures. In Proceedings of the Fifth
Australasian Symposium on Grid Computing and e-Research (AusGrid 2007), the Fifth Australasian Information Security
-----
_Appl. Sci. 2022, 12, 12415_ 19 of 20
Workshop (Privacy Enhancing Technologies) (AISW 2007), and the Australasian Workshop on Health Knowledge Management
and Discovery (HKMD 2007). Proceedings, Ballarat, VIC, Australia, 30 January-2 February 2007.
5. Panait, A.; Olimid, R.; Stefane, A. Identity Management on Blockchain—Privacy and Security Aspects. Proc. Rom. Acad.-Ser. A
_Math. Phys. Tech. Sci. Inf. Sci. 2021, 21, 45–52._
6. Alrodhan, W. Privacy and Practicality of Identity Management Systems: Academic Overview; Vdm Verlag Dr. Müller: Saarbrücken,
Germany, 2011.
7. Lim, S.Y.; Tankam Fotsing, P.; Almasri, A.; Musa, O.; Mat Kiah, M.L.; Ang, T.F.; Ismail, R. Blockchain Technology the Identity
Management and Authentication Service Disruptor: A Survey. Int. J. Adv. Sci. Eng. Inf. Technol. 2018, 8, 1735. Available online:
[http://insightsociety.org/ojaseit/index.php/ijaseit/article/view/6838 (accessed on 15 August 2022). [CrossRef]](http://insightsociety.org/ojaseit/index.php/ijaseit/article/view/6838)
8. Almeshal, T.A.; Alhogail, A.A. Blockchain for Businesses: A Scoping Review of Suitability Evaluations Frameworks. IEEE Access
**[2021, 9, 155425–155442. [CrossRef]](http://doi.org/10.1109/ACCESS.2021.3128608)**
9. Zhu, X. Research on blockchain consensus mechanism and implementation. IOP Conf. Ser. Mater. Sci. Eng. 2019, 569, 042058.
[[CrossRef]](http://doi.org/10.1088/1757-899X/569/4/042058)
10. Maldonado, F.C. Introduction to Blockchain and Ethereum: Use Distributed Ledgers to Validate Digital Transactions in a Decentralized and
_Trustless Manner; Packt Publishing: Birmingham, UK, 2018._
11. Joshi, J.; Nepal, S.; Zhang, Q.; Zhang, L. Blockchain—ICBC 2019. In Proceedings of the Second International Conference, held as
Part of the Services Conference Federation, SCF 2019, San Diego, CA, USA, 25–30 June 2019; Springer: Cham, Switzerland, 2019.
12. Bao, Z.; Wang, Q.; Shi, W.; Wang, L.; Lei, H.; Chen, B. When Blockchain Meets SGX: An Overview, Challenges, and Open Issues.
_[IEEE Access 2020, 8, 170404–170420. [CrossRef]](http://doi.org/10.1109/ACCESS.2020.3024254)_
13. Bouras, M.A.; Lu, Q.; Dhelim, S.; Ning, H. A Lightweight Blockchain-Based IoT Identity Management Approach. Future Internet
**[2021, 13, 24. [CrossRef]](http://doi.org/10.3390/fi13020024)**
14. Xu, R.; Chen, Y.; Blasch, E.; Chen, G. BlendCAC: A Smart Contract Enabled Decentralized Capability-Based Access Control
[Mechanism for the IoT. Computers 2018, 7, 39. [CrossRef]](http://doi.org/10.3390/computers7030039)
15. Ra, G.; Kim, T.; Lee, I. VAIM: Verifiable Anonymous Identity Management for Human-Centric Security and Privacy in the Internet
[of Things. IEEE Access 2021, 9, 75945–75960. [CrossRef]](http://doi.org/10.1109/ACCESS.2021.3080329)
16. Sun, S.; Du, R.; Chen, S.; Li, W. Blockchain-Based IoT Access Control System: Towards Security, Lightweight, and Cross-Domain.
_[IEEE Access 2021, 9, 36868–36878. [CrossRef]](http://doi.org/10.1109/ACCESS.2021.3059863)_
17. Cocco, L.; Tonelli, R.; Marchesi, M. Blockchain and Self Sovereign Identity to Support Quality in the Food Supply Chain. Future
_[Internet 2021, 13, 301. [CrossRef]](http://doi.org/10.3390/fi13120301)_
18. Cui, P.; Dixon, J.; Guin, U.; Dimase, D. A Blockchain-Based Framework for Supply Chain Provenance. IEEE Access 2019, 7,
[157113–157125. [CrossRef]](http://doi.org/10.1109/ACCESS.2019.2949951)
19. Wang, S.; Li, D.; Zhang, Y.; Chen, J. Smart Contract-Based Product Traceability System in the Supply Chain Scenario. IEEE Access
**[2019, 7, 115122–115133. [CrossRef]](http://doi.org/10.1109/ACCESS.2019.2935873)**
20. Ferdousi, T.; Gruenbacher, D.; Scoglio, C.M. A Permissioned Distributed Ledger for the US Beef Cattle Supply Chain. IEEE Access
**[2020, 8, 154833–154847. [CrossRef]](http://doi.org/10.1109/ACCESS.2020.3019000)**
21. Figueroa-Lorenzo, S.; Añorga Benito, J.; Arrizabalaga, S. Modbus Access Control System Based on SSI over Hyperledger Fabric
[Blockchain. Sensors 2021, 21, 5438. [CrossRef]](http://doi.org/10.3390/s21165438)
22. Zhu, Y.; Huang, C.; Hu, Z.; Al-Dhelaan, A.; Al-Dhelaan, M. Blockchain-Enabled Access Management System for Edge Computing.
_[Electronics 2021, 10, 1000. [CrossRef]](http://doi.org/10.3390/electronics10091000)_
23. Alsayed Kassem, J.; Sayeed, S.; Marco-Gisbert, H.; Pervez, Z.; Dahal, K. DNS-IdM: A Blockchain Identity Management System to
[Secure Personal Data Sharing in a Network. Appl. Sci. 2019, 9, 2953. [CrossRef]](http://doi.org/10.3390/app9152953)
24. Gutierrez-Aguero, I.; Anguita, S.; Larrucea, X.; Gomez-Goiri, A.; Urquizu, B. Burnable Pseudo-Identity: A Non-Binding
[Anonymous Identity Method for Ethereum. IEEE Access 2021, 9, 108912–108923. [CrossRef]](http://doi.org/10.1109/ACCESS.2021.3101302)
25. Gruner, A.; Muhle, A.; Meinel, C. ATIB: Design and Evaluation of an Architecture for Brokered Self-Sovereign Identity Integration
[and Trust-Enhancing Attribute Aggregation for Service Provider. IEEE Access 2021, 9, 138553–138570. [CrossRef]](http://doi.org/10.1109/ACCESS.2021.3116095)
26. Moreno, R.T.; Garcia-Rodriguez, J.; Bernabe, J.B.; Skarmeta, A. A Trusted Approach for Decentralised and Privacy-Preserving
[Identity Management. IEEE Access 2021, 9, 105788–105804. [CrossRef]](http://doi.org/10.1109/ACCESS.2021.3099837)
27. Wang, S.; Pei, R.; Zhang, Y. EIDM: A Ethereum-Based Cloud User Identity Management Protocol. IEEE Access 2019, 7, 115281–
[115291. [CrossRef]](http://doi.org/10.1109/ACCESS.2019.2933989)
28. Feng, X.; Cui, K.; Jiang, H.; Li, Z. EBAS: An Efficient Blockchain-Based Authentication Scheme for Secure Communication in
[Vehicular Ad Hoc Network. Symmetry 2022, 14, 1230. [CrossRef]](http://doi.org/10.3390/sym14061230)
29. Li, H.; Pei, L.; Liao, D.; Chen, S.; Zhang, M.; Xu, D. FADB: A Fine-Grained Access Control Scheme for VANET Data Based on
[Blockchain. IEEE Access 2020, 8, 85190–85203. [CrossRef]](http://doi.org/10.1109/ACCESS.2020.2992203)
30. Lu, Z.; Liu, W.; Wang, Q.; Qu, G.; Liu, Z. A Privacy-Preserving Trust Model Based on Blockchain for VANETs. IEEE Access 2018, 6,
[45655–45664. [CrossRef]](http://doi.org/10.1109/ACCESS.2018.2864189)
31. Javed, I.T.; Alharbi, F.; Bellaj, B.; Margaria, T.; Crespi, N.; Qureshi, K.N. Health-ID: A Blockchain-Based Decentralized Identity
[Management for Remote Healthcare. Healthcare 2021, 9, 712. [CrossRef]](http://doi.org/10.3390/healthcare9060712)
32. Stamatellis, C.; Papadopoulos, P.; Pitropakis, N.; Katsikas, S.; Buchanan, W.J. A Privacy-Preserving Healthcare Framework Using
[Hyperledger Fabric. Sensors 2020, 20, 6587. [CrossRef] [PubMed]](http://doi.org/10.3390/s20226587)
-----
_Appl. Sci. 2022, 12, 12415_ 20 of 20
33. Xiang, X.; Wang, M.; Fan, W. A Permissioned Blockchain-Based Identity Management and User Authentication Scheme for
[E-Health Systems. IEEE Access 2020, 8, 171771–171783. [CrossRef]](http://doi.org/10.1109/ACCESS.2020.3022429)
34. Akhter, A.F.M.S.; Ahmed, M.; Shah, A.F.M.S.; Anwar, A.; Kayes, A.S.M.; Zengin, A. A Blockchain-Based Authentication Protocol
[for Cooperative Vehicular Ad Hoc Network. Sensors 2021, 21, 1273. [CrossRef] [PubMed]](http://doi.org/10.3390/s21041273)
35. Bao, S.; Cao, Y.; Lei, A.; Asuquo, P.; Cruickshank, H.; Sun, Z.; Huth, M. Pseudonym Management Through Blockchain: Cost[Efficient Privacy Preservation on Intelligent Transportation Systems. IEEE Access 2019, 7, 80390–80403. [CrossRef]](http://doi.org/10.1109/ACCESS.2019.2921605)
36. Lin, C.; He, D.; Huang, X.; Khurram Khan, M.; Choo, K.-K.R. A New Transitively Closed Undirected Graph Authentication
[Scheme for Blockchain-Based Identity Management Systems. IEEE Access 2018, 6, 28203–28212. [CrossRef]](http://doi.org/10.1109/ACCESS.2018.2837650)
37. de Ponteves, H.; Eremenko, K.; Ligency Team. Blockchain A-Z™: Learn How To Build Your First Blockchain. September 2021.
[Available online: https://www.udemy.com/course/build-your-blockchain-az/#instructor-1 (accessed on 11 June 2022).](https://www.udemy.com/course/build-your-blockchain-az/#instructor-1)
38. Shobanadevi, A.; Tharewal, S.; Soni, M.; Kumar, D.D.; Khan, I.R.; Kumar, P. Novel identity management system using smart
[blockchain technology. Int. J. Syst. Assur. Eng. Manag. 2022, 13 (Suppl. 1), 496–505. [CrossRef]](http://doi.org/10.1007/s13198-021-01494-0)
39. Lastovetska, A. Blockchain Architecture Basics: Components, Structure, Benefits & Creation. 5 January 2021. Available online:
[https://mlsdev.com/blog/156-how-to-build-your-own-blockchain-architecture (accessed on 1 November 2022).](https://mlsdev.com/blog/156-how-to-build-your-own-blockchain-architecture)
40. [Buterin, V. The Meaning of Decentralization. [Online] Medium. 2017. Available online: https://medium.com/@VitalikButerin/](https://medium.com/@VitalikButerin/the-meaning-of-decentralization-a0c92b76a274)
[the-meaning-of-decentralization-a0c92b76a274 (accessed on 26 October 2022).](https://medium.com/@VitalikButerin/the-meaning-of-decentralization-a0c92b76a274)
41. Wüst, K. Do you need a Blockchain? In Proceedings of the Crypto Valley Conference on Blockchain Technology (CVCBT), Zug,
Switzerland, 20–22 June 2018.
42. [Alrodhan, W.; Mitchell, C. Improving the Security of CardSpace. EURASIP J. Inf. Secur. 2009, 2009, 1–8. [CrossRef]](http://doi.org/10.1155/2009/167216)
43. Alrodhan, W.; Mitchell, C. Enhancing User Authentication in Claim-Based Identity Management. In Proceedings of the 2010
International Symposium on Collaborative Technologies and Systems, Chicago, IL, USA, 17–21 May 2010.
44. Dai, Z.; Zhou, W. The Federated Identity and Access Management Architectures: A Literature Survey; Deakin University, School of
Information Technology: Geelong, VIC, Australia, 2005.
45. Sung, C.; Park, J. Understanding of blockchain-based identity management system adoption in the public sector. J. Enterp. Inf.
_[Manag. 2021, 34, 1481–1505. [CrossRef]](http://doi.org/10.1108/JEIM-12-2020-0532)_
46. Niu, J.; Ren, Z. A self-sovereign identity management scheme using smart contracts. MATEC Web Conf. 2021, 336, 08005.
[[CrossRef]](http://doi.org/10.1051/matecconf/202133608005)
47. Bouras, M.; Lu, Q.; Zhang, F.; Wan, Y.; Zhang, T.; Ning, H. Distributed Ledger Technology for eHealth Identity Privacy: State of
[the Art and Future Perspective. Sensors 2020, 20, 483. [CrossRef]](http://doi.org/10.3390/s20020483)
48. Ferdous, M.S.; Poet, R. A Comparative Analysis of Identity Management Systems. In Proceedings of the 2012 International
Conference on High Performance Computing & Simulation (HPCS), Madrid, Spain, 2–6 July 2012.
49. Stockburger, L.; Kokosioulis, G.; Mukkamala, A.; Mukkamala, R.; Avital, M. Blockchain-enabled Decentralized Identity Manage[ment: The Case of Self-sovereign Identity in Public Transportation. Blockchain Res. Appl. 2021, 2, 100014. [CrossRef]](http://doi.org/10.1016/j.bcra.2021.100014)
50. Outchakoucht, A.; Es-Samaali, H. Dynamic Access Control Policy based on Blockchain and Machine Learning for the Internet of
[Things. Int. J. Adv. Comput. Sci. Appl. 2017, 8, 417–424. [CrossRef]](http://doi.org/10.14569/IJACSA.2017.080757)
51. Liao, C.H.; Guan, X.Q.; Cheng, J.H.; Yuan, S.M.; Blockchain-Based Identity Management and Access Control Framework for
[Open Banking Ecosystem. pp. 450–466. Available online: https://ssrn.com/abstract=4039865 (accessed on 5 October 2022).](https://ssrn.com/abstract=4039865)
52. Desabathina, N.V.M.; Merugu, S.; Gunjan, V.K.; Kumar, B.S. Agricultural Crowdfunding Through Blockchain. In ICDSMLA
_2020; Kumar, A., Senatore, S., Gunjan, V.K., Eds.; Lecture Notes in Electrical Engineering; Springer: Singapore, 2022; Volume 783._
[[CrossRef]](http://doi.org/10.1007/978-981-16-3690-5_155)
53. Tetzlaff, J.; Page, M.; Moher, D. Pns154 the prisma 2020 statement: Development of and key changes in an updated guideline for
[reporting systematic reviews and meta-analyses. Value Health 2020, 23, S312–S313. [CrossRef]](http://doi.org/10.1016/j.jval.2020.04.1154)
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3390/app122312415?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/app122312415, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "GOLD",
"url": "https://www.mdpi.com/2076-3417/12/23/12415/pdf?version=1670316403"
}
| 2,022
|
[
"Review"
] | true
| 2022-12-04T00:00:00
|
[
{
"paperId": "e93cdaa30a2a08ef4cd1b02c016d40cca07c6395",
"title": "EBAS: An Efficient Blockchain-Based Authentication Scheme for Secure Communication in Vehicular Ad Hoc Network"
},
{
"paperId": "e6fa0025972678f7911c7555e374738ba8da67ec",
"title": "Blockchain-based identity management and access control framework for open banking ecosystem"
},
{
"paperId": "a02fbb024d20025ac6481106d1b43e81c992a93b",
"title": "Novel identity management system using smart blockchain technology"
},
{
"paperId": "94666e9231860cad4c6314c4cbfd6e9fb14313ef",
"title": "Blockchain and Self Sovereign Identity to Support Quality in the Food Supply Chain"
},
{
"paperId": "dda4fdfe5324d17aa37a1a3e1874c08e0169cc14",
"title": "Agricultural Crowdfunding Through Blockchain"
},
{
"paperId": "315d1f7b6effbc6edfe35cb1c8072f3b3bed59fe",
"title": "Understanding of blockchain-based identity management system adoption in the public sector"
},
{
"paperId": "91c4e342f6ce4e2086e19e82d2c790efde91225d",
"title": "Modbus Access Control System Based on SSI over Hyperledger Fabric Blockchain"
},
{
"paperId": "e63a9d5738e4c3f9d6ea71f290b8d72099bf8be7",
"title": "Health-ID: A Blockchain-Based Decentralized Identity Management for Remote Healthcare"
},
{
"paperId": "8ee6a504f7217b8797606e38b32ef9524b1eb9a8",
"title": "Blockchain-Enabled Decentralized Identify Management: The Case of Self-Sovereign Identity in Public Transportation"
},
{
"paperId": "8516ef3c32c934fd8f731efd954043fae390fa88",
"title": "Blockchain-Enabled Access Management System for Edge Computing"
},
{
"paperId": "c334ad5237bca17aa678e40fa1876e6f067bfa40",
"title": "A Blockchain-Based Authentication Protocol for Cooperative Vehicular Ad Hoc Network"
},
{
"paperId": "713a182d25a452966f9a68cc71888bf8a6b2daf4",
"title": "A Lightweight Blockchain-Based IoT Identity Management Approach"
},
{
"paperId": "4e5dae79996e1e6f77e726f00f0ba0aa15ae7a8c",
"title": "A Privacy-Preserving Healthcare Framework Using Hyperledger Fabric"
},
{
"paperId": "a31f659a46cecb0652d4b54f6ef7a303b0d87795",
"title": "FADB: A Fine-Grained Access Control Scheme for VANET Data Based on Blockchain"
},
{
"paperId": "ce99d3c263a4596bdb9298f760dd4def8d38fe5f",
"title": "PNS154 THE PRISMA 2020 STATEMENT: DEVELOPMENT OF AND KEY CHANGES IN AN UPDATED GUIDELINE FOR REPORTING SYSTEMATIC REVIEWS AND META-ANALYSES"
},
{
"paperId": "496a450d5921c13a369fb885c8c19400d5d27c00",
"title": "Identity Management on Blockchain - Privacy and Security Aspects"
},
{
"paperId": "d324c7d451061edb5f2d8e04afb051058bd5b4da",
"title": "Distributed Ledger Technology for eHealth Identity Privacy: State of The Art and Future Perspective"
},
{
"paperId": "cfa7edded27c9342abe3cd1f35ee137332c0c4f2",
"title": "A Blockchain-Based Framework for Supply Chain Provenance"
},
{
"paperId": "d91a912d8a959eb57ed494a0fac48441094b73a9",
"title": "Smart Contract-Based Product Traceability System in the Supply Chain Scenario"
},
{
"paperId": "5e6ff0d5a5808045905c7e86ecdf0b6d5bc0364e",
"title": "Research on blockchain consensus mechanism and implementation"
},
{
"paperId": "e9747eace33602d3f4b5f5c0fd35e33c31b022cc",
"title": "DNS-IdM: A Blockchain Identity Management System to Secure Personal Data Sharing in a Network"
},
{
"paperId": "f9a498c61698444c7b0924a84bc6a31a7436f9b7",
"title": "Pseudonym Management Through Blockchain: Cost-Efficient Privacy Preservation on Intelligent Transportation Systems"
},
{
"paperId": "11a5fa69fd443dce30eab25309c2727e45427834",
"title": "Blockchain Technology the Identity Management and Authentication Service Disruptor: A Survey"
},
{
"paperId": "6cc50a2fb28f7c8f810a5e1c919686e2d6bf9ed2",
"title": "A Privacy-Preserving Trust Model Based on Blockchain for VANETs"
},
{
"paperId": "83bc879120207d575fa92dbbc1d34f40351e6085",
"title": "BlendCAC: A Smart Contract Enabled Decentralized Capability-Based Access Control Mechanism for the IoT"
},
{
"paperId": "ad9760ea1568263d4f670edc52e8d91875c95e42",
"title": "Do you Need a Blockchain?"
},
{
"paperId": "54099b438aa0c825a009db810a5d9d86f1140311",
"title": "A New Transitively Closed Undirected Graph Authentication Scheme for Blockchain-Based Identity Management Systems"
},
{
"paperId": "2fb015911e45a39153c2006b35001a86f75f9c31",
"title": "A comparative analysis of Identity Management Systems"
},
{
"paperId": "2b925752f1085029c4124e5e0f360c431403902c",
"title": "Privacy and Practicality of Identity Management Systems: Academic Overview"
},
{
"paperId": "79834e55efa027feb52d492cb346b4211104c7d4",
"title": "Enhancing user authentication in claim-based identity management"
},
{
"paperId": "31f8d61d98c0192f44a1ccd85fbf23ec77e645f4",
"title": "Usability and Privacy in Identity Management Architectures"
},
{
"paperId": "63a9681663778913982f8f028e170be5ff36f532",
"title": "The Meaning of Decentralization"
},
{
"paperId": "73940007299f38fad090f45357c11d168a4f1de1",
"title": "ICDSMLA 2020"
},
{
"paperId": "eead3bf6565da57f622af2f160018a6db4311856",
"title": "A self-sovereign identity management scheme using smart contracts"
},
{
"paperId": "b2949146c822d2b0b3ad9f61e63f2f12df7b47ff",
"title": "A Trusted Approach for Decentralised and Privacy-Preserving Identity Management"
},
{
"paperId": "69501a0df726a5061079df41703ccc3a6cd7e4ed",
"title": "Blockchain for Businesses: A Scoping Review of Suitability Evaluations Frameworks"
},
{
"paperId": "cf80de04004d28c24ea2e551759e7e25ea288f37",
"title": "VAIM: Verifiable Anonymous Identity Management for Human-Centric Security and Privacy in the Internet of Things"
},
{
"paperId": "67b86a0ed3ff564cae9e58d2cf1b2aee4ee141bd",
"title": "ATIB: Design and Evaluation of an Architecture for Brokered Self-Sovereign Identity Integration and Trust-Enhancing Attribute Aggregation for Service Provider"
},
{
"paperId": "e907694d7bb6cdfda680d19066e0cf135b3a6a81",
"title": "Burnable Pseudo-Identity: A Non-Binding Anonymous Identity Method for Ethereum"
},
{
"paperId": "46eee47d659397cc5aa0df54544c24ce78948ab4",
"title": "Blockchain-Based IoT Access Control System: Towards Security, Lightweight, and Cross-Domain"
},
{
"paperId": null,
"title": "Ligency Team. Blockchain A-Z™: Learn How To Build Your First Blockchain"
},
{
"paperId": null,
"title": "Blockchain Architecture Basics: Components, Structure, Benefits & Creation"
},
{
"paperId": "811e4ec25ae9558b5daf057a8a3d60c603c3cbb7",
"title": "A Permissioned Distributed Ledger for the US Beef Cattle Supply Chain"
},
{
"paperId": "68751a6ada0290576996ce34cf229e1a4a582625",
"title": "A Permissioned Blockchain-Based Identity Management and User Authentication Scheme for E-Health Systems"
},
{
"paperId": "c60306cf289468331d8e4ec5b0b4a9d77eb58e38",
"title": "When Blockchain Meets SGX: An Overview, Challenges, and Open Issues"
},
{
"paperId": "20f0822a8a354809c7b34181e80a8c99412c08f3",
"title": "EIDM: A Ethereum-Based Cloud User Identity Management Protocol"
},
{
"paperId": "1ddae1695eb04990be95ff9038a218cfd4f14bb3",
"title": "Blockchain – ICBC 2019"
},
{
"paperId": null,
"title": "Introduction to Blockchain and Ethereum: Use Distributed Ledgers to Validate Digital Transactions in a Decentralized and Trustless Manner"
},
{
"paperId": "c3bd5f09b1f13c365c78101f65a77acb41dfcc20",
"title": "Dynamic Access Control Policy based on Blockchain and Machine Learning for the Internet of Things"
},
{
"paperId": null,
"title": "Identity Management Systems: Laws of Identity for Models (cid:48) Evaluation"
},
{
"paperId": "13ab07ab629844cae21f2ed86ff8e32bd355f1e2",
"title": "Digital Identity and Identity Management Technologies"
},
{
"paperId": "8e63dfee41b2a8d2baccda16577e7a6857486bac",
"title": "Improving the Security of CardSpace"
},
{
"paperId": null,
"title": "The Federated Identity and Access Management Architectures: A Literature Survey"
},
{
"paperId": "667cc1ac858c058b6f55797d427bb5174f712dba",
"title": "Journal of Network and Computer Applications"
}
] | 20,319
|
en
|
[
{
"category": "Mathematics",
"source": "external"
},
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Mathematics",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/02280183f81323acabd93adf831183a26abe12c0
|
[
"Mathematics",
"Computer Science"
] | 0.866114
|
A variant of Wiener’s attack on RSA
|
02280183f81323acabd93adf831183a26abe12c0
|
Computing
|
[
{
"authorId": "1759823",
"name": "A. Dujella"
}
] |
{
"alternate_issns": null,
"alternate_names": null,
"alternate_urls": null,
"id": "48433268-b689-457a-8e57-36895fd4f04e",
"issn": "0010-485X",
"name": "Computing",
"type": "journal",
"url": "https://link.springer.com/journal/607"
}
|
Wiener’s attack is a well-known polynomial-time attack on a RSA cryptosystem with small secret decryption exponent d, which works if d < n0.25, where n = pq is the modulus of the cryptosystem. Namely, in that case, d is the denominator of some convergent pm/qm of the continued fraction expansion of e/n, and therefore d can be computed efficiently from the public key (n, e). There are several extensions of Wiener’s attack that allow the RSA cryptosystem to be broken when d is a few bits longer than n0.25. They all have the run-time complexity (at least) O(D2), where d = Dn0.25. Here we propose a new variant of Wiener’s attack, which uses results on Diophantine approximations of the form |α − p/q| < c/q2, and “meet-in-the-middle” variant for testing the candidates (of the form rqm+1 + sqm) for the secret exponent. This decreases the run-time complexity of the attack to O(D log D) (with the space complexity O(D)).
|
## A variant of Wiener’s attack on RSA
### Andrej Dujella
Abstract
Wiener’s attack is a well-known polynomial-time attack on a RSA
cryptosystem with small secret decryption exponent d, which works if
d < n[0][.][25], where n = pq is the modulus of the cryptosystem. Namely,
in that case, d is the denominator of some convergent pm/qm of the
continued fraction expansion of e/n, and therefore d can be computed
efficiently from the public key (n, e).
There are several extensions of Wiener’s attack that allow the RSA
cryptosystem to be broken when d is a few bits longer than n[0][.][25]. They
all have the run-time complexity (at least) O(D[2]), where d = Dn[0][.][25].
Here we propose a new variant of Wiener’s attack, which uses results
on Diophantine approximations of the form α p/q < c/q[2], and
| − |
“meet-in-the-middle” variant for testing the candidates (of the form
rqm+1 + sqm) for the secret exponent. This decreases the run-time
complexity of the attack to O(D log(D)) (with the space complexity
O(D)).
## 1 Introduction
The most popular public key cryptosystem in use today is the RSA cryptosystem, introduced by Rivest, Shamir, and Adleman [8]. Its security is
based on the intractability of the integer factorization problem.
The modulus n of a RSA cryptosystem is the product of two large primes
p and q. The public exponent e and the secret exponent d are related by
ed 1 (mod ϕ(n)), (1)
≡
0 2000 Mathematics Subject Classification: Primary 94A60; Secondary 11A55, 11J70.
Key words: RSA cryptosystem, continued fractions, cryptanalysis
1
-----
where ϕ(n) = (p 1)(q 1). In a typical RSA cryptosystem, p and q have
− −
approximately the same number of bits, while e < n. The encryption and
decryption algorithms are given by C = M [e] mod n, M = C [d] mod n.
To speed up the RSA decryption one may try to use small secret decryption exponent d. The choice of a small d is especially interesting when
there is a large difference in computing power between two communicating
devices, e.g. in communication between a smart card and a larger computer.
In this situation, it would be desirable that the smart card has a small secret
exponent, while the larger computer has a small public exponent, to reduce
the processing required in the smart card.
In 1990, Wiener [13] described a polynomial time algorithm for breaking
a typical (i.e. p and q are of the same size and e < n) RSA cryptosystem if
the secret exponent d has at most one-quarter as many bits as the modulus
n. From (1) it follows that there is an integer k such that ed kϕ(n) = 1.
−
Since ϕ(n) ≈ n, we have that [k]d [≈] n[e] [. Wiener’s attack is usually described in]
the following form (see [2, 9]):
[√]4
If p < q < 2p, e < n and d < [1] n, then d is the denominator of some
3
convergent of the continued fraction expansion of [e]
n [.]
Indeed, under these assumptions it is easy to show that
e 1
<
n d 2d[2] [.]
���� [−] [k] ����
By the classical Legendre’s theorem, [k]d [is some convergent][ p]qm[m] [of the continued]
fraction expansion of ne [, and therefore][ d][ can be computed efficiently from]
the public key (n, e). Namely, the total number of convergents is of order
O(log n), and each convergent can be tested in polynomial time.
In 1997, Verheul and van Tilborg [12] proposed an extension of Wiener’s
attack that allows the RSA cryptosystem to be broken when d is a few
bits longer than n[0][.][25]. For d > n[0][.][25] their attack needs to do an exhaustive
search for about 2t+8 bits (under reasonable assumptions on involved partial
convergents), where t = log2(d/n[0][.][25]).
In [4], we proposed a slight modification of the Verheul and van Tilborg
attack, based on Worley’s result on Diophantine approximations [14], which
implies that all rationals [p]q [satisfying the inequality]
p
α
−
q
����
< c (2)
q[2] [,]
����
2
-----
for a positive real number c, have the form
p
(3)
q [=][ rp]rqm[m]+1[+1] ±[ ±] sq[ sp]m[m]
for some m 1 and nonnegative integers r and s such that rs < 2c. It has
≥−
been shown recently in [5] that Worley’s result is sharp, in the sense that the
condition rs < 2c cannot be replaced by rs < (2 ε)c for any ε.
−
In both mentioned extensions of Wiener’s attack, the candidates for the
secret exponent are of the form d = rqm+1+sqm. Then we test all possibilities
for d. The number of possibilities is roughly the product of the number
of possibilities for r and the number of possibilities for s, which is O(D[2]),
where d = Dn[0][.][25]. More precisely, the number of possible pairs (r, s) in
the Verheul and van Tilborg attack is O(D[2]A[2]), where A = max{ai : i =
m+1, m+2, m+3, while in our variant the number of pairs is O(D[2] log A)
}
(and also O(D[2] log D)).
Another modification of the Verheul and van Tilborg attack has been
recently proposed by Sun, Wu an Chen [11]. It requires (heuristically) an
exhaustive search for about 2t 10 bits, so its complexity is also O(D[2]).
−
We cannot expect drastic improvements here, since, by a result of Steinfeld,
Contini, Wang and Pieprzyk [10], there does not exist an attack in this class
with subexponential running time.
Boneh and Durfee [3] and Bl¨omer and May [1] proposed attacks based
on Coppersmith’s lattice-based technique for finding small roots of modular
polynomials equations using LLL-algorithm. The attacks work if d < n[0][.][292].
The conjecture is that the right bound below which a typical version of RSA
is insecure is d < n[0][.][5].
In the present paper, we propose a new variant of Wiener’s attack. It
also uses continued fractions and searches for candidates for the secret key
in the form d = rqm+1 + sqm. However, the searching phase of this variant is
significantly faster. Its complexity is O(D log D), and it works efficiently for
d < 10[30]n[0][.][25]. Although this bound is asymptotically weaker than the bounds
in the above mentioned attacks based on the LLL-algorithm (note however
that these bounds are not strictly proved since Coppersmith’s theorem in the
bivariate case is only a heuristic result - see also [6, 7]), for practical values
of n (e.g. for 1024-bits) these bounds are of comparable size.
3
-----
## 2 The Verheul and van Tilborg attack
In this section we briefly describe the Verheul and van Tilborg attack [12]
and its modification from [4].
We assume that p < q < 2p and e < n. Then it is easy to see that
e .122 e
< 2 (4)
n d n[√]n [.]
���� [−] [k] ����
Let m be the largest (odd) integer satisfying [p]qm[m] n [>][ 2]n[.][122][√]n[ e] [. Verheul and van]
[−] [e]
Tilborg proposed to search for [k]d [among the fractions of the form][ rp]rq[m]m[+1]+1[+]+[sp]sqm[m] [.]
This leads to the system
rpm+1 + spm = k,
rqm+1 + sqm = d.
The determinant of the system satisfies |pm+1qm −qm+1pm| = 1, and therefore
the system has (positive) integer solutions:
r = dpm − kqm,
s = kqm+1 − dpm+1.
If r and s are small, then they can be found by an exhaustive search. Let
[a0; a1, a2, . . .] be the continued fraction expansion of e/n and D = d/n[0][.][25].
In [4], the following upper bounds for r and s were derived:
r < max{�2.122(am+3 + 2)(am+2 + 1)D, �2.122(am+2 + 2)D},
s < max{2�2.122(am+3 + 2)D, �2.122(am+2 + 2)(am+1 + 1)D}.
The modified attack proposed in [4] searches for [k]d [among the fractions of]
the forms [rp]rq[m]m[+1]+1[+]+[sp]sqm[m] [,][ rp]rqm[m]+2[+2]−[−]sq[sp]m[m]+1[+1] [and][ rp]rq[m]m[+3]+3[+]+[sp]sqm[m]+2[+2] [. It results with bounds for]
r and s which are (almost) independent on the partial quotients am’s. Hence,
in both attacks bounds for r and s are of the form O(D), but in the case
of [4] the implied constants are much smaller (indeed, the table in Section 4
shows that with high probability we have r < 4D and s < 4D).
## 3 Testing the candidates
There are two principal methods for testing candidates for the secret exponent d.
4
-----
Method I ([13]): Compute p and q, assuming d is the correct guess,
using the following formulas:
ϕ(n) = (de 1)/k, p + q = n + 1 ϕ(n),
− −
(q p)[2] = (p + q)[2] 4n,
− −
p = [p][ +][ q], q = [p][ +][ q] + [q][ −] [p] .
− [q][ −] [p]
2 2 2 2
Method II ([9, Chapter 17]): Test the congruence (M [e])[d] M (mod n),
≡
for some random value of M, or simply for M = 2.
Both methods are very efficient. But in the situation where we have to
test huge amount of candidates for d of the form rqm+1 + sqm, there is a
significant difference between them. With the Method I it seems that we
cannot avoid testing separately all possible pairs (r, s). On the other hand,
here we present a new idea, which is to apply “meet-in-the-middle” to the
Method II.
We want to test whether
2[e][(][rq][m][+1][+][sq][m][)] 2 (mod n). (5)
≡
Note that m is (almost) fixed. Indeed, let m[′] be the largest odd integer such
that
pm′
qm′ [> e]n [+ 2]n[.][122][√]n [e][.]
Then m m[′], m[′] + 1, m[′] + 2 (see [4] for details).
∈{ }
Let 2[eq][m][+1] mod n = a, (2[eq][m])[−][1] mod n = b. Then we test the congruence
a[r] 2b[s] (mod n). (6)
≡
We can do it by computing a[r] mod n for all r, sorting the list of results,
and then computing 2b[s] mod n for each s one at a time, and checking if the
result appears in the sorted list.
This decreases the time complexity of the testings phase to O(D log D)
(with the space complexity O(D)).
5
-----
## 4 Implementation issues and improvements
The theoretic base for the extension of Wiener’s attack is Worley’s theorem
on Diophantine approximations of the form (2). We have already mentioned
a result from [5] which shows that Worley’s result is in some sense the best
possible. However, some improvements are possible if we consider unsymmetrical variants of Worley’s result (with different bounds on r and s). Roughly
speaking, in solutions of (2) in form (3), if r < s then we may take rs < c
instead of rs < 2c. Due to such unsymmetrical results, a space-time tradeoff
might be possible. The following table shows the chance of success of our
attack for various (symmetrical and unsymmetrical) bounds on r and s. We
can see that, with the same bound for rs, the better results are obtained for
smaller bounds on r and larger bounds on s. In the implementations, this
fact can be used to decrease the memory requirements (up to factor 16).
bound for r bound for s chance of success
4D 4D 98%
2D 2D 89%
D D 65%
D 4D 86%
4D D 74%
D/2 2D 70%
2D D/2 47%
D/4 4D 54%
4D D/4 28%
In the implementation of the proposed attack, we can use hash functions instead of sorting. Furthermore, it is not necessary to store all bits
of a[r] mod n in the hash table. Indeed, values of a[r] mod n are from the
set 0, 1, . . ., n, and the number of r’s is typically much smaller than n.
{ }
Therefore, around 2 log2 D stored bits will suffice in order to avoid too many
accidental collisions. Note that a reasonable number of collisions is not big
problem here, since each such collision can be efficiently tested by Method I.
Hash tables can be used to take into account the condition gcd(r, s) = 1. This
condition was easy to use in brute-force testing of all possible pairs (r, s), but
the direct application of our “meet-in-the-middle” variant seemingly ignores
it. But if we create rows in the hash table according to divisibility properties
6
-----
of exponents r modulo small primes, we may take again an advantage of this
condition and speed up the algorithm up to 39%.
We have implemented several variants of the proposed attack in PARI and
C++, and they work efficiently for values of D up to 2[30], i.e. for d < 2[30]n[0][.][25].
For larger values of D the memory requirements become too demanding
for ordinary computers.
The following table compares this bound with the bound of d in the best
known attacks on RSA with small secret exponent based on LLL-algorithm.
log2 n log2(2[30]n[0][.][25]) log2(n[0][.][292])
512 158 150
768 222 224
1024 286 299
2048 542 598
The attack can be also slightly improved by using better approximations
to [k]d [, e.g.] n+1−e 2[√]n [instead of][ e]n[. Namely,]
e .1221 e
< 0 . (7)
n + 1 2[√]n d n[√]n
���� − [−] [k] ����
Comparing (7) with (4), we see that by replacing n[e] [by] n+1−e 2[√]n [we can gain]
the factor 4 in bounds for r and s, so decreasing both, time and memory
requirements.
With these improvements, for 1024-bits RSA modulus n, the range in
which our attack can be applied becomes comparable and competitive with
best known attacks based on the LLL-algorithm.
Acknowledgements. The author would like to thank Vinko Petriˇcevi´c
for his help with C++ implementation of the various variants of the attack
described in this paper. The author was supported by the Ministry of Science,
Education and Sports, Republic of Croatia, grant 037-0372781-2821.
## References
[1] J. Bl¨omer, A. May, Low secret exponent RSA revisited, Cryptography and
Lattice - Proceedings of CaLC 2001, Lecture Notes in Comput. Sci. 2146
(2001), 4–19.
7
-----
[2] D. Boneh, Twenty years of attacks on the RSA cryptosystem, Notices Amer.
Math. Soc. 46 (1999), 203–213.
[3] D. Boneh, G. Durfee, Cryptanalysis of RSA with private key d less than
N [0][.][292], Advances in Cryptology - Proceedings of Eurocrypt ’99, Lecture Notes
in Comput. Sci. 1952 (1999), 1–11.
[4] A. Dujella, Continued fractions and RSA with small secret exponent, Tatra
Mt. Math. Publ. 29 (2004), 101–112.
[5] A. Dujella, B. Ibrahimpaˇsi´c, On Worley’s theorem in Diophantine approximations, Ann. Math. Inform., to appear.
[6] J. Hinek, Low Public Exponent Partial Key and Low Private Exponent Attacks on Multi-prime RSA, Master’s thesis, University of Waterloo, 2002.
[7] M. J. Hinek, M. K. Low, E. Teske, On some attacks on multi-prime RSA,
Proceedings of SAC 2002, Lecture Notes in Comput. Sci. 2595 (2003), 385–
404.
[8] R. L. Rivest, A. Shamir, L. Adleman, A method for obtaining digital signatures
and publi-key cryptosystems, Communications of the ACM 21 (1978), 120–
126.
[9] N. Smart, Cryptography: An Introduction, McGraw-Hill, London, 2002.
[10] R. Steinfeld, S. Contini, H. Wang, J. Pieprzyk, Converse results to the Wiener
attack on RSA, Public Key Cryptography - PKC 2005, Lecture Notes in
Comput. Sci. 3386 (2005), 184–198.
[11] H.-M. Sun, M.-E. Wu, Y.-H. Chen, Estimating the Prime-Factors of an RSA
Modulus and an Extension of the Wiener Attack, Applied Cryptography and
Network Security, Lecture Notes in Comput. Sci. 4521 (2007), 116–128.
[12] E. R. Verheul, H. C. A. van Tilborg, Cryptanalysis of ‘less short’ RSA secret
exponents, Appl. Algebra Engrg. Comm. Computing 8 (1997), 425–435.
[13] M. J. Wiener, Cryptanalysis of short RSA secret exponents, IEEE Trans.
Inform. Theory 36 (1990), 553–558.
[14] R. T. Worley, Estimating α p/q, Austral. Math. Soc. Ser. A 31 (1981),
| − |
202–206.
8
-----
Andrej Dujella
Department of Mathematics
University of Zagreb
Bijeniˇcka cesta 30
10000 Zagreb, Croatia
E-mail address: duje@math.hr
9
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/0811.0063, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "CLOSED",
"url": ""
}
| 2,008
|
[
"JournalArticle"
] | false
| 2008-11-01T00:00:00
|
[
{
"paperId": "9159e05170d7e2de593dca5f6318bd522b49f82a",
"title": "A New Lattice Construction for Partial Key Exposure Attack for RSA"
},
{
"paperId": "7ed6709c43cc052fa779cbf2488a7d1dd2315986",
"title": "Estimating the Prime-Factors of an RSA Modulus and an Extension of the Wiener Attack"
},
{
"paperId": "546db35cfd519822b9f635a10940fbccf08e6e2a",
"title": "Partial Key Exposure Attacks on RSA up to Full Size Exponents"
},
{
"paperId": "5df59cf84251633a661a71bd2284a8914c4f3105",
"title": "Converse Results to the Wiener Attack on RSA"
},
{
"paperId": "4c01e5b25b0dd9b86894c778051ab636fa429466",
"title": "Cryptography: An Introduction"
},
{
"paperId": "efbee2d0be5391a2245539c23ce86794db82d1f2",
"title": "Continued fractions and RSA with small secret exponent"
},
{
"paperId": "237624e4ebd4d23952680ef71be009ec4ad09389",
"title": "On Some Attacks on Multi-prime RSA"
},
{
"paperId": "7dcaed1b0ac58978ba3c92f476d6cfe9edcf3e30",
"title": "Low Secret Exponent RSA Revisited"
},
{
"paperId": "ccfafaf2ccf0e3607829d449b8cf3c444e85405d",
"title": "Cryptanalysis of RSA with private key d less than N0.292"
},
{
"paperId": "8fa97e90498cf02bfd5e75b9f619795f452dde03",
"title": "Cryptanalysis of ‘Less Short’ RSA Secret Exponents"
},
{
"paperId": "e1855c956271f84decba5972cf84e09e8932da89",
"title": "Cryptanalysis of Short RSA Secret Exponents (Abstract)"
},
{
"paperId": "6408f394f424660f430fdd6fa32520fc5bdf0847",
"title": "Estimating |α – p / q|"
},
{
"paperId": "b93923304173b662818588acc4554f7d46c9ed9c",
"title": "On Worley's theorem in Diophantine approximations"
},
{
"paperId": "2d859e8937fe652558a60e82ffd39cd4ab835e31",
"title": "On the security of some variants of rsa"
},
{
"paperId": null,
"title": "Low Public Exponent Partial Key and Low Private Exponent Attacks on Multi-prime RSA"
},
{
"paperId": "e85c8789db8a9379c60a1621878222bcbe7a11a3",
"title": "TWENTY YEARS OF ATTACKS ON THE RSA CRYPTOSYSTEM"
},
{
"paperId": "22d5ff0ad3dadaa713d07e00499587e76c80c3f2",
"title": "Cryptanalysis of RSA with Private Key d Less Than N 0"
},
{
"paperId": null,
"title": "A space-time tradeoff might be possible, by using unsymmetrical variants of Worley's result (with different bounds on r and s)"
},
{
"paperId": null,
"title": "Andrej Dujella Department of Mathematics University of Zagreb Bijenička cesta 30"
}
] | 4,797
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Economics",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/02292be913a5c03ba6177bc9b7e85eaed6c26cf7
|
[
"Computer Science"
] | 0.90421
|
Time is Money: Strategic Timing Games in Proof-of-Stake Protocols
|
02292be913a5c03ba6177bc9b7e85eaed6c26cf7
|
Conference on Advances in Financial Technologies
|
[
{
"authorId": "2133346796",
"name": "Caspar Schwarz-Schilling"
},
{
"authorId": "46459277",
"name": "Fahad Saleh"
},
{
"authorId": "13227476",
"name": "T. Thiery"
},
{
"authorId": "50617806",
"name": "Jennifer Pan"
},
{
"authorId": "1737249",
"name": "Nihar B. Shah"
},
{
"authorId": "9545425",
"name": "B. Monnot"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"AFT",
"Conf Adv Financial Technol"
],
"alternate_urls": null,
"id": "3f5bab6f-a198-4ad4-b2c1-f674f0d964d3",
"issn": null,
"name": "Conference on Advances in Financial Technologies",
"type": "conference",
"url": null
}
|
We propose a model suggesting that honest-but-rational consensus participants may play timing games, and strategically delay their block proposal to optimize MEV capture, while still ensuring the proposal's timely inclusion in the canonical chain. In this context, ensuring economic fairness among consensus participants is critical to preserving decentralization. We contend that a model grounded in honest-but-rational consensus participation provides a more accurate portrayal of behavior in economically incentivized systems such as blockchain protocols. We empirically investigate timing games on the Ethereum network and demonstrate that while timing games are worth playing, they are not currently being exploited by consensus participants. By quantifying the marginal value of time, we uncover strong evidence pointing towards their future potential, despite the limited exploitation of MEV capture observed at present.
|
## Time is Money: Strategic Timing Games in Proof-of-Stake Protocols
Caspar Schwarz-Schilling[1], Fahad Saleh[2], Thomas Thiery[1],
Jennifer Pan[3], Nihar Shah[3], and Barnabé Monnot[1]
1 Ethereum Foundation
```
{caspar.schwarz-schilling,thomas.thiery,barnabe.monnot}@ethereum.org
```
2 Wake Forest University
```
saleh@wfu.edu
```
3 Jump Crypto
```
{jpan,nshah}@jumptrading.com
```
**Abstract. We propose a model suggesting that honest-but-rational consensus participants may play tim-**
_ing games, and strategically delay their block proposal to optimize MEV capture, while still ensuring_
the proposal’s timely inclusion in the canonical chain. In this context, ensuring economic fairness among
consensus participants is critical to preserving decentralization. We contend that a model grounded in
honest-but-rational consensus participation provides a more accurate portrayal of behavior in economically incentivized systems such as blockchain protocols. We empirically investigate timing games on the
Ethereum network and demonstrate that while timing games are worth playing, they are not currently
being exploited by consensus participants. By quantifying the marginal value of time, we uncover strong
evidence pointing towards their future potential, despite the limited exploitation of MEV capture observed
at present.
### 1 Introduction
Consensus protocols are typically evaluated based on their ability to maintain liveness and safety [11], referring to
the regular addition of new transactions to the output ledger in a timely manner, and to the security of confirmed
transactions remaining in their positions within the ledger. However, beyond liveness and safety, blockchain
protocols require fairness of economic outcomes amongst consensus participants to preserve decentralization.
More specifically, a protocol should be designed to maximize profitability of honest participation, wherein
participants adhere to the prescribed rules. Otherwise, a deviating participant will outcompete their honest
peers, leading to centralization of the validation set over time and security implications for consensus itself.
However, the advent of Maximal Extractable Value (MEV) frustrates such fairness goals. It is defined as the
value that consensus participants, in their duties as block producers, accrue by selectively including, excluding
and ordering user transactions [14,6], MEV has equally substantial implications for the security of consensus
protocols. For a system in which transaction fee rewards are dominant, consensus may become unstable due
to increased variance in miner rewards [12]. Similarly, it was argued that a rational actor issuing a whale
_transaction with an abnormally large transaction fee can convince peers to fork the current chain, further_
destabilizing consensus [22]. More broadly, understanding and mitigating the impact of MEV on the security
and fairness of blockchain networks has become a central concern of protocol designers [30].
As the whale transaction highlights, potential MEV accrues over time as users submit transactions and the
value of the set of pending transactions increases for the block producer. As a consequence, time is valuable to
consensus participants, a feature obviated by the assumption of honest behavior in previous models of consensus.
However, we argue that protocols who wish to preserve properties such as economic fairness amongst consensus
participants must assume some share of honest-but-rational consensus participation. In particular, the effects
of MEV on the consensus participants’ incentives must be better understood.
In this paper, we investigate the possibility for block proposers to delay their block proposal as long as
possible while ensuring they become part of the canonical chain, aiming to maximize MEV extraction. The
reader may note that in Proof-of-Work (PoW)-based leader selection protocols, delaying a proposal bears the
risk of losing to a competing block proposer. PoW protocols exhibit an inherent racing condition that prevents
these types of strategic delay deviations, or at least make them unprofitable in expectation. Thus, we investigate
the implications of MEV on the incentives of consensus participants, particularly block proposers, in a Proof-ofStake (PoS) context. More specifically, we consider propose-vote type of PoS protocols, where in each consensus
-----
p g,, y,,,
round, one leader proposes a block, and a committee of consensus participants is selected in-protocol to vote
on the acceptance of that block. This effectively grants block proposers a short-lived monopoly as the only
valid proposer for some given round. During this time interval they can attempt to strategically deviate from
their assigned block proposal time and delay the release of their block as long as possible in order to extract
more MEV, while still ensuring that a sufficient share of attesters see the block in time to vote it into the
canonical chain. This behavior leads to an environment in which honest validators earn less than their deviating
counterparts, resulting in stake centralization and second-order effects for consensus stability.
_Related Work To the best of our knowledge, timing games have not been formally analyzed in previous literature_
on Proof-of-Stake. Selfish mining [19], studied in the context of Proof-of-Work, relies on appropriately timing
the release of a block, in order to waste computation of honest miners and earn an outsize share of the rewards.
Our timing games are also concerned with strategic behavior to capture a larger share of the total available
rewards to consensus participants, yet do not feature the same dynamics as selfish mining in Proof-of-Work,
since participants in many PoS-based consensus mechanisms are given a fixed time interval in which to perform
their duties.
The security of PoS-based mechanisms has been discussed in terms of chain growth [18] or focusing on the
safety and liveness properties of hybrid protocols such as Gasper [10,24]. The economics literature has also
examined Proof-of-Stake security with respect to particular attacks such as the double-spending attack [26] and
51% attacks [20]. Separately, incentive considerations in the presence of MEV led to the discovery of severe
attacks on the Gasper consensus [28,25] and protocol changes to address such attacks [15,16,17].
_Our Contributions Our work models the value of time to consensus participants and explores the potential_
emergence of timing games in Proof-of-Stake protocols. By understanding the strategic behavior of consensus
participants within this model, we gain insights into how these dynamics affect the robustness of consensus
protocols to exogenous incentives, and ultimately fairness.
**– Despite initial pessimism regarding the existence of equilibria in timing games [23], we formally show how**
to sustain equilibrium behavior, where it is individually irrational for proposers to deviate from a schedule
enforced by attesters, and reward-sharing is fair among participants (Sections 2 and 3).
**– We then investigate whether such timing games might occur in real-world systems (namely, the Ethereum**
network), using a large, granular data set recording the MEV offered to block proposers over time. We show
incidental deviations from the honest protocol specification, highlighting the feasibility of timing games, yet
we do not conclude on the existence of intentional timing games (Section 4).
### 2 Model
We model an infinite horizon game among block proposers and attesters. Time is partitioned into slots n ∈ N,
each of time length ∆> 0. Each slot n has a block proposer n and a unit measure of attesters An = {A(i,n)}i∈[0,1]
where A(i,n) refers to the ith attester within slot n.[4]
The game evolves as follows:
**– At the beginning of slot n, proposer n acts by deciding whether to build on top of the block of proposer**
_n −_ 1 and also when to release their own block. More formally, proposer n selects φn ∈{0, 1} and tn ≥ _n · ∆_
where φn = 1 (φn = 0) refers to proposer n (not) building on top of the block of proposer n − 1 and tn
denotes the time at which proposer n releases their own block. Note that we specify that proposer n cannot
release their block before the start of slot n (i.e., tn ≥ _n · ∆) but that they release their block after the end_
of the slot.
**– After proposer n acts, all slot n attesters act simultaneously. In particular, attester A(i,n) decides whether to**
attest to the block of proposer n and also the time to release their attestation. More formally, attester A(i,n)
select ν(i,n) ∈{0, 1} and τ(i,n) ≥ _n · ∆_ where ν(i,n) = 1 (ν(i,n) = 0) refers to attester A(i,n) (not) attesting to
proposer n’s block and τ(i,n) refers to the time that they release their attestation. Notably, attester A(i,n)
can attest to the block of proposer n only if they receive the block before releasing their attestation. We
4 Note that we have a continuum of attesters, rather than a discrete set. In Ethereum PoS, over 18,000 attesters emit a
vote per slot (as of 2023-05-12).
-----
y g g
let δn,(i,n) ∼ _exp(θ[−][1]) refer to the random time required for the block of proposer n to reach attester A(i,n)_
where θ > 0 denotes the average communication time across the network and we assume that the slot length
is at least double the average communication time across the network (i.e., ∆ _≥_ 2θ). In turn, the action of
attester A(i,n) is constrained by ν(i,n) = 1 =⇒ _τ(i,n) ≥_ _tn + δn,(i,n)._
**2.1** **Block proposers**
The pay-off for proposer n is given as follows:
_U_ _[P]_ (tn, φn) =
�
_α + µ · (tn −_ _tn−_ )[+] if χn = 1 (1)
0 otherwise
where α, µ > 0 are exogenous constants while tn− corresponds to the time of the most recent canonical
block before slot n and χn ∈{0, 1} corresponds to whether the block in slot n is canonical on the blockchain.
We introduce the conditions for a block to become canonical in our model in the following, and delay until
Section 3.2 its interpretation with respect to established consensus models.
Note that we assume that the reward of proposer n increases linearly with time relative to the most recent
canonical block so long as block n eventually becomes canonical. This assumption reflects that proposer n
accrues incremental MEV over time by delaying the release of their block but that they risk being skipped if
they delay release for too long. The time of the most recent canonical block, tn−, is endogenous where slot n−
refers to the most recent canonical slot and is thus given explicitly as follows:
_n−_ = max{k ∈ N : χk = 1, χk ≤ _n −_ 1} (2)
For a block to be canonical, we require both that it receives sufficiently many successful attestations and that
the subsequent block producer builds on top of it. More formally, letting _A[˜]n denote the successful attestations_
for block n, χn is given explicitly as follows:
_χn =_
�
1 if φn+1 = 1, _A[˜]n_ _γ_
_≥_ (3)
0 otherwise
where the number of successful attestations for block n is given as the measure of attesters in slot n voting for
block n:
_A˜n = |{i ∈_ [0, 1] : ν(i,n) = 1}| (4)
**2.2** **Attesters**
Attester (i, n) receives a pay-off if and only if two conditions are met:
**– Correctness: A vote by attester (i, n) is correct if their vote is consistent with the canonical blockchain.**
Recall that the vote of attester (i, n) is given by ν(i,n) and the eventual canonical status of the block is given
by χn; thus, this condition is equivalent to ν(i,n) = χn.
**– Freshness: A vote by attester (i, n) is fresh if it was received by proposer n + 1 soon enough that it could**
be included in the block in slot n + 1 and the block in slot n + 1 is eventually made canonical. We let
_δ(i,n),n+1 ∼_ _exp(θ[−][1]) denote the random communication time between attester (i, n) and proposer n + 1,_
implying that the first part of this condition equates with τ(i,n) + δ(i,n),n+1 _tn+1. Moreover, the second_
_≤_
part of this condition equates with χn+1 = 1.
For exposition, we normalize the pay-off for attester (i, n) to unity, implying that their pay-off function is
given explicitly as follows:
_U_ _[A](ν(i,n), τ(i,n)) =_
1 if ν(i,n) = χn, τ(i,n) + δ(i,n),n+1 _tn+1, χn+1 = 1_
_≤_
(5)
0 otherwise
-----
p g,, y,,,
### 3 Analysis
**3.1** **Equilibrium analysis**
There exists a multiplicity of Nash equilibria. In particular, attesters can coordinate to implement proposers
acting at any particular time ∆[⋆] _∈_ [0, ∆] within the slot. Formally, we have the following result:
**Proposition 1. Multiple Equilibria**
_For any ∆[⋆]_ _∈_ [0, ∆], there exists an equilibrium as follows:
_Proposer n selects tn as follows:_
_tn = n · ∆_ + ∆[⋆] (6)
_and selects φn as follows:_
_φn =_
_Attester (i, n) selects ν(i,n) as follows:_
�
1 _if tn−1 ≤_ (n − 1) · ∆ + ∆[⋆] (7)
0 _otherwise_
_ν(i,n) =_
�
1 _if (6) and (7) hold_
(8)
0 _otherwise_
_and selects τ(i,n) as follows:_
_τ(i,n) =_
�
_tn + δn,(i,n)_ _if (6) and (7) hold_
(9)
_n · ∆_ _otherwise_
Proposition 1 arises because a proposer receives a zero pay-off unless her block earns sufficiently many
attestations. In turn, if attesters coordinate on voting for a proposer’s block only if the proposer releases her
block at a particular time, then the proposer earns a strictly positive pay-off only if she releases her block at
that particular time. Thus, since a proposer prefers a strictly positive pay-off to a zero pay-off, each proposer
optimally releases her block at the release time on which attesters coordinate.
As an aside, we emphasize that the referenced coordination by attesters is equilibrium behavior. In particular,
an attester receives a strictly positive pay-off only if her attestation is correct, and her attestation is correct
only if it agrees with the majority of attesters in her slot. As a consequence, when all other attesters vote in
one direction, each attester optimally votes in that same direction to avoid a zero pay-off.
_Proof._
We begin by establishing that (8) - (9) are optimal responses for any attester (i, n). Formally, we take as given
that all attesters other than (i, n) follow the equilibrium actions (8) - (9) and also that all proposers follow the
equilibrium actions (6) - (7); in that context, we demonstrate that (8) - (9) maximize (5) and thus these are
equilibrium actions for each attester (i, n).
If (6) and (7) hold, then φn = 1 follows directly for all n ∈ N. Moreover, if all attesters other than (i, n)
follow (8), then (6) and (7) imply ν(−i,n) = 1 which implies _A[˜]n = 1 for all n ∈_ N. Then, since (6) and (7) imply
_φn = 1 for all n ∈_ N and also _A[˜]n = 1 ≥_ _γ, (3) therefore implies χn = 1 for all n ∈_ N. In turn, since ν(i,n) ̸= χn
implies the lowest possible pay-off in (5), we have that ν(i,n) = χn = 1 whenever (6) and (7) holds. If (6) and (7)
do not hold, then (8) implies ν(−i,n) = 0 which implies _A[˜]n = 0 for all n ∈_ N. Moreover, (3) implies χn = 0 for
all n ∈ N. In turn, since ν(i,n) ̸= χn implies the lowest possible pay-off, we have that ν(i,n) = χn = 0 whenever
the conjunction of (6) and (7) do not hold. Thus, ν(i,n) = 1 is an optimal response if (6) and (7) and ν(i,n) = 0
is an optimal response otherwise, thereby establishing (8) as the equilibrium action for any attester (i, n).
To establish (9) as an optimal response for attester (i, n), note that (5) pointwise decreases in τi,n and thus
it is optimal to set τ(i,n) as low as possible subject to feasibility. In general, τ(i,n) ≥ _n · ∆_ but ν(i,n) = 1 =⇒
_τ(i,n) ≥_ _tn + δn,(i,n) = n · ∆_ + ∆[⋆] + δn,(i,n) > n · ∆. As such, whenever ν(i,n) = 0, then τ(i,n) = n · ∆, whereas
whenever ν(i,n) = 1, then τ(i,n) = tn + δn,(i,n). Then, as per our proof of (8), (6) and (7) imply ν(i,n) = 1 which
-----
y g g
implies τ(i,n) = tn + _δn,(i,n), whereas if either (6) or (7) does not hold, then ν(i,n) = 0 which implies τ(i,n) = n_ _·_ _∆,_
which thereby establishes (9).
We conclude by demonstrating that (6) - (7) are an optimal response for any proposer n. More formally, we
take as given that all attesters follow the equilibrium actions (8) - (9) and also that all proposers other than
proposer n follow the equilibrium actions (6) - (7); in this context, we establish that (6) - (7) maximize (5) and
thus these are equilibrium actions for each proposer n.
Due to (8), any deviation in (6) or (7) implies ν(i,n) = 0 for all (i, n) which further implies _A[˜]n = 0. Then,_
under such a deviation, (3) implies χn = 0 which implies a zero pay-off as per (1). Finally, since pay-offs are
bounded below by zero, not deviating from (6) and (7) necessarily produces a higher pay-off than any such
deviation and thus (6) and (7) are equilibrium actions.
**3.2** **Model justification**
The model presented in Section 2 is an idealized description of a blockchain consensus mechanism. A sequence
of proposers is selected, each of which is given the right to produce a block for the slot they are assigned to
in the sequence. Once the block is released, a set of attesters assigned for the current slot gets to vote for the
presence or absence of the block.
When the proposer chooses to build on the previous block, they affirm its place in the canonical chain.
There is no block tree: either the current proposer recognizes the block produced by the proposer before them
as part of the canonical chain (φ = 1), or they recognize that the previous proposer failed to produce a block
which is part of the canonical chain (φ = 0). With the assumption of a continuum of attesters, at equilibrium,
sufficiently many votes reach the following proposer, allowing them to make the call on whether or not the
previous proposer’s block is canonical.
This model resembles the Streamlet protocol [13]. A proposer submits a block for consideration to the rest
of the network. If γ = 2/3 share of attesters vote the block in, the block is notarized. If attesters do not,
e.g., because the block is unavailable, the chain height is not increased, but the next slot starts, giving the
opportunity to the next block producer to submit a block for consideration. Leaders extend the longest chain
of notarized blocks they have seen.
The model also bears resemblance with the proposed (block, slot) fork choice rule of the Ethereum Gasper
protocol [1], specifically the dynamically available chain produced by the protocol, when γ = 1/2. In this model
of the fork choice, attesters submit a vote attesting to the presence or absence of a block at some given slot.
The canonicity of a block is however complicated by the LMD-GHOST rule for block weight accumulation.
Obtaining more than half of the attesters’ vote may then neither be a sufficient nor a necessary condition to be
part of the canonical chain.
Generally, we formulate the hypothesis that most Proof-of-Stake-based leader selection protocols will be
exposed to timing games. As long as duties are assigned according to an absolute (wall-clock) time schedule,
there exists no pressure to complete duties in a timely manner comparable to the random arrival process of
leaders in Proof-of-Work. For instance, PBFT-based finalization protocols such as Tendermint [21] or HotStuff
[31] do not perform a view change until some timeout is reached, which a leader may use to time their release
appropriately. While a sufficiently decentralized committee of validators is an existing feature of these protocols,
our model further highlights its role in enforcing timeliness at equilibrium, as described in Section 3.1.
### 4 An empirical case study: Ethereum
Following a formal analysis of the coordination game between proposers and attesters, we now investigate the
occurrence of such strategic timing games in real-world systems. To this end, we examine Ethereum, an ideal
candidate for the empirical analysis of potential timing games, owing to its mature MEV market structure and
the availability of accessible, informative data points.
We show that timing games are indeed worth playing. However, we find that proposers do not delay their
block release with the intention to capture more MEV. Instead, we find that delays are mostly due to latency
in their signing processes. Thus, we can conclude that timing games are rational to engage in, but do not yet
occur to their full possible extent.
-----
p g,, y,,,
**4.1** **Consensus mechanism**
The Ethereum consensus mechanism is a composite of two protocols: variants of LMD GHOST [29] and Casper
FFG [9], often referred to together as Gasper [10]. In this paper, we focus exclusively on Ethereum’s available
_chain that is built roughly following LMD GHOST. This is because timing games only occur on the available_
chain. Within this protocol, time progresses in 12-second slots [3]. For each slot, one consensus participant,
referred to as a validator, is selected as the block proposer. According to the honest validator specifications [4],
which define the rules for honest protocol participation, a block should be released at the beginning of the slot
(0 seconds into the slot). Furthermore, the protocol selects a committee of attesters from the validator set who
vote on what they consider to be the latest canonical block as soon as they hear a valid block for their assigned
slot, or 4 seconds into the slot, whichever comes first [4]. We refer to this 4-second mark as the attestation
_deadline. This dynamic, in which block proposers must release their block early enough for attesters to receive_
it via the peer-to-peer network before the attestation deadline, results from the attestation deadline serving as
a coordination Schelling point [27]. It is worth noting that the honest validator specification prescribes block
proposers to release their block at the beginning of the slot, while attesters only attest 4 seconds into the slot
(unless a valid block is heard prior to the attestation deadline). This opens up room for block proposers to
release their block strategically—i.e., as late as possible while ensuring they accumulate a sufficient share of
attestations.
**4.2** **Block production process**
To assess the potential benefits of timing games for block proposers, it is important to comprehend the value of
time and the process by which MEV opportunities are captured in the block proposing process. In Ethereum,
the MEV market structure evolved and matured significantly over time, turning the block production process
into an intricate interplay between specialized actors [30]. This division of labor enables validators to profit from
MEV without engaging in the complex process of identifying MEV opportunities themselves. Instead, validators
can outsource the task of building a maximally profitable block to an out-out-protocol block auction process
known as MEV-Boost [5].
_Searchers look for MEV opportunities (e.g., arbitrages), and submit bundles of transactions alongside bids_
to express their order preference to block builders. Block builders, in turn, specialize in packing maximally
profitable blocks using searcher bundles and other available user transactions before submitting their blocks
with bids to relays. Relays act as trust facilitators between block proposers and block builders, validating blocks
received by block builders and forwarding only valid headers to validators. This ensures validators cannot steal
the content of a block builder’s block, but can still commit to proposing this block by signing the respective
block header. In the long run, Ethereum’s plans include enshrining this currently out-of-protocol mechanism
into the protocol [7,8] to eliminate relay trust assumptions. It is worth noting that MEV-Boost is an opt-in
protocol, and validators can always choose to revert to local block building. Finally, when a validator is selected
to propose a block in a given slot, they request the highest-bidding block header from the relay, sign it, and
return the signed block header to the relay, which then releases the block to the peer-to-peer network.
In summary, searchers find MEV opportunities and express their transaction-ordering preferences within
a block via bids. Block builders aim to build maximally profitable blocks using searcher bundles and user
transactions, then submit their block content and bids to relays. Validators ultimately request the highestpaying block header, sign it and return it to relays, which release the signed block to the peer-to-peer network.
Due to competition at all levels in this block production process (except for block proposing monopoly), the
block proposer is able to capture most of the MEV via this block auction.
**MEV-Boost block auction Here, we granularly outline the sequence of events that take place during the**
block construction of MEV-Boost block auctions on the Ethereum network. Figure 1 illustrates these events
along with their corresponding timestamps, and is intended to serve as a reference for the remainder of this
empirical analysis.
The auction for block of slot n begins in slot n − 1 (at t = −12000ms), during which builders submit blocks
alongside bids to relays. This competitive process between block builders determines the right to construct the
block for slot n and secures potential MEV-derived profits (block building profit equates to extracted MEV
minus bid value). For each bid, the relay logs the timestamps of events at which the bid was received by the
relay (receivedAt). After some validity checks are completed by the relay, the bid is made available to the
-----
y g g
proposer (eligibleAt). When the proposer chooses to propose a block [5], the proposer requests getHeader to
receive the highest bidding, eligible block header from the relay. Upon receiving the header associated with the
winning bid, the proposer signs it and thereby commits to proposing this block built by the respective builder
in slot n. The signed block header is sent to the relay, along with a request to get the full block content from the
relay (getPayload). Finally, the relay receives the signed block header (signedAt) and publishes the full block
contents to the peer-to-peer network and proposer. As soon as peers see the new block, validators assigned to
the slot can attest to it. This cycle completes one round of consensus repeating every slot.
Fig. 1: Logical representation of the block production process for slot n. Builder bids begin streaming in during
slot n − 1, after which the proposer and relay interact through requests and responses.
**4.3** **Data sets**
The analysis utilizes data provided by the ultra sound relay from March 4, 2023, to April 11, 2023. This covers
just under 185,000 slots, interspersed from slot 5,965,398 to slot 6,282,397, and includes all bids placed by block
builders through this relay. There were over 800 bids per slot, for a total of over 150 million bids. The winning
block originated from the ultra sound relay for nearly 85,000 of these slots, and so we measure timestamps
and other properties for those slots when investigating winning bids specifically. Finally, we augmented the
winning slots with various on-chain measures from the execution layer (EL) and consensus layer (CL), such as
attestations and aggregations, using a combination of analytical tools like Dune and direct observation of the
peer-to-peer network.
**4.4** **Are timing games worth playing?**
**Marginal value of time Timing games offer potential for substantial profit due to the increased MEV**
opportunities they provide. First, we assess whether timing games are worth playing for proposers, by estimating
the incremental MEV gained per second. We utilize all bids submitted by builders from the ultra sound relay
to examine the relationship between the timestamp at which the relay received a bid submitted by a builder
(receivedAt timestamp relative to the slot boundary) and the bid value, residualized against slot fixed effects
to account for differences between low- and high-MEV regimes and other unobservable forms of heterogeneity.
5 An honest participant will request the block header shortly before slot n such that the block can be released on time,
at the beginning of slot n (t = 0ms).
-----
p g,, y,,,
We then fit a regression line to this relationship, obtaining a slope with a coefficient of 0.0065 ETH per second,
which represents our estimate for the marginal value of time. Figure 2 depicts the linear increase in median bid
values over the slot duration on a point-by-point basis, and the distribution of bid receival times, indicating
that most bids are submitted between four seconds before the slot boundary to one second after. This analysis
shows there exists a positive marginal value of time, indicating that a rational block proposer would participate
in timing games.
Fig. 2: Analysis of bid values and their distribution over slot duration. The histogram (in blue) shows the
distribution of bid counts across time in seconds. The dark green line represents the median bid value in Ether
(ETH) for each time bin (with its associated IQR in green), residualized against the slot fixed effects that are
estimated in a linear regression of bid on timestamp (dashed red line). The x-axis shows time in milliseconds
relative to the slot boundary, the left y-axis displays the residualized bid value in ETH, and the right y-axis
displays the count of bids.
**4.5** **Are block proposers playing timing games?**
Having shown that timing games are worth playing, we turn our attention to whether proposers are currently
taking advantage of the opportunity to accumulate more MEV by committing to a bid later than foreseen by
the honest validator specifications.
**Characterizing late block signing behavior First, we investigate whether block headers and associated**
bids are signed by proposers later than the slot boundary (t = 0), the time stipulated by the honest protocol
specifications to broadcast their block to the network. We observe that winning bids are signed by proposers
approximately 774 ms after the slot boundary (t(111573) = 575.5, p < 1×10[−][20], using a two-tailed paired Student
_t test) and about 513 ms after the relay made the bid eligible (t(111573) = 472.6, p < 1×10[−][20], using a two-tailed_
paired Student t-test). Figure 3a displays the distribution of timings for winning bids, based on ultra sound relay
timestamps for bid reception from the builder (receivedAt, median = 157ms), eligibility for proposer signing
(eligibleAt, median = 260ms), and the actual signing by proposers (signedAt, median = 774ms). To better
understand the reasons behind late-signing behavior by proposers, we map validator public keys to their staking
entities and CL clients, see Figure 3b and 3c respectively). Validator to staking entity mappings were obtained
via a combination of open source data sets [6], and validator to client mappings were obtained using blockprint
[6 Dune Spellbooks: https://dune.com/spellbook, Mevboost.pics Open Data: https://mevboost.pics/data.html](https://dune.com/spellbook)
-----
y g g
[2], an open source tool assigning client labels to validators based on their attestation packing on the Ethereum
beacon chain. We found that staking entities such as Kraken (t = 38.9, p < 1 × 10[−][20]) and Coinbase (t = 67.6,
_p < 1 × 10[−][20]), as well as proposers using the Lodestar client (t = 44.9, p < 1 × 10[−][20]) sign block headers_
significantly later than other block proposer types (results were obtained using two-tailed unpaired Student ttests). Notably, additional analyses are required to differentiate the interdependencies between validator entities
and clients to better understand their roles in late signing behavior. This analysis confirms that proposers are
signing blocks significantly later than expected, but it does not yet clarify the underlying reasons, which could
include participation in timing games or increased latency for independent reasons, e.g., longer signing processes.
(a)
(c)
Fig. 3: Analysis of event timestamps and their distributions among Validator Clients and Entities. (3a) Multiple
Kernel Density Estimation (KDE) distributions of event timestamps from the relay data, showing the probability
density functions for three event types: receivedAt (blue), eligibleAt (green), and signedAt (light green).
(3b-3c) Violin plots comparing the distribution of signedAt event timestamps for the top 7 validator entities and
clients. The x-axis represents time in milliseconds (ms) relative to the slot boundary, while the y-axis displays
validator clients and entities, ordered by the mean signedAt time. The width of each violin plot signifies the
kernel density estimation of the signedAt event timestamps, demonstrating the distribution and frequency of
the events within each group.
We subsequently collected data for 1241 slots (slot 6,200,251 to 6,204,957 on April 11, 2023) and used the
time difference between getHeader and getPayload calls by proposers as an approximation to estimate the duration of the signing process (see Figure 1). Figure 4 shows that the median difference between getHeader
and getPayload call is 418 ms. Interestingly, this delay, attributable to the signing process, accounts for
75.42% of the overall latency. This percentage was determined by calculating the difference between the signing time and the moment the bid was deemed eligible by the relay on a slot-by-slot basis, using the formula
median(getPayload−getHeader)
median(signedAt−eligibleAt) _[×][100][. We conclude that late signing behavior is primarily attributed to latency caused]_
by the signing process, rather than intentional delays to incorporate more MEV in blocks. This finding aligns
with the hypothesis that large US-based staking entities, such as Coinbase and Kraken, may prefer utilizing
-----
p g,, y,,,
sophisticated remote secure signing mechanisms, resulting in a lengthier signing process compared to other
parties.
(a) (b)
Fig. 4: Estimating the latency induced by the signing process. (4a) Histogram of getHeader and getPayload
call timestamps relative to slot boundary. The histogram displays the density of events occurring at different
times into the slot (in milliseconds) for getHeader (yellow) and getPayload (blue) calls. (4b) Histogram of the
time difference between getHeader and getPayload calls. The histogram shows the density of time differences
(in milliseconds) getHeader and getPayload calls. Vertical lines represent the 50[th] (solid), 90[th] (dashed), and
99[th] (dotted) percentiles of the distribution.
**4.6** **The impact of latency on the peer-to-peer network**
Our prior results indicate that validators are not engaging in timing games to accrue more MEV. Nonetheless,
we assess the implications of late signing of consensus messages on the peer-to-peer network. Specifically, we
examine the relationship between the relay timestamps, the timings at which blocks are (1) first seen by the rest
of the peer-to-peer network and (2) begin collecting attestations and aggregations. The consensus layer data was
obtained through nodes run by the Ethereum Foundation, for 2643 slots (slots 6,357,601 to 6,363,807 on May 3
and 4, 2023). Figure 5a shows the sequence of these event timestamps over the course of a slot. We subsequently
assess the correlations between each of these event pairs, as depicted in Figure 5b. Our analysis reveals high
correlations between the time at which blocks are signed by proposers (i.e., the signedAt relay timestamp) and
the time at which blocks (correlation coefficient = 0.986) and attestations (correlation coefficient = 0.971) are
initially observed by the peer-to-peer network. These findings underscore the significance of proposers signing
blocks promptly, as it considerably impacts the downstream processes at the consensus layer in the network.
Next, we evaluate the impact of latency induced by late signing behavior on attestations collected by winning
blocks proposed to the peer-to-peer network. We examine the relationship between the time at which blocks are
signed by proposers (signedAt), and the share of attestations included by blocks in their respective target slot
referred to as slot n in Figure 1. As a reminder, attestations collected on a given slot n are only included on-chain
one (slot n + 1) or more slots later. In our analysis, we focus on the attestations included in the subsequent slot
and compute a metric next-slot shares. This metric refers to the percentage of attestations for the winning block
in a given slot that appear in the next block (slot n + 1), out of the total number of attestations in the next
slot that refer to any block in the target slot. Our hypothesis is that if a block is signed too late by a proposer,
it will not propagate early enough for attesters to vote for it before their attestation deadline (t = 4000ms, see
Figure 1, and [4]). Hence, in such settings attesters vote for another block (e.g., the parent block), and this will
be reflected in the next-slot shares metric.
Figure 6 shows that latency does indeed cause a steep drop-off in the share of attestations received by the
winning block. We observe that the share value stays close to one as long as the block is signed within the
-----
y g g
(a) (b)
Fig. 5: Analysis of relay and consensus layer timestamps. (5a) Box plot of the time differences between relay and
consensus timestamps. The box plots display the distribution of time differences for receivedAt, elligibleAt,
and signedAt events, as well as blocks, attestations, and aggregations first seen by the peer-to-peer network.
The boxes represent the interquartile range (IQR) from the first quartile (Q1, 25[th] percentile) to the third
quartile (Q3, 75[th] percentile), while the whiskers extend to the minimum and maximum values within three
times the IQR. The horizontal lines within the boxes represent the median values. (5b) Bar plot of Pearson
correlation coefficients for each pair of event timestamps. The bars represent the mean correlation coefficient
for each relationship, while the error bars represent the 95% confidence intervals obtained via bootstrapping.
first two seconds of the slot. Once the two second threshold is crossed, there is a substantial drop-off and many
winning blocks earn fewer than half of the next-slot attestations, which continues to rapidly decrease towards
zero as we approach the theoretical t = 4000ms attestation deadline. These results demonstrate the impact of
latency on the rest of the peer-to-peer network and highlight the importance of signing and broadcasting blocks
on time to prevent missed slots and reorganizations.
We previously documented a private incentive for proposers to delay their block release, according to the
steady increase of MEV as time progresses through the slot. Such malicious behavior is not prevalent, as 75%
and 98% of all blocks are seen by our nodes after two and four seconds into the slot, respectively. Yet, our
analysis reveals on-chain evidence whenever latency degrades consensus formation.
### 5 Discussion
In this paper, we present an argument that consensus participants are subject to exogenous incentives, primarily
MEV, that exist outside the consensus mechanism itself. This highlights the imperative for blockchain protocols
to ensure economic fairness among all consensus participants. Specifically, it necessitates a design where honest
and honest-but-rational consensus participation become indistinguishable, and honesty within the protocol is
the most profitable strategy. This approach ensures that honest-but-rational participants have no incentive to
deviate from honest consensus participation.
We present a model that highlights the time-dependent value for consensus participants and probes into the
strategic timing considerations that block proposers face. Our model uncovers a spectrum of equilibria wherein
attesters can enforce any deadline for block proposals to achieve canonical status, thereby emphasizing the
crucial role of Schelling points as coordination mechanisms. For instance, in the Ethereum network we observe
the emergence of such Schelling points through the default settings of client software. The widespread use of
these default settings among consensus participants generally ensures their effectiveness.
We support our theoretical findings by observations of the Ethereum network. Our analysis demonstrates
that timing games are indeed worth playing for block proposers, enabling them to capture additional MEV by
delaying their block proposals beyond the timeframe prescribed by the honest validator specification. However,
we observe that current instances of delayed block proposals are primarily due to latency in the block signing
process, rather than a conscious strategy to maximize profits. The apparent lack of maximal MEV capture by
-----
p g,, y,,,
Fig. 6: Effects of block signing times on next-slot attestations. This scatter plot features the x-axis displaying the
time at which proposers signed the winning block relative to the slot n boundary, and the y-axis illustrating the
share of next-slot (n + 1) attestations for the winning block. Each point on the graph corresponds to the time
(in milliseconds) at which the winning block was signed within the slot and the average share of attestations it
received, included in the next slot, across all winning blocks signed that specific time.
honest proposers could be attributed to either a lack of common knowledge, existing social norms around this
practice. It’s clear, however, that these are not sustainable safeguards for maintaining economic fairness.
The implications of timing games are manifold and significant. An honest-but-rational participant who
engages in timing games will outperform honest participants, leading to a centralization of stake over time.
Hypothetically, this could culminate in a breach of consensus security. In a more practical sense, it may encourage
individual stakers to delegate their stake to professional entities adept at these practices, negatively impacting
the network’s decentralization. Moreover, timing games can overload the messaging system within a short time
span, potentially causing cascading failures at the peer-to-peer layer, particularly within client systems.
Essentially, timing games are facilitated by the monopolistic right that block proposers possess for a single
round of consensus. Introducing competition in block proposing, similar in effect to the exogenous randomness
in Proof of Work (PoW) systems, emerges as a potential solution. However, the challenge lies in deterministically selecting a winning proposer, or reverting to peer-to-peer latency races, which in itself is centralizing.
Alternatively, an on-chain heuristic for timely block proposals could incentivize timely participation, yet the
allure of MEV rewards might still outweigh any in-protocol consensus rewards. Tackling the root cause of timing
games remains an open challenge.
In the Ethereum context, a late-block reorging mechanism has been adopted in the fork choice, effectively
imposing a 4-second deadline for block proposers. This constraint significantly limits the extent to which block
delays are possible. Looking ahead, the adoption of (block, slot) type of attestations is likely, further refining the
protocol. However, it remains challenging to address the root cause of timing games, as it is deeply intertwined
with the fundamental workings of Proof of Stake (PoS). Although limiting the length of the proposer’s interval
is feasible, completely eliminating the monopolistic market structure of block proposers proves to be a difficult
task.
Consequently, it may prove valuable to find a more general abstraction for PoS type of protocols and further
explore the implications of consensus participants being exposed to incentives outside of consensus itself, such
as MEV. More generally, assuming honest-but-rational as opposed to honest type of consensus participation
should prove significant in designing economically fair blockchain protocols.
-----
y g g
### Acknowledgments
The authors acknowledge helpful discussions and comments from Francesco d’Amato and Anders Elowsson. We
also appreciate the significant contributions of Mike Neuder in obtaining the necessary data for this study.
### References
[1. (block, slot) fork choice. https://github.com/ethereum/consensus-specs/pull/2197, accessed: 2023-10-05](https://github.com/ethereum/consensus-specs/pull/2197)
[2. blockprint, https://github.com/sigp/blockprint, accessed: 2023-10-05](https://github.com/sigp/blockprint)
[3. Ethereum consensus specifications - beacon chain, https://github.com/ethereum/consensus-specs/blob/dev/specs/](https://github.com/ethereum/consensus-specs/blob/dev/specs/phase0/beacon-chain.md)
[phase0/beacon-chain.md, accessed: 2023-10-05](https://github.com/ethereum/consensus-specs/blob/dev/specs/phase0/beacon-chain.md)
[4. Ethereum consensus specifications - honest validator, https://github.com/ethereum/consensus-specs/blob/dev/](https://github.com/ethereum/consensus-specs/blob/dev/specs/phase0/validator.md)
[specs/phase0/validator.md, accessed: 2023-10-05](https://github.com/ethereum/consensus-specs/blob/dev/specs/phase0/validator.md)
[5. Mev-boost, https://github.com/flashbots/mev-boost, accessed: 2023-10-05](https://github.com/flashbots/mev-boost)
6. Babel, K., Daian, P., Kelkar, M., Juels, A.: Clockwork finance: Automated analysis of economic security in smart
contracts. arXiv preprint arXiv:2109.04347 (2021)
[7. Buterin, V.: Proposer/block builder separation-friendly fee market designs, https://ethresear.ch/t/proposer-block-](https://ethresear.ch/t/proposer-block-builder-separation-friendly-fee-market-designs/9725)
[builder-separation-friendly-fee-market-designs/9725, accessed: 2023-10-05](https://ethresear.ch/t/proposer-block-builder-separation-friendly-fee-market-designs/9725)
[8. Buterin, V.: Two-slot proposer/builder separation, https://ethresear.ch/t/two-slot-proposer-builder-separation/](https://ethresear.ch/t/two-slot-proposer-builder-separation/10980)
[10980, accessed: 2023-10-05](https://ethresear.ch/t/two-slot-proposer-builder-separation/10980)
[9. Buterin, V., Griffith, V.: Casper the friendly finality gadget. arXiv:1710.09437 [cs.CR] (2019), https://arxiv.org/abs/](https://arxiv.org/abs/1710.09437)
[1710.09437](https://arxiv.org/abs/1710.09437)
10. Buterin, V., Hernandez, D., Kamphefner, T., Pham, K., Qiao, Z., Ryan, D., Sin, J., Wang, Y., Zhang, Y.X.: Combining
ghost and casper. arXiv preprint arXiv:2003.03052 (2020)
11. Cachin, C., Vukolić, M.: Blockchain consensus protocols in the wild. arXiv preprint arXiv:1707.01873 (2017)
12. Carlsten, M., Kalodner, H., Weinberg, S.M., Narayanan, A.: On the instability of bitcoin without the block reward.
In: Proceedings of the 2016 acm sigsac conference on computer and communications security. pp. 154–167 (2016)
13. Chan, B.Y., Shi, E.: Streamlet: Textbook streamlined blockchains. In: Proceedings of the 2nd ACM Conference on
Advances in Financial Technologies. pp. 1–11 (2020)
14. Daian, P., Goldfeder, S., Kell, T., Li, Y., Zhao, X., Bentov, I., Breidenbach, L., Juels, A.: Flash boys 2.0: Frontrunning,
[transaction reordering, and consensus instability in decentralized exchanges. CoRR abs/1904.05234 (2019), http:](http://arxiv.org/abs/1904.05234)
[//arxiv.org/abs/1904.05234](http://arxiv.org/abs/1904.05234)
15. D’Amato, F., Neu, J., Tas, E.N., Tse, D.: No more attacks on proof-of-stake ethereum? arXiv preprint
arXiv:2209.03255 (2022)
16. D’Amato, F., Zanolini, L.: Recent latest message driven ghost: Balancing dynamic availability with asynchrony
resilience. arXiv preprint arXiv:2302.11326 (2023)
17. D’Amato, F., Zanolini, L.: A simple single slot finality protocol for ethereum. arXiv preprint arXiv:2302.12745 (2023)
18. Dembo, A., Kannan, S., Tas, E.N., Tse, D., Viswanath, P., Wang, X., Zeitouni, O.: Everything is a race and nakamoto
always wins. In: Proceedings of the 2020 ACM SIGSAC Conference on Computer and Communications Security. pp.
859–878 (2020)
19. Eyal, I., Sirer, E.G.: Majority is not enough: Bitcoin mining is vulnerable. Communications of the ACM 61(7),
95–102 (2018)
20. John, K., Rivera, T., Saleh, F.: Economic implications of scaling blockchains: Why the consensus protocol matters.
NYU Stern Working Paper (2023)
21. Kwon, J.: Tendermint: Consensus without mining. Draft v. 0.6, fall 1(11) (2014)
22. Liao, K., Katz, J.: Incentivizing blockchain forks via whale transactions. In: Financial Cryptography and Data
Security: FC 2017 International Workshops, WAHC, BITCOIN, VOTING, WTSC, and TA, Sliema, Malta, April 7,
2017, Revised Selected Papers 21. pp. 264–279. Springer (2017)
[23. Monnot, B.: Timing games in proof-of-stake, https://ethresear.ch/t/timing-games-in-proof-of-stake/13980, accessed:](https://ethresear.ch/t/timing-games-in-proof-of-stake/13980)
2023-10-05
24. Neu, J., Tas, E.N., Tse, D.: Ebb-and-flow protocols: A resolution of the availability-finality dilemma. In: 2021 IEEE
Symposium on Security and Privacy (SP). pp. 446–465. IEEE (2021)
25. Neu, J., Tas, E.N., Tse, D.: Two more attacks on proof-of-stake ghost/ethereum. In: Proceedings of the 2022 ACM
Workshop on Developments in Consensus. pp. 43–52 (2022)
26. Saleh, F.: Blockchain without waste: Proof-of-stake. Review of Financial Studies 34(3), 1156–1190 (2021)
27. Schelling, T.C.: The Strategy of Conflict: with a new Preface by the Author. Harvard university press (1980)
28. Schwarz-Schilling, C., Neu, J., Monnot, B., Asgaonkar, A., Tas, E.N., Tse, D.: Three attacks on proof-of-stake
ethereum. In: Financial Cryptography and Data Security: 26th International Conference, FC 2022, Grenada, May
2–6, 2022, Revised Selected Papers. pp. 560–576. Springer (2022)
-----
p g,, y,,,
29. Sompolinsky, Y., Zohar, A.: Secure high-rate transaction processing in Bitcoin. In: International Conference on
Financial Cryptography and Data Security. pp. 507–527. Springer (2015)
30. Yang, S., Zhang, F., Huang, K., Chen, X., Yang, Y., Zhu, F.: Sok: Mev countermeasures: Theory and practice (2022).
[https://doi.org/10.48550/ARXIV.2212.05111, https://arxiv.org/abs/2212.05111](https://doi.org/10.48550/ARXIV.2212.05111)
31. Yin, M., Malkhi, D., Reiter, M.K., Gueta, G.G., Abraham, I.: Hotstuff: Bft consensus with linearity and responsiveness. In: Proceedings of the 2019 ACM Symposium on Principles of Distributed Computing. pp. 347–356 (2019)
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/2305.09032, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "GREEN",
"url": "http://arxiv.org/pdf/2305.09032"
}
| 2,023
|
[
"JournalArticle"
] | true
| 2023-05-15T00:00:00
|
[
{
"paperId": "3d6d7416b8249c2dc6f47ec127274ff0d63c4a9b",
"title": "A Simple Single Slot Finality Protocol For Ethereum"
},
{
"paperId": "e04584b4e478c93f2137d83b768d1a1caaa7c50b",
"title": "Recent Latest Message Driven GHOST: Balancing Dynamic Availability With Asynchrony Resilience"
},
{
"paperId": "71495ab57440c34578f5a01813699b98fb3f984d",
"title": "Two More Attacks on Proof-of-Stake GHOST/Ethereum"
},
{
"paperId": "8b647a12f0ff831db7e6489b1166710841d66a88",
"title": "Goldfish: No More Attacks on Ethereum?!"
},
{
"paperId": "36382648cb6df633c981e91cc557c90575781c30",
"title": "Three Attacks on Proof-of-Stake Ethereum"
},
{
"paperId": "647fa7aaa4eac4fafa56ac86ebafc5cf5944a8eb",
"title": "Clockwork Finance: Automated Analysis of Economic Security in Smart Contracts"
},
{
"paperId": "d4b4f0c642358004982e75eb85da9d6fbcb11602",
"title": "Economic Implications of Scaling Blockchains: Why the Consensus Protocol Matters"
},
{
"paperId": "ea17b3d47f1d95a1c4215da19cd7efeefd7f5b1a",
"title": "Streamlet: Textbook Streamlined Blockchains"
},
{
"paperId": "7ebc49a3bc102c24a46bf79ce94b3213711fe2d3",
"title": "Ebb-and-Flow Protocols: A Resolution of the Availability-Finality Dilemma"
},
{
"paperId": "03d1b883e9d8474212094e5764646bc6450cf565",
"title": "Blockchain Without Waste: Proof-of-Stake"
},
{
"paperId": "0a6b4af44cba14f11add0df09e5da33af0ff18e1",
"title": "Everything is a Race and Nakamoto Always Wins"
},
{
"paperId": "5c89ae258f209fc2ff3464b4c01e6e88eae10852",
"title": "Combining GHOST and Casper"
},
{
"paperId": "50dd47f615068b0eac1eeed60c91e95549aed3d4",
"title": "HotStuff: BFT Consensus with Linearity and Responsiveness"
},
{
"paperId": "393ab84a86631d5fda128c3aac0bf5476da07791",
"title": "Flash Boys 2.0: Frontrunning, Transaction Reordering, and Consensus Instability in Decentralized Exchanges"
},
{
"paperId": "fdbebd67c8a9671efabf4e53d6267789cd91d96c",
"title": "Casper the Friendly Finality Gadget"
},
{
"paperId": "26a286e447cd78227eddc801a5d9816e0215834b",
"title": "Blockchain Consensus Protocols in the Wild"
},
{
"paperId": "6d818bf1b05a07b79e73f5b29b4c5a34ad2bc456",
"title": "Incentivizing Blockchain Forks via Whale Transactions"
},
{
"paperId": "557d145a725e3b29336f004194e1e6bbb17fe8e9",
"title": "On the Instability of Bitcoin Without the Block Reward"
},
{
"paperId": "728b60c04afb5b87853b59265e49f430dbf631db",
"title": "Secure High-Rate Transaction Processing in Bitcoin"
},
{
"paperId": "86bf53b9c87265e42152d40a9d5ce6a5cd02f055",
"title": "Countermeasures"
},
{
"paperId": null,
"title": "Proposer/block builder separation-friendly fee market designs, https://ethresear.ch/t/proposer-blockbuilder-separation-friendly-fee-market-designs/9725"
},
{
"paperId": null,
"title": "Majority is not enough: Bitcoin mining is vulnerable"
},
{
"paperId": "df62a45f50aac8890453b6991ea115e996c1646e",
"title": "Tendermint : Consensus without Mining"
},
{
"paperId": null,
"title": "The Strategy of Conflict: with a new Preface by the Author"
},
{
"paperId": null,
"title": "Time is Money: Strategic Timing Games in Proof-of-Stake Protocols"
},
{
"paperId": null,
"title": "Two-slot proposer/builder separation"
},
{
"paperId": null,
"title": "Mev-boost"
},
{
"paperId": null,
"title": "Timing games in"
},
{
"paperId": null,
"title": "Ethereum consensus specifications - beacon chain"
},
{
"paperId": null,
"title": "blockprint"
}
] | 12,907
|
en
|
[
{
"category": "Medicine",
"source": "external"
},
{
"category": "Environmental Science",
"source": "s2-fos-model"
},
{
"category": "Political Science",
"source": "s2-fos-model"
},
{
"category": "Medicine",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/022ad80bb2cf8922a5d0b62cb56c4b7beb08450c
|
[
"Medicine"
] | 0.910237
|
Mitigation Planning and Policies Informed by COVID-19 Modeling: A Framework and Case Study of the State of Hawaii
|
022ad80bb2cf8922a5d0b62cb56c4b7beb08450c
|
International Journal of Environmental Research and Public Health
|
[
{
"authorId": "2110859392",
"name": "Thomas H. Lee"
},
{
"authorId": "2120238704",
"name": "Bobby Do"
},
{
"authorId": "2165752670",
"name": "Levi Dantzinger"
},
{
"authorId": "47889503",
"name": "Joshua R. Holmes"
},
{
"authorId": "143824683",
"name": "M. Chyba"
},
{
"authorId": "80974909",
"name": "Steven D Hankins"
},
{
"authorId": "69452483",
"name": "E. Mersereau"
},
{
"authorId": "118840200",
"name": "Kenneth S. Hara"
},
{
"authorId": "33931275",
"name": "V. Fan"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"Int J Environ Res Public Health"
],
"alternate_urls": null,
"id": "3096eb5c-d18c-4877-94cd-28edd3a9c357",
"issn": "1660-4601",
"name": "International Journal of Environmental Research and Public Health",
"type": "journal",
"url": "http://www.mdpi.com/journal/ijerph/"
}
|
In the face of great uncertainty and a global crisis from COVID-19, mathematical and epidemiologic COVID-19 models proliferated during the pandemic. Yet, many models were not created with the explicit audience of policymakers, the intention of informing specific scenarios, or explicit communication of assumptions, limitations, and complexities. This study presents a case study of the roles, uses, and approaches to COVID-19 modeling and forecasting in one state jurisdiction in the United States. Based on an account of the historical real-world events through lived experiences, we first examine the specific modeling considerations used to inform policy decisions. Then, we review the real-world policy use cases and key decisions that were informed by modeling during the pandemic including the role of modeling in informing planning for hospital capacity, isolation and quarantine facilities, and broad public communication. Key lessons are examined through the real-world application of modeling, noting the importance of locally tailored models, the role of a scientific and technical advisory group, and the challenges of communicating technical considerations to a public audience.
|
International Journal of
**_[Environmental Research](https://www.mdpi.com/journal/ijerph)_**
**_and Public Health_**
_Article_
# Mitigation Planning and Policies Informed by COVID-19 Modeling: A Framework and Case Study of the State of Hawaii
**Thomas H. Lee** **[1,2], Bobby Do** **[1], Levi Dantzinger** **[1], Joshua Holmes** **[1], Monique Chyba** **[3]** **, Steven Hankins** **[4],**
**Edward Mersereau** **[5], Kenneth Hara** **[6]** **and Victoria Y. Fan** **[1,7,]***
1 Thompson School of Social Work & Public Health, University of Hawaii at Manoa, Honolulu, HI 96822, USA;
tlee@hawaiidata.org (T.H.L.); bdo7@hawaii.edu (B.D.); levidantzinger@gmail.com (L.D.);
jrholmes@hawaii.edu (J.H.)
2 Hawaii Data Collaborative, Honolulu, HI 96813, USA
3 Department of Mathematics, College of Natural Sciences, University of Hawaii at Manoa,
Honolulu, HI 96822, USA; chyba@hawaii.edu
4 John A. Burns School of Medicine, University of Hawaii at Manoa, Honolulu, HI 96813, USA;
hankinss@hawaii.edu
5 Behavioral Health Administration, Hawaii Department of Health, Honolulu, HI 96813, USA;
phac@hawaii.edu
6 Hawaii Department of Defense, Honolulu, HI 96816, USA; kenneth.s.hara@hawaii.gov
7 Center for Global Development, Washington, DC 20036, USA
***** Correspondence: vfan@hawaii.edu
**Citation: Lee, T.H.; Do, B.;**
Dantzinger, L.; Holmes, J.; Chyba, M.;
Hankins, S.; Mersereau, E.; Hara, K.;
Fan, V.Y. Mitigation Planning and
Policies Informed by COVID-19
Modeling: A Framework and Case
Study of the State of Hawaii. Int. J.
_Environ. Res. Public Health 2022, 19,_
[6119. https://doi.org/10.3390/](https://doi.org/10.3390/ijerph19106119)
[ijerph19106119](https://doi.org/10.3390/ijerph19106119)
Academic Editor: Fernando Augusto
Lima Marson
Received: 29 March 2022
Accepted: 12 May 2022
Published: 18 May 2022
**Publisher’s Note: MDPI stays neutral**
with regard to jurisdictional claims in
published maps and institutional affil
iations.
**Copyright:** © 2022 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
[Attribution (CC BY) license (https://](https://creativecommons.org/licenses/by/4.0/)
[creativecommons.org/licenses/by/](https://creativecommons.org/licenses/by/4.0/)
4.0/).
**Abstract: In the face of great uncertainty and a global crisis from COVID-19, mathematical and**
epidemiologic COVID-19 models proliferated during the pandemic. Yet, many models were not
created with the explicit audience of policymakers, the intention of informing specific scenarios,
or explicit communication of assumptions, limitations, and complexities. This study presents a
case study of the roles, uses, and approaches to COVID-19 modeling and forecasting in one state
jurisdiction in the United States. Based on an account of the historical real-world events through lived
experiences, we first examine the specific modeling considerations used to inform policy decisions.
Then, we review the real-world policy use cases and key decisions that were informed by modeling
during the pandemic including the role of modeling in informing planning for hospital capacity,
isolation and quarantine facilities, and broad public communication. Key lessons are examined
through the real-world application of modeling, noting the importance of locally tailored models,
the role of a scientific and technical advisory group, and the challenges of communicating technical
considerations to a public audience.
**Keywords: COVID-19; pandemic; modeling; epidemiology; isolation and quarantine; media and**
communication; public health planning; governance; hospital; pandemic preparedness
**1. Introduction**
The health and economic toll of the global coronavirus disease 2019 (COVID-19)
pandemic posed unprecedented challenges for how public health authorities respond and
mitigate a public health emergency and crisis globally. Yet, the local public health response
to COVID-19 was challenging for many reasons, including the widespread uncertainty as
well as ever-evolving science and knowledge of a new disease, the wide-ranging mitigation
measures and health and socioeconomic impacts, the disproportionate impact on vulnerable
populations, the highly charged political context and polarized communication challenge,
the lack of preparedness and capacity for public sector response, among others [1–3].
However, in many countries, COVID-19 was mitigated through local or subnational efforts
and responses in addressing the multidimensional impacts of COVID-19. As a result, local
policymakers and public health authorities were challenged to make timely decisions and
deploy a variety of tools to best respond to the emergency.
-----
_Int. J. Environ. Res. Public Health 2022, 19, 6119_ 2 of 14
In many countries, one key tool used and deployed by policymakers was mathematical
modeling and epidemiologic forecasting of infectious diseases. There was widespread
use of a variety of models that forecasted and made predictions about the spread of
[COVID-19 and its subsequent health impacts. The COVID Forecast Hub (https://COVID1](https://COVID19forecasthub.org/)
[9forecasthub.org/ (accessed on 11 May 2022)) curated a partial list of more than 40 models,](https://COVID19forecasthub.org/)
predominantly from universities and research institutes and with a wide range of predictions on case counts, hospitalizations, intensive care unit use, and deaths. The landscape of
models can be dizzying for a non-technical user including policymakers.
Policymakers can use mathematical and epidemiologic modeling to inform a variety
of pressing policy, programmatic, and planning questions and decisions for mitigating
COVID-19. These decisions can be grouped into two main categories along the slopes of
the COVID-19 curve—the surge up and the decline down. Key questions for policymakers seeking to make decisions informed by best available science and evidence through
modeling include:
Given the nature of COVID-19 to surge exponentially, are there sufficient health care
_•_
resources in our jurisdiction or state, including hospital beds, ventilators, personal
protective equipment, medication, isolation and quarantine facilities, contact tracers,
and other health workers? Will capacity be sufficient, and if not, when will they run
out? Does the state need to enforce stronger mitigation measures including at its
extreme “shutdown”, i.e., mass quarantine and isolation for the state?
_•_ As COVID-19 ostensibly declines, what policy decisions should authorities undertake
to reopen and relax measures? What mitigation measures need to be maintained
and what measures can be dispensed including testing, tracing, isolation, masking,
distancing, and vaccination?
This study presents a real-world historical case study of the roles, uses, and approaches
to COVID-19 modeling and forecasting for policy decisions and policy use cases, drawing
from the historical perspectives in one state jurisdiction in the United States. This study does
not present a micro-level analysis of detailed modeling and its mathematical specifications,
but rather provides a macro-level historical and policy perspective on the ways in which
modeling informs policymaking. The methodology and data used for this case study
rely on a review of the historical facts and real-world events through lived experiences of
the authors of this paper who are members of the Hawaii Pandemic Applied Modeling
[Work Group (HiPAM) (https://www.hipam.org (accessed on 11 May 2022)), the Hawaii](https://www.hipam.org)
Department of Defense, Hawaii Emergency Management Agency (HI-EMA), or the Hawaii
Department of Health in a variety of roles during the COVID-19 pandemic from March 2020
to May 2022.
This case study is intended to and may help future policymakers seeking to navigate
this complex landscape of models and draw upon practical lessons learned on how to
make appropriate evidence-based decisions using models. As such, this paper is structured
as follows. The first part of this case study focuses on the major technical considerations
of mathematical and epidemiologic models and how models were selected given realworld limitations of time and resources. The second part of this case study reviews and
summarizes the real-world policy use cases and key policy decisions informed by modeling
during the pandemic surge and decline, including the role of modeling in informing
planning for hospital capacity and isolation and quarantine facilities, and deploying a broad
public communication strategy that navigates the complexities and pitfalls of modeling.
We then reflect on the key lessons and discussion from the use cases that may be relevant
for other jurisdictions seeking to use modeling to inform decision making.
**2. COVID-19 Models Used to Inform Policymakers in Hawaii**
_2.1. Model Selection_
A “model” refers to a mathematical or logical representation of the biology and
epidemiology of disease transmission and its associated processes [4]. To date, there are
more than 40 COVID-19 models available. In many locations around the world, there
-----
_Int. J. Environ. Res. Public Health 2022, 19, 6119_ 3 of 14
was a need for timely public health decisions pressing against the lack of available time,
resources, and a severe dearth of expertise such as epidemiologists in the jurisdiction (as
was the case in Hawaii). Thus, it was not feasible nor practical for policymakers and their
technical support teams to comprehensively review all models in order to make decisions.
Instead, policymakers were continuously forced to make strategic decisions to select and
use tools in order to make the best available information at hand. Nevertheless, even
the selection of tools in order to make policy decisions requires technical expertise and
communication savvy in order to wade through the science and complexity and minimize
creation of additional confusion.
The first part of this case study focuses on the major criteria for selecting a model.
Beginning in April of 2020, the HI-EMA tasked a technical team (including a physician,
a lead epidemiologic adviser and a technical analyst) who in turn sought guidance from
a newly developed HiPAM work group to review and use models to inform a variety of
specific policy decisions, described in the second part of this case study. The four models
that were ultimately chosen and used for informing Hawaii public authorities for decision
making were based on selective review rather than comprehensive review of models. The
four models used by the technical team were the following: the University of Washington Institute for Health Metrics and Evaluation (IHME) model [5], the Imperial College
London model [6], the Epidemic Calculator [7], and the University of Basel model [8]. The
dimensions for reviewing and selecting these models are described in Section 2.2.
Upon review of the documentation and source code, if available, of these models, in
2020, the technical team identified some of the key assumptions of these models. These
assumptions were crucial in understanding the limitations or applicability of a given model
to a particular jurisdiction, and in this case, the state of Hawaii. These assumptions and
limitations are discussed in Section 2.3.
_2.2. Criteria for Model Selection_
There are several criteria that could be considered for selecting a model. In this
case study, the technical team in Hawaii was prompted with questions from policymakers
relying on wide media coverage on two models in particular—the University of Washington
and the Imperial College London model. Yet, as the technical team discovered, these two
models were not completely suitable or customizable for the situation in the local state
jurisdiction. The technical team then identified two more models (Epidemic Calculator
and the University of Basel models) and reviewed these four models based on publicly
available documentation (noted in the aforementioned references), and in some cases, data
visualizations and source code, along five key dimensions. At the time, the COVID-19
modeling hub had not yet been available in the early part of the pandemic, and thus the
models chosen were selective and purposive.
The key dimensions used to select and use these four models were the following:
(1) model objective, (2) interactivity and local parameter customizability, (3) age distribution, (4) type of model, and (5) open source (see Table 1). Given limitations of time and
resources, the technical team made purposive decisions on which models to consider and
use in 2020, and compared and contrasted the models along these dimensions. These
dimensions were argued to be relevant for decision making in Hawaii based on the issues
of the assumptions and limitations of the models. While these are not comprehensive of all
considerations, they reflect the historical events in the Hawaii case.
**Model Objective. Each model had a different objective. The IHME model intended**
to estimate COVID-19 hospital impacts, whereas the Imperial College London model
sought to illustrate how public health measures such as physical distancing and protecting
vulnerable populations affected the spread of COVID-19. Understanding the objective of
the model is an important but incomplete aspect to its appropriate use.
**Local Parameter Customizability. Some models allowed for interactivity and cus-**
tomizability of the model parameters. The Epidemic Calculator had sliders to allow for a
user to modify parameters driving the transmission and clinical dynamics underpinning
-----
_Int. J. Environ. Res. Public Health 2022, 19, 6119_ 4 of 14
the model (e.g., the population size and the basic reproduction number R0) and to add
an intervention to decrease transmission by a specified amount from a given day. The
Basel model allowed for the user to modify various model parameters, age-group-specific
parameters, and isolation measures, and add multiple interventions to reduce transmission. In contrast, the IHME model had limited local parameter customizability. Although
it generated state-specific estimates, it did not allow for state-specific parameters to be
incorporated. Moreover, although the IHME model used a wide variety of data sources,
not all states had their data reflected in the model. In the case of Hawaii, the IHME model
initially did not appear to utilize data from Hawaii but instead utilized average estimates of
time from hospitalization to death from other states, despite widely different demographic,
epidemiologic, and socioeconomic considerations.
**Age Distribution. Age is well documented as one of the largest and most significant**
risk factors for COVID-19, with older adults at increased risk of being hospitalized and
dying due to COVID-19 [9,10]. Each state has different age distributions and demographic
and population age structure, and so it is important for models to account for age to
project the case, hospitalization, and fatality numbers more accurately. The technical team
during their rapid review identified the University of Basel model as Susceptible Infected
Recovered/Susceptible Exposed Infected Recovered (SEIR) compartment-based models
that accounted for age distributions, allowing for the user to adjust the age distribution
and age-group-specific parameters to reflect the population of interest.
**Type of Model. Models can be broadly categorized into two types—mechanistic**
and statistical. Mechanistic models make assumptions about how the actual process of
COVID-19 disease transmission occurs and include the SEIR compartmental models and
their modified variants. In contrast, statistical models fit curves using existing data, the
main example being the IHME model which early on used the existing data from China
and Italy to predict what would happen in the United States and elsewhere. This means
that while statistical models can forecast what will happen in the near future, mechanistic
models can make assumptions on the transmission dynamics of COVID-19 and forecast
longer-term scenarios based on different interventions and policy changes [11].
**Table 1. Landscape of selected models for informing COVID-19 control and mitigation, 2020.**
**Localized** **Local Age**
**Objective of Model** **Type of Model** **Open Source**
**Customizability** **Distribution**
IHME [5] Estimate hospital impacts No Unknown [1] Statistical No
Imperial College Assess public health
No [2] No [2] Mechanistic No [2]
London [6] measures on spread
Epidemic
Calculator [7]
University of
Basel [8]
Estimate change in epi
curve after reduction Yes No Mechanistic Yes
in transmission
Planning tool with features
such as imported cases and Yes Yes Mechanistic Yes
age groups
1 The IHME model was closed source so it was unknown how local age distribution was taken into account. 2 The
source code was not available when the original Report 9 was released. The updated source code was eventually
made available much later with limited documentation, making localized use of the model difficult.
Incorrectly utilizing a statistical model to create long-term scenarios can produce
results that “may suffer from the fallacy of Farr’s law, a similar non-mechanistic method
in which epidemics are assumed to follow a normal distribution shifted and scaled to fit
data” [12]. This was a common and widespread criticism of the IHME, as simply fitting a
curve to historical data and extrapolating into the future can produce dramatic over- or
underestimates of the epidemic’s impact [13].
However, mechanistic models also have shortcomings. Parameters available to a
model are finite, meaning any output will be inherently flawed. In addition, assignment of
values to the parameters available in each model may only be viable with respect to a given
historical situation, but relatively meaningless considering even a small shift in the makeup
-----
_Int. J. Environ. Res. Public Health 2022, 19, 6119_ 5 of 14
or habit of the population or any immediate major policy change in the future, including
upon seeing the results of the model. Therefore, it is important to recognize and establish
degrees of uncertainty within the parameters themselves for mechanistic models. Models
without such extended boundaries should be caveated generously or avoided completely.
**Open Source. Models that are open source, defined as having the source code made**
publicly available for use and modification, are models that enable users to “open up the
hood of the car” or “look into the sausage-making machine”. This transparency in model
assumptions and limitations should have appropriate interpretation by an epidemiologist
to policymakers to ensure appropriate planning. Most importantly, open source incurs little
to no additional cost and offers support to states with limited technical and epidemiologic
capacity. For example, the IHME model was not open source which made it very challenging to assess even basic assumptions such as how it incorporates age-specific distributions.
Policymakers should approach model interpretations cautiously and not make assumptions
of the data.
_2.3. Model Assumptions and Limitations_
This section reviews the key assumptions identified by the technical team of these
models, which were used to inform the applicability of a given model to a particular
jurisdiction, and in this case, the state of Hawaii. Table 2 presents explicitly some of the
assumptions in the four models. Ultimately, the technical team chose to use the Epidemic
Calculator and University of Basel model for several key decisions in 2020, as described
in Section 3. In Hawaii in 2020, the Hawaii Data Collaborative in partnership with local
hospitals and HiPAM built on and modified the open-source Epidemic Calculator model to
show how policy measures on reopening and resuming travel could impact the spread of
COVID-19. Later beginning in 2021, HiPAM relied on a locally customized model.
**Table 2. Selected model assumptions for informing COVID-19 control and mitigation, 2020.**
**Key Assumption #2:**
**Age Distribution**
Uses actual data and are
based on results for
specific age
distributions (for China
and Italy) applied and
adapted to other
populations
Agent based model has
individuals that reflect
the population’s age
distribution
Does not take age or age
distributions into
account and unclear the
reference population or
data used to benchmark
(e.g., China)
Divides population into
age groups with
age-group-specific
parameters (such as
how severe, critical, and
fatal the infection is)
**Other Assumptions**
Assumes changes in
transmission are
reflected through
mobility of the
population
Puts imported cases
into the Exposed
compartment, which
can be interpreted as the
cases coming from
outside are all
incubating/recently
infected and not
symptomatic
**Underestimate or**
**Overestimate on Total**
**Severity (Cases,**
**Deaths)**
As the model is not
open source, it is
unapparent to what
extent asymptomatic vs.
symptomatic is
considered
Same as for Epidemic
Calculator (see below)
May underestimate
total severity as
asymptomatic
individuals are more
likely to spread
COVID-19 as they are
unaware, they are
infected and/or
infectious
Same as for Epidemic
Calculator (see above)
**Underestimate or**
**Overestimate on Total**
**Severity (Cases,**
**Deaths)**
As the model is not
open source, it is
unapparent how the
age-specific
distributions are
incorporated and
applied
Not applicable
May overestimate
hospitalizations and
fatalities if population is
younger, as increased
age significantly
increases risk [1]
Depends on whether
the user correctly selects
the age distribution and
age-group-specific
parameters of
geographic location of
interest
IHME [5]
Imperial College [6]
Epidemic Calculator [7]
University of Basel [8]
**Key Assumption #1:**
**Asymptomatic vs.**
**Symptomatic**
As the model is not
open source, it is
unapparent to what
extent asymptomatic vs.
symptomatic is taken
into account
Does not appear to
distinguish between
asymptomatic and
non-hospitalized
symptomatic
individuals
Does not appear to
distinguish between
asymptomatic and
non-hospitalized
symptomatic
individuals
Does not appear to
distinguish between
asymptomatic and
non-hospitalized
symptomatic
individuals
1 The United States has a younger age distribution compared to China, so models that use aggregate estimates of
mortality for China may overestimate mortality for the United States unless age-specific mortality distributions
are accounted for.
-----
_Int. J. Environ. Res. Public Health 2022, 19, 6119_ 6 of 14
**3. Policy Use Cases of Applying Models to Specific State Policy Decisions**
In this section of the policy case study, we provide a historical account for how
models were used to inform key policy decisions in the jurisdiction, both in terms of
managing capacity during a surge and reopening amidst a decline. The policy case study
reflects actual lived experiences along with (now historical) observations and perspectives
of the technical team and the HiPAM work group who were providing information to
policymakers seeking to make decisions informed by modeling. Thus, the detailed microlevel analyses for each use case are not presented herein, but rather, the specific translation
of evidence to knowledge and communication with a variety of stakeholders including
policymakers, media, and the public are described.
_3.1. Using Models for Managing Resources and Capacity in a Surge_
Many states did not have an established epidemic or pandemic response plan for
COVID-19, let alone a plan for how to use modeling for informing policy decisions. Amidst
this context, a pressing and overarching question was how quickly COVID-19 would spread
in their state or community. As such, the IHME model was utilized because it provided
early state-specific estimates. It was one of the only models at the time that gave a hard
deadline by which a state’s bed surge capacity might be reached because of the speed by
which COVID-19 spreads and leads to hospitalization. It was also widely disseminated in
the news and prompted several policymakers to inquire whether decisions could be made
based on what was circulated in the media. Policymakers rarely have a background in
infectious disease or epidemiology, and the wide coverage of the models in media does
not guarantee their appropriate use. Thus, this case study reflects the occasion in which
some policymakers had the foresight and humility to seek out information and inputs from
technical experts for three use cases described herein:
_•_ **Use Case 1: Determining whether there was adequate hospital bed capacity in the**
state and adequate PPE in the state.
_•_ **Use Case 2: Assessing the need for isolation and quarantine facilities from the surge**
of the second wave in the fall of 2020.
_•_ **Use Case 3: The role of public communication during the Delta surge in the summer**
of 2021, and the Omicron surge in the fall of 2021.
3.1.1. Use Case 1: Adequacy of Hospital Bed and Personal Protective Equipment Capacity
The IHME model was initially used to plan for ensuring adequate bed capacity and
to decide whether to put up additional acute care facilities. In Hawaii, policymakers
pondered challenging decisions of whether to retrofit existing hotel rooms or outfit a
convention center. Either option would require collaboration with the US Army Corps
of Engineers with an expensive price tag. This policy decision required COVID-19 case
and hospitalization projections specific for Hawaii. In the beginning of the pandemic,
with no other available guidance or tools as well as limited or no epidemiologic advisors,
policymakers turned to the web-accessible IHME model for guidance on when Hawaii
would be hit with a “surge” of cases.
However, at the onset of COVID-19 in the US, many states had yet to fully understand
how the virus was spreading through their individual communities and how measures
such as requiring face mask use in public would affect the spread [14]. Through the month
of March and early April, many states did not yet have a high case count and fatality
count to get a sense of the trend of COVID-19 within their state. The IHME model used
the hospitalization to death ratio from seven locations within the US with the most cases
to create a weighted average for their ratio and applied it to states with fewer than five
fatalities, which included Hawaii. This resulted in Hawaii expecting to see a surge in cases
and hospitalizations that was projected to overwhelm the local healthcare system.
Yet, when the technical team with the HiPAM work group and utilized a basic SEIR
model with Hawaii-specific parameters, no surge was estimated within the same time
frame that IHME was predicting. The modeling team in Hawaii understood that Hawaii’s
-----
_Int. J. Environ. Res. Public Health 2022, 19, 6119_ 7 of 14
unique geography and early mitigation efforts, most significantly restricting air and sea
travel, drastically reduced the Rt below the value of 2.2 that was used by most models.
The technical team stated all the limitations and assumptions of their early model to the
decision makers in HI-EMA and others. Based on the recommendation of state-specific data
and use of a locally tailored epidemiologic model, the decision was made to not retrofit
the Hawaii Convention Center into an acute care facility at that time, and to re-evaluate at
a future date. Ultimately, Hawaii was never hit with a surge at the level predicted by the
IHME model, i.e., the predictive validity of the IHME model was poor for Hawaii. Due to
the information provided by the technical team, the state of Hawaii made a policy decision
that avoided millions of USD in costs.
While the IHME model influenced policymakers and emergency management leaders’
decisions about imposing public health measures to stop COVID-19 spread (e.g., through
closing business and halting travel), it is important to note that the actual death totals
from COVID-19 were outside the IHME model’s 95% confidence interval 70% of the
time [15]. This fact is notable given the importance of deaths as a measurement of COVID-19
spread during the early days of the pandemic [16]. Yet, predictive validity is an ex post
consideration for model selection. Policymakers must make best possible decisions without
knowing the future or the predictive validity of any given model. Thus, in the historical
case study, the selection of models best used for a given jurisdiction was based on the
factors noted in Section 2.
Allocation, logistics, and utilization of personal protective equipment (PPE) during
the initial response to COVID-19 was another use of COVID-19 models by policymakers.
Hospital administrators and policymakers need to accurately account for burn rates of PPE
(e.g., masks, surgical gowns, and facemasks) to request appropriate funding from their
funding sources. The technical team used the University of Basel model for informing
PPE. Regarding stockpiling respirators, estimations of need were essential to decreasing
over-stocking which may diminish the supply in other areas of need, or under-stocking
respirators, which would have had severe consequences.
3.1.2. Use Case 2: Isolation and Quarantine Capacity Planning
In the second surge that Hawaii experienced in the fall of 2020, the models adapted and
used through the HiPAM work group were used to communicate the forecasted number of
cases, hospitalizations, and deaths, primarily through behind-the-scenes communications
to senior state policymakers including the Governor’s office and the county Mayors, among
others. Whereas the models used for hospital bed planning were informing HI-EMA as
a key state agency, the departure of the epidemiologic advisor to HI-EMA in July of 2020
resulted in the HiPAM work group stepping in to serve as the go-to local institutional
contact for modeling, supported by the Hawaii Data Collaborative and Hawaii Department
of Health Behavioral Health Administration. HiPAM had formed in April of 2020, bringing
together health professionals, data scientists, mathematicians, and agency staff to convene
around an agenda on COVID-19 modeling. Given the limitations in resources in a small
remote state, there was a need to pool resources and efforts together to reduce duplication
and confusion. The interdisciplinary HiPAM work group was structured on past work
of the HiPAM chair, who had previous experience using work groups at a think tank in
Washington, DC (the Center for Global Development). Based on the work of the HI-EMA
epidemiologic advisor and technical team with support from HiPAM, a need for an ongoing
forecast for the state was identified. By July of 2020, HiPAM launched an online two-week
COVID-19 forecast, accessible publicly.
As the local response evolved including increasing capacity for testing, tracing, and
isolation and quarantine, the models were also used to inform isolation and quarantine
capacity which had a particular emphasis on vulnerable populations including homeless
individuals, Native Hawaiian and Other Pacific Island communities, as well as individuals
with co-occurring mental illness and substance use challenges. The Hawaii Department of
Health’s Behavioral Health Administration (BHA) was designated to lead isolation and
-----
_Int. J. Environ. Res. Public Health 2022, 19, 6119_ 8 of 14
quarantine beginning in August 2020 as Hawaii experienced its second surge. The BHA
was also the sole DOH unit to establish the standalone Temporary Quarantine & Isolation
Center specifically for homeless individuals and later for medically needy individuals [17].
As the BHA leadership actively participated in HiPAM, BHA leadership had sought inputs
and guidance from HiPAM models to monitor adequacy of bed capacity for isolation and
quarantine real-time case counts. Models from HiPAM were also used to estimate the
adequacy of shelter capacity for homeless populations.
There was a policy need for a simple benchmark to identify whether there was adequate isolation and quarantine capacity and, specifically, enough beds procured by the State
of Hawaii. With limited time and resources available to conduct a detailed epidemiologic
and demographic study, there was a need to identify in a simple manner how many people might need isolation and quarantine services. Eligibility for isolation and quarantine
services in a government-procured hotel was determined in part based on whether an
individual was able to safely isolate at home and whether the individual lived in a shared
bedroom with someone. In Hawaii, the percentage of the population living in a shared
bedroom was identified to be nearly 10%. When applied to the total number of active
COVID-19 cases at any given time, this benchmark helped to inform the planning for
the total beds procured by the State of Hawaii for isolation and quarantine operational
activities in 2020. Although ‘active cases’ as a concept was challenging to independently
measure due to lack of capacity for verification of individuals released from isolation and
quarantine, the rolled-up cumulative case count over the last 14 days was used as a proxy
for active case count for the state, on which the ten percent was applied to estimate need
for isolation and quarantine outside of one’s home.
3.1.3. Use Case 3: Broad Public and Media Communications
In the Delta and Omicron surges in 2021 in the summer and winter, respectively,
HiPAM took a direct public communications strategy to communicate the results of the
model and forecast. Rather than use only backdoor communication with senior policymakers and government authorities, HiPAM emphasized direct communications with the
media, similar to the weatherman, as well as the release of regular and timely reports sent
to all key policymakers and news outlets in the state, supplementing the online web tool
hosting the two-week advance forecast, which had been launched in July of 2020.
By 2021, a locally developed and customized model led by mathematicians (Chyba
et al.) became the de facto and well-accepted model by HiPAM for the state, and the
other models by the University of Basel and IHME were abandoned by 2021 [18]. The
Chyba et al. model fulfilled the key considerations for the models including local parameter
customizability, local age distribution, use for assessing different policy scenarios, and being
customizable and potentially open source because it was developed in-house. Developing
local in-house mathematical and epidemiologic modeling capacity is extremely challenging
and dependent upon the availability of scientific experts willing to engage in real-world
policy challenges and was spearheaded by funding from the National Science Foundation
competitively awarded to Chyba et al. [18].
The direct public dissemination of the model results to the state in 2021 and 2022 was
vastly different from the 2020 approach of behind-the-scenes information provided to senior
leaders. Nevertheless, this public communication strategy also had challenges and risks in
terms of the ways in which the modeling results and information was communicated and
the kinds of questions and concerns posed by the media, policymakers, and the public. The
media and policymakers, for example, repeatedly asked HiPAM representatives challenging
questions about the specific policy guidance that should be made based on the modeling
results. Yet in order to ensure and maintain the scientific credibility of the HiPAM models,
HiPAM repeatedly emphasized its role as a scientific body that focused on high-quality
models based on best-available evidence and ever-evolving science. It had to remind
the media, the public, and policymakers that while HiPAM’s information was important,
policymakers would need to use multiple sources of information, in order to make decisions.
-----
_Int. J. Environ. Res. Public Health 2022, 19, 6119_ 9 of 14
In doing so, HiPAM reinforced its role as a scientific body and not as a body making policy
recommendations or actions. Maintaining a scientifically neutral and unbiased position
was essential for ensuring HiPAM public credibility and recognition.
A second key question repeatedly raised by the public, the media, and policymakers
was about the time horizon of the model. There was a longstanding desire for understanding the forecast or projection of COVID-19 well into the future by several months.
HiPAM, however, maintained a stance of emphasizing a two-week forecast horizon, and
that anything longer than that would be subject to change. Seeing the real-world mistakes
of models communicated at the national level, HiPAM made a deliberate choice to focus
on a limited time horizon and repeatedly emphasized the ways in which individual and
policy actions could easily influence the forecast beyond two weeks.
A third key challenge of direct communications was the emphasis on the dynamic
nature of the modeling results. HiPAM repeatedly noted that upon release of the forecast,
the projection would immediately change based on the fact that knowledge and information
about the situation would result in changes in individual behavior as well as policy changes
and action. Unlike weather forecasting, dissemination of a COVID-19 forecast changes
the forecast itself. This difficulty in communicating the dynamic nature of modeling was
challenging throughout the pandemic, even with seasoned media reporters and engaged
policymakers and legislators.
The media and communications engagements was also broad through numerous
media engagements through print, television, radio, social media, state and county government hearings and fora, and so on, raising public awareness of the COVID-19 modeling
writ-large beyond the closed circles of policymakers. The media strategy led to wide
acceptance, recognition and use of the COVID-19 modeling not only by state agencies but
also other health care provider organizations including the local hospital association. The
credibility and validity of the HiPAM model work was emphasized primarily through
maintaining a neutral stance on any specific policy recommendation but focusing on the
specific technical result or information that the model provided.
The interactive communication loop also enabled immediate feedback from the policymakers, the media, and the public to ask questions including about understanding the
potential impacts of a particular policy or intervention scenario. These communication engagements also enabled policymakers to request and clarify potential scenarioing requests
for any given policy.
_3.2. Using Models for Reopening Amidst Decline_
Models were also used to inform policy decisions for reopening, and these decisions
were equally pressing due to the economic impacts of COVID-19. In the case of Hawaii, one
major question facing policymakers seeking to reopen is how and when travel volumes,
both domestic and international, can increase. Some of the early mechanistic models
only accounted for a population where the total size stayed the same as well as for how
COVID-19 would progress under certain mitigation efforts scenarios pre-programmed
into the model. Moreover, there continues to be uncertainty and evolving understanding
about the basic scientific facts and assumptions of COVID-19 (e.g., extent of screening
for asymptomatic transmission [19] and the infection fatality rate [20]), making policy
decisions difficult. As travel volumes return to higher levels, models that factor in imports
of new cases can provide more accurate estimates of travel impacts on overall disease
spread. Teams engaged in epidemic forecasting can estimate metrics for different travel
volume scenarios and demonstrate how the range of new cases is dependent on how many
imported cases are brought into their community.
There are many assumptions built into the various COVID-19 models, such as whether
symptomatic travelers will restrict themselves from traveling and whether they will be
identified at the port of departure. Arguably, one of the largest considerations for developing travel scenarios is that of asymptomatic and pre-symptomatic cases—how assumptions
about these parameters are incorporated into a given model, the distribution of COVID-19
-----
_Int. J. Environ. Res. Public Health 2022, 19, 6119_ 10 of 14
cases which are asymptomatic or pre-symptomatic, and the rate of spread from these
cases [21–23]. Reopening strategies based on one or multiple tests have been suggested
without any numerical estimations of possible infected travelers slipping through. Modeling can provide policymakers with an educated guess when comparing reopening strategies
based on frequency and type of tests.
Testing, contact tracing, and isolation and quarantine represent major public health
tools for policymakers responding to COVID-19. States can consider how tests such as
body temperature and symptom screens as well as standard polymerase chain reaction
(PCR) tests can be linked to travel policies, and use models to help estimate the potential
impacts and consequences of different testing strategies.
Most models at present do not account for health impacts beyond the immediate
COVID-19 health impacts such as those pertaining to mental health, reductions in use
of other essential health services, or long-term care facilities or other congregate settings.
Mental health and substance use, already important public health issues prior to COVID-19,
have become exacerbated by secondary and tertiary impacts due to COVID-19 [24].
COVID-19 will continue to directly impact communities and indirectly for decades
to come. Policymakers will need to shift from the use of models that focus on hospital
capacity and reopening, to models identifying long-term health and economic impacts of
COVID-19, such as mental health, access to non-COVID-19 health care services, education,
and other dimensions of the social determinants of health. Most of the COVID-19 models
used at present have not directly incorporated these long-term impacts. The use of Bayesian
modeling and synthetic population models can and have begun to be used to examine
these longer-term health impacts and policy implications, as this type of modeling accounts
for additional differences in a population such as economic status and race.
**4. Discussion**
Epidemiologic models used for COVID-19 are numerous and complex, requiring
subject matter experts to appropriately utilize data, interpret, and communicate results.
The study used a case study approach to provide a historical account of the events and
reflect on the lessons learned from one jurisdiction in the United States. The historical events
reflected in this case study demonstrate the real-world challenges that policymakers and
subject matter experts face when deciding which model to use, including demonstrating
how even with accurate data, utilization of an inappropriate model or considerations has
the potential to lead to inappropriate interpretation of results. COVID-19 models vary
by their designed intent and understanding these differences, including their differences
in geographic application and applicability to specific policy decisions, is necessary for
policymakers to better utilize them in making decisions [25–27].
There are several key lessons that can be drawn from this case study documenting the
historical application of mathematical and epidemiologic models for key policy decisions.
First, COVID-19 modeling in Hawaii benefited from the incorporation of state-specific data
which were historically argued to directly result in cost savings from decreased unnecessary
spending, particularly in the case of the hospital capacity planning in the early part of 2020
during the pandemic. This model incorporated two of the most important factors that
assist local leaders in modeling local issues, age distribution and customization that was
specific to Hawaii [9,10]. It also helped to inform isolation and quarantine planning and
adequacy of facilities available in order to meet demand and need in the fall of 2020, as well
as helped to inform the media, the public, and policymakers of the potential magnitude of
the Delta and Omicron surges in 2021 to 2022.
Second, regardless of model selection, it is essential that model outputs be interpreted
directionally, not as a forecast of hard, immutable numbers, and with a clearly delineated
time horizon. The numerous known and unknown factors and their combinations thereof
impacting the spread of COVID-19 means that no model, no matter the level of sophistication, can concurrently accurately account for all factors. Further, the nature of the models,
easily influenced by actions of individuals and policies today, makes the models dynamic
-----
_Int. J. Environ. Res. Public Health 2022, 19, 6119_ 11 of 14
and uncertain rather than static, despite a desire for static and definitive answers. Therefore,
numbers produced regarding cases, hospitalizations, deaths, etc. should be communicated
and understood as a possible scenario should current trends continue forward into the future assuming no change in policy or human behavior, which is an impossible assumption
to begin with as soon as a forecast is released and disseminated.
Third, when the above factors are properly considered and both model outputs—
projected trends and subsequent reductions by means of interventions—are combined, it
is essential that warnings are heeded and action be taken as soon as possible. Based on
appropriate interpretation of a model, policymakers can be advised if a policy intervention
may avert critical thresholds such as hospital and ICU capacity. With these crucial timings
in mind, policymakers can then use other models to help more accurately understand
how an intervention may impact the Rt in a given population and, subsequently, what
scenarios of intervention combinations and efficacies might result in faster control or even
elimination of COVID-19 [28]. Hesitation in implementation or inadequate interventions
can have dramatic effects on disease spread, such as the delayed and scattered approach to
mask wearing early on in the pandemic [29]. However, failure to grasp model limitations
can also result in hasty and expensive overreactions.
Fourth, these models require a firm grasp of epidemiologic concepts. As such, policymakers are advised to seek out and involve an interdisciplinary scientific advisory group
as early as possible to translate modeled outcomes into actionable context. Because of
the complexity of models, significant unpredictable impact of human behavior, and the
potential for misinterpretation, it can be argued that these models do more harm than
good. Rather than dismiss the use of models because of their complexity, policymakers
should incorporate into the scientific and technical advisory group or ‘brain trust’ as early
as possible to help inform and navigate the difficult policy decisions that can have positive
impacts on their constituents and communities.
In the case study herein, the brain trust was a diverse team that provided input
from various areas of expertise (e.g., epidemiology, data science, behavioral health, and
mathematics). The role of the local university to work in close collaboration and partnership
with the community and government authorities is essential. While it remains to be seen
what the long-run impact of the communication of COVID-19 was modeling and forecasting
in the state of Hawaii, it cannot be denied that there is value of having mathematical and
epidemiologic capacity of scientific experts who are willing and able to communicate and
contribute to real-world policy challenges in service of the public and community during
an extended period of confusion and crisis.
The work in Hawaii of using a brain trust may be contextualized to the work in many
countries around the world which used COVID-19 modeling and forecasting to inform
decision making. In Hawaii, the creation of HiPAM included a range of local experts from
epidemiology, public health, data science, and mathematics who were able to contribute to
modeling and forecasting locally. Other countries such as Ireland, United Kingdom, New
Zealand, and several others had technical advisory groups that provided inputs and information to policymakers who ultimately made the policy calls and decisions. For example,
in New Zealand, a COVID-19 technical advisory group comprised medical, public health,
and academic advisors, which provided advice to the ministry of health. In Australia,
the COVID-19 Expert Database hosted by the Australian Academy of Science provided
a mechanism for governments and decision makers to have easier access to expertise
in COVID-19. The UK also established a Scientific Advisory Group for Emergencies as
the entity responsible for providing scientific advice to UK decision makers while not
representing official government policy.
More research is needed to examine how to create and then institutionalize these
bodies with blended technical expertise, savvy communication skills, and linkages to
policymakers making decisions. What is the optimal composition of these bodies? To what
extent are these bodies linked and connected to decision-making? What is the role and
balance of neutrality in balancing scientific facts relative to policy recommendations? The
-----
_Int. J. Environ. Res. Public Health 2022, 19, 6119_ 12 of 14
work in Hawaii serves as a basis for the hypothesis that the composition of a body that
draws from a wide range of expertise beyond clinical medicine or public health fields can
help to bridge challenges of mathematics, applied or real-world epidemiology, behavioral
health, and data science. Existing public health institutions such as those pertaining to
health technology assessment or epidemic intelligence are also relevant and should be
considered before new bodies are duplicated, creating more institutional fragmentation,
siloization, and duplication of effort.
Further, there is a need for communication and practitioners who can help to translate
and communicate complex ideas into simple concepts for policymakers and the public.
We would also hypothesize that whether forecasting and modeling are sidelined or are
integral to decision-making depends on leadership and governance in formally supporting,
building, seeking guidance from, and incorporating information from technical advisory
groups. During a time when there is controversy and doubt about science, the ways in
which scientific and technical advisory bodies monitor ever-evolving science and evidence
and how such bodies intersect with and communicate with policymakers who then make
decisions for policies, programs, and practice merits further study.
There is some research in the field of political science and public administration
which examines the ways in which policy decisions and actions are determined and
implemented. Political science theories have been applied to understand how political
actors make policy actions and political decisions, including Kingdon’s Multiple Streams
Model [30] and Reich’s work on political economy [31]. Work by Walt et al. noted that
rigorous health policy research methods have much to be desired for understanding the
policy process [32]. In particular, a major limitation of this historical case study is its focus
on a single jurisdiction—the state of Hawaii—one that lacks a comparison or historical
“counterfactual” for what might have happened in the absence of this work on modeling
in the state. We argue that drawing on historical perspectives and chronology from lived
experiences of those engaged in real-world implementation and operations (albeit informed
by modeling and evidence) is a research methodology.
This case study also did not explicitly examine cases of misappropriation and misuse
of models in Hawaii or cases in which the modeling outputs were ignored or otherwise not
used for specific policy actions. Determining what constitutes misuse and misappropriation
is beyond the scope of this paper, but we acknowledge that the complexity of models makes
inappropriate or poor application quite possible, if not the default. Future research would be
valuable to examine the different ways in which modeling informed or did not inform key
policy decisions in multiple states and jurisdictions and the variations in communication
about modeling.
**5. Conclusions**
This article has emphasized the role of localizing knowledge that can be translated
and used to inform local decisions. With tremendous uncertainty about a novel disease, the
need for thoughtful application of scientific knowledge is ever more pressing. Although the
specific use cases and policy window and moment for critical decisions described herein
have now passed, the lessons from this case study may be relevant for jurisdictions seeking
to make smarter decisions informed by modeling. The knowledge and experience that
was gained through these lived experiences may be applicable to island countries and
states with age, ethnicity, and other sociodemographic distributions similar to Hawaii.
The knowledge and experience from this case study may also help to inform jurisdictions
experiencing limitations in resources, time, and scientific expertise for COVID-19 modeling
in informing policymaking.
**Author Contributions: Conceptualization, T.H.L. and V.Y.F.; methodology, T.H.L., B.D., L.D., J.H.,**
M.C. and V.Y.F.; resources, S.H., E.M., K.H. and V.Y.F.; writing—original draft preparation, T.H.L.;
writing—review and editing, T.H.L., B.D., L.D., J.H., M.C., S.H., E.M., K.H. and V.Y.F.; supervision,
project administration and funding acquisition, E.M., K.H. and V.Y.F. All authors have read and
agreed to the published version of the manuscript.
-----
_Int. J. Environ. Res. Public Health 2022, 19, 6119_ 13 of 14
**Funding: T.H. Lee and V.Y. Fan gratefully acknowledge extramural funds from the Hawaii State**
Department of Health Behavioral Health Administration Alcohol and Drug Abuse Division (ADADMOA-SP-19-01, ASO Log No. 15-074). M. Chyba gratefully acknowledges extramural funds from the
National Science Foundation, award #2030789. V.Y. Fan and M. Chyba gratefully acknowledge support from the Coronavirus State Fiscal Recovery Funds via the Governor’s Office Hawaii Department
of Defense and the University of Hawaii at Manoa Provost’s Office.
**Institutional Review Board Statement: Not applicable.**
**Informed Consent Statement: Not applicable.**
**Data Availability Statement: Not applicable.**
**Acknowledgments: The authors gratefully acknowledge support from the members of the Hawaii**
Pandemic Applied Modeling Work Group (Adriann Gin, Baseem Missaghi, Brian Wu, Curtis Toma,
David Chow, Francis Chan, Istvan Szapudi, Janet Berreman, Kendrick Leong, Kiyoshi Shiraishi, Lee
Altenberg, Marguerite Butler, Nick Redding, Noah Hafner, Peter Fuleky, Roy Esaki, Rukiyah Walker,
Tiana Tran, Tom Blamey), ACES Lab, Applied Research Laboratory, Margo Edwards, Harry Kim,
Aimee Grace, Vassilis Syrmos, Velma Kameoka, Michael Bruno, David Lassner, and John Valera.
The authors also gratefully acknowledge the public, the media, and the federal, state, and county
policymakers for their interest in this work. Any omissions or errors are our own.
**Conflicts of Interest: The authors declare no conflict of interest.**
**References**
1. Logan, R.I.; Castañeda, H. Addressing Health Disparities in the Rural United States: Advocacy as Caregiving among Community
[Health Workers and Promotores de Salud. Int. J. Environ. Res. Public Health 2020, 17, 9223. [CrossRef] [PubMed]](http://doi.org/10.3390/ijerph17249223)
2. Shelus, V.S.; Frank, S.C.; Lazard, A.J.; Higgins, I.C.A.; Pulido, M.; Richter, A.P.C.; Vandegrift, S.M.; Vereen, R.N.; Ribisl, K.M.; Hall,
M.G. Motivations and Barriers for the Use of Face Coverings during the COVID-19 Pandemic: Messaging Insights from Focus
[Groups. Int. J. Environ. Res. Public Health 2020, 17, 9298. [CrossRef] [PubMed]](http://doi.org/10.3390/ijerph17249298)
3. Zhang, X.; Warner, M.E. COVID-19 Policy Differences across US States: Shutdowns, Reopening, and Mask Mandates. Int. J.
_[Environ. Res. Public Health 2020, 17, 9520. [CrossRef]](http://doi.org/10.3390/ijerph17249520)_
4. Dubé, C.; Garner, G.; Stevenson, M.; Sanson, R.; Estrada, C.; Willeberg, P. The use of epidemiological models for the management
of animal diseases. Conf. OIE 2007, 1, 13–23.
5. Team IC 19 Health Service Utilization Forecasting; Murray, C.J. Forecasting the impact of the first wave of the COVID-19 pandemic
[on hospital demand and deaths for the USA and European Economic Area countries. medRxiv 2020, 2020, 20074732. [CrossRef]](http://doi.org/10.1101/2020.04.21.20074732)
6. Ferguson, N.; Laydon, D.; Nedjati Gilani, G.; Imai, N.; Ainslie, K.; Baguelin, M.; Bhatia, S.; Boonyasiri, A.; Cucunuba Perez, Z.;
Cuomo-Dannenburg, G. Impact of Non-Pharmaceutical Interventions (NPIs) to Reduce COVID19 Mortality and Healthcare Demand;
[Imperial College London: London, UK, 2020. [CrossRef]](http://doi.org/10.25561/77482)
7. [Goh, G. Epidemic Calculator. Available online: https://gabgoh.github.io/COVID/index.html (accessed on 9 June 2020).](https://gabgoh.github.io/COVID/index.html)
8. Noll, N.B.; Aksamentov, I.; Druelle, V.; Badenhorst, A.; Ronzani, B.; Jefferies, G.; Neher, R.A. COVID-19 Scenarios: An interactive
[tool to explore the spread and associated morbidity and mortality of SARS-CoV-2. medRxiv. 2020, 2020, 20091363. [CrossRef]](http://doi.org/10.1101/2020.05.05.20091363)
9. Bialek, S.; Boundy, E.; Bowen, V.; Chow, N.; Cohn, A.; Dowling, N.; Ellington, S.; Gierke, R.; Hall, A.; MacNeil, J.; et al. Severe
Outcomes Among Patients with Coronavirus Disease 2019 (COVID-19)—United States, February 12–March 16, 2020. MMWR
_[Morb. Mortal. Wkly. Rep. 2020, 69, 343–346. [CrossRef]](http://doi.org/10.15585/mmwr.mm6912e2)_
10. Garg, S.; Kim, L.; Whitaker, M.; O’Halloran, A.; Cummings, C.; Holstein, R.; Prill, M.; Chai, S.J.; Kirley, P.D.; Alden, N.B.;
et al. Hospitalization Rates and Characteristics of Patients Hospitalized with Laboratory-Confirmed Coronavirus Disease
[2019—COVID-NET, 14 States, March 1–30, 2020. MMWR Morb. Mortal. Wkly. Rep. 2020, 69, 458–464. [CrossRef]](http://doi.org/10.15585/mmwr.mm6915e3)
11. Holmdahl, I.; Buckee, C. Wrong but Useful—What COVID-19 Epidemiologic Models Can and Cannot Tell Us. N. Engl. J. Med.
**[2020, 383, 303–305. [CrossRef]](http://doi.org/10.1056/NEJMp2016822)**
12. Jewell, N.P.; Lewnard, J.A.; Jewell, B.L. Caution Warranted: Using the Institute for Health Metrics and Evaluation Model for
[Predicting the Course of the COVID-19 Pandemic. Ann. Intern. Med. 2020, 173, 226–227. [CrossRef]](http://doi.org/10.7326/M20-1565)
13. Brown, G.; Cavanaugh, J.; Miller, A.; Oleson, J.; Pentella, M.; Perencevich, E.; Sewell, D. Critique of the IHME Model for COVID-19
[Projections. Available online: https://governor.iowa.gov/sites/default/files/documents/IDPH%20Whitepaper%201%20-%](https://governor.iowa.gov/sites/default/files/documents/IDPH%20Whitepaper%201%20-%20Critique%20of%20IHME%20Model%20for%20COVID-19%20Projections%20%281%29.pdf)
[20Critique%20of%20IHME%20Model%20for%20COVID-19%20Projections%20%281%29.pdf (accessed on 11 April 2021).](https://governor.iowa.gov/sites/default/files/documents/IDPH%20Whitepaper%201%20-%20Critique%20of%20IHME%20Model%20for%20COVID-19%20Projections%20%281%29.pdf)
14. Lyu, W.; Wehby, G.L. Community Use Of Face Masks And COVID-19: Evidence From A Natural Experiment Of State Mandates
[In The US. Health Aff. 2020, 39, 1419–1425. [CrossRef] [PubMed]](http://doi.org/10.1377/hlthaff.2020.00818)
15. Marchant, R.; Samia, N.I.; Rosen, O.; Tanner, M.A.; Cripps, S. Learning as We Go: An Examination of the Statistical Accuracy of
[COVID19 Daily Death Count Predictions. arXiv 2020, arXiv:2004.04734. [CrossRef]](http://doi.org/10.1101/2020.04.11.20062257)
16. Subbaraman, N. Why daily death tolls have become unusually important in understanding the coronavirus pandemic. Nature
**[2020. [CrossRef] [PubMed]](http://doi.org/10.1038/d41586-020-01008-1)**
-----
_Int. J. Environ. Res. Public Health 2022, 19, 6119_ 14 of 14
17. Fan, V.Y.; Fontanilla, T.M.; Yamaguchi, C.T.; Geib, S.M.; Holmes, J.R.; Kim, S.; Do, B.; Lee, T.H.; Talagi, D.K.P.; Sutton, Y.; et al.
[Experience of isolation and quarantine hotels for COVID-19 in Hawaii. J. Travel Med. 2021, 28, taab096. [CrossRef] [PubMed]](http://doi.org/10.1093/jtm/taab096)
18. Kunwar, P.; Markovichenko, O.; Chyba, M.; Mileyko, Y.; Koniges, A.; Lee, T. A study of computational and conceptual complexities
[of compartment and agent based models. N. Heterog. Media 2022, 17, 11546. [CrossRef]](http://doi.org/10.3934/nhm.2022011)
19. Park, M.; Cook, A.R.; Lim, J.T.; Sun, Y.; Dickens, B.L. A Systematic Review of COVID-19 Epidemiology Based on Current Evidence.
_[J. Clin. Med. 2020, 9, 967. [CrossRef]](http://doi.org/10.3390/jcm9040967)_
20. Basu, A. Estimating The Infection Fatality Rate Among Symptomatic COVID-19 Cases In The United States. Health Aff. 2020, 39,
[1229–1236. [CrossRef]](http://doi.org/10.1377/hlthaff.2020.00455)
21. Gandhi, M.; Yokoe, D.S.; Havlir, D.V. Asymptomatic Transmission, the Achilles’ Heel of Current Strategies to Control COVID-19.
_[N. Engl. J. Med. 2020, 382, 2158–2160. [CrossRef]](http://doi.org/10.1056/NEJMe2009758)_
22. He, X.; Lau, E.H.Y.; Wu, P.; Deng, X.; Wang, J.; Hao, X.; Lau, Y.C.; Wong, J.Y.; Guan, Y.; Tan, X.; et al. Temporal dynamics in viral
[shedding and transmissibility of COVID-19. Nat. Med. 2020, 26, 672–675. [CrossRef]](http://doi.org/10.1038/s41591-020-0869-5)
23. [Oran, D.P.; Topol, E. Prevalence of Asymptomatic SARS-CoV-2 Infection. Ann. Intern. Med. 2020, 173, 362–367. [CrossRef]](http://doi.org/10.7326/M20-3012)
24. Hartnett, K.P.; Kite-Powell, A.; Devies, J.; Coletta, M.A.; Boehmer, T.K.; Adjemian, J.; Gundlapalli, A.V. Impact of the COVID-19
Pandemic on Emergency Department Visits—United States, January 1, 2019–May 30, 2020. MMWR. Morb. Mortal. Wkly. Rep.
**[2020, 69, 699–704. [CrossRef] [PubMed]](http://doi.org/10.15585/mmwr.mm6923e1)**
25. Ouerfelli, N.; Vrinceanu, N.; Coman, D.; Cioca, A.L. Empirical Modeling of COVID-19 Evolution with High/Direct Impact on
[Public Health and Risk Assessment. Int. J. Environ. Res. Public Health 2022, 19, 3707. [CrossRef] [PubMed]](http://doi.org/10.3390/ijerph19063707)
26. Ganasegeran, K.; Jamil, M.F.A.; Appannan, M.R.; Ch’Ng, A.S.H.; Looi, I.; Peariasamy, K.M. Spatial Dynamics and Multiscale
Regression Modelling of Population Level Indicators for COVID-19 Spread in Malaysia. Int. J. Environ. Res. Public Health 2022,
_[19, 2082. [CrossRef] [PubMed]](http://doi.org/10.3390/ijerph19042082)_
27. Rypdal, K.; Bianchi, F.M.; Rypdal, M. Intervention Fatigue is the Primary Cause of Strong Secondary Waves in the COVID-19
[Pandemic. Int. J. Environ. Res. Public Health 2020, 17, 9592. [CrossRef]](http://doi.org/10.3390/ijerph17249592)
28. [Siegenfeld, A.F.; Bar-Yam, Y. The impact of travel and timing in eliminating COVID-19. Commun. Phys. 2020, 3, 1–8. [CrossRef]](http://doi.org/10.1038/s42005-020-00470-7)
29. Howard, J.; Huang, A.; Li, Z.; Tufekci, Z.; Zdimal, V.; van der Westhuizen, H.-M.; von Delft, A.; Price, A.; Fridman, L.; Tang, L.-H.;
[et al. An evidence review of face masks against COVID-19. Proc. Natl. Acad. Sci. USA 2021, 118, e2014564118. [CrossRef]](http://doi.org/10.1073/pnas.2014564118)
30. Kingdon, J.W. Agendas, Alternatives, and Public Policies; Little, Brown: Boston, MA, USA, 1984.
31. [Reich, M.R. Political economy analysis for health. Bull. World Health Organ. 2019, 97, 514. [CrossRef]](http://doi.org/10.2471/BLT.19.238311)
32. Walt, G.; Shiffman, J.; Schneider, H.; Murray, S.F.; Brugha, R.; Gilson, L. ‘Doing’ health policy analysis: Methodological and
[conceptual reflections and challenges. Health Policy Plan. 2008, 23, 308–317. [CrossRef]](http://doi.org/10.1093/heapol/czn024)
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://pmc.ncbi.nlm.nih.gov/articles/PMC9140577, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "GOLD",
"url": "https://www.mdpi.com/1660-4601/19/10/6119/pdf?version=1652854362"
}
| 2,022
|
[
"JournalArticle",
"Review"
] | true
| 2022-05-01T00:00:00
|
[
{
"paperId": "b71197b3384e5ce4cd75b1a2fba999416692e76a",
"title": "Empirical Modeling of COVID-19 Evolution with High/Direct Impact on Public Health and Risk Assessment"
},
{
"paperId": "d3f99b2ed1d23d37277228937fe96cf92570f5f2",
"title": "Spatial Dynamics and Multiscale Regression Modelling of Population Level Indicators for COVID-19 Spread in Malaysia"
},
{
"paperId": "bb5fcc2c315f2d8570bdfe0a8198e2e19acb1146",
"title": "A study of computational and conceptual complexities of compartment and agent based models"
},
{
"paperId": "7a630f287498fba98ec1f9bdde82a3fef6e83ba9",
"title": "Experience of isolation and quarantine hotels for COVID-19 in Hawaii"
},
{
"paperId": "01db2ee5d371a513a911cc5144c76dd735b896ec",
"title": "Prevalence of Asymptomatic SARS-CoV-2 Infection"
},
{
"paperId": "c3b6e328dca9c135a50edd6b33f541a5a7405143",
"title": "An evidence review of face masks against COVID-19"
},
{
"paperId": "17ff703ad8ce6e2bea88584193adceafef4c137e",
"title": "Motivations and Barriers for the Use of Face Coverings during the COVID-19 Pandemic: Messaging Insights from Focus Groups"
},
{
"paperId": "48560927862f49035fb08a5942a31154a570e850",
"title": "COVID-19 Policy Differences across US States: Shutdowns, Reopening, and Mask Mandates"
},
{
"paperId": "9556743ebd32eaf70aa5add82200b45ff6431490",
"title": "Addressing Health Disparities in the Rural United States: Advocacy as Caregiving among Community Health Workers and Promotores de Salud"
},
{
"paperId": "f0d1df1c588eec17f881730b0087b776d49a60f8",
"title": "Intervention Fatigue is the Primary Cause of Strong Secondary Waves in the COVID-19 Pandemic"
},
{
"paperId": "b8a0ff6de3d9469d67fdbfd402ca0d3c972666ad",
"title": "The impact of travel and timing in eliminating COVID-19"
},
{
"paperId": "1a2e2784eb0df80643c7f1adac6079ea5e0254b3",
"title": "Community Use Of Face Masks And COVID-19: Evidence From A Natural Experiment Of State Mandates In The US."
},
{
"paperId": "a0a9184e748a58fe25e051bab4a6c338b358e825",
"title": "Impact of the COVID-19 Pandemic on Emergency Department Visits — United States, January 1, 2019–May 30, 2020"
},
{
"paperId": "0e7c2681bc8b0350f563663774fbe19dce3c6b27",
"title": "Wrong but Useful - What Covid-19 Epidemiologic Models Can and Cannot Tell Us."
},
{
"paperId": "0c5a3cbb9069454e95dc38099e979ae80b815525",
"title": "Estimating The Infection Fatality Rate Among Symptomatic COVID-19 Cases In The United States."
},
{
"paperId": "526625cf5000ab29d70641c70564d1f49b42946a",
"title": "COVID-19 Scenarios: an interactive tool to explore the spread and associated morbidity and mortality of SARS-CoV-2"
},
{
"paperId": "9712e670200fbc081c1115a5789086a59b0edd50",
"title": "Forecasting the impact of the first wave of the COVID-19 pandemic on hospital demand and deaths for the USA and European Economic Area countries"
},
{
"paperId": "e836dabc407abecd889905eca49f2ed07a702e58",
"title": "Asymptomatic Transmission, the Achilles’ Heel of Current Strategies to Control Covid-19"
},
{
"paperId": "9f138bc04b187093dffd7a041ad35495ce8aa5f8",
"title": "Hospitalization Rates and Characteristics of Patients Hospitalized with Laboratory-Confirmed Coronavirus Disease 2019 — COVID-NET, 14 States, March 1–30, 2020"
},
{
"paperId": "4fce764b00de92faae0af5386a55350db9a46dc8",
"title": "Caution Warranted: Using the Institute for Health Metrics and Evaluation Model for Predicting the Course of the COVID-19 Pandemic"
},
{
"paperId": "1e1b26fb478f39cca24208072bc12a7519548d46",
"title": "Why daily death tolls have become unusually important in understanding the coronavirus pandemic"
},
{
"paperId": "4dd6fd8bb608ec3e9d6835578f0d16fe320b605d",
"title": "Learning as We Go: An Examination of the Statistical Accuracy of COVID19 Daily Death Count Predictions"
},
{
"paperId": "4c970b611ff10313bd88b7c3e197c45c06a67feb",
"title": "A Systematic Review of COVID-19 Epidemiology Based on Current Evidence"
},
{
"paperId": "0813446cd9dcf350a7212280b8db7ee3fd05b970",
"title": "Severe Outcomes Among Patients with Coronavirus Disease 2019 (COVID-19) — United States, February 12–March 16, 2020"
},
{
"paperId": "a5a77eb6c25592824e322c47e4c2de04676bd8ae",
"title": "Temporal dynamics in viral shedding and transmissibility of COVID-19"
},
{
"paperId": "0874966fc2b887ece8ce43d4dd09ec180cb11615",
"title": "Report 9: Impact of non-pharmaceutical interventions (NPIs) to reduce COVID19 mortality and healthcare demand"
},
{
"paperId": "88479b44437b28f96cb357a0ce15a71da77fe0d6",
"title": "Political economy analysis for health"
},
{
"paperId": "ed860409ad73160875dcb38e972d302f7b336dfe",
"title": "‘Doing’ health policy analysis: methodological and conceptual reflections and challenges"
},
{
"paperId": "df03e1844e385324d61af7ca3587d4411ee81ae4",
"title": "THE USE OF EPIDEMIOLOGICAL MODELS FOR THE MANAGEMENT OF ANIMAL DISEASES"
},
{
"paperId": "8488b9d69fa47093b6cf77562473d0333ece1896",
"title": "Agendas, alternatives, and public policies"
},
{
"paperId": null,
"title": "Critique of the IHME Model for COVID-19 Projections"
},
{
"paperId": null,
"title": "Team IC 19 Health Service Utilization Forecasting"
},
{
"paperId": null,
"title": "Epidemic Calculator"
}
] | 15,361
|
en
|
[
{
"category": "Economics",
"source": "external"
},
{
"category": "Economics",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/022baa2ab01c85223ced326773cfec28142ff784
|
[
"Economics"
] | 0.894598
|
Gauging Market Responses to Monetary Policy Communication
|
022baa2ab01c85223ced326773cfec28142ff784
|
The Review
|
[
{
"authorId": "69896945",
"name": "Kevin L. Kliesen"
},
{
"authorId": "50153875",
"name": "Brian Levine"
},
{
"authorId": "49372646",
"name": "Christopher J. Waller"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"International Conference on Remote Engineering and Virtual Instrumentation",
"Rev",
"REV",
"Int Conf Remote Eng Virtual Instrum"
],
"alternate_urls": null,
"id": "86ecfb72-3296-4092-bd46-03d007d47bd0",
"issn": "1941-532X",
"name": "The Review",
"type": "conference",
"url": null
}
|
The modern model of central bank communication suggests that central bankers prefer to err on the side of saying too much rather than too little. The reason is that most central bankers believe that clear and concise communication of monetary policy helps achieve their goals. For the Federal Reserve, this means to achieve its goals of price stability, maximum employment, and stable long-term interest rates. This article examines the various dimensions of Fed communication with the public and financial markets and how Fed communication with the public has evolved over time. We use daily and intraday data to document how Fed communication affects key financial market variables. We find that Fed communication is associated with changes in prices of financial market instruments such as Treasury securities and equity prices. However, this effect varies by type of communication, by type of instrument, and by who is doing the speaking.
|
The modern model of central bank communication suggests that central bankers prefer to err on
the side of saying too much rather than too little. The reason is that most central bankers believe
that clear and concise communication of monetary policy helps achieve their goals. For the Federal
Reserve, this means to achieve its goals of price stability, maximum employment, and stable longterm interest rates. This article examines the various dimensions of Fed communication with the
public and financial markets and how Fed communication with the public has evolved over time.
We use daily and intraday data to document how Fed communication affects key financial market
variables. We find that Fed communication is associated with changes in prices of financial market
instruments such as Treasury securities and equity prices. However, this effect varies by type of
communication, by type of instrument, and by who is doing the speaking. (JEL E52, E58, E61, G10)
Federal Reserve Bank of St. Louis Review, Second Quarter 2019, 101(2), pp. 69-91.
https://doi.org/10.20955/r.101.69-91
**KEYNES: Arising from Professor Gregory’s questions, is it a practice of the Bank of England never to**
explain what its policy is?
**HARVEY: Well, I think it has been our practice to leave our actions to explain our policy.**
**KEYNES: Or the reasons for its policy?**
**HARVEY: It is a dangerous thing to start to give reasons.**
**KEYNES: Or to defend itself against criticism?**
**HARVEY: As regards criticism, I am afraid, though the Committee may not all agree, we do not admit**
there is a need for defense; to defend ourselves is somewhat akin to a lady starting to defend her virtue.
Exchange between John Maynard Keynes and Bank of England Deputy Governor Sir Ernest Harvey,
December 5, 1929.[1]
Kevin L. Kliesen is a business economist and research officer, Brian Levine was a senior research associate, and Christopher J. Waller is executive
vice president and director of research at the Federal Reserve Bank of St. Louis.
© 2019, Federal Reserve Bank of St. Louis. The views expressed in this article are those of the author(s) and do not necessarily reflect the views of
the Federal Reserve System, the Board of Governors, or the regional Federal Reserve Banks. Articles may be reprinted, reproduced, published,
distributed, displayed, and transmitted in their entirety if copyright notice, author name(s), and full citation are included. Abstracts, synopses,
and other derivative works may be made only with prior written permission of the Federal Reserve Bank of St. Louis.
**F d** **l R** **B** **k f St L** **i REVIEW** **S** **d Q** **t** **2019** **69**
-----
**Kliesen, Levine, Waller**
## INTRODUCTION
Central bank communication has come a long way since the Bank of England’s motto
ostensibly was “Never explain, never apologize.”[2] Today, the motto of central bankers might
instead be “Can you hear me now?” The modern model of central bank communication sug
gests that central bankers prefer to err on the side of saying too much rather than too little. In
this vein, central bank communication takes many forms, from economic forecasts and official
reports, to speeches, interviews, testimonies before governmental bodies, and policy statements
and press conferences immediately after policy meetings. In the United States, enhancements
in central bank communication are most pronounced in the realm of speeches and other
remarks (e.g., television interviews) by Federal Reserve (hereafter, Fed) governors and Reserve
Bank presidents. These forms of communication have become more prominent since the
recession and Financial Crisis. In an era of increased communication by Federal Open Market
Committee (FOMC) participants, one may ask whether additional information is useful for
financial market participants who carefully monitor monetary policy developments. Indeed,
some economists and analysts have argued that Fed officials talk too much.[3] There are many
nuances to this argument, but the primary claim is that more information increases the proba
bility of market mispricing. Shin (2017) discusses some of these issues.
There are at least two counterarguments to the market mispricing view. The first, as
enunciated by Kocherlakota (2017), is that the price of an independent central bank is a set
of independent voices to insure against group think. The second counterargument is that the
pricing of financial instruments in markets is more efficient with more, not less, information.
Regardless, central bank communication is important because individuals’ economic decisions
are based on expectations of future policies. Thus, clear communication of its policies and
actions may help the Fed achieve its mandated goals of stable prices, maximum employment,
and moderate long-term interest rates.
The purpose of this article is twofold. The first part examines the various dimensions of
Fed communication with the public and financial markets. This includes documenting how
communication with the public has evolved over time. The second part empirically analyzes
the economic effects of Fed communication on key financial market variables. Our analysis
uses daily and intraday data. We find that Fed communication can affect prices of financial
market instruments such as equities and Treasury securities. However, this effect varies by
type of communication, by type of instrument, and by who is doing the speaking. We also
find that larger financial market reactions tend to be associated with communication from
the Fed Chair, non-Chair Fed governors, and FOMC meetings without an associated press
conference. We further find that financial market reactions following press conferences after
FOMC meeting statements are not significant.
## HOW DOES THE FED COMMUNICATE?
As the exchange between John Maynard Keynes and Bank of England Deputy Governor
Sir Ernest Harvey demonstrated, the principles of central bank communication have evolved
**70** **S** **d Q** **t** **2019** **F d** **l R** **B** **k f St L** **i REVIEW**
-----
**Kliesen, Levine, Waller**
### Table 1
**Types of Fed Communication**
**Type** **Communicator** **Frequency** **Release timing**
Policy statement FOMC 8 times per year After each FOMC meeting, ~2 PM EST
Minutes FOMC 8 times per year 3 weeks after each FOMC meeting, ~2 PM EST
Press conference Chair 8 times per year* After designated FOMC meeting, ~2:30 PM EST
Summary of Economic Projections FOMC 4 times per year After designated FOMC meeting, ~2 PM EST
Monetary Policy Report to Congress Chair 2 times per year ~February and July of each year
Speeches and other public remarks FOMC Continuous[†] NA
Statement of Longer-Run Goals
FOMC 1 time per year Reaffirmed each January
and Policy Strategy
Policy Normalization Principles
FOMC Updated periodically After associated FOMC meeting, ~2 PM EST
and Plans[‡]
NOTE: Table reflects the present-day FOMC procedure. The timing and frequency of each event has changed over the past 20 years. ~Indicates
times are approximations and may differ slightly from event to event. *During the period analyzed, press conferences were held only four times
per year. Beginning in January 2019, press conferences are held after every FOMC meeting. [†]Excludes FOMC “blackout periods,” which begin the
second Saturday preceding an FOMC meeting and end the Thursday following the meeting. [‡]Initially released in September 2014. An addendum
[was adopted in March 2015 and augmented in June 2017. For a history of revisions, see https://www.federalreserve.gov/monetarypolicy/time](https://www.federalreserve.gov/monetarypolicy/timeline-policy-normalization-principles-and-plans.htm)
[line-policy-normalization-principles-and-plans.htm.](https://www.federalreserve.gov/monetarypolicy/timeline-policy-normalization-principles-and-plans.htm)
over time. A modern comparison describing the evolution of Fed communication was noted in
2003 by then Fed Governor Janet Yellen when she said that the FOMC “had journeyed from
‘never explain’ to a point where sometimes the explanation is the policy.”[4] Some have termed
this policy “open-mouth operations.”[5] Although views may differ between policymakers and
across central banks, the fundamental principles of central bank communication are founded
on the dual notions that increased transparency enhances the effectiveness of policy and the
accountability of policymakers in a democratic society.[6] In this article, we focus on Fed com
munication, though the principles and practices are similar among many of the world’s central
banks.
When analyzing central bank communication, the following questions come to mind: First,
who should do the talking; second, what should the central bank talk about; and, third, who
should the central bank talk to? There is a vast economic literature that attempts to answer
these questions. One notable early effort was a cross-country study by Blinder et al. (2001), who
surveyed communication methods and tactics, among other things. A subsequent article by
Blinder et al. (2008) argued that there was large variation in strategies but no consensus on the
best-practice approach to communicating monetary policy to the public. Woodford (2001)
was an early proponent of using communication to influence market expectations. This view
influenced several subsequent Fed officials, most notably former Fed Chairman Ben Bernanke.[7]
Finally, in the aftermath of the Financial Crisis of 2008, several event studies were published
that analyzed the FOMC’s unconventional policy actions on prices of financial market instru
ments, macroeconomic outcomes, and the expectations about future monetary policy actions.[8]
**F d** **l R** **B** **k f St L** **i REVIEW** **S** **d Q** **t** **2019** **71**
-----
**Kliesen, Levine, Waller**
In sum, the academic literature offers more support for the modern view of central bank
communication: More is generally better. Table 1 lists the primary methods that the Fed uses
to communicate its policies, procedures, and policy expectations to the public.[9] These methods
include the policy statement released at the end of the eight regularly scheduled FOMC meeting,
the minutes released three weeks after each of the eight regularly scheduled FOMC meetings,
the Chair’s quarterly press conference, along with speeches, testimonies, and media interviews
by Fed governors and Reserve Bank presidents. Some of these innovations are long standing,
such as the FOMC minutes, while others are more recent, such as the Chair’s press conferences.[10]
Given the prominence of FOMC policy statements as a communication instrument, the fol
lowing discussion will first briefly focus on their history and role.
### Policy Statements: Length and Readability
The Fed’s principle medium of communication is the policy statement released after each
FOMC meeting. The policy statement has evolved over time. From 1967 to 1992, the FOMC
issued a “Record of Policy Actions” (ROPA), which were initially released with a 90-day lag.[11]
Beginning under Chairman Alan Greenspan, the FOMC began to issue policy statements
immediately after the February 4, 1994, meeting. The first policy statement was rather short,
at 99 words, and made no mention of the intended federal funds target rate. Instead, the
inaugural statement indicated that the Committee decided to “increase slightly the degree of
pressure on reserve positions” in financial markets. In taking this action, the FOMC noted
that they expected an “associated small increase in short-term money market interest rates.”[12]
Following the release of the inaugural statement, the FOMC released a post-meeting state
ment four additional times in 1994. Three post-meeting statements were released in 1995,
including the statement released after the July 6, 1995, meeting, which was the first instance
that the FOMC specifically mentioned the federal funds rate. The FOMC continued to issue
post-meeting statements over the next few years, but only at meetings where a policy change
occurred. However, beginning with the May 18, 1999, meeting, statements were released after
each FOMC meeting.[13] The public focus on the policy statement was such that the financial
press developed a “briefcase barometer.”[14]
The post-meeting FOMC statements have evolved over time. Prior to the Financial Crisis,
the post-meeting policy statement mostly focused on the state of the economy and the Com
mittee’s rationale for raising or lowering the policy rate or reasons why the policy rate was not
changed. In general, less was said about the future path of interest rate changes. The policy
statement evolved to take on a larger role in communicating the stance of monetary policy
during the Financial Crisis after the federal funds rate reached the zero lower bound (ZLB)
on December 16, 2008.[15] Figure 1 shows that the word count of the policy statements began
to increase steadily in 2007 during the early stages of the Financial Crisis. The word count
continued to increase during the adoption of quantitative easing (QE) policies that both
increased the size of the balance sheet and changed its composition. Prior to the ZLB period,
the number of words in each statement averaged 223. During the ZLB period, the count was
more than twice as much, averaging 580 words.
After the nominal federal funds target rate reached the ZLB in December 2008, the Fed pro
vided the largest amount of monetary accommodation through balance sheet adjustments and
**72** **S** **d Q** **t** **2019** **F d** **l R** **B** **k f St L** **i REVIEW**
-----
**Kliesen, Levine, Waller**
### Figure 1
**FOMC Statement Word Count**
Number of Words in Each Statement
QE1 QE2 MEP QE3
1,000
900
800
700
600
500
400
300
200
100
0
1994 1999 2001 2002 2004 2006 2008 2009 2011 2013 2015 2017
NOTE: Shaded area indicates the period of the FOMC’s unconventional monetary policy with interest rates at the effective ZLB. MEP, Maturity
Extension Program. Under the MEP, the Fed sold or redeemed shorter-term Treasury securities and used the proceeds to buy longer-term
Treasury securities, thereby extending the average maturity of the securities in the Fed’s portfolio. Updated through 2017.
SOURCE: Board of Governors of the Federal Reserve System.
other unconventional policies.[16] But as the U.S. economy transitioned from recession to a
slower-than-average recovery, the Fed’s policy approach also changed. The new approach
focused instead on influencing the public’s expectations of the future direction and level of the
federal funds target rate. This approach, in its current form, is referred to as forward guidance.[17]
For example, following the August 9, 2011, meeting, the policy statement stated the following:
The Committee currently anticipates that economic conditions—including low rates of
resource utilization and a subdued outlook for inflation over the medium run—are likely
to warrant exceptionally low levels for the federal funds rate at least through mid-2013.
In this case, the FOMC’s intent was to signal to the public that its policy rate would remain
low for a long time in order to spur the economy’s recovery. This signal was meant to be taken
as a public commitment, what Campbell et al. (2012) termed “Odyssean” policy. Using lan
guage from Greek mythology, Odyssean policy is meant to convey a public commitment not
to change policy for a certain period—in this case, for more than two years. Instead, the public
appeared to view this statement as a forecast, what Campbell et al. (2012) termed “Delphic”
policy. In effect, the Delphic statement strongly suggested that, in the FOMC’s view, the eco
nomic weakness would persist for more than two years. However, at the June 2011 meeting
two months earlier, the Summary of Economic Projections (SEP) indicated that real gross
domestic product (GDP) would increase by 3.5 percent in 2012 and by 3.9 percent in 2013
(each measure is the midpoint of the central tendency).[18] Thus, by August, the Committee
appeared to have concluded that it, like most private sector forecasters, had been much too
optimistic about the pace of real GDP growth during the early stages of the expansion. Indeed,
by the January 2012 meeting, forecasts for real GDP growth in 2012 and 2013 had been marked
down to 1.7 percent and 2.5 percent, respectively.
**F d** **l R** **B** **k f St L** **i REVIEW** **S** **d Q** **t** **2019** **73**
-----
**Kliesen, Levine, Waller**
### Figure 2
**FOMC Statement Complexity**
Flesch-Kincaid Reading Grade Level
QE1 QE2 MEP QE3
24
22
20
18
16
14
12
10
8
6
1994 1999 2001 2002 2004 2006 2008 2009 2011 2013 2015 2017
NOTE: Shaded area indicates the period of the FOMC’s unconventional monetary policy with interest rates at the effective ZLB. MEP, Maturity
Extension Program. Under the MEP, the Fed sold or redeemed shorter-term Treasury securities and used the proceeds to buy longer-term
Treasury securities, thereby extending the average maturity of the securities in the Fed’s portfolio. Updated through December 2017.
SOURCE: Board of Governors of the Federal Reserve System (FOMC statements) and Educational Testing Service (word count).
To accomplish the Fed’s goals and objectives in a slow-growth economy, the post-meeting
statement changed in two dimensions. The first change, as noted above, was that the length
increased. The statements included more discussion of the economic situation and its impli
cation for the near-term direction of policy (changes in the federal funds target rate).[19] Second,
the statements incorporated more complex economic terms and analysis. This is shown in
Figure 2, which uses text evaluation software to measure the Flesch-Kincaid reading grade
level of the policy statement. A higher grade level is assumed to reflect increased complexity
of the statement. Prior to the ZLB period, the median grade level was 13.5, indicating com
prehension accessible to someone reading at a college undergraduate level. But by late 2013,
when the FOMC was in the midst of increasing the size of its balance sheet through asset pur
chases, the grading level rose to 20, which is commensurate to a graduate school reading level.
For the entire ZLB period, the grade level rose to 16 (median), but then fell to 15 (median)
during the post-ZLB period.[20] Researchers find that the readability of central bank policy
statements and remarks are an important factor in how they are received by financial markets.
Not surprisingly, clearer statements lead to lower volatility.[21]
This section has highlighted how the FOMC changed the length and composition of the
policy statement during the period of unconventional monetary policy. But the policy state
ment is only one form of central bank communication. Speeches and other public remarks
are another form of communication that policymakers have deployed to increase the public’s
knowledge of the prevailing monetary policy regime. The next two sections will delve into
monetary policy communication strategies by Fed officials, both old and new.
**74** **S** **d Q** **t** **2019** **F d** **l R** **B** **k f St L** **i REVIEW**
-----
**Kliesen, Levine, Waller**
### Figure 3
**Number of Public Remarks by Type of Fed Official**
Total Remarks Per Year
250
ZLB Period Begins
(December 2007)
200
**Bank Presidents**
150
100
**Non-Chair Governors**
50
**FOMC Chair**
0
1998 2000 2002 2004 2006 2008 2010 2012 2014 2016
NOTE: Through 2017.
SOURCE: Board of Governors of the Federal Reserve System, the 12 Federal Reserve Banks, Bloomberg, and authors’
calculations.
### Public Remarks by Fed Officials
Fed officials have long used other forms of public communication besides policy state
ments.[22] Public remarks can take many forms, including formal speeches, Congressional
testimonies, interviews with the financial media, or published articles and commentaries.
Sometimes, Fed officials do not comment on monetary policy issues that may be discussed at
recent or upcoming FOMC meetings. In those instances, policymakers may instead choose
to focus on other issues, such as local economic conditions, economic education, community
development, or banking and financial market regulation.
The ZLB period witnessed an unprecedented rate of spoken and written communication
with the public by Fed governors and Reserve Bank presidents. Figure 3 shows the annual
number of public remarks by the Fed Chair, non-Chair governors, and Reserve Bank presi
dents since 1998.[23] From 1998 to 2004, the total number of public remarks by Reserve Bank
presidents remained roughly constant at about 150 per year. A slightly different pattern
occurred with governors and the Fed Chair. Total remarks over this period steadily fell, but
then rebounded, so that the numbers of public remarks in 2004 were close to the 1998 totals.
Beginning in 2005, the total number of public remarks by Reserve Bank presidents began to
increase, reaching a peak in 2013 of a little more than 220 public remarks. Interestingly, though,
the FOMC Chair and governors delivered public remarks slightly less frequently over the
ZLB period. Some of the reduced frequency of public remarks by members of the Board of
Governors (excluding the Chair) reflects the fact that the Board has rarely operated with a
full complement of Governors (seven). From 1998 to 2017, there has only been four years
when there were seven governors present at the last formal meeting of the year. Indeed, at
the end of 2017, there were only four governors at the December meeting. At the March 2018
meeting, the number of governors had dwindled to three.
**F d** **l R** **B** **k f St L** **i REVIEW** **S** **d Q** **t** **2019** **75**
|al Remarks Per Year|Col2|
|---|---|
|ZLB Period Begins (December 2007)||
||Bank Presidents|
|||
||Non-Chair Governors|
|FOMC Chair||
-----
**Kliesen, Levine, Waller**
### Figure 4
**Number of Times More Than One Bank President or Governor Spoke on the Same Day, 1998-2017**
Frequency
80
70
60
50
40
30
20
Governors
Bank Presidents
60
10
3
0
1998 2000 2002 2004 2006 2008 2010 2012 2014 2016
SOURCE: Board of Governors of the Federal Reserve System, Bloomberg, and authors’ calculations.
Speeches have become important communication events. Chairman Greenspan’s new
economy speech in 1995 and his “irrational exuberance” speech in 1996 were among his more
notable speeches. Chairman Ben Bernanke also gave notable speeches during his tenure. Two
that standout are his “Deflation: Making Sure ‘It’ Doesn’t Happen Here” speech in 2002 and
his global saving glut speech in 2005. Days with multiple Fed communication events have
become more numerous over time—particularly since the Financial Crisis. Figure 4 shows
the increase in multiple Fed communication events on the same day stems from an increase
in more than one Reserve Bank president speaking on the same day. For example, in 2017,
there were 60 days when more than one Reserve Bank president spoke. In 2004, it was about
half as much. By contrast, in 2017 there were only three days when more than one Fed
governor spoke publicly on the same day. This is down sharply from 2003, when there were
19 days when multiple Fed governors spoke on the same day.[24]
In separate analysis, we looked at the annual number of public remarks by Reserve Bank
presidents from January 1998 to December 2017. We separated the sample into roughly two
10-year periods: January 1998 to August 2008 (pre-Financial Crisis) and September 2008 to
December 2017 (post-Financial Crisis). The number of public remarks by Reserve Bank
presidents increased in all but three Fed Districts (Chicago, New York, and Richmond). The
average increase in volume across these nine Districts was 46 percent. We did not examine
whether the nature of the remarks by Reserve Bank presidents has changed over time. We did,
however, analyze the number of speeches and public remarks given by presidents of the Fed
Bank of St. Louis since January 1929. We have documented this in the boxed insert.
### Other Forms of Fed Communication
In the past several years, chiefly under the Bernanke regime, the FOMC has adopted
several new forms of communication to further increase transparency. As noted earlier, the
Chair’s quarterly press conference, beginning under Chairman Bernanke’s term in January
**76** **S** **d Q** **t** **2019** **F d** **l R** **B** **k f St L** **i REVIEW**
-----
**Kliesen, Levine, Waller**
2012, is one key innovation. Current Chairman Jerome Powell expanded on this innovation,
announcing that press conferences will be held after every FOMC meeting beginning in January
2019. Other innovations include the FOMC’s “Statement of Longer-Run Goals and Monetary
Policy Strategy,” “Policy Normalization Principles and Plans,” and “Summary of Economic
Projections” (SEP). These are also listed in Table 1. The first two are meant to provide clarity
on the Fed’s dual mandate and balance sheet, respectively, while the SEP conveys projections
for four key macroeconomic variables. In addition, the SEP conveys each FOMC participant’s
assessment of appropriate monetary policy, as indicated by their federal funds rate projections
over short-, medium-, and longer-term horizons.
**F d** **l R** **B** **k f St L** **i REVIEW** **S** **d Q** **t** **2019** **77**
-----
**Kliesen, Levine, Waller**
### Grading Fed Communication
The Hutchins Center on Fiscal and Monetary Policy at Brookings conducted a survey of
academics and private sector Fed watchers to assess the effectiveness of different forms of
Fed communication.[25] Survey participants viewed the FOMC policy statement, speeches by
the FOMC Chair, and quarterly press conferences as the most useful forms of Fed communi
cation. On net, academics generally found these forms of communication more useful than
did the private sector economists and Fed watchers.
One of the key communication innovations during the Bernanke tenure was the public
release of individual FOMC participants’ expectations of the future level of the federal funds
rate. Once a quarter, with the release of the SEP, each FOMC participant—anonymously—
indicates their preference for the level of the federal funds rate at the end of the current year,
at the end of the next two to three years, and over the “longer run.” According to the survey,
these projections are often termed the FOMC “dot plots.” Both academics and those in the
private sector found the dot plots of limited use as an instrument of Fed communication
(more “useless” than “useful”). One-third of the respondents found the dot plots “useful or
extremely useful,” 29 percent found them “somewhat useful,” and 38 percent found them
“useless or not very useful.”
The limited usefulness of the dot plots probably reflects many factors. First, each partici
pant’s projection is conditioned on the highly restrictive assumption of “appropriate monetary
policy.” Each participant’s appropriate monetary policy stance is conditioned on their view
of the outlook for real GDP growth, inflation, and the unemployment rate over the medium
term. Moreover, the range of participants’ views may not dovetail with the policy path out
lined in the FOMC statement, which can further complicate the communicated outlook and
diminish the tool’s effectiveness. The regular presence of dissents suggests that appropriate
policy can differ sharply across the Committee.
Second, the participants may have other vastly different assumptions that influence their
outlook, such as the equilibrium real interest rate, the future path of crude oil prices, the for
eign exchange value of the dollar, or their outlook for foreign economic growth. For these
reasons and more, FOMC participants persistently over-projected the federal funds target
rate path during the early years of the current expansion. [See earlier discussion on page 73.]
These persistent one-sided forecast errors may have impaired the credibility of the dot plots to
the extent that the projections were important inputs in establishing expectations about future
monetary policy.
Finally, the Brookings study revealed that survey participants believe that Reserve Bank
presidents’ speeches are slightly less useful than the dot plots, but still more useful than Fed
reports to Congress, such as the semi-annual Monetary Policy Report.[26] This finding is per
haps striking given that the number of public remarks by Reserve Bank presidents has been
trending up over time, especially during the ZLB period, while the number of public remarks
by the Chair and non-Chair governors has been trending down.
**78** **S** **d Q** **t** **2019** **F d** **l R** **B** **k f St L** **i REVIEW**
-----
**Kliesen, Levine, Waller**
## EMPIRICAL ANALYSIS
The final section of the article assesses how financial market participants respond to vari
ous forms of Fed communication. Admittedly, this is a difficult empirical exercise for many
reasons. First, public remarks by Fed senior officials are often context- and perspectivedependent. Each individual brings their own perspective, model of the economy, and view of
the monetary policy transmission mechanism. These views naturally inform their assessments
of appropriate monetary policy going forward, which are then conveyed in public remarks.
For their part, financial market participants may become familiar with a given policymaker’s
view or assume a given outcome for a particular FOMC meeting. If so, markets may react only
to views that are sufficiently different from expectations. Past research has demonstrated that
monetary policy surprises can have significant effects on high-frequency asset prices.[27] We
acknowledge the importance of monetary policy surprises, but use a different approach to
assess the significance of Fed communication events.
Second, when attempting to gauge the significance of public remarks, markets do not
usually assign equal weights to all FOMC participants. Certainly, markets carefully parse
remarks by the Chair, who is typically viewed as the public voice of the FOMC and the one
who sets the policy agenda. Moreover, while the Chair’s views often convey the consensus view
of the Committee, the Chair nonetheless also has a policy preference. Although the Chair’s
preference invariably prevails, dissents still occur periodically. Indeed, Reserve Bank presidents
sometimes use their public remarks, or dissents, with the intention of signaling future policy
preferences or advocating for alternative frameworks.[28] Still, markets may discount the views
of the presidents, on average, because they believe their views unnecessarily distort market
signals or future policy intentions. For example, Lustenberger and Rossi (2017) claim that
remarks by Reserve Bank presidents worsen the accuracy of private sector forecasts.
With these caveats in mind, we adopt a two-pronged empirical exercise. The first exercise
uses daily data to examine whether Fed communication events are associated with significant
movements in key financial market variables. Admittedly, this approach has some drawbacks.
First, daily financial market data tend to be more volatile compared with monthly or quarterly
data. Second, this volatility arises, in part, because financial markets trade on many types of
information, such as macroeconomic data or global financial or geopolitical developments.
Thus, while Fed communication comprises one set of information the market uses to price
assets, there are potentially many other sources of information that the market uses that we
can’t readily account for. Our intent is to assess market reactions to Fed communication
events and not to model changes in asset price movements at a high frequency.
The second empirical exercise uses intraday data at 5-minute frequencies. Using intra
day data allows us to more closely match the timing of Fed communication events with the
responses in financial markets. This is the approach adopted by most of the aforementioned
event studies. Our intent is to determine if the empirical results using the daily data are con
sistent with those from the intraday data. Before presenting the results, we provide a detailed
description of our data sources and approach.
**F d** **l R** **B** **k f St L** **i REVIEW** **S** **d Q** **t** **2019** **79**
-----
**Kliesen, Levine, Waller**
### Data Sources and Approach
We study the effects of seven types of Fed communication events: FOMC meeting state
ments;[29] FOMC minutes; Fed Chair press conferences; public remarks by the Fed Chair,
non-Chair Fed governors, and Reserve Bank presidents; and unconventional monetary policy
announcements.[30] Five of the seven categories are included in the Brookings study. It is
important to note that there is an overlap between FOMC meetings and six of the seven
unconventional policy announcements we include.[31] Initially, our data set included public
remarks made after market hours and on weekends. Consistent with some of the literature,
we initially moved an after-hours communication event to the following trading day to gauge
the market’s reaction to the remark. However, this approach ended up producing large reac
tions that were probably not tied to the public remark itself. For example, many key data
releases are often issued before the market opens.[32] In this case, it is difficult to determine
whether the market is responding to the public remarks by a Fed official or to economic data
releases that may be a surprise.[33]
### Empirical Analysis: Daily Data
We create a series of dummy variables for the Fed communication events. Because the
Brookings study found that survey participants viewed the Fed Chair press conferences as a
useful form of communication, we identify regularly scheduled FOMC meetings with and
without an associated press conference. In recent years, FOMC press conferences have
occurred after the March, June, September, and December meetings. Since the liftoff from
the ZLB at the December 2015 meeting, increases in the FOMC’s federal funds target rate
have occurred at meetings with an associated press conference by the Fed Chair. Our sample
period is January 6, 1998, to December 29, 2017. There are nine types of communication
events:
- Non-press conference FOMC meeting statements
- Press conference FOMC meeting statements
- Releases of FOMC minutes
- Remarks by the FOMC Chair
- Remarks by all other Fed governors
- Remarks by Reserve Bank presidents
- Days when there are multiple Fed communication events (e.g., speeches)
- Unconventional policy actions (e.g., large-scale asset purchases)
- Key macroeconomic data releases (e.g., industrial production)
We evaluate the market reaction for three financial instruments: the absolute value of
the daily change in the yield on 2-year Treasury notes, the yield on 10-year Treasury notes,
and the Chicago Board Options Exchange equity market volatility index (VIX). Changes in
2-year Treasury yields are widely viewed as being sensitive to expected changes in FOMC
policy. The 10-year Treasury yield is the most liquid, long-term, risk-free interest rate in the
financial markets. It is also sensitive to changes in inflation expectations and longer-term
expectations about short-term interest rates. Finally, the VIX, which is often termed the mar
**80** **S** **d Q** **t** **2019** **F d** **l R** **B** **k f St L** **i REVIEW**
-----
**Kliesen, Levine, Waller**
ket’s “fear gauge,” is sometimes viewed as signaling changes in economic uncertainty. This
exercise can be represented by the following equation:
ΔYi,t =α + β1ΔYi,t−1 + β2NPCi,t + β3PCi,t + β4MINi,t + β5CHAIRi,t + β6GOVi,t
+β7PRESi,t + β8MULTi,t + β9UNCONVi,t + β10MACROi,t,
where ∆Yi,t represents the absolute value of the daily change in financial variable i (either the
2-year Treasury yield, 10-year Treasury yield, or VIX) on day t. The independent variables
include a constant, a one-day lag of the dependent variable, and a series of dummy variables
(specified earlier in this section) that take the value of 1 if that event occurs on day t or are
zero if the event does not occur on day t.
We analyze daily data with three ordinary least-squares regressions. We use the absolute
value of the daily changes because some communication events will cause yields to increase
or decrease, while others will generate no market response. Using absolute values are a more
effective way to gauge the effects of communication events on financial market activity.[34] We
also include another dummy variable (MACRO) on days when key economic statistics are
released. The motivation for this is that the market trades on information contained in these
reports. Our economic statistic dummy variable takes the value of 1 when the following
monthly economic reports are released (and is zero on all other days): the consumer price
index, monthly employment situation, industrial production, retail sales, the Institute for
Supply Management Report on Manufacturing, and the three GDP releases (advance, second,
and third estimates).
Table 2 shows the results of our analysis using daily data. Daily data allow us to make a
few noteworthy observations. First, for the change in the 2-year Treasury yield, markets react
significantly (at the 1 or 5 percent level) to Fed Chair and Fed governor communication events
and also to FOMC statements at non-press conference meetings. Table 2 further indicates that
changes in 2-year Treasury securities do not react significantly on days when one Reserve
Bank president speaks, but they do react significantly on days when there are multiple Fed
speakers. (Recall from Figure 4 that the number of days with multiple Fed speakers has increased
since the Financial Crisis). Finally, 2-year yields also react significantly to macroeconomic
data releases. Unconventional policy actions are marginally significant (at the 8 percent level).
With the exception of days with multiple Fed speakers, the signs of the coefficients on the
significant variables are positive.
The second and third sets of regressions in Table 2 show results for the change in the
10-year Treasury yield and in the VIX. Traders of longer-term Treasury securities react broadly
similarly to Fed communication events and data releases as traders of 2-year Treasury secu
rities. For instance, 10-year yields react significantly to the Fed Chair’s remarks, on days when
there are multiple Fed speakers, and to macroeconomic data releases; the coefficients generally
have the same signs and magnitudes as those from the regression using 2-year yields. However,
there are some differences between the 2-year and 10-year responses. For example, the change
in the 10-year yield is significantly associated with unconventional policy actions. Moreover,
10-year yields do not react significantly to remarks by Fed governors or to non-press confer
ence FOMC statements.
**F d** **l R** **B** **k f St L** **i REVIEW** **S** **d Q** **t** **2019** **81**
-----
**Kliesen, Levine, Waller**
### Table 2
**Federal Reserve Communication Events and Financial Market Responses Using Daily Data**
**Dependent variables**
Independent variables **2-year Treasury** **10-year Treasury** **VIX**
0.023
Constant
(0.000)**
0.248
Lagged dependent variable
(0.000)**
0.013
Non-press conference FOMC meetings
(0.001)**
0.004
Press conference FOMC meetings
(0.594)
0.004
FOMC minutes
(0.154)
0.641
(0.000)**
0.321
(0.000)**
0.059
(0.588)
0.092
(0.626)
–0.078
(0.315)
0.018
(0.787)
0.020
(0.645)
0.093
(0.111)
–0.068
(0.265)
0.626
(0.176)
0.077
(0.031)*
0.006
FOMC Chair remarks
(0.001)**
0.004
Fed governor remarks
(0.003)**
0.000
Fed president remarks
(0.775)
–0.004
Multiple Fed speakers
(0.032)*
0.036
Unconventional policy actions
(0.080)
0.009
Macroeconomic data releases
(0.000)**
0.035
(0.000)**
0.132
(0.000)**
0.006
(0.187)
0.001
(0.880)
0.004
(0.309)
0.005
(0.012)*
0.001
(0.347)
0.000
(0.877)
–0.004
(0.018)*
0.044
(0.021)*
0.009
(0.000)**
Adjusted R-squared 0.089 0.040 0.109
Durbin-Watson statistic 2.107 2.043 2.264
NOTE: p-values listed in parentheses. The sample period is January 6, 1998 to December 29, 2017.
- and ** indicate significance at the 5 percent and 1 percent levels, respectively. Dependent variables are
expressed as the absolute value of their one-day changes.
Column 3 presents the results for the change in the VIX. Equity market volatility does
not react significantly to Fed communication events. The closest variable of significance
(p = 0.11) are remarks by Reserve Bank presidents. Equity market volatility does, however,
react significantly to macroeconomic data releases. Finally, in all three regressions, the con
stant and the lagged dependent variable are significant at the 1 percent level.
Figure 5 provides some visual evidence for the behavior of equity market volatility around
FOMC meetings: From January 1994 to December 2017, the VIX begins to rise about a week
before an FOMC meeting. The VIX then drops relatively sharply (nearly 3 percent) on the
day the FOMC statement is released. This finding suggests that equity markets appear to be
increasingly uncertain about the meeting outcome, or its effects on financial markets, in the
run-up to FOMC meetings.[35] Likewise, we see a noticeable reduction in market volatility
**82** **S** **d Q** **t** **2019** **F d** **l R** **B** **k f St L** **i REVIEW**
-----
**Kliesen, Levine, Waller**
### Figure 5
**Relative Changes in the VIX Near FOMC Announcement Days**
VIX = 100 on FOMC Announcement Day
104
103
102
101
100
99
98
97
–10 –9 –8 –7 –6 –5 –4 –3 –2 –1 FOMC 1 2 3 4 5 6 7 8 9 10
NOTE: Sample includes all regularly scheduled FOMC meetings between January 1994 and December 2017.
SOURCE: Haver Analytics and authors’ calculations.
after the policy announcement (statement), perhaps indicating a decline in uncertainty and
clearer understanding of the Fed’s reaction function. Finally, other than the lagged dependent
variable, the dummy variable that accounts for the release of key economic reports is the only
other independent variable that is statistically significant.
We now turn to the second approach of our empirical exercise, namely, examining the
effects of communication events on financial market outcomes using intraday data.
### Empirical Analysis: Intraday Data
We use intraday data to estimate the effects of Fed communication on key financial market
variables. Many researchers have used intraday data to gauge market reactions to monetary
policy surprises or to the Fed’s announcements of unconventional polices after the Financial
Crisis. These event studies, as they are often called, are intended to measure the financial
market’s response to news at intervals measured in minutes. Our analysis of the market’s
response to Fed communication events generally follows the form and practice of the event
study literature.
Event studies can be criticized for many reasons. First, the studies gauge only the initial
announcement responses rather than the responses across time. Second, the results can be
sensitive to the choice of window size—that is, responses evaluated over a 1-minute window
versus a 5- or 10-minute window. Third, responses could be affected by non-announcement
effects, such as from economic data releases or geopolitical events. In view of these concerns,
we tested several different window sizes for robustness and used minute-by-minute asset price
data for the S&P 500 stock prices and 10-year Treasury futures prices.[36] For FOMC meeting
statements, FOMC minutes, and unconventional policy announcements (non-speaker events),
a window of plus or minus 15 minutes is used. For FOMC press conferences and other public
remarks (speaker events), a window of 15 minutes before to 60 minutes after the event is used.
We do not find that the interpretation of the results meaningfully changes when the event
window is adjusted.[37]
**F d** **l R** **B** **k f St L** **i REVIEW** **S** **d Q** **t** **2019** **83**
-----
**Kliesen, Levine, Waller**
Since these two variables are both prices, we calculate the percent change in each series
over each event window.[38] We then summarize this information via two metrics: mean
absolute change and cumulative change. For non-speaker events, where the event window is
+/–15 minutes, the mean absolute change can be represented as
MACNonj -Speaker = N[1] i∑N=1 ⎛⎝[⎜] YYii,,jjtt+−1515 −1⎞⎠[⎟] [*] [100,]
where Yi,t[j] [ represents, for each non-speaker event category ][j][, the asset price associated with ]
observation i at time t, and N represents the total number of observations for each nonspeaker event category j over the sample.
Likewise, for speaker events, where the event window is –15/+60 minutes, the mean
absolute change can be represented as
MACSpeakerj = N[1] i∑N=1 ⎛⎝[⎜] YYii,,jjtt+−1560 −1⎞⎠[⎟] [*] [100,]
using the same notation as before.
The cumulative change is calculated similarly, but we are now summing (instead of
averaging) over our sample, and we are not taking the absolute value beforehand. We represent
this as
CCNonj -Speaker = ∑iN=1 ⎡⎢⎢⎣⎛⎝[⎜] YYii,,jjtt+−1515 −1⎞⎠[⎟] [*] [100]
⎤
⎥
⎥⎦
and
CCSpeakerj = ∑iN=1 ⎢⎢⎣⎡⎛⎝[⎜] YYii,,jjtt+−1560 −1⎞⎠[⎟] [*] [100]
⎤
⎥
⎥⎦
for non-speaker and speaker events, respectively, again using the same notation as before.
The results are shown in Figure 6A, Figure 6B, and Figure 7. The grouping on the left
side of Figure 6A shows the mean absolute changes in the S&P 500 index in response to Fed
communication events not associated with an individual Fed official (non-speaker events),
while the grouping on the right side of Figure 6A shows those is response to events with public
remarks by a Fed official (speaker events).[39] On the left side, we find that stock prices react
most strongly to unconventional policy actions—indeed, twice as strong as the next-largest
event (FOMC meeting statements). This finding appears consistent with the event study litera
ture cited earlier. On the right side, stock prices react the most to the Chairs’ press conferences
and their remarks. In contrast, stock price changes in response to Fed communication events
by Reserve Bank presidents and Governors are similar in magnitude.
Figure 6B shows the same calculation for 10-year Treasury bond futures prices. The results
in Figure 6B are broadly similar to those in Figure 6A. In particular, responses to unconven
tional policies are substantially larger than to other forms of Fed communication, such as
FOMC meeting statements. As with stock prices, bond markets appear to react more strongly
**84** **S** **d Q** **t** **2019** **F d** **l R** **B** **k f St L** **i REVIEW**
-----
**Kliesen, Levine, Waller**
### Figure 6A
**Mean Absolute Changes in S&P 500 Index**
Percent
0.70
0.7
0.6
0.5
0.4 0.35 0.35 0.33
0.3 0.28 0.26 0.25
0.21
0.2 0.16
0.1
0.0
NOTE: Underlying data are expressed as a percent change in the S&P 500 index over an event window of –15/+15
minutes (non-speaker events) or of –15/+60 minutes (speaker events). The absolute values of these percent changes
are then averaged, for each event category, over the full sample. Non-event days are days with no Fed communication
event. Non-speaker events and speaker events have different non-event day controls because they are associated with
different window sizes.
|Col1|Col2|Non-Speakers|Col4|Col5|Col6|Col7|Col8|Col9|Speakers|Col11|Col12|Col13|Col14|Col15|Col16|Col17|Col18|Col19|Col20|
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
|||||||||||||||||||||
|||||||||||||||||||||
|||0.35|||||||0.35 0.33 0.28|||||||||||
|||||||||||||||||||||
|||||0.21 0.16|||||||||0.26 0.25|||||||
|||||||||||||||||||||
|||||||||||||||||||||
|||||||||||||||||||||
|||||||||||||||||||||
### Figure 6B
**Mean Absolute Changes in 10-Year Treasury Futures Prices**
Percent
0.68
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
|Col1|Col2|Non-Speakers|Col4|Col5|Col6|Col7|Speakers|Col9|Col10|Col11|Col12|Col13|Col14|Col15|Col16|Col17|Col18|
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
|||||||||||||||||||
|||||||||||||||||||
|||||||||||||||||||
|||0.22|||||0.18|||||||||||
|||||||||||||||||||
|||||0.09|||||0.14 0.12 0.10 0.08|||||||||
|||||||0.05||||||||||||
NOTE: Underlying data are expressed as a percent change in 10-year Treasury futures price over an event window of
–15/+15 minutes (non-speaker events) or of –15/+60 minutes (speaker events). The absolute values of these percent
changes are then averaged, for each event category, over the full sample. Non-event days are days with no Fed com
munication event. Non-speaker events and speaker events have different non-event day controls because they are
associated with different window sizes.
**F d** **l R** **B** **k f St L** **i REVIEW** **S** **d Q** **t** **2019** **85**
-----
**Kliesen, Levine, Waller**
### Figure 7
**Cumulative Changes for Fed Communication Events**
Percentage Points
4.0
S&P 500
2.0
0.0
|S&P 500|Col2|Col3|3.19|3.39|Col6|
|---|---|---|---|---|---|
|10-Year Treasury Futures 1.76||||||
|1.32 1.32||||||
|||||||
|||–0.07 –0.54||||
–2.0
–1.38
Meetings Minutes Press Conferences Unconventional
NOTE: Underlying data are expressed as a percent change in the index (S&P 500) or price (10-year Treasury futures) over
the event window. These percent changes are then summed, by category, over the full sample. For FOMC meetings,
minutes, and unconventional policy measures, the window is –/+15 minutes. For press conferences, the window is
–15/+60 minutes. For illustrative purposes, other public remarks were removed from the figure because of very high
cumulative change values.
to meeting statements than the release of FOMC minutes. The right side of Figure 6B shows
that the bond market’s responses to the Chair’s press conferences and the Chair’s remarks
are appreciably larger than to non-Chair Fed governors and Reserve Bank presidents.
Figure 7 plots the cumulative changes for FOMC meeting statements, minutes, press
conferences, and unconventional monetary policy announcements. We exclude other events
for illustrative purposes, as they exhibit very high cumulative change values. Similar to the
findings in Figures 6A and 6B, unconventional policies are associated with large stock and
bond market responses during our sample. The cumulative change in stock prices associated
with FOMC press conferences is also relatively large and positive. However, for FOMC meeting
statements and the release of FOMC minutes, the cumulative response of stock prices is neg
ative, with the response of the latter more than double the former. The response of bond futures
prices to FOMC meeting statements is of the same magnitude as the minutes, but, again, far
smaller than to unconventional policies. For Chair press conferences, the near-zero cumula
tive change is not a function of the bond futures market ignoring this information; rather, it
is the result of large, positive price reactions negating large, negative price reactions over the
sample. In summary, the empirical analysis presented in this article suggests that stock and
bond markets respond to a variety of Fed communication events, especially FOMC meeting
statements, FOMC press conferences, and remarks by the Fed Chair.
## CONCLUSION
Clear and concise communication of monetary policy helps the Fed achieve its congres
sionally mandated goals of price stability, maximum employment, and stable long-term
**86** **S** **d Q** **t** **2019** **F d** **l R** **B** **k f St L** **i REVIEW**
-----
**Kliesen, Levine, Waller**
interest rates. It does so by helping to reduce uncertainty about the future direction of policy.
This helps to reduce distortions in market pricing, thereby improving the efficient allocation
of resources by firms, households, and governments. This article has examined the various
dimensions of Fed communication with the public and financial markets. This includes docu
menting how the Fed’s communication with the public has evolved over time. Using both
daily and intraday data, our empirical analysis documents how Fed communication affects
key financial market variables. We find that Fed communication is associated with changes
in prices of financial market instruments such as Treasury securities and equity prices. How
ever, this effect varies by type of communication, by type of instrument, and by who is doing
the speaking. Perhaps not surprisingly, we find that the largest financial market reactions tend
to be associated with communication by Fed Chairs rather than by other Fed governors and
Reserve Bank presidents and with FOMC meeting statements rather than FOMC minutes. n
## NOTES
1 The occasion was a hearing of the Committee on Finance and Industry. According to Ahamed (2009), this was a
select committee to investigate the British banking system in the aftermath of the 1929 collapse in stock prices
and the poor performance of the British economy. See Ahamed (2009, pp. 371-72).
2 Ahamed (2009, p. 371).
3 For example, see Cochrane (2017), Cogan and Shultz (2017), and Derby (2017).
4 From a 2003 speech by Governor Yellen, as quoted in Holmes (2013). Holmes argues that central bankers, both in
the United States and elsewhere, have increasingly (even before the Financial Crisis) moved away from traditional
instruments, such as interest rates or exchange rates, toward “communicative experiments” designed to influence
public sentiments and expectations.
5 For an early discussion of this phenomenon applied to the Reserve Bank of New Zealand and the FOMC, see
Guthrie and Wright (2000) and Thornton (2004), respectively.
6 See Blinder et al. (2001).
7 See Bernanke, Reinhart, and Sack (2004). A synthesis of Bernanke’s views was presented in a 2013 speech,
“Communication and Monetary Policy.”
8 See, for example, Neely (2015) or Bauer and Rudebusch (2014).
9 We define the public as anyone who uses expectations about future monetary policy actions as an input into
their decisionmaking process.
10 Current Chairman Jerome Powell expanded on this innovation, announcing that press conferences will be held
after every FOMC meeting beginning in January 2019.
11 For example, the ROPA for the January 15, 1970, meeting was released on April 15, 1970, a three-month lag. The
FOMC ceased publication of the ROPA after the December 22, 1992, meeting. Beginning in 1993, the ROPA was
effectively folded into the FOMC minutes and released with a much shorter lag. For more historical detail, see
https://www.federalreserve.gov/monetarypolicy/fomc_historical.htm.
12 This statement, and subsequent policy statements, can be found on the Board of Governors of the Federal Reserve
[System website: https://www.federalreserve.gov/monetarypolicy/fomc_historical_year.htm.](https://www.federalreserve.gov/monetarypolicy/fomc_historical_year.htm)
13 Wynne (2013) provides a short history of the FOMC’s communication practices.
14 See Gavin and Mandal (2000).
15 The ZLB is the period when the target range for the intended federal funds rate was 0 percent to 0.25 percent.
The ZLB period ended at the December 2015 FOMC meeting.
**F d** **l R** **B** **k f St L** **i REVIEW** **S** **d Q** **t** **2019** **87**
-----
**Kliesen, Levine, Waller**
16 The monetary easing commenced in August 2007, when the Board of Governors voted to reduce the discount
[rate by 50 basis points. See https://www.stlouisfed.org/financial-crisis/full-timeline.](https://www.stlouisfed.org/financial-crisis/full-timeline)
17 Wynne (2013) documented that the Fed used forward-looking language to shape expectations before the Financial
Crisis. For example, in 2003, the FOMC noted that “policy accommodation can be maintained for a considerable
period” in its post-meeting statement. Most economists and policymakers, though, would probably agree that
the use was most pronounced during the ZLB era. The FOMC’s forward guidance policy was influenced importantly
by Woodford (2001) and Eggertsson and Woodford (2003).
18 The midpoint of the central tendency excludes the three highest and three lowest projections for each variable in
each year.
19 Moreover, following the November 3, 2010, meeting, the policy statements crafted under the leadership of Chair
man Bernanke began to emphasize the economy’s current performance and expected outcome relative to the
Fed’s “statutory mandate” of price stability and maximum employment. This was a departure from the Greenspan
era, when the statement rarely—if ever—mentioned the Fed’s statutory mandate. The November 2010 statement
was also noteworthy because it announced the second round of the large-scale asset purchase program (QE2).
20 The average Flesch-Kincaid scores during this period were very close to the reported medians.
21 See also Jansen (2011). Others have found similar findings for other major central bank communications. See
Coenen at al. (2017), Haldane (2017), and Ehrmann and Talmi (2017).
22 Meltzer (2009) documents a 1962 FOMC meeting where communication with the public was discussed. Then
Chairman Martin favored increased communication with the public as a way to counter academic critics of Fed
policy who he believed were mistaken in their analysis. However, Martin opposed regular (quarterly) policy reviews
because there were instances where the FOMC would not wish to explain its decision. See discussion on p. 337 of
Meltzer (2009).
23 The source of this repository is Bloomberg. More detail on this source, and its limitations, is provided in the
empirical analysis section.
24 As noted above, the declining number of Fed governors speaking on the same day reflects to some extent the
dwindling number of years when there was full complement of governors (seven) serving on the FOMC.
25 See Olson and Wessel (2016).
[26 See https://www.federalreserve.gov/monetarypolicy/mpr_default.htm.](https://www.federalreserve.gov/monetarypolicy/mpr_default.htm)
27 See Fawley and Neely (2014).
28 See Bullard (2016), Evans (2017), and Kashkari (2017).
29 Conference calls and unscheduled FOMC meetings were excluded from the analysis.
30 For simplicity, we only focus on announcements directly related to a large-scale asset purchase program. These
include the following: QE1 announcement and expansion, QE2 announcement, Maturity Extension Program
announcement and expansion, and QE3 announcement and expansion.
31 The initial QE1 announcement, which was made on November 25, 2008, did not coincide with an FOMC meeting.
32 For example, the release of nonfarm payroll employment, CPI inflation, and GDP (advance, second, and third esti
mates) all occur before or at the market open.
33 As previously mentioned, our database for public remarks comes from Bloomberg; it begins in 1998. For consis
tency, we start all Fed communication event categories at this date, where applicable. Only public remarks made
during market hours are included in the event study. If Bloomberg did not provide a time for an event, and this
time could not be identified by other sources, the event was removed from the sample. We considered merging
Bloomberg’s repository with other databases, but since there was not a consistent time horizon or speaker overlap,
we did not proceed with this approach. In particular, we examined databases from the Board of Governors of the
Federal Reserve System and the Federal Reserve Bank of St. Louis’s “FOMC Speak.” The Board’s database does not
include public remarks made by Bank presidents, while “FOMC Speak” only begins in 2010. Merging either data
base with Bloomberg’s would result in an upward estimate of governors’ remarks (for the former scenario) or an
upward estimate of remarks over the 2010-17 period (for the latter scenario), which would also affect Figure 3.
**88** **S** **d Q** **t** **2019** **F d** **l R** **B** **k f St L** **i REVIEW**
-----
**Kliesen, Levine, Waller**
Nevertheless, we acknowledge that the Bloomberg database is only a proxy for public remarks when presenting
this analysis.
34 Our dependent variable is very similar to the approach used by Andersson (2010), who used intraday data to
analyze financial market responses to Federal Reserve and European Central Bank monetary policy decisions.
35 Andersson (2010) studied intraday volatility in the bond futures market and in the equity market (S&P 500 index)
around FOMC statement releases from April 1999 to May 2006. He found that intraday volatility rises sharply at
the time of the release of FOMC meeting statements.
36 TickWrite is the source for the intraday data used in this analysis.
37 The one exception is for press conferences, where expanding our event window noticeably increased the market
reaction relative to other events. One possible explanation is that press conferences are often more than an hour
long. However, a closer inspection reveals that the press conferences driving this jump in magnitude are those on
June 22, 2011, and June 19, 2013. The latter was noteworthy because this is when Chairman Bernanke discussed
the so-called taper tantrum that had developed in the markets in response to his Congressional testimony a month
earlier. In that testimony, he raised the possibility of the FOMC beginning to taper asset purchases later that year.
38 It is not our intent to examine whether stock and bond prices may react differently to Fed communication events.
We refer the reader to numerous studies on the effects of these dynamics in the interactions with monetary policy
actions. For example, see Campbell and Ammer (1993), Bernanke and Kuttner (2005), Andersen et al. (2007), and
Connolly, Stivers, and Sun (2005).
39 The non-event day controls in Figures 6A and 6B are constructed to have similar response windows to the events
they are compared with. For example, in Figure 6A, we use a rolling event window of 30 and 75 minutes to calcu
late a benchmark for non-speakers and speakers, respectively. Windows that either include an event or overlap
days are removed before calculating the benchmark mean absolute changes. We follow the same procedure for
Figure 6B.The authors thank Chris Neely for helpful comments in this regard.
## REFERENCES
Ahamed, Liaquat. Lords of Finance: The Bankers Who Broke the World. Penguin Books, 2009 (paperback version).
Andersen, Torben G.; Bollerslev, Tim; Diebold, Francis X. and Vega, Clara. “Real-Time Price Discovery in Global Stock,
Bond and Foreign Exchange Markets.” Journal of International Economics, November 2007, 73(2), pp. 251-77;
[https://doi.org/10.1016/j.jinteco.2007.02.004.](https://doi.org/10.1016/j.jinteco.2007.02.004)
Andersson, Magnus. “Using Intraday Data to Gauge Financial Market Responses to Federal Reserve and ECB
Monetary Policy Decisions.” International Journal of Central Banking, June 2010, 6(2), pp. 117-46.
Bauer, Michael D. and Rudebusch, Glenn D. “The Signaling Channel for Federal Reserve Bond Purchases.”
_International Journal of Central Banking, September 2014, 10(3), pp. 233-89._
Bernanke, Ben S. “Deflation: Marking Sure ‘It’ Doesn’t Happen Here.” Remarks before the National Economists Club,
[Washington, D.C., November 21, 2002; https://www.federalreserve.gov/boarddocs/speeches/2002/20021121/.](https://www.federalreserve.gov/boarddocs/speeches/2002/20021121/)
Bernanke, Ben S. “The Global Saving Glut and the U.S. Current Account Deficit.” Remarks at the Sandridge Lecture,
Virginia Association of Economists, Richmond, Virginia, March 10, 2005;
[https://www.federalreserve.gov/boarddocs/speeches/2005/200503102/.](https://www.federalreserve.gov/boarddocs/speeches/2005/200503102/)
Bernanke, Ben S. and Kuttner, Kenneth N. “What Explains the Stock Market’s Reaction to Federal Reserve Policy?”
_[Journal of Finance, May 2005, 60(3), 1221-257; https://doi.org/10.1111/j.1540-6261.2005.00760.x.](https://doi.org/10.1111/j.1540-6261.2005.00760.x)_
Bernanke, Ben S.; Reinhard, Vincent R. and Sack, Brian P. “Monetary Policy Alternatives at the Zero Bound: An
Empirical Assessment.” Brookings Papers on Economic Activity, January 2004;
[https://www.brookings.edu/wp-content/uploads/2004/06/2004b_bpea_bernanke.pdf.](https://www.brookings.edu/wp-content/uploads/2004/06/2004b_bpea_bernanke.pdf)
Blinder, Alan; Ehrmann, Michael; Fratzscher, Marcel; De Haan, Jakob and Jansen, David-Jan. “Central Bank
Communication and Monetary Policy: A Survey of Theory and Evidence.” ECB Working Paper No. 898, European
[Central Bank, May 2008; https://www.ecb.europa.eu/pub/pdf/scpwps/ecbwp898.pdf.](https://www.ecb.europa.eu/pub/pdf/scpwps/ecbwp898.pdf)
**F d** **l R** **B** **k f St L** **i REVIEW** **S** **d Q** **t** **2019** **89**
-----
**Kliesen, Levine, Waller**
Blinder, Alan; Goodhart, Charles; Hildebrand, Philipp; Lipton, David and Wyplosz, Charles. “How Do Central Banks
Talk?” Geneva Report on the World Economy, No. 3, International Center for Monetary and Banking Studies, 2001.
Bullard, James. “A New Characterization of the U.S. Macroeconomic and Monetary Policy Outlook.” Remarks
delivered at the Society of Business Economists Annual Dinner, London, United Kingdom, June 30, 2016;
[https://www.stlouisfed.org/from-the-president/speeches-and-presentations/2016/a-new-characterization.](https://www.stlouisfed.org/from-the-president/speeches-and-presentations/2016/a-new-characterization)
Campbell, Jeffrey R.; Evans, Charles L.; Fisher, Jonas D.M. and Justiano, Alejandro. “Macroeconomic Effects of
Federal Reserve Forward Guidance.” Brookings Papers on Economic Activity, Spring 2012, pp. 1-63;
[https://doi.org/10.1353/eca.2012.0004.](https://doi.org/10.1353/eca.2012.0004)
Campbell, John Y. and Ammer, John. “What Moves the Stock and Bond Markets? A Variance Decomposition for
Long-Term Asset Returns.” Journal of Finance, March 1993, 48(1), pp. 3-37;
[https://doi.org/10.1111/j.1540-6261.1993.tb04700.x.](https://doi.org/10.1111/j.1540-6261.1993.tb04700.x)
Cochrane, John H. “Taylor for Fed.” The Grumpy Economist Blog, October 20, 2017;
[https://johnhcochrane.blogspot.com/2017/10/taylor-for-fed.html.](https://johnhcochrane.blogspot.com/2017/10/taylor-for-fed.html)
Coenen, Günter; Ehrmann, Michael; Gaballo, Gaetano; Hoffmann, Peter; Nakov, Anton; Nardelli, Stefano; Persson,
Eric and Strasser, Georg. “Communication of Monetary Policy in Unconventional Times.” ECB Working Paper No.
[2080, European Central Bank, June 2017; https://www.ecb.europa.eu/pub/pdf/scpwps/ecb.wp2080.en.pdf.](https://www.ecb.europa.eu/pub/pdf/scpwps/ecb.wp2080.en.pdf)
Cogan, John F. and Shultz, George P. “The Fed Chief America Needs.” Commentary, Wall Street Journal, October 18,
2017.
Connolly, Robert; Stivers, Chris and Sun, Licheng. “Stock Market Uncertainty and the Stock-Bond Return Relation.”
_Journal of Financial and Quantitative Analysis, March 2005, 40(1), pp. 161-94;_
[https://doi.org/10.1017/S0022109000001782.](https://doi.org/10.1017/S0022109000001782)
Derby, Michael S. “Fed’s Fischer Laments ‘The Cacophony’ of Too Much Fed Speak.” WSJ Pro, April 25, 2017.
Eggertsson, G. and Woodford, Michael. “The Zero Bound on Interest Rates and Optimal Monetary Policy.” Brookings
_Papers on Economic Activity, 2003, 34(1), pp. 139-235;_
[https://www.brookings.edu/wp-content/uploads/2003/01/2003a_bpea_eggertsson.pdf.](https://www.brookings.edu/wp-content/uploads/2003/01/2003a_bpea_eggertsson.pdf)
Ehrmann, Michael and Talmi, Jonathan. “Starting from a Blank Page? Semantic Similarity in Central Bank
Communication and Market Volatility.” ECB Working Paper No. 2023, European Central Bank, February 2017;
[https://www.ecb.europa.eu/pub/pdf/scpwps/ecbwp2023.en.pdf?561f7f26fea2285effb36dae431188d1.](https://www.ecb.europa.eu/pub/pdf/scpwps/ecbwp2023.en.pdf?561f7f26fea2285effb36dae431188d1)
Evans, Charles L. “Rationale for My Dissent at the December 2017 Meeting.” Federal Reserve Bank of Chicago,
[December 15, 2017. https://www.chicagofed.org/publications/speeches/2017/12-15-2017-evans-rationale-for-](https://www.chicagofed.org/publications/speeches/2017/12-15-2017-evans-rationale-for-dissent-december-fomc)
[dissent-december-fomc.](https://www.chicagofed.org/publications/speeches/2017/12-15-2017-evans-rationale-for-dissent-december-fomc)
Fawley, Brett W. and Neely, Christopher J. “The Evolution of Federal Reserve Policy and the Impact of Monetary
Policy Surprises on Asset Prices.” Federal Reserve Bank of St. Louis Review, First Quarter 2014, 96(1), pp. 73-109;
[https://files.stlouisfed.org/files/htdocs/publications/review/2014/q1/fawley.pdf.](https://files.stlouisfed.org/files/htdocs/publications/review/2014/q1/fawley.pdf)
Gavin, William T. and Mandal, Rachel J. “Inside the Briefcase: The Art of Predicting the Federal Reserve.” Federal
[Reserve Bank of St. Louis Regional Economist, July 2000, pp. 4-9; https://www.stlouisfed.org/publications/region](https://www.stlouisfed.org/publications/regional-economist/july-2000/inside-the-briefcase-the-art-of-predicting-the-federal-reserve)
[al-economist/july-2000/inside-the-briefcase-the-art-of-predicting-the-federal-reserve.](https://www.stlouisfed.org/publications/regional-economist/july-2000/inside-the-briefcase-the-art-of-predicting-the-federal-reserve)
Greenspan, Alan. “Remarks before the Economic Club of Chicago.” Chicago, Illinois, October 19, 1995;
[https://fraser.stlouisfed.org/scribd/?item_id=8549&filepath=/files/docs/historical/greenspan/](https://fraser.stlouisfed.org/scribd/?item_id=8549&filepath=/files/docs/historical/greenspan/Greenspan_19951019.pdf)
[Greenspan_19951019.pdf.](https://fraser.stlouisfed.org/scribd/?item_id=8549&filepath=/files/docs/historical/greenspan/Greenspan_19951019.pdf)
Greenspan, Alan. “The Challenge of Central Banking in a Democratic Society.” Remarks at the Annual Dinner and
Francis Boyer Lecture of the American Enterprise Institute for Public Policy Research, Washington, D.C.,
[December 5, 1996; https://fraser.stlouisfed.org/scribd/?item_id=8581&filepath=/files/docs/historical/green](https://fraser.stlouisfed.org/scribd/?item_id=8581&filepath=/files/docs/historical/greenspan/Greenspan_19961205.pdf)
[span/Greenspan_19961205.pdf.](https://fraser.stlouisfed.org/scribd/?item_id=8581&filepath=/files/docs/historical/greenspan/Greenspan_19961205.pdf)
Guthrie, Graeme and Wright, Julian. “Open Mouth Operations.” Journal of Monetary Economics, October 2000, 46(2),
[pp. 489-516; https://doi.org/10.1016/S0304-3932(00)00035-0.](https://doi.org/10.1016/S0304-3932(00)00035-0)
Haldane, Andrew G. “A Little More Conversation, A Little Less Action.” Remarks at the Macroeconomics and
Monetary Policy Conference, Federal Reserve Bank of San Francisco, March 2017;
[https://www.bankofengland.co.uk/speech/2017/a-little-more-conversation-a-little-less-action.](https://www.bankofengland.co.uk/speech/2017/a-little-more-conversation-a-little-less-action)
**90** **S** **d Q** **t** **2019** **F d** **l R** **B** **k f St L** **i REVIEW**
-----
**Kliesen, Levine, Waller**
Holmes, Douglas R. Economy of Words: Communicative Imperatives in Central Banks. University of Chicago Press,
[2013; https://doi.org/10.7208/chicago/9780226087764.001.0001.](https://doi.org/10.7208/chicago/9780226087764.001.0001)
Jansen, David-Jan. “Does the Clarity of Central Bank Communication Affect Volatility in Financial Markets?
Evidence from Humphrey-Hawkins Testimonies.” Contemporary Economic Policy, 2011, 29, pp. 494-509;
[https://doi.org/10.1111/j.1465-7287.2010.00238.x.](https://doi.org/10.1111/j.1465-7287.2010.00238.x)
Kashkari, Neel. “Why I Dissented a Third Time.” Federal Reserve Bank of Minneapolis, December 18, 2017;
https://www.minneapolisfed.org/news-and-events/messages/why-i-dissented-a-third-time.
Kocherlakota, Narayana. “Maybe Central Banks Are Too Independent.” Bloomberg View, August 7, 2017;
[https://www.bloomberg.com/view/articles/2017-08-07/maybe-central-banks-are-too-independent.](https://www.bloomberg.com/view/articles/2017-08-07/maybe-central-banks-are-too-independent)
Lustenberger, Thomas and Rossi, Enzo. “Does Central Bank Transparency and Communication Affect Financial and
Macroeconomic Forecasts?” Swiss National Bank Working Paper, December 2017;
[https://www.snb.ch/n/mmr/reference/working_paper_2017_12/source/working_paper_2017_12.n.pdf.](https://www.snb.ch/n/mmr/reference/working_paper_2017_12/source/working_paper_2017_12.n.pdf)
Meltzer, Allan H. A History of the Federal Reserve. Volume 2, Book 1: 1951-1969. University of Chicago Press, 2009.
Neely, Christopher J. “Unconventional Monetary Policy Had Large International Effects.” Journal of Banking and
_[Finance, March 2015, 52, pp. 101-11; https://doi.org/10.1016/j.jbankfin.2014.11.019.](https://doi.org/10.1016/j.jbankfin.2014.11.019)_
Olson, Peter and Wessel, David. “Federal Reserve Communications: Survey Results.” Hutchins Center on Fiscal and
Monetary Policy at Brookings, November 2016;
[https://www.brookings.edu/wp-content/uploads/2016/11/fed-communications-survey-results.pdf.](https://www.brookings.edu/wp-content/uploads/2016/11/fed-communications-survey-results.pdf)
Shin, Hyun Song. “Can Central Banks Talk Too Much?” Speech at the European Central Bank conference
“Communications Challenges for Policy Effectiveness,” Frankfurt, November 14, 2017.
Thornton, Daniel L. “The Fed and Short-Term Rates: Is It Open Market Operations, Open Mouth Operations or
Interest Rate Smoothing?” Journal of Banking and Finance, 2004, 28(4), pp. 475-98;
[https://doi.org/10.1016/S0378-4266(02)00409-0.](https://doi.org/10.1016/S0378-4266(02)00409-0)
Woodford, Michael. “Monetary Policy in the Information Economy.” Presented at the Federal Reserve Bank of
Kansas City symposium “Economic Policy for the Information Economy,” August 30–September 1, 2001;
[https://www.kansascityfed.org/publicat/sympos/2001/papers/S02wood.pdf.](https://www.kansascityfed.org/publicat/sympos/2001/papers/S02wood.pdf)
Wynne, Mark A. “A Short History of FOMC Communication.” Federal Reserve Bank of Dallas Economic Letter,
[September 2013, 8(8); https://www.dallasfed.org/research/eclett/2013/el1308.cfm#n8.](https://www.dallasfed.org/research/eclett/2013/el1308.cfm#n8)
**F d** **l R** **B** **k f St L** **i REVIEW** **S** **d Q** **t** **2019** **91**
-----
**92** **S** **d Q** **t** **2019** **F d** **l R** **B** **k f St L** **i REVIEW**
-----
|
{
"disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.20955/r.101.69-91?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.20955/r.101.69-91, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "BRONZE",
"url": "https://files.stlouisfed.org/research/publications/review/2019/04/15/gauging-market-responses-to-monetary-policy-communication.pdf"
}
| 2,019
|
[] | true
| null |
[] | 19,709
|
en
|
[
{
"category": "Law",
"source": "s2-fos-model"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/022f64f1cbb0b6dd859736f162cff1130501ec1b
|
[] | 0.888863
|
Dispute Resolution Mechanism for Smart Contracts
|
022f64f1cbb0b6dd859736f162cff1130501ec1b
|
Masaryk University Journal of Law and Technology
|
[
{
"authorId": "94054909",
"name": "M. Kasatkina"
}
] |
{
"alternate_issns": [
"1802-5951"
],
"alternate_names": [
"Masaryk Univ J Law Technol",
"Masaryk University journal of law and technology",
"Masaryk Univ j law technol"
],
"alternate_urls": null,
"id": "9e7f3023-e1fa-4121-9043-ef2a46353fbf",
"issn": "1802-5943",
"name": "Masaryk University Journal of Law and Technology",
"type": "journal",
"url": null
}
|
Disputes regarding smart contracts are inevitable, and parties will need means for dealing with smart contract issues. This article highlights the need for dispute resolution mechanisms for smart contracts. The author provides analysis of the possible mechanisms to solve disputes arising from smart contracts, namely dispute resolution by traditional arbitration institutions and blockchain arbitration. Article acknowledges the benefits and challenges of both mechanisms. In the light of this, the author concludes about instituting a hybrid approach aimed at resolving disputes that will not stymie efficiencies of smart contracts.
|
_2022]_ _M. Kasatkina: Dispute Resolution Mechanism for Smart Contracts_ _143_
_DOI 10.5817/MUJLT2022-2-2_
# DISPUTE RESOLUTION
MECHANISM FOR SMART CONTRACTS
_by_
## MARINA KASATKINA[*]
_Disputes regarding smart contracts are inevitable, and parties will need means for_
_dealing with smart contract issues. This article highlights the need for dispute_
_resolution mechanisms for smart contracts. The author provides analysis_
_of the possible mechanisms to solve disputes arising from smart contracts, namely_
_dispute resolution by traditional arbitration institutions and blockchain_
_arbitration. Article acknowledges the benefits and challenges of both mechanisms._
_In the light of this, the author concludes about instituting a hybrid approach aimed_
_at resolving disputes that will not stymie efficiencies of smart contracts._
## KEY WORDS
_Smart Contracts, Blockchain Technology, Digital Disputes, Dispute Resolution_
_Mechanism, Off-chain, On-chain._
## 1. INTRODUCTION
With the rapid development of new technologies occurring during
the fourth industrial revolution, new types of disputes with significant
specifics are gradually beginning to form. A special category among them
belongs to disputes arising from smart contracts based on blockchain
technology. Smart contracts are not really “contracts” in the true sense
of the word, understood by most as negotiated terms in an arms-length
transaction (or “meeting of the minds”).[1] Enforcement is automatic,
and the code is immutable. Therefore, smart contracts on the blockchain
- m.kasatkina@maastrichtuniversity.nl, Ph.D. candidate, Maastricht University, Netherlands
1 Schmitz, A. J. and Rule, C. (2019) Online Dispute Resolution for Smart Contracts. _Journal_
_of Dispute Resolution. University of Missouri School of Law Legal Studies Research Paper,_
2019 (11). Available from: https://ssrn.com/abstract=3410450 [Accessed 12 April 2022].
-----
_144_ _Masaryk University Journal of Law and Technology_ _[Vol. 16:2_
present a different set of challenges due to the inflexibility of the code-based
executions.
It has to be noted that there is a close interaction between the real world
and the software transaction world. Smart contracts inherently interfere
with real-world people or institutions, which would result in legal issues
due to the nature of our societies.[2] For the reason that virtual experiences
lead to specific actions in the real world, disputes are inevitable. Possible
scenarios in which disputes may arise include changing of circumstances,
creating undesirable results for one party, absence of legal capacity to enter
into the smart contract. Smart contracts may not be accurately coded
to encompass the parties’ original intentions. Moreover, coders may be sued
for liability as a result of inaccurate smart contracts, or hackers may be
prosecuted for interfering with or manipulating smart contracts.[3] In this
respect, the potential need for dispute resolution mechanism is inevitable.
But nowadays, there exist no well-defined system of rules applicable
to smart contracts. All these aspects show that there is room for identifying
dispute resolution mechanisms for smart contracts.
Generally speaking, there are two possible ways to resolve such
disputes. According to the first approach, they are subject to review
by traditional courts. The second approach assumes that arbitration
institutions lend to resolve disputes arising out of smart contracts. They,
in turn, are divided into two groups:
a) “off-chain” arbitration, meaning dispute resolution by traditional
arbitration institutions guided by the usual rules;
b) “on-chain” arbitration that assumes to create innovative applications
based on blockchain technology and designed to resolve disputes arising
in a digital decentralized environment (blockchain arbitration).[4]
My focus in this article is on the possible mechanisms to solve disputes
arising from smart contracts. I have two aims: first, to outline a framework
for dispute resolution by traditional arbitration institutions and blockchain
arbitration, and second, based on advantages and disadvantages of both
2 Clément, M. (2019) Smart Contracts and the Courts. In: DiMatteo, L., Cannarsa, M.
and Poncibò, C. (eds.) The Cambridge Handbook of Smart Contracts, Blockchain Technology
_and Digital Platforms. Cambridge University Press, pp. 271–287._
3 Zaslowsky, D. (2018) What to Expect When Litigating Smart Contract Disputes. [online]
Available from: https://www.law360.com/articles/1028009/what-to-expect-when-litigatingsmart-contract-disputes [Accessed 02 May 2022].
4 International Chamber of Commerce (2018). ICC Dispute Resolution Bulletin. Issue 1.
Available from: https://www.hoganlovells.com/~/media/hogan-lovells/pdf/2018/
2018_12_13_icc_robots_arbitrator.pdf [Accessed 02 May 2022].
-----
_2022]_ _M. Kasatkina: Dispute Resolution Mechanism for Smart Contracts_ _145_
mechanisms I introduce a new hybrid approach to blockchain dispute
resolution, that combines both on and off-blockchain components.
## 2. ANALYSIS OF THE POSSIBLE DISPUTE RESOLUTION MECHANISMS
The first question while considering dispute resolution mechanisms should
be asked whether traditional courts could adjudicate disputes arising from
smart contracts. In this respect, the following should be mentioned. Firstly,
a smart contract is the code, which is understandable to programmers, not
lawyers and judges. Courts may be substantially challenged in interpreting
smart contracts, written in a coded language, that is not understandable
to a human observer. Furthermore, a court could not intervene to prevent
or reverse an automatic contract, since the execution of smart contracts does
not allow for modifications.[5] As James Grimmelmann notes,
_“…as long as the code does what it is supposed to and blockchain nodes_
_achieve consensus, the intent and actions of one’s counterpart do not matter;_
_once triggered, the contract moves forward as defined at the time of its_
_writing, regardless of either party’s change in circumstances,_
_misunderstandings, or otherwise.”[6]_
In this regard, it is important to distinguish between two main models
of smart contracts: external and internal.[7] External smart contracts are those
that are governed by traditional, natural language contracts with the smart,
code-driven part of the contract merely automating the performance
of terms as appropriate (e.g. payment, shipment, etc.). If there is any
disagreement between the parties, the traditional, non-code version
of the contract prevails. An external smart contract must be clear about
which version of the contract prevails in order to successfully put
the natural-language terms first and foremost. However, when such clarity
is lacking in multi-language contracts, the UNIDROIT Principles stipulate
5 Rodrigues, U. (2018) Law and the Blockchain. Iowa Law Review, 104. Available from:
https://ilr.law.uiowa.edu/print/volume-104-issue-2/law-and-the-blockchain/ [Accessed
02 May 2022].
6 Grimmelmann, J. (2019) All Smart Contracts are Ambiguous. Journal of Law & Innovation,
2 (1). Available from: https://www.law.upenn.edu/live/files/9782-grimmelmann-all-smartcontracts-are-ambiguous [Accessed 02 May 2022].
7 Chamber of Digital Commerce. (2018) Smart Contracts: Is the Law Ready? Available from:
https://www.theblockchaintest.com/uploads/resources/CDC%20-%20Smart%20Contract-Is
%20the%20Law%20Ready%20-%202018%20-%20Sep.pdf [Accessed 02 May 2022].
-----
_146_ _Masaryk University Journal of Law and Technology_ _[Vol. 16:2_
that preference should be given to the contract that was originally
drawn up. Presumably, the same can apply to smart contracts; if the code
was written first and the natural-language contract second, the code
prevails. Inversely, one may say that code is not a “human” language of any
kind and therefore should be interpreted as an appendix for the natural
language contract, but not the main, binding part of any agreement. This
approach may work in certain contexts, however, given that the code
creates an outcome automatically, its interpretive value seems more
relevant to the main body of most external smart contract.[8]
In the internal smart contracts, the code is supreme and any natural
-language portion of the agreement is secondary. Therefore, while
the natural-language portion of the contract may help courts understand
the parties’ intent, they will still have to interpret code to understand what
consensus was reached. While this has been raised as a problem for courts
wishing to exert power over smart contracts, the use of expert witnesses
who can read and inform the court what the code “says”, can quickly and
easily remedy this issue (e.g. bringing a programmer to the stand to testify
what the outcome of the code, as written, would be).[9] Thus, regardless
of the specific type of smart contract, the inflexibility of code-based
executions presents potential challenges.
Secondly, the anonymous nature of smart contracts and the fluidity
of online identities make it difficult to determine the identities of the parties.
The aforementioned anonymity gained by the use of public-key encrypted
identities and VPNs. Nodes that contain the blockchain and all of its
information are located all over the world. Transactions in the blockchain
are fully networked and present only in cyberspace. The nodes hold
imperfect partial copies of the blockchain; no particular node holds
the entire blockchain.[10] And the decentralized nature of smart contracts
prevents courts from establishing jurisdiction and determining the choice
of law based on traditional rules.
For all of these reasons, it can be concluded that smart contract disputes
should not be resolved by any national court. This leads to the demand for
8 Sillanpaa, T. (2020) Freedom to (Smart) Contract: The Myth of Code and Blockchain
Governance Law. _IALS Student Law Review,_ 7 (2). Available from:
https://journals.sas.ac.uk/lawreview/issue/view/582 [Accessed 02 May 2022].
9 Ibid.
10 Kaal, W. A. and Calcaterra, C. (2018) Crypto Transaction Dispute Resolution. Business
_Lawyer. Available from: http://dx.doi.org/10.2139/ssrn.2992962 [Accessed 05 May 2022]._
-----
_2022]_ _M. Kasatkina: Dispute Resolution Mechanism for Smart Contracts_ _147_
resolving smart contract disputes with cross-jurisdictional, extra-legal, and
efficient remedies.
Therefore, international arbitration presents a well-suited alternative for
smart contract disputes as they have many common features, such as
functioning in a decentralized manner, flexibility, confidentiality
of proceedings. Nowadays, there exist two main approaches for dealing
with smart contract issues, namely “on-chain” and “off-chain” arbitration.[11]
### 2.1 “OFF-CHAIN” ARBITRATION (DISPUTE RESOLUTION BY TRADITIONAL ARBITRATION INSTITUTIONS)
According to this approach, smart contracts can operate within the existing
contract law framework, and disputes arising from them are subject
to the arbitration institutions.[12] In this regard, a special arbitration center
dealing with the resolution of digital disputes is being created
or a specialized board in the existing arbitration institutions is being
formed. Generally speaking, “off-chain” dispute resolution system could be
characterized as a combination of traditional forms of dispute resolution
process, lacking a mechanism for the automatic enforcement of the award.
For instance, on the 8th of November 2018 was opened the Court
_of Arbitration of the Polish Blockchain and New Technology Chamber of Commerce_
(hereinafter “Court of Arbitration”) which purpose is to resolve disputes
related to digital technologies.[13] It is Europe’s first and the world’s second
(after Japan) arbitral tribunal specializing in blockchain. Court
of Arbitration applies the provisions of the Rules of the Court of Arbitration
of the Polish Blockchain and New Technology Chamber of Commerce
(hereinafter “Rules”).[14] According to paragraph 3 of the Rules, the Court
of Arbitration has jurisdiction over a dispute if the parties conclude
a written agreement (arbitration agreement) in the following forms:
11 Szczudlik, K. (2019) _“On-chain” and “off-chain” arbitration: Using smart contracts to amicably_
_resolve disputes. [online] Available from: https://newtech.law/en/on-chain-and-off-chain-_
arbitration-using-smart-contracts-to-amicably-resolve disputes [Accessed 02 May 2022].
12 De Filippi, P. and Wright A. (2018) Blockchain and the Law: The Rule of Code. Cambridge,
Cambridge, MA: Harvard University Press, 300; Holden R. and Malani A. (2018) Can
Blockchain Solve the Holdup Problem in Contracts? University of Chicago Coase-Sandor
_Institute for Law & Economics. Working Paper, 846._
13 _The Court of Arbitration of the Polish Blockchain and New Technology Chamber of Commerce._
[online] Available from: https:// blockchaincourt.org/ [Accessed 02 May 2022].
14 The Court of Arbitration of the Polish Blockchain and New Technology Chamber
of Commerce (2019). The Rules of the Court of Arbitration of the Polish Blockchain and New
_Technology Chamber of Commerce._ Available from: https://blockchaincourt.org/
wp-content/uploads/2019/07/The-Rules-of-the-Court-of-Arbitration-ENG.pdf [Accessed
04 May 2022].
-----
_148_ _Masaryk University Journal of Law and Technology_ _[Vol. 16:2_
a) a clause included in letters exchanged between the parties
or declarations made by the parties by means of remote communication that
enable the content of such declarations to be recorded
b) a reference made in a written agreement to a document containing
a provision on submitting disputes to resolution by the Court
of Arbitration. The dispute resolution process is carried out according
to the standard arbitration procedure with certain exceptions. Firstly,
the number of arbitrators for resolution of the dispute could be 5 or 7 in
contrast to “traditional” arbitration (paragraph 19 of the Rules), where
the number of arbitrators is limited (1 or 3). Secondly, an award made by
the Court of Arbitration shall be pronounced at the same hearing at which
the trial is closed. When pronouncing the award, the presiding arbitrator
shall state orally the main reasons upon which such award is based
(paragraph 45 of the Rules). Whereas the traditional arbitration ends
without the announcement of the decision, which is sent to the parties later.
This approach also includes the creation of specialized boards
in the existing arbitration institutions. For example, in 2018, the Arbitration
_center of the Russian Union of Industrialists and Entrepreneurs (RSPP)_
announced the formation of a new Panel on disputes in the digital economy.
The panel was created to resolve disputes arising from transactions
involving automatic execution, including using information systems based
on a distributed registry (blockchain); disputes arising from the issuance,
accounting and circulation of digital rights and disputes over transactions
made using and (or) in relation to digital financial assets.[15]
Due to the absence of special rules, the proceedings on such disputes are
conducted according to the Rules of the arbitration center
at the RSPP 2018.[16]
The above-mentioned approach to the disputes arising from smart
contracts is considered the mainstream view. Although in the legal
literature it is often criticized.[17] Instead, it is proposed to create special
methods of dispute resolution based on technology- blockchain arbitration.
15 _Arbitraznyu zentr pri RSPP._ [online] Available from: https://arbitration
rspp.ru/about/structure/boards/digital-disputes/ [Accessed 04 May 2022].
16 Ibid.
17 Schmitz, A. and Rule C. (2019) Online Dispute Resolution for Smart Contracts.
_Journal of Dispute Resolution, 2, pp. 103–125._
-----
_2022]_ _M. Kasatkina: Dispute Resolution Mechanism for Smart Contracts_ _149_
### 2.2 “ON-CHAIN” ARBITRATION (BLOCKCHAIN ARBITRATION)
This group includes projects that provide for the creation of new
mechanisms specifically designed to resolve disputes arising from smart
contracts. “On-chain” arbitration contains solutions in which the equivalent
of a traditional arbitration decision is automatically executed by a smart
contract without the involvement of any third parties. For instance, this
could be realized with the help of certain assets, which, upon the occurrence
of a defined condition, are transferred from one party to the other.[18]
This approach contemplates smart contracts as distinct legal tools, rather
than digital alternatives to traditional legal contracts. From this perspective,
blockchain technologies and smart contracts may create new legal systems,
or a new Lex Cryptographia.[19] Several characteristics of blockchain-based
technologies and smart contracts, such as its anonymity, automatic
execution, and tamper-resistance, mean that
_“existing legal infrastructure cannot address legal challenges presented_
_by crypto transaction disputes”.[20]_
Instead, these disputes require a “distributed jurisdiction” created
through a process of institutional innovations.
Currently, there exist more than 20 projects that use blockchain
to automate dispute resolution. All these projects could be divided into two
groups:
a) Special on-line arbitration (CodeLegit, Cryptonomica, Juris,
Mattereum, SAMBA);
b) Crowdsourced dispute resolution (Aragon, BitCad, CrowdJury,
Confideal, Jur, Kleros, Oath).
In this article, I examined the most noteworthy projects, which have
already been tested by end users.
18 Szczudlik, K. (2019) _“On-chain” and “off-chain” arbitration: Using smart contracts to amicably_
_resolve disputes. [online] Available from: https://newtech.law/en/on-chain-and-off-chain-_
arbitration-using-smart-contracts-to-amicably-resolve disputes [Accessed 02 May 2022].
19 De Filippi, P. and Wright A. (2018) Blockchain and the Law: The Rule of Code. Cambridge,
Cambridge, MA: Harvard University Press, 300.
20 Kaal, W. A. and Calcaterra, C. (2018) Crypto Transaction Dispute Resolution. Business
_Lawyer. Available from: http://dx.doi.org/10.2139/ssrn.2992962 [Accessed 05 May 2022]._
-----
_150_ _Masaryk University Journal of Law and Technology_ _[Vol. 16:2_
### 2.2.1 SPECIAL ON-LINE ARBITRATION
This group includes platforms that enable the creation of a special
arbitration combining the advantages of international commercial
arbitration and blockchain technology. They presume the automation
of certain elements of the proceedings. However, the mechanism of their
action is in many ways similar to international arbitration, as the rules
of many such projects are based on the UNCITRAL Arbitration Rules. In
this case, the decision made by the arbitrators is executed in the traditional
way or is automatically executed with a smart contract.
For instance, a Juris project that presents a blockchain-based development
system, operating on the basis of the Juris Protocol Mediation and
Arbitration.[21] A prerequisite for considering a dispute is the existence
of an arbitration agreement, integrated into a smart contract via a coded
clause. In case of a dispute between the parties, a user initiates a protocol
by filing a complaint (Formal Complaint). The system suspends further
execution of the smart contract generation and notifies the other party about
the dispute. After that, the following three procedures are possible:
1) Self Mediation – through which the parties get access to a number
of tools, specially designed for self-regulation dispute resolution with
the help of Self-Enforced Library Functions (or Self layer). These tools
enable the execution of basic operations that alter the outcome of a smart
contract implementation (such as contract cancellation and asset transfer).
In the case of impossibility to resolve the dispute, parties could escalate
to the second stage.
2) SNAP (Simple Neutral Arbitrator Poll) means the consideration
of the dispute by independent arbitrators. Results of the voting are reported
to the parties. Based on this information, the parties still may try to resolve
the dispute by using Self layer or applying to the third tool.
3) PANEL (Juris Peremptory Agreement for Neutral Expert Litigation) is
the analogue of traditional arbitration proceedings based
on the UNCITRAL Arbitration Rules. The dispute is reviewed by three
arbitrators selected on the basis of their reputation and compliance with
the requirements specified by the parties while entering into the contract.
21 Kerpelman, A. J. (2018) Introducing the Juris Protocol: Human-Powered Dispute Resolution
_for Blockchain_ _Smart_ _Contracts._ [online] Available from:
https://medium.com/jurisproject/introducing-the-juris-protocol-human-powered-disputeresolution-for-blockchain-smart-contracts-bc574b50d8e1 [Accessed 05 May 2022].
-----
_2022]_ _M. Kasatkina: Dispute Resolution Mechanism for Smart Contracts_ _151_
After hearing the parties and considering evidence, the arbitrators within 30
days make a decision that is binding and subject to automatic execution
by smart contract.
Another project based on the blockchain technology is Mattereum, which
presents the layer of the legal, technological, and commercial infrastructure
that governs on-chain rights control and transfer for tangible, intangible,
and digital assets. Mattereum supports a decentralized commercial law
system, the Smart Property Register, that executes through automated
smart contracts that ensure property rights, as well as dispute resolution
and enforcement. This register facilitates “on-chain property transfer”
through a smart contract that in effect becomes a “legal contract” without
the need for legislative support.[22] A distinctive feature of this project is
the “Ricardian Contracts” on which the contract protocol is based.[23]
Ricardian Contracts are cryptographically verified documents signed with
a digital signature and available for reading both in electronic and text
form. The project involves the creation of a decentralized arbitration court,
meeting the requirements of the New York Convention on Recognition and
Enforcement of Foreign Arbitral Awards of 1958 (hereinafter referred
to as the New York Convention). Therefore, awards of such decentralized
commercial arbitration court will be enforced by national courts in nearly
all of the countries in the world.[24]
A separate point must be made about OpenBazaar Dispute Resolution
_(notary). It is a distributed program that provides an on-line trading_
platform for any type of merchandise using cryptocurrencies.[25] It is
a distributed network where the parties and transactions are anonymous.[26]
A core element of the OpenBazaar dispute resolution mechanism concerns
the possibility of appealing to a notary who becomes an arbitrator and
determines the dispute based on the evidence presented. Notaries
in the OpenBazaar system are randomly chosen to provide anonymity for
keeping the system secure. An important feature of OpenBazaar’s approach
22 Allen, D., Lane, A. M. and Poblet, M. (2019) The Governance of Blockchain Dispute
Resolution. Harvard Negotiation Law Review, 25, pp. 75–101.
23 Zagaynova, M. (2018) _Obzor ICO proekta Mattereum._ [online] Available from:
https://ffc.media/ru/overviews/ico-mattereum-project-review/ [Accessed 21 June 2022].
24 Allen, D., Lane, A. M. and Poblet, M. (2019) The Governance of Blockchain Dispute
Resolution. Harvard Negotiation Law Review, 25, pp. 75–101.
25 Sanchez Dr W. _Dispute Resolution in OpenBazaar._ [online] Available from:
http://docs.openbazaar.org/03.-OpenBazaar-Protocol/ [Accessed 21 June 2022].
26 Kaal, W. A. and Calcaterra, C. (2018) Crypto Transaction Dispute Resolution. Business
_Lawyer. Available from: http://dx.doi.org/10.2139/ssrn.2992962 [Accessed 05 May 2022]._
-----
_152_ _Masaryk University Journal of Law and Technology_ _[Vol. 16:2_
is connected with the ability of the parties to choose the notary pools
as an expert in certain fields of law. Besides, OpenBazaar has an appeal
system that includes randomly selecting new notaries from the pool chosen
by the parties earlier.
### 2.2.2 CROWDSOURCED DISPUTE RESOLUTION
This group includes projects that provide for the establishment
of fundamentally new, unique platforms based on blockchain technology
and specifically designed to resolve disputes arising from smart contracts.
Their essence is an attempt to create a quasi-judicial system, where
the judges (members of the jury) are registered on the relevant platform
users who are elected through the method of generating random numbers,
remaining anonymous to the parties. Each of the judges votes separately;
after the voting is completed, the system counts the votes and determines
the outcome of the dispute. Then the decision is automatically executed
using a smart contract. Another important characteristic of such projects is
the use of codes of non-state regulation in the dispute resolution process.[27]
It has to be noted that crowdsourced dispute resolution is not new. For
example, more than twenty years ago iCourthouse pioneered the notion
of online crowdsourcing in civil cases and ten years ago eBay India’s
Community Court leveraged the best judgement of other eBay users
to decide whether a contested eBay review should be deleted. The following
examples of crowdsourced dispute resolution on the blockchain go even
further with this model, however, by tokenizing the process. In other
words, jurors vote with funds (generally cryptocurrency), which they lose if
they are on the losing side. In contrast, jurors on the winning side generally
gain some reward. This creates a market for accurate crowdsourced
resolution outcomes.[28]
One example is Oath, a project based on the Ethereum platform.
The model of OATH’s dispute resolution is related to the idea of a jury trial.
When entering into a smart contract, the parties can use the provided
dispute resolution protocol (Smart Arbitration Plan). In the case
of a dispute, the protocol is converted into a Smart Arbitration Case. After
27 Zasemkova O. (2020) Methods of Resolving Disputes Arising from Smart Contracts. Lex
_Russica, 73 (4), p. 20._
28 Rule, C. and Nagarajan, C. (2011) Crowdsourcing Dispute Resolution Over Mobile Devices.
In: Poblet, M. (ed.) Mobile Technologies for Conflict Management. Law, Governance and
_Technology Series, vol 2. Dordrecht: Springer, pp. 93–100._
-----
_2022]_ _M. Kasatkina: Dispute Resolution Mechanism for Smart Contracts_ _153_
that, the parties set the parameters for resolving dispute: the number
of jurors (any odd number in the range from 11 to 101); the percentage
of votes required to make a decision (from 51 to 100 %). Juries are selected
randomly from the users of the blockchain platform. The decision is made
solely on the basis of common sense (Common sense), based on the study
of the terms of the contract, witness statements and other evidence.
The decision can be appealed within 5 days from the date of its issuance
by repeating the procedure but with other jurors.[29]
Like Oath, Kleros promises inexpensive and transparent, online dispute
resolution using crowdsourcing theory. The mechanism is similar to Oath,
advocating for an opt-in court platform that uses “crowdsourced jurors”.
First, smart contracts have to designate Kleros as their arbitrator in cases
of dispute, including the type of court (Kleros is developing an ecosystem
of specialized courts) and the number of juries to be involved (idem). When
a dispute arises, Kleros randomly assigns the dispute to a jury
of crowdsourced, self-selected experts, who analyze the evidence and vote
for a verdict. Jurors are penalized for communicating with each other, and
must “justify” their votes so that the parties can later understand their
decisions. A smart contract then transfers the money to the winning party.
Oracles are engaged to provide real-world data to assist dispute
resolution.[30]
A similar platform is Jur.io that advertises itself as a free service to users
for creating and securing smart contracts and resolving contract disputes
within 24 hours. Accordingly, Jur’s key promise seems to be speed and
security in smart contracting.[31] Its unique feature is the opportunity
to create their own hub (a “specialized oracle”) which operates on special
rules for users in particular industries.[32] Additionally, the Jur platform
provides tools for signing contracts, and creating and reselling contract
templates.[33]
29 _OATH Protocol. Blockchain Alternative Dispute Resolution Protocol. Version 2.6.0. Available_
from: https://oaths.io/files/OATH-Whitepaper-EN.pdf [Accessed 15 June 2022].
30 Allen, D., Lane, A. M. and Poblet, M. (2019) The Governance of Blockchain Dispute
Resolution. Harvard Negotiation Law Review, 25, pp. 75–101.
31 _JUR.Io – platforma kotoray pomozet razreshit finansovye spory mezdy investorami i srartupami._
(2018) [online] Available from: https://invest4all.ru/obzory-i-otchyoty/obzory-kraudsejlovico/jur-io-platforma-kotoraya-pomozhet-razreshit-finansovye-spory-mezhdu-investorami-istartapami [Accessed 15 June 2022].
32 Ibid.
33 Ibid.
-----
_154_ _Masaryk University Journal of Law and Technology_ _[Vol. 16:2_
It is worth pointing out that the above-mentioned platforms have
a dispute resolution mechanism with the following characteristics:
(i) adjudicator expertise in dispute resolution and law; (ii) independence
(neutral and anonymous adjudicators); (iii) impartiality (random selection
of judges without vested interests); and (iv) transparency (all procedures
are documented and rationalized).[34]
## 3. SHORTCOMINGS OF THE TRADITIONAL ARBITRATION INSTITUTIONS AND BLOCKCHAIN ARBITRATION
There are several drawbacks associated with “off-chain” arbitration. Firstly,
courts could only force the parties to execute a secondary transaction
or otherwise pay remedies for a smart contract that created damages for one
of the parties. Courts are not able to change the terms of the given smart
contract that was executed according to its parameters and added
to the blockchain because they could not change the existing code. Because
of these inherent limitations, courts are not able to render resolutions
to disputes arising from blockchain-based smart contracts. Secondly, it is
worth mentioning that high price is another disadvantage of traditional
arbitration institutions. In particular, Tang Z. S. states that the average
online consumer contract value is USD60, whereas an exemplary UK
provider of ODR services charges between GBP25 and GBP850 for
a resolution of consumer disputes. Therefore, even the lowest charge
of GBP25 will be disproportionately expensive compared with the average
value of the consumer disputes.[35]
Moreover, traditional arbitration institutions are characterized by a slow
speed of dispute resolution. However, in the online environment, people
would often like to get a quick decision. In relation to the incapability
of traditional dispute resolution to resolve numerous online disputes,
it should be pointed out that when the number of disputes runs into
the millions, human-powered dispute resolution cannot handle the scale
of disputes.[36]
34 Allen, D., Lane, A. M. and Poblet, M. (2019) The Governance of Blockchain Dispute
Resolution. Harvard Negotiation Law Review, 25, pp. 75–101.
35 Tang, Z. S. (2015) Electronic consumer contracts in the conflict of laws. 2[nd] ed. Oxford: Hart
Publishing, p. 373.
36 Dimov, D. (2017) Crowdsourced Online Dispute Resolution. [online] Ph.D. Leiden University.
-----
_2022]_ _M. Kasatkina: Dispute Resolution Mechanism for Smart Contracts_ _155_
Therefore, traditional arbitration mechanisms could not be the only
possible recourse for smart contract disputes.[37]
The first drawback of “on chain” arbitration concerns the enforceability
of awards. In other words, arbitral awards rendered through online
arbitration may not be recognized and enforced under the New York
Convention because, pursuant to Article 2 of the New York Convention, it
applies only to agreements “in writing”.[38] However, online arbitral
agreements would appear to satisfy the writing requirements
of the convention. The reason is that, under most national legislation,
electronic writings are considered equivalent to traditional writings.[39]
As a corollary, it is uncertain whether an award issued pursuant
to an arbitration agreement contained in the code of a smart contract would
be capable of being enforced.
The second drawback is the lack of trust in the procedures caused
by non-face-to-face communication. People who do not trust each other
may act tentatively and keep important information to themselves.
As a result, disputants participating in ODR processes may not disclose all
the relevant information to online arbitrators.[40] Moreover, criminals may
exploit the information security vulnerabilities of the ODR platform
in order to obtain unauthorized access to information related to the dispute
and the disputants. That is why the ODR provider should use information
security practices.[41]
The third drawback concerns the parties who may not be familiar and
comfortable with the relevant technology. Besides, it should be noted that
the legal qualification of arbitrators may be crucial for parties who want to
choose arbitrators with the special technical knowledge to adjudicate certain
disputes.
37 Kaal, W. A. and Calcaterra, C. (2018) Crypto Transaction Dispute Resolution. Business
_Lawyer. Available from: http://dx.doi.org/10.2139/ssrn.2992962 [Accessed 05 May 2022]._
38 _Convention on the Recognition and Enforcement of Foreign Arbitral Awards, 10 June 1958._
Available from: http://www.newyorkconvention.org/11165/web/files/original/1/5/15432.pdf
[Accessed 23 June 2022].
39 Cortes, P. (2010) Online Dispute Resolution for Consumers in the European Union. Routledge
_Research in IT and E-commerce Law. London: Routledge, Taylor & Francis Group. Available_
from: https://www.econstor.eu/bitstream/10419/181972/1/391038.pdf [Accessed 23 June
2022].
40 Ibid.
41 Lodder, A. R. and Zeleznikow, J. (2005) Developing an Online Dispute Resolution
Environment: Dialogue Tools and Negotiation Support Systems in a Three-Step Model.
_Harvard Negotiation Law Review, 10, pp. 287–337._
-----
_156_ _Masaryk University Journal of Law and Technology_ _[Vol. 16:2_
In addition, the described method of dispute resolution is obviously
devoid of a standard of efficiency, since there is no possibility to limit
in advance the range of checks used by arbitrators, who may not respect
the accumulated experience in resolving similar cases. As a result,
a decentralized court decision will become more and more resource-intensive over time, as the parties will try to determine all possible
circumstances in the program code. In other words, the parties will have to
discuss each dispute from the very beginning, without any knowledge
of the previous cases.
Besides, problems arise with the method of selection of the arbitrators
as well as ways of making their decisions. Arbitrators are selected
randomly, but from a certain group of specialists in the blockchain area,
which is not very big now. For that reason, there is a risk that the arbitrators
will not be independent of the parties.
To sum up, neither of these two alternative mechanisms can provide
an adequate environment for resolving disputes arising from smart
contracts. Therefore, in the next paragraph, I introduce the design and
implementation of a hybrid for the digital dispute.
## 4. HYBRID APPROACH
In light of the shortcomings of the available dispute resolution mechanisms
for the crypto economy, it is possible to talk about instituting a hybrid
approach. It means the creation of an independent, decentralized platform
that integrates both approaches to the smart contract dispute resolution
problem. This framework recognizes internal mechanisms of the smart
contract system that will regulate disputes depending on the precise nature
of the case and certain circumstances.
Parties should incorporate a mandatory dispute settlement clause
directly in the smart contract code.
Such a clause may include the following provisions:
a) automatic adoption of interim measures (for example, suspension
of performance of obligations under a smart contract, blocking of funds);
b) rules and deadlines for the creation of arbitration;
c) procedure and deadlines for dispute resolution;
d) procedure for the execution of arbitration awards; it means technical
standards that allow smart contracts to be reversed;
-----
_2022]_ _M. Kasatkina: Dispute Resolution Mechanism for Smart Contracts_ _157_
e) an agreement between the parties to resolve disputes using on-chain
resolution platforms. The lack of agreement between the parties should lead
to resolve the dispute with an on-chain system;
f) a clause regulating dispute resolution. For instance, by including
an ICC Arbitration Clause in a contract, the parties agree that their dispute
will be resolved by arbitration and that the arbitration proceedings will be
governed by the procedural rules in the ICC Rules of Arbitration, given
the finality and binding effect of an arbitral award for the parties.
Even if the dispute was resolved with “on chain” mechanisms,
the interested party should still have the right to appeal to the off-chain
arbitration. In these cases, decisions reached by way of blockchain
arbitration should not rise to the level of “off chain” arbitration.
To be specific, the off-chain arbitration should be viable for the following
cases:
- the disputes where one party is a consumer (taking into account
the level of consumer protection existing in the EU and its Member-States);
- the complex disputes (i.e. it is necessary to examine additional
evidence, to assign an expert examination or to hear witness testimony);
- the procedure may lead to the disclosure of commercial secrets;
- the disputes where fundamental rights are at stake.
This last condition is due to the impossibility to predict at the moment
of drafting the contract, what kind of disputes may arise between
the parties in the interpretation and performance of the contract. Therefore,
it should be possible for the parties to consider the dispute using traditional
arbitration.
Generally speaking, on-chain resolution platforms could be used for
resolving minor disputes (with a small cost), for instance cross-border
consumer disputes. Moreover, they could be used for technical disputes,
such as gas or share price determination and construction schedule
disputes. In other words, an “on chain” arbitration system could act
as an expert to resolve factual issues, such as whether a contract
performance complied with technical specifications, to calculate the market
value of shares or commodities, or to calculate damages. In these cases,
the parties may agree that the “on chain” arbitration award will be binding.
The ability of the parties to resolve disputes with online forms is of high
importance due to several benefits. Firstly, the high speed of online
procedures. Off-chain arbitration is not able to cope with the huge number
-----
_158_ _Masaryk University Journal of Law and Technology_ _[Vol. 16:2_
of online disputes. Secondly, the absence of on-chain resolution would
negate key blockchain benefits and would undermine the evolution
of the crypto economy.
However, on-chain arbitration requires the adaptation to the existing
legal regulation, primarily to the requirement of the New York Convention
to an arbitration agreement to be in writing. Otherwise, smart contracts run
the risk of not being enforced under the New York Convention, unless they
have an equivalent traditional word-format contract signed by both parties.
In this regard, it seems appropriate to have a hybrid version of smart
contracts, whereby there is a text-based version of the same force
in addition to the encrypted-coded-language smart contract.
All these considerations are compelling and favor a hybrid approach.
Given the current legal framework, fully “on chain” arbitration will not
become a reality in the nearest future. At the same time, prospects
of a hybrid approach are much more likely. It will reflect the complex
nature of blockchain technologies and the diversity of smart contracts used
in a dynamically competitive environment. On the one hand, the possibility
of using “on chain” arbitration will lead to speedy, less-costly awards,
to the benefit of parties in various specific sectors. Thus, the essence
of a smart contract will be reflected in comparison with a traditional
contract. On the other hand, “off chain” arbitration in certain cases seems
to be unavoidable given the legal realities of the modern world.
## 5. CONCLUSION
All in all, building and implementation of the effective dispute resolution
into smart contracts will be a crucial step in achieving level of certainty
in crypto transactions and facilitating the broadening evolution
of the crypto economy. Different mechanisms described above for resolving
smart contracts demonstrate various possibilities, opting human-driven
resolution systems or crowdsourced systems.
The development and introduction of new technologies should be
convenient for the participants, diminishing their risks and making it
possible to protect their rights in a faster manner. Besides, the use
of technology could be advantageous for the justice system, which could be
relieved of the burden of deciding certain kinds of disputes.
The hybrid approach that I suggest in this article addresses problems
that neither the “on chain” nor “off chain” approaches can address
-----
_2022]_ _M. Kasatkina: Dispute Resolution Mechanism for Smart Contracts_ _159_
separately. I argue that for some reasons, hybrid solutions are more
adequate given the framework of the Internet Age. The world is rapidly
changing, and laws will have to adapt to this rising tide. As such,
the growth of smart contracts will require adaptation by the legal profession
and modification of approaches to dispute resolution. In doing so, though,
contract law should operate according to its traditional canons and
categories, through a modification and supplementation of existing rules
and procedures.[42] And these technologies should be seen
as an improvement of existing contractual structures in terms of their
effectiveness. They cannot definitely change the essence of dispute
resolution relationships between the parties.
Without a doubt, using a hybrid architecture can substantially improve
the dispute resolution from smart contracts while retaining existing
traditional law rules and principles. However, there is a room for
specification of the individual conditions of “on chain” and “off chain”
arbitration.
## LIST OF REFERENCES
[1] Allen, D., Lane, A. M. and Poblet, M. (2019) The Governance of Blockchain Dispute
Resolution. Harvard Negotiation Law Review, 25, pp. 75–101.
[2] _Arbitraznyu zentr pri RSPP. [online] Available from: https://arbitration-rspp.ru/_
about/structure/boards/digital-disputes/ [Accessed 04 May 2022].
[3] Chamber of Digital Commerce. (2018) Smart Contracts: Is the Law Ready? Available from:
https://www.theblockchaintest.com/uploads/resources/CDC%20-%20Smart%20Contract
Is%20the%20Law%20Ready%20-%202018%20-%20Sep.pdf [Accessed 02 May 2022].
[4] Clément, M. (2019) Smart Contracts and the Courts. In: DiMatteo, L., Cannarsa, M. and
Poncibò, C. (eds.) The Cambridge Handbook of Smart Contracts, Blockchain Technology and
_Digital Platforms. Cambridge University Press, pp. 271–287._
[5] _Convention on the Recognition and Enforcement of Foreign Arbitral Awards, 10 June 1958._
Available from: http://www.newyorkconvention.org/11165/web/files/original/
1/5/15432.pdf [Accessed 23 June 2022].
42 Pardolesi, R. and Davola, A. (2019) What Is Wrong in the Debate About Smart Contracts.
_SSRN_ _Electronic_ _Journal._ Available from: https://www.researchgate.net/
publication/331834837_What_Is_Wrong_in_the_Debate_About_Smart_Contracts [Accessed
23 June 2022].
-----
_160_ _Masaryk University Journal of Law and Technology_ _[Vol. 16:2_
[6] Cortes, P. (2010) Online Dispute Resolution for Consumers in the European Union. Routledge
_Research in IT and E-commerce Law. London: Routledge, Taylor & Francis Group._
Available from: https://www.econstor.eu/bitstream/10419/181972/1/391038.pdf [Accessed
23 June 2022].
[7] _The Court of Arbitration of the Polish Blockchain and New Technology Chamber of Commerce._
[online] Available from: https:// blockchaincourt.org/ [Accessed 02 May 2022].
[8] The Court of Arbitration of the Polish Blockchain and New Technology Chamber
of Commerce (2019). The _Rules of the Court of Arbitration of the Polish Blockchain and New_
_Technology Chamber of Commerce. Available from: https://blockchaincourt.org/wp-_
content/uploads/2019/07/The-Rules-of-the-Court-of-Arbitration-ENG.pdf [Accessed
04 May 2022].
[9] Dimov, D. (2017) Crowdsourced Online Dispute Resolution. [online] Ph.D. Leiden
University.
[10] De Filippi, P. and Wright A. (2018) Blockchain and the Law: The Rule of Code. Cambridge,
Cambridge, MA: Harvard University Press, 300.
[11] Grimmelmann, J. (2019) All Smart Contracts are Ambiguous. Journal of Law & Innovation,
2 (1). Available from: https://www.law.upenn.edu/live/files/9782-grimmelmann-all
smart-contracts-are-ambiguous [Accessed 02 May 2022].
[12] International Chamber of Commerce (2018). ICC Dispute Resolution Bulletin. Issue 1.
Available from: https://www.hoganlovells.com/~/media/hogan-lovells/pdf/2018/
2018_12_13_icc_robots_arbitrator.pdf [Accessed 02 May 2022].
[13] JUR.Io – platforma kotoray pomozet razreshit finansovye spory mezdy investorami i srartupami.
(2018) [online] Available from: https://invest4all.ru/obzory-i-otchyoty/obzory
kraudsejlov-ico/jur-io-platforma-kotoraya-pomozhet-razreshit-finansovye-spory
mezhdu-investorami-i-startapami [Accessed 15 June 2022].
[14] Kaal, W. A. and Calcaterra, C. (2018) Crypto Transaction Dispute Resolution. Business
_Lawyer. Available from: http://dx.doi.org/10.2139/ssrn.2992962 [Accessed 05 May 2022]._
[15] Kerpelman, A. J. (2018) _Introducing the Juris Protocol: Human-Powered Dispute Resolution_
_for Blockchain Smart Contracts. [online] Available from: https://medium.com/jurisproject/_
introducing-the-juris-protocol-human-powered-dispute-resolution-for-blockchain-smart
contracts-bc574b50d8e1 [Accessed 05 May 2022].
-----
_2022]_ _M. Kasatkina: Dispute Resolution Mechanism for Smart Contracts_ _161_
[16] Lodder, A. R. and Zeleznikow, J. (2005) Developing an Online Dispute Resolution
Environment: Dialogue Tools and Negotiation Support Systems in a Three-Step Model.
_Harvard Negotiation Law Review, 10, pp. 287–337._
[17] OATH Protocol. Blockchain Alternative Dispute Resolution Protocol. Version 2.6.0. Available
from: https://oaths.io/files/OATH-Whitepaper-EN.pdf [Accessed 15 June 2022].
[18] Pardolesi, R. and Davola, A. (2019) What Is Wrong in the Debate About Smart Contracts.
_SSRN Electronic Journal. Available from: https://www.researchgate.net/publication/_
331834837_What_Is_Wrong_in_the_Debate_About_Smart_Contracts [Accessed 23 June
2022].
[19] Rodrigues, U. (2018) Law and the Blockchain. Iowa Law Review, 104. Available from:
https://ilr.law.uiowa.edu/print/volume-104-issue-2/law-and-the-blockchain/ [Accessed
02 May 2022].
[20] Rule, C. and Nagarajan, C. (2011) Crowdsourcing Dispute Resolution Over Mobile
Devices. In: Poblet, M. (ed.) Mobile Technologies for Conflict Management. Law, Governance
_and Technology Series, vol 2. Dordrecht: Springer, pp. 93–100._
[21] Sanchez Dr W. _Dispute Resolution in OpenBazaar._ [online] Available from:
http://docs.openbazaar.org/03.-OpenBazaar-Protocol/ [Accessed 21 June 2022].
[22] Schmitz, A. J. and Rule, C. (2019) Online Dispute Resolution for Smart Contracts. Journal
_of Dispute Resolution. University of Missouri School of Law Legal Studies Research Paper,_
2019 (11). Available from: https://ssrn.com/abstract=3410450 [Accessed 12 April 2022].
[23] Schmitz, A. and Rule, C. (2019) Online Dispute Resolution for Smart Contracts. Journal
_of Dispute Resolution, 2, pp. 103–125._
[24] Sillanpaa, T. (2020) Freedom to (Smart) Contract: The Myth of Code and Blockchain
Governance Law. _IALS Student Law Review,_ 7 (2). Available from:
https://journals.sas.ac.uk/lawreview/issue/view/582 [Accessed 02 May 2022].
[25] Szczudlik, K. (2019) “On-chain” and “off-chain” arbitration: Using smart contracts to amicably
_resolve disputes. [online] Available from: https://newtech.law/en/on-chain-and-off-chain-_
arbitration-using-smart-contracts-to-amicably-resolve disputes [Accessed 02 May 2022].
[26] Tang, Z. S. (2015) Electronic consumer contracts in the conflict of laws. 2[nd] ed. Oxford: Hart
Publishing, p. 373.
[27] Zagaynova, M. (2018) Obzor ICO proekta Mattereum. [online] Available from:
https://ffc.media/ru/overviews/ico-mattereum-project-review/ [Accessed 21 June 2022].
-----
_162_ _Masaryk University Journal of Law and Technology_ _[Vol. 16:2_
[28] Zasemkova O. (2020) Methods of Resolving Disputes Arising from Smart Contracts. Lex
_Russica, 73 (4), p. 20._
[29] Zaslowsky, D. (2018) What to Expect When Litigating Smart Contract Disputes. [online]
Available from: https://www.law360.com/articles/1028009/what-to-expect-when
litigating-smart-contract-disputes [Accessed 02 May 2022].
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.5817/mujlt2022-2-2?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.5817/mujlt2022-2-2, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "GOLD",
"url": "https://journals.muni.cz/mujlt/article/download/18580/26353"
}
| 2,022
|
[] | true
| 2022-09-30T00:00:00
|
[
{
"paperId": "79538bf4a093fdc52fe6cea592045bcc23fc900a",
"title": "Freedom to (Smart) Contract: The Myth of Code and Blockchain Governance Law"
},
{
"paperId": "9697cd944da32cbc5e8758fe56f589740e89bb34",
"title": "Methods of Resolving Disputes Arising from Smart Contracts"
},
{
"paperId": "3379525cc9ce816fa06fedcc7edc74840fda48de",
"title": "Smart Contracts and the Courts"
},
{
"paperId": "0731a6799b22a5a92b5c49d2227b0939156d1421",
"title": "Blockchain and the Law: The Rule of Code"
},
{
"paperId": "4589fa535d8f3c6c086f7977f5d8ba28aa771e62",
"title": "Online Dispute Resolution for Smart Contracts"
},
{
"paperId": "921fe5b1d9f66845c878289d942e1c4229619cc8",
"title": "What Is Wrong in the Debate About Smart Contracts"
},
{
"paperId": "044dd0452e813aac098d1de11e899cb9ab071b53",
"title": "The Governance of Blockchain Dispute Resolution"
},
{
"paperId": "f922f790930b5795f525af54f21c8932529fe8a9",
"title": "All Smart Contracts Are Ambiguous"
},
{
"paperId": "a3095f085b07903fe4b9018f3cc6a2e9ef42c670",
"title": "Law and the Blockchain"
},
{
"paperId": "2f40c7c8b8ea5573a7e616c7d7a23b4024f71927",
"title": "Crowdsourced Online Dispute Resolution"
},
{
"paperId": "166996fc47f25e1a3918d28166a5ac907f43ebc5",
"title": "Crypto Transaction Dispute Resolution"
},
{
"paperId": "14f8a7d1993977c485cc3e32fa690c99ac96606d",
"title": "Online Dispute Resolution for Consumers in the European Union"
},
{
"paperId": "e11136635fff95aac5bf8c8ab7be0cd1b1ab2def",
"title": "Electronic Consumer Contracts in the Conflict of Laws"
},
{
"paperId": "026fe761d9c2171d9c7dc5cafaf882f9ab19e014",
"title": "Developing an Online Dispute Resolution Environment: Dialogue Tools and Negotiation Support Systems in a Three-Step Model"
},
{
"paperId": null,
"title": "Arbitraznyu zentr pri RSPP"
},
{
"paperId": "99741b9fca17a05abc21aa7d14bd43e3107d2fdf",
"title": "The Cambridge Handbook of Smart Contracts, Blockchain Technology and Digital Platforms"
},
{
"paperId": null,
"title": "“On-chain” and “off-chain” arbitration: Using smart contracts to amicably resolve disputes"
},
{
"paperId": null,
"title": "ICC Dispute Resolution Bulletin. Issue 1"
},
{
"paperId": null,
"title": "Obzor ICO proekta Mattereum"
},
{
"paperId": null,
"title": "Smart Contracts: Is the Law Ready? Available from"
},
{
"paperId": null,
"title": "Introducing the Juris Protocol: Human-Powered Dispute Resolution for Blockchain Smart Contracts"
},
{
"paperId": null,
"title": "What to Expect When Litigating Smart Contract Disputes"
},
{
"paperId": null,
"title": "JUR.Io – platforma kotoray pomozet razreshit finansovye spory mezdy investorami i srartupami"
},
{
"paperId": "67076814f2b7335cea04f31b1f33815ab34c8117",
"title": "Crowdsourcing Dispute Resolution over Mobile Devices"
},
{
"paperId": null,
"title": "The Rules of the Court of Arbitration of the Polish Blockchain and New Technology Chamber of Commerce"
},
{
"paperId": null,
"title": "Convention on the Recognition and Enforcement of Foreign Arbitral Awards, 10 June 1958"
},
{
"paperId": null,
"title": "Blockchain Alternative Dispute Resolution Protocol. Version 2.6.0"
}
] | 12,282
|
en
|
[
{
"category": "Medicine",
"source": "external"
},
{
"category": "Biology",
"source": "s2-fos-model"
},
{
"category": "Medicine",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/02306c29e74698aa4e2ac5e75c872514872ffbe5
|
[
"Medicine"
] | 0.879291
|
Genome-wide analysis of genes encoding core components of the ubiquitin system during cerebral cortex development
|
02306c29e74698aa4e2ac5e75c872514872ffbe5
|
Molecular Brain
|
[
{
"authorId": "4799055",
"name": "A. Bouron"
},
{
"authorId": "3657994",
"name": "M. Fauvarque"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"Mol Brain"
],
"alternate_urls": [
"http://www.molecularbrain.com/"
],
"id": "239c22f5-f478-44e5-877f-8fb0e2daedb4",
"issn": "1756-6606",
"name": "Molecular Brain",
"type": "journal",
"url": "https://molecularbrain.biomedcentral.com/"
}
|
Ubiquitination involves three types of enzymes (E1, E2, and E3) that sequentially attach ubiquitin (Ub) to target proteins. This posttranslational modification controls key cellular processes, such as the degradation, endocytosis, subcellular localization and activity of proteins. Ubiquitination, which can be reversed by deubiquitinating enzymes (DUBs), plays important roles during brain development. Furthermore, deregulation of the Ub system is linked to the pathogenesis of various diseases, including neurodegenerative disorders. We used a publicly available RNA-seq database to perform an extensive genome-wide gene expression analysis of the core components of the ubiquitination machinery, covering Ub genes as well as E1, E2, E3 and DUB genes. The ubiquitination network was governed by only Uba1 and Ube2m , the predominant E1 and E2 genes, respectively; their expression was positively regulated during cortical formation. The principal genes encoding HECT (homologous to the E6-AP carboxyl terminus), RBR (RING-in-between-RING), and RING (really interesting new gene) E3 Ub ligases were also highly regulated. Pja1 , Dtx3 (RING ligases) and Stub1 (U-box RING) were the most highly expressed E3 Ub ligase genes and displayed distinct developmental expression patterns. Moreover, more than 80 DUB genes were expressed during corticogenesis, with two prominent genes, Uch-l1 and Usp22, showing highly upregulated expression. Several components of the Ub system overexpressed in cancers were also highly expressed in the cerebral cortex under conditions not related to tumour formation or progression. Altogether, this work provides an in-depth overview of transcriptomic changes during embryonic formation of the cerebral cortex. The data also offer new insight into the characterization of the Ub system and may contribute to a better understanding of its involvement in the pathogenesis of neurodevelopmental disorders.
|
p g
## RESEARCH
## Open Access
# Genome‑wide analysis of genes encoding core components of the ubiquitin system during cerebral cortex development
### Alexandre Bouron[1,2*] and Marie‑Odile Fauvarque[1]
**Abstract**
Ubiquitination involves three types of enzymes (E1, E2, and E3) that sequentially attach ubiquitin (Ub) to target pro‑
teins. This posttranslational modification controls key cellular processes, such as the degradation, endocytosis, subcel‑
lular localization and activity of proteins. Ubiquitination, which can be reversed by deubiquitinating enzymes (DUBs),
plays important roles during brain development. Furthermore, deregulation of the Ub system is linked to the patho‑
genesis of various diseases, including neurodegenerative disorders. We used a publicly available RNA-seq database to
perform an extensive genome-wide gene expression analysis of the core components of the ubiquitination machin‑
ery, covering Ub genes as well as E1, E2, E3 and DUB genes. The ubiquitination network was governed by only Uba1
and Ube2m, the predominant E1 and E2 genes, respectively; their expression was positively regulated during cortical
formation. The principal genes encoding HECT (homologous to the E6-AP carboxyl terminus), RBR (RING-in-betweenRING), and RING (really interesting new gene) E3 Ub ligases were also highly regulated. Pja1, Dtx3 (RING ligases) and
_Stub1 (U-box RING) were the most highly expressed E3 Ub ligase genes and displayed distinct developmental expres‑_
sion patterns. Moreover, more than 80 DUB genes were expressed during corticogenesis, with two prominent genes,
_Uch-l1 and Usp22, showing highly upregulated expression. Several components of the Ub system overexpressed in_
cancers were also highly expressed in the cerebral cortex under conditions not related to tumour formation or pro‑
gression. Altogether, this work provides an in-depth overview of transcriptomic changes during embryonic formation
of the cerebral cortex. The data also offer new insight into the characterization of the Ub system and may contribute
to a better understanding of its involvement in the pathogenesis of neurodevelopmental disorders.
**Keywords: Rodent, Brain, Cerebral cortex, Ubiquitin, Ubiquitination, Deubiquitinating enzymes**
**Introduction**
Ubiquitination is a multistep process during which ubiquitin (Ub), a versatile and highly conserved 76 amino-acid
polypeptide, is covalently conjugated to target substrates.
It is one of the most common posttranslational modifications of proteins [1] and requires the sequential action
of three types of enzymes: Ub-activating (E1) enzymes,
*Correspondence: alexandre.bouron@cea.fr
2 Genetics and Chemogenomics Lab, Building C3, CEA, 17 rue des Martyrs,
38054 Grenoble Cedex 9, France
Full list of author information is available at the end of the article
Ub-conjugating (E2) enzymes and Ub-ligases (E3) [2].
Ubiquitination is counterbalanced by the action of deubiquitinating enzymes (or deubiquitinases, DUBs) that
can reverse the conjugation of Ub to substrates. Mammalian DUBs are classified into seven categories: Ub-specific proteases (USPs), Ub carboxyl-terminal hydrolases
(UCHs), otubain proteases (OTUs), Machado-Joseph
disease protein domain proteases (MJDs or Josephins),
JAB1/MPN/Mov34 metallopeptidases (JAMMs), motifinteracting with Ub-containing novel DUB family
(MINDY) and ZUP1 [3, 4]. Ubiquitination is described as
a quality control system devoted to protein homeostasis
© The Author(s) 2022. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which
permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the
original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or
other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line
to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory
regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this
[licence, visit http://creativecommons.org/licenses/by/4.0/. The Creative Commons Public Domain Dedication waiver (http://creativeco](http://creativecommons.org/licenses/by/4.0/)
[mmons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data.](http://creativecommons.org/publicdomain/zero/1.0/)
-----
since it targets damaged or misfolded proteins for degradation via the Ub–proteasome system. However, ubiquitination shows a wider physiological importance because
the conjugation of Ub can modify the activity of their
targets, changing their subcellular localization or involvement in the formation of multiple protein complexes.
Therefore, ubiquitination is involved in the regulation of
various key cellular processes, such as endocytosis, cell
signalling, autophagy and DNA repair [5].
A large number of proteins in the brain are ubiquitinated. For instance, an analysis of the ubiquitome (ubiquitinated proteins) revealed 921 targets in the rat brain
with numerous pre- and postsynaptic actors [6]. The
Ub pathway controls multiple neuronal processes, such
as neuron migration, growth and synaptic transmission
[7–10]. The Ub system also governs fundamental mechanisms controlling memory reorganization [11]. Moreover,
alterations of the Ub pathway are thought to contribute
to neurodevelopmental, cognitive and age-related neurodegenerative diseases [8, 12]. It is thus of paramount
importance to understand how the Ub system participates in normal brain formation and development. The
aim of this study was to provide an extensive and detailed
overview of the expression pattern of genes encoding
major factors mediating ubiquitination and deubiquitination during the formation of the cerebral cortex in mice.
The data presented rely on a published RNA-seq database [13] covering 4 stages of cortical development corresponding to the beginning (embryonic Day 11, E11), the
peak (E13), and the end of neurogenesis (E17), followed
by the beginning of the maturation process and neuronal
circuit assembly (postnatal Day 1, PN1) [14, 15]. These
four periods cover stages of profound cell division (E11E13) followed by stages characterized by the growth and
morphological differentiation of the postmitotic neurons
and the establishment of neural networks (E17-PN1)
[14, 15]. The analysis of gene expression patterns during
development will help to understand the extraordinary
complexity of the Ub conjugation/deconjugation system.
This study not only presents a description of transcription profile changes during embryonic cerebral cortex
formation but also provides an in-depth overview of the
core components of the Ub system as discerned through
published data.
**Materials and methods**
The analysis was based on a published RNA-seq gene
expression dataset on E11, E13, E17 and PN1 [13], reflecting the period of cerebral cortex formation in the mouse
brain. The complete dataset is freely accessible from the
GEO repository with the accession number GSE154677.
Throughout this study, the results are expressed in
transcripts per million (TPM) and the mean ± standard
error of mean (SEM).
**Results and discussion**
The data obtained from the genome-scale profiling of
gene expression are organized into two parts: the first
part covers the ubiquitination process, and the second
part is devoted to the genes encoding DUBs, a class of
enzymes participating in the novo synthesis of Ub and
are responsible of the deubiquitination of substrates.
**Ubiquitination**
This first section is subdivided into 4 sections covering
the following topics: (1) genes involved in the synthesis of
Ub and Ub-like proteins, followed by the genes encoding
(2) Ub-activating (E1) enzymes; (3) Ub-conjugating (E2)
enzymes; and (4) Ub ligases (E3).
**_Ubiquitin genes_**
In mammals, two classes of genes are involved in de novo
synthesis of Ub: the monomeric Ub-ribosomal fusion
genes _Uba52 and_ _Rps27a (Uba80) and the stress-induc-_
ible poly-Ub genes _Ubb and_ _Ubc [16]. In the immature_
cerebral cortex, _Ubb and_ _Rps27a were the most highly_
expressed Ub genes (Fig. 1A). Ubb transcripts accounted
for 30–40% of the total Ub transcripts, which is actually
in close agreement with previously reported data [17],
indicating a high abundance of total _Ubb transcripts in_
the brain. _Ubb and_ _Rps27a followed dissimilar expres-_
sion patterns: transcripts of Rps27a predominated at the
onset of corticogenesis (E11–13) before their number significantly decreased, whereas the Ubb gene was the most
highly Ub gene expressed at the end of corticogenesis
(E17-PN1). Notably, the number of transcripts (expressed
in transcripts per million, TPM) was on the order of
700–1700, reflecting a very high abundance of _Ubb,_
_Rps27a and_ _Uba52 transcripts at all stages (Fig._ 1A). In
comparison, the TPM values of H2afz (encoding histone
2A, member z), one of the most abundant cellular proteins ubiquitously expressed [18], were 1740 (at E11) and
400 (on PN1). H2af was the 66th most highly expressed
gene on E11, whereas Rps27a was the 68th mostly highly
expressed. On PN1, _Ubb, the most prominent Ub gene_
(with TPM values of 1400) was the 60th most highly
expressed gene in the immature cerebral cortex. These
data are in line with the known high abundance of the Ub
protein in biological samples, in which it has been shown
to comprise up to 5% of total protein [19]. These 4 genes
generate single (Uba52 and Rps27a) or multiple copies (9
for Ubc, and 3 for Ubb) of Ub. Therefore, to better assess
the contribution of these genes to the total pool of Ub,
we calculated the theoretical production of Ub molecules
assuming a similar translation efficiency for the four
-----
|Col1|Col2|
|---|---|
|||
transcripts [20]. The Ubc gene accounted for nearly 65%
of the total pool of Ub in the embryonic cortex, followed
by Ubb, Rps27a and Uba52, accounting for 18%, 11% and
6%, respectively. These data are in line with a previous
report showing that UBC accounts for 64% of the total
Ub pool in HeLa cells [20]. Thus, the poly-ubiquitin gene
_Ubc is the major cellular contributor of Ub molecules._
In the mouse brain, 60% of the Ub pool is in the free
form (i.e., not attached to target substrates) [19]. The
level of free Ub is important to neuronal functions and
survival [16, 19]. The morphology, neurite outgrowth and
synaptic development are impaired in cultured neurons
isolated from the brains of Ubb-knockout mice [21].
A recent study reported 52 Ub pseudogenes in humans
[22]. Moreover, some of these genes, such as human
_Ubb pseudogene 4 (Ubbp4),_ _Rps27a pseudogene 16_
(Rps27ap16), and _Uba52 pseudogene 8, encode pro-_
teins [22]. Here, the expression of the following murine
Ub pseudogenes was examined: _Gm1821 (Ubb-pseudo_
_gene or_ _Ubb-ps),_ _Rps27a-ps1,_ _Rps27a-ps2 and_ _Gm7866_
(Uba52-ps). Only transcripts of the pseudogene Gm1821
were found with TPM values of ⁓ 15, which was nearly
-----
⁓ 100-fold less abundant than _Ubb transcripts. In con-_
clusion, our data indicated a high expression level of
the three Ub-encoding genes _Ubb,_ _Rps27a and_ _Uba52._
In comparison, the expression of Ub pseudogenes was
negligible.
**_Ubiquitin‑like genes_**
Similar to Ub, various Ub-like proteins can be covalently
conjugated to target substrates via an enzymatic cascade
involving E1, E2, and E3 enzymes [23], [24]. The expression of this set of genes was analysed, and the results are
reported in Fig. 1B. The following 10 genes were identified: _Sumo1-3, Nedd8, Isg15, Ubd (Fat10), Ufm1, Atg8_
_(Map1lc3b), Atg12, and_ _Urm1. No transcript for_ _Sumo2_
or _Ubd was found (TPM < 2) [25]. All the other genes_
were, however, significantly expressed, with Atg8, Sumo3
and _Nedd8 displaying the highest levels of expression._
For this set of genes, the TPM ranged from 70 to ⁓ 200
(Fig. 1B), a number that is approximately tenfold lower
than that of Ub genes (Fig. 1A). Atg8 was the major gene,
and its expression was highly upregulated during development, with a number of transcripts showing expression
increases by a factor of 3 between E11 and E17, suggesting activation of Atg8-dependent physiological processes,
such as autophagy activation, at the end of cortical development (Fig. 1B). A recent proteomic analysis showed
that the levels of conjugated and free NEDD8 (or ISG15)
in the mouse brain were at least 60- and 20-fold lower
than those of Ub [6], which is consistent with the profoundly lower gene expression levels of these genes, particularly Isg15, whose expression was negligible.
**_Ub‑activating (E1) enzymes_**
Ubiquitination is a three-step enzymatic reaction. During
the initial step, Ub is activated in an adenosine triphosphate-dependent manner by the Ub-activating (E1)
enzyme before being transferred to a Ub-conjugating
(E2) enzyme [2, 26]. Figure 1C shows the expression profile of the two genes Uba1 and Uba6 (Ube1l2) encoding
mammalian Ub-activating E1 enzymes and seven genes
encoding Ub-like proteins that activate E1 enzymes
(Uba2-3, _Uba5,_ _Uba7,_ _Nae1, and_ _Sae1) [24]. The_ _Uba1_
gene was by far the most prominently expressed E1
gene in the cerebral cortex (Fig. 1C). Its TPM values
were ⁓ 200 on E11 and ⁓ 390 on E17, thus showing a
nearly twofold increase in transcript abundance during
embryonic development. Abundant expression of UBA1
may be a common feature of many cellular types, as the
UBA1 protein is among the top 2% of the most highly
expressed proteins in HeLa cells [18], reflecting the crucial requirement of this E1 enzyme in Ub-dependent cell
processes. In the cerebral cortex, _Uba1 was the 597th_
and 343th most highly expressed gene on E11 and PN1,
respectively, which confirms the relatively high abundance of _Uba1 transcripts. The_ _Uba1 gene product is_
abundant in the nucleus and cytoplasm [27], whereas the
other mammalian E1 _Uba6 gene product is found only_
in the cytoplasm, which may ensure much more specific
and restricted functions. Uba6 was expressed at very low
levels, with TPM values ranging from ⁓ 9 on E11 to ⁓ 4
on PN1. In the cerebral cortex, the ratios of Uba1:Uba6
transcript abundance were > 20:1 and 90:1 on E11 and
PN1, respectively. This differential expression was consistent with proteomic data showing that the relative
abundance ratio of the UBA1 and UBA6 proteins is > 10:1
in HeLa cells [18], further suggesting a restricted function for UBA6 compared to that of UBA1. Collectively,
the expression profile data of Uba1 showed that it is the
primary E1 gene in the cerebral cortex of mice. Due to its
central role in Ub homeostasis, UBA1 is likely to regulate
a wide range of neurobiological processes [28].
Far below the level of expression observed for _Uba1,_
the expression of a set of six genes encoding Ub- and
Ub-like proteins activating E1 enzymes (Uba2-3, _Uba5,_
_Uba7, Nae1, and Sae1) exhibited TPM values < 100. Uba2_
encodes an E1 enzyme specific for the Ub-like molecule
SUMO. This Uba2 gene product is thought to form heterodimers with SAE1 [24, 26]. Interestingly, the expression of both genes (Uba2 and _Sae1) was downregulated_
during development, with an abundance of transcripts
reduced by nearly 50% between E11 and PN1 (Fig. 1C).
Notably, no transcript for the Ub-like E1 gene _Atg7 was_
detected. Despite its low level of expression in the developing cortex, _Uba6 plays important roles in neuronal_
development, dendritic spine architecture, and mouse
behaviour, and its deficiency is lethal [29]. Moreover,
_Uba6 is required for neuronal viability in primary hip-_
pocampal neuronal cultures. Collectively, our data led to
the identification of the highly regulated Uba1 and Uba2
genes as the major E1 and E1-like enzyme genes, respectively, expressed during cortical brain development.
**_Ub‑conjugating (E2) enzymes_**
The analysis of the genes encoding Ub- and Ub-like
protein-conjugating E2 enzymes [30, 31] is shown in
Fig. 1D, E. Transcripts for _Ube2d4,_ _Ube2e2,_ _Cdc34b,_
_Atg10 and Ube2u were not found. This observation rein-_
forces the validity of our results since Ube2u transcripts
were detected specifically in tissues of the urogenital
tract [32]. All the other genes were expressed at significant levels (35 of 40 genes), particularly Ube2m (Ubc12),
encoding a Nedd8-conjugating enzyme, with TPM values
increasing from ⁓270 on E11 to ⁓320 on PN1 (Fig. 1D).
The most important expressed genes encoding Ub-conjugating E2 enzymes were _Ube2c and_ _Ube2r1 (cdc34)_
(Fig. 1D) and, to a lesser extent, _Ube2q1,_ _Ube2ql1, and_
-----
_Ube2z (Use1) (Fig. 1E). Expression of Ube2c was inhibited_
during embryonic development, with high levels of transcripts evident in the neurogenic period (E11–E13) and a
marked decrease from E17 and later. Overall, TPM values
of the _Ube2c gene decreased by a factor of 25 between_
E11 (> 150 TPM) and PN1 (⁓ 7 TPM) (Fig. 1D). This
decline represented the most downregulated genes in the
Ub pathway during corticogenesis. The UBE2C protein
is an exclusive partner of APC/C E3 ligases and controls
cell cycle progression. Its mRNA and protein levels were
low in quiescent cells but greatly increased and peaked
during mitosis [31, 33]. Ube2c mRNA was thought to be
barely detectable in tissues except under oncogenic conditions, with high levels in various cancers, such as brain
and breast cancers [33]. The data presented in Fig. 1D
show that high levels of Ube2c mRNA were found in nontumorous tissue under conditions not related to cancer
onset or progression, similar to many developmentalspecific genes whose re-expression is associated with
carcinogenesis. Further studies are required to verify
whether the UBE2C protein is a marker of neurogenesis
in the brain.
_Ube2r1 was another prominently expressed Ub-con-_
jugating E2 enzyme-encoding gene. Similar to that of
_Ube2c, the expression of_ _Ube2r1 was repressed, with an_
abundance of transcripts reduced by a factor of 4 between
E11 and PN1 (from 126 to 34 TPM) (Fig. 1D). The E2
enzyme CDC34 (encoded by _Ube2r1) is the primary E2_
for cullin-RING E3 ligases (CRLs). Two members of the
_Ube2q gene family were expressed at moderate levels:_
_Ube2q1 and Ube2ql1. The expression of the Ube2q1 gene_
was constant, showing no sign of developmental regulation. Western blot and immunohistochemical experiments showed the presence of UBE2Q1 proteins in the
rat brain cortex, mainly in neurons [34]. UBE2Q1has
been postulated to play an anti-apoptotic role, at least in
pathological states such as traumatic brain injury [34].
The expression of Ube2ql1 was not detected before E13,
and it peaked in the E17-PN1 period, with TPM values
increasing from ⁓8 to 85 TPM from E13 to E17. This
increase represented an 11-fold increase in transcript
abundance, making _Ube2ql1 the second most induced_
gene (in the Ub system) during embryonic development.
In HeLa cells, UBE2QL1 exhibited a dual function: it is
required for the efficient clearance of damaged lysosomes
by lysophagy and maintains lysosome integrity [35].
The expression of minor genes (TPM values < 40)
encoding Ub- and Ub-like proteins-conjugating E2
enzymes is reported in Additional file 1: Fig. S1A, B. The
vast majority of these genes were expressed at constant
levels, except _Ube2g2,_ _Ube2s,_ _Ube2t and_ _Ube2l6, which_
were downregulated. In particular, the expression Ube2l6
was profoundly repressed, with TPM values decreasing
from ⁓ 18 to ⁓ 2 (a ⁓ninefold reduction in transcript
abundance) (Additional file 1: Fig. S1B), making it one
of the most important downregulated genes observed in
this study. Mutations in the Ube2a gene lead to neurodevelopmental disorders such as X-linked syndromic mental retardation. The precise roles _Ube2a plays in brain_
formation are unknown. In the rodent brain, the _Ube2a_
gene was expressed at low levels, with few transcripts
evident throughout cortical development (Additional
file 1: Fig. S1A). It has been proposed that some of the
cellular effects of UBE2A involve the E3 Ub ligase Parkin.
This might be the case in adults, but this enzyme was not
expressed during embryonic development of the cerebral
cortex, suggesting the involvement of other E3 partners
(see below, “RING E3 Ub ligases”). UBE2A exerts some of
its actions via the E3 Ub ligases RAD18 and RNF20. The
UBE2A/RAD18 complex is at least partially responsible
for the pathogenesis of mental retardation (in association
with the proliferating cell nuclear antigen, PCNA [12]).
The rad18 gene was expressed uniquely on E11 and E13,
but the abundance of transcripts was very low (< 6 TPM),
suggesting that its presence is required for brain development exclusively during neurogenesis.
This transcriptomic analysis provides a detailed overview of the expression patterns of E2 genes that are central players in the Ub pathway. Overall, these genes were
expressed at low levels during embryonic development of
the mouse cerebral cortex. As in HeLa cells or Swiss 3T3
cells, Ube2m (encoding UBE2M/UBC12 and involved in
neddylation) was the prominent E2. However, the pattern
of expression of the Ub- and Ub-like protein-conjugating
E2 genes in the cerebral cortex did not completely overlap with that reported in cell lines. For instance, UBE2I/
UBC9 (encoding SUMO) and UBE2N were abundant
E2 proteins in cell lines [18]. Together with UBE2V1,
UBE2N represents > 50% of the Ub-dedicated E2
enzymes in HeLa cells, and it is associated with UBE2V2
in Swiss 3T3 cells [18]. UBE2L3 is another abundant E2
that is expressed at levels twofold higher that all HECT
and RBR E3 ligase genes in HeLa cells combined [18].
This expression pattern is profoundly different than that
of the cerebral cortex, where the Ube2l3 and Ube2n genes
were expressed at moderate levels (TPM values of 20–30)
(Additional file 1: Fig. S1). The E2 enzyme UBE2L3
works in concert with the E3 ligase Ube3a (also known
as E6-associated protein or E6-AP) [36]. Mutations and
genetic defects in the Ube3a gene are associated with the
Angelman syndrome, a neurodevelopmental disease [37].
The profile of E2 enzyme expression in nontumorous tissue is likely to differ from that in immortalized cells. Our
data clearly illustrate that the pattern of E2 gene expression was temporally regulated. This pattern may, however, differ from one brain area to another.
-----
**_Ub ligases (E3)_**
Ub ligase (E3) enzymes exert two crucial functions: they
target a specific type of ubiquitinated substrate and enable the final transfer of Ub (to the substrate) [38]. In this
study, E3 Ub ligases are grouped into 3 families according
to [39]: the HECT (homologous to the E6-AP carboxyl
terminus), RBR (RING-in-between-RING), and RING
(really interesting new gene) E3 families. Depending on
the E3 ligase, the transfer of Ub from the E2 enzyme to
the target substrate can occur directly or via a 2 step
process. For instance, RING E3 ligases enable a direct
transfer of Ub from E2 to the target whereas the ubiquitination involving HECT develops via a 2 step process
with Ub carried by the E2 enzyme binds first to a HECT
domain before being transferred to the target protein [38,
39]. Although RBR E3 ligases have two RING domains
and could be categorised as a sub-class of RING-type
ligases, they are described as RING-HECT hybrids catalysing ubiquitination not directly like RING-type ligases
but via a two-step reaction like HECT-type ligases during
which Ub is transferred to the RING2 domain and then
to the target [38–41].
_HECT Ub ligases Genes were subdivided into 3 groups:_
Nedd, Herc and other HECT ligases [42]. Transcripts of
twenty-four HECT genes were found (Fig. 2), with four
genes (Ube2cbp, Hace1, Herc6, Hecw2) below the detection limit. The group of HECT E3 Ub ligases was dominated by the high abundance of _Nedd4 transcripts. The_
TPM values decreased from 249 (on E11) to 84 (on PN1),
reflecting a ⁓ threefold reduction in transcript abundance
throughout corticogenesis. _Nedd4 was the most highly_
expressed HECT E3 gene during neurogenesis (E11–
E13) and the second most highly expressed gene at the
end of corticogenesis (E17-PN1), after Hectd3, the other
prominent HECT gene. Our data were in line with the
first study reporting the isolation of both a set of Nedd4
cDNA clones and the corresponding mRNA expression [43]. This later study also showed a gradual mRNA
decrease during embryonic development in the brain.
_Nedd4l, which is closely related to_ _Nedd4, the found-_
ing and most ancient member of the Nedd4 family, was
expressed on E13 and onwards at a very low level (TPM
values < 9). These data illustrate the temporal regulation
of these E3 Ub ligases, which play essential roles in neuronal cell fate determination and survival, neurite outgrowth, axon guidance and branching [44].
Compared to HECT, E3 Ub ligases in the Herc group
were expressed at much lower levels, with only _Herc1_
and _Herc2 showing a TPM reaching a maximum of 23,_
on E17 (Fig. 2). In the third category (i.e., the HECT E3
Ub ligases that differ from Nedd and Herc), Hectd3 was
the predominant gene, with highly regulated expression
during corticogenesis, with TPM values increasing by
a factor of 3 from E11 to PN1 (from ⁓ 30 to ⁓ 94 TPM)
(Fig. 2). _Huwe1 was the third most highly expressed_
gene of this subfamily of HECT E3 ligases, with TPM
values of 43–60. Interestingly, knocking down HUWE1
expression in cortical tissue with a siRNA resulted in an
increase in the fraction of proliferating cells in the developing brain and blockade of neuronal differentiation
[45]. These results show that HUWE1 controls neural
differentiation and proliferation. All the other genes in
this subgroup were expressed at low levels, with TPM
values < 30 (Fig. 2). The HECT E3 Ub ligase UBE3A has
been well characterized because mutations in the Ube3a
gene cause Angelman syndrome, a neurodevelopmental
disease [37]. However, _Ube3a was expressed at low lev-_
els throughout the cortex (TPM values of 16–25, Fig. 2).
Notably, the highest abundance of HECT Ub ligase gene
transcripts was noted at the peak of neurogenesis (E13)
and then declined.
-----
_RING E3 Ub ligases RING E3 ligases constitute the larg-_
est family of E3 Ub ligases [8] [38]. For instance, a previous analysis of the mouse genome identified 398 putative
E3 enzymes [46]. In the following sections, RING E3 Ub
ligase genes are classified into three main subgroups:
(1) single subunit, (2) multiple subunit RING E3 and (3)
U-box RING E3 ligases.
A. Single subunit RING E3:
A list compiled by [47] was used for the analysis of
the major E3 ligase subgroups: _Cbl, Deltex, Goliath,_
_IAP, Listerin, Makorin, MARCH, Neuralized, Pel-_
_lino, Pex, Polycomb, Praja, RBR, Siah, Traf, Trim_
and _Ubr. Notable heterogeneity in expression levels_
was observed among these genes. The most highly
expressed subgroup genes included _Deltex, Goliath,_
_Makorin, March, Neuralized, Praja, Polycomb, Traf_
(Fig. 3) and _Trim genes (Fig._ 4). The expression levels of the other minor gene groups (Cbl, IAP, Listerin,
_Pellino, Pex, Siah, and Ubr) are presented in Table 1._
The TPM values of all the genes in this set were < 37,
except the Ubr7 gene, for which the TPM value was
⁓ 60 on E11 and E13.
_Deltex E3 Ub ligases_ _Dtx3 and Dtx4 were the major_
_deltex genes, with TPM values increasing from 179_
to 351 (Dtx3) and from 40 to 91 (Dtx4) (Fig. 3A),
revealing strong positive regulation during corticogenesis. However, the most highly regulated gene in
this group was _Dtx1: no transcripts were detected_
on E11, but it was clearly strongly induced later, with
TPM values increasing from 9 (on E13) to 60 (on
PN1), a nearly sevenfold increase (Fig. 3A). Deltex E3
has been principally studied in the context of tumorigenesis and tumour cell invasion [48], but very little is known about the roles played by Deltex E3 Ub
ligases in the developing or adult brain in mammals.
The marked enhancement of _deltex gene expression_
supports the notion suggesting key roles in neuronal
growth and differentiation in mammals.
_Goliath E3 Ub ligases Twenty-nine orthologous_
genes of the Drosophila Goliath E3 Ub ligases were
[identified in mice (https://flybase.org/reports/FBgg0](https://flybase.org/reports/FBgg0000104.html)
[000104.html). However, nine of these genes were not](https://flybase.org/reports/FBgg0000104.html)
expressed (Rnf43, _Rnf128 (Grail),_ _Rnf133, Rnf148,_
_Rnf150, Znrf3, Znrf4, Zswim2 and 4930595M18Rik)._
In the expressed gene subgroup, _Rnf44 (68–115_
TPM), _Rnf167 (72–125 TPM), and_ _Rnf126 (46–70_
TPM) were the major genes (Fig. 3B). Notably,
_Rnf215 was the most highly regulated gene, with the_
abundance of its transcripts decreasing from 74 to
27 TPM from E11 to PN1, a 2.7-fold decrease during corticogenesis. The contribution of RNF44 to
brain formation and functions is unknown. The E3
ligase RNF167 plays important roles in neuronal
cells. Although principally found in lysosomes, a
fraction of RNF167 is present at the cell surface,
where it participates in the ubiquitination of AMPA
receptors. Ubiquitination modulates the number of
AMPA receptors at the cell surface as well as synaptic currents. Therefore, RNF167 is an important
physiological modulator of glutamatergic neurotransmission [49]. This RNF167-dependent ubiquitination of AMPA receptors was recently shown to
be mediated by the E2 enzymes Ube2D1 and Ube2N
[50]. RNF126, another prominent factor in this subgroup, has been shown to be involved in Friedreich
ataxia, a severe genetic neurodegenerative disease
characterized by reduced expression of the essential mitochondrial protein frataxin. The E3 Ub ligase
RNF126 specifically mediates frataxin ubiquitination,
which induces its degradation [51]. Our results point
towards a role played by Goliath E3 Ub ligases in
neuronal function from early embryonic stages.
_Makorin E3 Ub ligases All three makorin genes_
(Mkrn1-3) were expressed in the immature cerebral
cortex (Fig. 3A). _Mkrn1 was the predominant gene,_
with TPM values increasing from 32 (on E11) to 116
(on E17). Consistent with our findings, _Mkrn1 had_
been originally identified as a highly expressed gene
during mouse embryonic development, with a high
level of mRNA expression in the developing brain
[52]. Low levels of Mkrn1 proteins were found in the
brain despite its relatively high mRNA abundance
due to the autoubiquitination properties of this E3
Ub ligase, which induces its own proteasomal degradation [53]. Experiments performed with Xenopus
embryos showed that Mkrn2 proteins inhibit neurogenesis by acting downstream of phosphatidylinositol
3-kinase (PI3K) and Akt [54]. Thus, Mkrn proteins
clearly play major roles in the developing nervous
system.
_MARCH E3 Ub ligases The family of proteins of the_
membrane-associated RING-CH (MARCH) comprises eleven E3 Ub ligases (MARCH-1 to -11) [55].
Four March genes of the eleven analysed in this study
were not expressed: _March-1,_ _-3, -10 and_ _-11. The_
major gene in the group was March 9. Its TPM value
was approximately 60 on E13, which was profoundly
increased on E17 (192 TPM) and PN1 (217 TPM),
corresponding to a > 3.5-fold increase (Fig. 3A). On
E17 and PN1, March 9 transcripts represented more
-----
than 50% of all March transcripts. In dendritic cells,
MARCH-9 proteins localize to the trans-Golgi network (TGN) and controls a TGN‐to‐endosome
transport step [56]. In previous studies, MARCH-9
expression had been mainly found in immune cells
and organs such as the lung, lymph nodes, and the
spleen, not neuronal cells [57]. Our data, however,
indicate that MARCH-9 was highly expressed at
the end of neurogenesis. Clearly, additional work
is needed to delineate the neuronal functions of
-----
MARCH-9 proteins in the brain. Although expressed
at a considerably lower level, the other major March
gene was March-5, a mitochondrial-associated E3 Ub
ligase. Its TPM values were 58–62, and showed no
sign of development regulation. Of note, transcripts
of March 4 were detected on E17 and onwards with
TPM values of ⁓ 12 (Fig. 3A). MARCH-4, a Golgiassociated E3 Ub ligase, was the only member of the
MARCH family previously known to be expressed
in the brain [57], but the TPM values were low (less
than 13 TPM) (Fig. 3A). Our transcriptomic analysis revealed a large repertoire of factors, with seven
_March genes expressed throughout corticogenesis_
and both high and highly regulated expression of
_March-9 (Fig. 3A)._
_Neuralized E3 Ub ligases Transcripts of three neu-_
ralized genes were measured: _Neurl1b,_ _Neurl2 and_
_Neurl4. The expression of the major gene Neurl4 was_
enhanced, with TPM values increasing from ⁓ 90 to
⁓ 200 (Fig. 3A). The protein NEURL4, **found in the**
developing rodent cerebellum [58], is a p53-interacting protein that when overexpressed inhibits cellular
growth [59]. Furthermore, previous experiments preformed with NEURL4-knockdown animals showed a
reduced number of presynaptic boutons, indicating
that NEURL4 regulates synapse development in the
brain [58], consistent with its upregulated expression
during corticogenesis.
_Praja E3 Ub ligases The expression of the two Praja_
genes _Pja1 and_ _Pja2 was positively regulated dur-_
ing corticogenesis. The abundance of _Pja1 and_
_Pja2 transcripts increased by factors of 1.6 and 3.8,_
respectively (Fig. 3A). With expression less regulated than that of _Pja2, the_ _Pja1 gene was the pre-_
dominant _Praja gene and one of the most highly_
expressed genes in the Ub system. Its TPM values
were on the order of 270 (on E11) and ⁓ 430 (PN1),
peaking on E17 (⁓ 480) (Fig. 3A). On average, _Pja1_
transcripts were 11- and 5–sixfold more abundant
than _Pja2 transcripts on E11-E13 and E17-PN1,_
respectively. The human and mouse _Pja1 genes are_
highly expressed in the brain, particularly in the cerebral cortex [60]. Northern blot experiments showed
_Pja1 mRNA in the immature brain on E11.5. Sup-_
pression of _Pja1 expression led to a high apoptosis_
rate, indicating that the protein exerts a prosurvival
anti-apoptotic effect. In line with the function of the
encoded protein in cell survival, _Pja1 mRNA has_
been previously found to be overexpressed in twentynine cancer types, with particularly high expression
in gliomas [60]. Altogether, these results support the
idea that Praja1 proteins play important roles in brain
development and regulation of cell apoptosis.
_Polycomb complexes Polycomb-containing complexes_
possess E3 Ub ligase activity due to its RING1A
(Ring1) or RING1B (Rnf2) member. The abundance
of _Rnf2 transcripts did not vary during corticogen-_
esis, whereas a reduction in _Ring1 transcripts was_
observed between E11 and E17, with TPM values
decreasing from 80 to 50 (Fig. 3A).
-----
**Table 1 It gives the list of the genes encoding RING E3 ligases that were found to be weakly expressed during the formation of the**
cerebral cortex
**E3 families** **Genes** **TPM values (mean values)** **Expression**
**pattern**
**E11** **E13** **E17** **PN1**
Cbl _Cbl_ 17.2 18.7 15.2 10.8 ↘
_Cblb_ 4.9 4.5 5.1 3.1 =
_Cblc_ n.d n.d n.d n.d
_Cbll1_ 20.7 23.5 30.2 23.0 ↗
IAP _Birc2_ 23.5 29.7 24.1 17.0 ↗
_Birc3_ n.d n.d n.d n.d
_Xiap_ 18.4 19.9 17.8 16.0 =
_Birc7_ n.d n.d n.d n.d
Listerin _Ltn1_ 12.7 14.8 14.3 12.0 =
Pellino _Peli1_ 11.0 15.6 24.2 14.7 ↗
_Peli2_ n.d 3.1 3.2 3.2 =
_Peli3_ n.d 3.2 13.8 14.5 ↗
Pex _Pex2_ 22.9 22.7 16.6 16.0 ↘
_Pex10_ 17.3 20.6 17.1 11.6 ↘
_Pex12_ 12.4 16.0 19.1 15.3 ↗
Siah _Siah1a_ 14.1 16.5 23.2 22.5 ↗
_Siah1b_ 36.6 31.0 13.0 9.2 ↘
_Siah2_ 5.7 4.8 9.4 11.2 ↗
_Siah3_ n.d 2.7 n.d n.d
Ubr _Ubr1_ 4.4 5.4 7.3 5.6 =
_Ubr2_ 8.0 10.0 12.2 9.9 =
_Ubr3_ 7.9 8.4 11.4 11.8 ↗
_Ubr4_ 19.0 22.4 27.8 27.0 ↗
_Ubr7_ 61.4 58.3 35.2 27.7 ↘
They all displayed TPM values < 65
n.d.: not detected, below the detection threshold. Depending on the gene, the abundance of transcripts increased, decreased or was nearly constant (↗, ↘ and =,
respectively) during corticogenesis
_Traf E3 Ub ligases Transcripts of four_ _Traf genes_
were identified, but their levels varied, with the predominant being _Traf4 with TPM values of 120–180_
(Fig. 3A). In contrast to the other _Traf members,_
_Traf4 expression was increased during corticogen-_
esis. TRAF4 proteins are essential for neural crest
development and neural folding in Xenopus [61]. In
mice, TRAF4 deficiency can induce defects in neural
tube closure [62]. This protein also participates in the
control of myelination [62]. However, transcripts of
two _Traf genes (Traf1 and_ _Traf5) were not detected_
in the present study.
_TRIM E3 Ub ligases Proteins of the tripartite motif_
(TRIM) family are engaged in multiple cellular processing through their E3 Ub ligase activity. Absent
in yeast, TRIM proteins are required for activation
of mammalian autophagy and critical for the regulation of innate immunity [41]. Several families and
subfamilies of TRIM proteins have been identified
(C-I to C-XI), in addition to a group of unclassified
TRIM proteins lacking a RING-finger domain [41].
More than eighty genes were analysed in this study.
Taken together, the data revealed that transcripts of
thirty-two genes encoding RING-finger domaincontaining TRIM and only one TRIM without a
RING-finger domain were found (Fig. 4). Six _Trim_
genes were very highly expressed: _Trim28,_ _Trim32,_
_Trim35, Trim46, Trim59 and Trim67. The latter was_
both the most highly expressed _Trim gene and one_
of the most highly upregulated genes analysed in the
present study. No transcripts were detected before
E13, and the TPM values increased from 12 (on E13)
to 253 (on PN1). Overall, the transcript abundance
increased by a factor of 21 during embryonic development. The most significant increase was noted
between E13 and E17, indicating that TRIM67 is a
dispensable ligase during neurogenesis but is crucial
-----
for postmitotic cell functions and the maturation of
the cerebral cortex (Fig. 4). These data are in line with
a previous report showing that TRIM67 proteins
are highly expressed in the developing and mature
brain but not found in nonneuronal tissues [63].
The TRIM67 protein expression peaked late in the
embryonic and perinatal stages, indicating that it is
involved in neuronal development after the proliferative period. Deletion of the Trim67 gene causes malformations in several brain regions associated with
cognitive and behavioural impairments [63]. The
molecular role played by TRIM67 in brain development as well as the nature of its substrates are, however, unknown.
_Trim35 expression was not regulated to the same_
extent as that of Trim67, but Trim35 was nevertheless
expressed at all ages. _Trim35 TPM values increased_
from ⁓ 170 to 230 from E11 to PN1, reflecting a 35%
augmentation in transcript abundance (Fig. 4). The
third most prominent gene in this family was Trim28.
Its expression was downregulated, with TPM values
decreasing from ⁓ 240 to 120 between E11 and PN1.
The repression of _Trim28 expression was evident_
after E13, indicating that TRIM28 (KAP1 or TIF1b)
exerts important effects during the proliferative
period. TRIM28 is an epigenetic corepressor protein
highly expressed both in the developing and adult
brain [64]. Its absence in mice is embryonically lethal
(on approximately E5.5). TRIM28 has been proposed
to be a SUMO E3 ligase [65]. In murine and human
brains, TRIM28 functions as a transcriptional regulator of neurodevelopmental gene programmes important for brain development [64].
The other main Trim genes were found to be Trim32,
_Trim46, and_ _Trim59. The expression of_ _Trim32_
and _Trim46 was upregulated: the abundance of_
their transcripts increased markedly after E13. For
instance, the TPM values increased by a factor of 2.6
and 8.7 for Trim32 and Trim46, respectively, between
E11 and PN1 (Fig. 4). _Trim46 was one of the most_
induced genes (an ⁓ ninefold increase). Accumulation of TRIM32 proteins into neural cells favours
their commitment to the neuronal lineage [66]. Following its translocation to the nucleus, TRIM32 targets c‐Myc for proteasomal degradation, which initiates neuronal differentiation [66]. TPM values for
_Trim59, another highly regulated gene, decreased by_
a factor of ⁓ 9 (from 149 to 17 TPM) (Fig. 4). These
changes in transcript abundance were primarily identified after the peak of neurogenesis (on E13). The
mRNA levels were much higher during the prolif
erative periods of corticogenesis. TRIM59 proteins
are abundantly expressed in certain organs, such
as the spleen, stomach and ovary, but they are also
found at lower levels in the brain, lung, kidney, muscle and intestine [67]. Again, it is interesting to note
the high expression level of factors known to regulate carcinogenesis. For instance, TRIM28, TRIM32,
and TRIM59, which have been found to be aberrantly overexpressed in certain cancers, were highly
abundant in this study. Specifically, TRIM28 has been
associated with proteins of the melanoma-associated
antigen (MAGE) family and favours the progression of carcinogenesis via suppression of autophagy
[68]. Notably, many E3 Ub ligases, such as MARCH
and TRIM proteins, known for the roles they play
in immune responses, were highly expressed in the
developing cerebral cortex.
B. Multisubunit RING E3 ligases:
Three families of multimeric RING E3 ligases were
considered in the present report: (1) cullin RING
ligases, (2) the APC/C E3 ligase, and (3) the Fanconi
anaemia complex.
B.1 Cullin RING ligases (CRLs)
Cullin RING ligases (CRLs) represent the largest
family of E3 Ub ligases. They are complex molecular entities with several independent subunits.
CRLs (CRL1-9) comprise a cullin (Cul) scaffold
associated with a RING-box protein and an adaptor protein. They also require a substrate recognition element that is an interchangeable subunit
that indicates the target protein to be ubiquitinated. CRL3 is a notable exception, because
the same molecular entity (Broad complex,
Tramtrack, Bric-a-brac, the BTB domain) is both
an adaptor and substrate receptor [39]. Several
cullins, RING-box proteins, adaptor proteins,
and hundreds of substrate recognition proteins
contribute to the generation of a wide range of
combinations giving rise to a multitude of functionally distinct CRLs [69]. Table 2 presents an
overview of the multisubunit structure of CRLs
and their modularity.
B.1.1 Cullin scaffold proteins
Transcripts of nine cullin genes (Cul1-3, _4a,_
_4b, 5, 7, and Cul9 or Parc) were detected, with_
_Cul7 being the predominant member of this_
group. _Cul7 TPM values slightly decreased_
from 103 to 82 from E11 to PN1 (Fig. 5A). The
CUL7 protein, present only in chordates, par
-----
**Table 2 Gives an overview of the multi-subunit structure of CRLs and their modularity**
**Type of CRL** **Cullin scaffold** **RING-finger protein** **Adaptor protein** **Substrate**
**recognition**
**protein**
CRL1 CUL1 ROC1 (Rbx1) Skp1 F-box
CRL2 CUL2 ROC1 (Rbx1) Elongin B/Elongin C VHL-box
CRL3 CUL3 ROC1 (Rbx1) BTB
CRL4A CUL4A ROC1 (Rbx1) DDB1 DCAF
CRL4B CUL4B ROC1 (Rbx1) DDB1 DCAF
CRL5 CUL5 ROC2 (Rbx2) Elongin B/Elongin C SOCS-box
CRL7 CUL7 ROC1 (Rbx1) Skp1 Fbw8
CRL9 CUL9 (PARC) ? ? ?
ticipates in the control of embryonic development. CUL7-knockout mice displayed neonatal lethality, and mutations in the human Cul7
gene were linked to the growth retardation
disorder 3-M syndrome [69]. The expression
level of the human _Cul7 gene is increased_
in glioblastoma tissues compared to normal
brain tissues. Furthermore, _Cul7 facilitates_
the proliferation, invasion and migration of
glioma cells by activating the NF-κB pathway [70]. These effects are consistent with
the current view that CUL7 is an oncogene
[71]. _Cul7 expression is elevated in healthy_
(nontumorigenic) brain tissue. The neuronal
functions of the scaffold protein CUL7 are
not clear but it has been found to be highly
abundant in the developing rat brain [72, 73].
Only two F-box proteins are known to interact with CUL7: FBXW8 and FBXW11 [71];
the gene expression of these proteins showed
an opposite pattern: the abundance of Fbxw8
transcripts decreased from 16 to 10 TPM,
whereas the abundance of Fbxw11 transcripts
increased from 12 to 19 TPM from E11 to
PN1. CUL7 is, together with the F-box protein FBXW8, associated with the Golgi apparatus in neuronal cells and is required for the
growth of dendrites (but not axons) in neurons in the mammalian brain [72]. CUL7 is
also found at synaptic sites, controlling the
degradation of Eag1, a potassium channel in
the plasma membrane that participates in
the regulation of membrane excitability [73].
**Fig. 5** Expression of genes encoding cullin scalfolds, adaptor Due to its high expression level and synapproteins and RING finger proteins. Transcripts of cullin (Cul) (A), tic localization, CUL7 may be an important
adaptor proteins (B) and RINGER proteins (C) genes are shown. D, modulator of neuronal excitability in the
**E and F show the pattern of expression of the major Fbxw, Fbxl and** brain. Within the CRL7 E3 ligase complex,
_Fbxo genes. They all had TPM values ≥ 40. The minor Fbxw, Fbxl and_ the scaffold protein CUL7 is associated with
_Fbxo genes are shown in Additional file 1: Fig. S2A–C_
-----
the adaptor protein Skp1 and the RING finger
protein ROC1, which contains a E2 enzymebinding domain. The expression of the _Skp1_
and Rbx1 genes is discussed below. The other
_Cul genes showed low TPM values (approxi-_
mately 10–20 TPM). The expression of the
minor _Cul9 gene was developmentally regu-_
lated: its transcript abundance was increased
by a factor of ⁓ eight between E11 and PN1
(from ⁓ 2 to 19 TPM) (Fig. 5A). It was one of
the most upregulated genes analysed in this
study.
B.1.2 Adaptor proteins
Adaptor proteins are attached to the cullin
scaffold. The following four genes were identified in this study: Skp1 (Skp1a), Elob (elongin
_B), Eloc (elongin C) and Ddb1. Three of these_
genes were highly expressed with TPM values ≥ 180: _Skp1a,_ _Elob, and_ _Ddb1 (Fig._ 5B).
They displayed distinct patterns of expression: _Elob was expressed at a constant level,_
whereas _Ddb1 expression was negatively_
regulated (Fig. 5B). Transcripts of _Ddb1, the_
major gene in this subgroup, were reduced
drastically during embryonic development,
with the TPM value decreasing from 351 to
162 (Fig. 5B).
B.1.3 RING finger proteins
The RING finger E3 ligases function as
docking sites for E2 enzymes. The TPM values of the _Rbx1 and_ _Rbx2 (Rnf7) genes were_
nearly identical on E11 (89 and 86, respectively) (Fig. 5C). These genes displayed differing expression patterns: the abundance of
_Rbx1 transcripts increased (with a peak on_
E17 with 109 TPM), but that of _Rbx2 was_
decreased (with 43 TPM on PN1). Overall,
_Rbx1 was the predominant RING finger pro-_
tein-encoding gene in the cerebral cortex during development (Fig. 5C).
B.1.4 Substrate receptors
In this chapter, we have analysed the expression of genes encoding F-box proteins. They
are important components of the Skp1-Cullin 1-F-box complex (SCF E3 Ub ligases).
F-box proteins play critical roles as substrate
receptors and have been classified into three
groups: FBXW, FBXL and FBXO F-box proteins [74]. The analysis of the genes (Fbxw,
_Fbxl and_ _Fbxo) followed this classification._
The most highly expressed genes (with TPM
values ≥ 40) are shown in Fig. 5D–F. The
expression of the minor F-box genes (TPM
values < 40) is presented in Additional file 1:
Fig. S2.
_Fbxw genes: The major genes of this subgroup_
were _Fbxw2,_ _Fbxw5 and_ _Fbxw9 (Fig._ 5D).
They were all upregulated during corticogenesis, particularly _Fbxw9 with TPM val-_
ues increasing ⁓ fourfold from 22 (on E11)
to 93 TPM (on PN1). _Fbxw5 was the most_
highly expressed _Fbxw gene. The TPM val-_
ues increased from 98 to 193, a ⁓ twofold
increase in transcript abundance (Fig. 5D).
This is another example of a member of the
Ub system known for its contribution in tumorigenesis [75] that is highly expressed in a
nontumorous brain tissue. However, the roles
of FBXW5 proteins in the brain are unknown.
The minor Fbxw genes displayed TPM values
ranging from 2 to 20. Similar to major _Fbxw_
genes, the expression of the minor _Fbxw_
genes was positively regulated during corticogenesis except _Fbxw8 (Additional file_ 1:
Fig. S2A). Of note, two Fbxw genes were not
expressed (Fbxw10 and Fbxw12).
_Fbxl genes: The major Fbxl genes were Fbxl6,_
_Fbxl16, Fbxl19 associated with the predomi-_
nant _Fbxl14 gene (Fig._ 5E). This gene had
elevated TPM values that decreased during
corticogenesis (from 202 to 98 TPM). The
proteins FBXL14 has been reported to associate with Hes1 (hairy and enhancer of split
1), a repressor of proneural genes. Furthermore, the loss or overexpression of FBXL14,
respectively, stabilizes Hes1 or decreases its
protein levels [76]. In stem cells, FBXL14 controls the proteasomal degradation of Hes1,
which favours neuronal differentiation [76].
The temporal pattern of expression of the
_Hes1 gene is reported in Fig. 9B. Both genes,_
_Fbxl14 and Hes1, were downregulated during_
corticogenesis. The other major _Fbxl genes_
had however a different pattern of expression
with an abundance of transcripts increasing significantly during embryogenesis. The
expression of the _Fbxl16 gene was strongly_
induced with TPM values increasing from 19
(on E13) to 130 (on PN1), a nearly sevenfold
increase (Fig. 5E). The temporal pattern of
expression of the minor Fbxl genes is shown
in Additional file 1: Fig. S2B. No transcripts
of the following six Fbxl genes (Fbxl4, Fbxl7,
_Fbxl8, Fbxl13, Fbxl17, and Fbxl21) were_
detected.
-----
_Fbxo genes: Amongst the six genes of this_
group that had TPM values > 40: _Ccnf_
(Fbxo1), _Fbxo5, 21, 41, 44, and_ _45, two_ _Fbxo_
genes predominated: _Fbxo5 and_ _Fbxo21_
(Fig. 5F). They had however differing expression patterns: the expression of _Fbxo5 was_
strongly reduced (TPM decreasing from 84
to 5, a 17-fold reduction in transcript abundance) while the expression of _Fbxo21 was_
induced during corticogenesis (TPM values
increasing from 44 to 79, a 1.8-fold increase)
(Fig. 5F). Although expressed at lower levels
(TPM values ranging from 46 to 3), Ccnf was
also strongly downregulated. The abundance
of _Ccnf transcripts decreased nearly 15-fold_
from E11 to PN1 (Fig. 5F). _Ccnf and_ _Fbxo5_
seemed to play critical roles at the onset
of neurogenesis. Fbxo5 proteins have been
shown to control cell proliferation [77]. The
expression of _Fbxo41 was strongly induced:_
no transcripts were detected at E11 but the
abundance of transcripts increased from 4
TPM (on E13) to 45 TPM (on PN1) (Fig. 5F),
a 11-fold increase. Fbxo45 proteins are found
exclusively in the brain [78, 79] and Fbxo45
mRNA is detected as early as E12 [79]. The
loss of Fbxo45 is postnatally lethal and is
associated with an abnormal embryonic
neural development [78]. Fbxo45 proteins
play important roles in the brain by regulating neurotransmission [79]. The Fbxo45 gene
expression was significantly enhanced at the
end of neurogenesis (Fig. 5F). It is however
important to note that Fbxo45 fails to associate with Cul1 and does not form an SCF complex but associates with a RING finger-type
Ub ligase [80]. The expression of the minor
_Fbxo genes is shown in Additional file 1: Fig._
S2C. The number of transcripts of seven Fbxo
genes were below the detection threshold
(Fbxo15, Fbxo16, Fbxo24, Fbxo36, Fbxo39,
_Fbxo40, and Fbxo23)._
B.2. The anaphase-promoting complex/
cyclosome (APC/C) E3 ligase
The E3 Ub ligase anaphase-promoting complex/cyclosome (APC/C) is well known for its
control of the cell cycle because it regulates
mitotic progression and exit. It is highly abundant in postmitotic neurons, where it plays
a role in dendrite and axon arborization and
in synaptogenesis [9]. The APC/C E3 ligase
is a multi-subunit complex displaying a similar structure to the Skp1/Cul1/F-box protein
Ub ligases. Both are composed of three fixed
subunits (a catalytic RING protein, a scaffold
protein and an adaptor protein) and another
component conferring substrate specificity (an
F-box protein for SCF, and Cdh1 or Cdc20 for
APC/C) [81].
The three sub-complexes of the APC/C ligase
consist in a catalytic core (with APC2, APC10,
or APC11), a scaffolding platform (APC1,
APC4, APC5, or APC15), and a substrate recognition module (or tetratricopeptide repeat
lobe, TPR) (consisting of APC3, APC6, APC7,
APC8, APC12, APC13, of APC16). In addition, CDC20 and CDH1 are coactivators (also
considered to be substrate receptors) essential
for the activity of an APC/C ligase [82, 83]. The
expression of fourteen APC-encoding genes
(Anapc) and two coactivator-encoding genes
(Cdc20 and Cdh1) was analysed.
_Anapc2 (100–115 TPM) and_ _Anapc11 (⁓ 49_
TPM) were the major genes of the catalytic
core. Transcripts of the other component
(Anapc10) were near the detection level (2–3
TPM) (Fig. 6A). With TPM values ranging
from 128 to 107 from E11 to PN1, _Anapc5_
was the predominant gene in the scaffolding
platform. Anapc1 and Anapc4 presented comparable low levels of expression (TPM values
of 20–40), whereas _Anapc15 was expressed_
at even lower levels (TPM values of 6–12)
(Fig. 6A). Anapc6 (Cdc16) and Anapc8 (Cdc23)
were the most highly expressed genes in the
substrate recognition module (Fig. 6A). Taken
together, the transcripts of this subgroup had
low or moderate abundance, with TPM values ranging from 12 to 64 TPM. As shown in
Fig. 6A, the expression of genes in the APC/C
subgroup displayed nearly constant transcript
numbers throughout cortical formation, suggesting their basic functions in cell physiology.
Notably, the extremely elevated expression of
the Cdc20 gene at early stages of cortical development was followed by a sharp decrease on
E17. Its TPM values declined from ⁓ 350–310
on E11–E13 to 44–18 on E17-PN1. Overall, the
abundance of _Cdc20 transcripts was reduced_
by a factor ⁓ 20, showing that this gene exhibited a marked temporal pattern of expression.
The high abundance of Cdc20 transcripts cor
-----
responded to periods of cell production (E11E13), strongly suggesting a role for _Cdc20 in_
cell proliferation. The expression of the other
coactivator-encoding gene _Cdh1 (Fzr1) was_
not developmentally regulated. Constant levels
of Cdh1 transcripts were found during embryonic development (TPM values of 89–90).
Chd1 proteins are required for neurogenesis
in vivo [84]. It seemed, however, that the regulation of the coactivator gene _Cdc20 expres-_
sion was a central determinant affecting the
functionality of the APC/C E3 Ub ligase in the
embryonic cerebral cortex. The APC/C E3 Ub
ligase ubiquitinates its substrates in conjunction with a limited set of E2 Ub-conjugating
enzymes: UBE2S, UBCH10 (UBE2C) and, to
a lesser extent, UBCH5 (UBE2D1) [82, 83]. As
shown in Fig. 1D, _Ube2c was the most highly_
expressed of these three E2 enzymes. Interest
ingly, _Ube2c and_ _Cdc20 displayed similar pat-_
terns of expression (Fig. 6B). The decline in
_Cdc20 expression mirrored the marked repres-_
sion of Ube2c expression.
B.3. Fanconi anaemia (FA) E3 ligases
The classification of the components of the
Fanconi anaemia (FA) complex was established according to [85] and the Fanconi anaemia mutation database [https://www2.rocke](https://www2.rockefeller.edu/fanconi/)
[feller.edu/fanconi/). The FA complex is com-](https://www2.rockefeller.edu/fanconi/)
monly described as a machine recruited to
DNA lesions and playing a role in DNA repair.
FANCL is the only protein of the FA complex
displaying a ligase activity. However, no transcripts of its gene (Fancl) were found, suggesting poor or no activity under physiological
developmental conditions. Of note, Ube2T, the
E2 working in concert with FA E3 ligases, was
also expressed at very low levels with TPM values decreasing from 17 to 4 between E11 and
E17, and no _Ube2T transcripts were detected_
on PN1 (Fig. 1D).
C. U-box RING E3 ligases:
U-box RING E3 enzymes form another prominent class of E3 Ub ligases. They are characterized by a peculiar protein domain named the
U-box and are structurally related to the RING
finger family [86, 87]. U-box E3 Ub ligases are
scaffolds that recruit a Ub-charged E2 and
its colocalized substrate. Interestingly, mammalian U-box E3 Ub proteins interact with
molecular chaperones or cochaperones such
as Hsp90, Hsp70, DnaJc7, EKN1, CRN, and
VCP [88]. U-box E3 Ub ligases can be found as
monomers (i.e., UBE4) or homodimers (CHIP
and PRPF19) [89]. Some U-box E3 Ub proteins
have been identified as E4 enzymes due to
their involvement in the assembly of poly-Ub
chains on substrates that are first ubiquitinated
by a non-U-box E3 Ub enzyme.
Nine genes were analysed: Stub1 (Chip), Prpf19
_(Prp19), Ube4a (Ufd2b), Ube4b (Ufd2a), Ppil2_
_(Cyc4), Ubox5 (Uip5), Wdsub1, Act1 (Traf3ip2)_
and _Aff4 (Fig._ 7). Except for a transcript of
_Act1, transcripts of all U-box E3 Ub-encoding_
genes were found. No clear developmental
regulation pattern was observed for the U-box
ligases except for _Prpf19 (Prp19), the second_
most highly expressed U-box E3 Ub gene.
-----
The _Prpf19 TPM values decreased from 250_
to 180 from E11 to PN1, a ⁓ 30% reduction in
transcript abundance during corticogenesis.
The functions of the _Prpf19 gene product are_
unknown, but it is an essential protein since
mouse Prpf19-null mutants show lethality [90].
With TPM values of 300–350, the Stub1 (Chip)
gene was the predominant gene in this group
and one of the most highly expressed E3 Ub
ligase genes. This high expression highlights
its physiological relevance during brain formation and development. The U-box E3 Ub ligase
CHIP can tag misfolded or damaged proteins
for subsequent proteasomal degradation. A
previously performed proteomic analysis identified hundreds of potential CHIP substrates in
HEK 293 cells [91]. The very high level of Stub1
_(Chip) expression underscores the physiologi-_
cal importance of CHIP during the protein
quality control process and clearance of abnormal proteins throughout embryonic development.
The UFD2a protein (coded by _Ube4b/Ufd2a)_
has been found to be highly abundant in some
brain areas, such as the cerebrum and cerebellum, of 8-week-old C57Bl6 mice [92]. Furthermore, immunohistochemical data have indicated that in the cerebral cortex, the UFD2a
protein is localized mainly in the cytoplasm
of neurons. Kaneko et al. [92] proposed that
UFD2a contributes to the ubiquitination of
specific substrates related to neuronal function. The high abundance of UFD2a proteins
previously observed in the adult mouse brain
differs from the low mRNA abundance of
_Ube4b (Ufd2a) transcripts in the embryonic_
brain.
_RBR Ub ligases The RBR Ub ligase family includes_
a few E3 Ub ligases. Fourteen RBR E3 genes were analysed (Arih1, Arih2, Ankib1, Park2, Rbck1, Rnf14, Rnf19a,
_Rnf19b, Rnf31, Rnf144a, Rnf144b, Rnf216, and_ _Rnf217)_
(Fig. 8). The number of _Park2 and_ _Rnf144b genes tran-_
scripts was lower than the detection limit. Notably, Park2
encodes Parkin, a protein controlling mitophagy via the
ubiquitination of mitochondrial proteins. Mitochondria
can, however, be recycled via a Ub-independent pathway involving the specific autophagy receptors FUNDC1,
BNIP3, and NIX [93]. Interestingly, the RNA-seq dataset indicated that, albeit sometimes at low levels, the
_Fundc1, Bnip3 and Nix (Bnip3l) genes were all expressed_
with TPM values of ⁓ 10 (Fundc1), ⁓ 11 (Bnip3), and ⁓ 50
(Nix/Bnip3l). Based on these results, it is proposed that
in the embryonic cerebral cortex, mitophagy is initiated independently of Parkin but via a Ub-independent
process. As previously pointed out by [94], most of the
data describing the regulation of mitophagy have been
obtained with cells overexpressing Parkin and through
the use of mitochondrial-depolarizing agents, which may
not be relevant under the basal conditions of mitochondrial clearance.
_Rnf14 (ring finger protein 14, also known as Triad2),_
encoding transcriptional regulator RNF14, was the major
RBR E3 gene expressed during cortex development. Its
expression was positively regulated, with TPM values
increasing from 27 to 113 from E11 to PN1 (Fig. 8A), representing a nearly > fourfold increase in Rnf14 transcript
abundance. RNF14 is an oncoprotein that promotes cell
cycle progression and proliferation by inducing cyclin
D1 expression [95]. In the developing (nontumorigenic)
cerebral cortex, the expression of the cyclin D1 gene
(Ccnd1) decreased from 333 to 32 TPM from E11 to PN1,
a ⁓ tenfold reduction in _Ccnd1 transcript abundance._
This decrease revealed an inverse relationship between
the expression of the Ccnd1 and Rnf14 genes that contradicts the suggestion that RNF14 exerts a positive effect
-----
on cyclin D1 expression, at least under physiological conditions. The biological functions of RNF14 in the brain
are unknown; however, our data indicated that RNF14
likely plays important functions in the cerebral cortex,
particularly after the cessation of cell production (on E17
and onwards), during the maturation of neurons and the
establishment of synaptic networks.
The other major RBR E3 gene in the embryonic cerebral cortex, although expressed at low levels, was Rbck1,
encoding RanBP-type and C3HC4-type zinc finger containing 1 (HOIL-1 or HOIL-1L). Its TPM values were on
the order of 45 at the onset of development and 60 at the
latest stage, indicating a modest upregulation during corticogenesis (Fig. 8B). A recent report showed that RBCK1
plays a role in the linear ubiquitin assembly complex
(LUBAC) [96] comprising the adaptor protein SHARPIN
and two RBR E3 ligases: HOIP and HOIL-1L. HOIP is the
main E3 catalytic centre of LUBAC and is necessary for
linear ubiquitination. The gene encoding HOIP (Rnf31)
was expressed at low levels (TPM values of 18–27)
(Fig. 8B). HOIL-1L, which is the second most active
and minor ligase of LUBAC, exerts a regulatory role in
the complex by negatively regulating HOIL activity [96].
LUBAC is recruited to different protein aggregates associated with neurodegenerative diseases. LUBAC-dependent linear ubiquitination decreases the toxic potential of
misfolded protein species and promotes their removal via
the proteasome [97]. The linear ubiquitination catalysed
by HOIP is antagonized by the DUB OTULIN [97]. Low
HOIP and HOIL-1L levels in mice cause early embryonic
lethality (on approximately E10.5) [98]. The other RBR E3
genes were also expressed at low levels, and their patterns
of expression were not found to be developmentally regulated, except for _Rnf19b, whose number of transcripts_
increased by a factor of ⁓5 between E11 and PN1, highlighting a putative function in neuronal development that
remains to be discovered (Fig. 8B).
**Deubiquitinating enzymes (DUBs)**
This section covers the gene expression of seven families
of DUBs: USPs, UCHs, OTUs, MJDs, JAMMs, MINDYs
and ZUP1 [3, 4]. DUBs are Ub hydrolases responsible of
the deubiquitination process. Additionally, some of them
such as USP5, UCH-L3, USP9X, USP7, and Otulin participate in the processing of Ub precursors [99].
**_Ub‑specific proteases (USPs)_**
With approximately fifty members, USPs constitute the
largest subfamily of DUBs [4]. In a recent report analysing the expression of _Usp genes in the rat cerebellar_
cortex, only 32 USP-encoding genes were retained for
analysis [100]. In the present study, we first selected fiftyfour genes for analysis, but five of these genes (Usp50, 39,
_53, 54 and_ _Pan2) showed no catalytic activity [18] and_
were therefore not analysed further. Ultimately, fortynine genes were analysed. The transcripts of forty-five
_Usp genes were quantified, implying that four transcripts_
of _Usp genes were undetected (Usp13, Usp17, Usp18,_
and Usp26). Large heterogeneity in gene expression was
observed. For the sake of clarity, the five most highly
expressed Usp genes are shown as a group in Fig. 9A, and
the other members, with much lower transcript abundance, are shown in Additional file 1: Fig. S3A-B. Of
note, USP9X was shown to play roles in neurodevelopment. However, a moderate level of gene expression was
observed with TPM values ranging from 20 to 30 (Additional file 1: Fig. S3A). Furthermore, [101] found an agerelated up-regulation of USP9X protein expression in the
mouse brain with much higher protein levels in the adult
brain, suggesting that USP9X could play important roles
postnatally rather than during embryonic development.
With TPM values increasing from ⁓ 115 to ⁓ 390
between E11 and PN1, _Usp22 was the major gene and_
the most highly upregulated _Usp gene (Fig._ 9A). It was
also one of the five most highly expressed DUB genes
throughout corticogenesis. Its expression was developmentally regulated and showed a marked increase on
E17. At this point, Usp22 was the most highly expressed
_Usp gene and the second most highly expressed DUB_
gene after Uch-l1 (see below). Our data were in line with
previous reports showing high abundance of USP22 in
-----
the mouse embryonic brain [102, 103] and further indicating the specific expression of its gene in the cortex.
USP22 proteins are critically required for embryogenesis since their loss leads to early embryonic lethality (at
approximately E10.5) [103, 104]. USP22 proteins interfere with SOX2 and Hes1 activity, as well as that of other
targets. _Sox2 is a pluripotency gene, and_ _Hes1 represses_
the expression of proneural genes, contributing to the
regulated maintenance of neural/stem progenitor cells.
In embryonic stem cells, an inverse correlation has been
identified between SOX2 and USP22 protein levels [105].
Specifically, USP22 occupies the _Sox2 promoter and_
represses _Sox2 transcription [105]. Hes1, which under-_
goes a fast turnover rate due to its degradation by the
proteasome, is deubiquitinated and stabilized by USP22
[102]. We therefore measured the expression of the Sox2
and Hes1 genes to gain further understanding of USP22dependent regulatory mechanisms during corticogenesis.
Interestingly, both genes were highly expressed when
the Usp22 expression was the lowest. In contrast, a profound increase in _Usp22 expression coincided with a_
marked reduction in Sox2 and Hes1 transcript abundance
(Fig. 9B). Thus, a clear inverse correlation between Usp22
(which promotes neuronal differentiation) and Sox2 and
_Hes1 (genes necessary for the maintenance of neural/_
stem progenitor cells) was identified.
_Usp1 was another highly expressed_ _Usp gene for_
which a high abundance of transcripts was found on
E11 and E13 (⁓ 160–130 TPM), corresponding to periods of intense cell division. _Usp1 was the second most_
highly expressed DUB gene on E11. The TPM values
were ⁓ 40–30 on E17 and PN1 (Fig. 9A). This decrease in
_Usp1 expression at E17 indicated that the gene may play_
a role in proliferation but not in the growth or maturation of neurons. In osteosarcoma cells, USP1 knockdown
triggers osteogenic differentiation, whereas USP1 overexpression enhances proliferation, suggesting that, in
this cell type, USP1 is involved in the maintenance of a
stem cell state [106]. A similar finding was observed with
glioblastoma cells [107]. The transcription of Usp1 is regulated in a cell cycle-dependent manner, with transcription peaking during the S phase. The transcriptomic data
clearly support the notion that _Usp1 is highly regulated_
during embryonic cortical development, showing high
mRNA expression levels during stages of cell division and
neurogenesis. Deletion of the Usp1 gene has been associated with 80% perinatal lethality, and the surviving Usp1deficient mice exhibited growth retardation [108].
In addition to _Usp1 and_ _Usp22,_ _Usp5,_ _Usp19 and_
_Usp21 were the other major Usp genes expressed in the_
embryonic cortical wall (Fig. 9A). However, these genes
displayed no clear pattern of developmental expression. Similar to most DUBs, USP19 is a soluble cytosolic
protein, but one prominent USP19 isoform possesses a
C-terminal transmembrane domain, which enables its
translocation to the endoplasmic reticulum (ER). USP19
seems to function in ER-associated degradation (ERAD).
ER stress induction upregulates USP19 expression, and
its biological relevance has been studied in muscle cells,
where it plays a role in metabolic regulation and controls
muscle mass [109]. Little is known regarding the neurobiological functions of Usp19 and Usp21. The Usp21 gene
was one of the most highly _Usp genes expressed during_
corticogenesis. Previously experiments conducted with
embryonic stem cells showed that USP21 proteins control the balance between stem cell self-renewal and differentiation [110].
Notably, Usp5 is continually highly expressed throughout embryonic development. Its protein product USP5, is
primarily located in the cytosol and nucleoplasm, where
it recognizes poly-Ub chains not conjugated to target
proteins and contributes to maintaining the pool of free
Ub monomers by removing Ub from the proximal end of
these unanchored chains [111]. USP5, which is an important contributor of Ub precursors processing [99], has
been studied extensively in relation to cancer, but this
DUB is widely expressed. For example, USP5 has been
shown to play a role in inflammatory and neuropathic
pain by regulating the cell surface abundance of the Cav3.2
protein, a T-type voltage-gated Ca[2][+] channel that plays
important roles in nociception [112]. USP5 counterbalances the action of the E3 Ub ligase WWP1 [112]. Our
data emphasize that certain components of the Ub system that are generally associated with cancers, such as
USP5, are also highly expressed during development in
nontumorous tissue.
As shown in Additional file 1: Fig S3A-B, many _Usp_
genes were expressed in the cortical tissue throughout embryonic cortical development. For 10 _Usp genes_
(Usp4, 7, 9, 10, 14, 24, 28, 30, 36, and 38), the abundance
of transcripts was nearly constant throughout corticogenesis, indicating that their expression was not developmentally regulated. In addition to Usp1, the expression
of twelve genes was downregulated: _Usp3, 8, 21, 25, 37,_
_39, 40, 44, 45, 49, 51 and 54 (Additional file 1: Fig. S3A-_
B). Notably, _Usp25 and_ _Usp44 were only expressed at_
the beginning of corticogenesis. The expression of other
_Usp genes was upregulated, although to moderate levels._
Interestingly, the expression of 4 Usp genes was induced
at the end of corticogenesis (Usp2, 29, 43, and 53), with
no transcripts detected before E17. This finding points
to a potential role of these DUBs in neuronal growth and
the establishment of neural circuits, whereas the _Usp25_
and _Usp44 gene products exert their biological effects_
during the neurogenesis period.
-----
**_Ub carboxyl‑terminal hydrolases (UCHs)_**
Four _Uch genes,_ _Uch-l1, Uch-l3, Uch-l5 and_ _Bap1_
_(BRCA1-associated protein 1), were expressed during_
corticogenesis, although the levels were considerably different (Fig. 9C). With TPM values ranging from ⁓155 (on
E11) to ⁓550 (on PN1), _Uch-l1 clearly showed the high-_
est expression levels in this subfamily, at least in the two
latest stages. Its expression was markedly upregulated,
with an abundance of transcripts increasing by a factor
of 3.3 and 3.5 on E17 and PN1, respectively, compared
to the abundance on E11. This observation is in line with
the fact that UCH-L1 (also named PGP 9.5) is one of the
most abundant brain proteins, representing up to 1–5%
of total soluble brain proteins [113]. Isolated from brain
extracts, UCH-L1 was originally described as a neuronal
marker [113]. In a previous study on the brain, _Uch-l1_
mRNA was detected in early stages of embryonic development [114] and was found in progenitor cells and
neurons [115]. UCH-L1 has been postulated to facilitate
neurogenesis and determine the morphology of progenitor cells [115]. The precise roles UCH-L1 plays in neuronal physiology are, however, poorly understood, but
UCH-L1 dysfunction has been associated with several
age-related neurodegenerative processes, such as Alzheimer’s and Parkinson’s diseases [116].
A high and constant abundance of _Bap1 transcripts_
was observed (TPM values of ⁓ 200) with no evidence
of developmental regulation. _Bap1 was the most highly_
expressed _Uch gene on E11-E13, whereas_ _Uch-l1 was_
the major gene expressed at the end of corticogenesis
(E17-PN1) (Fig. 9C). The protein BAP1 was originally
described as a nuclear DUB that exhibited tumour-suppressing properties. It regulates transcription and the
DNA repair response. Additionally, BAP1 modulates
intracellular Ca[2][+] signalling by deubiquitinating (and
stabilizing) inositol 1,4,5-trisphosphate (IP3) receptors,
prominent Ca[2][+] release channels in the ER [117]. Hence,
BAP1 displays a basal prosurvival function by inhibiting
the unfolded protein response induced by glucose deprivation [118]. Our data, together with results found in
the literature, suggest that the protein product of _Bap1,_
a highly expressed gene, plays important roles during the
production, survival and differentiation of neural cells in
the cortical wall during embryonic development.
**_MJD_**
The TPM values of the MJD genes Atxn3, Josd1 and Josd2
[4, 119] ranged from 4 to 60, revealing low to moderate
transcript abundance (Fig. 9D). Josd2 was the main MJD
gene. Its expression was repressed throughout cortical
development.
**_Otubain proteases (OTUs)_**
The analysis encompassed fifteen genes, and for two
of them, no transcript was found: _Otud6a and_ _Otud7a_
(Cezanne2). In addition to the _A20 gene (Tnfaip3) that_
was expressed exclusively (and at low level) on E17 and
PN1, all the other _Otu genes were expressed at low and_
moderate levels on the order of 5 to 30 TPM at all time
points with no clear pattern of developmental regulation
(Fig. 10A), except for _Otud1. Its transcript abundance_
increased by nearly sixfold from E11 to PN1 (TPM values
ranging from 4 to 23) (Fig. 10A). The gene Otub1 was the
only member of this family showing relatively high levels of expression (TPM values of 120–140). OTUB1 has
been described as one of the most abundant DUBs in
cells with ubiquitous tissue expression [18]. To date, the
neuronal functions of _Otub1 have been poorly charac-_
terized. OTUB1 is found in the brain and is expressed in
neurons but not in microglia or astrocytes [120]. OTUB1
attenuates the apoptosis of neuronal cells after intracerebral haemorrhage [120]. Hence, it is coenriched with
α-synuclein [121], the major component of Lewy bodies,
which constitute a hallmark of Parkinson’s disease. The
pathogenicity of OTUB1 has been underscored by [122],
who showed that OTUB1 is an amyloidogenic protein
that could contribute to the development of Parkinson’s
-----
disease. Notably, Otud1, although expressed at low levels
(TPM values ranging from 4 to 23), was the most highly
regulated gene of this subfamily; the abundance of Otud1
transcripts increased nearly sixfold from E11 to PN1
(Fig. 10A).
**_Machado‑Joseph disease protein domain proteases (JaMMs_**
**_or Josephins)_**
First, we focused our analysis on the following seven
JaMM genes: _Cops5 (Csn5), Psmd14, Brcc3, Mpnd,_
_Mysm1, Stambp (Amsh), and Stambpl1, which all encode_
DUBs with enzymatic activity. These genes were all
expressed, except Stambpl1 (Fig. 10B). Cops5 and Mpnd
were the major genes in this group, with TPM values of
⁓ 60 and ⁓ 70–90, respectively. For the other members,
namely, _Psmd14, Brcc3, Mysm1, and_ _Stambp (Amsh),_
a low but continuous abundance of transcripts was
observed (< 15 TPM) (Fig. 10B); however, among these
genes, _Stambp (Amsh) was found to be strongly and_
positively regulated, as indicated by a fourfold increase
in transcript abundance between E11 and PN1 (from 3
to 14 TPM). Many JaMMs fail to display catalytic activity and are thus classified as pseudoenzymes [4]. The following pseudoenzyme genes were selected for analysis:
_Cops6 (Csn6), Eif3f, Eif3h, Prpf8, and_ _Psmd7. As shown_
in Fig. 10C, they were all highly expressed. For instance,
the TPM values of the 3 major genes _Cops6, Eif3f, and_
_Prpf8 were on the order of 190–200. These transcripts_
were thus 3- to tenfold more abundant than the transcripts of the JaMMs genes Cops5, Psmd14, Brcc3, Mpnd,
_Mysm1, or Stambp. This finding indicates that the protein_
products of the _Cops6, Eif3f, and_ _Prpf8 genes may exert_
important nonenzymatic biological functions in cells of
the cortical wall during embryonic development.
**_Motif‑interacting with Ub‑containing novel DUB family_**
**_(MINDY)_**
MINDY is a family of DUBs with four members:
FAM63A (MINDY-1), FAM63B (MINDY-2), FAM188A
(MINDY-3) and FAM188B (MINDY-4) [123]. The expression of these genes was investigated (Fig. 10D), and the
TMP values were found to be on the order of ⁓ 2–10 for
_Fam63b and Fam188a and ⁓ 20–30 for Fam63a. No tran-_
script of the Fam188b gene was found. Compared to that
of the other DUB families, the lowest abundance of transcripts was found in this group of genes. The neuronal
functions of members of the MINDY family are currently
unknown.
**_ZUP1_**
ZUP1 (or ZUFSP, zinc finger with UFM1-specific peptidase domain) was identified as a seventh family of human
DUB [124]. The murine _Zufsp gene was expressed at_
extremely low levels (TPM values of 2–4, not shown).
Nothing is known about the biological roles played by
_Zufsp in the rodent brain. In humans, the protein ZUFSP,_
which is mainly localized in the nucleus, is thought to be
a putative DNA repair and/or replication factor involved
in Ub signalling at DNA lesions [124].
**Conclusions**
The contribution of the Ub system has been studied
using various lines of embryonic stem cells and their differentiation into neural precursor cells. In contrast, in
this study, no cell lines were employed, and data were
extracted from an RNA-seq database [13], allowing us to
detect transcriptomic changes in the core components of
the Ub system during the formation of the cerebral cortex
in mice. This strategy permitted us to describe the transcriptomic landscape of the whole tissue. This approach
also revealed the large repertoire of functional components of the Ub system in embryogenesis. One important result indicated that the expression of Ub genes,
notably Ubb and Rps27a, was extremely high. These two
genes were among the 100 most highly expressed genes
of the cortical wall. Our findings illustrate that the intricate ubiquitination network was governed by the E1 gene
_Uba1, which was more highly expressed, from 20- to_
90-fold, than the other E1 gene _Uba6. The most promi-_
nent E2 gene was _Ube2m, encoding a Nedd8-conjugat-_
ing E2 enzyme. The major Ub-conjugating E2 gene was
_Ube2c, the expression of which was profoundly down-_
regulated during embryonic development. A large diversity of E3 Ub ligase gene transcripts was detected with
distinct temporal patterns of expression. _Pja1, Trim67_
(RING E3-encoding genes), _Stub1 (U-box E3-encoding_
gene), and _Nedd4 (HECT E3-encoding gene) were the_
most prominent E3 genes. A previous report analysed the
expression of thirty DUB-encoding genes in the rat cerebellum, out of approximately one hundred DUB genes
[100], and thirty DUBs have also been independently
described as being involved in the nervous system [119].
In this study, an extensive genome-wide gene expression
analysis of the core components of the Ub machinery
showed that more than 80 DUB genes were expressed
during the formation of the cerebral cortex. This outcome provides a comprehensive survey of the large diversity of DUB gene expression and further indicates some
important candidate products that may play major roles
in cortex development. For instance, _Uch-l1 was one of_
the most highly expressed genes. It was also positively
regulated during corticogenesis.
This study was based on a bulk transcriptomic analysis
that did not discriminate between cell type or the cell lineages within the whole tissue sample. Moreover, certain
cells, such as neurons, are highly polarized with several
-----
subcellular compartments (i.e., dendrites, cell body,
axon) with distinct biological functions. Cellular polarity
requires precise spatial targeting of the factors participating in the Ub pathway. In most (if not all) instances,
the mechanisms governing the spatial targeting of Ub
components are unknown. Despite these limitations,
this study provides novel insights into the complex transcriptomic changes occurring during cerebral cortex
formation.
One interest of the present work is the identification of
several components of the Ub system known to be overexpressed in cancers that correspond to developmental
genes highly expressed in the embryonic cerebral cortex under physiological conditions but are not related to
tumour formation or progression, for instance Ube2c (E2
gene), Trim28, Trim32, and Trim59 (E3 genes). The data
collected may be used as a starting point for future functional studies of the rodent brain.
**Supplementary Information**
[The online version contains supplementary material available at https://doi.](https://doi.org/10.1186/s13041-022-00958-z)
[org/10.1186/s13041-022-00958-z.](https://doi.org/10.1186/s13041-022-00958-z)
**Author details**
1 Université Grenoble Alpes, Inserm, CEA, UMR 1292, 38000 Grenoble, France.
2 Genetics and Chemogenomics Lab, Building C3, CEA, 17 rue des Martyrs,
38054 Grenoble Cedex 9, France.
Received: 3 June 2022 Accepted: 2 August 2022
**References**
1. Gross S, Rahal R, Stransky N, Lengauer C, Hoeflich KP. Targeting cancer
with kinase inhibitors. J Clin Investig. 2015;125(5):1780–9.
2. Komander D, Rape M. The ubiquitin code. Annu Rev Biochem.
2012;81:203–29.
3. Komander D, Clague MJ, Urbe S. Breaking the chains: struc‑
ture and function of the deubiquitinases. Nat Rev Mol Cell Biol.
2009;10(8):550–63.
4. Clague MJ, Urbe S, Komander D. Breaking the chains: deubiquity‑
lating enzyme specificity begets function. Nat Rev Mol Cell Biol.
2019;20(6):338–52.
5. Damgaard RB. The ubiquitin system: from cell signalling to dis‑
ease biology and new therapeutic opportunities. Cell Death Differ.
2021;28(2):423–6.
6. Na CH, Jones DR, Yang Y, Wang X, Xu Y, Peng J. Synaptic protein ubiqui‑
tination in rat brain revealed by antibody-based ubiquitome analysis. J
Proteome Res. 2012;11(9):4722–32.
7. DiAntonio A, Hicke L. Ubiquitin-dependent regulation of the synapse.
Annu Rev Neurosci. 2004;27:223–46.
8. Yamada T, Yang Y, Bonni A. Spatial organization of ubiquitin ligase
pathways orchestrates neuronal connectivity. Trends Neurosci.
2013;36(4):218–26.
9. Kawabe H, Brose N. The role of ubiquitylation in nerve cell develop‑
ment. Nat Rev Neurosci. 2011;12(5):251–68.
10. Todi SV, Paulson HL. Balancing act: deubiquitinating enzymes in the
nervous system. Trends Neurosci. 2011;34(7):370–82.
11. Lee SH, Choi JH, Lee N, Lee HR, Kim JI, Yu NK, Choi SL, Lee SH, Kim H,
Kaang BK. Synaptic protein degradation underlies destabilization of
retrieved fear memory. Science. 2008;319(5867):1253–6.
12. Sewduth RN, Baietti MF, Sablina AA. Cracking the monoubiquitin code
of genetic diseases. Int J Mol Sci. 2020;21(9):3036.
13. Hasna J, Bohic S, Lemoine S, Blugeon C, Bouron A. Zinc uptake and stor‑
age during the formation of the cerebral cortex in mice. Mol Neurobiol.
2019;56(10):6928–40.
14. Jabaudon D. Fate and freedom in developing neocortical circuits. Nat
Commun. 2017;8:16042.
15. Kriegstein AR, Noctor SC. Patterns of neuronal migration in the embry‑
onic cortex. Trends Neurosci. 2004;27(7):392–9.
16. Hallengren J, Chen PC, Wilson SM. Neuronal ubiquitin homeostasis. Cell
Biochem Biophys. 2013;67(1):67–73.
17. Ryu KY, Garza JC, Lu XY, Barsh GS, Kopito RR. Hypothalamic neurode‑
generation and adult-onset obesity in mice lacking the Ubb polyubiq‑
uitin gene. Proc Natl Acad Sci U S A. 2008;105(10):4016–21.
18. Clague MJ, Heride C, Urbe S. The demographics of the ubiquitin system.
Trends Cell Biol. 2015;25(7):417–26.
19. Park CW, Ryu KY. Cellular ubiquitin pool dynamics and homeostasis.
BMB Rep. 2014;47(9):475–82.
20. Bianchi M, Giacomini E, Crinelli R, Radici L, Carloni E, Magnani M.
Dynamic transcription of ubiquitin genes under basal and stressful
conditions and new insights into the multiple UBC transcript variants.
Gene. 2015;573(1):100–9.
21. Ryu HW, Park CW, Ryu KY. Restoration of cellular ubiquitin reverses
impairments in neuronal development caused by disruption of
the polyubiquitin gene Ubb. Biochem Biophys Res Commun.
2014;453(3):443–8.
22. Dubois ML, Meller A, Samandi S, Brunelle M, Frion J, Brunet MA, Toupin
A, Beaudoin MC, Jacques JF, Levesque D et al. UBB pseudogene 4
encodes functional ubiquitin variants. Nat Commun. 2020; 11(1).
23. Cappadocia L, Lima CD. Ubiquitin-like protein conjugation: structures,
chemistry, and mechanism. Chem Rev. 2018;118(3):889–918.
**Additional file 1. Supplementary Figure 1. It shows the expression**
of the minor genes (TPM values <40) encoding Ub-(A) and Ub-like (B)
proteins conjugating E2 enzymes. Supplementary Figure 2. It shows
the expression of the minor Fbxw (A), Fbxl (B) and Fbxo genes (TPM values
<40). Supplementary Figure 3. It shows the expression of the minor Usp
genes.
**Acknowledgements**
We thank Dr Helen Walden for her help with the Fanconi anaemia com‑
plex and Dr Claudio Joazeiro for his comments on a preliminary version of
this manuscript. We also wish to thank Dr Sophie Lemoine and Dr Corinne
Blugeon for their help with the transcriptomic analysis.
**Author contributions**
Data curation and analysis, AB; writing, AB; review and editing, AB, MOF. Both
authors read and approved the final manuscript.
**Funding**
The work receives support from the Centre National de la Recherche Scienti‑
fique (CNRS), Commissariat à l’Energie Atomique et aux Energies Alternatives
(CEA), Université de Grenoble Alpes (UGA). This project received funding from
GRAL, a programme of the Chemistry Biology Health (CBH) Graduate School
of University Grenoble Alpes (ANR-17-EURE-0003).
**Availability of data and materials**
The complete dataset is freely accessible on the GEO repository with the
accession number GSE154677.
**Declarations**
**Ethics approval and consent to participate**
Not applicable.
**Consent for publication**
Not applicable.
**Competing interests**
The authors declare that they have no competing interests.
-----
24. Schulman BA, Harper JW. Ubiquitin-like protein activation by E1
enzymes: the apex for downstream signalling pathways. Nat Rev Mol
Cell Bio. 2009;10(5):319–31.
25. Wagner GP, Kin K, Lynch VJ. A model based criterion for gene expres‑
sion calls using RNA-seq data. Theory Biosci. 2013;132(3):159–64.
26. Barghout SH, Schimmer AD. E1 enzymes as therapeutic targets in
cancer. Pharmacol Rev. 2021;73(1):1–58.
27. Lambert-Smith IA, Saunders DN, Yerbury JJ. The pivotal role of ubiqui‑
tin-activating enzyme E1 (UBA1) in neuronal health and neurodegen‑
eration. Int J Biochem Cell B. 2020;123:105746.
28. Groen EJN, Gillingwatert TH. UBA1: at the crossroads of ubiquitin home‑
ostasis and neurodegeneration. Trends Mol Med. 2015;21(10):622–32.
29. Lee PC, Dodart JC, Aron L, Finley LW, Bronson RT, Haigis MC, Yankner
BA, Harper JW. Altered social behavior and neuronal development
in mice lacking the Uba6-Use1 ubiquitin transfer system. Mol Cell.
2013;50(2):172–84.
30. van Wijk SJL, Timmers HTM. The family of ubiquitin-conjugating
enzymes (E2s): deciding between life and death of proteins. FASEB J.
2010;24(4):981–93.
31. Ye YH, Rape M. Building ubiquitin chains: E2 enzymes at work. Nat Rev
Mol Cell Bio. 2009;10(11):755–64.
32. van Wijk SJ, de Vries SJ, Kemmeren P, Huang A, Boelens R, Bonvin AM,
Timmers HT. A comprehensive framework of E2-RING E3 interactions of
the human ubiquitin-proteasome system. Mol Syst Biol. 2009;5:295.
33. Xie CL, Powell C, Yao M, Wu JM, Dong QH. Ubiquitin-conjugating
enzyme E2C: a potential cancer biomarker. Int J Biochem Cell B.
2014;47:113–7.
34. Wan C, Chen J, Hu B, Zou H, Li A, Guo A, Jiang J. Downregulation of
UBE2Q1 is associated with neuronal apoptosis in rat brain cortex fol‑
lowing traumatic brain injury. J Neurosci Res. 2014;92(1):1–12.
35. Koerver L, Papadopoulos C, Liu B, Kravic B, Rota G, Brecht L, Veenendaal
T, Polajnar M, Bluemke A, Ehrmann M, et al. The ubiquitin-conjugating
enzyme UBE2QL1 coordinates lysophagy in response to endolysosomal
damage. EMBO Rep. 2019;20(10): e48014.
36. Nuber U, Schwarz S, Kaiser P, Schneider R, Scheffner M. Cloning of
human ubiquitin-conjugating enzymes UbcH6 and UbcH7 (E2–F1) and
characterization of their interaction with E6-AP and RSP5. J Biol Chem.
1996;271(5):2795–800.
37. Kishino T, Lalande M, Wagstaff J. UBE3A/E6-AP mutations cause Angel‑
man syndrome. Nat Genet. 1997;15(1):70–3.
38. Zheng N, Shabek N. Ubiquitin ligases: structure, function, and regula‑
tion. Annu Rev Biochem. 2017;86(86):129–57.
39. Morreale FE, Walden H. Types of ubiquitin ligases. Cell.
2016;165(1):248–248.
40. Dove KK, Klevit RE. RING-between-RING E3 ligases: emerging themes
amid the variations. J Mol Biol. 2017;429(22):3363–75.
41. Deshaies RJ, Joazeiro CA. RING domain E3 ubiquitin ligases. Annu Rev
Biochem. 2009;78:399–434.
42. Wang Y, Argiles-Castillo D, Kane EI, Zhou A, Spratt DE. HECT E3 ubiquitin
ligases - emerging insights into their biological roles and disease
relevance. J Cell Sci. 2020; 133(7).
43. Kumar S, Tomooka Y, Noda M. Identification of a set of genes with
developmentally down-regulated expression in the mouse brain.
Biochem Biophys Res Commun. 1992;185(3):1155–61.
44. Donovan P, Poronnik P. Nedd4 and Nedd4-2: ubiquitin ligases at work in
the neuron. Int J Biochem Cell Biol. 2013;45(3):706–10.
45. Zhao X, Heng JI, Guardavaccaro D, Jiang R, Pagano M, Guillemot F,
Iavarone A, Lasorella A. The HECT-domain ubiquitin ligase Huwe1 con‑
trols neural differentiation and proliferation by destabilizing the N-Myc
oncoprotein. Nat Cell Biol. 2008;10(6):643–53.
46. Hou X, Zhang W, Xiao Z, Gan H, Lin X, Liao S, Han C. Mining and char‑
acterization of ubiquitin E3 ligases expressed in the mouse testis. BMC
Genomics. 2012;13:495.
47. Li W, Bengtson MH, Ulbrich A, Matsuda A, Reddy VA, Orth A, Chanda
SK, Batalov S, Joazeiro CA. Genome-wide and functional annotation of
human E3 ubiquitin ligases identifies MULAN, a mitochondrial E3 that
regulates the organelle’s dynamics and signaling. PLoS ONE. 2008;3(1):
e1487.
48. Wang L, Sun X, He J, Liu Z. Functions and molecular mechanisms of
Deltex family ubiquitin E3 ligases in development and disease. Front
Cell Dev Biol. 2021;9: 706997.
49. Lussier MP, Herring BE, Nasu-Nishimura Y, Neutzner A, Karbowski M,
Youle RJ, Nicoll RA, Roche KW. Ubiquitin ligase RNF167 regulates AMPA
receptor-mediated synaptic transmission. Proc Natl Acad Sci U S A.
2012;109(47):19426–31.
50. Ghilarducci K, Cabana VC, Desroches C, Chabi K, Bourgault S, Cappado‑
cia L, Lussier MP. Functional interaction of ubiquitin ligase RNF167 with
UBE2D1 and UBE2N promotes ubiquitination of AMPA receptor. FEBS J.
2021;288(16):4849–68.
51. Benini M, Fortuni S, Condo I, Alfedi G, Malisan F, Toschi N, Serio D,
Massaro DS, Arcuri G, Testi R, et al. E3 ligase RNF126 directly ubiquit‑
inates frataxin, promoting its degradation: identification of a potential
therapeutic target for Friedreich ataxia. Cell Rep. 2017;18(8):2007–17.
52. Gray TA, Hernandez L, Carey AH, Schaldach MA, Smithwick MJ, Rus
K, Marshall Graves JA, Stewart CL, Nicholls RD. The ancient source of
a distinct gene family encoding proteins featuring RING and C(3)H
zinc-finger motifs with abundant expression in developing brain and
nervous system. Genomics. 2000;66(1):76–86.
53. Miroci H, Schob C, Kindler S, Olschlager-Schutt J, Fehr S, Jungenitz T,
Schwarzacher SW, Bagni C, Mohr E. Makorin ring zinc finger protein 1
(MKRN1), a novel poly(A)-binding protein-interacting protein, stimu‑
lates translation in nerve cells. J Biol Chem. 2012;287(2):1322–34.
54. Yang PH, Cheung WK, Peng Y, He ML, Wu GQ, Xie D, Jiang BH, Huang
QH, Chen Z, Lin MC, et al. Makorin-2 is a neurogenesis inhibitor down‑
stream of phosphatidylinositol 3-kinase/Akt (PI3K/Akt) signal. J Biol
Chem. 2008;283(13):8486–95.
55. Ohmura-Hoshino M, Goto E, Matsuki Y, Aoki M, Mito M, Uematsu M,
Hotta H, Ishido S. A novel family of membrane-bound E3 ubiquitin
ligases. J Biochem. 2006;140(2):147–54.
56. De Angelis RF, De Gassart A, Pforr C, Cano F, N’Guessan P, Combes A,
Camossetto V, Lehner PJ, Pierre P, Gatti E. MARCH9-mediated ubiq‑
uitination regulates MHC I export from the TGN. Immunol Cell Biol.
2017;95(9):753–64.
57. Lin H, Li S, Shu HB. The membrane-associated MARCH E3 ligase family:
emerging roles in immune regulation. Front Immunol. 2019;10:1751.
58. Valnegri P, Huang J, Yamada T, Yang Y, Mejia LA, Cho HY, Oldenborg A,
Bonni A. RNF8/UBC13 ubiquitin signaling suppresses synapse forma‑
tion in the mammalian brain. Nat Commun. 2017;8(1):1271.
59. Cubillos-Rojas M, Schneider T, Bartrons R, Ventura F, Rosa JL. NEURL4
regulates the transcriptional activity of tumor suppressor protein p53
by modulating its oligomerization. Oncotarget. 2017;8(37):61824–36.
60. Shin J, Mishra V, Glasgow E, Zaidi S, Chen J, Ohshiro K, Chitti B, Kapadia
AA, Rana N, Mishra L, et al. PRAJA is overexpressed in glioblastoma
and contributes to neural precursor development. Genes Cancer.
2017;8(7–8):640–9.
61. Kalkan T, Iwasaki Y, Park CY, Thomsen GH. Tumor necrosis factor-recep‑
tor-associated factor-4 is a positive regulator of transforming growth
factor-beta signaling that affects neural crest formation. Mol Biol Cell.
2009;20(14):3436–50.
62. Blaise S, Kneib M, Rousseau A, Gambino F, Chenard MP, Messadeq N,
Muckenstrum M, Alpy F, Tomasetto C, Humeau Y, et al. In vivo evidence
that TRAF4 is required for central nervous system myelin homeostasis.
PLoS ONE. 2012;7(2): e30917.
63. Boyer NP, Monkiewicz C, Menon S, Moy SS, Gupton SL. Mammalian
TRIM67 functions in brain development and behavior. eNeuro. 2018;
5(3).
64. Brattas PL, Jonsson ME, Fasching L, Nelander Wahlestedt J, Shahsavani
M, Falk R, Falk A, Jern P, Parmar M, Jakobsson J. TRIM28 controls a gene
regulatory network based on endogenous retroviruses in human
neural progenitor cells. Cell Rep. 2017;18(1):1–11.
65. Liang Q, Deng H, Li X, Wu X, Tang Q, Chang TH, Peng H, Rauscher FJ
3rd, Ozato K, Zhu F. Tripartite motif-containing protein 28 is a small
ubiquitin-related modifier E3 ligase and negative regulator of IFN
regulatory factor 7. J Immunol. 2011;187(9):4754–63.
66. Hillje AL, Worlitzer MM, Palm T, Schwamborn JC. Neural stem cells main‑
tain their stemness through protein kinase C zeta-mediated inhibition
of TRIM32. Stem cells. 2011;29(9):1437–47.
67. Zhao X, Liu Q, Du B, Li P, Cui Q, Han X, Du B, Yan D, Zhu X. A novel
accessory molecule Trim59 involved in cytotoxicity of BCG-activated
macrophages. Mol Cells. 2012;34(3):263–70.
68. Hatakeyama S. TRIM family proteins: roles in autophagy, immunity, and
carcinogenesis. Trends Biochem Sci. 2017;42(4):297–311.
-----
69. Sarikas A, Hartmann T, Pan ZQ. The cullin protein family. Genome Biol.
2011;12(4):220.
70. Xu J, Zhang Z, Qian M, Wang S, Qiu W, Chen Z, Sun Z, Xiong Y, Wang
C, Sun X, et al. Cullin-7 (CUL7) is overexpressed in glioma cells and
promotes tumorigenesis via NF-kappaB activation. J Exp Clin Cancer
Res CR. 2020;39(1):59.
71. Shi L, Du D, Peng Y, Liu J, Long J. The functional analysis of Cullin 7 E3
ubiquitin ligases in cancer. Oncogenesis. 2020;9(10):98.
72. Litterman N, Ikeuchi Y, Gallardo G, O’Connell BC, Sowa ME, Gygi SP,
Harper JW, Bonni A. An OBSL1-Cul7Fbxw8 ubiquitin ligase signaling
mechanism regulates Golgi morphology and dendrite patterning. PLoS
Biol. 2011;9(5): e1001060.
73. Hsu PH, Ma YT, Fang YC, Huang JJ, Gan YL, Chang PT, Jow GM, Tang CY,
Jeng CJ. Cullin 7 mediates proteasomal and lysosomal degradations of
rat Eag1 potassium channels. Sci Rep. 2017;7:40825.
74. Kipreos ET, Pagano M. The F-box protein family. Genome Biol.
2000;1(5):REVIEWS3002.
75. Puklowski A, Homsi Y, Keller D, May M, Chauhan S, Kossatz U, Grunwald
V, Kubicka S, Pich A, Manns MP, et al. The SCF-FBXW5 E3-ubiquitin
ligase is regulated by PLK4 and targets HsSAS-6 to control centrosome
duplication. Nat Cell Biol. 2011;13(8):1004–9.
76. Chen F, Zhang C, Wu H, Ma Y, Luo X, Gong X, Jiang F, Gui Y, Zhang H,
Lu F. The E3 ubiquitin ligase SCF(FBXL14) complex stimulates neuronal
differentiation by targeting the Notch signaling factor HES1 for prote‑
olysis. J Biol Chem. 2017;292(49):20100–12.
77. Uddin S, Bhat AA, Krishnankutty R, Mir F, Kulinski M, Mohammad RM.
Involvement of F-BOX proteins in progression and development of
human malignancies. Semin Cancer Biol. 2016;36:18–32.
78. Saiga T, Fukuda T, Matsumoto M, Tada H, Okano HJ, Okano H, Nakayama
KI. Fbxo45 forms a novel ubiquitin ligase complex and is required for
neuronal development. Mol Cell Biol. 2009;29(13):3529–43.
79. Tada H, Okano HJ, Takagi H, Shibata S, Yao I, Matsumoto M, Saiga T,
Nakayama KI, Kashima H, Takahashi T, et al. Fbxo45, a novel ubiquitin
ligase, regulates synaptic activity. J Biol Chem. 2010;285(6):3840–9.
80. Zhou W, Wei W, Sun Y. Genetically engineered mouse models for func‑
tional studies of SKP1-CUL1-F-box-protein (SCF) E3 ubiquitin ligases.
Cell Res. 2013;23(5):599–619.
81. Fasanaro P, Capogrossi MC, Martelli F. Regulation of the endothe‑
lial cell cycle by the ubiquitin-proteasome system. Cardiovasc Res.
2010;85(2):272–80.
82. Yamano H. APC/C: current understanding and future perspectives.
F1000Research. 2019;8:725.
83. Schrock MS, Stromberg BR, Scarberry L, Summers MK. APC/C ubiquitin
ligase: functions and mechanisms in tumorigenesis. Semin Cancer Biol.
2020;67(Pt 2):80–91.
84. Delgado-Esteban M, Garcia-Higuera I, Maestre C, Moreno S, Almeida A.
APC/C-Cdh1 coordinates neurogenesis and cortical size during devel‑
opment. Nat Commun. 2013;4:2879.
85. Walden H, Deans AJ. The Fanconi anemia DNA repair pathway:
structural and functional insights into a complex disorder. Annu Rev
Biophys. 2014;43:257–78.
86. Hatakeyama S, Nakayama KI. U-box proteins as a new family of ubiqui‑
tin ligases. Biochem Bioph Res Co. 2003;302(4):635–45.
87. Cyr DM, Hohfeld J, Patterson C. Protein quality control: U-box
containing E3 ubiquitin ligases join the fold. Trends Biochem Sci.
2002;27(7):368–75.
88. Hatakeyama S, Matsumoto M, Yada M, Nakayama KI. Interaction of
U-box-type ubiquitin-protein ligases (E3s) with molecular chaperones.
Genes Cells. 2004;9(6):533–48.
89. Nordquist KA, Dimitrova YN, Brzovic PS, Ridenour WB, Munro KA, Soss
SE, Caprioli RM, Klevit RE, Chazin WJ. Structural and functional charac‑
terization of the monomeric U-box domain from E4B. Biochemistry.
2010;49(2):347–55.
90. Marin I. Ancient origin of animal U-box ubiquitin ligases. BMC Evol Biol.
2010;10:331.
91. Bhuripanyo K, Wang YY, Liu XP, Zhou L, Liu RC, Duong D, Zhao B, Bi YT,
Zhou H, Chen G et al. Identifying the substrate proteins of U-box E3s
E4B and CHIP by orthogonal ubiquitin transfer. Sci Adv. 2018; 4(1).
92. Kaneko C, Hatakeyama S, Matsumoto M, Yada M, Nakayama K,
Nakayama KI. Characterization of the mouse gene for the U-box-type
ubiquitin lipase UFD2a. Biochem Bioph Res Co. 2003;300(2):297–304.
93. Grumati P, Dikic I. Ubiquitin signaling and autophagy. J Biol Chem.
2018;293(15):5404–13.
94. Jacomin AC, Taillebourg E, Fauvarque MO. Deubiquitinating enzymes
related to autophagy: new therapeutic opportunities? Cells. 2018; 7(8).
95. Wang P, Dai X, Jiang W, Li Y, Wei W. RBR E3 ubiquitin ligases in tumori‑
genesis. Semin Cancer Biol. 2020;67(Pt 2):131–44.
96. Fuseya Y, Fujita H, Kim M, Ohtake F, Nishide A, Sasaki K, Saeki Y, Tanaka
K, Takahashi R, Iwai K. The HOIL-1L ligase modulates immune signal‑
ling and cell death via monoubiquitination of LUBAC. Nat Cell Biol.
2020;22(6):663–73.
97. van Well EM, Bader V, Patra M, Sanchez-Vicente A, Meschede J, Furth‑
mann N, Schnack C, Blusch A, Longworth J, Petrasch-Parwez E et al.
A protein quality control pathway regulated by linear ubiquitination.
EMBO J. 2019; 38(9).
98. Peltzer N, Darding M, Montinaro A, Draber P, Draberova H, Kupka S,
Rieser E, Fisher A, Hutchinson C, Taraborrelli L, et al. LUBAC is essential
for embryogenesis by preventing cell death and enabling haemat‑
opoiesis. Nature. 2018;557(7703):112–7.
99. Grou CP, Pinto MP, Mendes AV, Domingues P, Azevedo JE. The de novo
synthesis of ubiquitin: identification of deubiquitinases acting on
ubiquitin precursors. Sci Rep. 2015;5:12836.
100. Anckar J, Bonni A. Regulation of neuronal morphogenesis and posi‑
tioning by ubiquitin-specific proteases in the cerebellum. PLoS ONE.
2015;10(1): e0117076.
101. Xu J. Age-related changes in Usp9x protein expression and DNA meth‑
ylation in mouse brain. Brain Res Mol Brain Res. 2005;140(1–2):17–24.
102. Kobayashi T, Iwamoto Y, Takashima K, Isomura A, Kosodo Y, Kawakami
K, Nishioka T, Kaibuchi K, Kageyama R. Deubiquitinating enzymes
regulate Hes1 stability and neuronal differentiation. FEBS J.
2015;282(13):2411–23.
103. Koutelou E, Wang L, Schibler AC, Chao HP, Kuang X, Lin K, Lu Y, Shen J,
Jeter CR, Salinger A et al. USP22 controls multiple signaling pathways
that are essential for vasculature formation in the mouse placenta.
Development. 2019; 146(4).
104. Lin Z, Yang H, Kong Q, Li J, Lee SM, Gao B, Dong H, Wei J, Song J, Zhang
DD, et al. USP22 antagonizes p53 transcriptional activation by deubiq‑
uitinating Sirt1 to suppress cell apoptosis and is required for mouse
embryonic development. Mol Cell. 2012;46(4):484–94.
105. Sussman RT, Stanek TJ, Esteso P, Gearhart JD, Knudsen KE, McMahon
SB. The epigenetic modifier ubiquitin-specific protease 22 (USP22)
regulates embryonic stem cell differentiation via transcriptional
repression of sex-determining region Y-box 2 (SOX2). J Biol Chem.
2013;288(33):24234–46.
106. Williams SA, Maecker HL, French DM, Liu J, Gregg A, Silverstein LB,
Cao TC, Carano RA, Dixit VM. USP1 deubiquitinates ID proteins to
preserve a mesenchymal stem cell program in osteosarcoma. Cell.
2011;146(6):918–30.
107. Lee JK, Chang N, Yoon Y, Yang H, Cho H, Kim E, Shin Y, Kang W, Oh YT,
Mun GI, et al. USP1 targeting impedes GBM growth by inhibiting stem
cell maintenance and radioresistance. Neuro Oncol. 2016;18(1):37–47.
108. Kim JM, Parmar K, Huang M, Weinstock DM, Ruit CA, Kutok JL, D’Andrea
AD. Inactivation of murine Usp1 results in genomic instability and a
Fanconi anemia phenotype. Dev Cell. 2009;16(2):314–20.
109. Wing SS. Deubiquitinating enzymes in skeletal muscle atrophy-an
essential role for USP19. Int J Biochem Cell Biol. 2016;79:462–8.
110. Pei D. Deubiquitylating Nanog: novel role of USP21 in embryonic stem
cell maintenance. Signal Transduct Target Ther. 2017;2:17014.
111. Ning F, Xin H, Liu J, Lv C, Xu X, Wang M, Wang Y, Zhang W, Zhang X.
Structure and function of USP5: insight into physiological and patho‑
physiological roles. Pharmacol Res. 2020;157: 104557.
112. Garcia-Caballero A, Gadotti VM, Stemkowski P, Weiss N, Souza IA, Hodg‑
kinson V, Bladen C, Chen L, Hamid J, Pizzoccaro A, et al. The deubiquit‑
inating enzyme USP5 modulates neuropathic and inflammatory pain
by enhancing Cav3.2 channel activity. Neuron. 2014;83(5):1144–58.
113. Wilkinson KD, Lee KM, Deshpande S, Duerksen-Hughes P, Boss JM, Pohl
J. The neuron-specific protein PGP 9.5 is a ubiquitin carboxyl-terminal
hydrolase. Science. 1989;246(4930):670–3.
114. Schofield JN, Day IN, Thompson RJ, Edwards YH. PGP9.5, a ubiqui‑
tin C-terminal hydrolase; pattern of mRNA and protein expression
during neural development in the mouse. Brain Res Dev Brain Res.
1995;85(2):229–38.
-----
115. Sakurai M, Ayukawa K, Setsuie R, Nishikawa K, Hara Y, Ohashi H,
Nishimoto M, Abe T, Kudo Y, Sekiguchi M, et al. Ubiquitin C-terminal
hydrolase L1 regulates the morphology of neural progenitor cells and
modulates their differentiation. J Cell Sci. 2006;119(Pt 1):162–71.
116. Bishop P, Rocca D, Henley JM. Ubiquitin C-terminal hydrolase L1 (UCH
L1): structure, distribution and roles in brain function and dysfunction.
Biochem J. 2016;473(16):2453–62.
117. Bononi A, Giorgi C, Patergnani S, Larson D, Verbruggen K, Tanji M, Pel‑
legrini L, Signorato V, Olivetto F, Pastorino S, et al. BAP1 regulates IP3R3mediated Ca(2+) flux to mitochondria suppressing cell transformation.
Nature. 2017;546(7659):549–53.
118. Dai F, Lee H, Zhang Y, Zhuang L, Yao H, Xi Y, Xiao ZD, You MJ, Li W,
Su X, et al. BAP1 inhibits the ER stress gene regulatory network and
modulates metabolic stress response. Proc Natl Acad Sci U S A.
2017;114(12):3192–7.
119. Ristic G, Tsou WL, Todi SV. An optimal ubiquitin-proteasome pathway
in the nervous system: the role of deubiquitinating enzymes. Front Mol
Neurosci. 2014;7:72.
120. Xie L, Li A, Shen J, Cao M, Ning X, Yuan D, Ji Y, Wang H, Ke K. OTUB1
attenuates neuronal apoptosis after intracerebral hemorrhage. Mol Cell
Biochem. 2016;422(1–2):171–80.
121. Xia Q, Liao L, Cheng D, Duong DM, Gearing M, Lah JJ, Levey AI, Peng J.
Proteomic identification of novel proteins associated with Lewy bodies.
Front Biosci. 2008;13:3850–6.
122. Kumari R, Kumar R, Kumar S, Singh AK, Hanpude P, Jangir D, Maiti TK.
Amyloid aggregates of the deubiquitinase OTUB1 are neurotoxic, sug‑
gesting that they contribute to the development of Parkinson’s disease.
J Biol Chem. 2020;295(11):3466–84.
123. Abdul Rehman SA, Kristariyanto YA, Choi SY, Nkosi PJ, Weidlich S, Labib
K, Hofmann K, Kulathu Y. MINDY-1 is a member of an evolutionarily
conserved and structurally distinct new family of deubiquitinating
enzymes. Mol Cell. 2016;63(1):146–55.
124. Kwasna D, Abdul Rehman SA, Natarajan J, Matthews S, Madden R, De
Cesare V, Weidlich S, Virdee S, Ahel I, Gibbs-Seymour I, et al. Discovery
and characterization of ZUFSP/ZUP1, a distinct deubiquitinase class
important for genome stability. Mol Cell. 2018;70(1):150–64.
**Publisher’s Note**
Springer Nature remains neutral with regard to jurisdictional claims in pub‑
lished maps and institutional affiliations.
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://pmc.ncbi.nlm.nih.gov/articles/PMC9380329, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "GOLD",
"url": "https://molecularbrain.biomedcentral.com/counter/pdf/10.1186/s13041-022-00958-z"
}
| 2,022
|
[
"JournalArticle",
"Review"
] | true
| 2022-08-16T00:00:00
|
[
{
"paperId": "e5bb5a26ee9153ca09c8ab39050dcf4ad7f62010",
"title": "Functions and Molecular Mechanisms of Deltex Family Ubiquitin E3 Ligases in Development and Disease"
},
{
"paperId": "7214f8e5d30b2d5e155691b9eaa0135861a46969",
"title": "Functional interaction of ubiquitin ligase RNF167 with UBE2D1 and UBE2N promotes ubiquitination of AMPA receptor"
},
{
"paperId": "cd219831de40cc36539f3d273b73edec489e081a",
"title": "The ubiquitin system: from cell signalling to disease biology and new therapeutic opportunities"
},
{
"paperId": "08aea2619ffc7e585bbf4ad193fb67a5c823b0f8",
"title": "E1 Enzymes as Therapeutic Targets in Cancer"
},
{
"paperId": "f4b402405a230a4481f66db04367b0f65b3fd206",
"title": "The functional analysis of Cullin 7 E3 ubiquitin ligases in cancer"
},
{
"paperId": "3ede35d399b20c158eb3da46654f3dcf63784841",
"title": "RBR E3 ubiquitin ligases in tumorigenesis."
},
{
"paperId": "1b1c256ad65d78a46884b7d36478b6a00359a114",
"title": "The HOIL-1L ligase modulates immune signalling and cell death via monoubiquitination of LUBAC"
},
{
"paperId": "e1aeb071a9c6d8ceff9bbc63406e98fe735a2183",
"title": "Cracking the Monoubiquitin Code of Genetic Diseases"
},
{
"paperId": "700f7c6c26141e30cb91ed2a94937cbd2c7b3348",
"title": "The pivotal role of Ubiquitin-activating enzyme E1 (UBA1) in neuronal health and neurodegeneration."
},
{
"paperId": "ae3d57ecaea2fc10237152c2ab0e94279651f7ae",
"title": "Cullin-7 (CUL7) is overexpressed in glioma cells and promotes tumorigenesis via NF-κB activation"
},
{
"paperId": "9afb032bdbbd520a04d2a28722b0dc9899008dad",
"title": "HECT E3 ubiquitin ligases – emerging insights into their biological roles and disease relevance"
},
{
"paperId": "ccc50605fca2b8348eceb597f116f8f0ee61a7e8",
"title": "UBB pseudogene 4 encodes functional ubiquitin variants"
},
{
"paperId": "8c491425087a1e8ed1e695683ba949219fec848b",
"title": "APC/C ubiquitin ligase: functions and mechanisms in tumorigenesis."
},
{
"paperId": "321530efeaba7cc9fa5eb23b317d0a7d2a930819",
"title": "Amyloid aggregates of the deubiquitinase OTUB1 are neurotoxic, suggesting that they contribute to the development of Parkinson's disease"
},
{
"paperId": "709a7fe874a114a8c6eaef010df22cc3e951a5ed",
"title": "Structure and function of USP5: Insight into physiological and pathophysiological roles."
},
{
"paperId": "f9a77002177403d781576aaa4a615d3ee15b2314",
"title": "The ubiquitin‐conjugating enzyme UBE2QL1 coordinates lysophagy in response to endolysosomal damage"
},
{
"paperId": "87fab1ec24a0f8573705e992817efab7667ac8cb",
"title": "The Membrane-Associated MARCH E3 Ligase Family: Emerging Roles in Immune Regulation"
},
{
"paperId": "7573b3fba1117ca9d50bfa322409ed20eaf62d31",
"title": "APC/C: current understanding and future perspectives"
},
{
"paperId": "1cdb5ff4c43772c8da4160c9cbbf89a9b3c9b19a",
"title": "Zinc Uptake and Storage During the Formation of the Cerebral Cortex in Mice"
},
{
"paperId": "874ee9ad1981d43567c7187809908668f23a8c14",
"title": "A protein quality control pathway regulated by linear ubiquitination"
},
{
"paperId": "979e678f53ecad79d68c5e8729d7426c9c40251e",
"title": "Breaking the chains: deubiquitylating enzyme specificity begets function"
},
{
"paperId": "3d244aacc7ff2667b33b827db2012179864d0a5f",
"title": "USP22 controls multiple signaling pathways that are essential for vasculature formation in the mouse placenta"
},
{
"paperId": "c9cf89e7642b3da75a33d482d26a330212677479",
"title": "Deubiquitinating Enzymes Related to Autophagy: New Therapeutic Opportunities?"
},
{
"paperId": "d0510e7a011d3257b5c816b6e8d2c09cb0026454",
"title": "Mammalian TRIM67 Functions in Brain Development and Behavior"
},
{
"paperId": "ca68595bf945b22156e2f9b1b074973017c4cf4f",
"title": "Discovery and Characterization of ZUFSP/ZUP1, a Distinct Deubiquitinase Class Important for Genome Stability"
},
{
"paperId": "1eb9c18190345b1be5dd8a0013e3bbf29c4d04b2",
"title": "LUBAC is essential for embryogenesis by preventing cell death and enabling haematopoiesis"
},
{
"paperId": "ab283aed475305e44f0ef5db38ae766d2df051db",
"title": "Identifying the substrate proteins of U-box E3s E4B and CHIP by orthogonal ubiquitin transfer"
},
{
"paperId": "e746befcb34fb62ba614c0a39f6ef261fd5ccd1c",
"title": "Ubiquitin signaling and autophagy"
},
{
"paperId": "15250730127427b824a0f2d067d0da2f0ccec33c",
"title": "RING-Between-RING E3 Ligases: Emerging Themes amid the Variations."
},
{
"paperId": "a40e6ed9811a106337f5ceaf0973562859fca9f3",
"title": "RNF8/UBC13 ubiquitin signaling suppresses synapse formation in the mammalian brain"
},
{
"paperId": "e69310387cd5d7940960a11681328f516fe5956d",
"title": "MARCH9‐mediated ubiquitination regulates MHC I export from the TGN"
},
{
"paperId": "806b06ef786edcedf1570c714678077c61f1a157",
"title": "Fate and freedom in developing neocortical circuits"
},
{
"paperId": "7ba20f1ffa00acba47a9f7d48530850b23a196de",
"title": "PRAJA is overexpressed in glioblastoma and contributes to neural precursor development"
},
{
"paperId": "43ce931c66dd7bedd00a0c1178d30e20f6f3f239",
"title": "Ubiquitin Ligases: Structure, Function, and Regulation."
},
{
"paperId": "b13bd118e7c7d8c681012eeb23944ab47865e21c",
"title": "NEURL4 regulates the transcriptional activity of tumor suppressor protein p53 by modulating its oligomerization"
},
{
"paperId": "6482873df64a490c03e12353ac7a40d9c1bbd6d4",
"title": "BAP1 regulates IP3R3-mediated Ca2+ flux to mitochondria suppressing cell transformation"
},
{
"paperId": "cc70e10bcedf714eafbe892ffb1aa689d6f02292",
"title": "Deubiquitylating Nanog: novel role of USP21 in embryonic stem cell maintenance"
},
{
"paperId": "bd7104c4cca4260dd43c7e5c823685336b8d7498",
"title": "TRIM Family Proteins: Roles in Autophagy, Immunity, and Carcinogenesis."
},
{
"paperId": "c12ec8449b6f45662f3a2743ca2cb8b18cf51688",
"title": "BAP1 inhibits the ER stress gene regulatory network and modulates metabolic stress response"
},
{
"paperId": "e897253e028a0637cfd4e822efeaa981de30f06b",
"title": "Ubiquitin-like Protein Conjugation: Structures, Chemistry, and Mechanism"
},
{
"paperId": "c1ef19883567eb1a735bf7979fce6f062691b0b8",
"title": "E3 Ligase RNF126 Directly Ubiquitinates Frataxin, Promoting Its Degradation: Identification of a Potential Therapeutic Target for Friedreich Ataxia"
},
{
"paperId": "77569da96d2ccbe6a969d1488bd2dd2bd52446c5",
"title": "Cullin 7 mediates proteasomal and lysosomal degradations of rat Eag1 potassium channels"
},
{
"paperId": "9524d00faf53b53459447294f8d3827fc3528b81",
"title": "TRIM28 Controls a Gene Regulatory Network Based on Endogenous Retroviruses in Human Neural Progenitor Cells."
},
{
"paperId": "e81cba8231c650e65ca22e09ff95870e4084874d",
"title": "Deubiquitinating enzymes in skeletal muscle atrophy-An essential role for USP19."
},
{
"paperId": "351cafd4582a53fe006e3664b242267d5d3933f2",
"title": "OTUB1 attenuates neuronal apoptosis after intracerebral hemorrhage"
},
{
"paperId": "d053b5ad410489efc193f333112b988a77b555a8",
"title": "Ubiquitin C-terminal hydrolase L1 (UCH-L1): structure, distribution and roles in brain function and dysfunction"
},
{
"paperId": "6effad3ba483e7996465d07e017db86f1f63480d",
"title": "MINDY-1 Is a Member of an Evolutionarily Conserved and Structurally Distinct New Family of Deubiquitinating Enzymes"
},
{
"paperId": "1dedae5007832dedb60a491bfac559e9811e275b",
"title": "Types of Ubiquitin Ligases"
},
{
"paperId": "1ad545f66256176b53c18a9136dccc9e2a9f9a76",
"title": "Involvement of F-BOX proteins in progression and development of human malignancies."
},
{
"paperId": "76d0c6fe797444a8e63aa63254c3d667de2957a7",
"title": "Dynamic transcription of ubiquitin genes under basal and stressful conditions and new insights into the multiple UBC transcript variants."
},
{
"paperId": "57264f204e21a558088a9adb47540e4c63aceee7",
"title": "UBA1: At the Crossroads of Ubiquitin Homeostasis and Neurodegeneration"
},
{
"paperId": "8068195a216b357b40a6cb460bd99a0d8ad98772",
"title": "The de novo synthesis of ubiquitin: identification of deubiquitinases acting on ubiquitin precursors"
},
{
"paperId": "94b10d376283e1699e87c865ac3d40d0354c9300",
"title": "The demographics of the ubiquitin system."
},
{
"paperId": "4150e439eab9c8681268deca6ab4504f5e5eca5a",
"title": "Deubiquitinating enzymes regulate Hes1 stability and neuronal differentiation"
},
{
"paperId": "4f7e7ee9d8eaaecdd99f18773a4136c97446d054",
"title": "Targeting cancer with kinase inhibitors."
},
{
"paperId": "b6d81ee46a752067888fa0aa52a0f0eb94323285",
"title": "Regulation of Neuronal Morphogenesis and Positioning by Ubiquitin-Specific Proteases in the Cerebellum"
},
{
"paperId": "cb5cd5e43fc9518c1a7e260a32379f056a8c1034",
"title": "Restoration of cellular ubiquitin reverses impairments in neuronal development caused by disruption of the polyubiquitin gene Ubb."
},
{
"paperId": "3f0bbc89970ff0f708920d4d41154fc10b1ccfd4",
"title": "The Deubiquitinating Enzyme USP5 Modulates Neuropathic and Inflammatory Pain by Enhancing Cav3.2 Channel Activity"
},
{
"paperId": "484fa8d90cd34dd4c0509cfc4c71172e3a8d7993",
"title": "Cellular ubiquitin pool dynamics and homeostasis"
},
{
"paperId": "ae142c9b88d09a35efd1311b0b44c60155e7e87b",
"title": "An optimal ubiquitin-proteasome pathway in the nervous system: the role of deubiquitinating enzymes"
},
{
"paperId": "c7795dc786298fd24b1a145399a1582a44b81d17",
"title": "The Fanconi anemia DNA repair pathway: structural and functional insights into a complex disorder."
},
{
"paperId": "6e9876806a4f48adb1c7aed6dded69ab03071fdd",
"title": "Neurobehavioural effects of developmental toxicity"
},
{
"paperId": "5956f58b9b340846cf82e1c745e6b639ad295851",
"title": "Ubiquitin-conjugating enzyme E2C: a potential cancer biomarker."
},
{
"paperId": "471cb1ac76a6f8903eb38e708cf57862fed36047",
"title": "Downregulation of UBE2Q1 is associated with neuronal apoptosis in rat brain cortex following traumatic brain injury"
},
{
"paperId": "fadb6c833ed2cd00d96f98013244320838e6127e",
"title": "APC/C-Cdh1 coordinates neurogenesis and cortical size during development"
},
{
"paperId": "92a97e9ffc6e43fc50b687ab7cb8c391fcf7dbef",
"title": "The Epigenetic Modifier Ubiquitin-specific Protease 22 (USP22) Regulates Embryonic Stem Cell Differentiation via Transcriptional Repression of Sex-determining Region Y-box 2 (SOX2)*"
},
{
"paperId": "d01501b743ce2ce2637c2d79fa368fd5a7233467",
"title": "Neuronal Ubiquitin Homeostasis"
},
{
"paperId": "c89ec9c0034d0e79a053073c8f4f34bb339a1396",
"title": "Altered social behavior and neuronal development in mice lacking the Uba6-Use1 ubiquitin transfer system."
},
{
"paperId": "0b32f919568991e440f92d2685ce3063e1764a24",
"title": "A model based criterion for gene expression calls using RNA-seq data"
},
{
"paperId": "142fb42d200bb840462f5f6c7b5554f85b350ef3",
"title": "Spatial organization of ubiquitin ligase pathways orchestrates neuronal connectivity"
},
{
"paperId": "846ca56a7ae19680a4e6544be19d175f399fc194",
"title": "Genetically engineered mouse models for functional studies of SKP1-CUL1-F-box-protein (SCF) E3 ubiquitin ligases"
},
{
"paperId": "81659835f9b09ac4abc3ec8dd2a23dfdc483f81c",
"title": "Nedd4 and Nedd4-2: ubiquitin ligases at work in the neuron."
},
{
"paperId": "0d44c103b7050b9a39bf07172326c5eaa8400acf",
"title": "Ubiquitin ligase RNF167 regulates AMPA receptor-mediated synaptic transmission"
},
{
"paperId": "1cb02df5084781ec988fa2d73b8a1e53199af473",
"title": "Mining and characterization of ubiquitin E3 ligases expressed in the mouse testis"
},
{
"paperId": "29c412d2aed1a78d33806a57679588226dc9551c",
"title": "Mining and characterization of ubiquitin E3 ligases expressed in the mouse testis"
},
{
"paperId": "8c471432d5a7926bdab78d15ceee0378442ef988",
"title": "Synaptic protein ubiquitination in rat brain revealed by antibody-based ubiquitome analysis."
},
{
"paperId": "91aaac570b2142db5a4a8c6e8b7b9a3f6758fce5",
"title": "A novel accessory molecule Trim59 involved in cytotoxicity of BCG-activated macrophages"
},
{
"paperId": "5b2953bfd8ea399abbd4d0c58a7da91d859914e6",
"title": "The ubiquitin code."
},
{
"paperId": "72c81d60057c35d34bf6d87395692f46a97c73d1",
"title": "USP22 antagonizes p53 transcriptional activation by deubiquitinating Sirt1 to suppress cell apoptosis and is required for mouse embryonic development."
},
{
"paperId": "6845d753ba4e9d7d239fe47884f5a7871ca6c7b1",
"title": "In Vivo Evidence That TRAF4 Is Required for Central Nervous System Myelin Homeostasis"
},
{
"paperId": "cfb083749fefbbc49afe7570f520117a93b18c82",
"title": "Iron toxicity in neurodegeneration"
},
{
"paperId": "3bd32cddfd7d73bbb40e178dd5b6f9a9af321978",
"title": "Makorin Ring Zinc Finger Protein 1 (MKRN1), a Novel Poly(A)-binding Protein-interacting Protein, Stimulates Translation in Nerve Cells*"
},
{
"paperId": "585f26ca3943a4b238d192f9ece631343ee9c826",
"title": "Tripartite Motif-Containing Protein 28 Is a Small Ubiquitin-Related Modifier E3 Ligase and Negative Regulator of IFN Regulatory Factor 7"
},
{
"paperId": "6d288f5ddac9c345af8cedfaaf11f63d9deb5dda",
"title": "USP1 Deubiquitinates ID Proteins to Preserve a Mesenchymal Stem Cell Program in Osteosarcoma"
},
{
"paperId": "1f44a87617d2a35dd41822090c1d2ae39c1caa0a",
"title": "Neural Stem Cells Maintain Their Stemness through Protein Kinase C ζ‐Mediated Inhibition of TRIM32"
},
{
"paperId": "0d7829a89bf927e95edd4d7af8a571be454104f6",
"title": "The SCF–FBXW5 E3-ubiquitin ligase is regulated by PLK4 and targets HsSAS-6 to control centrosome duplication"
},
{
"paperId": "28c3cb1dd49cffb3ca5e5d43bc2dd7b2a5180d29",
"title": "Balancing act: deubiquitinating enzymes in the nervous system"
},
{
"paperId": "22929bafb6748508a0c3a9487b90962f2d169dc5",
"title": "The role of ubiquitylation in nerve cell development"
},
{
"paperId": "e6ac047c89940e731791542c6cc22087e0ddd835",
"title": "An OBSL1-Cul7Fbxw8 Ubiquitin Ligase Signaling Mechanism Regulates Golgi Morphology and Dendrite Patterning"
},
{
"paperId": "9a1d270baff22f4ed145b7e6cfd539f24b484295",
"title": "The cullin protein family"
},
{
"paperId": "c8261e5508c0b2d95e9419dd97ef2185a534f8a8",
"title": "Ancient origin of animal U-box ubiquitin ligases"
},
{
"paperId": "fbb329dd2e56cc61bc65c613a127eca7f9fb1adb",
"title": "The family of ubiquitin‐conjugating enzymes (E2s): deciding between life and death of proteins"
},
{
"paperId": "a79a40260468441d0c56d819ab2ccdd0f83885f6",
"title": "Structural and functional characterization of the monomeric U-box domain from E4B."
},
{
"paperId": "1038b0cff47071f527f69d5714278d8bffd2fe73",
"title": "Regulation of the endothelial cell cycle by the ubiquitin-proteasome system."
},
{
"paperId": "7f2d7d4f14540a4accf18f0815b77f6b3eee46e7",
"title": "Fbxo45, a Novel Ubiquitin Ligase, Regulates Synaptic Activity*"
},
{
"paperId": "e8324db196f5eef7c5f93d73b087269eabd6f7e6",
"title": "Building ubiquitin chains: E2 enzymes at work"
},
{
"paperId": "0f6045c62ca785ff117b95bb77d500855a6869c7",
"title": "A comprehensive framework of E2–RING E3 interactions of the human ubiquitin–proteasome system"
},
{
"paperId": "502e7ab3db4438511bd316d0bab6efec474cb684",
"title": "Breaking the chains: structure and function of the deubiquitinases"
},
{
"paperId": "5eed0dc4886c43ba9b90db1b9a6a039412dfe828",
"title": "Tumor necrosis factor-receptor-associated factor-4 is a positive regulator of transforming growth factor-beta signaling that affects neural crest formation."
},
{
"paperId": "bad3f08c363426e09b0089582deb2cecfa364e86",
"title": "RING domain E3 ubiquitin ligases."
},
{
"paperId": "44db31c41ab0845a121901fe8ca24a5709b8ef3b",
"title": "Ubiquitin-like protein activation by E1 enzymes: the apex for downstream signalling pathways"
},
{
"paperId": "39e6b7b4a563db11144bb0a7011c159bd6a4ac76",
"title": "Fbxo45 Forms a Novel Ubiquitin Ligase Complex and Is Required for Neuronal Development"
},
{
"paperId": "5250c7a201c0aed2ea2893a3e047785168a316d5",
"title": "Inactivation of murine Usp1 results in genomic instability and a Fanconi anemia phenotype."
},
{
"paperId": "6511fab711aae31dd4aefa5243cabe0972759946",
"title": "The HECT-domain ubiquitin ligase Huwe1 controls neural differentiation and proliferation by destabilizing the N-Myc oncoprotein"
},
{
"paperId": "610daf80fe883271e0b5d451a1ee41ccbfa71ee9",
"title": "Proteomic identification of novel proteins associated with Lewy bodies."
},
{
"paperId": "a69069067161c28dbb821046af02eec33fe40291",
"title": "Makorin-2 Is a Neurogenesis Inhibitor Downstream of Phosphatidylinositol 3-Kinase/Akt (PI3K/Akt) Signal*"
},
{
"paperId": "17cf9c5da0c75344be782f84b08f7a479efb7208",
"title": "Hypothalamic neurodegeneration and adult-onset obesity in mice lacking the Ubb polyubiquitin gene"
},
{
"paperId": "92955849e0401f762ec2e88c9e82b42839049022",
"title": "Synaptic Protein Degradation Underlies Destabilization of Retrieved Fear Memory"
},
{
"paperId": "286b4740717f21abcc48d2f5a13b287f19295ee4",
"title": "Genome-Wide and Functional Annotation of Human E3 Ubiquitin Ligases Identifies MULAN, a Mitochondrial E3 that Regulates the Organelle's Dynamics and Signaling"
},
{
"paperId": "3fdeb142fa7d79648f146caa0758f7ab04707fea",
"title": "A novel family of membrane-bound E3 ubiquitin ligases."
},
{
"paperId": "edbbdaced3b7665315d3e5c79571e06af2843fe6",
"title": "Ubiquitin C-terminal hydrolase L1 regulates the morphology of neural progenitor cells and modulates their differentiation"
},
{
"paperId": "c2208857e4a6d29aa65d370cb2eca5ad0dc6d171",
"title": "Age-related changes in Usp9x protein expression and DNA methylation in mouse brain."
},
{
"paperId": "ef326d9e76f9a152d54f649f06dbb3e70c71b4aa",
"title": "Patterns of neuronal migration in the embryonic cortex"
},
{
"paperId": "0473b85a9d2deb40ca455a8e6e523e19eae740b6",
"title": "Ubiquitin-dependent regulation of the synapse."
},
{
"paperId": "16fca93d7d9d216271d933753564203965a133a4",
"title": "Interaction of U‐box‐type ubiquitin‐protein ligases (E3s) with molecular chaperones"
},
{
"paperId": "fb6be405704e37aaf8edc521fcf19f7f8ce80044",
"title": "U-box proteins as a new family of ubiquitin ligases."
},
{
"paperId": "cb0ed3da02549d2740e37773f01257dc91740e8b",
"title": "Characterization of the mouse gene for the U-box-type ubiquitin ligase UFD2a."
},
{
"paperId": "411639b4cb9253fe486b2471ab3dedaac723a1e6",
"title": "Protein quality control: U-box-containing E3 ubiquitin ligases join the fold."
},
{
"paperId": "741988991b95cf84bd968f0835cd4bfad1d2bc77",
"title": "The F-box protein family"
},
{
"paperId": "ea0296945068cfa177ed84a30db31ceaed411c8e",
"title": "The ancient source of a distinct gene family encoding proteins featuring RING and C(3)H zinc-finger motifs with abundant expression in developing brain and nervous system."
},
{
"paperId": "85bc9e16b847648c6329167b2ccb6eabc009125a",
"title": "UBE3A/E6-AP mutations cause Angelman syndrome"
},
{
"paperId": "22896ca2db2bdefd1778e9af02bcdeedb02adab6",
"title": "Cloning of Human Ubiquitin-conjugating Enzymes UbcH6 and UbcH7 (E2-F1) and Characterization of Their Interaction with E6-AP and RSP5 (*)"
},
{
"paperId": "a60832cbef8ac6904058c9c24c3bcdc8a8bf4ec0",
"title": "PGP9.5, a ubiquitin C-terminal hydrolase; pattern of mRNA and protein expression during neural development in the mouse."
},
{
"paperId": "b394eff16abda393c9968842ebc8f9b0744dceb4",
"title": "Identification of a set of genes with developmentally down-regulated expression in the mouse brain."
},
{
"paperId": "b193d1e5d99bed859ab3d62c748ecf2a085f0a39",
"title": "The neuron-specific protein PGP 9.5 is a ubiquitin carboxyl-terminal hydrolase."
},
{
"paperId": null,
"title": "The E3 ubiquitin ligase SCF(FBXL14) complex stimulates neuronal differentiation by targeting the Notch signaling factor HES1 for prote‐ olysis"
},
{
"paperId": "4c6c581316a825c69ed4a22ac67c8a78dca07d6a",
"title": "USP1 targeting impedes GBM growth by inhibiting stem cell maintenance and radioresistance."
},
{
"paperId": "34b9635d7779e219e9d60e0d3d33919ca9bc123c",
"title": "Publisher's Note"
},
{
"paperId": null,
"title": "Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations"
}
] | 30,868
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/023095b6c75a66623876dbc6bca0dfe6b78291f3
|
[
"Computer Science"
] | 0.906266
|
On the composition of authenticated byzantine agreement
|
023095b6c75a66623876dbc6bca0dfe6b78291f3
|
Symposium on the Theory of Computing
|
[
{
"authorId": "1682750",
"name": "Yehuda Lindell"
},
{
"authorId": "1783286",
"name": "Anna Lysyanskaya"
},
{
"authorId": "1693109",
"name": "T. Rabin"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"Symp Theory Comput",
"STOC"
],
"alternate_urls": null,
"id": "8113a511-e0d9-4231-a1bc-0bf5d0212a4e",
"issn": null,
"name": "Symposium on the Theory of Computing",
"type": "conference",
"url": "http://acm-stoc.org/"
}
| null |
# On the Composition of Authenticated Byzantine Agreement
## Yehuda Lindell�
### Dept. of Computer Science Weizmann Institute of Science Rehovot 76100, ISRAEL lindell@wisdom.weizmann.ac.il
## ABSTRACT
�
## Anna Lysyanskaya
### MIT LCS 200 Technology Square Cambridge, MA 02139 USA anna@theory.lcs.mit.edu
## 1. INTRODUCTION
## Tal Rabin
### IBM T.J.Watson Research PO Box 704, Yorktown Heights NY 10598, USA talr@watson.ibm.com
1
The Byzantine Generals (Byzantine Agreement ) problem
A fundamental problem of distributed computing is that
of simulating a (secure) broadcast channel, within the set
ting of a p oint-to-p oint network. This problem is known
as Byzantine Agreement and has b een the fo cus of much
research. Lamp ort et al. showed that in order to achieve
Byzantine Agreement in the standard mo del, more than 2=3
of the participating parties must b e honest. They further
showed that by augmenting the network with a public-key
infrastructure, it is p ossible to obtain secure proto cols for
any numb er of faulty parties. This augmented problem is
called \authenticated Byzantine Agreement".
In this pap er we consider the question of concurrent, par
allel and sequential comp osition of authenticated Byzantine
Agreement proto cols. We present surprising imp ossibility
results showing that:
1. Authenticated Byzantine Agreement cannot b e com
p osed in parallel or concurrently (even twice), if 1=3
or more of the parties are faulty.
2. Deterministic authenticated Byzantine Agreement pro
to cols that run for r rounds and tolerate 1=3 or more
faulty parties, can only b e comp osed sequentially less
than 2r times.
In contrast, we present randomized proto cols for authen
ticated Byzantine Agreement that comp ose sequentially for
any p olynomial numb er of times. We exhibit two such proto
cols: The �rst proto col tolerates corruptions of up to 1=2 of
the parties, while In the �rst proto col, the numb er of faulty
parties may b e any numb er less than 1=2. On the other
hand, the second proto col can tolerate any numb er of faulty
parties, but is limited to the case that the overall numb er of
parties is O (log k ), where k is a security parameter. Finally,
we show that when the mo del is further augmented so that
unique and common session identi�ers are assigned to each
concurrent session, then any p olynomial numb er of authen
ticated Byzantine agreement proto cols can b e concurrently
executed, while tolerating any numb er of faulty parties.
�
This work was carried out while the �rst and second au
thors were visiting IBM Research.
##### Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. STOC’02, May 19-21, 2002, Montreal, Quebec, Canada. Copyright 2002 ACM 1-58113-495-9/02/0005 ...$5.00.
is one of the most researched areas in distributed computing.
Numerous variations of the problem have b een considered
under di�erent communication mo dels, and b oth p ositive
results, i.e. proto cols, and negative results, i.e. imp ossibil
ity and lower b ounds on eÆciency and resources, have b een
established. The reason for this vast interest is the fact that
the Byzantine Generals problem is the algorithmic imple
mentation of a broadcast channel within a p oint-to-p oint
network. In addition to its imp ortance as a stand-alone
primitive, broadcast is a key to ol in the design of secure
proto cols for multiparty computation.
Despite the imp ortance of this basic functionality and the
vast amount of research that has b een directed towards it,
our understanding of the algorithmic issues is far from com
plete. As is evident from our results, there are still key
questions that have not yet b een addressed. In this pap er,
we provide solutions to some of these questions.
The problem of Byzantine Generals is (informally) de�ned
as follows: There are n parties, one of which is the General
who holds an input x. In addition, there is an adversary who
controls up to t of the parties and can arbitrarily deviate
from the designated proto col sp eci�cation. The (honest)
parties need to agree on a common value. Furthermore, if
the General is not faulty, then this common value must b e
his original input x.
Pease et al. [15, 13] provided a solution to the Byzantine
Generals problem in the standard mo del, i.e. the information
theoretic mo del with p oint-to-p oint communication lines (and
no setup assumptions). For their solution, the numb er of
faulty parties, t, must b e less than n=3. Furthermore, they
complemented this result by showing that the requirement
for t < n=3 is in fact inherent. That is, no proto col which
solves the Byzantine Generals problem in the standard mo del
can tolerate a third or more faulty parties.
The ab ove b ound on the numb er of faulty parties in the
standard mo del is a severe limitation. It is therefore of great
imp ortance to intro duce a di�erent (and realistic) mo del in
which it is p ossible to achieve higher fault tolerance. One
p ossibility involves augmenting the standard mo del such
that messages sent can b e authenticated. By authentica
tion, we mean the ability to ascertain that a message was
in fact sent by a sp eci�c party, even when not directly re
ceived from that party. This can b e achieved using a trusted
prepro cessing phase in which a public-key infrastructure for
digital signatures (e.g. [18, 11]) is set up. (We note that this
1
These two problems are essentially equivalent.
-----
requires that the adversary b e computationally b ounded.
However, there exist prepro cessing phases which do not re
quire any computational assumptions; see [16].) Indeed,
Pease et al. [15, 13] use such an augmentation and obtain a
proto col for the Byzantine Generals problem which can tol
erate any numb er of faulty parties (this is very dramatic con
sidering the limitation to 1/3 faulty in the standard mo del).
The Byzantine Generals problem in this mo del is referred to
as authenticated Byzantine Generals.
A common use of Byzantine Generals is to substitute a
broadcast channel. Therefore, it is clear that the settings
in which we would want and need to run it, involve many
invo cations of the Byzantine Generals proto col. The ques
tion of whether these proto cols remain secure when executed
concurrently, in parallel or sequentially is thus an imp ortant
one. However, existing work on this problem (in b oth the
standard and authenticated mo dels) fo cused on the security
and correctness of proto cols in a single execution only.
It is easy to see that the unauthenticated proto col of Pease
et al. [15], and other proto cols in the standard mo del, do
comp ose concurrently (and hence in parallel and sequen
tially). However, this is not the case with resp ect to authen
ticated Byzantine Generals. The �rst to notice that comp o
sition in this mo del is problematic were Gong, Lincoln and
Rushby [12], who also suggest metho ds for overcoming the
problem. Our work shows that these suggestions and any
others are futile; in fact comp osition in this mo del is imp os
sible (as long as 1/3 or more of the parties are faulty). (We
note that by comp osition, we refer to stateless comp osition;
see Section 2.3 for a formal discussion.)
Our Results. Our �rst theorem, stated b elow, shows that
authenticated Byzantine Generals proto cols, b oth determin
istic and randomized, cannot b e comp osed in parallel (and
thus concurrently). This is a surprising and p owerful state
ment with resp ect to the issue of enhancing the standard
mo del by the addition of authentication. The theorem shows
that this enhancement does not provide the ability to over
come the imp ossibility result when comp osition is required.
That is, if there is a need for parallel comp osition, then
the numb er of faulty players cannot b e n=3 or more, and
hence the authenticated mo del provides no advantage over
the standard mo del.
Theorem 1. No protocol for authenticated Byzantine
Agreement that composes in paral lel (even twice) can tol
erate n=3 or more faulty parties.
Regarding the question of sequential comp osition, we show
di�erent results. We �rst prove another (weaker) lower b ound
for deterministic proto cols:
Theorem 2. Let � be a deterministic protocol for au
thenticated Byzantine Agreement that terminates within r
rounds of communication. Then, � can be sequential ly com
posed at most 2r � 1 times.
In contrast, for randomized proto cols we obtain p ositive re
sults and present a proto col which can b e comp osed sequen
tially (any p olynomial numb er of times), and which toler
ates t < n=2 faulty parties. The proto col which we present
is based on a proto col of Fitzi and Maurer [8] that toler
ates t < n=2 faulty parties, and is in the standard mo del
augmented with an ideal three-party broadcast primitive.
We show that this primitive can b e replaced by an authen
ticated proto col for three parties that can b e comp osed se
quentially (and the resulting proto col also comp oses sequen
tially). Thus, we prove:
Theorem 3. Assume that there exists a signature scheme
that is existential ly secure against chosen message attacks.
Then, there exists a randomized protocol for authenticated
Byzantine Generals with a bounded number of rounds, that
tolerates t < n=2 faulty parties and composes sequential ly
any polynomial number of times.
We also present a randomized Byzantine Generals proto col
that tolerates any numb er of faulty parties, and comp oses
sequentially any p olynomial numb er of times. However, the
numb er of messages sent in this proto col is exp onential in
the numb er of parties. Therefore, it can only b e used when
the overall numb er of parties is logarithmic in the security
parameter of the signature scheme.
On the Use of Unique Session Identi�ers. As will b e
apparent from the pro ofs of the lower b ounds (Theorems 1
and 2), what prevents agreement in this setting is the fact
that honest parties cannot tell in which execution of the
proto col a given message was authenticated. This allows
the adversary to \b orrow" messages from one execution to
another, and by that attack the system. In Section 5, we
show that if we further augment the authenticate mo del so
that unique and common indices are assigned to each exe
cution, then security under many concurrent executions can
b e achieved (for any numb er of faulty parties).
Thus, on the one hand, our results strengthen the com
mon b elief that session identi�ers are necessary for achieving
authenticated Byzantine Generals. On the other hand, we
show that such identi�ers cannot be generated within the sys
tem. Typical suggestions for generating session identi�ers in
practice include having the General cho ose one, or having
the parties exchange random strings and set the identi�er
to b e the concatenation of all these strings. However, The
orem 1 rules out all such solutions (notice that just coming
up with a common identi�er involves reaching agreement).
Rather, one must assume the existence of some trusted ex
ternal means for coming up with unique and common in
dices. This seems to b e a very diÆcult, if not imp ossible,
assumption to realize in many natural settings.
A natural question to ask here relates to the fact that
unique and common session identi�ers are anyway needed
in order to carry out concurrent executions. In particular,
parties need to b e able to allo cate messages to proto col exe
cutions, and this requires a way of distinguishing executions
from each other. Indeed, global session identi�ers solve this
problem. However, it also suÆces for each party to allo cate
local identi�ers for itself. That is, when a party b egins a new
execution, it cho oses a unique identi�er sid and informs all
parties to concatenate sid to any message they send him
within this execution. It is then guaranteed that any mes
sage sent by an honest party to another honest party will b e
directed to the execution it b elongs to. We thus conclude
that for the purp oses of carrying out concurrent executions,
global identi�ers are not needed.
Implications for Secure Multiparty Computations.
As we have stated ab ove, one imp ortant use for Byzantine
Generals proto cols is to substitute the broadcast channel
in a multiparty proto col. In fact, most known solutions
-----
for multiparty computations assume a broadcast channel,
claiming that it can b e substituted by a Byzantine Generals
proto col without any complications. Our results therefore
imply that multiparty proto cols that rely on authenticated
Byzantine Generals to replace the broadcast channel, cannot
b e comp osed in parallel or concurrently.
Another imp ortant implication of our result is due to the
fact that any secure proto col for solving general multiparty
tasks can b e used to solve Byzantine Generals. Therefore,
none of these proto cols can b e comp osed in parallel or con
currently, unless more than 2=3 of the parties are honest or
a physical broadcast channel is available.
Our Work vs. Comp osition of Secure Multiparty
Proto cols. There has b een much work on the topic of
proto col comp osition in the context of multiparty computa
tion [1, 14, 2, 5, 3]. Much of this work has fo cused on zero
knowledge and concurrent zero-knowledge proto cols [10, 6,
17, 4]. For example, Goldreich and Krawczyk [10] show that
there exist proto cols that are zero-knowledge when executed
stand-alone, and yet do not comp ose in parallel (even twice).
However, proto cols that comp ose do exist (see, for example,
Goldreich [9] and references therein). In contrast, we show
that it is imp ossible to obtain any protocol that will comp ose
twice in parallel.
## 2. DEFINITIONS
2.1 Computational Model
versary (for whom the set of faulty parties is �xed b efore
the execution b egins).
Note that our proto cols for authenticated Byzantine Agree
ment that comp ose sequentially rely on the security of sig
nature schemes, and thus assume probabilistic p olynomial
time adversaries only. On the other hand, our imp ossibility
results hold for adversaries (and honest parties) whose run
ning time is of any complexity. In fact, the adversary that
we construct to prove our lower b ounds is of the same com
plexity as the honest parties.
## 2.2 Byzantine Generals/Agreement
The existing literature de�nes two related problems: Byzan
tine Generals and Byzantine Agreement. In the �rst prob
lem, there is one designated party, the General, who wishes
to broadcast its value to all the other parties. In the second
problem, each party has an input and the parties wish to
agree on a value, with a validity condition that if a ma jority
of honest parties b egin with the same value, then they must
terminate with that value. These problems are equivalent
in the sense that any proto col solving one can b e used to
construct a proto col solving the other, while tolerating the
same numb er of faulty parties. We relax the standard re
quirements on proto cols for the ab ove Byzantine problems
in that we allow a proto col to fail with probability that is
negligible in some security parameter. This relaxation is
needed for the case of authenticated Byzantine proto cols
where signature schemes are used (and can always b e forged
with some negligible probability). Formally,
We consider a setting involving n parties, P1
; : : : ; P
n
, that
Definition 1. (Byzantine Generals): Let P1
; : : : ; Pn�1
interact in a synchronous p oint-to-p oint network. In such
a network, each pair of parties is directly connected, and
it is assumed that the adversary cannot mo dify messages
sent b etween honest parties. In this setting, each party is
formally mo deled by an interactive Turing machine with n �
1 pairs of communication tap es. The communication of the
network pro ceeds in synchronized rounds, where each round
consists of a send phase followed by a receive phase. In the
send phase of each round, the parties write messages onto
their output tap es, and in the receive phase, the parties read
the contents of their input tap es.
This pap er refers to the authenticated mo del, where some
typ e of trusted preprocessing phase is assumed. This is mo d
eled by all parties also having an additional setup-tap e that
is generated during the prepro cessing phase. Typically, in
such a prepro cessing phase, a public-key infrastructure of
signature keys is generated. That is, each party receives its
own secret signing key, and in addition, public veri�cation
keys asso ciated with all other parties. (This enables parties
to use the signature scheme to authenticate messages that
they receive, and is thus the source of the name \authen
ticated".) However, we stress that our lower b ound holds
for all prepro cessing phases (even those that cannot b e eÆ
ciently generated).
In this mo del, a t-adversary is a party that controls t < n
and G = Pn
be n parties and let G be the designated party
with input x. In addition there is an adversary who may cor
rupt up to t of the parties including the special party G. A
protocol solves the Byzantine Generals problem if the fol low
ing two properties hold (except with negligible probability):
1. Agreement: Al l honest parties output the same value.
2. Validity: If G is honest, then al l honest parties output x.
We denote such a protocol by BGn;t
.
In the setting of Byzantine Agreement it is not straight
forward to formulate the validity prop erty. Intuitively, it
should capture that if enough honest parties b egin with the
same input value then they will output that value. By \hon
est," we mean the parties that follow the prescrib ed proto
col exactly, ignoring the issue that the �rst step of the party
might b e to change its lo cal input.
Definition 2. (Byzantine Agreement): Let P1
; : : : ; Pn
be n parties, with associated inputs x1
; : : : ; xn
. In addition
of the parties P1
; : : : ; Pn
, where the corruption strategy de
there is an adversary who may corrupt up to t of the parties.
Then, a protocol solves the Byzantine Agreement problem if
the fol lowing two properties hold (except with negligible prob
ability):
1. Agreement: Al l honest parties output the same value.
2. Validity: If max(n � t; bn=2c + 1) of the parties have the
same input value x and fol low the protocol speci�cation, then
al l honest parties output x.
We note that for the information-theoretic setting, the valid
ity requirement is usually stated so that it must hold only
when more than two thirds of the parties have the same
p ends on the adversary's view (i.e., the adversary is adap
tive). Since the adversary controls these parties, it receives
their entire views and determines the messages that they
send. In particular, these messages need not b e according
to the proto col execution, but rather can b e computed by
the adversary as an arbitrary function of its view. We note
that our imp ossibility results hold even against a static ad
-----
input value, b ecause in the information-theoretic setting,
n � t - 2n=3.
Authenticated Byzantine Proto cols: In the mo del for
authenticated Byzantine Generals/Agreement, some trusted
prepro cessing phase is run b efore any executions b egin. In
this phase, a trusted party distributes keys to every partic
ipating party. Formally,
Definition 3. (Authenticated Byzantine Generals and
Agreement): A protocol for authenticated Byzantine Gener
als/Agreement is a Byzantine Generals/Agreement protocol
with the fol lowing augmentation:
� Each party has an additional setup-tap e.
� Prior to any protocol execution, an ideal (trusted) party
� Run the preprocessing phase associated with � and ob
tain the strings s1
; : : : ; sn
. Then, for every j, set the
setup-tap e of P
.
j
to equal sj
� Repeat the fol lowing process a polynomial number of times
sequential ly (resp., in paral lel).
1. The adversary A chooses an input vector x
; : : : ; xn .
.
1
2. Fix the input tape of every honest Pj
to be xj
and
the random-tape to be a uniformly (and indepen
dently) chosen random string.
3. Invoke al l parties for an execution of � (using the
strings generated in the preprocessing phase above).
The execution is such that for i 2 I, the messages
chooses a series of strings s1 ; : : : ; sn
according to some
distribution, and sets party Pi
(for every i = 1; : : : ; n).
's setup-tap e to equal si
Fol lowing the above preprocessing stage, the protocol is run
in the standard communication model for Byzantine Gener
als/Agreement protocols.
As we have mentioned, a natural example of such a prepro
cessing phase is one where the strings s1
; : : : ; sn
constitute a
sent by party Pi are determined by A (who also sees
Pi 's view). On the other hand, al l other parties
fol low the instructions as de�ned in �.
We stress that the prepro cessing phase is executed only once
and all executions use the strings distributed in this phase.
Furthermore, we note that De�nition 4 implies that all hon
est parties are oblivious of the other executions that have
taken place (or that are taking place in parallel). This is
implicit in the fact that in each execution the parties are
invoked with no additional state information, b eyond the
contents of their input, random and key tap es. On the
other hand, the adversary A can co ordinate b etween the
executions, and its view at any given time includes all the
2
messages received in all other executions.
Before pro ceeding, we show that any Byzantine Generals
(or Agreement) proto col in the standard model comp oses
concurrently.
Proposition 2.1. Any protocol � for Byzantine Gener
als (or Agreement) in the standard mo del, remains secure
under concurrent composition.
Proof: We reduce the security of � under concurrent
comp osition to its security for a single execution. Assume
by contradiction that there exists an adversary A who runs
N concurrent executions of �, such that with non-negligible
probability, in one of the executions the outputs of the par
ties are not according to the requirements of the Byzantine
0
public-key infrastructure. That is, the trusted party cho oses
key-pairs (pk1
; sk1
); : : : ; (pkn
; skn ) from a secure signature
scheme, and sets the contents of party Pi
's tap e to equal
si
= (pk
1 ; : : : ; pki�1
; ski
; pki+1
; : : : ; pkn
). That is, all par
ties are given their own signing key and the veri�cation keys
of all the other parties.
We remark that the ab ove-de�ned prepro cessing phase is
very strong. First, it is assumed that it is run completely
by a trusted party. Furthermore, there is no computational
b ound on the p ower of the trusted party generating the keys.
Nevertheless, our imp ossibility results hold even for such a
prepro cessing phase.
## 2.3 Composition of Protocols
This pap er deals with the security of authenticated Byzan
tine Agreement proto cols, when the proto col is executed
many times (rather than just once). We de�ne the comp osi
tion of proto cols to b e stateless. This means that the honest
parties act up on their view in a single execution only. In par
ticular, this means that the honest parties do not store in
memory their views from previous executions or co ordinate
b etween di�erent executions o ccurring at the current time.
Furthermore, in stateless comp osition, there is no unique
session identi�er that is common to all participating par
ties. (See the Intro duction for a discussion on session iden
ti�ers and their role.) We note that although the parties are
stateless, the adversary is allowed to maliciously co ordinate
b etween executions and record its view from previous exe
cutions. Formally, comp osition is captured by the following
pro cess:
Definition 4. (sequential and parallel comp osition): Let
Generals. We construct an adversary A
who internally in
corp orates A and attacks a single execution of �. Intuitively,
0
A
simulates all executions apart from the one in which A
0
succeeds in its attack. Formally, A
b egins by cho osing an
th
index i 2R
f1; : : : ; N g. Then, for all but the i
0
execution of
the proto col, A
plays the roles of the honest parties in an
0
interaction with A (this simulation is internal to A
). On
externally inter
th
the other hand, for the i
0
execution, A
acts with the honest parties and passes messages b etween
them and A (which it runs internally). The key p oint in the
pro of is that the honest parties hold no secret information
(and do not co ordinate b etween executions). Therefore, the
0
simulation of the concurrent setting by A
th
for A is perfect.
Thus, with probability 1=N, the i
execution is the one in
0
which A succeeds. However, this means that A
succeeds
0
P1
; : : : ; Pn
be parties for an authenticated Byzantine Gener
in breaking the proto col for a single execution (where A
's
als/Agreement protocol �. Let I � [n] be an index set such
that for every i 2 I, the adversary A controls the party Pi
.
success probability equals 1=N times the success probability
of A.) This contradicts the stand-alone security of �.
2
The analogous de�nition for the comp osition of unauthenticated
Byzantine Generals/Agreement is derived from De�nition 4 by re
moving the reference to the prepro cessing stage and setup-tap es.
Over time, indices are added to I as the adversary chooses to
corrupt additional parties, with the restriction that jI j � t.
Then, the sequential (resp., parallel) comp osition of � in
volves the fol lowing process:
-----
## 3. IMPOSSIBILITY RESULTS
In this section we present two imp ossibility results regard
ing the comp osition of authenticated Byzantine Agreement
proto cols. Recall that we are concerned with stateless com
p osition. First, we show that it is imp ossible to construct an
authenticated Byzantine Agreement proto col that comp oses
in parallel (or concurrently), and is secure when n=3 or more
parties are faulty. This result is analogous to the Fischer et
al. [7] lower b ound for Byzantine Agreement in the standard
mo del (i.e., without authentication). We stress that our
result do es not merely show that authenticated Byzantine
Agreement proto cols do not necessarily comp ose; rather, we
show that one cannot construct proto cols that will comp ose.
Since there exist proto cols for unauthenticated Byzantine
Agreement that are resilient for any t < n=3 faulty parties
and comp ose concurrently, this shows that the advantage
gained by the prepro cessing step in authenticated Byzantine
Agreement proto cols is lost when comp osition is required.
Next, we show a lower b ound on the numb er of rounds re
quired for deterministic authenticated Byzantine Agreement
that comp oses sequentially. (Note that the imp ossibility of
paral lel comp osition holds even for randomized proto cols.)
We show that if an authenticated Byzantine Agreement pro
to col that tolerates n=3 or more faulty parties is to comp ose
sequentially r times, then there are executions in which it
runs for more than r =2 rounds. Thus, the numb er of rounds
in the proto col is linear in the numb er of times it is to com
p ose. This rules out any practical proto col that will comp ose
for a (large) p olynomial numb er of times.
Intuition. Let us �rst provide some intuition into why
the added p ower of the prepro cessing step in authenticated
Byzantine Agreement do es not help when comp osition is re
quired. (Recall that in the stand-alone setting, there exist
authenticated Byzantine Agreement proto cols that tolerate
any numb er of faulty parties. On the other hand, under par
allel comp osition, more than 2n=3 parties must b e honest.)
An instructive step is to �rst see how authenticated Byzan
tine Agreement proto cols typically utilize the prepro cess
ing step, in order to increase fault tolerance. A public-key
infrastructure for signature schemes is used and this helps
in achieving agreement for the following reason. Consider
three parties A, B and C participating in a standard (unau
thenticated) Byzantine Agreement proto col. Furthermore,
assume that during the execution A claims to B that C sent
it some message x. Then, B cannot di�erentiate b etween
the case that C actually sent x to A, and the case that C
did not send this value and A is faulty. Thus, B cannot b e
sure that A really received x from C . Indeed, such a mo del
has b een called the \oral message" mo del, in contrast to the
\signed message" mo del of authenticated Byzantine Agree
ment [13]. On the other hand, the use of signature schemes
helps to overcome this exact problem: If C had signed the
value x and sent this signature to A, then A could forward
the signature to B . Since A cannot forge C 's signature,
this would then constitute a pro of that C indeed sent x to
A. Therefore, utilizing the unforgeability prop erty of signa
tures, it is p ossible to achieve Byzantine Agreement for any
numb er of faulty parties.
However, the ab ove intuition holds only in a setting where
a single execution of the agreement proto col takes place.
Sp eci�cally, if a numb er of executions were to take place,
then A may send B a value x along with C 's signature on
x, yet B would still not know whether C signed x in this
execution, or in a di�erent (concurrent or previous) execu
tion. Thus, the mere fact that A pro duces C 's signature on
a value do es not provide pro of that C signed this value in
this execution. As we will see in the pro of, this is enough
to render the public-key infrastructure useless under some
typ es of comp osition.
We remark that it is p ossible to achieve concurrent comp o
sition, using state in the form of unique and common session
identi�ers. However, as we have mentioned, there are many
scenarios where this do es not seem to b e achievable (and
many others where it is undesirable).
Theorem 1. No protocol for authenticated Byzantine
Agreement that composes in paral lel (even twice) can tol
erate n=3 or more faulty parties.
Proof: The pro of of Theorem 1 is based on some of the
ideas used by Fischer et al. [7] in their pro of that no unau
thenticated Byzantine Agreement proto col can tolerate n=3
or more faulty parties. We b egin by proving the following
lemma:
Lemma 3.1. There exists no protocol for authenticated
Byzantine Agreement for three parties, that composes in par
al lel (even twice) and can tolerate one faulty party.
Proof: Assume, by contradiction, that there exists a
proto col � that solves the Byzantine Agreement problem
for three parties A, B and C, where one may b e faulty. Fur
thermore, � remains secure even when comp osed in parallel
twice. Exactly as in the pro of of Fischer et al. [7], we de
�ne a hexagonal system S that intertwines two indep endent
copies of �. That is, let A1 ; B1
, C1
and A2, B2
and C2
b e indep endent copies of the three parties participating in
�. By indep endent copies, we mean that A1
and A
2
are
the same party A with the same key tap e, that runs in two
di�erent parallel executions of �, as de�ned in De�nition 4.
The system S is de�ned by connecting party A1 to C2 and
B1 (rather than to C1 and B1 ); party B1 to A1 and C1 ;
party C1 to B1 and A2 ; and so on, as in Figure 1.
###### A1 B1
###### A
###### C2 1 S 0 C1
##### Π
###### B C
###### B2
###### A2
Figure 1: Combining two copies of � in a hexagonal
system S .
In the system S, parties A1
, B1
, and C
1
have input 0;
while parties A
2, B
2
have input 1. Note that within
and C2
S, all parties follow the instructions of � exactly. We stress
that S is not a Byzantine Agreement setting (where the
parties are joined in a complete graph on three no des),
and therefore the de�nitions of Byzantine Agreement tell
us nothing directly of what the parties' outputs should b e.
However, S is a well-de�ned system and this implies that the
parties have well-de�ned output distributions. The pro of
-----
pro ceeds by showing that if � is a correct Byzantine Agree
ment proto col, then we arrive at a contradiction regarding
the output distribution in S . We b egin by showing that B1
and C1 output 0 in S . We denote by rounds(�) the up
p er b ound on the numb er of rounds of � (when run in a
Byzantine Agreement setting).
Claim 3.2. Except with negligible probability, parties B1
and C1 halt within rounds(�) steps and output 0 in the sys
tem S .
Proof: We prove this claim by showing that there exists
a faulty party (or adversary) A who participates in two par
allel copies of � and simulates the system S, with resp ect to
1. Send outgoing messages of round i: A obtains mes
� In �2, A sends B2
sages msg
) and msg
(A
sages msg i (A1 ; B1 )
and messages msg
; B2
) from A1
; C
in �1,
and messages msg i (A2 ; B2 ) and msg i (A2 ; C2 ) from A2
in �2 (these are the round i messages sent by A1 and
A2 to the other parties; as we have mentioned, A1 and
A2 compute these messages according to the proto col
(A2 ;
msg i (A1 ; C1 )
) and msg
(A2 ;
2
de�nition and based on their view).
� In �1, A sends B1
the message msg
msg i (A1 ; B1 ) and
) (and thus the
sends C1 the message msg i (A2 ; C2 ) (and thus the
(A1 ; C1 ) directed edge is replaced by the directed
edge (A2
; C1
)).
the message msg
msg i (A2 ; B2 ) and
) (and thus the
B1
and C1 's view. The faulty party A (and the other honest
sends C2 the message msg i (A1 ; C1 ) (and thus the
(A2 ; C2 ) directed edge is replaced by the directed
parties participating in the parallel execution) work within
a Byzantine Agreement setting where there are well-de�ned
requirements on their output distribution. Therefore, by an
alyzing their output in this parallel execution setting, we are
able to make claims regarding their output in the system S .
edge (A1
; C2
)).
Let A1
, B
1
b e parties running an execution of �,
and C1
denoted �
, where B1
and C1
b oth have input 0. Further
1
more, let A
, B2
and C2
b e running a parallel execution of
2
�, denoted �2, where B2
and C2
b oth have input 1. Recall
that B1
and B2
are indep endent copies of the party B with
the same key tap e (as de�ned in De�nition 4); likewise for
C1
and C2 .
Now, let A b e an adversary who controls b oth A1 in �1
and A2
in �2
(recall that the faulty party can co ordinate
b etween the di�erent executions). Party A's strategy is to
maliciously generate an execution in which B1
1 's and C1
's
2. Obtain incoming messages from round i: A receives
messages msg i (B1 ; A1 ) and msg i (C1 ; A1 ) from B1 and
C1 in round i of �1, and messages msg i (B2 ; A2 ) and
msg i (C2 ; A2 ) from B2 and C2 in round i of �2 .
� A passes A1 in �1 the messages msg i (B1 ; A1 ) and
msg i (C2 ; A2 ) (and thus the (C1 ; A1 ) directed edge
is replaced by the directed edge (C2 ; A1 )).
� A passes A2 in �2 the messages msg i (B2 ; A2 ) and
msg i (C1 ; A1 ) (and thus the (C2 ; A2 ) directed edge
is replaced by the directed edge (C1 ; A2 )).
We now claim that B1 and C1 's view in �1 is identical to
3
view in �1
is identical to their view in S . A achieves this by
B1
and C1 's view in S .
This holds b ecause in the parallel
redirecting edges of the two parallel triangles (representing
the parallel execution), so that the overall system has the
same b ehavior as S ; see Figure 2.
execution of �
, all parties follow the proto col de�ni
1
and �2
tion (including A1
and A2
). The same is true in the system
S, except that party A1
is connected to B1
instead
and C1
###### B1
of to B1 and C1
instead of to B2
. Likewise, A
and C2
2
connected to B1 and C2
is connected to B2
###### A 1
###### B1
###### A 1
. However, by the de�nition of A,
###### 0
Π1
the messages seen by all parties in the parallel execution of
###### C2
###### 0
�1
and �2 are exactly the same as the messages seen by the
parties in S (e.g., the messages seen by C1 in �1
###### C2 1 S
###### C1
###### 1
###### 0
Π2
###### C1
are those
sent by B1
and A2
, exactly as in S ). Therefore, the views of
###### 1 1
B2 A2
###### A 2
B1
and C1
in the parallel execution maliciously controlled
4
by A, are identical to their views in S .
By the assumption that � is a correct Byzantine Agree
ment proto col that comp oses twice in parallel, we have that,
Figure 2: Redirecting edges of �1
hexagon.
###### B2
and �2
to make a
except with negligible probability, in �1
b oth B1
and C1
halt within rounds(�) steps and output 0. The fact that
they b oth output 0 is derived from the fact that B
and C1
1
Sp eci�cally, the (A1
; C1
) and (A2 ;
; C2
2 ) edges of �1
are an honest ma jority with the same input value 0. There
fore, they must output 0 in the face of any adversarial A1 ;
in particular this holds with resp ect to the sp eci�c adver
sp ectively are removed, and the (A1 ; C2
) and (A2
and �2 re
; C1 ) edges
of S are added in their place. A is able to make such a
mo di�cation b ecause it only involves redirecting messages
sary A describ ed ab ove. Since the views of B1
and C1
in S
to and from parties that it controls (i.e., A1
and A2 ).
are identical to their views in �1, we conclude that in the
Before pro ceeding, we present the following notation: let
msg i
th
(A1
; B
1
1 ) denote the message sent from A1
to B1
in the
system S, they also halt within rounds(�) steps and out
3
In fact, the views of al l the parties in the parallel execution with
A are identical to their view in the system S . However, in order to
i
i
round of the proto col execution. We now formally show
how the adversary A works. A invokes parties A1 and A2,
up on inputs 0 and 1 resp ectively. We stress that A1 and A2
obtain Claim 3.2, we need only analyze the views of B1
4
and C1 .
We note the crucial di�erence b etween this pro of and that of Fischer
follow the instructions of proto col � exactly. However, A
provides them with their incoming messages and sends their
outgoing messages for them. The only malicious b ehavior of
a single execution of � with B1
et al. [7]: the faulty party A is able to simulate the entire A1
�
C2 � B2
� A2
segment of the hexagon system S by itself. Thus, in
and C1
, party A can simulate the
hexagon. Here, due to the fact that the parties B2 and C2
have secret
A is in the redirection of messages to and from A1
and A2
. A
full description of A's co de is as follows (we recommend the
reader to refer to Figure 2 in order to clarify the following):
information that A do es not have access to, A is unable to simulate
their b ehavior itself. Rather, A needs to redirect messages from the
parallel execution of �
in order to complete the hexagon.
2
-----
put 0 (except with negligible probability). This completes
the pro of of the claim.
Using analogous arguments, we obtain the following two
claims:
Claim 3.3. Except with negligible probability, parties A2
and B2 halt within rounds(�) steps and output 1 in the sys
tem S .
In order to prove this claim, the faulty party is C and it works
in a similar way to A in the pro of of Claim 3.2 ab ove. (The
only di�erence is regarding the edges that are redirected.)
Claim 3.4. Except with negligible probability, parties A2
proto col. (More generally, r rounds of k parallel executions
of a proto col can b e simulated in k � r sequential executions.)
Thus, essentially, the deterministic sequential lower b ound
is derived by reducing it to the parallel comp osition case of
Theorem 1. That is,
Theorem 2. Let � be a deterministic protocol for authen
ticated Byzantine Agreement that concludes after r rounds
of communication. Then, � can be sequential composed at
most 2r � 1 times.
## 4. SEQUENTIALLY COMPOSABLE RAN- DOMIZED PROTOCOLS
In this section we present two results. The �rst one is a
proto col which tolerates any t < n=2 faulty parties and has
p olynomial communication complexity (i.e., bandwidth). The
second one is a proto col that can tolerate any numb er of
faulty parties but is exp onential in the numb er of partici
pating parties.
The building blo ck for b oth of the ab ove proto cols will b e
and C1
halt within rounds(�) steps and output the same
value in the system S .
Similarly, this claim is proven by taking the faulty party as B
who follows a similar strategy to A in the pro of of Claim 3.2
ab ove.
Combining Claims 3.2, 3.3 and 3.4 we obtain a contradic
a randomized (sequentially comp osable) proto col, ABG3;1
,
tion. This is b ecause, on the one hand C1
must output 0 in
S (Claim 3.2), and A2
must output 1 in S (Claim 3.3). On
the other hand, by Claim 3.4, parties A2
and C1
must out
put the same value. This concludes the pro of of the lemma.
Theorem 1 is derived from Lemma 3.1 in the standard way [15,
13] by showing that if there exists a proto col that is correct
for any n � 3 and n=3 faulty parties, then one can construct
a proto col for 3 parties that can tolerate one faulty party.
This is in contradiction to Lemma 3.1, and thus Theorem 1
is implied.
The following corollary, referring to concurrent comp osition,
is immediately derived from the fact that parallel comp osi
tion (where the scheduling of the messages is �xed and syn
chronized) is merely a sp ecial case of concurrent comp osition
(where the adversary controls the scheduling).
Corollary 1. No protocol for authenticated Byzantine
Agreement that composes concurrently (even twice) can tol
erate n=3 or more faulty parties.
Sequential Comp osition of Deterministic Proto cols.
We now show that there is a signi�cant limitation on de
terministic Byzantine Agreement proto cols that comp ose
sequential ly. Sp eci�cally, any proto col which terminates
within r rounds can only b e comp osed sequentially for at
most 2r � 1 times. The lower b ound is derived by show
ing that for any deterministic proto col �, r rounds of the
hexagonal system S (see Figure 1) can b e simulated in 2r
sequential executions of �. As we have seen in the pro of of
Theorem 1, the ability to simulate S results in a contradic
tion to the correctness of the Byzantine Agreement proto col
�. However, a contradiction is only derived if the system
S halts. Nevertheless, since � terminates within r rounds,
the system S also halts within r rounds. We conclude that
the proto col � can b e sequentially comp osed at most 2r � 1
times.
We remark that in actuality, one can prove a more gen
eral statement that says that for any deterministic proto
col, r rounds of 2 parallel executions of the proto col can b e
p erfectly simulated in 2r sequential executions of the same
for authenticated Byzantine Generals b etween 3 parties and
tolerating one faulty party. Recall that ABGn;t denotes an
authenticated Byzantine Generals proto col for n parties that
tolerates up to t faults. We �rst present the proto col ABG3;1
and then show how it can b e used to achieve the ab ove
describ ed results.
## 4.1 Sequentially Composable ABG3;1
For this proto col we assume three parties: the general, G,
As we will wish to incorp orate Proto col 3 into a proto col
with n parties we state a broader claim for the comp osition
than for a simple three party setting.
Lemma 4.1. Assume that the signature scheme � is ex
istential ly secure against adaptive chosen message attacks.
Then, Protocol 3 is a secure protocol for ABG3;1 that can be
composed sequential ly within a system of n parties, in which
t may become faulty, for any t < n.
and the recipients P1
; P2
. The General has an input value x.
According to De�nition 1, parties P1
0
and P2
need to output
0
the same value x
, and, if G is not faulty, then x
= x. As
is evident from the pro ofs of the imp ossibility, what hinders
a solution is that faulty parties can import messages from
previous executions, and there is no means to distinguish b e
tween those and the current messages. Thus, if some fresh
ness could b e intro duced in the signatures, then this would
foil the adversary's actions. Yet, agreeing on such freshness
would put us in a circular problem. Nevertheless, the case
of three parties is di�erent: here there are only two parties
who need to receive each signature. Furthermore, it turns
out that it suÆces if the parties who are receiving a signature
can jointly agree on a fresh string. Fortunately, two parties
can easily agree on a new fresh value: they simply exchange
messages and set the fresh string to equal the concatenation
of the exchanged values. Now, in the proto col which follows
for three parties, we require that whenever a party signs a
message, it uses freshness generated by the two remaining
parties. We note that in the proto col, only the General, G,
signs a message, and therefore only it needs a public key.
The proto col is describ ed in Figure 3. For simplicity, we as
sume that the signature scheme is de�ned such that �pk
also contains the value z .
(z )
-----
Intuitively, with probability 1=n, this is the party who plays
the general when A foils the agreement. For all other parties,
the forger F cho oses a key pair, for which it knows b oth the
signing and veri�cation keys. Then, F gives the adversary
A the key pairs for all the initially corrupted parties. F
now invokes A and simulates the roles of the honest parties
in the sequential executions of Proto col 3, with A as the
adversary. In particular, F works as follows:
� In all executions where the recipient/s P1
and/or P2
are not corrupted, F plays their role, following the
proto col exactly as sp eci�ed. This is straightforward
as the recipients do not use signing keys during such
an execution.
� In all executions where the general is some uncor
rupted party Pl
6= Pj,
, the forger F plays the role of Pl,
,
following the proto col and using the signing-key which
it asso ciated with Pl initially.
� In all executions where the general is the uncorrupted
Pj
, the forger F plays the role of Pj
following the pro
to col. However, in this case, F do es not have the as
so ciated signing-key. Nevertheless, it do es have access
to the signing oracle asso ciated with pk (which is Pj 's
public veri�cation-key). Therefore, F executes these
signatures by accessing its oracle. In particular, for
lab els `1
; `2
that it receives during the simulation, it
queries the signature oracle for �pk
(x; `1
; `
2 ).
� Corruptions: If at any p oint, A corrupts a party Pl
6=
Pj
, then F hands A the signing-key that is asso ciated
with Pl
(this is the only secret information that Pl
has). On the other hand, if at any p oint A corrupts
Proof: We prove the theorem by contradiction. Assume
Pj
, then F ab orts (and do es not succeed in forging).
that a series of ABG3;1
proto cols are run sequentially, such
that in some (or all) of them, the adversary succeeded in
foiling agreement with non-negligible probability. We will
show that in such a case, we can construct a forger F for
the signature scheme who succeeds with non-negligible prob
ability. This will then b e in contradiction to the security of
the signature scheme.
As there are n parties and the adversary can control up
to t of them, there may b e executions where two or three of
the parties are corrupted. However, in such a case, agree
ment holds vacuously. On the other hand, any execution in
which all three parties are honest must b e correct. There
fore, agreement can only b e foiled in the case that exactly
one participating party is corrupted.
We �rst claim that when A plays the General in an ex
ecution, it cannot foil the agreement. This is b ecause P1
Pj
Throughout the ab ove-describ ed simulation, F monitors each
execution and waits for an execution in which exactly one
party is corrupt and the agreement is foiled. If no such ex
ecution o ccurs, then F ab orts. Otherwise, in the �rst foiled
execution, F checks if the uncorrupted Pj
is the general in
this execution. If not, then F ab orts (without succeeding in
generating a forgery). Otherwise, we have an execution in
which Pj
is the general and agreement is foiled. In such a
case, F succeeds in generating a forgery as follows.
As we have mentioned, agreement can only b e foiled if
exactly one party is faulty. Since by assumption Pj
is not
corrupted, we have that one of the recipients P1
or P2
are
corrupted; without loss of generality, let P1
b e the corrupted
party. (We note that F plays the roles of b oth honest parties
and P2
in the simulation.) Now, since the agreement was
and P2
's views of the messages sent by A (playing G) are
foiled, we know that P2
do es not output P
j
's input value
identical. Furthermore, their decision making pro cess based
on their view is deterministic. Therefore, they must output
the same value. We stress that this is irresp ective of how
many executions have passed (and is also not dep endent at
all on the security of the signature scheme b eing used).
Thus, it must b e the case that the foiled execution is one
where the general is an honest party. As we have mentioned,
we build a forger F for the signature scheme � who uses
A. The forger F receives as input a public veri�cation
key pk, and access to a signing oracle asso ciated with this
key. F b egins by cho osing at random one of the parties, say
x, which means that it defaulted in Step 5. This can only
happ en if P2 received two valid signatures on the lab el `
which it sent Pj in this execution. Now, P2 clearly received
b ecause P2
a correct signature m on Pj 's input using the lab el ` from Pj
itself. (In fact, by the simulation, this signature is generated
by F accessing its signature oracle.) However, in addition,
0
P2
0
m
must have received a valid signature m
constitutes Pj
from P1, where
's signature on a string that contains lab el
0
` and a di�erent message x
. With overwhelming probabil
ity the lab el ` did not app ear in any previous execution,
is honest and cho oses its p ortion of the lab el
Pj
, and asso ciating the veri�cation-key pk with this party.
at random. Thus, previously in the simulation, the signing
-----
oracle was never queried with a string containing `. Further
0
more, by the assumption that x
6= x, the oracle query by
F in this execution was di�erent to the string up on which
0
m
0
is a signature. We conclude that m
is a valid signature
on a message, and that F did not query the signing oracle
0
with this message. Therefore, F outputs m
successful forgery.
and this is a
setting, the proto col can only b e carried out for n = log k
parties (where k is a security parameter). We stress that
the fact that the numb er of parties must b e logarithmic in
the security parameter is due to two reasons. First, we wish
the proto col to run in p olynomial time. Second, we use a
signature scheme and this is only secure for p olynomial-time
adversaries, and a p olynomial numb er of signatures.
Our proto col is constructed by presenting a transforma
tion that takes a sequentially comp osable ABG proto col for
n�1 parties which tolerates n�3 faulty parties, ABGn�1;n�3,
and pro duces a sequentially comp osable ABG proto col for
It remains to analyze the probability that F succeeds in
this forgery. First, it is easy to see that when F do es not
ab ort, the simulation of the sequential executions is perfect,
and that A's view in this simulation is identical to a real exe
.
cution. Furthermore, the probability that Pj
is the identity
n parties which tolerates n � 2 faulty parties, ABGn;n�2
of the (uncorrupted) general in the �rst foiled agreement
Then, given our proto col for broadcast among three parties
, we can apply our
equals 1=n exactly. The fact that Pj
is chosen ahead of
which tolerates one faulty party, ABG3;1
time makes no di�erence b ecause the simulation is p erfect.
Therefore, the choice of Pj
by F do es not make any di�er
transformation and obtain ABGn;n�2 for any n.
The idea for the transformation is closely related to the
ideas b ehind the proto col for Byzantine Generals for three
parties. The solution for the three-party broadcast assumes
two-party broadcast (which is trivial). Using two-party broad
cast, agreement on a fresh lab el can b e reached. Having
agreed on this lab el, the two p oint communications with
the General are suÆcient. Each party sends its claimed
fresh lab el to the General, and the General includes the two
received lab els inside any signature that it pro duces. Our
general transformation will work in the same manner. We
ence to the b ehavior of A. We conclude that F succeeds
in forging with probability 1=n times the probability that A
succeeds in foiling agreement (which is non-negligible). This
contradicts the security of the signature scheme.
## 4.2 Sequentially Composable ABGn;n=2
Fitzi and Maurer [8] present a proto col for the Byzan
tine Generals problem that tolerates any t < n=2 faulty
parties. Their proto col is for the information-theoretic and
unauthenticated mo del. However, in addition to the p oint
to-p oint network, they assume that every triplet of parties
is connected with an ideal (3-party) broadcast channel. As
we have shown in Section 4.1, given a public-key infrastruc
ture for signature schemes, it is p ossible to implement secure
broadcast among three parties that comp oses sequentially.
use the ABGn�1;n�3
proto col to have all parties (apart from
Thus, a proto col for ABGn;n=2
is derived by substituting
the General) agree on a random lab el. Then, each party pri
vately sends this lab el to the General, who then includes all
lab els in its signatures. Thus, we prove:
Theorem 4. Assume that there exists a signature scheme
that is existential ly secure against chosen message attacks,
for adversaries running in time p oly (k ). Then, there exists a
Byzantine Generals protocols for O (log k ) parties, that toler
ates any number of faulty parties and composes sequential ly.
The formal description of the proto col and the pro of of the
theorem are omitted due to lack of space in this abstract.
## 5. AUTHENTICATED BYZANTINE AGREE- MENT USING UNIQUE IDENTIFIERS
In this section we consider an augmentation to the authen
ticated mo del in which each execution is assigned a unique
and common identi�er. We show that in such a mo del, it
is p ossible to achieve Byzantine Agreement/Generals that
comp oses concurrently, for any numb er of faulty parties. We
stress that in the authenticated mo del itself, it is not p ossi
ble for the parties to agree on unique and common identi
�ers, without some external help. This is b ecause agreeing
on a common identi�er amounts to solving the Byzantine
Agreement problem, and we have proven that this cannot b e
achieved for t � n=3 when comp osition is required. There
fore, these identi�ers must come from outside the system
(and as such, assuming their existence is an augmentation
to the authenticated mo del).
Intuitively, the existence of unique identi�ers helps in the
authenticated mo del for the following reason. Recall that
our lower b ound is based on the ability of the adversary
to borrow signed messages from one execution to another.
Now, if each signature also includes the session identi�er,
then the honest parties can easily distinguish b etween mes
sages signed in this execution and messages signed in a dif
ferent execution. It turns out that this is enough. That is,
the ideal 3-party broadcast primitive in the proto col of Fitzi
and Maurer [8] with Proto col 3. Since Proto col 3 and the
proto col of Fitzi and Maurer [8] b oth comp ose sequentially,
the resulting proto col also comp oses sequentially.
Theorem 3 Assume that there exists a signature scheme
that is existential ly secure against chosen message attacks.
Then, there exists a randomized protocol for authenticated
Byzantine Generals that tolerates t < n=2 faulty parties and
composes sequential ly any polynomial number of times.
As we show in Section 5, it is p ossible to execute many
copies of an authenticated Byzantine Generals proto col con
currently, by allo cating each execution a unique identi�er
that is common and known to all parties. Now, inside the
Fitzi-Maurer proto col we can allo cate unique indices to each
invo cation of the ABG3;1
proto col. We can therefore run the
ABG3;1
proto cols in parallel (rather than sequentially), im
proving the round complexity of the resulting proto col. In
particular, our proto col is of the same round complexity as
the underlying Fitzi-Maurer proto col. (We stress that the
fact that the ABG3;1 subproto cols can b e executed in par
allel within the ABGn;n=2 proto col does not imply that the
ABGn;n=2 proto col itself can comp ose in parallel. Rather,
by our imp ossibility result, we know that it indeed cannot
b e comp osed in parallel.)
## 4.3 Sequentially Composable ABGn;t [for any] t
In this section we describ e a proto col for the Byzantine
Generals problem for n parties, which can tolerate any num
ber of faulty parties. However, this proto col is exp onential
in the numb er of participating parties. Therefore, in our
-----
we give a transformation from almost any Byzantine Agree
ment proto col based on signature schemes, to a proto col
that comp oses concurrently when unique identi�ers exist.
By \almost any proto col," we mean that this transforma
tion applies for any proto col that uses the signature scheme
for signing and verifying messages only. This is the natural
use of the signature scheme and all known proto cols indeed
work in this way.
More formally, our transformation works as follows. Let
� b e a proto col for authenticated Byzantine Agreement. We
de�ne a mo di�ed proto col �(id) that works as follows:
� Each party is given the identi�er id as auxiliary input.
� If a party Pi
has an instruction in � to sign a given
message m with its secret key ski
, then Pi
signs up on
## Acknowledgments
We would like to thank Oded Goldreich for p ointing out a
simpler pro of of Theorem 5. And Matthias Fitzi for discus
sions ab out [8].
## 7. REFERENCES
[1] D. Beaver. Secure multiparty proto cols and
zero-knowledge pro of systems tolerating a faulty
minority. Journal of Cryptology, 4:75{122, 1991.
[2] R. Canetti. Security and comp osition of multiparty
cryptographic proto cols. Journal of Cryptology,
13(1):143{202, 2000.
[3] R. Canetti. Universally Comp osable Security: A New
Paradigm for Cryptographic Proto cols. In 42st FOCS,
pages 136{145. 2001.
[4] R. Canetti, J. Kilian, E. Petrank, and A. Rosen.
Black-Box Concurrent Zero-Knowledge Requires
Omega(log n) Rounds. In 33th STOC, pages 570{579.
2001.
[5] Y. Do dis and S. Micali. Parallel Reducibility for
Information-Theoretically Secure Computation. In
Crypto '00, pages 74{92, 2000. LNCS No. 1880.
[6] C. Dwork, M. Naor, and A. Sahai. Concurrent
zero-knowledge. In 30th STOC, pages 409{418. 1998.
[7] M. Fischer, N. Lynch, and M. Merritt. Easy
Imp ossibility Pro ofs for Distributed Consensus
Problems. Distributed Computing, 1(1):26{39, 1986.
[8] M. Fitzi and U. Maurer. From partial consistency to
global broadcast. In 32th STOC, pages 494{503. 2000.
[9] O. Goldreich. Concurrent Zero-Knowledge With
Timing Revisited. In 34th STOC. 2002.
[10] O. Goldreich and H. Krawczyk. On the comp osition of
zero-knowledge pro of systems. SIAM. J. Computing,
25(1):169{192, 1996.
[11] S. Goldwasser, S. Micali, and R. L. Rivest. A digital
signature scheme secure against adaptive
chosen-message attacks. SIAM J. Computing,
17(2):281{308, Apr. 1988.
[12] L. Gong, P. Lincoln, and J. Rushby. Byzantine
Agreement with Authentication: Observations and
Applications in Tolerating Hybrid and Link Faults. In
Dependable Computing for Critical Applications, 1995.
[13] L. Lamp ort, R. Shostack, and M. Pease. The
Byzantine generals problem. ACM Trans. Prog. Lang.
and Systems, 4(3):382{401, 1982.
[14] S. Micali and P. Rogaway. Secure computation. In
Crypto '91, pages 392{404, 1991. LNCS No. 576.
[15] M. Pease, R. Shostak, and L. Lamp ort. Reaching
Agreement in the Presence of Faults. Journal of the
ACM, 27(2):228{234, 1980.
[16] B. P�tzmann and M. Waidner. Information-Theoretic
Pseudosignatures and Byzantine Agreement for
t >= n=3. Technical Rep ort RZ 2882 (#90830), IBM
Research, 1996.
[17] R. Richardson and J. Kilian. On the concurrent
comp osition of zero-knowledge pro ofs. In Eurocrypt
'99, pages 311{326, 1999. LNCS No. 1592.
[18] R. L. Rivest, A. Shamir, and L. M. Adleman. A
metho d for obtaining digital signatures and public-key
cryptosystems. Communication of the ACM,
21(2):120{126, 1978.
id Æ m instead (where Æ denotes concatenation).
� If a party Pi has an instruction in � to verify a given
signature � on a message m with a public key pkj, then
Pi
veri�es that � is a valid signature for the message
id Æ m.
We now state our theorem:
Theorem 5. Let � be a secure protocol for authenticated
Byzantine Agreement which uses an existential ly unforge
able signature scheme. Furthermore, this scheme is used
for generating and verifying signatures only. Let the pro
tocol �(id) be obtained from � as described above, and let
id1
; : : : ; id`
be a series of ` unique strings. Then, the pro
tocols �(id
) al l solve the Byzantine Agreement
1
); : : : ; �(id`
problem, even when run concurrently.
We conclude by noting that it is not at all clear how such an
augmentation to the authenticated mo del can b e achieved
in practice. In particular, requiring the on-line participation
of a trusted party who assigns identi�ers to every execution
is clearly impractical. (Furthermore, such a party could just
b e used to directly implement broadcast.) However, we do
note one imp ortant scenario where Theorem 5 can b e ap
plied. As we have mentioned, secure proto cols often use
many invo cations of a broadcast primitive. Furthermore, in
order to improve round eÆciency, in any given round, many
broadcasts may b e simultaneously executed. The key p oint
here is that within the secure proto col, unique identi�ers can
b e allo cated to each broadcast (by the proto col designer).
Therefore, authenticated Byzantine Agreement can b e used.
Of course, this do es not change the fact that the secure
proto col itself will not comp ose in parallel or concurrently.
However, it do es mean that its security is guaranteed in the
stand-alone setting, and a physical broadcast channel is not
necessary.
## 6. OPEN PROBLEMS
Our work leaves op en a numb er of natural questions. First,
an unresolved question is whether or not it is p ossible to
construct randomized proto cols for authenticated Byzantine
Generals that sequentially comp ose, for any n and any num
b er of faulty parties. Second, it is unknown whether or not
it is p ossible to construct a deterministic proto col that ter
minates in r rounds and sequentially comp oses ` times, for
some 2 � ` � 2r � 1. Another question that arises from this
work is to �nd a realistic computational mo del for Byzantine
Agreement that does allow parallel and concurrent comp o
sition for n=3 or more faulty parties.
-----
|
{
"disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references', 'abstract'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1145/509907.509982?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1145/509907.509982, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "GREEN",
"url": "http://www.cs.brown.edu/research/pubs/pdfs/2002/Lindell-2002-CAB.pdf"
}
| 2,002
|
[
"JournalArticle"
] | true
| 2002-05-19T00:00:00
|
[] | 18,021
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Mathematics",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/02318c72964c65db5dc32b3997968c673c3bef9a
|
[
"Computer Science"
] | 0.8153
|
Private-Key Algebraic-Coded Cryptosystems
|
02318c72964c65db5dc32b3997968c673c3bef9a
|
Annual International Cryptology Conference
|
[
{
"authorId": "2093398212",
"name": "T. Rao"
},
{
"authorId": "72321938",
"name": "Kil-Hyun Nam"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"Int Cryptol Conf",
"Annu Int Cryptol Conf",
"CRYPTO",
"International Cryptology Conference"
],
"alternate_urls": null,
"id": "212b6868-c374-4ba2-ad32-19fde8004623",
"issn": null,
"name": "Annual International Cryptology Conference",
"type": "conference",
"url": "http://www.iacr.org/"
}
| null |
been studied previously. Private-key cryptosystems `using` simpler codes have _,
###### PRIVATEKEY
ALGEBRAIC-CODED
```
CRYFTOSYSTEMS *
T. R. N. Rao Kil-Hyun Nam **
```
The Center `for Advanced Computer Studies` National Defense College
University of Southwestern Louisiana Seoul, Korea
Lafayette, Louisiana **70504**
```
ABSTRACT
```
Public-key cryptosystems using very large distance algebraic codes have
been studied previously. Private-key cryptosystems `using` simpler codes have _,
**also** been subject of some study recently. `This paper proposes` a new ap-
proach to the private-key cryptosystems which allows use of very simple codes
such `as distance` `3 and 4` Hamming codes.
This new approach gives not only very efficient encoding/decoding and
very high information rates but also appears to be secure even under chosen-
plaintext attacks.
Keywords : cryptosystems, public-key cryptosystems, private-key cryptosp-
tems, Algebraic codes, crypt-complexity, chosen-plaintext attack,
Joint Encryption and Error-control `Coding`
###### * This research is supported by a grant from National Security Agency grant
# MDA 904-84-H-005.
** This author's research was performed while he was with the Center for Ad-
vanced Computer Studies at University of Southwestern Louisiana.
-----
`1.` **INTRODUCTION**
McEliece introduced a public-key cryptosystem based on algebraic coding
theory using t-error correcting Goppa codes [McEliece **’781.** But McEliece
Public-key Cryptosystem (MPBC) requires large block lengths with capabilities
to correct large number of errors (n = 1000 bits, t **x 50** bits) to be effective.
This involves very large computational (encryption and decryption) overhead
to be practical in computer communications.
Private-key Algebraic-coded Cryptosystems (PRAC) were suggested by
###### Rao [Rao ’84b] using the same techniques as MPBC but keep the public gen-
erator matrix `as` private. PRAC provides better security with simpler error
correcting codes, hence, requires relatively low computational overhead. Howev-
er, we show that PRAC can be broken easily by a chosen-plaintext attack.
Both MPBC and PRPLC are classified `as` Algebraic-Coded Cryptosystems
(ACC) here.
This paper introduces a new approach to PRAC, which requires simple
error correcting codes (i.e. distance **3** codes) and also provides much higher
security level.
###### 1.1. McEliece Public-key Cryptosystems (MPBC)
Encryption
Let G be a t-error correcting **k*n** generator matrix of a linear code over
GF(2) capable of t-error correction. The rate of the code is -. k We can
**n**
select a random `k*k nonsingular matrix` S called scrambler and a random n*n
permutation matrix P. Having G, `S and` P, we can compute the public gen-
erator matrix G’ such that G’ = SGP, which is combinatorially equivalent to
G.
Then the encryption **is** done by:
C = MG’ + `Z`
where `C : ciphertext` of length n,
-----
M : plaintext message of length `k,`
Z : random error vector of length n with weight t.
Note that the vectors are italic lettered, and weight meam Hamming weight.
**Decryption**
The decryption is very straight forward.
###### Fro= the encryption equation
G' = SGP
###### c = MG' i- z
= MSGP + Z = M' GP + Z where M' = MS
Hence, we can recover M `as given by the following steps.`
###### Step 1 compute C' :
`C'` = C P T = M ' G + Z P T
###### = M ' G + 2' where Z' = Z P T
(Note: `2'` `has same weight` `as Z` since
###### P and P T are permutation matrices)
Step 2 Decoding and error correction:
(Patterson Algorithm [MCEL **771).**
###### Step 3 recover plaintext M :
_M = M ' S - '_
**Cryptanalysis** `of MPBC`
`As suggested by McEliece in his paper [McEliece '781,` there could be two
kinds of basic attacks for the cryptanalyst to try.
(a) Factoring S, G and P from G'
Since the number of codes which are combinatorially equivalent to a
given code is astronomical, it is hopeless `task to` find out exact keys s,
G and P used for G'. However, the cryptanalyst needs only some
-----
`Si` Gi and
code. For the given _G’,_ the cryptanalyst can obtain `Sj , G,` and Pi
satisfying the equation `S, G,` P, = G’, where G, is a generator in `s y ~ -`
tematic form. G, is obtained from G’ by.elementary row operations
(row canonical reduction) and column operations. G’, G, `and G` are all
said to be combinatorially equivalent. Where `as G` corresponds directly
to a Goppa code which has well understood and well-known decoding
algorithms, `no such would be possible for` `G,. Trial and error manipu-`
lation to obtain a G, coinciding with an equivalent Alternant code
generator would require an astronomically large work factor.
(b) Recovering **_M_** from `C directly without keys`
Another approach involves solving a set of k-unknowns from `n simul-`
taneous equations for all possible `Z values.`
Let `M` and `C be a plaintext pair`
**_M_** = **_m l m 2 m 3_** . . _ mk
###### c = c 1 c 2 c 3 . . . ck . . . cn
z = z l b 2 z 3 . . . t k . . . Zn
G’ = [ Gij’ 1 i = 1, ... , k
j = 1, ... , n
(t-error correcting algebraic code)
Then, for `j=` `1, ... , n`
**_C 1_** = m1G11’ **_fm2G21’_** +...+ mkGk{ **_+ Z I_**
**_C 2_** = m1G12’ **_f m z G 2 2 ’_** +...+ mkGk$ **_+_** **_~_** **_2_**
To solve `k unknowm` `(rn l,m2,` . . . , mk ), k operations are required
because `k equations are sufficient` to solve the equations if the code
`is` maximal distance separable (MDS) code. Otherwise, at most
k’ = n-d+l equations are required to solve for k-unknowns [Pless **,821.**
-----
Since t is smaller than n-k, it is possible that the cryptanalyst could
select k equations containing `no errom from` n equations. Therefore,
the cryptanalyst could repeat solving equations by selecting arbitrary k
equations from n simultaneous equations with- the assumption of `no er-`
```
rors in selected equations until a meaningful plaintext is obtained.
```
The probability of `no errors in` `k` equations, `is:`
and the average number of repetition is `Pk-'.`
Hence, the average work factor, T is:
###### T = k 3 * Pk-1
However, this does not include the work factor to check whether the
plaintext **_M_** obtained by solving equations is correct (Le., meaningful)
or not. It is assumed that the plaintexts are from a source such `as`
natural language or a programming language which contains an enor-
###### mous amount of redundancy penning '821. Redundancy in M heIps to
determine the validity of the plaintext derived.
**1.2.** **Private-key** `Algebraic-coded Cryptosystems (PF2AC)`
###### To increase information rate and to reduce computational (encryption and
decryption) overhead of `MPBC, Private-key Algebraic-coded Cryptovstem`
###### (PRAC) were suggested [Rao '84b]. PRAC can provide better security with .
simpler error correcting codes, hence, require relatively low computational over-
head compared to MPBC.
PRAC keeps G' private `as well` `as` S, P and G to provide higher security
level. **A** known-plaintext attack to PFL4C is feasible by solving matrices for
each column vector of G' independently but this method requires a very large
set of known **_( M , C ) pairs. Hence, this attack can be foiled_** by periodic change
or modification of the keys by the cryptographer. However, the analysis given
below shows that PRAC still requires large t to be secure from a chosen-
plaintext attack.
-----
**Chosen- Plaint ext** **Attack**
The cryptanalyst is required to go through two steps.
Step 1 : Solve for G’ from a large set of _( M , C ) pairs._
Step **2** : Determine _M_ from `C using` G’ obtained in Step 1 (same **work** fac-
tor `as in MPBC).`
It can be safely assumed that a chosen plaintext of the form _M_ = (00 - -
`.010 . . .O)` with only one `1 in` ith position (for i = 1, . . . ,k) is not allowed
by the cryptosystem. However, a chosen-plaintext attack may proceed `as fol-`
lows.
Let **M1 and** `M 2 are two plaintext differing in one p i t i o n only, that` is,
_M , -_ **_M ,_** = `(00 . . . 010` . . . 0)
##### ith position for i = 1, . . . ,k
then,
_C, -_ `C,` = **_gj’_** + ( 2 1 - `22)` **(Es. 1)**
where **_gi’_** `is` the ith row vector of G’.
The Hamming weight of _(2, -_ _2,)_ is at most 2t. Since t is much smaller than
n, the majority of the `bits of` the vector C, - C2 correspond directly **with** **_9;’_** .
We can let C1 - C2 represent one estimate of **_gj’_** . By repeating the step
several times a number of estimates of **g;’** can be obtained. From these esti-
mates of **_g j ’_** and by majority voting for each position, the vector **_gj’_** `can be`
correctly determined. `This step repeated for` all i = 1’2,. . . .k will give us G’,
which can be used to break the code by step 2. This step **2** will require a re-
latively small work factor because t is small.
However, a chosen-plaintext attack of the above nature can succeed only when
t
n - is small and it **will** not if t - - 2’
-----
**2.** **MODIFIED CRYPTOSYSTEMS**
**2.1.** **Introduction**
Our intent here is to obtain private-key cryptosystems using simple alge-
braic codes such `as Hamming codes` or distance **_5_** `BCH codes. Furthermore,`
we would still want the _Z_ vector to have a weight t sufficiently large to prc-
vide good security. By a clever design we will show that we could obtain t
###### e L. Obviously it would not be possible unless we change or modiy the origi-
**2**
nal encryption method.
Here we develop such a modification and show that it is indeed possible
to use simple (i-e., short distance) algebraic codes for `PRAC which are very`
secure from chosen-plaintext attacks. Clearly a system that is secure from such
`an attack` is **also** secure from other attacks including known-plaintext attacks.
**2.2.** **Encryption** **of** **Modified** **PRAC**
This approach uses a minimum distance **3** code generator G `(as an` exam-
ple) and uses specific error patterns for the random error vector Z of which
the average `Hamming` weight is approximately `f.` Encryption method- ‘L3
**2**
modified `as follow.`
Let G’ = SG
where S : k*k nonsingular matrix
G : k*n distance **3** code generator matrix
G’: k*n encryption matrix
Then
C = _(MG’ + Z)P_
where _M_ : plaintext of length `k`
C : ciphertext of length n
###### P : n*n permutation matrix
_Z_ : a random ATE (Method **1)**
or an entry of the Syndrome-error table (Method 2)
-----
(Method `1 and` **2** are destribed below.)
Since the security of `PRAC crucially depends` on the weight of _2 , the selec-_
tion of Z is very important. We introduce two kinds of error patterns.
###### Method 1 : Use adjacent t errors for Z.
Definition 1 : Adjacent t Errors (ATE)
**An** ATE `is` a vector of length n with t `(5 i) adjacent errors, i.e.,`
an ATE consists of n-t 0’s and t consecutive 1’s. ATE must not
be a codeword.
**A** random ATE can be used for _2 ._ There exist exactly n-t+l ATE’s
for the given n and t (and n ATE’s for cyclic codes).
###### Method 2 : Use of predetermined set of vectors (Syndrome-error table).
A predetermined set of vectors consisting one from each coset of the
standard array decoding table can be used for Z. Each coset has a
distinct syndrome and there are exactly **2n-k** cosets [Blahut `‘83, Lin`
**’831.** Therefore, we could select any set of vectors one from each of
the `2n“` cosets. The set is predetermined in the sense the decryptor
knows the Syndrome-error table used for `2.` Fig.1 shows an example
of standard array and Syndrome-error table. The vectors in the rec-
tangular `boxes are selected` `as z` - vectors.
0 1 1 1 0 0
Coset leader Syndrome
000000 I 001110 010101 limiil 011011 101101 110110 111000 000
```
-A-
```
000001 I 001111 010100 100010 r-1 101100 110111 111001 001
000010 I 001100 010111 100001 011001 101111 im1 111010 010
000100 I 001010 010001 100111 011111 (lolooll 110010 111100 100
001000 I 000110 101011 010011 100101 111110 110000 110
010000 I 011110 000101 m l 001011 111101 100110 101000 101
100000 1 110101 000011 111011 001101 010110 011000 011
## [m-
001001 I 000111 011100 101010 010010 100100 111111 plooorl 111
Fig. 1. Standard array for the (6, 3, 3) code
-----
G and P are secret encryption keys, and the Syndrome-error table is also
secret in the Method **2.**
**2.3.** **Decryption** **of** `Modified Cryptosystems`
From the encryption algorithm `(Eq. 2)`
###### c = (MG' + z)P
= M S G P + ZP = M ' G P + ZP. (M' = M S )
Decryption can be done using secret keys S-l, `HT (GHT` = 0) and p
through following steps.
###### Step 1 Obtain C' :
C' = c p = = M ' G f Z
###### Step 2 : Find the error pattern and recover M' :
`C' HT =` `M' GHT` + ZHT
###### = ZHT (Syndrome)
Identify the error pattern.
(use the Syndrome-error table look-up for the Method **2).**
Recover `M'` by correcting for the error pattern.
###### Step 3 Recover plaintext M :
_M_ = _M'_ s-1
###### Note : It appears that this approach requires long keys (S, P, G and the
Syndrome-emr table for the Method **2).** However, the keys could be
generated by using a pseudo-random number generator algorithm. In
that case the user may require only short seeds for keys S, P and the
Syndromeerror table. This problem is not addressed here and it would
be a topic for future work.
-----
**2.4.** **Application to JOEEC**
Recently Joint Encryption and Error-control Coding (JOEEC) **was** sug-
gested `pi-.. ’84a].` This approach combines data .encryption and error-control
coding steps into one step to gain speed and efficiency in implementation.
The modified cryptosystems could also be implemented `as JOEEC by` us-
ing higher distance codes. But the application to JOEEC of this approach `is`
presently being studied.
`3.` **CRYPTANALYSIS** `OF MODIFIED` **CRYPTOSYSTEMS**
The encryption algorithm **(Eq.** **2)** can be rewritten `as follows.`
```
C = (MG’ + Z ) P
###### = MG” + Z P
```
where G” = G’P = [ **g;”** ] for i = 1, ... ,k,
###### and gin is a row vector.
The following lemmas help us to establish the high level of security **pro-**
vided by this new approach.
###### Lemma 1 : The number of P’s that transform ATE’S into non-ATE’S is at
least (n - - `I)!` if 2 < t <_ _2_ where n is the length of
**2 ’**
###### ATE and t is the length of adjacent errors.
Outline of Proof: Let vector V be an ATE of length n. We select a set of
positions, (1, **2,** t, 2t, ..., bt}, from V where b = 121. We reorder
**_t_**
these positions as an ordered set, B = (1, t, 2t, ..., bt, **2).** This map-
ping is illustrated in the figure below.
###### v- = I-+--+ ---- + +--- - - - ---+____I
(ATE)
**1** **2** t 2t 3t bt n b =
###### B = (1, t, 2t, . . . , bt, 2)
V’ = I-- - - - ---+ ____ B_--+ ______ - - - _-_I (non-ATE)
We consider a permutation map of vector V to V’ with B embedded in
V’ The purpose is to make V’ a non-ATE This is achieved because B
-----
0 s.
any position of V' and therefore, we have n-b-1 choices. In addition the
number of permutations possible for V --> V of the remaining positions
is (n-b-2)!. Thus the total number of permutations of transforming an
ATE vector V to non-ATE vector V', N, can be shown to be at least
###### Np = (n-b-1) * (n-b-2)!
= (n-b-l)! QED.
This formula gives a lower bound for N, of (n - **3)!** when t = Lt].
###### Lemma 2 : The number of code generators combinatorially equivalent to a
(n, k, 3) code generator is at least `k!.`
Proof: Let G be a `(n, k,` **3)** code generator in systematic form.
###### G = [Ik Pk,n-k 1
where `Ik` `is` an identity matrix and
P k ,n -k is a parity check matrix.
Then, there are k! row combinations of parity check matrix, which are
distinct `(n, k, 3) code generators also.` All of these code generators can
be obtained by row exchange and column permutation of G, `and hence,`
are combinatorially equivalent to G [Peterson '721.
###### Lemma 3: The number of k*k non-singular matrices over GF(2), Ns is given
by
Proof: We can start with any non-zero vector for the first row of non-
singular matrix S and we have **_zk_** - 1 choices. The second row must be
linearly independent of the first. That is **we** have **2"** - 2 choices for the
second row. For the third row the choice is any vector linearly indepen-
dent of the first two. Clearly it has **_(zk - 27_** choices. Continuing this
way, the number of non-singular matrices are given by the equality (Eq.
**3).** Since there are k terms in the product, the smallest of which is **_Z k - l ,_**
-----
**An** attack by exhaustive search for S, G and P is considered hopeless task due
to the results of above Lemmas. The previously described method of the
chosen-plaintext attack (described in Section **1.2.1)' can not be applied here be-**
**n**
cause the average Hamming weight of (2,-Z,)P is about **_2, which_** is very
large. Therefore, we have to look for a different method to cryptanalysb and
it could be `as follows.`
Let `Cj and` c k be two distinct ciphertexts obtained for the same plaintext `M.`
Then `Cj =` _MG"_ + Zip
`Ck` = _MG"_ + ZkP
###### cj - c, = (Zj -&)P
The above step provides `one value` for _(Zj - zk)P._ This step needs to be re-
peated until all possible pairs of 2's are used. The number of distinct 2's is
given by
###### N = 2L for the Method 1,
**2**
###### > - n for the Method 2;
```
N *-N
```
`and the number` of possible distinct values of **(Zi** `-zj)P is` -.
**2**
**An** expression for **gin** by a computation `as described in Section` 1.2.1 `is` given
by
**_C,-Cz_** = **gin + (Z1-ZJP**
**_g i n_** = ~ **1** `C Z -` - (2, - zJP. (Es. 4 )
Hence, every possible value of (Zi - Zj)P should be tested for `(2, - Z,)P` of Eq.
**_4._** Since the correctness of each row vector of G", **_g i ,_** can not be verified in-
dependently, the complete solution of G" should be obtained and verified.
This involves on the average work factor, T given by
```
k
###### T ?&] 1 N2 -
```
Substituting for N, T can be shown to be **(nu).** Thus we establish the fol-
lowing.
-----
###### Claim : To determine G from a chosen- plaintext attack (as discussed
above) `has a work factor` T = fl ( n").
It can be easily shown that the above step, namely, the determination of
G" is the really dominant factor. Determination of P and `Z vectors` are
straight forward after that. `As of` now, the analysis and procedure ex-
plained seem to be the only possible approach to break the code and it
requires an enormous work factor _0_ **_( n 2 k ) ._**
**4.** **CONCLUSION**
We have introduced a new approach to the private-key algebraic-coded
cryptosystems requiring only simple codes such `as` distance **3** codes. These
systems will be very efficient because of high information rates and low over-
head for encoding and decoding logic. The chosen-plaintext attack given here
appears to be the only plausible approach for cryptanalyst.
It requires a work factor R **(,a2&)** and is therefore, computationally secure
even for small **_k w a . It will be_** a chalIenge to find alternate methods of attack
which can be successful.
###### REFERENCES
plahut **'831** Richard E. Blahut, _Theory and Implementation_ `of Error Correct-`
_ing Code, Addison-Wesley, 1983._
penning **'821** Dorothy E. Denning, _Cryptography and Data Security,_ Addison
Wesley, **1982.**
bin **'831** Shu Lin, Daniel J. Costello, Jr., _Error_ _Control Coding: Fundarnentab_
_and Applications, Prentice-Hall,_ **1983.**
WcEliece '771 McEliece R. J. "The theory of Information and coding," **_(vol._** **_9_**
`of` _the encyclopedia_ of _mathematics and_ **_its_** _Applications) Reading,_ Mass
Addison-Wesley, 1977.
-----
Coding Theory," DSN Progress _Report,_ Jet Propulsion Laboratory, **CA.,**
Jan. & Feb. 1978, pp **42** - 44.
Peterson **'721** W. Wesley Peterson and E. J. Weldon, Jr., _Error-Correcting_
_Codes, Second edition, The MIT_ Press, **1972.**
###### pi.. '84a] T.R.N. Rao, "Joint Encryption and Error Correction Schemes,"
_Proc._ _11th Inti. Symp. on_ _Cornp. Arch., Ann_ Arbor, Mich., May 1984.
###### pm '84bl T.R.N. Rao, "Cryptosystems Using Algebraic Codes," Inti. Conf.
_on Computer Systems_ `6 Signal` _Processing, Bangalore, India,_ Dec. 1984.
-----
|
{
"disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references', 'abstract'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1007/3-540-47721-7_3?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1007/3-540-47721-7_3, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "BRONZE",
"url": "https://link.springer.com/content/pdf/10.1007%2F3-540-47721-7_3.pdf"
}
| 1,986
|
[
"JournalArticle",
"Conference"
] | true
| null |
[] | 6,407
|
en
|
[
{
"category": "Education",
"source": "s2-fos-model"
},
{
"category": "Economics",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/0231f3a936aa6807a7d215c8ef49bb96affd5f3f
|
[] | 0.913903
|
The Equity and Inclusion in Higher Education: A Proposed Model for Open Data
|
0231f3a936aa6807a7d215c8ef49bb96affd5f3f
|
Social Science Research Network
|
[
{
"authorId": "133763011",
"name": "Carla Hamida"
},
{
"authorId": "2081117",
"name": "A. Landi"
},
{
"authorId": "1739188795",
"name": "Ziyi Liu"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"SSRN, Social Science Research Network (SSRN) home page",
"SSRN Electronic Journal",
"Soc Sci Res Netw",
"SSRN",
"SSRN Home Page",
"SSRN Electron J",
"Social Science Electronic Publishing presents Social Science Research Network"
],
"alternate_urls": [
"www.ssrn.com/",
"https://fatcat.wiki/container/tol7woxlqjeg5bmzadeg6qrg3e",
"https://www.wikidata.org/wiki/Q53949192",
"www.ssrn.com/en",
"http://www.ssrn.com/en/",
"http://umlib.nl/ssrn",
"umlib.nl/ssrn"
],
"id": "75d7a8c1-d871-42db-a8e4-7cf5146fdb62",
"issn": "1556-5068",
"name": "Social Science Research Network",
"type": "journal",
"url": "http://www.ssrn.com/"
}
|
Recently, governmental institutions and private industries in power have been pushed to be more transparent so that more people can have ownership of their data. Another type of institution with a large amount of power over data are educational institutions. Colleges and Universities around the globe store a significant amount of data on millions of students, such as financial aid, grades, dropout or graduation, successes after graduation. Each institution is rated with respect to these items and more, and potential students are making decisions to go to the school based on these ratings. Therefore, it is imperative for students, who invest their time and their money into the school of their choice, to know the truth. In 2017, the College Transparency Act and the Student Right to Know Before You Go Act were passed, which were created to push transparency for data in higher education. The openness of data in higher education will be beneficial to prospective students. The push for these two bills coincided with the bitcoin bubble. In the past three years, experts in economics, medicine, and supply chain management have been researching methods on how to implement blockchains to create optimal and decentralized data systems. In this paper, we propose a model for open data in higher education inspired by the Bitcoin, which uses blockchain. When used together with InterPlanetary File System, a peer-to-peer distributed file system, we can create a decentralized platform that increases accessibility of data and autonomy of prospective students.
|
RESEARCH
ASSOCIATION for
# **R A I S**
INTERDISCIPLINARY DOI: 10.5281/zenodo.3267518
JUNE 2019 STUDIES
## **The Equity and Inclusion in Higher Education: ** **A Proposed Model for Open Data **
### **Carla Hamida [1], Amanda Landi [2], Ziyi Liu [3]**
*[1]* *Bard College at Simon’s Rock, Great Barrington, USA (Indonesia), chamida16@simons-rock.edu*
*2* *Bard College at Simon’s Rock, Great Barrington, USA, alandi@simons-rock.edu*
*3* *Bard College at Simon’s Rock, Great Barrington, USA (China), ziyiliu16@simons-rock.edu*
ABSTRACT: Recently, governmental institutions and private industries in power have been pushed to be
more transparent so that more people can have ownership of their data. Another type of institution with a
large amount of power over data are educational institutions. Colleges and Universities around the globe
store a significant amount of data on millions of students, such as financial aid, grades, dropout or
graduation, successes after graduation. Each institution is rated with respect to these items and more, and
potential students are making decisions to go to the school based on these ratings. Therefore, it is imperative
for students, who invest their time and their money into the school of their choice, to know the truth. In 2017,
the College Transparency Act and the Student Right to Know Before You Go Act were passed, which were
created to push transparency for data in higher education. The openness of data in higher education will be
beneficial to prospective students. The push for these two bills coincided with the bitcoin bubble. In the past
three years, experts in economics, medicine, and supply chain management have been researching methods
on how to implement blockchains to create optimal and decentralized data systems. In this paper, we propose
a model for open data in higher education inspired by the Bitcoin, which uses blockchain. When used
together with InterPlanetary File System, a peer-to-peer distributed file system, we can create a decentralized
platform that increases accessibility of data and autonomy of prospective students.
KEYWORDS: open data, higher education, blockchain, IPFS, transparency
### **Introduction ** In today’s society, data is currency. Many stepped into the market by collecting data, e.g. Google, Facebook, National Security Agency. Others, still, monetized controlling the access and use of the data, e.g. Facebook, government spending budgets. Open data is defined by the Open Data Institute as data that anyone can access, use, or share (Open Data Institute 2017). Across the globe, nonprofit organizations are pushing for empowering citizens with data. For example, the Open Data Charter was founded in 2015, and it is a collaboration of more than 70 governments, experts, and organizations whose sole goal is to make governmental data more available and accessible to citizens of the world. The Open Data Charter proposed six principles, and they are meant “for improved governance and citizen engagement” and “for inclusive development and innovation” (Open Data Charter 2015). While governments and private for-profit companies certainly play a role in the monitoring and controlling of data publication, education institutions make huge profits from their management of student data. Transparency and accountability is imperative in higher education. Prospective students need accurate information with respect to financial aid, program success statistics, job- obtained- after-graduation data, demographic statistics, and other forms of cost such as living and food. Educational researchers, accreditation teams, and governments investing financial aid need granulated data on student success so that inclusivity of marginalized groups can be improved (Koch 2018). However, transparency does not mean simply listing summarized data online. In fact, every college or university that receives federal aid from the United States is legally required to submit raw data regarding their demographics and financial aid reports annually (Schneider 2017). This information is available to the public, as it is on the The Integrated Postsecondary Education Data System (IPEDS) website. However, navigating through the website itself is a hassle, and large chunks of data must be downloaded in order to attain the raw data for each institution. We need for these institutions to publish simplified aggregate data in order to fully understand how much change has been made.
-----
### RAIS Conference Proceedings, June 10-11, 2019
### 101
### In Section 1 of this paper, we expand on why transparency and accessibility is required of higher education institutions. In Section 2, we discuss the major issue of privacy of student data with respect to making data more granular. In Section 3, we explain the current blockchain technology so that in section 4 we can propose our solution to the issue of privacy as a stumbling block to complete transparency of higher education data. Finally, in section 5 we conclude our paper and state paths for future work. **Section 1. Transparency in Higher Education ** In 2019, there is still a lack of representation of various people in higher education institutions. Organizational change is slow and it only happens effectively when all members involved in and affected by the change see the value in implementation of the change (Berg and Hanson 2018). Although there have been efforts to inspire this change, such as Affirmative Action which first appeared in the Supreme Court in the 1978 case Regents of California v. Bakke and scholarships for underrepresented people in higher education (West 1998), one reason for such little change in the last several decades is the still present sexism, racism, classism, and homophobia among the student body as well as the admissions process. Another huge reason is the lack of access to data that tracks minority students, providing educational researchers the ability to determine weaknesses in a program and allowing curricular developers the chance to improve courses. In 2017, the STEM Research and Education Effectiveness and Transparency Act was passed. The purpose of the bill is to promote inclusion of marginalized groups, specifically women, in participation of research in STEM. Section 2 Article 2 of the bill emphasizes the need to continually collect “information on student outcomes using all available data, including dropout rates, enrollment in graduate programs, internships or apprenticeships, and employment” so the development of marginalized groups in STEM can be tracked (US House 2017). We often read that universities are becoming more inclusive on the news, e.g. (Association of American Colleges & Universities 2015), (Esters), and (Smith 2018). However, many people either do not have access, or have little access to, the actual data informing the demographics at universities. Moreover, the data available may not break demographics down into specific fields and undergraduate v. graduate programs. Universities such as Harvard and Cambridge publish annual reports on their demographics. Even after these universities publish annual reports, it is still inconvenient for readers to open each annual report to compare the progress between these higher education institutions. Given the continued existence of institutional marginalization, there is a great need to create and implement new policies and solutions. We need to implement a more optimal allocation of resources that can provide real impact to young lives. Understanding the issues within the higher education system, and how these issues affect students, could be done in a systematic matter if all the information was collected on one network in an accessible manner. **Section 2. Transparency v. Privacy: Efforts to Protect Student Data** Despite our need to publish accurate and granular data, we still need to protect the identities of the students represented in the data (ensure anonymity). While there is concern for the misuse of existing data in higher education, it does not mean that is a reason to abandon the idea of sharing. Rather, it means that we need to build systems and establish unambiguous policies in place to protect the data. In 2017, the College Transparency Act was introduced; the bill requires that the National Center for Education Statistics create a data system that analyzes financial costs and student enrollment patterns, customizes information for users accessing the data system, and have the ability to link with other federal data systems (US Senate 2017). In addition to the College Transparency Act, in November 2017, the Student Right to Know Before You Go Act was introduced by Senators Marco Rubio, Mark Warner, and Ron Wyden (US House 2017). The purpose of this bill is to publish granular and uncomplicated data on higher education institutions in order for prospective students to create informed decisions when applying to colleges and universities, while maintaining privacy standards. The bill requires the data
-----
### RAIS Conference Proceedings, June 10-11, 2019
### 102
### platform use encryption technology that “includes the use of secure multi-party computation, which generates statistical data based on information provided by colleges and universities as well as loan and income information from government agencies like the IRS” in order to keep published information anonymous (Ortega 2017). The push for more open data in higher education has not only come from congress. Private organizations such as Data Quality Campaign have been advocating the need to make student data more accessible since 2005. In 2014, the Data Quality Campaign and the Consortium for School Networking established the “Student Data Principles.” There are 10 principles that promote the openness of data in higher education and the use of data to create inclusion in the academic world. The technological platforms being used for higher education data need to be advanced enough to meet the needs in the future. As seen with IPEDs, handling student data can be complex since different governmental organizations and schools have unique ways of collecting their data. Given that the data could also be misused, those holding the complete and raw data are responsible to ensure that identities remain anonymous. We need to establish a coherent system in which information can spread across the network and each party has the ability to access the appropriate information while being able to update the network systematically with complete information. Handling a large amount of information often leads to complications with storage. In the past, establishing such a network was a more difficult task. In the present, however, decentralized and transparent data systems have been created and implemented. One such data system we next discuss is blockchain technology. **Section 3. The Model ** ***Section 3.1. Blockchain *** Blockchain was originally designed to store Bitcoin transactions (Zheng 2017). At the basic level, it is a list of blocks that contains certain information. Figure 1 illustrates a simple blockchain model. 1. Index: the index of the block 2. Timestamp: the time when the particular block was created 3. BPM: pulse rate, an example of the kind of data that can be stored in a block 4. Hash [1] : a unique hash of the block, which is calculated based on all the information stored in the block 5. PrevHash: the hash of the previous block to link the blocks together Figure 1. Simple Blockchain Model (Coral Health 2018)
1 Hashes are calculated with a hash function such as SHA and RIPEMD. The function takes a input string, performs a
series of operations, and output another string of fixed length, the hash. (Madeira 2019)
-----
### RAIS Conference Proceedings, June 10-11, 2019
### 103
### Blockchain has several important characteristics. First of all, blockchain is immutable, which means once a block is added to chain, it cannot be changed. If a block is tampered, its hash will be different from the PrevHash stored in the next block. Therefore, no one can secretly change the data stored on the chain. Moreover, it doesn’t allow single point of authority, which means no single party in the network has complete control over the data stored on the blockchain. Before a new block is added to the chain, an algorithm checks that the block satisfies all agreed-upon features, and this also ensures that all parties on the network has the same chain (Zheng 2017). This makes blockchain a decentralized technology, and is extremely useful for increasing data transparency. To store data with blockchain, there are two options: on-chain and off-chain. Since blockchain was originally designed to store bitcoin transactions, its protocols or big transaction fees limits that only a small amount of data can be stored on-chain, usually in range of kilobytes or less with one kilobytes equals to approximately 500 words (Marx 2018). Therefore, a reasonable solution is storing the hash of the data on-chain, while storing the actual files and the corresponding block hashes (TX-ID) off-chain as shown in Figure 2. There are two common options for storing data off-chain. The first one is traditional database or cloud storage. However, there are several problems with this first option. Once the files are uploaded to cloud or inserted into database, they are once again controlled by one central point of authority, such as Google, Microsoft, Oracle and so on. Not only transparency could be lost, but also if the company decided to close down the storage service, data itself could be lost as well (Marx 2018). This leads to the second option - decentralized storage. In decentralized storage, data is distributed across many nodes on the network, and files are broken apart and stored on various nodes. So no single node has the entire file and breakdown with one node will not affect the others, so files are at a much lower chance of being lost permanently (Marx 2018). One such project is the InterPlanetary File System (IPFS). Figure 2. Storing data with Blockchain (Marx 2018) ***Section 3.2 IPFS *** Just like the Hyper Text Transfer Protocol (HTTP) which the internet is based on today, IPFS is an internet protocol. However, unlike the location addressed HTTP where users get information from central servers according to the IP address, IPFS uses content addressing and a peer-to-peer (P2P) network in which users can share files directly with others in the network (Curran 2018). This is illustrated in Figure 3. This gives IPFS several advantages. Since HTTP is location based addressing, if the server is down or the webpages are deleted, the files are not available anymore and useful information could be lost
-----
### RAIS Conference Proceedings, June 10-11, 2019
### 104
### (FortKnoxster 2017). Also HTTP is centralized, so access to data can be slow depending on where the server is, or even restricted, such as Google, Youtube, Facebook and so on are all blocked by the government of China (Carson 2015). Figure 3. HTTP vs. IPFS (Curran 2018) On the other hand, IPFS is content based addressing, meaning that the link of each file is composed of its unique hash, and every node in the network can choose to keep the files it is interested in. Each file is broken into multiple IPFS objects and linked together by an empty object as illustrated in Figure 4 (Fanzil 2019). This makes sharing and downloading files much faster, since users not only can get data from the closest node which has a copy of it, but also can download parts of a file from different nodes at the same time instead of downloading the entire file from a single server (Curran 2018). Moreover, since IPFS is decentralized, all files on the network is publicly visible and cannot be blocked. Thus, transparency is preserved. A real life example happened in Turkey, 2017. Turkish authorities blocked access to Wikipedia throughout Turkey, but activists created a copy of Wikipedia on IPFS and made it available again (Dale 2017). Figure 4. IPFS model (Fazil 2019) **Section 4. Discussion of the Model and Open Data ** Our proposed model is to use blockchain together with IPFS to create a completely decentralized application that holds important college data. The issue with publishing data while worrying that
-----
### RAIS Conference Proceedings, June 10-11, 2019
### 105
### private information can get in the wrong hands can be absolved using the model we proposed. All the files will be stored on IPFS, and the immutable, permanent IPFS links will be placed into blockchain as shown in Figure 5 (FortKnoxster 2017). Since IPFS tracks version history, using blockchains and IPFS can ensure that annual data will be preserved permanently on the network. In addition, every modification to the data will be visible to the public. Since the consensus algorithm will check the information to be added on the chain, sensitive information is highly unlikely to be published or accessed by parties outside of the network. Therefore, IPFS makes a good candidate for our purpose -- increase transparency of college data. This model can benefit three parties involved in higher education: the students, universities/colleges, and the government. Figure 5. Use of IPFS with Blockchain (Coral Health 2018) Prospective students can see previous versions of the college annual data and discover differences, improvements or setbacks on the network. Thus, having access to more holistic information, these students can make well informed decisions about their future. Universities and the government will have access to organized and clearly presented data. This will make it easier for them to analyze trends, discover issues, and fix problems within higher education. Moreover, governments can determine which universities and colleges will be a positive investment for placing their government aid. An effective data system will lead to effective decision making, for all parties involved. **Section 5. Conclusion** Federal organizations, in general, need to promote the existence and importance of the availability of data. There is no use in making data more accessible to the public without active citizen engagement, because only involvement can push for development. An ideal future next step is to implement this data network through a decentralized application built with blockchain and IPFS. When building this application, we can learn from research or existing implementations of blockchains or IPFS in different fields. Since the use of this decentralized data system is flexible, we strongly encourage governments to store other federal data on the network, which would increase comparability of data and minimize the time in finding problems. **Acknowledgements ** We would like to thank our close friends and family for the support they have given us throughout the year.
-----
### RAIS Conference Proceedings, June 10-11, 2019
### 106
### **References **
Association of American Colleges & Universities. n.d. “Indicators of Equity in Higher Education in the United States.”
Accessed May 4, 2019. https://www.aacu.org/aacu-news/newsletter/indicators-equity-higher-education-united
states.
Berg, E.A. and Hanson, M. 2018. “Putting the ‘Evidence’ in Evidence-Based: Utilizing Institutional Research to Drive
Gateway-Course Reform” in *Improving Teaching, Learning, Equity, and Success in Gateway Courses: New*
*Directions for Higher Education*, Number 180. Koch, A.K. (Ed.). John Wiley & Sons.
Carson, Biz. 2015. “9 Incredibly Popular Websites that are Still Blocked in China.” *Business Insider* . Accessed April
10, 2019. https://www.businessinsider.com/websites-blocked-in-china-2015-7.
Coral Health. 2018. “Learn to securely share files on the blockchain with IPFS!” *Medium.* Accessed April 2,
2019.https://medium.com/@mycoralhealth/learn-to-securely-share-files-on-the-blockchain-with-ipfs219ee47df54c.
Curran, Brian. 2018. “What is Interplanetary File System IPFS? Complete Beginner’s Guide.” *Blockonomi.* Accessed
April 2, 2019. https://blockonomi.com/interplanetary-file-system/
Dale, Brady. 2017. “Turkey Can’t Block This Copy of Wikipedia.” *Observer* . Accessed April 2,
2019.https://observer.com/2017/05/turkey-wikipedia-ipfs/
Esters, Lorenzo L. “Making an Impact in Higher Education Equity.” *Strada Education Network* . Accessed May 4, 2019.
http://www.stradaeducation.org/making-an-impact-in-higher-education-equity/.
Fazil, Usman. 2019. “IPFS: A Distributed File Store.” *Block360* . Accessed April 2, 2019. https://block360.io/ipfs-a
distributed-file-store/.
FortKnoxster. 2017. “How the IPFS Concept Can Change the Internet and the Storage Distribution.” *Medium.* Accessed
April 2, 2019. https://medium.com/fortknoxster/how-the-ipfs-concept-can-change-the-internet-and-the-storagedistribution-c6c13283f12d.
Gandhi, Rohith. 2018. “InterPlanetary File System(IPFS) — Future of the Web.” *Medium.* Accessed April 2, 2019.
https://medium.com/coinmonks/interplanetary-file-system-ipfs-future-of-the-web-c45c955e384c
National Center for Education Statistics (NCES). Integrated Postsecondary Education Data System. Accessed April 10,
2019. https://nces.ed.gov/ipeds/use-the-data.
Koch, A.K. (Ed.). 2018. *Improving Teaching, Learning, Equity, and Success in Gateway Courses: New Directions for*
*Higher Education*, Number 180. John Wiley & Sons.
Madeira, Antonio. 2019. “How Does a Hashing Algorithm Work.” *CryptoCompare.* Accessed April 10, 2019.
https://www.cryptocompare.com/coins/guides/how-does-a-hashing-algorithm-work/.
Marx, Lukas. 2018. “Storing Data on the Blockchain: The Developers Guide.” *Malcoded.* Accessed April 2,
2019.https://malcoded.com/posts/storing-data-blockchain.
Open Data Charter. 2015. “Principles.” Open Data Charter. Accessed July 21, 2018.
https://opendatacharter.net/principles/
Open Data Institute. 2017. “What is Open Data and Why Should We Care?” Accessed May 2, 2019.
https://theodi.org/article/what-is-open-data-and-why-should-we-care/.
Ortega, Jennifer. 2017. “Student Right to Know Before You Go Bill Introduced.” *EDUCAUSE* . Accessed May 4, 2019.
https://er.educause.edu/blogs/2017/12/student-right-to-know-before-you-go-bill-introduced
Schneider, M. 2017. Reforms to Increase Transparency in Higher Education: Testimony before the House
Subcommittee on Higher Education.
Smith, Ashley A, 2018. “States Attempt Closing Racial Gaps to Improve Graduation.” *Inside Higher Ed* . Accessed May
4, 2019. https://www.insidehighered.com/news/2018/08/21/states-showing-some-progress-closing-racial-equity
gaps.
Student Data Principles. “The Principles.” Accessed April 15, 2019. https://studentdataprinciples.org/the-principles/
US House. 115th Congress, 1st Session. *H.R. 4375, STEM Research and Education Effectiveness and Transparency*
*Act.* Act. Washington, Government Printing Office, 2017. Passed in 2017.
US House. 115th Congress, 1st Session. *H.R. 4479, Student Right to Know Before You Go Act of 2017.* Act.
Washington, Government Printing Office, 2017. Introduced in 2017.
US Senate. 115th Congress, 1st Session. S. 1121, College Transparency Act. Act. Washington, Government Printing
Office, 2017. Introduced in 2017.
West, M. S. 1998. The Historical Roots of Affirmative Action. *Berkeley La Raza Law Journal*, *10* (2): 607.
Zheng, Zibin. 2017. “An Overview of Blockchain Technology: Architecture, Consensus, and Future Trends.” *6th IEEE*
*International Congress on Big Data* . 10.1109/BigDataCongress.2017.85.
-----
|
{
"disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.2139/ssrn.3434064?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.2139/ssrn.3434064, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "GREEN",
"url": "http://rais.education/wp-content/uploads/2019/07/012HC.pdf"
}
| 2,019
|
[] | true
| 2019-06-30T00:00:00
|
[] | 5,243
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Political Science",
"source": "s2-fos-model"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/02323347b6dc92c516ef6bc456db5aa647627895
|
[
"Computer Science"
] | 0.906707
|
The Democracy to Come? An Enquiry Into the Vision of Blockchain-Powered E-Voting Start-Ups
|
02323347b6dc92c516ef6bc456db5aa647627895
|
Frontiers in Blockchain
|
[
{
"authorId": "122026692",
"name": "M. Imperial"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"Front Blockchain"
],
"alternate_urls": null,
"id": "17d7865f-0af7-472c-b174-60948bf06d11",
"issn": "2624-7852",
"name": "Frontiers in Blockchain",
"type": null,
"url": "https://www.frontiersin.org/journals/blockchain#"
}
|
This research sets out to analyze the message promoted by start-up enterprises that apply blockchain technologies for the purpose of e-voting [blockchain-powered e-voting (BPE)], and their perceived effects of this technological solution on democratic outcomes. Employing Norman Fairclough’s critical discourse analysis (CDA), I examined the written output of seven BPE start-ups (Agora, DemocracyEarth, Follow My Vote, Polys, Voatz, Votem, and VoteWatcher), as displayed in their websites. The close attention of CDA to power relations brought out relevant topics of discussion for analysis. Notably, these included: voting as an expression of democracy; technological determinism; individual versus communitarian understandings of democracy; the prominence of neoliberalism and the economic sphere; and technological literacy. Findings from the literature suggest that the assumptions of BPE start-ups about a blockchain-powered democracy diverge from widely accepted understandings of democracy. BPE start-ups envision a democracy determined by positions and institutions of power, by the technologically able, and by economic interests. This research argues that this conception of democracy disempowers voters from any form of decision-making regarding how democracy is run beyond their expression in the form of a vote decided by these established powers. The widespread addresses to existing elites to enable BPE, as well as what is left unsaid about community, collective rights and the not so technologically literate population, imply that BPE developers display concern for one particular expression among the many diverse and heterogeneous understandings of democracy, while disregarding outstanding privacy, security and accountability concerns associated to implementations of the technology for BPE. This work is a contribution to much needed research on technology and democracy’s deepening intersections, at a time of rapid technological innovation and turbulent democratic scepticism.
|
_Edited by:_
_Marta Poblet,_
_RMIT University, Australia_
_Reviewed by:_
_Vanessa Teague,_
_Australian National University,_
_Australia_
_Jake Goldenfein,_
_The University of Melbourne, Australia_
_*Correspondence:_
_Miranda Imperial_
_mci30@cam.ac.uk;_
_miranda.imperial@gmail.com_
_Specialty section:_
_This article was submitted to_
_Blockchain for Good,_
_a section of the journal_
_Frontiers in Blockchain_
_Received: 24 July 2020_
_Accepted: 19 March 2021_
_Published: 09 April 2021_
_Citation:_
_Imperial M (2021) The Democracy_
_to Come? An Enquiry Into the Vision_
_of Blockchain-Powered E-Voting_
_Start-Ups._
_Front. Blockchain 4:587148._
_[doi: 10.3389/fbloc.2021.587148](https://doi.org/10.3389/fbloc.2021.587148)_
p p
[doi: 10.3389/fbloc.2021.587148](https://doi.org/10.3389/fbloc.2021.587148)
# The Democracy to Come? An Enquiry Into the Vision of Blockchain-Powered E-Voting Start-Ups
_Miranda Imperial*_
_Department of Sociology, University of Cambridge, Cambridge, United Kingdom_
This research sets out to analyze the message promoted by start-up enterprises
that apply blockchain technologies for the purpose of e-voting [blockchain-powered
e-voting (BPE)], and their perceived effects of this technological solution on democratic
outcomes. Employing Norman Fairclough’s critical discourse analysis (CDA), I examined
the written output of seven BPE start-ups (Agora, DemocracyEarth, Follow My Vote,
Polys, Voatz, Votem, and VoteWatcher), as displayed in their websites. The close
attention of CDA to power relations brought out relevant topics of discussion for
analysis. Notably, these included: voting as an expression of democracy; technological
determinism; individual versus communitarian understandings of democracy; the
prominence of neoliberalism and the economic sphere; and technological literacy.
Findings from the literature suggest that the assumptions of BPE start-ups about
a blockchain-powered democracy diverge from widely accepted understandings
of democracy. BPE start-ups envision a democracy determined by positions and
institutions of power, by the technologically able, and by economic interests. This
research argues that this conception of democracy disempowers voters from any
form of decision-making regarding how democracy is run beyond their expression in
the form of a vote decided by these established powers. The widespread addresses
to existing elites to enable BPE, as well as what is left unsaid about community,
collective rights and the not so technologically literate population, imply that BPE
developers display concern for one particular expression among the many diverse and
heterogeneous understandings of democracy, while disregarding outstanding privacy,
security and accountability concerns associated to implementations of the technology
for BPE. This work is a contribution to much needed research on technology and
democracy’s deepening intersections, at a time of rapid technological innovation and
turbulent democratic scepticism.
Keywords: blockchain, critical discourse analysis, democracy, e-voting, start-up, technological determinism,
technological literacy
-----
## INTRODUCTION
The possibilities of use of blockchain technologies in the public
sector have been thoroughly reviewed recently (Berryhill et al.,
2018; Thomason et al., 2019). Besides any possible incorporation
to the public sector of financial applications of blockchains, a
field that is spearheading blockchain adoption in the private
sector (Arslanian and Fischer, 2019), governments appear to be
aware of the “transformative and potentially disruptive nature
of this emerging technology” (Berryhill et al., 2018, p. 20),
and several hundred initiatives are underway (Berryhill et al.,
2018, pp. 20–22), many of them taking advantage of publicprivate partnerships as a means to jumpstart access to the
technology (Berryhill et al., 2018, pp. 22–23). According to the
American Council for Technology-Industry Advisory Council
(American Council for Technology Industry Advisory Council
(ACT-IAC), 2017) most areas of the public sector could benefit
from the use of blockchains (Berryhill et al., 2018; Thomason
et al., 2019). But it is perhaps the application of blockchains
for voting that has become the most advocated (Allen et al.,
2019), although outstanding security risks remain (National
Academies of Sciences, Engineering, and Medicine, 2018; Park
et al., 2020) that could hinder its widespread application
and that have resulted in early failures (Juels et al., 2018;
Goodman and Halderman, 2020).
Many abstract controversies regarding blockchain-powered
e-voting (BPE) systems are likely to be tempered in the landscape
emerging in the aftermath of the current COVID-19 pandemic,
where “new normal” (Beck, 1992, p. 79) regulations are likely
to make current methods of synchronous, in-person voting
exceedingly cumbersome or unadvisable (Blessing et al., 2020).
On the other hand, the need for alternatives to in-person voting
is likely to intensify the scrutiny over privacy and security
issues with those systems (National Academies of Sciences,
Engineering, and Medicine, 2018; Park et al., 2020), as the recent
November 2020 United States presidential election, with its
widespread–although not necessarily well-founded–allegations
of voting irregularities starkly highlighted.
Beyond technological, privacy and security aspects, there is
a paucity of reports on the relationship between BPE intended
applications and current understandings of social relations,
as it can be expected from a relatively new technology that
is beginning to enter a widespread adoption phase. This
is particularly true regarding the implications that adopting
this new technology may have for power relations. In this
work, I center on the study of the power relations covert
in communication between technology developers and users
by analyzing how BPE start-ups communicate the means and
ends of their products, their views on how their technology
can impact democracy, and how they are shaped by the
power relations implicit to the development of BPE. To that
end, I will first review the relevant sociological literature on:
(i) the impact of new technologies on society, particularly
those related to technological determinism and its criticism;
and (ii) current understandings of democracy and how they
are impacted by alternative modes of voting, as these will
become relevant in the subsequent empirical analysis of the
messages BPE start-ups are advancing to promote their products
and technology.
## LITERATURE REVIEW
Technological Determinism
Scholarly attempts to discern the relationship between new
technologies and society have tended to fall into the trap of
technological determinism, that is, describing a purely causal
relationship between technology and its–generally positive–
important effects on society. There is a “strong tendency,
especially when technologies are new, to view them as causal
agents, entering societies as active forces of change that humans
have little power to resist,” communication around them
becoming “deterministic” (Baym, 2015, p. 26). Examples abound
in the literature (McLuhan, 1964; Fischer, 1992). Technological
determinism is an “optimistic theory” that either fails to
recognize the misgivings of technology, or believes negative
outcomes stemming from technology can be eliminated by new
innovation (Markus, 1994).
An informed discussion of technological determinism is
crucial to a discourse analysis of communication of BPE, since
descriptions of the technology are determinant in promoting
its adoption. It is important to analyze the messaging around
it, its developers’ intentions and assumptions, conveyed in
messaging: whether they conceive of BPE as an innovation to
induce radical, positive change unto the world, and whether they
adhere to logics similar to those of technological determinism.
Such technologically “utopian” ideas (in Nye, 1997) tend
to see technologies as “natural societal developments,” as
“improvements to daily life,” and as “forces that will transform
reality for the better” (Baym, 2015, p. 28). The yardstick of
democratic ideals that most Western liberal societies adhere to–
and that the world now adheres to through the West’s colonialist
imposition of its aspirations–is something which BPE addresses,
albeit covertly, in their messaging.
Technological deterministic narratives have been associated
to the promotion of democratic ideals in the literature before.
Some academics have deemed the creation of the Internet
a “renaissance of democracy” (Agre, 1994, cited in Curran.
2016), and a “revitalized democracy characterized by a more
active informed citizenry” (Corrado, 1996, in Levin. 2002,
p. 81). Similarly, Castells has stated that “dictatorships could
be overthrown with the bare hands of the people” thanks to
the power of a technology: the Internet (2012, p. 1). This is
not a new thought, as much has been written about social
media’s promotion of platforms and media where “participatory”
agency is created (Fuchs, 2013, p. 26). Others, like Jenkins
(2008, p. 137) and Shirky (2009, p. 107) have further remarked
how the Internet goes beyond economic facilitation and
empowers “consumer participation” (Jenkins, 2008, p. 137)
and “participatory engagement” (Deuze, 2007, p. 95), enabling
conversation and action.
I draw on this theory because the conveyance of blockchain’s
new and improved privacy, anonymity and efficiency capabilities
might well fall into technologically deterministic writing, or at
-----
least, follow a trend in theories linking emerging technology and
the current transformation of democracy. Directly relevant to
voting, the emergence of Internet-mediated platforms showed,
for many, that “the public would gain unprecedented access to
information, and be better able to control government” (Toffler
and Toffler, 1995, in Curran, 2016), and “empower [.] previously
excluded groups” (Poster, 2001, p. 175) via more horizontal
avenues of communication. Ideas about the Information Society’s
“democratic and decentralizing” power (Garnham, 1994, p. 43)
are widespread. Some have even called the Internet a “liberation
technology” (Diamond, 2010, p. 71).
Regarding e-voting, academics have discussed it as a potential
end of Internet-mediated democracy. It has been linked to direct
democracy (Grossman, 1996, p. 250) as, through it “voters will
have a voice that reaches directly to the highest levels of both
parties and the government” and it will “bring accountability
directly to bear on elected officials” (Rash, 1997, in Levin, 2002,
p. 82). It remains to be seen whether communication around
blockchain’s adoption for e-voting purposes will fall into these
technologically utopian commentaries or whether it will bring
about different implications for the democratic aims to be sought
after in modern societies.
## Criticism of Technological Determinism
Critics of this kind of optimism over the web 2.0 regard it as
spreading an ideology serving corporate interests (Van Dijck and
Nieborg, 2009; Fuchs, 2011). This line of criticism of academics
overly relying on technological innovation for its explanation of
societal trends, have also developed different ideas on the side,
beyond circular narratives of “technology shap(ing) technology”
and society (Ellul, 1964, pp. 85–94; Winner, 1977, pp. 57–73, in
MacKenzie and Wajcman, 1989). Many of these social theories
are important for this research, as they, unlike technologically
deterministic accounts, do not fail to explore power relations
emerging from the creation and adoption of new technologies.
For some social scientists, the consequences of technology
are not just innovation in hardware and software or economic
benefits, rather “they apply to all areas of social life” (Schroeder,
2007, p. 9). Social construction of technology approaches
developed by Wiebe Bijker and Trevor Pinch (Bijker et al., 1987;
Bijker, 1995), to the contrary from technological determinism, see
“technology and society as continually influencing one another”
(Baym, 2015, p. 26). Placing agency in people’s hands as creators
of technology (Nye, 1997, p. 151), these theories are more fitted
to observing decisions taken by technology developers, as they
are “seen as dependent on their social contexts which are, in
turn, shaped [. . .] by communication” (Baym, 2015, p. 44). This
is why this approach is particularly relevant to an analysis of
BPE. There has further been much written about the “prosumer,”
with the “blurring of the line that separates producer from
consumer” (Toffler, 1980, p. 267), showing how individuals,
even beyond the creators themselves, can shape technology in
their everyday life usage. However, the fact that BPE is not a
platform for conversation might make these theories less relevant
to this particular avenue of research. I adopt Winner’s thesis
that “technologies [.] can be inherently political” (Winner, 1989,
p. 33), since they “can be designed, consciously or unconsciously,
to open certain social options and close others” (MacKenzie
and Wajcman, 1989, p. 4; see Winner, 1989, p. 32). This is an
important idea that will emerge as pertinent to an analysis of
BPE. Ultimately, Mackenzie and Wacjman’s thesis, that “it is
mistaken to think of technology and society as separate spheres
influencing each other: technology and society are mutually
constitutive” (MacKenzie and Wajcman, 1989, p. 23) will inform
the analysis of my research.
## Blockchain-Powered E-Voting and Democracy
I decided to explore democracy as an analytical benchmark
since blockchain applications for e-voting deal with voting as a
concept, and voting is crucial to democratic aims, values and
ends, albeit not the one key defining feature of democracy, as
we will see below.
The relationship between voting and democracy is
delimited by the concept and fundamentals of democracy.
Different definitions of democracy exist, with some recent
understandings of the term taking into consideration data
and algorithms (“linked democracy,” Poblet et al., 2019) or
blockchain technologies (“cryptodemocracy,” Allen et al., 2019).
Within the context of this research, Bernard Crick’s remarks
seem appropriate:
[A]ll can participate if they care (and care they should), but they
must then mutually respect the equal rights of fellow citizens
within a regulatory legal order that defines, protects, and limits
those rights. This is what most people today [. . .]. ordinarily
mean by democracy–let us call it “modern democracy,” ideally a
fusion [. . .] of the idea of power of the people and the idea of
legally guaranteed individual rights. The two should, indeed, be
combined, but they are distinct ideas, and can prove mutually
contradictory in practice (Crick, 2002, p. 13).
At the root of common conceptions of democracy, citizens
can freely exercise their opinions on how to be governed
via vote, either to choose representatives or to give their
opinion on an issue, and they will abide by the decision of
the majority. However, individual voters can have very little
influence on the outcome of an election, and this can discourage
them from voting. As early noted by Condorcet (McLean and
Hewitt, 1994, pp. 245–246), or Hegel (1991), this can discourage
voting and pose hindrances to participation and commitment
to democratic ideals. Notably, I observe an underlying tension
in defining democracy that Crick has masterfully outlined: that
between the participatory, communal and emotional “power
of the people,” versus “individual rights.” Both of these are
needed for a functioning and normatively “good” democracy
according to Crick’s widely accepted definition. Therefore,
I will take this definition and these two elements as the
benchmark against which BPE start-ups’ assumptions about
democracy will be judged.
In his 1957 attempt at modeling political decision-making in
democracies, Anthony Downs highlighted the high opportunity
cost of voting, which thereby led to a paradox: by voting, rational
citizens do not maximize the expected utility. Therefore, why
do they vote? (1957, pp. 244–246). Followers of Downs have
-----
supported instrumental theories of the rationality of voting
(Grossback et al., 2007; Noel, 2010).
Critics of this rational choice framework hold that voters wish
to voice their opinions and their ideas through their voting. This
is at the root of the expressive theory of voting, advanced by
Brennan and Buchanan (1984); Brennan and Lomasky (1993),
and Hillman (2010), that “sees the vote as expressing support
for one or other electoral options, rather like cheering at a
football match” (Brennan and Hamlin, 1998, p. 149). Despite all
of the above, most people believe they have a moral obligation to
vote (Mackie, 2015), although the exact reasons why this belief
arises are controversial (Brennan, 2016). In fact, expressivist
theorists believe there is no duty to vote (Brennan and Lomasky,
1993) in countries where voting is a right rather than a duty.
Besides the above, is there a moral obligation regarding how
citizens vote? Influential theorists in the history of democracy,
such as Mill and Rousseau have argued that voters should
cast their vote for the common good, beyond self-interest
(Mansbridge, 1990). Along similar lines, expressivist theorists
claim that voters become attached, in a morally significant way,
to the ideas defended–and implemented in case of victory–
by their candidate (Brennan and Lomasky, 1993, p. 186). If
the conclusion is that voting is morally important, these views
would vary greatly from Downs’ rationalist approach. Further,
by virtue of ascribing a need and a morality to voting, voting
becomes, by necessity, something crucial for a community, as
its importance is greater than that of one choice in many in
an individual’s day. These considerations are fundamental for
this research, as I will be analyzing with which connotations
blockchain e-voting producers speak about and approach voting.
Whether voting is considered an individual or a collective action,
and its importance, will be essential in relation to the democratic
outcomes these technology developers wish for.
Within voting, we must look at e-voting more concretely,
as this is the technological innovation that BPE is advancing.
There is a large body of work regarding voting methods. Current
research suggests that no voting method outperforms the rest in
all situations (reviewed by List, 2013).
E-voting is just the last of a series of “convenience voting”
solutions implemented over time (Gronke et al., 2008). In
e-voting, “voters are provided a method of signing into
a secure website [. . .] and cast their votes using a web
browser” (Gronke et al., 2008, p. 441), although a more
comprehensive definition should reflect that, in e-voting, the
voter’s intention is recorded electronically rather than on
paper. Gibson et al. (2016) have recently highlighted the many
challenges of e-voting, more demanding than those of, for
example, e-banking: (a) authentication, (b) anonymity/privacy;
(c) verifiability/auditability. Most importantly, voter coercion can
be an issue in remote e-voting.
Proponents of e-voting and advancements in voting
technologies, like Krimmer (2012), believe that ICT radically
changes the framework of representation, as it allows new and
more direct interaction with representatives. It also provides
solutions to old problems (like voting from far away or remote
places). Finally, it offers the promise of increasing turnout both
through its facilitation of the voting process (Krimmer, 2012)
and by increasing participation of the youth in the political
process (McAllister, 2016), an effect that is not observed with
late adopters (Richey and Zhu, 2015). A recent example of these
benefits can be found in Indonesia, with the world’s fourth largest
population (close to 300 million) spread over 17,000 islands, and
where democratic voting processes are severely hampered by
weak and insufficient electoral infrastructures. The application
of a BPE technology has shown much promise of improvement
over traditional voting systems (Van Niekerk, 2019). Many of
these potential advantages are presently offset by outstanding
limitations of current technologies centering on security–vote
preservation and certification–and privacy–authentication and
anonymity–issues (Gibson et al., 2016; National Academies of
Sciences, Engineering, and Medicine, 2018; Park et al., 2020).
Owing to these, some electoral systems–such as Switzerland
(Kuenzi, 2019)–have halted any further development of BPE
systems, whereas others, such as Russia, appear to keep on
pushing for BPE, despite initial, security-related setbacks
(Kapilkov, 2020).
In this work, I argue that a vision of technology and democracy
through power relations is crucial to uncover what might be
hidden behind assumptions made by the communication of
innovation. I show that BPE start-ups, rather than focusing on
procedural matters, such as appealing to the efficiency and ease
of their offered BPE product, choose to focus on the democratic
achievements of their technology, and on its creation for the
betterment of democracy.
Through critical discourse analysis (CDA), I investigated
the perceptions of BPE start-ups on technological determinism
vis-à-vis democratic ends and outcomes, relating back to
understandings of power relations and to democracy, an
elusive, heterogeneous concept where tensions between
communal participative, bottom-down perspectives coexist
with institutionalized practices granting an individualized set of
rights. All of this will help us address how new technology for
democracy (in the form of BPE) confronts this intersection.
This research centers on the application of blockchains to
e-voting, a form of voting online mediated by the blockchain.
Voting is intrinsically related to democracy, and thus, to citizen
participation and representation in government. Representation
is all important in democracy, but it is biased by inequality.
In fact, rampant inequality is at the basis of the ills of present
democratic systems (Fitzi et al., 2018; International IDEA, 2019).
Power relations here are crucial, and must be explored when
dealing with the emergence of a new technology (intrinsically
linked with economic rationale, profit accumulation, . . .) and
its relationship to democracy. As summarized above, there is
a dearth of theoretical studies linking BPE and democracy,
however, such a paucity can be in itself a consideration for
“power relations” (Robins and Webster, 1988, p. 52). Through
a CDA of how developers of this new technology market and
communicate their products I aim at uncovering assumed power
relations (who this technology is directed to, who has the right to
use it, . . .). I also aim at contributing to the academic literature
linking BPE technologies and democracy by analyzing how BPE
developers view the state of democratic values today and where
democracy is headed.
-----
## MATERIALS AND METHODS
Conceptual Framework
The fact that BPE is a technology in a stage of recent
development and being newly adopted by consumers means
that there is little content addressing and describing it besides
that crafted by its creators. Therefore, I embarked on examining
textual content produced by BPE start-ups, describing their
products and the reasons for using them. I believed the
assumptions comprehended in what is written (and what
is not), in tone, directed audience and structure, would be
indicative of the thoughts and stance of the developers of
this technology, that so directly seeks to impact democratic
outcomes. A preliminary examination of these materials made
it clear that texts provided by BPE start-ups focused on
democracy itself, the technology’s ultimate achievements, rather
than on implementation, deployment logistics or technological
functionality. Surprisingly, little attention appeared to be paid
to technology adoption by final users, the voters, suggesting
an underlying assumption that “code is law” (Lessig, 2006; De
Filippi and Hassan, 2016), that people can use the technology.
This involved a complexity in the textual analysis that would be
best approached through the use of CDA as the methodology
of analysis because it includes rich, detailed, in-depth textual
analysis of a carefully selected number of sources (Fairclough,
1995; Wodak and Meyer, 2009).
The intrinsic interest of this research approach was
compounded by the fact that there is a significant gap in
CDA approximations to blockchain applications for any
industry. I anticipated the assumptions around power relations
made within texts marketing these technologies to be especially
covert by the use of technical jargon in the form of complex
technological lexis. All these reasons set the scope of my research
around the analysis of final, published texts of BPE for textual,
discursive and societal features (Fairclough, 1995).
Critical discourse analysis as an analytical method relies on
broader academic theory surrounding discourse analysis and its
identification of the embodiment of power relations in language
(Fairclough, 1989, 1992). Past research around discourse analysis
has put forward the idea that “our use of language in particular
(is) bound up with causes and effects which we may not be aware
of under normal conditions” (Fairclough, 1995, p. 54). Language
is perceived to be more than just an innate and natural function
enabling communication. Proponents of discourse analysis have
theorized that language conveys a set of assumptions and
understandings about the world that are “historically and
culturally specific and relative” (Gill, 2000, p. 173). Because of
this importance of contextual factors, the discursive “is a space
for dispersion, it is an open field of relationships” (Foucault,
1968, p. 10). Discourse analysis further brings about a clearer
understanding of “assumptions” (Altheide and Schneider, 2013,
p. 2) that might appear within the construction of concepts
(Gill, 1996, p. 144) in the language employed. This study of
assumptions presents itself as essential to perceive factors such
as the “limits and the forms of expressibility,” as well as of
“conservation: [. . .] those which are repressed and censured”
(Foucault, 1968, p. 14). The study of what is not said in a text,
as well as the apparent purpose of the content and the people it
is addressed to, distinguish discursive forms of textual analysis
from other methods. This is the main reason why CDA was
chosen to conduct detailed, in-depth research of assumptions in
the texts advanced by BPE start-ups. The small number (seven)
of search-engine discoverable start-ups for BPE existing in the
market suggested that an in-depth critical appraisal of their
online materials according to discourse analysis was feasible
and should be done.
## Research Design
Within available practices of discourse analysis, CDA as theorized
by Fairclough (1995), appeared to be particularly fit to analyze the
power relations established between the views of technological
developers as reflected in text form, and democratic goals
and values. CDA goes further than observing the “content,
organization and functions of texts” (Gill, 2000, p. 187) and
interprets that all social interactions are mediated through
language (Fairclough, 1995). Its capacity for “relationality”
(Fairclough, 1989, p. 3) and “transdisciplin(arity)” (Fairclough,
1989, p. 3), make it well suited to find implicit connections
between expressions as diverse thematically as those concerning
technological innovation and democracy. Furthermore, CDA
goes beyond other discourse analysis genres (such as narrative
analysis) in being particularly “sensitive to power relations”
(Fairclough, 1992). Unlike other discourse analysis frameworks,
it picks up “power, ideology and inequality issues” (Blommaert
and Bulcaen, 2000, p. 447) in a close reading of texts. Fairclough’s
CDA (1992; 1995) revolves around three areas of analysis: the
Textual (consisting of discourse-as-text and micro language
choices), the Discursive (looking at contextual, speech acts and
intertextuality) and the Societal (viewing discourse-as-socialpractice, within ideological and hegemonic processes). Altogether
these are particularly perceptive of power relations implied
within the forms of communication of texts. Academic research
exploring technology and democracy that has foregone power
relations in the past has been criticized for this omission, as I have
observed above. The research fields of technology and democracy
and their interaction, having much to do with representation of
voices, which voices matter most and who has the power to design
creative technological outcomes that end up “mattering,” show
that a method, such as CDA, that prioritizes power relations is
needed for this research.
Critical discourse analysis has intrinsic limitations due to the
influence of context and researcher bias. Many of these have
been highlighted in the literature (Wodak and Meyer, 2009;
Wodak, 2014, p. 311). Despite these limitations, CDA is still
the only method providing the necessary amount of detail and
“self-reflection” to an emergent topic that deeply requires “new
responses and new thoughts” (Wodak and Meyer, 2009, p. 32),
such as BPE. In this instance, CDA can help uncover associations
between topics according to context. It also shows that “there
is no neat separation between the meanings in language and
in the social world” (Taylor, 2013, p. 78). This is important
to the nature of this research: rather than valuing the benefits
or shortcomings of BPE in itself, it is what BPE technology
-----
developers express about BPE that will be analyzed. Thus, it is
only through CDA that such an acute drawing out of social
relations as is needed can be undertaken. Along these lines,
criticisms of CDA as producing interpretation rather than fact
feed into a “notion of truth” (Taylor, 2013, p. 82), dichotomizing
facts and interpretation into binary opposites. This criticism
hardly applies here, given that, as mentioned above, I will be
focusing on production and construction of meaning far beyond
the mere reporting of facts, and exploring how technological
innovators describe their products.
To use CDA empirically, I adopted Gill’s (2000, pp. 188–
189) systematization of the steps to follow in order to undertake
discourse analysis, as she provides a good foundation on how
to approach a broad research question. This was useful in
countering the aforementioned lack of order or clarity in process
in conducting CDA with texts.
## Sample Selection
A Google web search engine exploration for BPE start-ups was
carried out during the month of April 2019. After a thorough
search for the most prominent start-ups, and scouring through
some online media articles talking about different BPE emerging
start-ups, seven start-ups with Internet presence that used the
Web to communicate about their BPE products in English were
chosen. This small number of start-ups was well-suited for CDA
and ensured the possibility to fully examine all of the content
in their websites, including attached pdfs, white papers and
blog posts concerning their products. The chosen start-ups were
Agora, Democracy Earth, Follow My Vote, Polys, Voatz, Votem,
and VoteWatcher (Table 1), and all the materials from the web
pages that were the subject of my analysis were collected during
the months of April and May, 2019. As of May 1st, 2020, the
relevant contents had not been changed.
## Analytical Procedures
Since web materials did not require transcription, I moved
on to “skeptically read and interrogate the text” (Gill, 2000,
p. 189), familiarizing myself with the content and keeping the
research objectives in mind. Following Fairclough, language was
appraised as Textual, Discursive or Social. According to these
categories, texts were analyzed and annotated for implicit and
explicit references having to do with technology’s role in the
government of society and in democratic values, and technology
and power/agency. This involved comparing variability in data
(frequency and presence of different elements, placement on the
web, visibility, rhetorical force, . . .) as well as forming hypotheses
about what I uncovered.
## RESULTS AND DISCUSSION
The CDA conducted covers the bulk of the publicly available
information offered in their web pages by seven BPE startups: Agora; Democracy Earth; Follow My Vote; Polys; Votem;
Voatz; and VoteWatcher (Table 1), following Fairclough’s threedimensional canonical model: textual, discursive, and societal
(Fairclough, 1995). Due to spatial constraints and to the copious
amount of suitable materials, mainly themes of interest that recur
throughout the seven start-ups will be highlighted. However,
some points of nuance and division among these that contribute
toward general conclusions will be included.
Some challenges to the CDA are worthy of mention, notably,
the abundance of data in the seven websites, as well as
the general lack of discursive data due to the “newness” of
blockchain for e-voting. Despite these limitations, analysis of
BPE start-ups’ content allowed me to scrutinize the vision that
technology entrepreneurs have for the future with relation to
existing power relations. The content of the analysis largely
focused on discourses of change, governance, technological
ability and links between democracy and productivity, and
between democracy and capitalism.
## Failing Methods, Technological Solutions: Democracy Reduced to Voting
On the whole, the seven start-ups studied introduced widespread
claims of problems that plague current democratic organizations.
Most of them highlighted flaws in current voting systems and
their subsequent hinderance to democracy. Agora, Vote Watcher
and Follow My Vote, especially, outlined the problems existing
with current voting technologies. Paper ballots and Electronic
Voting Machines were denounced as being “slow, costly and
exposed to many vulnerabilities that can inhibit free and fair
elections” (Agora, 2019b, p. 4). Textual analysis here identified
the use of a highly descriptive lexis, including adjectives or
descriptive phrases loaded with negative connotations referring
to existing voting technologies. Similarly, vocabulary such as
“voting machines used in the 2012 election were over a decade
old, running outdated software that took only 15 min to hack
into?” (VoteWatcher, 2019) was directly followed by “We are
using the latest operating system with the most up to date
software” (VoteWatcher, 2019) as an effective juxtaposition. Like
Agora, Vote Watcher displayed the new technology as necessary:
the flaws of the past technological advancements in voting
prescribe the need for a new technology that solves all of these
issues. Follow My Vote went as far as giving a granular, page-long
analysis of different voting machines’ use cases (Follow My Vote,
2019c), from which the following takeaway was drawn:
Several things can be learned from these system failures. First, the
machines must be physically sound. Second, the programming
must not have holes that can be exploited. Third, it’s not
best practice to have extremely simple and guessable passwords
hardwired into a voting machine.
This statement presents data in a simplified, matter-of-fact
manner. The repeated use of modal verb “must,” as well as
the simple and very direct wording of the list (with very
factual adverbs like “First. . .” “Second” and “Third” introducing
the “things (that) can be learned from these system failures”)
are important to carry meaning forward. This enumeration
prescribed what a voting technological ideal was for Follow
My Vote. Through dichotomy and drawing out particular
existing issues in voting technologies, start-ups were more
effective at pushing the importance of their technology onto
readers’ perceptions, while purposefully ignoring unresolved
-----
TABLE 1 | Blockchain-powered e-voting startups chosen for study.
Startup Country Description
Agora Switzerland Initiated in 2017 (Swiss Lab & Foundation for Digital Democracy, Leonardo Grammar).
[(https://www.agora.vote)](https://www.agora.vote) Wide media exposure after their technology was used in a recent general election in
Sierra Leone
Democracy Earth International Open-source, collaborative enterprise, started in Argentina by Santiago Siri and Pia
[(https://democracy.earth)](https://democracy.earth) Mancini (NET liquid democracy political party). In 2015 they joined other developers
and hacktivists to start the Democracy Earth Foundation, an international collective
united by the vision that distributed ledger technologies can reverse some of the ills of
democracy today: “low participation, polarized endogamy and eroded trust in
[governing institutions” (https://words.democracy.earth/about).](https://words.democracy.earth/about)
[Follow My Vote (https:](https://followmyvote.com) United States A “non-partisan public benefit corporation [. . .], founded on the principles of freedom,
[//followmyvote.com)](https://followmyvote.com) as a tribute to the Founding Fathers of the United States [. . .] to promote truth and
freedom by empowering individuals to communicate effectively and implement
non-coercive solutions to societal problems.” It aims at “improving the integrity
standards of voting systems used in elections worldwide” through the use of
[blockchain technology (https://followmyvote.com/about-us/). The brainchild of Adam](https://followmyvote.com/about-us/)
Ernest, Nathan Hourt and Will Long, it is based in Longmont, CO, United States
[Polys (https://polys.me)](https://polys.me) Russia An “online voting platform based on blockchain technology and backed with
transparent crypto algorithms.” Promoted by the Kaspersky Software Lab (Moscow).
[Voatz (https://voatz.com)](https://voatz.com) United States Founded in 2015 and devoted to the development of blockchain-powered mobile
voting systems that allow voters to cast their e-vote from their mobile phones. Based
in Boston, MA, United States
Votem United States Offering a “revolutionary mobile voting platform designed to securely cast votes in
[(https://votem.com)](https://votem.com) elections across the globe.” It was started in 2014 by Pete Martin and it is based in
Cleveland, OH, United States.
VoteWatcher United States A voting platform launched by Blockchain Technologies (MA, United States), “a voting
[(http://votewatcher.com)](http://votewatcher.com) system for the 21st century [. . .] focused on transparency and efficiency and all of the
code is open source or available for inspection. It runs on off-the-shelf hardware.
Simple paper ballots make it easy for the voter. Detailed election records are posted
online and on the blockchain. Every step in the process is highly auditable.”
privacy, security and overall accountability issues that have
been repeatedly associated to these technologies and recently
summarized by Park et al. (2020). Also significant to broader
research is, perhaps, a societal analysis of what remained unsaid
in BPE start-ups’ content dealing with the flaws of modern
voting. Causal links drawn between the failures of democracy
and the necessary solutions that BPE offered imply that: (a) how
voting is currently conducted is the main problem existing for
democracy, and (b) a reform to how voting is done will be the
answer. This line of argument thus ignored other pressing issues
outside of the boundaries of voting, such as the irruption of
populism, the rise of democratic discontent, or corruption by
elected representatives. Importantly in my CDA, I observed that
many of the existing power relations in society were ignored
by the seven start-ups scrutinized. Imbalances of power such
as the aforementioned were disregarded by a reductionist line
of argument that brings down democracy to one of its many
expressions: voting. This subject, the conflation of democracy
with voting, appeared recurrently throughout my analysis.
## Efficiency, Capitalism–Monetary Concerns? Beyond the Democratic?
As aforementioned, existing problems in technologically aided
voting were a focal point of most of the analyzed websites.
BPE start-ups presented these chaotic methods against the
orderly, scientific promise of the blockchain. Such expressions
could be found within the start-ups’ mission and vision
sections of their websites. For Democracy Earth, this was
“the need to make our shared home a place of peaceful
coexistence” (Democracy Earth, 2019b, p. 2). For Follow My
Vote, it was “to promote truth and freedom by empowering
individuals to communicate effectively and implement noncoercive solutions to societal problems” (Follow My Vote,
2019a). Polys wanted to “change the way people vote” (Polys,
2019b, p. 2). However, despite stating these ideas, expressed
by abstract nouns, an emphasis on changes to proceduralism
over form and ideas was a common thread that could be
observed across different start-ups’ written expression. Within
the textual dimension of CDA, I encountered many instances
where the democratic process was referred to in a highly
detached, scientific manner. The little concern for emotion and
emotional language showed BPE start-ups’ focus on productivity,
efficiency and securitization of the means for democracy. But
no reference to ideas of communitarianism, equality or justice
conveyed in democratic thought (e.g., Rousseau) was made.
This emphasis on such a means of democracy for success
was repeated throughout the texts. Certain lexical choices
employed throughout the corpus displayed this, e.g., “electoral
procedure” (Agora, 2019b, p. 4). This noun, “procedure,” conveys
a potential for mechanical, technological operationalization, to
enact more effective and productive action, in order to facilitate
democratic outcomes.
It could be argued that a focus on operationalizing means
and efficiency may be associated to scientific language because
BPE start-ups were, at their core, proposing a technological
product. However, the appearance of lexis like “incentivizes”
(Democracy Earth, 2019a), “voting systems” (Democracy Earth,
2019b, p. 2), “Governance as a service” (Democracy Earth,
2019a), . . . indicates something different. The societal dimension
of CDA links language signaling more efficient and cost-effective
-----
operations to capitalism and the economic sphere. What this
indicates in relation to power relations is that BPE start-ups
operate beyond ends purely focused on democratic outcomes,
and hints at the economic forming a large part of how their
ideal voting “procedures” are to be developed and deployed.
The languages of technology and capitalism converged in this
frame. Beyond occasional expressions of the more emotional
values behind democratic theory: “While money is the language
of self-interest, votes express the shared views of a community.
Political currency is not strictly meant for trade but for
social choice” (Democracy Earth, 2019b, p. 8), start-ups were
primarily concerned with both democracy and the economic, as
exemplified by the following Democracy Earth quotation:
Although politics and economics are often perceived as different
realms, history teaches that money means power and power
means votes. In order to effectively promote democracy it is
essential to address both (Democracy Earth, 2019b, p. 15).
As I noted previously, this acute concern with the
antidemocratic flaws of existing voting procedures, coupled
with ideas of vote “incentivization” and efficiency, display a
stark reality. The current state of the economy is a capitalist
one, whereby start-ups present business models whose main
purpose is to develop a profit-making mechanism for themselves
and for their investors. Discursively, I found a similar trend:
investors were a crucial warrant of accountability for the subjects
of my analysis. Though most websites avoided referring to
them, venture capital funds such as Fenbushi Capital (powering
Agora) were mentioned.
Papacharissi (2014) identified the crucial role of affect and
the emotional for democracy, especially in times of election
and electoral campaigning. BPE start-ups, however, neglected
this dimension and presented a radically opposite view: they
considered the pure act of voting as the expression of democracy.
By doing so, they proposed a mechanistic, operational ideal,
and ultimately showed a highly polarized view within the
political scenario.
The observed connection between the development of
new technologies for democracy and economic motivations
requires further exploration, and opens up grounds for
research in future work.
## Who? Audiences and the Issue of Voice
Discursively, CDA allows for a meticulous insight into the
treatment of voice in BPE start-ups’ literature. A common trend
that I will outline here is that, interestingly, the perceived existing
“flaws” of democracy singled out previously, tended to be framed
in the texts from the lens of the individual subject. Quotations
such as “Every eligible individual should be able to actively
participate in democracy by easily and safely voting when,
how, and where they want” (Votem, 2019b, p. 3) were dotted
throughout the texts, thus centering the problem and potential
solution around the rights of individual citizens. The promise
behind BPE start-ups’ services appeared as an improvement for
the individual citizen within a democracy. Quotes such as the
aforementioned, and similar ones, like “Follow My Vote’s mission
is to promote truth and freedom by empowering individuals to
communicate effectively and implement non-coercive solutions
to societal problems” (Follow My Vote, 2019a), focused on
the individual as the main subject existing vis-à-vis democratic
institutions and being addressed by the radical positive changes
of the blockchain. This notion has important implications for
power relations. BPE start-ups generally, and Votem and Follow
My Vote more specifically, conveyed that they conceive of
voting and democratic outcomes as, ultimately, an atomized,
individual choice, echoing rational choice theory models such as
that proposed by Downs (1957). The start-ups under scrutiny
thus showed little interest in comprehending or adopting more
communitarian understandings of democracy and participation.
This point reiterates and aligns well with the previous finding of
a lack of emotional and social participation in the ideal future
of democracy that BPE start-ups foresaw. Rather, these startups sought to transform democratic practice for the better by
guaranteeing the fulfillment of democratic ideals to individuals
and individuals alone. This, again, represented a way of avoiding
communitarian ideas of democracy. In doing so, BPE start-ups
appeared to very much side with the current statu quo, a statu
quo that is being questioned by citizens in the wake of the last
global economic crisis (Ancelovici et al., 2016).
In addition, and very importantly for the discursive sphere of
the CDA, is who (the Audience) these messages are subliminally
designed to be read by. I explored who it is that the BPE startups were attempting to reach with their literature, who it was that
would be interested enough to read and consider their output.
In several instances, there was written content in their websites
that was under lockdown for the general Internet public, and
could not be accessed unless you were a client or in touch with
the business (e.g., Voatz, with a mostly locked down page). Voatz
openly marketed itself toward electoral administrators, with “Are
you an election administrator interested in trying out Voatz at
your next federal, state or local election? Contact Us” (Voatz,
2019) as the question introducing their contact form. This makes
it clear that the expected audience of the product were people
already involved in the implementation of democratic practices.
Other websites presented similar focal points. Agora “work(s)
together with vote administrators and politically neutral thirdparty organizations to implement fair and trusted voting systems”
(Agora, 2019a) and alluded to their authority as making their
“consensus framework” run right, over choosing to assign this
role to, say, impartial voters. Follow My Vote attempted to
include voters in its rhetoric more than other more institutionally
minded websites (perhaps due to the fact that its code is opensource). It did so with expressions such as:
No one except for election officials really knows what happens to
your vote once you cast it, so it’s not surprising that more and
more research is showing that citizens don’t vote because they
don’t believe their votes count. Understandably, these frustrated
voters are losing confidence in our democratic system (Follow My
Vote, 2019a).
However, and despite this claim, there were areas of the
website where authority and decision-making power were not
as radically shared. Quotations such as “after all, in an election,
it’s not who votes that counts, it’s who counts the votes!”
-----
(Follow My Vote, 2019a) emphasize this, as do web sections such
as “Benefits for Candidates” and “Benefits for Registrars,” only
countered by a single “Benefits to Voters.” Therefore, I argue
that it is possible to trace a democratic asymmetry inherent
in these start-ups’ “democratic” ends. This analysis challenges
the extent to which the products that BPE start-ups introduced
are that ground-breaking, or even, “democratic,” given that
existing institutions continued to be thoroughly involved with
guaranteeing the running of democratic outcomes, and average
citizens continued to be excluded from decisions surrounding
electoral processes. The Voatz website stated the product will
benefit “overseas” and “military” voters who found it difficult
to participate in elections before (Voatz, 2019), however, the
discursive aspects of my analysis display that the text (the
solution to their troubles) is clearly not including them as
active participants in the creation of opportunity for their
involvement. What this signals, accordingly, is an inherent power
asymmetry in the way these start-ups address audiences. By
restricting BPE implementation to existing institutions and not
involving citizens in the process, the start-ups under scrutiny
fell into representing a product that has not been democratically
implemented and decided upon, but rather, one that would be
imposed by existing institutions, the same existing institutions
that are being attacked by present-day critics of the democratic
_status quo. Overall, the acute emphasis these start-ups placed on_
individualistic interpretations of democratic processes, together
with their dialog with existing institutions and representatives in
power, display a replication of present-day power relations. This
questions to what extent BPE is a ground-breaking force with a
great potential to democratize governance.
## Technological Determinism
A common aspect in the message put forward by BPE startups is the emphasis on the positive overtones of the relationship
between technology and normative goodness. An example is
given in the following quote:
With internet growth reaching over 3 billion lives [. . .] there’s no
reason stopping mankind from building a borderless commons
that can help shape the next evolutionary leap for democratic
governance at any scale (Democracy Earth, 2019b, p. 3).
This fits into societal discursive analysis and displays a concern
for the power relations embedded at the crux of technology and
democracy. Democracy Earth also stated that “The next Silicon
Valley is not in a faraway land or on any land at all, but a
new frontier of the internet itself rising as the one true open,
free and sovereign network of peers” (Democracy Earth, 2019b,
p. 24). Both the direct causal link between “internet growth” and
“shap(ing) the next evolutionary leap for democratic governance”
shown in the first quote, and the hyperbolic language: “the
one true open, free and sovereign network of peers,” employed
in the second, act similarly. These passages depict a causal
relationship between technology and democracy, with technology
being expressed as the solution to democratization and giving
voice to a population.
The intrinsic relationship established between e-voting startups and technology and its widespread benefits reminisce of the
technologically deterministic discourses of technological hype
and utopia surrounding the origin of the web and the web 2.0,
stemming from Silicon Valley (Castells, 1998, 2012). Therefore,
whilst it is in the start-ups’ interest to put forward a claim
where their technological innovation is seen as indispensable
for the future of our democratic values (“it is impossible to
envision the future of democracy where digital elections are not
the global standard” Votem, 2019b, p. 3), there are overarching
discourses identifying broader ideas about the world at play.
One such allusion was displayed in the following example:
“the mere existence of risk should not preclude technological
progress” (Votem, 2019b, p. 3). The intrinsic relationship
between technological innovation and capitalism was alluded to
through connotation, through words evoking investment, such
as “risk.” This is representative of the power embedded in the
funding and expertise required to develop these technologies.
This, and its relationship with a better future says a lot about who
they envision as bringers of a promising future and what kind of
skills and resources are needed for this.
Much of my discussion around technology in the textual
dimension of CDA follows this societal enduring discourse too.
My analysis highlights that start-ups believed in “empowering”
voters through their technology:
We are tapping on delivering a human right that can effectively
empower individuals that will have to face the coming challenges
of automation (Democracy Earth, 2019b, p. 20).
Follow My Vote’s mission is to promote truth and freedom
by empowering individuals to communicate effectively and
implement non-coercive solutions to societal problems (Follow
My Vote, 2019b,c, p. 2).
This indicates that BPE platforms view voters as a largely
disempowered commons that can achieve the empowering that
democracy should bring about through technological innovation
(as argued above) powered by capitalism and Venture Capital.
## Technological Literacy
Finally, voice and technology meet around the issue of
technological literacy. There was a multiplicity of statements
that indicate the existence of complex power relations between
technological innovation and the promise of democratic, fair and
equal political futures. Several pressing problems were singled
out in these start-ups’ literature, problems that must be solved to
uphold the feasibility and realization of democratic government.
One of these is participation. Sentences like:
Democracy can’t function if almost half of citizens aren’t voting;
and in this regard Follow My Vote is striving to restore the
democratic tradition (Follow My Vote, 2019a).
are evocative of many meanings directly linked to power
relations surrounding technology and democracy. The uses of
language here did not include conditional verb tenses: rather, it
was stated, in the present tense, that democracy “can’t function”
as it currently is. This statement hints at the fact that it is
technological ability and competence that allows for Follow My
Vote to “striv(e) to restore the democratic tradition and that
will find the solutions necessary for democracy to be “fixed.”
-----
Thus, the sentence “Follow My Vote is striving to restore
the democratic tradition” indicates, through a gerund implying
continuity, that Follow My Vote is attempting this via their
technology. This particular use of language may well imply
a trend seen throughout this CDA, textually, discursively and
societally: that technological ability gives developers agency to be
able to perform this “striving to restore the democratic tradition”
through their creative means and solutions. This conclusion ties
up with the academic line of inquiry identified as technological
determinism. Connotations of technology being able to solve
a perceived democratic deficit demonstrate this. However, it
is through a discursive analysis of the quote above that the
implications of these start-ups’ narratives for technological
literacy can be perceived.
The discursive aspects of CDA raise many questions
pertaining to intertextuality and audiences. The quote above
suggests the Follow My Vote technology has the potential
to “restore the democratic tradition.” However, it also poses
important concerns regarding the agency of others to contribute
to the preservation of democracy. These queries are elicited
because the agency of citizens with no technological skill
or knowledge remained unaddressed and unaccounted for.
There was a perceived scarcity of statements about the
participation and contribution of citizens without technological
skills, beyond actively voting for politicians and institutions
within the boundaries set by these (as seen in my audience
analysis in section “Who? Audiences and the Issue of Voice”).
Follow My Vote assumed that democracy would be preserved
if most or all citizens were voting, and thus assigned
voters, who would employ the technology developed by the
start-up, a passive role in shaping democracy’s functional
future. The text’s connotations imply that voters with no
technological skills or without institutional power are resigned
to vote and nothing more. On the other hand, technologically
skilled individuals, as well as the institutions they seek
to collaborate with, can actively shape the governmental
outcomes of society.
Among all the start-ups analyzed, Polys stood out as claiming
to provide a platform where “no specific training or IT literacy
is needed” and emphasizing that it “is a flexible application
and can be easily customized for your particular needs” (Polys,
2019b, p. 3). They also emphasized the importance that
complementary information and knowledge have in order to
realize the promise of a more democratic society that technology
can potentially bring about. By stating that “any attempts to
improve the electoral system and democracy with the help
of new technologies are meaningless without raising the level
of voter awareness” (Polys, 2019a), Polys demonstrated a care
and a need for additional qualities. Beyond individualism and
beyond technological literacy it is the actual normative values
that are important to maintain democracy. This demonstrates,
on the part of Polys specifically, an acute emphasis on access
to democracy in a way which, for them, is virtually made
harder by bureaucracy and the impositions of an inefficient
system. Regarding technology, Polys commented: “People make
democracy–an online voting system is just an instrument
for facilitating honest, transparent elections” (Polys, 2019a).
Further and similar to this, Democracy Earth argued that “No
technology will ever be able to satisfy democratic aspirations
if it can only be understood by an elite” (Democracy Earth,
2019b, p. 8), and put Facebook and Google as examples of
monopolistic technological companies that do not manage to
ensure the rights to privacy and transparency of their user
bases. Finally, Votem also emphasized the ease of the process:
“With just a few taps, you can give voters a more powerful
way to have their say from their mobile phone or secure web
browser. . .” (Votem, 2019a).
As referenced above, it is interesting to note how these
start-ups acknowledged the need for convenient, user-friendly
technology for everyone in order to achieve democratic
outcomes. However, most of the societal discourses leading to
an improvement of democratic outcomes idealized technology
and presented it as able to channel democracy in the right
way, as aforementioned, in a highly technologically deterministic
mode. They barely considered the fact that this technology might
not initially be reachable by everyone. In fact, they disregarded
discourses dealing with the digital divide (Warschauer, 2003)
on the basis that most of the population is connected to the
internet (International Telecommunications Union (ITU), 2018).
As a result, BPE start-ups made little reference to promoting
the skills that are necessary to participate in the creation of
“solutions” to democracy through the use of their platforms.
This is clearly a paradox in wide-reaching projects such as those
I have analyzed.
## CONCLUSION
In analyzing texts made available by BPE start-ups in their
web pages, omissions related to widespread concerns regarding
current implementations of BPE technologies stand out. These
concerns have been recently summarized as follows:
(1) Blockchain technology does not solve the fundamental
security problems suffered by all electronic voting systems. (. . .)
(2) Electronic, online, and blockchain-based voting systems are
more vulnerable to serious failures than available paper-ballotbased alternatives. (. . .) (3) Adding new technologies to systems
may create new potential for attacks (Park et al., 2020, p. 19).
The above considerations led the authors to conclude that
“blockchain-based voting methods fail to live up to their apparent
promise” (Park et al., 2020, p. 19). Perhaps unsurprisingly for
start-ups that wish to promote their products, these issues were
not touched upon in their texts. This is especially poignant
in the case of Voatz, perhaps the most secretive of the BPE
start-ups analyzed (see above). An independent analysis of
their BPE server carried out by Specter et al. (2020) through
reverse engineering of their mobile app uncovered extreme
security and privacy vulnerabilities that should preclude its use
in electoral processes.
Likewise, allusions and mentions of community and
empowering voters can be spotted sparingly, presenting ideas of
-----
radical change to existing power relations governing democracy
that have made it fail. However, CDA findings overwhelmingly
suggest that, rather than wishing to profoundly alter existing
power relations to transform and revitalize democracy, BPE
start-ups are ready to make little change beyond switching
the ways in which citizens vote, essentially promoting the
adoption of their technological solutions. Their understanding
of democracy challenges existing definitions of the term.
Those employed throughout this work emphasize a care
for voters’ rights, as well as a concern for upholding them
after and beyond election (voting) time. For this to happen,
a combination of the “power of the people,” along with a
legal enshrinement of individual rights, are necessary. This
tension between communal, bottom-up power and institutional,
reified power appears in academic research about democracy.
Nevertheless, BPE start-ups do not address it and, instead,
reduce the extent of democracy to its most procedural expression
(voting) and address the blockchain’s benefits to individual
voters, foregoing any mention of the positives to life as a
community within democracy. This raises some doubts as
to what community and assembly in participation would
look like (if at all) under BPE start-ups’ ideal of democracy.
Further to this, textual features show that BPE start-ups display
a consideration for economics, with constant reference to
their products’ potential to increase incentive and efficiency.
Altogether, they present an image whereby the economy and
its individualistic concerns under capitalism take precedence
over communitarian, emotional understandings of democratic
association. Finally, the recurring idea that technology has
a complete and utter capacity for transforming democracy
for the better is found as a common trend throughout, and
strongly echoes academic research trends on technological
determinism in the 1990s. BPE start-ups’ belief is so strong
that it relegates any activities to promote democracy on the
part of voters with no technological background to just voting.
This seriously problematizes power relations, entrenched in
prior positions and stagnant at a moment when world history
and political history advance much faster than ever before
(Virilio, 1986).
Overall, it can be concluded that BPE start-ups’ conceptions
of democracy from an individualized, atomized, economic
perspective, solely enabled by technological operations are
antidemocratic according to current definitions. They show
a considerable lack of concern for the communal, emotional
domain that has been crucial in present-day participative
criticism of democracy seeking to reduce the democratic
deficit (e.g., the Spanish Indignados movement; Errejón
and Mouffe, 2016). In summary, while BPE platforms have
a potential to solve problems related to the mechanics
of voting, it is unlikely that, in their present design,
they will contribute to revitalize democracy or advance
democratizing aims.
These conclusions are relevant because the main frame
employed by the selected BPE start-ups emphasizes a broad
and normative message of democratic improvement through
BPE technology, beyond the specific qualities of their product.
The use of Fairclough’s CDA allowed delimitation of the
wide network of power relations involved in choosing
certain framings of the products over others. The authority
and expertise asserted by the texts describing technological
products, such as those analyzed, creates a convoluted
relationship between the creators of the technology, the
targeted ‘buyers’ (e.g., Governments) and the ultimate “users”
of the technology (citizens). My conclusions suggest that this
relationship is one where the final users of the technology are
referred to assiduously, and where the functionality of the
technology in less technologically literate households is not
even considered.
As a continuation of this research, it would be interesting
to conduct a parallel CDA on governments’ views and
understanding of the use of new technologies, such as BPE, for
the future of democratic systems. Setting up a thorough, attentive
and informed dialog between both sides would make us gain
a better appreciation of where views coming from technology
and views coming from government overlap, and whether they
are more mindful of other areas within the broad conception of
democracy sustained here.
To conclude, this work opens up several avenues for future
research, including those on public perceptions of politics
on online platforms and their impact on participation and
voting, online modes of civic engagement with partisan politics
and their democratic outcomes, techno-politics and cyberactivism for a new democratic culture. It may also address
and interrogate internet-mediated channels of communal
participation in local and national politics and its consequences
for current democracy, and might benefit from more overarching
methodological understandings–grounded theory, mixed
methods qualitative research, ethnographies, quantitative
methods, . . .–to achieve more concrete and sound conclusions
within this field of research.
## DATA AVAILABILITY STATEMENT
The original contributions presented in the study are included
in the article/supplementary material, further inquiries can be
directed to the corresponding author/s.
## AUTHOR CONTRIBUTIONS
MI designed, carried out research, and wrote the manuscript.
## ACKNOWLEDGMENTS
MI is the recipient of a “La Caixa” post-graduate fellowship.
MI thanks Dr. Dylan Mulvin (LSE) for early advice and
guidance, the participants in the “Crisis and Challenges of
Democracy” workshop (CES, University of Coimbra, Portugal)
for feedback on a preliminary version of this manuscript, and two
knowledgeable reviewers for their critical insights into current
limitations of BPE technologies regarding fundamental privacy,
security and accountability rights.
-----
## REFERENCES
[Agora (2019a). Agora Web. Available online at: https://agora.vote (accessed July 1,](https://agora.vote)
2019).
[Agora (2019b). Agora Whitepaper. Available online at: https://www.agora.vote/s/](https://www.agora.vote/s/Agora_Whitepaper.pdf)
[Agora_Whitepaper.pdf (accessed July 1, 2019).](https://www.agora.vote/s/Agora_Whitepaper.pdf)
Agre, P. (1994). Networking and Democracy. The Network Observer 1.4. Available
[online at: http://polaris.gseis.ucla.edu/pagre/tno/april-1994.html (accessed July](http://polaris.gseis.ucla.edu/pagre/tno/april-1994.html)
1, 2019).
Allen, D. W. E., Berg, C., and Lane, A. M. (2019). Cryptodemocracy: How
_Blockchain Can Radically Expand Democratic Choice. Lanham, MD: Lexington_
Books.
Altheide, D. L., and Schneider, C. J. (2013). Qualitative Media Analysis, 2nd Edn.
London: SAGE.
American Council for Technology Industry Advisory Council (ACT-IAC)
(2017). Blockchain Primer: Enabling Blockchain Innovation in the U.S. Federal
_[Government, ACT-IAC Whitepaper. Available online at: https://www.actiac.org/](https://www.actiac.org/act-iac-white-paper-enabling-blockchain-innovation-us-federal-government)_
[act-iac-white-paper-enabling-blockchain-innovation-us-federal-government](https://www.actiac.org/act-iac-white-paper-enabling-blockchain-innovation-us-federal-government)
(accessed July 1, 2019).
Ancelovici, M., Dufour, P., and Nez, H. (2016). Street Politics in the Age of Austerity:
_From the Indignados to Occupy. Amsterdam: Amsterdam University Press._
Arslanian, H., and Fischer, F. (2019). The Future of Finance: The Impact of FinTech,
_AI, and Crypto on Financial Services. Switzerland: Palgrave Macmillan._
Baym, N. K. (2015). Personal Connections in the Digital Age. 2nd ed. Cambridge,
MA: Polity Press.
Beck, U. (1992). Risk Society: Towards a New Modernity. London: Sage.
Berryhill, J., Bourgery, T., and Hanson, A. (2018). Blockchains Unchained:
_Blockchain Technology and its Use in the Public Sector, OECD Working Papers_
_on Public Governance, no. 28. Paris: OECD Publishing._
Bijker, W. E. (1995). Of Bicycles, Bakelites, and Bulbs: Toward A Theory of
_Sociotechnical Change. Cambridge, MA: MIT Press._
Bijker, W. E., Hughes, T. P., and Pinch, T. (1987). The Social Construction
_of Technological Systems: New Directions in the Sociology and History of_
_Technology. Cambridge, MA: MIT Press._
Blessing, J., Gomez, J., Patiño, M., and Nguyen, T. (2020). Security Survey and
[Analysis of Vote-by-Mail Systems. arXiv [Preprint]. Available at: https://arxiv.](https://arxiv.org/pdf/2005.08427.pdf)
[org/pdf/2005.08427.pdf (accessed June 30, 2020).](https://arxiv.org/pdf/2005.08427.pdf)
Blommaert, J., and Bulcaen, C. (2000). Critical Discourse Analysis. Annu. Rev.
_Anthropol. 29, 447–466._
Brennan, G., and Buchanan, J. (1984). Voter Choice. Am. Behav. Sci. 2, 185–201.
Brennan, G., and Hamlin, A. (1998). Expressive voting and electoral equilibrium.
_Public Choice 95, 149–175._
Brennan, G., and Lomasky, L. (1993). Democracy and Decision: The Pure Theory of
_Electoral Preference. Cambridge, MA: Cambridge University Press._
Brennan, J. (2016). “The Ethics and Rationality of Voting,” in The Stanford
_Encyclopedia of Philosophy, ed. E. N. Zalta.(Stanford, CA: Stanford University)._
[doi: 10.4159/harvard.9780674497764.c8](https://doi.org/10.4159/harvard.9780674497764.c8)
Castells, M. (1998). The Information Age: Economy, Society and Culture. 3. Malden,
MA: Blackwell Publishing.
Castells, M. (2012). Networks of Outrage and Hope: Social Movements in the Internet
_Age. London: Polity Press._
Corrado, A. (1996). “Elections in Cyberspace: Prospects and Problems,” in Elections
_in Cyberspace: Toward A New Era in American Politics, eds A. Corrado and_
C. M. Firestone (Washington, DC: Aspen Institute).
Crick, B. (2002). Democracy. A Very Short Introduction. Oxford: Oxford University
Press.
Curran, J. (2016). “The Internet of Dreams: Reinterpreting the Internet,” in
_Misunderstanding the Internet, eds J. Curran, N. Fenton, and D. Freedman_
[(London: Routledge), 1–47. doi: 10.4324/9781315695624-1](https://doi.org/10.4324/9781315695624-1)
De Filippi, P., and Hassan, S. (2016). Blockchain Technology as a Regulatory
[Technology: From Code Is Law to Law Is Code. First Monday 21, 12. doi:](https://doi.org/10.5210/fm.v21i12.7113)
[10.5210/fm.v21i12.7113](https://doi.org/10.5210/fm.v21i12.7113)
[Democracy Earth (2019a). Democracy Earth Web. Available online at: https://](https://democracy.earth/)
[democracy.earth/ (accessed July 1, 2019).](https://democracy.earth/)
Democracy Earth. (2019b). Democracy Earth Social Smart Contract Whitepaper.
Available online [at:https://github.com/DemocracyEarth/paper/blob/master/](https://github.com/DemocracyEarth/paper/blob/master/The%20Social%20Smart%20Contract.pdf)
[The%20Social%20Smart%20Contract.pdf (accessed July 1, 2019).](https://github.com/DemocracyEarth/paper/blob/master/The%20Social%20Smart%20Contract.pdf)
Deuze, M. (2007). Media Work. Cambridge, UK: Polity Press.
Diamond, L. (2010). Liberation Technology. J. Democracy 21, 61–83.
Downs, A. (1957). An Economic Theory of Democracy. New York, NY: Harper.
Ellul, J. (1964). The Technological Society. New York, NY: Vintage.
Errejón, I., and Mouffe, C. (2016). Podemos: In the Name of the People. London:
Lawrence & Wishart.
Fairclough, N. (1989). Language and Power. London: Longman.
Fairclough, N. (1992). Discourse and Social Change. Cambridge, UK: Polity Press.
Fairclough, N. (1995). Critical Discourse Analysis: The Critical Study of Language.
London: Longman.
Fischer, C. S. (1992). America Calling: A Social History of the Telephone to 1940.
Berkeley, CA: University of California Press.
Fitzi, G., Mackert, J., and Turner, B. S. (2018). Populism and the Crisis of Democracy.
_3. London: Routledge._
[Follow My Vote (2019a). Follow My Vote Web. Available online at: https://](https://followmyvote.com)
[followmyvote.com (accessed July 1, 2019).](https://followmyvote.com)
Follow My Vote (2019b). Follow My Vote: The Future of Voting. Available online
[at: https://followmyvote.com/the-future-of-voting/ (accessed July 1, 2019).](https://followmyvote.com/the-future-of-voting/)
Follow My Vote (2019c). Follow My Vote: Voting Systems Vulnerabilities. Available
[online at: https://followmyvote.com/voting-system-vulnerabilities/ (accessed](https://followmyvote.com/voting-system-vulnerabilities/)
July 1, 2019).
Foucault, M. (1968). Politics and the Study of Discourse. Ideol. Conscious. 3, 7–26.
Fuchs, C. (2011). Foundations of Critical Media and Information Studies. New York,
NY: Routledge.
Fuchs, C. (2013). “Social Media and Capitalism,” in Producing the Internet:
_Critical Perspective of Social Media, ed. T. Olsson (Gothenburg: Nordicom),_
25–44.
Garnham, N. (1994). “Whatever Happened to the Information Society?,” in
_Management of Information and Communication Technologies: Emerging_
_Patterns of Control, ed. R. Mansell (London: ASLIB), 42–51._
Gibson, J. P., Krimmer, R., Teague, V., and Pomares, J. (2016). A Review of
[E-voting: The Past, Present and Future. Ann. Telecommun. 71, 279–286. doi:](https://doi.org/10.1007/s12243-016-0525-8)
[10.1007/s12243-016-0525-8](https://doi.org/10.1007/s12243-016-0525-8)
Gill, R. (1996). “Discourse Analysis: Practical Implementation,” in Handbook of
_Qualitative Research Methods for Psychology and the Social Sciences, ed. J. T. E._
Richardson (London: BPS-Blackwell), 141–156.
Gill, R. (2000). “Discourse Analysis,” in Qualitative Researching with Text, Image
_and Sound: A Prectical Handbook, eds M. W. Bauer and G. Gaskell (London:_
SAGE), 172–190.
Goodman, R., and Halderman, J. A. (2020). Internet Voting is Happening Now.
_Slate 15:2020._
Gronke, P., Galanes-Rosenbaum, E., Miller, P. A., and Toffey, D. (2008).
Convenience Voting. Annu. Rev. Pol. Sci. 11, 437–455.
Grossback, L. J., Peterson, D. A. M., and Stimson, J. A. (2007). Mandate Politics.
Cambridge, UK: Cambridge University Press.
Grossman, L. K. (1996). The Electronic Republic: Reshaping American Democracy
_for the Information Age. New York, NY: Viking._
Hegel, G. W. F. (1991). Elements of the Philosophy of Right. Cambridge, UK:
Cambridge University Press.
Hillman, A. L. (2010). Expressive Behavior in Economics and Politics. Eur. J. Pol.
_[Econ. 26, 403–418. doi: 10.1016/j.ejpoleco.2010.06.004](https://doi.org/10.1016/j.ejpoleco.2010.06.004)_
International IDEA (2019). The Global State of Democracy 2019: Addressing the Ills,
_[Reviving the Promise. Stockholm: International IDEA, doi: 10.31752/idea.2019.](https://doi.org/10.31752/idea.2019.31)_
[31](https://doi.org/10.31752/idea.2019.31)
International Telecommunications Union (ITU) (2018). _Measuring_ _the_
_Information Society Report 2018, Executive Summary.Available online at:_
[https://www.itu.int/en/ITU-D/Statistics/Documents/publications/misr2018/](https://www.itu.int/en/ITU-D/Statistics/Documents/publications/misr2018/MISR2018-ES-PDF-E.pdf)
[MISR2018-ES-PDF-E.pdf (accessed August 1, 2019).](https://www.itu.int/en/ITU-D/Statistics/Documents/publications/misr2018/MISR2018-ES-PDF-E.pdf)
Jenkins, H. (2008). Convergence Culture: Where Old and New Media Collide.
New York, NY: New York University Press.
Juels, A., Eyal, I., and Naor, O. (2018). Blockchains won’t fix internet voting
security– and could make it worse. Conversation 18:2018.
Kapilkov, M. (2020). Russia Pilots Federal Voting on Waves Blockchain.
_Cointelegraph 19:2020._
Krimmer, R. (2012). The Evolution of E-voting: Why Voting Technology Is Used
_and How It Affects Democracy. Ph.D. Thesis Tallinn: Tallinn University of_
Technology.
Kuenzi, R. (2019). These Are the Arguments That Sank E-voting in Switzerland.
_Swissinfo 2:2019._
-----
Lessig, L. (2006). Code: And Other Laws of Cyberspace. Version 2.0. New York, NY:
Basic Books.
Levin, Y. (2002). Politics After the Internet. Public Interest 149, 80–94.
List, C. (2013). ““Social Choice Theory,”,” in The Stanford Encyclopedia of
_Philosophy, ed. E. N. Zalta.(Stanford, CA: Stanford University)._
MacKenzie, D., and Wajcman, J. (eds) (1989). The Social Shaping of Technology,
2nd Edn. Buckingham, UK: Open University.
Mackie, G. (2015). “Why It’s Rational to Vote,” in Rationality, Democracy,
_and Justice: The Legacy of Jon Elster, eds C. López-Guerra and J._
[Maskivker (Cambridge, UK: Cambridge University Press), 21–49. doi: 10.1017/](https://doi.org/10.1017/cbo9781107588165.005)
[cbo9781107588165.005](https://doi.org/10.1017/cbo9781107588165.005)
Mansbridge, J. J. (ed.) (1990). Beyond Self-Interest. Chicago, IL: University of
Chicago Press.
Markus, M. L. (1994). Finding the Happy Medium: Explaining the Negative Effects
of Electronic Communication on Social Life at Work. ACM T. Inform. Syst. 12,
[119–149. doi: 10.1145/196734.196738](https://doi.org/10.1145/196734.196738)
McAllister, I. (2016). Internet Use, Political Knowledge and Youth Electoral
[Participation in Australia. J. Youth Stud. 19, 1220–1236. doi: 10.1080/13676261.](https://doi.org/10.1080/13676261.2016.1154936)
[2016.1154936](https://doi.org/10.1080/13676261.2016.1154936)
McLean, I., and Hewitt, F. (1994). Condorcet: Foundations of Social Choice and
_Political Theory. Cheltenham, UK: Edward Elgar._
McLuhan, M. (1964). Understanding Media: The Extensions of Man. New York, NY:
McGraw-Hill.
National Academies of Sciences, Engineering, and Medicine (2018). Securing
_the Vote: Protecting American Democracy. Washington, DC: The National_
[Academies Press, doi: 10.17226/25120](https://doi.org/10.17226/25120)
Noel, H. (2010). Ten Things Political Scientists Know that You Don’t. The Forum
[8:12. doi: 10.2202/1540-8884.1393](https://doi.org/10.2202/1540-8884.1393)
Nye, D. E. (1997). Narratives and Spaces: Technology and the Construction of
_American Culture. New York, NY: Columbia University Press._
Papacharissi, Z. (2014). Affective Publics: Sentiment, Technology, and Politics.
_New York. Oxford: Oxford University Press._
Park, S., Specter, M., Narula, N., and Rivest, R. L. (2020). Going from Bad
to Worse: From Internet Voting to Blockchain Voting. Semantic Scholar
_[Preprint]._ Available online at: [https://www.semanticscholar.org/paper/](https://www.semanticscholar.org/paper/Going-from-Bad-to-Worse%3A-From-Internet-Voting-to-Park-Narula/d60e045731228b1118918cbd8cc41f2e19ac143f)
[Going-from-Bad-to-Worse%3A-From-Internet-Voting-to-Park-Narula/](https://www.semanticscholar.org/paper/Going-from-Bad-to-Worse%3A-From-Internet-Voting-to-Park-Narula/d60e045731228b1118918cbd8cc41f2e19ac143f)
[d60e045731228b1118918cbd8cc41f2e19ac143f (accessed June 30, 2020).](https://www.semanticscholar.org/paper/Going-from-Bad-to-Worse%3A-From-Internet-Voting-to-Park-Narula/d60e045731228b1118918cbd8cc41f2e19ac143f)
Poblet, M., Casanovas, P., and Rodríguez-Doncel, V. (2019). Linked Democracy:
_Foundations, Tools, and Applications. Switzerland: Springer Open._
[Polys. (2019a). Polys Web. Available online at: https://polys.me (accessed July 1,](https://polys.me)
2019).
[Polys. (2019b). Polys Whitepaper. Available online at: https://polys.blob.core.](https://polys.blob.core.windows.net/site/Polys_whitepaper.pdf)
[windows.net/site/Polys_whitepaper.pdf (accessed July 1, 2019).](https://polys.blob.core.windows.net/site/Polys_whitepaper.pdf)
Poster, M. (2001). What’s the Matter with the Internet. Minneapolis, MN: University
of Minnesota Press.
Rash, W. Jr. (1997). Politics on the Nets: Wiring the Political Process. San Francisco,
CA: W.H. Freeman.
Richey, S., and Zhu, J. (2015). Internet Access Does Not Improve Political Interest,
[Efficacy, and Knowledge for Late Adopters. Pol. Commun. 32, 396–413. doi:](https://doi.org/10.1080/10584609.2014.944324)
[10.1080/10584609.2014.944324](https://doi.org/10.1080/10584609.2014.944324)
Robins, K., and Webster, F. (1988). “Cybernetic Capitalism: Information,
Technology, Everyday Life,” in The Political Economy of Information, eds V.
Mosko and J. Wasko (Madison, WI: University of Wisconsin press), 45–75.
Schroeder, R. (2007). Rethinking Science, Technology, and Social Change. Palo Alto,
CA: Stanford University Press.
Shirky, C. (2009). Here Comes Everybody: The Power of Organizing Without
_Organizations. London: Penguin._
Specter, M., Koppel, J., and Weitzner, D. (2020). The Ballot is Busted Before the
_Blockchain: A Security Analysis of Voatz, the First Internet Voting Application_
_[Used in U.S.Federal Elections. Available online at: https://internetpolicy.mit.](https://internetpolicy.mit.edu/securityanalysisofvoatz_public/)_
[edu/securityanalysisofvoatz_public/ (accessed November 27, 2020).](https://internetpolicy.mit.edu/securityanalysisofvoatz_public/)
Taylor, S. (2013). What Is Discourse Analysis?. London: Bloomsbury.
Thomason, J., Bernhardt, S., Kansara, T., and Cooper, N. (2019). Blockchain
_Technology for Global Social Change. Hershey, PA: IGI Global._
Toffler, A. (1980). The Third Wave. New York, NY: Bantam.
Toffler, A., and Toffler, H. (1995). Creating A New Civilization: The Politics of The
_Third Wave. Atlanta, GA: Turner._
Van Dijck, J., and Nieborg, D. (2009). Wikinomics and its Discontents: A Critical
[Analysis of Web 2.0 Business Manifestos. New Media Soc. 5, 855–874. doi:](https://doi.org/10.1177/1461444809105356)
[10.1177/1461444809105356](https://doi.org/10.1177/1461444809105356)
Van Niekerk, M. (2019). How Blockchain Strengthened Indonesian Democracy
_(And Could Do The Same Elsewhere). Forbes, May 23, 2019. Available_
online at: [https://www.forbes.com/sites/worldeconomicforum/2019/05/23/](https://www.forbes.com/sites/worldeconomicforum/2019/05/23/how-blockchain-strengthened-indonesian-democracy-and-could-the-same-elsewhere/#6d32cf252c3a)
[how-blockchain-strengthened-indonesian-democracy-and-could-the-same-](https://www.forbes.com/sites/worldeconomicforum/2019/05/23/how-blockchain-strengthened-indonesian-democracy-and-could-the-same-elsewhere/#6d32cf252c3a)
[elsewhere/#6d32cf252c3a (accessed November 28, 2020).](https://www.forbes.com/sites/worldeconomicforum/2019/05/23/how-blockchain-strengthened-indonesian-democracy-and-could-the-same-elsewhere/#6d32cf252c3a)
Virilio, P. (1986). Speed and Politics: An Essay on Dromology. New York, NY:
Columbia University
[Voatz (2019). Voatz Web. Available online at: https://voatz.com (accessed July 1,](https://voatz.com)
2019).
[Votem (2019a). Votem Web. Available online at: https://votem.com (accessed July](https://votem.com)
1, 2019).
[Votem (2019b). Votem Proof of Vote Whitepaper. Available online at: https:](https://github.com/votem/proof-of-vote/blob/master/proof-of-vote-whitepaper.pdf)
[//github.com/votem/proof-of-vote/blob/master/proof-of-vote-whitepaper.pdf](https://github.com/votem/proof-of-vote/blob/master/proof-of-vote-whitepaper.pdf)
(accessed July 1, 2019).
[VoteWatcher (2019). VoteWatcher Web. Available online at: https://votewatcher.](https://votewatcher.com)
[com (accessed July 1, 2019).](https://votewatcher.com)
Warschauer, M. (2003). Technology and Social Inclusion: Rethinking the Digital
_Divide. Cambridge, MA: MIT Press._
Winner, L. (1977). Autonomous Technology: Technics-out-of-Control as a Theme in
_Political Thought. Cambridge, MA: MIT Press._
Winner, L. (1989). “Do Artifacts Have Politics?,” in The Social Shaping of
_Technology, eds D. MacKenzie and J. Wajcman (Buckingham, UK: Open_
University), 28–40.
Wodak, R. (2014). “Critical Discourse Analysis,” in The Routledge Companion
_to English Studies, eds C. Leung and B. V. Street (London: Routledge),_
302–317.
Wodak, R., and Meyer, M. (2009). “Critical Discourse Analysis: History, Agenda,
Theory, and Methodology,” in Methods for Critical Discourse Analysis, eds R.
Wodak and M. Meyer (London: SAGE), 1–33.
**Conflict of Interest: The author declares that the research was conducted in the**
absence of any commercial or financial relationships that could be construed as a
potential conflict of interest.
_Copyright © 2021 Imperial. This is an open-access article distributed under the_
_[terms of the Creative Commons Attribution License (CC BY). The use, distribution](http://creativecommons.org/licenses/by/4.0/)_
_or reproduction in other forums is permitted, provided the original author(s) and_
_the copyright owner(s) are credited and that the original publication in this journal_
_is cited, in accordance with accepted academic practice. No use, distribution or_
_reproduction is permitted which does not comply with these terms._
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3389/fbloc.2021.587148?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3389/fbloc.2021.587148, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "GOLD",
"url": "https://www.frontiersin.org/articles/10.3389/fbloc.2021.587148/pdf"
}
| 2,021
|
[
"JournalArticle"
] | true
| 2021-04-09T00:00:00
|
[
{
"paperId": "64e4b6f346ecefcf2a6dc6d6ec1a7fb69a467bff",
"title": "Agora"
},
{
"paperId": "d60e045731228b1118918cbd8cc41f2e19ac143f",
"title": "Going from bad to worse: from Internet voting to blockchain voting"
},
{
"paperId": "7c2e557b73c6e35918dbd0c6b2f0e7c258a63aa7",
"title": "Security Survey and Analysis of Vote-by-Mail Systems"
},
{
"paperId": "c819aa103ab314e07e3195ca78cb94e91edc5035",
"title": "SUMMARY: The Global State of Democracy 2019: Addressing the Ills, Reviving the Promise"
},
{
"paperId": "358aacb6585533dbbd7c9d9edf132d1a102e965d",
"title": "The Global State of Democracy 2019: Addressing the Ills, Reviving the Promise"
},
{
"paperId": "bcd1120eb19d1005bbd107629a256ac1d4848a22",
"title": "Critical Discourse Analysis"
},
{
"paperId": "a6e68731e074bba0809e0f1bf33211fab816ec48",
"title": "Correction to: The Future of Finance"
},
{
"paperId": "b46498b49b1c57fc172a2f6c16fc358d26f5a1ac",
"title": "Blockchain Technology for Global Social Change"
},
{
"paperId": "9341439ce1ef804745f45c54d29faf4894ef22d1",
"title": "Linked Democracy"
},
{
"paperId": "884cf84d7ca40cf21d9e241182b9e7723eb819be",
"title": "Discourse Analysis"
},
{
"paperId": "448d003323009d5c2f00b8f6540cef11b6b5c595",
"title": "Populism and the Crisis of Democracy"
},
{
"paperId": "c4ee3016d74645f68c53f3b7c42edbe1568922f3",
"title": "Securing the Vote"
},
{
"paperId": "6d7edb69dd35d75d4e3797fbd36f78f56626c85f",
"title": "Do Artifacts Have Politics?"
},
{
"paperId": "b6d70017be9f13e1406f32b2dbdf53e5d702209b",
"title": "Language and Power"
},
{
"paperId": "2b00630d77423fca5204ec91a6f8c6b280df0a6e",
"title": "Blockchain Technology as a Regulatory Technology: From Code is Law to Law is Code"
},
{
"paperId": "81ac8abee548ab937b68b3bd03d27e57b2705f5a",
"title": "A review of E-voting: the past, present and future"
},
{
"paperId": "4db268e806c077f00c7e5de5bf576da7d453f933",
"title": "Internet use, political knowledge and youth electoral participation in Australia"
},
{
"paperId": "8981142c5fe8f5350e7af234579fc115130fd359",
"title": "The internet of dreams"
},
{
"paperId": "b818c6b15d50dc92b2a4b9aa9b890a36ea4bd9dc",
"title": "Internet Access Does Not Improve Political Interest, Efficacy, and Knowledge for Late Adopters"
},
{
"paperId": "5b752d7cf98bf39fd6183ee19a28c7b59a467640",
"title": "Rationality, Democracy, and Justice: Why It’s Rational to Vote"
},
{
"paperId": "2900d2e581445f80eb3ecb75b587473e6dcedd08",
"title": "Affective Publics: Sentiment, Technology, and Politics"
},
{
"paperId": "f9247a56c2a88372e62f288505fc74858444e081",
"title": "NETWORKS OF OUTRAGE AND HOPE. SOCIAL MOVEMENTS IN THE INTERNET AGE, Manuel Castells, Cambridge, Polity Books, 2012, 200 pages"
},
{
"paperId": "6e75e39143e6a9e815fa6a4e33a293f9b208efb1",
"title": "The Evolution of E-voting: Why Voting Technology is Used and How it Affects Democracy."
},
{
"paperId": "9bddc8909096e0b42372a91d1dd880ed3ee561f5",
"title": "Foundations of Critical Media and Information Studies"
},
{
"paperId": "fc40c483e9641570651d7a52b95e70a4e6849954",
"title": "Expressive behavior in economics and politics"
},
{
"paperId": "e0411b7d700be9220ab97522b32713cef483de12",
"title": "Ten Things Political Scientists Know that You Don't"
},
{
"paperId": "c34b6a75becaac5f7eabed8b1b676806468ce15e",
"title": "Liberation Technology"
},
{
"paperId": "764306bbe00cdd72f8499d1e7f994b037759188c",
"title": "Wikinomics and its discontents: a critical analysis of Web 2.0 business manifestos"
},
{
"paperId": "baa56de8baeeb9936a4bc8f1033cc1e9d503622e",
"title": "The Risk Society: Towards a new modernity"
},
{
"paperId": "0e0f61eccd6b7c4f6557a7a28a17ac3ea44c1315",
"title": "Here Comes Everybody: The Power of Organizing Without Organizations"
},
{
"paperId": "c1146e3da50846519731d7bf24e01b9df61d3129",
"title": "Rethinking Science, Technology, and Social Change"
},
{
"paperId": "101a822431ac3566abc045a0dbe3c5d3850f5566",
"title": "Convergence Culture: Where Old and New Media Collide"
},
{
"paperId": "932f077b69006b951a9e7eda7407281c61b43074",
"title": "Mandate Politics"
},
{
"paperId": "af46eae8afb0838019958fba220158a7705e71ed",
"title": "Technology and Social Inclusion. Rethinking the Digital Divide"
},
{
"paperId": "53e02324935b45df6a31c79888a56cb680635ae4",
"title": "Code and Other Laws of Cyberspace"
},
{
"paperId": "8d19aa993deb672b118f3ff6d0dda16e096d73d2",
"title": "The Social Shaping of Technology (2nd ed.)"
},
{
"paperId": "fc46b74c477643dbbeef3f9ea7498728e7d3f29c",
"title": "Democracy: A Very Short Introduction"
},
{
"paperId": "38939cc515ac0abf6581c884fb6215754062f58d",
"title": "The Information Age: Economy, Society and Culture, 3 vols ‐ Vol. 1: The Rise of the Network Society"
},
{
"paperId": "5801ee5b383624a8e79689e3f3210f18b0e7b33a",
"title": "Democracy"
},
{
"paperId": "55fe27dd0fcb64464c55834f13e3b3a275099612",
"title": "The ‘Third Wave’"
},
{
"paperId": "98c02040fc5f8332ff32c0d447e40d483bb8bc2e",
"title": "CRITICAL DISCOURSE ANALYSIS"
},
{
"paperId": "a137d5157f16b2359df7194ea8c5c5517d7badb9",
"title": "Expressive voting and electoral equilibrium"
},
{
"paperId": "8e778dd7b2b6ab815d917954965e21097ca1bc92",
"title": "Democracy and Decision: The Pure Theory of Electoral Preference"
},
{
"paperId": "7b50810157e299a92bdf1872f4935b82f723c98c",
"title": "Politics on the nets: wiring the political process"
},
{
"paperId": "2882f1830e4e6f37c5bc30496a5542bf06d06dd8",
"title": "Of Bicycles, Bakelites, and Bulbs: Toward a Theory of Sociotechnical Change"
},
{
"paperId": "c8a263a051cd2eccadb7d0c51871883e0e4a335b",
"title": "Condorcet: Foundations of Social Choice and Political Theory"
},
{
"paperId": "ec69dd0c01b254f65a9573bee1e5baa9bcc4d264",
"title": "Democracy and Decision: The Pure Theory of Electoral Preference. By Geoffrey Brennan and Loren Lomasky. New York: Cambridge University Press, 1993. 237p. $44.95."
},
{
"paperId": "80b43e1c43a04491b6dc2763190cdf6e8f4e4ed8",
"title": "Personal Connections in the Digital Age"
},
{
"paperId": "f28cc2ba39a71d9a915f5ddfc33a422f1e5ae126",
"title": "Finding a happy medium"
},
{
"paperId": "2d10897a990b306e7203f387e981daa47060739d",
"title": "America Calling: A Social History of the Telephone to 1940"
},
{
"paperId": "633557d9e954b95d9a7da2d6faea1a8b786f69aa",
"title": "America calling: A social history of the telephone to 1940 by Claude S. Fischer University of California Press, Berkeley, CA, 1992, 424 pp, $25.00"
},
{
"paperId": "7389a18d0c18d12dac817871e46206b57f2f89d2",
"title": "Megatrends or megamistakes?: whatever happened to the information society?"
},
{
"paperId": "eabbd687a5258ab39975b690ff389e961375f18e",
"title": "Hegel: Elements of the Philosophy of Right"
},
{
"paperId": "be1eb642f5d6d7cf2d3f4e9b9a6274df6596ae83",
"title": "Beyond self-interest"
},
{
"paperId": "f5a1804de8b7cd550190a10f9b5120b74a434e28",
"title": "The social construction of technological systems: new directions in the sociology and history of technology"
},
{
"paperId": "d58ca9814f94799451f1a8a9d9ff5d575815cd27",
"title": "Speed and Politics: An Essay on Dromology"
},
{
"paperId": "3ed5797d655ab5e45ef416532236f2aee74e5ed5",
"title": "Voter Choice"
},
{
"paperId": null,
"title": "Conflict of Interest: The author declares that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest"
},
{
"paperId": "d6a5afd47a5bddc669399dc299c11ab8ac3368c2",
"title": "The Ballot is Busted Before the Blockchain: A Security Analysis of Voatz, the First Internet Voting Application Used in U.S. Federal Elections"
},
{
"paperId": null,
"title": "Russia Pilots Federal Voting on Waves Blockchain."
},
{
"paperId": null,
"title": "Internet Voting is Happening Now"
},
{
"paperId": "47d632f1083801c4411a87378a24ad038a24bda1",
"title": "”Networks of Outrage and Hope. Social Movements in the Internet Age”."
},
{
"paperId": null,
"title": "Votem Proof of Vote Whitepaper"
},
{
"paperId": null,
"title": "Voatz"
},
{
"paperId": null,
"title": "These Are the Arguments That Sank E-voting in Switzerland."
},
{
"paperId": null,
"title": "The Future of Finance: The Impact of FinTech, AI, and Crypto on Financial Services"
},
{
"paperId": null,
"title": "Cryptodemocracy: How Blockchain Can Radically Expand Democratic Choice"
},
{
"paperId": null,
"title": "Follow My Vote (2019a)"
},
{
"paperId": null,
"title": "Polys"
},
{
"paperId": null,
"title": "VoteWatcher"
},
{
"paperId": null,
"title": "Blockchains Unchained: Blockchain Technology and its Use in the Public Sector"
},
{
"paperId": null,
"title": "Measuring the Information Society Report 2018, Executive Summary"
},
{
"paperId": null,
"title": "Blockchains won’t fix internet voting security– and could make it worse."
},
{
"paperId": null,
"title": "Blockchain Primer: Enabling Blockchain Innovation in the U.S. Federal Government"
},
{
"paperId": "9d67bf0ee1c58c8aaa82ed71d8a38bccd000c06e",
"title": "Street Politics in the Age of Austerity"
},
{
"paperId": "848f86ba7996e3730b3ca486d3f320c50b150c15",
"title": "Podemos: In the name of the people"
},
{
"paperId": null,
"title": "StreetPoliticsintheAgeofAusterity: From the Indignados to Occupy"
},
{
"paperId": "984e25e67cbddfccdac7dcca320b621a0e64a81c",
"title": "An Economic Theory of Democracy"
},
{
"paperId": "6594dc608fb02d1d96b8074c8cff65afddfd9b11",
"title": "Social media and capitalism"
},
{
"paperId": "73f801d960e994768eca44df83d0b7ad19b613f5",
"title": "Convenience Voting"
},
{
"paperId": null,
"title": "Wasko (Madison, WI: University of Wisconsin press), 45–75"
},
{
"paperId": "129add587072f63a35729e324469c77025411215",
"title": "Social Construction"
},
{
"paperId": "6e8c43edc9b1a30bd955921592db41b5379ed81c",
"title": "What's the Matter with the Internet?"
},
{
"paperId": "e74e34dc50982ebc8df913498061f4bebb6c490b",
"title": "Code and Other Laws of Cyberspace"
},
{
"paperId": "e34404b00f0d3d91e53eba5c7b85a99eae6033c5",
"title": "Narratives And Spaces: Technology and the Construction of American Culture"
},
{
"paperId": "a4a7698ebd88ab0b8f2183371901323ddd66ceaf",
"title": "Qualitative Media Analysis"
},
{
"paperId": null,
"title": "The Electronic Republic: Reshaping American Democracy for the Information Age"
},
{
"paperId": null,
"title": "Elections in Cyberspace: Prospects and Problems"
},
{
"paperId": "70fd2e37b5869499374e486c49ccbf9be88b407c",
"title": "Critical Discourse Analysis: The Critical Study of Language"
},
{
"paperId": "93f9f42f347989ae79b0624c8a2376767ebe0b76",
"title": "Creating a new civilization : the politics of the Third Wave"
},
{
"paperId": "d0e3b23f1b4701d269ec4325c4f7f358c61441ea",
"title": "Democracy and Decision"
},
{
"paperId": "4abd90f1c03b88c675e341c028ad2a64abf90f55",
"title": "Discourse and social change"
},
{
"paperId": "b5c44f899edde1b5676d63d632ce912e54bec46e",
"title": "The social shaping of technology"
},
{
"paperId": null,
"title": "Cybernetic Capitalism: Information"
},
{
"paperId": "ce11759cc4047695a847c7def648e5ad9ef0dbf7",
"title": "Social Choice Theory"
},
{
"paperId": "ee2f5b689495f2eaca6efb6ca9cabb726e6dbf99",
"title": "Autonomous Technology: Technics-out-of-Control as a Theme in Political Thought"
},
{
"paperId": "02a83c02fde6852681304d430649a87122627b57",
"title": "CHAPTER 6. CONCLUSIONS"
},
{
"paperId": null,
"title": "UnderstandingMedia:TheExtensionsofMan"
},
{
"paperId": "66496d400a82b87cb706bff8e9b9ca3e86d599a1",
"title": "Elements of the Philosophy of Right"
},
{
"paperId": null,
"title": "Frontiers in Blockchain | www."
},
{
"paperId": null,
"title": "Blockchain-Powered E-Voting Start-Ups"
},
{
"paperId": null,
"title": "Networking and Democracy. The Network Observer 1.4"
},
{
"paperId": null,
"title": "Follow My Vote: Voting Systems Vulnerabilities"
},
{
"paperId": null,
"title": "Follow My Vote: The Future of Voting"
},
{
"paperId": null,
"title": "Votem Web"
},
{
"paperId": null,
"title": "How Blockchain Strengthened Indonesian Democracy (And Could Do The Same Elsewhere)"
}
] | 21,545
|
en
|
[
{
"category": "Economics",
"source": "s2-fos-model"
},
{
"category": "Business",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/02324eda5842b7144859126bf152495e7c8a415d
|
[] | 0.881789
|
International Capital Flows, Dynamic Changes in Cryptocurrency and Noble Metal Markets
|
02324eda5842b7144859126bf152495e7c8a415d
|
BCP Business & Management
|
[
{
"authorId": "2068168295",
"name": "Ruize Sun"
}
] |
{
"alternate_issns": null,
"alternate_names": null,
"alternate_urls": null,
"id": null,
"issn": null,
"name": null,
"type": null,
"url": null
}
|
In 2022, with the implementation of tightening monetary policies by FOMC, US dollar is experiencing a dramatic appreciation in a very short period. Though numerous studies have demonstrated the connection between the traditional currency market, cryptocurrency market, and precious metal market, rare studies are exploring the relationships between the three markets under a special political environment. This paper selects USDCNY exchange rate, gold and silver, and bitcoin as the representatives of three markets and then tests the volatility response of return on gold & silver and return on bitcoin to the change of return on USDCNY exchange rate. By employing impulse response function and ARMA-GARCHX model, the paper verifies the change of exchange rate will exacerbate the volatility of returns on gold & silver and bitcoin significantly, which suggests high risk and uncertainty of the cryptocurrency market and precious metal market in a complex and extreme political environment. Investors and speculators should take prudent investment strategies in such environment.
|
BCP Business & Management **AFTEM 2022**
Volume **32** (2022)
# **International Capital Flows, Dynamic Changes in ** **Cryptocurrency and Noble Metal Markets **
## Ruize Sun*
School of Business, University of Leicester, Leicester, LE2 7RH, UK
*Corresponding author: rs673@student.le.ac.uk
**Abstract.** In 2022, with the implementation of tightening monetary policies by FOMC, US dollar is
experiencing a dramatic appreciation in a very short period. Though numerous studies have
demonstrated the connection between the traditional currency market, cryptocurrency market, and
precious metal market, rare studies are exploring the relationships between the three markets under
a special political environment. This paper selects USDCNY exchange rate, gold and silver, and
bitcoin as the representatives of three markets and then tests the volatility response of return on gold
& silver and return on bitcoin to the change of return on USDCNY exchange rate. By employing
impulse response function and ARMA-GARCHX model, the paper verifies the change of exchange
rate will exacerbate the volatility of returns on gold & silver and bitcoin significantly, which suggests
high risk and uncertainty of the cryptocurrency market and precious metal market in a complex and
extreme political environment. Investors and speculators should take prudent investment strategies
in such environment.
**Keywords:** Exchange Rate, Precious Metal, Bitcoin, Volatility, ARMA-GARCHX.
## **1. Introduction **
### Intending to suppress the price pressure generated from high inflation, the Federal Open Market Committee (FOMC) planned to take a series of aggressive monetary contractions, and thus FOMC announced to raise the interest rate at FOMC meetings and scheduled to implement this monetary policy in March, May, July, September, and November. Until the 28th, of July, FOMC has raised 25 basic points, 50 basic points, and 75 basis points in March, May, and June respectively. With the tightening monetary policies, predictably, the US dollar is experiencing a soring appreciation. Since the Chinese Central Bank keeps the interest rate constant in 2022, this paper proposes to set CNY as a proxy to reflect the real-time value of USD. As shown in fig 1, the exchange rate of CNYUSD was maintained at around 6.34 before the first interest rise occurred in March. Corresponding with the interest rise, the exchange rate increased dramatically in March, April, and May, and then been stable in June, at around 6.72. With the skyrocketing appreciation of the dollar in such a short period, not just the commodity sector, but all the financial sectors are experiencing a huge shock, via various channels, for instance, price volatility, risk, and expectative return of financial assets might change greatly with a sudden appreciation of the dollar. Figure 1 the exchange rate between USD and CNY
231
-----
BCP Business & Management **AFTEM 2022**
Volume **32** (2022)
### Even Bretton Woods System has become a far-distance memory, the precious metals are still connected with USD closely, as the price of precious metals is normally bided by USD. Generally, precious metals, especially gold and silver, are recognized as both a kind of commodity and a special currency. From commodity aspect, gold and silver are scarce with great intrinsic value, which are also important industrial raw materials; from currency aspect, gold and silver have been perceived as a sample of wealth for thousands of years, which indicated they have the capacity of value storage and more they are used as mediums of exchange or quote of goods’ value. Because of the bivariate attributes of precious metals, the USD price of gold and silver might be sensitive to the change in the USD exchange rate and various macroeconomic factors. Becher and Soenen discover the USD price of gold rise accompanied by the depreciation of USD relative to other foreign currencies, this relationship has also been verified by Sjasstad and Scacciavillani through a comparison of the USD price of gold with DM (Deutsche mark) price of gold when the exchange rate of USDDM decreased [1, 2]. Then, they uncover the volatility of gold prices generated from floating exchange rates among the major currencies mainly. However, Pukthuanthong, and Rol, by using the VAR model and Granger Causality Analysis, prove that the negative relationship between change in the gold price and the change of the value of bided currencies is trivial, while the volatility of USD price, JPY price, DM price, and GBP price are similar [3]. Otherwise, via the asymmetry-powered GARCH model, Tully and Lucey confirm that the dollar effective exchange rate is the most influential factor to impact the mean of return on gold, while the exchange rate falls to explain the variance and fluctuation of gold price [4]. However, interest rate cannot explain either mean or variance of the return on gold, demonstrating gold price has limited relation to the interest rate. Interestingly, the coefficient of 1- lagged autoregression conditional variance is significant at 1% level and has the largest absolute value among all the coefficients, implying the high volatility persistence of gold price. This result is identical to the finding of Hammoudeh and Yuan, who also verify the high volatility persistence of return on gold and silver [5]. But on the contrary, the impact of interest rate on gold and silver is significant and dampening, not just mean but variance, similar to the conclusion of Hashim et al. [6]. Consequently, the interest rate and USD exchange rate have an impact on gold and silver, both price and volatility, however, the strength of this impact is unstable, and might shift over time. Bitcoin is firstly introduced by Nakamoto in 2008, to be designed as a new substitute for conventional currency. However, the nature of bitcoin still confused scholars and economists today because it is ambiguous and overlaps with various financial fields. In the article of Dyhrberg, she describes bitcoin as “the asset between gold and traditional currency”, since bitcoin shares many similar attributes with gold [7]. Like gold, bitcoin is scarce as the total amount of bitcoins has been decided by the algorithm, but it lacks intrinsic value, at this point, bitcoin approaches more with conventional currency, like dollar. Additionally, bitcoin has some unique features from both gold and traditional currency, first, it is decentralized and has no related departures or organizations to monitor the market of bitcoin. Although decentralization endows the very high liquidity for bitcoin, the risk of bitcoin is also high because of the high possibility of fraud and manual manipulation; plus, the value or credit of bitcoin is not guaranteed by any legal regimes or blocs, which implies the value of bitcoin are utterly determined by the market, and thus this feature brings high instability and volatility of its price. Geuder et al. have demonstrated that cryptocurrencies are a kind of special speculative asset, as their price is associated with bubble behaviors significantly [8]. Baur et al. and Corber et al. respectively confirm that bitcoin is insolated and has limited connection with other financial assets, for example, gold and oil futures, which manifests bitcoin’s failure to be a hedge asset but more suitable as risk diversification [9] [10]. The latest research from Kwon, via exploring the tail behaviors of bitcoin, gold, and dollar, illustrates the significant negative correlation between bitcoin and dollar while rejecting the similarity between bitcoin and gold, according to different tail features [11]. The conclusion may emphasize the currency and investment attribute of bitcoin. Consequently, the feature of bitcoin indicates it is a speculative asset mainly but not either a traditional currency or commodity, hence its price is highly uncertain and will be volatile with the fluctuation of other financial factors, such as interest rate and exchange rate.
232
-----
BCP Business & Management **AFTEM 2022**
Volume **32** (2022)
### With previous research, the general interactions between exchange market, precious metal market, and cryptocurrency market are clear, however, there is rare research to test the interaction between those three markets, within some unusual scenarios. As in a specific period, the interaction model and volatility relativity are quite different from normal time, for exchange rate, bitcoin, and gold. For both investors and policymakers, risk management is always important, therefore scholars must clarify the risk of assets in some unusual periods to avoid irrational investment or policy making. Whereby, the main purpose of this paper is to examine the risk sensitivity of returns on gold, silver, and bitcoin with the change of return on USDCNY exchange rate, within the monetary contraction period of 2022. By employing the impulse response function and ARMA-GARCH model, the author discovers that: the change of return on the USDCNY exchange rate has a significant impact on both means and volatility of return on gold, silver, and bitcoin. And the conclusion further manifests that investors shall be more prudent to invest financial assets in such period. The rest of this paper is organized as follows: Part 2 is research design, including data sources, unit root test, and identification strategy; Part 3 reports empirical results; Part 4 is discussion and part 5 is conclusion.
## **2. Research Design **
### **2.1 Data sources ** As FOMC starts to implement tightening monetary policies in 2022, to avoid interference from previous period, this paper collects data from January 4th to July 28th with 137 observations for each variable, and the four selected variables are the logarithm of the daily closing price of gold, silver, bitcoin, and USDCNY exchange rate. Intending to ensure the continuance of time-series with exchange rate, the trading price of gold, silver, and bitcoin at weekends have been excluded from the sample. **2.2 Unit Root Tests ** **2.2.1 ADF-test ** When making a time-series examination, the necessary and sufficient condition of the test shall be that the sequence is time-series stationery. As if the sequence is nonstationary and drifts randomly, the sequence will fail to cluster at the expectation of the sequence, and the past information and innovation will interfere with the result permanently. Dicky D.A and Fuller W.A. propose a method to test the unit root process of sequence, also known as DF test, [12]. 𝑝 𝑡 = 𝜙 0 + 𝜙 1 𝑝 𝑡−1 + 𝜀 𝑡 (1) Equation (1) is a standard 1 lag autoregression model, the stationary condition is that the coefficient 𝜙 1 < 1 Using ordinary least squares estimate, the coefficient 𝜙 1 is:
### 𝜙̂ = 1
𝑇
### ∑ 𝑡 = 1 𝑝 𝑡−1 𝑝 𝑡 = ∑ 𝑇𝑡=1 𝑝 𝑡 [2], 𝜎̂ [2]
### 1 𝑇1 ̂𝑝 𝑡−1 ) [2] (2) 𝑇−1 [∑(𝑝] [𝑡] [−𝜙] [1]
### As P0=0, and T is the sample quantity, then making DF-test:
𝑇
∑ 𝑡 = 1 𝑝 𝑡−1 𝑒 𝑡
𝑇
𝑡=1 𝑝 𝑡−1 [2]
𝜎̂√∑
### 𝐷𝐹=
𝜙 ̂ −1 1
### 𝑆𝐸(𝜙̂ ) 1 [=]
### (3)
### And hypothesis is 𝐻 0 : 𝜙 1 = 1, and 𝐻 1 : 𝜙 1 < 1 when p-value is small enough to reject hypothesis null, that indicates there is no unit root in sequence. However, many sequences in finance and
233
-----
BCP Business & Management **AFTEM 2022**
Volume **32** (2022)
### economy cannot be described as random drafting processes and they are fitted with autoregressive integrated moving average model (ARIMA), therefore the formula of such sequence. (4) Therefore, when 𝛽= 1, equation (4) is AR (p-1) model of Δ𝑋 𝑡 When 𝛽< 1, equation (4) is AR (p) model of 𝑋 𝑡 And then adjusted DF-test is:
### 𝐴𝐷𝐹=
𝛽 [̂] −1
### 𝑆𝐸(𝛽 [̂] ) [ (5) ]
### And hypothesis is: 𝐻 0 : 𝛽= 1, and 𝐻 1 : 𝛽< 1 rejects null hypothesis when there is no unit root process. **2.2.1 Test results ** Table 1 exhibits the result of ADF-test of samples and their first-order differences. The origin sample sequences fail to reject the null hypothesis of ADF-test and indicate that random drafting process is existing in those sequences, while the t-values of first-order differences are significant at 1% level and demonstrate the stationary of logarithm of assets’ yield rate. Table 1 ADF test Variables t - statistic p - value Price Gold -1.831 0.6895 Silver -1.970 0.6174 BTC -2.079 0.5579 Exchange rate -1.804 0.7030 Yield Gold -9.571 0.0000 [***] Silver -8.317 0.0000 [***] BTC -7.623 0.0000 [***] Exchange rate - 7.618 0.0000 [***] **2.3 Identification strategy ** As a variable could be influenced by not just other independent variables, but also the variables from past time, a autoregressive model could be built to describe this process: 𝑋 𝑡 = 𝜙 0 + ∑ 𝑃𝑗=1 𝜙 𝑗 𝑋 𝑡−𝑗 + 𝜀 𝑡 (6) And 𝜙 0 is constant, 𝜙 𝑗 is j-lagged autoregressive coefficient, only decided by lagged order but not time t; 𝜀 𝑡 is innovation of 𝑋 𝑡, which is also white noisy with independent and identically distribution, usually obey {0,1} normal distribution. But as the sequence is a weakly stationary process, the expectation of each 𝑋 𝑡 shall be all the same, therefore 𝑋 𝑡 could be perceived as result of both expectation and accumulation of innovations from past time. Generally, such a sequence could be described by Moving Average model (MA), and the formula of it is:
234
-----
BCP Business & Management **AFTEM 2022**
Volume **32** (2022)
### 𝑋 𝑡 = 𝜃 0 + ∑ 𝑝𝑗=1 𝜃 𝑗 𝜀 𝑡−𝑗 + 𝜀 𝑡 (7) And 𝜃 0 is expectation of variance X, 𝜃 𝑗 is the coefficient of j-lagged innovation, 𝜀 𝑡−𝑗 is the innovation of 𝑋 𝑡−𝑗, which is white noisy with independent and identical distribution, like 𝜀 𝑡 . When combining both AR(p) model and MA(p) model to describe sequence, the autoregressive moving average model (ARMA) is 𝑋 𝑡 = 𝜙 0 + ∑ 𝑃𝑗=1 𝜙 𝑗 𝑋 𝑡−𝑗 + ∑ 𝑞𝑙=1 𝜃 𝑙 𝜀 𝑡−𝑙 + 𝜀 𝑡 (8) Set 𝐵 is backshift, as 𝐵 [𝑗] 𝑋 𝑡 = 𝑋 𝑡−𝑗, therefore ARMA (p,q) could be rewritten to be MA(q): 𝑋 𝑡 −∑ 𝑃𝑗=1 𝜙 𝑗 𝑋 𝑡−𝑗 = 𝜙 0 + ∑ 𝑞𝑙=1 𝜃 𝑙 𝜀 𝑡−𝑙 + 𝜀 𝑡 (9) (10) (11) 𝑋 𝑡 = 𝜇+ (∑ ∞𝑗=0 𝜓 𝑗 𝐵 [𝑗] )𝜀 𝑡 (12)
𝜕𝑋 𝑡+𝑙
### And 𝜓 0 = 1, equation (12) is also known as Wold expression of ARMA (p, q), and 𝜓 𝑗 = 𝜕𝜀 𝑡 is impulse response function of ARMA (p, q), it means to bring 𝑋 𝑡+𝑙 an additional variable 𝜓 𝑗 when 𝜀 𝑡 = 1. When considering multiple vectors in AR(p) model, that is vector autoregressive model (VAR) and its formula is: (13) (14) And 𝑟 𝑡 is vector of k variables at time t, Φ 𝑗 is j-lagged cross-correlation matrix between vector 𝑟 𝑡 and vector 𝑟 𝑡−𝑗, and single unit of it is 𝜙 𝑎𝑏,𝑡−𝑗 = 𝑐𝑜𝑟𝑟(𝑟 𝑎,𝑡, 𝑟 𝑏,𝑡−𝑗 ), when a=b, the cross correlative coefficient is j-lagged autoregressive coefficient of variable 𝑟 𝑎,𝑡 ; 𝜙 0 is a 1*k constant matrix; 𝑎 𝑡 is innovation matrix Similar to AR(p) model, the characteristic function of VAR(p) is:
235
-----
BCP Business & Management **AFTEM 2022**
Volume **32** (2022)
### I −∑ 𝑝𝑗=1 Φ 𝑗 𝑍 [𝑗] (15) Therefore, as VAR(p) is a stationary sequence, the cross-correlative matrix should cluster to the null matrix when 𝑗= ∞, the stability condition of VAR(p) is that: First, transfer VAR(p) into VAR (1)
### 𝑟 𝑡 𝑟 𝑡−1 𝑟́ 𝑡 ⋮ 𝑟 = [ 𝑡−𝑝
𝑡
### ], 𝑎́ = [
### 𝑎 𝑡 𝑎 𝑡−1 ⋮ 𝑎 𝑡−𝑝
### = ] Φ [∗]
### Φ 2 ⋯ Φ 𝑝−1 Φ 𝑝 [Φ] [1] 𝐼 0 ⋯ 0 0 0 𝐼 ⋯ 0 0 𝑟́ 𝑡 = Φ [∗] 𝑟́ 𝑡−1 + 𝑏 [́] 𝑡 (16) ⋮ ⋮ ⋱ ⋮ ⋮ [ 0 0 ⋯ 𝐼 0 ]
### Establish the characteristic polynomial of equation (16). 𝜆𝐼−Φ [∗] (17) Calculate the solution of equation (17) when det(Φ [∗] −𝜆𝐼) = 0, the VAR(p) model is stable if all the solutions are inside the unit circle. Following equation (12) the Wold expression of ARMA (p, q), the Wold expression of VAR(p) is:
∞
### 𝑟 𝑡 = 𝜇+ ∑ 𝑗=0 Ψ 𝑗 𝑎 𝑡−𝑗 (18) And thus the Ψ 𝑗 is impulse response matrix of 𝑟 𝑡 . Generally, in a typical AR process, shown in equation (6), the residual 𝜀 𝑡 should be white noisy with independent and identical distribution, however, in financial research, some sequences have volatility clustering phenomenon, because the residual 𝑎 𝑡 is usually heteroskedastic and correlative with each other, despite they are nonlinear. Engle first develops introgressive conditional heteroskedasticity model (ARCH) and proposes to decompose the innovation 𝑎 𝑡 in to two parts: conditional standard variance 𝜎 𝑡 and white noisy 𝜀 𝑡, while the conditional variance could be linearly expressed as [13]: 𝑎 𝑡 = 𝜎 𝑡 𝜀 𝑡 2 2 2 (19) {𝜎 𝑡 = 𝛼 0 + 𝛼 1 𝑎 𝑡−1 + ⋯+ 𝛼 𝑚 𝑎 𝑡−𝑚 𝛼 0 is constant of variance, and 𝛼 𝑚 is the coefficient of past innovation, which means the extent of present volatility could be explained by past fluctuations. However, in empirical work, conditional variance may need high lags to be described with ARCH model and that will bring a negative impact on model from two sides: 1. High lags need more parameters in model and it will reduce the simple’s degree of freedom significantly, 2. Information penalty from a high degree of complexity. Bollerslev expends the ARCH to generalized ARCH (GARCH), considering autoregression of conditional variance itself. Interestingly, standard GARCH model has highly constructive similarity with ARMA model, as [14]:
2 𝑚 2 𝑠 2
### 𝜎 𝑡 = 𝛼 0 + ∑ 𝑙=1 𝛼 𝑙 𝑎 𝑡−𝑙 + ∑ 𝑗=1 𝛽 𝑗 𝜎 𝑡−𝑗 (20) The 𝛽 𝑗 is autoregressive coefficient of conditional variance. Further, in order to explore the contribution of exogenous variables to the volatility of financial assets, GARCH model with additional distributed lag term (GARCHX) could be better than standard GARCH model, and formula of GARCHX is:
236
-----
BCP Business & Management **AFTEM 2022**
Volume **32** (2022)
### 𝜎 𝑡2 = 𝛼 0 + ∑ 𝑚𝑙=1 𝛼 𝑙 𝑎 𝑡−𝑙2 + ∑ 𝑠𝑗=1 𝛽 𝑗 𝜎 𝑡−𝑗2 + ∑ 𝑘𝑖=1 𝜌 𝑖 𝑋 𝑖 (21) And 𝑋 𝑖 is exogenous which contributes volatility to variance, and 𝜌 𝑖 is the related coefficient of term 𝑋 𝑖 . By combining ARMA model and GARCHX model, the joint model of ARMA-GARCHX is:
### 𝑌 𝑡 = 𝜙 0 + ∑ 𝑃𝑗=1 𝜙 𝑗 𝑌 𝑡−𝑗 + ∑ 𝑞𝑙=1 𝜃 𝑙 𝜀 𝑡−𝑙 + 𝑎 𝑡 𝑎 𝑡 = 𝜎 𝑡 𝜀 𝑡 {𝜎 𝑡2 = 𝛼 0 + ∑ 𝑚𝑙=1 𝛼 𝑙 𝑎 𝑡−𝑙2 + ∑ 𝑠𝑗=1 𝛽 𝑗 𝜎 𝑡−𝑗2 + ∑ 𝑘𝑖=1 𝜌 𝑖 𝑋 𝑖
### (22)
## **3. Empirical results **
### **3.1 VAR ** For VAR order selection, the paper anticipates 12 lagged orders for the model and makes lags length test via STATA, and the results are shown in Table 2. For various information criteria, the test marks illustrate that 0 lag is the best option for the model, which implies the malfunction of information criteria to select an appropriate lag for the model. For likelihood-ratio statistic, lag-4, lag- 7, lag-8, lag-9, lag-12 are all significant at 5% level, and the recommendation from varsoc confirms the lag-12 is the optimal selection for the model, hence, VAR (12) model will be built. Table 2 VAR model identification Lag LL LR df p FPE AIC HQIC SBIC 0 1584.38 1.0e- -25.49* - -25.399* 16* 25.4531* 1 1588.99 9.2108 16 0.904 1.2e-16 - -25.1214 -24.8513 25.3062 2 1597.73 17.495 16 0.354 1.4e-16 - -24.8566 -24.3705 25.1893 3 1605.79 16.106 16 0.446 1.5e-16 - -24.5806 -23.8784 25.0611 4 1622.02 32.462 16 0.009 1.5e-16 - -24.4365 -23.5182 25.0648 5 1629.62 15.208 16 0.509 1.8e-16 - -24.1533 -23.0189 24.9294 6 1637.49 15.745 16 0.471 2.0e-16 - -23.8744 -22.5239 24.7983 7 1651.21 27.429 16 0.037 2.1e-16 - -23.6897 -22.1231 24.7614 8 1664.77 27.121 16 0.040 2.3e-16 - -23.5025 -21.7199 24.7221 9 1679.76 29.973 16 0.018 2.4e-16 - -23.3383 -21.3396 24.7057 10 1686.52 13.531 16 0.634 2.8e-16 - -23.0416 -20.8268 24.5568 11 1698.68 24.31 16 0.083 3.1e-16 - -22.8317 -20.4008 24.4948 12 1712.43 27.513* 16 0.036 3.4e-16 - -22.6477 -20.0007 24.4586
237
-----
BCP Business & Management **AFTEM 2022**
Volume **32** (2022)
### The solutions of characteristic polynomial of VAR (12) model have been illustrated in Figure 2. The visual results verify the roots of matrix inside the unit circle and thus manifesting the stability of VAR system. Figure 2 VAR stability Setting return on gold, silver, and bitcoin as response variables and return on exchange rate as impulse variables, the results of IRF are shown in Figure 3. Obviously, with one unit change in exchange rate, the other variables all have visible oscillation to respond to the impulse, and maximum amplitude for bitcoin, gold, and silver are 0.6%, 0.1%, and 0.2% respectively. That shall be a reasonable result, whereby gold and silver are usually perceived as hedge assets and have resistance against shocks. On the contrary, as a speculative asset, bitcoin is more likely affected by a positive or negative shock, consistent with the research of Geuder et al. and Glaster et al. [8, 15]. The impact of impulse decays around 20 lags and vanishes gradually, fitting the basic feature of time-series clustering sequence. Figure 3 Impulse and response
238
-----
BCP Business & Management **AFTEM 2022**
Volume **32** (2022)
### **3.2 ARMA-GARCHX estimation results ** Before establishing ARMA-GARCHX model, it is essential to select the order for AR model and MA model. Calculating the autocorrelative function and partial auto-correlative function of returns on gold, silver, and bitcoin, the results are visualized in Figure 4, 5 lagged, 19 lagged and 35 lagged PACFs and 5 lagged ACF of bitcoin are significant at 5% level and reject to be white noisy; however, all the lagged PACFs and ACFs of gold are insignificant and cannot reject white noisy hypothesis; for silver, only 19 lagged PACF is significant, while all the ACFs of it are insignificant as well. Consequently, the possible ARMA models for bitcoin, silver, and gold are ARMA (5,5), AR (19), and ARMA (0,0), and more ARMA (0,0) is also a kind of white noisy. Noticeably, with empirical test, the AR (19) model is not effective to adapt the sequence of silver and the program fails to estimate specific coefficients of this model, so the sequence of return on silver also could be considered as a white noisy. For GARCH model, there is still no theoretical method to select the order of model appropriately, but according to the experience, GARCH (1,1) model is usually available and efficient, hence, anticipating the GARCH (1,1) model is reasonable to describe the volatility of the assets. Figure 4 PACF and ACF
239
-----
BCP Business & Management **AFTEM 2022**
Volume **32** (2022)
### The results of ARMA-GARCHX model are exhibited in Table 3. Since at least one coefficient of ARCH and GARCH of each item are significant at 10% level, the effectiveness of ARMA-GARCHX model is verified. Surprisingly, the related coefficients of exchange rate for gold, silver and bitcoin are 155.18, 206.34, and 60.92, all significant at 5% level. This finding has not just meaningful in statistics, but more economic, as it demonstrates the gold, silver and bitcoin are quite sensitive to the value of dollar, in the environment with aggressive monetary policies. Table 3 ARMA - GARCH estimation results, variance equation AG BTC Variables [GOLD ] Coef. Std. err Coef. Std. err Coef. Std. err Exchange 155.1860 [**] 78.8116 206.3454 [***] 55.4349 60.9215 [***] 20.8007 rate ARCH (- 0.1331 [*] 0.0803 0.1536 [***] 0.0517 0.0450 0.0560 1) GARCH 0.0047 0.2207 0.7392 [***] 0.1513 0.5707 [***] 0.1510 (-1) Constant - 9.6370 [***] 0.3473 - 9.5033 0.3680 - 5.8147 [***] 0.1786
## **4. Conclusion **
### By employing VAR model and ARMA-GARCHX model, the paper successfully verifies that gold, silver, and bitcoin will be volatile fiercely, corresponding the change of exchange rate from two channels. First, the result from VAR model and impulse response function suggests that the change of exchange rate has a constant long-term influence on expectations of return on silver and gold, and bitcoin; then, via ARMA-GARCHX model, the change of exchange rate will exacerbate the volatility of return on gold & silver and bitcoin. The coefficients of exchange rate for gold, silver, and bitcoin are 155.18, 206.34, and 60.92, all significant at 5% level. This fact might warn that the potential risk of the precious metal market and cryptocurrency market, which relate to conventional currency market, could raise magnificently when the monetary environment becomes complex and extreme. Exposed to the shock, investor shall apply more prudent investment strategies to avoid volatility and uncertainty.
## **References **
[1] Beckers S, Soenen L. Gold: More attractive to non‐US than to US investors? Journal of Business Finance
& Accounting, 1984, 11(1): 107-112.
[2] Sjaastad L A, Scacciavillani F. The price of gold and the exchange rate. Journal of International Money
and Finance, 1996, 15(6):879-897.
[3] Pukthuanthong K, Roll R. Gold and the Dollar (and the Euro, Pound, and Yen. Journal of Banking &
Finance, 2011, 35(8):2070-2083.
[4] Tully E, Lucey B M. A power GARCH examination of the gold market. Research in International
Business and Finance, 2007, 21.
[5] Hammoudeh S, Yuan Y. Metal volatility in presence of oil and interest rate shocks. Energy Economics,
2008, 30(2):606-620.
[6] Hashim S L, Ramlan H, Razali N H, et al. Macroeconomic variables affecting the volatility of gold price.
Journal of Global Business and Social Entrepreneurship (GBSE), 2017, 3(5): 97-106.
[7] Dyhrberg, Haubo A. Bitcoin, gold and the dollar – A GARCH volatility analysis. Finance Research
Letters, 2015, 16:85-92.
[8] Geuder J, Kinateder H, Wagner N F. Cryptocurrencies as financial bubbles: The case of Bitcoin. Finance
Research Letters, 2019, 31:179-184.
240
-----
BCP Business & Management **AFTEM 2022**
Volume **32** (2022)
[9] Baur D G, Dimpfl T, Kuck K. Bitcoin, gold and the US dollar – A replication and extension. Finance
Research Letters, 2018, 25.
[10] Corbet S, Meegan A, Larkin C, et al. Exploring the dynamic relationships between cryptocurrencies and
other financial assets. Economics Letters, 2017, 165:28-34.
[11] Ji H K. Tail behavior of Bitcoin, the dollar, gold and the stock market index. Journal of International
Financial Markets Institutions and Money, 2020, 67:101202.
[12] Fuller D W A. Distribution of the Estimators for Autoregressive Time Series with a Unit Root. Journal
of the American Statal Association, 1979, 79(366):355-367.
[13] Engle R E. Autoregressive Conditional Heteroskedasticity with Estimates of the Variance of United
Kingdom Inflation. Econometrica, 1982, 50(50).
[14] Bollerslev T. Generalized autoregressive conditional heteroskedasticity. Eeri Research Paper, 1986,
31(3):307-327.
[15] Glaser F, Zimmermann K, Haferkorn M, et al. Bitcoin-asset or currency? revealing users' hidden
intentions[J]. Revealing Users' Hidden Intentions (April 15, 2014). ECIS, 2014.
241
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.54691/bcpbm.v32i.2894?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.54691/bcpbm.v32i.2894, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "HYBRID",
"url": "https://bcpublication.org/index.php/BM/article/download/2894/2859"
}
| 2,022
|
[] | true
| 2022-11-22T00:00:00
|
[
{
"paperId": "9c1637b5adcd3e69731b48e0dcb329d9d4a83687",
"title": "Tail behavior of Bitcoin, the dollar, gold and the stock market index"
},
{
"paperId": "d23fafcfa0f448f063397a18e53af9e16b35c8ae",
"title": "Cryptocurrencies as financial bubbles: The case of Bitcoin"
},
{
"paperId": "3b539f3d38dab6a25aea93f918ea1d0d8dac6c05",
"title": "Exploring the Dynamic Relationships between Cryptocurrencies and Other Financial Assets"
},
{
"paperId": "6eb7bb60ded45918cc3e3821c58b277cb53b7009",
"title": "Bitcoin, gold and the US dollar – A replication and extension"
},
{
"paperId": "268f2253c50dd1a897d2d5a767a8c1f65aab80fd",
"title": "Bitcoin, gold and the dollar – A GARCH volatility analysis"
},
{
"paperId": "3c7d998b88bf48c88cf693625d2852706e7cb8e4",
"title": "Bitcoin - Asset or Currency? Revealing Users' Hidden Intentions"
},
{
"paperId": "779b6486afb3c315fe8675c6dd8c60889f110192",
"title": "Gold and the Dollar (and the Euro, Pound, and Yen)"
},
{
"paperId": "8009664766bce3ede9cf44ff5e7cca1c60ab3c2e",
"title": "Metal volatility in presence of oil and interest rate shocks"
},
{
"paperId": "f694fa64d5bd760b044cc664e0f377c9b435756a",
"title": "A power GARCH examination of the gold market"
},
{
"paperId": "86c50fb4f22e46a55c6aa329dd356b981b12745e",
"title": "The price of gold and the exchange rate"
},
{
"paperId": "584c7954ebb89d6155fa50e5bcf44098fb881faa",
"title": "Generalized autoregressive conditional heteroskedasticity"
},
{
"paperId": "f3beebc8ecbcbba3eeefa5bbd9c9191c92b7edfd",
"title": "GOLD: MORE ATTRACTIVE TO NON-U.S. THAN TO U.S. INVESTORS?"
},
{
"paperId": "2ee6cb87fc81ecd78d161c4a92c9dfce00c8961c",
"title": "Autoregressive conditional heteroscedasticity with estimates of the variance of United Kingdom inflation"
},
{
"paperId": "5cbbb5deb4d92dc0504fb7f2af0f6fe7da355d98",
"title": "Distribution of the Estimators for Autoregressive Time Series with a Unit Root"
},
{
"paperId": null,
"title": "Macroeconomic variables affecting the volatility of gold price"
}
] | 8,456
|
en
|
[
{
"category": "Economics",
"source": "external"
},
{
"category": "Business",
"source": "s2-fos-model"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Economics",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/023383352dccd15dc24113f894f33efd983da33c
|
[
"Economics"
] | 0.909747
|
Empirical Research on the Fama-French Three-Factor Model and a Sentiment-Related Four-Factor Model in the Chinese Blockchain Industry
|
023383352dccd15dc24113f894f33efd983da33c
|
Sustainability
|
[
{
"authorId": "134903717",
"name": "Ziyang Ji"
},
{
"authorId": "2113941384",
"name": "Victor I. Chang"
},
{
"authorId": "121319087",
"name": "H. Lan"
},
{
"authorId": "2126737803",
"name": "Ching-Hsien Robert Hsu"
},
{
"authorId": "144586410",
"name": "Raul Valverde"
}
] |
{
"alternate_issns": null,
"alternate_names": null,
"alternate_urls": [
"http://mdpi.com/journal/sustainability",
"http://nbn-resolving.de/urn/resolver.pl?urn=urn:nbn:ch:bel-172127"
],
"id": "8775599f-4f9a-45f0-900e-7f4de68e6843",
"issn": "2071-1050",
"name": "Sustainability",
"type": "journal",
"url": "http://www.e-helvetica.nb.admin.ch/directAccess?callnumber=bel-172127"
}
|
As one of the most significant components of financial technology (FinTech), blockchain technology arouses the interests of numerous investors in China, and the number of companies engaged in this field rises rapidly. The emotion of investors has an effect on stock returns, which is a hot topic in behavioral finance. Blockchain is an essential part of FinTech, and with the fast development of this technology, investors’ sentiment varies as well. The online information that directly reflects investors’ mood could be utilized for mining and quantifying to construct a sentiment index. For a better understanding of how well some factors adequately explain the return of stocks related to blockchain companies in the Chinese stock market, the Fama-French three-factor model (FFTFM) will be introduced in this paper. Furthermore, sentiment could be a new independent variable to enhance the explanatory power of the FFTFM. A comparison between those two models reveals that the sentiment factor could raise the explanatory power. The results also indicate that the Chinses blockchain industry does not own the size effect and book-to-market effect.
|
## sustainability
_Article_
# Empirical Research on the Fama-French Three-Factor Model and a Sentiment-Related Four-Factor Model in the Chinese Blockchain Industry
**Ziyang Ji** **[1], Victor Chang** **[2]** **, Hao Lan** **[1]** **[, Ching-Hsien Robert Hsu](https://sciprofiles.com/profile/587004)** **[3,4,5,]* and Raul Valverde** **[6]**
1 International Business School Suzhou, Xi’an Jiaotong-Liverpool University, Suzhou 215123, China;
ji.ziyang@outlook.com (Z.J.); Hao.Lan@xjtlu.edu.cn (H.L.)
2 School of Computing, Engineering and Digital Technologies, Teesside University, Middlesbrough TS1 3BX,
UK; V.Chang@tees.ac.uk
3 Department of Computer Science and Information Engineering, Asia University, Taichung 400-439, Taiwan
4 Department of Medical Research, China Medical University, Taichung 400-439, Taiwan
5 School of Mathematics and Big Data, Foshan University, Foshana 528000, China
6 John Molson School of Business, Concordia University, Montreal, QC G1X 3X4, Canada;
raul.valverde@concordia.ca
***** Correspondence: chh@cs.ccu.edu.tw
Received: 2 April 2020; Accepted: 11 June 2020; Published: 24 June 2020
[����������](https://www.mdpi.com/2071-1050/12/12/5170?type=check_update&version=2)
**�������**
**Abstract: As one of the most significant components of financial technology (FinTech), blockchain**
technology arouses the interests of numerous investors in China, and the number of companies
engaged in this field rises rapidly. The emotion of investors has an effect on stock returns, which is
a hot topic in behavioral finance. Blockchain is an essential part of FinTech, and with the fast
development of this technology, investors’ sentiment varies as well. The online information that
directly reflects investors’ mood could be utilized for mining and quantifying to construct a sentiment
index. For a better understanding of how well some factors adequately explain the return of stocks
related to blockchain companies in the Chinese stock market, the Fama-French three-factor model
(FFTFM) will be introduced in this paper. Furthermore, sentiment could be a new independent
variable to enhance the explanatory power of the FFTFM. A comparison between those two models
reveals that the sentiment factor could raise the explanatory power. The results also indicate that the
Chinses blockchain industry does not own the size effect and book-to-market effect.
**Keywords:** financial technology (FinTech); blockchain; Fama-French three-factor model;
sentiment index
**1. Introduction**
Financial technology (FinTech) consists of several technologies, such as blockchain,
cloud computing, big data, and machine learning. Blockchain is an advanced technology extracted
from the bitcoin, which was first promoted by Nakamoto [1]. As one of the most innovative and
important components of FinTech, it could now tackle challenges such as digital currency, asset
securitization, cross-border payment and settlement, and insurance management. As part of FinTech,
blockchain has produced a series of extremely promising applications because of its characteristics,
such as decentralization, immutability, and anonymity. Blockchain can not only play a role in FinTech,
but can also be applied to diverse industries, such as supply chain, intellectual property, estate, and the
Internet of Things (IoT). Blockchain technology is highly valued in China. Even the People’s Bank of
China (PBOC) began planning to issue CBDC (Central Bank Digital Currency) based on blockchain
technology, and the excellent design has been basically completed [2]. Mu et al. claim that the People’s
-----
_Sustainability 2020, 12, 5170_ 2 of 22
Bank of China owns the most blockchain patents among central banks in the world [3]. Due to the
innovative nature of this technology and the high level of interest, the number of companies in this field
is also increasing. It is necessary to detect the value of these firms to understand this industry better.
This paper will try to make use of the FFTFM (Fama-French Three-Factor Model) for the analysis
of stocks of Chinese blockchain firms and to detect the existence of size effect and book-to-market ratio
effect (BM effect) in this field. The capital asset pricing model is a popular topic attracting numerous
researchers for a long time. Markowitz first proposed portfolio theory to balance the risk and return [4].
Sharpe, Lintner, and Mossion built the capital asset pricing model in the 1960s, and this model considers
the market return as the unique variable to explain the return. Fama and French proposed the FFTFM,
that adding size factor and book-to-market ratio factor into the CAPM (Capital Asset Pricing Model)
enhances the explanatory power [5]. After the model was released, Chinese researchers began to
utilize the FFTFM to analyze Chinese stock market performance and found it gains a better result
than CAPM [5–8]. Some studies pay attention to the stock market, for instance, the stocks belonging
to the A-shares’ market and Growth Enterprise Market [9,10]. Some Chinese researchers focus on a
particular industry via the FFTFM, such as the real estate industry, the electric power industry, steel
industry, and bank industry [11–14]. It could be noticed that many studies emphasize the traditional
industry [12,14], whereas the blockchain industry is an innovation and the research of implementation
of FFTFM in this field is lacking.
Blockchain owns its noticeable position, especially when it comes to concepts like FinTech. There is
a saying that “one day in the blockchain industry, one year in real life”, which reveals the extremely
rapid changes in this field. Blockchain technology was first applied in the financial field. Since it
revolutionizes centralization and simplifies a series of the transaction process, it is recognized as a
particularly useful tool of FinTech, which arouses a lot of interest. The rapid development also has
an impact on investors’ expectations and sentiment to the blockchain companies including but not
limited to the firms that take blockchain as FinTech, and the relationship between emotion and stock
return is an indispensable topic of behavioral finance.
Behavioral finance researchers study the impact of capital market participants’ psychological
and behavioral characteristics on capital markets based on the assumptions of limited arbitrage
opportunities and bounded rationality. The “emotion” in psychology is the expression of external
attitudes generated by individual cognitive processes. Investor sentiment in behavioral finance is
caused by investors’ limited rationality and can be interpreted as investors’ expected bias, subjective
preferences, investment beliefs, and speculative needs. When investor sentiment affects enough
investment demand, it will cause the stock price to deviate from its value. According to empirical
studies, investor sentiment has an essential impact on financial behaviors such as stock price and
income fluctuations, stock market anomalies, and corporate investment decisions and earnings
management [15,16]. Liu and Zhang summarize that the Chinese stock market is mainly composed of
individual investors with relatively weak investment skills, keen subjective awareness, and low-risk
perception ability [17]. Investors are more inclined to pursue short-term capital gains and are keen on
short-term investment projects to gain speculative profits. This determines that investor sentiment
owns a more powerful influence in China than in mature capital markets.
Under this background, this paper tries to investigate the influence of Internet information related
to the stock performance by mining and quantifying Internet public opinion information. Then the
sentiment factor would be added into the traditional FFTFM for research.
The data are collected from the China Stock Market Accounting Research platform and China
Research Data Services platform. Sentiment factor results from the Guba public comments of each
stock by online users. Stocks that related to blockchain technology, including but not limited to the
listed firms that treat blockchain as FinTech, would be grouped to construct portfolios with different
characteristics for the research. While comparing the results of the FFTFM and improved four-factor
model, the sentiment factor could present a better explanation of the return of Chinese blockchain
-----
_Sustainability 2020, 12, 5170_ 3 of 22
stocks. It also notices that size effect and BM effect could not be found in this industry, and the portfolios
constructed by big-size companies and low book-to-market ratios companies gain the best return.
Due to the creativity and the bright prospect of blockchain technology, this paper focused on
Chinese listed firms related to blockchain technology and demonstrated the valuation via the FFTFM,
which is supposed to describe the risk yield characteristics of the blockchain industry in China.
Many scholars have studied China‘s asset pricing based on FFTFM. This paper also drew on the
FFTFM to empirically research the Chinese A-share market, to try to explain the stock performance
of the Chinese A-share blockchain firms and verify whether the industry has scale effects, value
premium effects, and profitability effects. We concluded that the BM effect does not exist in the Chinese
blockchain industry. We also built a sentiment factor using data mining to improve the traditional
FFTFM in order to present a better explanatory power for this industry.
**2. Literature Review**
_2.1. Fama-French Model Research_
Fama and French found that the beta value of CAPM could not explain the difference of excess
return, so they proposed a three-factor model that divides the main factors into three factors, namely
market factor, scale factor, and value factor, for a better explanatory power of excess return [4]. In order
to explore whether the model can be applied to the stock markets of other countries, Fama and French
studied the stock returns and pricing factors in different countries and claimed that FFTFM is better
than CAPM [18].
Carhart believed that the FFTFM could not explain the difference of excess returns well and added
a new variable-momentum factor to construct the Fama-French four-factor model [19]. Xu and Xiong
used A-share listed companies as samples from 2004 to 2005 and found that the four-factor explanatory
ability has been improved but cannot fully explain the stock fluctuations in yield [20]. Xue and Guan
conducted empirical research through a four-factor model and found that only a few funds can perform
slightly better than the whole market index [21]. This paper aimed to construct a model based on
FFTFM with a sentiment factor and to detect a more explanatory power, which has a similar purpose
as the Carhart model.
Fama and French analyzed profitability and investment information and added them to the
FFTFM to construct a new Fama-French five-factor model (FFFFM) [22]. However, the FFFFM is still
an imperfect model. Fama and French found that the FFFFM mainly has two defects: The first is that
the model lacks the ability to describe the average return of small stocks, and the second weakness is
that the HML (High Minus Low) factor is a redundancy factor. Racicot et al. studied the FFFFM with
traditional illiquidity measures and found the weakness of this model, especially for the endogenous
illiquidity measures [23]. The robust instrumental variables (RIV) algorithm conducted by GMM
(Generalized Method of Moment) was taken into consideration for correction. Racicot et al. transferred
FFFFM into the dynamic specification and used Kalman filtering and a recursive robust instrumental
variables (IV) algorithm to detect the estimation of alpha and beta [24]. They noticed that illiquidity is
a significant factor in the Kalman filter approach and that market risk premium is the only effective
factor in a dynamic context based on the GMM approach.
Sembiring applied the Fama-French model in the Indonesian securities market under market
overreaction conditions and found that the market, size, and value factor are accurate to explain
portfolios’ returns [25]. Cox and Britten utilized the FFFFM in the Johannesburg securities market
and concluded that size and value factors are significant, but the market factor presents a negative
relationship [26]. Bangash, Khan, and Jabeen disagreed that size pattern performs well by empirical
research on the Pakistan equity market [27]. In China, scholars not only use the FFTFM to investigate
the Chinese stock market but also on the replacement of indicators based on the combination of
the Chinese stock market’s actual situation. Tian, Wang, and Zhang compared the FFTFM between
the securities market of China and the United States [28]. They concluded that the FFTFM could
-----
_Sustainability 2020, 12, 5170_ 4 of 22
explain the excess returns of the two countries’ investment portfolios, whereas their applicability in
the Chinese and American stock markets is different. The Chinese market risks are more significant
than other factors, and SMB (Small minus Big) has explanatory power for small-cap stocks. Yang and
Fan indicated that the FFTFM is interpretable to the stock markets of developed and developing
countries [29]. Several researchers use the FFTFM application in the Chinese securities market to test
the effectiveness of the whole market [9,30,31]. Liu, Zhu, and Li argued that the FFTFM is suitable for
China’s Growth Enterprise Market [10]. Yin added the sentiment factor into the Fama-French model
and found that prices of small-cap stocks, high P/E (Price/Earnings) ratios, and high-priced stocks
are more sensitive to investor sentiment [32]. Hu, Tu, and Zhu believe that the five-factor model of
Fama-French stock pricing is suitable for use in China’s stock market [33]. Yuan and Cong found that
the FFTFM is suitable for the listed companies in the Chinese HA-DA-QI (Harbin-Daqing-Qiqihar)
region [34]. For the stock returns in Chinese, a particular industry, some researchers take the FFTFM to
explain the returns. You found the market factor effect and BM effect in the estate industry and Cheng
and Fang conducted an empirical research of stock returns in the auto industry [11,35]. For energy listed
companies, Li and Zhao pointed out that the FFTFM model applies to the prediction of market returns
of China’s listed power companies and that FFTFM could also be used in Chinese iron industry [12,13].
Gou, Wang, and Zhu drew the same conclusion that small-cap stocks own scale effects, and stocks
with high book-to-market value ratios have BM effect [14,36].
_2.2. Conventional Investor Sentiment Research_
Behavioral finance arouses numerous academics’ interest and they hope to find a principle for
better decision making by using investor sentiment. Baur, Quintero, and Stevens used stock market
data from 1986 to 1988 to explore the relationship between investor sentiment and linkage with
the 1987 securities’ market crash [37]. Mehra and Sah summarized the three conditions in which
the investor sentiment affects the stock price in the arbitrage market: Firstly, there is a systematic
fluctuation of investor sentiment; secondly, investors make decisions based on emotions; thirdly,
investor ignores the subjective influence brought by emotions [38]. Brown and Cliff collected data and
compiled investor sentiment index [39]. The study found that the lag effect of market yield has a more
significant impact on investor sentiment, but, in turn, investor sentiment is not efficient in predicting
market returns. Cheng and Liu used a blue-chip index to reflect the bullish situation and found that the
stock market’s mid-term sentiment was more affected than the short-term sentiment [40]. Wang, Zhao,
and Fang claimed that investor sentiment leads to share prices in the early stage of IPOs (Initial Public
Offering), which will cause listed companies to use investor sentiment to maximize profitability [41].
Wen et al. used the Shanghai Securities’ Market data to construct an investor sentiment index to study
the characteristics of investor behavior under different emotions [42].
_2.3. Investor Sentiment Research Based on Internet Information_
Currently, how to use social network information to predict economic behavior has gradually
become one of the research hotspots in various fields. Tetlock uses the one column of the Wall Street
Journal as the investment sentiment analysis basis, and analyzes the relationship between investor
sentiment and stock market returns [43]. He believes that the large fluctuations in investor sentiment
will cause an increase in volume, and pessimistic forecasts will lead to a fall in stock prices. Chen et al.
indicated that online information helps investors make better financial decisions [44]. Meng, Meng,
and Hu constructed the investor sentiment index using factor analysis, based on the data of CSSCI
(Chinese Social Sciences Citation Index), Sina Weibo text, and Baidu’s keyword recommendation
system [45]. Luo, Wang, and Fang considered investor sentiment index when establishing the CAPM,
and the investor sentiment index was constructed on the sentiment analysis of the stock forum
posting [46]. Investor sentiment against the stock index based on the ordinary linear regression model
was found. Xu uses text analysis and machine learning to construct a new investor sentiment indicator
system based on Sina stock evaluation information and a long-term survey [47].
-----
_Sustainability 2020, 12, 5170_ 5 of 22
**3. Data**
_3.1. Stock and Financial Data_
This paper selected listed blockchain companies in Shanghai and Shenzhen stock markets.
The Special Treatment (st) company cannot be considered as a regular listed company because of its
business difficulties, so the data only included non-st companies’ data. Because the data of listed
blockchain companies included in the Wind blockchain index were complete since 2016, and to test the
authenticity and comprehensiveness of the samples, monthly stocks’ data during June 2016 to June 2019
were collected from China stock market and accounting research database and Chinese research data
services platform (CNRDS). After screening, there were 50 sample companies. A large sample time
interval, with sufficient and new sample data, could provide a practical and meaningful result.
This paper selected the monthly information of stocks related to the Chinese blockchain industry
from June 2016 to June 2019, which was collected from the Chinese stock market and CNRDS.
These stocks included, but were not limited to, some listed companies that use blockchain as a FinTech.
The way to determine whether a company is involved in the blockchain industry is based on the
components of Wind’s blockchain industry index. If a firm was collected as a component of the index,
this firm was considered for the research. There was an overlap between the stocks belonging to the
blockchain industry index and the stocks belonging to the FinTech index.
The traditional Fama-French model will reclassify the portfolios at the end of June each year,
but to better reflect the performance, this paper regrouped the portfolios monthly. The reason is that
the blockchain industry in China is an emerging industry, and several companies are embracing this
innovative technology, including, but not limited to, FinTech firms. Other companies involved in
FinTech also show the same trend. In 2016, there were merely eight listed firms related to blockchain
technology according to Wind database, whereas there were more than 150 firms in 2019 [48]. More data
details will be shown in Sections 3.1.1, 3.1.2 and 3.1.4.
3.1.1. Stock Returns
The stock returns were selected by the monthly return that after cash dividend reinvestment.
Transaction costs was not considered. Stock returns are the basic element for constructing SMB and
HML factors [49].
3.1.2. Market Returns
Monthly market returns would be seen as a market index, which is comprehensively calculated
based on the returns of China A-shares market, B-shares market, and China’s growth enterprise
market. This is due to the complexity of the blockchain industry in China. These listed companies have
different characteristics and are distributed in different stock markets. Some of them are FinTech firms.
Therefore, this comprehensive index can more comprehensively and objectively reflect the overall
price change trend of the market and provide investors with more valuable indicators. This element
could reflect the situation of the whole market, which is also an essential data for calculating beta [49].
3.1.3. Risk-Free Rate
In practice, there is no absolute risk-free interest rate. Researchers choose those financial products
with better liquidity and less default risk to represent the risk-free interest rates, such as the national
debt rate and bank savings deposit rate. The Chinese banking system is dominated by state-owned
banks with little default risk, no market segmentation issues, or any individual or corporate institution
that can deposit in the bank [50]. This paper intended to select the one-year-term savings deposit
interest rate and convert it into a monthly return rate by using the continuous compounding method.
-----
_Sustainability 2020, 12, 5170_ 6 of 22
3.1.4. Size
The size of a listed company is determined by its market capitalization, obtained by multiplying
the stock price by the number of shares outstanding. According to China’s unique conditions, stocks
are divided into tradable shares and nontradable shares. Due to the special historical background of
the restructuring of state-owned enterprises, the existence of nontradable shares is a major feature of
China’s securities market and nontradable shares cannot be circulated on the secondary market [51].
So only tradable shares will be used and market capitalization calculated by the tradable shares as well.
3.1.5. Value
The measurement of the value of the company is the number of book equity to the market value
of equity. The financial statement of the listed company does not directly present this number, but it
could be gained from price-to-book value. This paper took the reciprocal of the PB (Price-to-Book)
value on the last trading day of each month.
3.1.6. Sentiment Data
There are various resources to construct investor sentiment factors. Zhang and Liu summarized
that the simple sentiment index mainly adopts a direct survey method or data mining method,
and compound indicators are constructed by selecting multiple single objective indicators, or a
combination of single objective indicators and subjective indicators to construct investor sentiment [52].
This paper used the data mining method, and Guba comments were used to present investors’
sentiment. Unlike news reports from newspapers or traditional news websites, Guba is a free medium,
and the content of their posts is mainly the expression of investors’ subjective wishes, which are
relatively random and irregular. For example, Guba comments may contain a few simple words,
or some irrelevant expressions and meaningless text expressions. These noises will affect the accuracy
of our sentiment judgment on posts. Comments on blockchain companies, including the firms, treat
this technology as FinTech and would be collected for further analysis.
There are several platforms for analyzing the text and extracting the emotional tendencies from the
content, for example, cloud natural language from Google, Baidu AI (Artificial Intelligence) platform,
and Yuyi data platform. In order to keep data consistent, the Guba media data analysis database
from the Chinese research data services platform was used in this paper. According to the Guba
database description, the platform uses a supervised learning model to judge the post’s sentiment.
The application of supervised learning in the post classification of the database includes the following
steps: �1 Define the categories (including positive, negative, and neutral) in advance, manually label
the content of the posts, and obtain positive, negative tendency. Score 1 is positive, −1 is negative,
and 0 is neutral. �2 Automatically obtain data from a dataset with category information. This part
of the data is called “training data”. �3 Supervised learning algorithm support vector machine is
introduced to learn the classification model on the training dataset. Use classification models to predict
the categories of a test dataset automatically. It is noticeable that the Guba comments data from the
Chinese research data services platform were merely from 2008 to 2018, so the 2019 comments data
were collected from the China stock market and accounting research database.
**4. Methodology**
_4.1. Build Portfolios According to Size and Value_
The stocks will be grouped monthly from two dimensions, size and value.
4.1.1. Size
There are many ways to determine in which group the company should be involved, and there are
various ways to deal with it. Fama and French divided the stocks from three American stock exchanges
-----
_Sustainability 2020, 12, 5170_ 7 of 22
into a small or big group based on the median of the size factor [5]. This paper used the FFTFM division
method. After sorting the circulating market capitalization in ascending order, the stocks were evenly
divided into two groups: Small (S) and big (B).
4.1.2. Value
Fama and French sorted the book-to-market ratio in ascending order and divided the sample data
into low (L) group, medium (M) group, and high (H) group, according to the proportions of 30%, 40%,
and 30%, which are called growth group, medium group, and value group [5]. Since the number of
blockchain companies change rapidly, including but not limited to the firms that take blockchain as
FinTech, so, based on the traditional Fama-French division method, the number of shares in a certain
group might be zero. In order to deal with this issue, the stocks were evenly classified into two groups,
high group (H) and low group (L). In the future, there will be more companies using blockchain
technology and more companies embracing FinTech. By then, this weak point can be compensated.
4.1.3. Portfolios
After the above two classifications, each stock had two indicators: Size and value. Those stocks
will be cross-combined to build four portfolios based on those two dimensions. They are S/L, S/H, B/L,
and B/H. The research on the Fama-French pricing model was based on the data of these four portfolios.
The details of these four portfolios are:
1. Portfolio S/L: Refers to those stocks which both belong to the small-size group and low
book-to-market ratio group at the same time.
2. Portfolio S/H: Refers to those stocks which both belong to the small-size group and high
book-to-market ratio group at the same time.
3. Portfolio B/L: Refers to those stocks which both belong to the big-size group and low
book-to-market ratio group at the same time.
4. Portfolio B/H: Refers to those stocks which both belong to the big-size group and high
book-to-market ratio group at the same time.
4.1.4. Construction of Independent Variables
Ri
Ri is the return of the portfolio that is calculated according to the ratio of the circulation market
value of each stock to the sum of the circulation market value of the combination.
Market Risk Premium Factor (Rm-Rf)
The market risk premium is obtained by subtracting the risk-free interest rate from the market
rate of return. As mentioned above, the market return (Rm) is a monthly return of A-shares, B-shares,
and China growth enterprise market. It is a comprehensive monthly return that is considered after
cash dividend reinvestment and obtained by using a weighted average market capitalization method.
The risk-free rate is the coupon rate of the one-year bank saving deposits and then turns an annual
risk-free interest rate into a monthly one.
SMB
SMB factor is obtained by comparing the average portfolio return of a small-sized company and
the average portfolio return of a big-sized company. This factor measures the difference in returns
due to the size of the listed companies. The construction method is to sort all companies according to
market capitalization from low to high, select the first 50% of the stocks to form a small market value
group, select the last 50% of the stocks to form a large market value group, and calculate the return of
small market value group and the return of large market value group, respectively. Then calculate
-----
_Sustainability 2020, 12, 5170_ 8 of 22
the difference between those two return rates. Repeat the above process every month to get the SMB
factor sequence.
The specific calculation formula is as follows:
_SMBt = [(][S][/][L][) + (][S][/][H][)]_ − [(][B][/][L][) + (][B][/][H][)]
2 2
HML
HML factor is obtained by comparing the monthly return rate of the high book-to-market ratio
portfolio and the monthly return rate of the low book-to-market ratio portfolio. This factor measures
the difference in returns due to different book-to-market ratios of listed companies. The construction
method sorts all companies according to the book-to-market ratio from high to low, selects the top 50%
of the stocks to form a group with a high book-to-market value ratio, and selects the bottom 50% of the
stocks to form a group with a low book-to-market value ratio. Then calculate the market-weighted
return of the two groups, respectively. Repeat the above process every month to get the HML
factor sequence.
The specific calculation formula is as follows:
_HMLt = [(][H][/][S][) + (][H][/][B][)]_ − [(][L][/][S][) + (][L][/][B][)]
2 2
Sentiment Factor
Antweiler and Frank introduced a method to measure the effect of investors’ sentiment [53].
Bu et al. proposed a measure of investor sentiment that integrates bullish and bearish expectations of
investors based on the Guba comments and naive Bayesian method [54]. This paper referred to this
method to construct the model for analyzing the negative or positive sentiment based on the Guba
comment website. The formulation is:
_M[pos]t_ − _M[neg]t_
_Sentt =_
_M[pos]t_ + M[neg]t
The M[c]t [=][ �]i∈D(t) _[w]i[x][c]i_ [means the sum of one emotion during the period D(t). The “c” belongs to]
positive, negative, or neutral. The x[c]
_i_ [means if a comment “i” is one of “c”, then][ x]i[c] [equals 1. If a comment]
“i” is not one of “c”, then x[c]
_i_ [equals 0. The “pos” represents positive emotions, “neg” represents negative]
emotions, and “neu” represents neutral emotions. The sentiment index “Sentt ” is between −1 to 1,
which indicates investor expectations. Every stock has a “Sent” value every month. According to
which portfolios they belong (S/L, S/H, B/L, B/H), the “sent” value of each portfolio will be obtained.
Fama-French Model
After collecting and processing the data above, the traditional Fama-French model could
be presented:
E(Ri) − Rf = bi[E(Rm) − Rf] + siE(SMB) + hiE(HML)
The regression equation of the model is expressed as follows:
Ri − Rft = αi + βi(Rmt − Rft) + siSMBt + hiHMLt + εit
According to the idea of the FFTFM, based on the model, the sentiment index reflecting investor
sentiment was constructed according to the above sentiment analysis method. Then, add the sentiment
factor to the model, and finally get a four-factors model:
Ri − Rft = αi + βi(Rmt − Rft) + siSMBt + hiHMLt + sentiSentimentit + εit
-----
_Sustainability 2020, 12, 5170_ 9 of 22
Ri − Rft is the excess return on portfolio “i”. It is the difference between the weighted return of
portfolio “I” and the risk-free rate during the same period ‘t’.
Rmt − Rft is the difference between the market return and risk-free rate during the same period “t”.
SMBt is the difference between the portfolio returns of the small companies and the portfolio
returns the big companies constructed during the period “t”.
_HMLt is the difference between the return of portfolios with high book-to-market ratio companies_
and the return of portfolios with low book-to-market ratio companies during the period “t”.
_Sentimentit is the sentiment score of the portfolio “i” during the period “t”._
**5. Analysis of the Model**
_5.1. Descriptive Analysis_
Based on the collected data and indicators constructed before, we could perform descriptive
statistical analysis. Related data processing was conducted in Python.
Table 1 shows the basic monthly returns of four different portfolios from 2016 to 2019, and it could
be noticed that portfolios with low book-to-market ratios had a positive average return. Additionally,
the investment portfolio with the lowest average rate of return was B/H, which was −0.01115.
Portfolio B/L owned the highest average rate of return, which was 0.01505. This portfolio had the
lowest standard deviation, indicating the smallest fluctuation of the performance. S/L portfolio owned
the highest standard deviation, which meant that the small-size company with low book-to-market
ratios carried the most top variations. Table 2 shows the correlation between parameters and portfolios’
returns. The correlation coefficient between SMB and market premium was 0.19, indicating that the
two were positively correlated, where that between HML and market premium was −0.1, showing the
negatively correlated.
**Table 1. Return of each portfolio’s descriptive analysis.**
**Test Focus** **R_SL** **R_SH** **R_BL** **R_BH** **Rm-Rf** **SMB** **HML**
count 37 37 37 37 37 37 37
mean 0.007166 −0.01992 0.01505 −0.01115 −0.12047 −0.00833 −0.02664
std 0.124866 0.091218 0.108523 0.081383 0.044723 0.049832 0.05708
min −0.19904 −0.14461 −0.15265 −0.12005 −0.20996 −0.11855 −0.1435
25% −0.06269 −0.07561 −0.07093 −0.08007 −0.14663 −0.03881 −0.06217
50% −0.02073 −0.04207 0.005983 −0.01998 −0.11772 −0.00401 −0.0235
75% 0.035004 0.017523 0.081635 0.0212 −0.09386 0.016986 0.013356
max 0.395968 0.278853 0.327864 0.245898 0.024608 0.101769 0.089053
**Table 2. Correlation among parameters and portfolios’ return.**
**Test Variables-** **rm-rf** **SMB** **HML** **R_SL** **R_SH** **R_BL** **R_BH**
rm-rf 1 0.19 −0.1 0.56 0.68 0.54 0.67
SMB 0.19 1 −0.082 0.54 0.34 −0.13 0.16
HML −0.1 −0.082 1 −0.59 −0.12 −0.59 −0.16
R_SL 0.56 0.54 −0.59 1 0.79 0.72 0.79
R_SH 0.68 0.34 −0.12 0.79 1 0.75 0.92
R_BL 0.54 −0.13 −0.59 0.72 0.75 1 0.78
R_BH 0.67 0.16 −0.16 0.79 0.92 0.78 1
-----
_Sustainability 2020, 12, x FOR PEER REVIEW_ 10 of 22
_Sustainability 2020, 12, 5170_ 10 of 22
_5.2. Market Risk Factor_
_5.2. Market Risk Factor_
Generally, the trend of market risk factors is the same as the trend of the average return of the
four portfolios. Among them, the market risk factor better reflects the changing direction of the S/H, Generally, the trend of market risk factors is the same as the trend of the average return of the
B/H portfolio. Changes in other combinations can also be presented, but there will be some four portfolios. Among them, the market risk factor better reflects the changing direction of the
discrepancies in specific periods. The portfolio S/L and B/L excess market factored a lot during the S/H, B/H portfolio. Changes in other combinations can also be presented, but there will be some
period from February 2018 to April 2018. This showed that the market risk factor was one of the discrepancies in specific periods. The portfolio S/L and B/L excess market factored a lot during the
essential variables to explain the difference in stock returns, but it was not enough to rely on the period from February 2018 to April 2018. This showed that the market risk factor was one of the
market risk factor alone to explain the changes. This also reflected the defects of the CAPM model essential variables to explain the difference in stock returns, but it was not enough to rely on the market
from the side, as shown in Figure 1. risk factor alone to explain the changes. This also reflected the defects of the CAPM model from the
side, as shown in Figure 1.
**Figure 1. Figure 1.Tendencies of return rate of four portfolios and market risk factor. Tendencies of return rate of four portfolios and market risk factor.**
_5.3. Size Factor_
_5.3. Size Factor_
Figures 2 and 3 show the comparison of portfolios with different companies’ size, given the same
Figures 2 and 3 show the comparison of portfolios with different companies’ size, given the same
book-to-market ratio. The orange line presents the portfolios constructed by the big companies, while
book-to-market ratio. The orange line presents the portfolios constructed by the big companies, while
the blue line displays the one built by the small companies. From June 2016 to February 2019, the trend
the blue line displays the one built by the small companies. From June 2016 to February 2019, the
of the monthly average return of a portfolio with large-size listed companies was consistent with the
trend of the monthly average return of a portfolio with large-size listed companies was consistent
direction of the one constructed by small-size firms. After February 2019, the average monthly yield
with the direction of the one constructed by small-size firms. After February 2019, the average
on stocks of small-size listed companies was higher than the portfolios of large-scale listed companies.
monthly yield on stocks of small-size listed companies was higher than the portfolios of large-scale
This may be because there were fewer companies engaged in the blockchain industry before 2019,
listed companies. This may be because there were fewer companies engaged in the blockchain
and the size of the company cannot be an essential factor affecting the portfolio yield. After 2019, there
industry before 2019, and the size of the company cannot be an essential factor affecting the portfolio
were more than 50 companies engaged in the blockchain industry, which more clearly reflected the
yield. After 2019, there were more than 50 companies engaged in the blockchain industry, which
difference in yields caused by size. This trend may be because blockchain is still a new technology,
more clearly reflected the difference in yields caused by size. This trend may be because blockchain
and the number of companies, including the firms that treat blockchain as FinTech, was relatively
is still a new technology, and the number of companies, including the firms that treat blockchain as
small in the early stage.
FinTech, was relatively small in the early stage.
-----
_Sustainability Sustainability2020 2020,,12 12, x FOR PEER REVIEW, 5170_ 11 of 22 11 of 22
**Figure 2. The return of portfolio S/L (Small-and-Low) and B/L (Big-and-Low).**
**Figure 2. Figure 2.The return of portfolio S/L (Small-and-Low) and B/L (Big-and-Low). The return of portfolio S/L (Small-and-Low) and B/L (Big-and-Low).**
**Figure 3. Figure 3.The return of portfolio S/H (Small-and-High) and B/H (Big-and-High). The return of portfolio S/H (Small-and-High) and B/H (Big-and-High).**
_5.4. Related TestFigure 3. The return of portfolio S/H (Small-and-High) and B/H (Big-and-High)._
ADF (Augmented Dickey–Fuller) Test
Generally, the first step is to perform a stationary test when studying on a time series data.
The Fama-French model is based on the return of stocks, which is a kind of time series dataset.
In addition to the method of visual inspection, the more commonly used statistical test method is the
augmented Dickey–Fuller (ADF) test, and it is an extended form of Dickey–Fuller test. The ADF test is
also known as the unit root test. If the significance test statistic obtained is less than three confidence
levels at 10%, 5%, or 1%, then there should be 90%, 95%, or 99% certainty to reject the null hypothesis
accordingly. Since the difference between the FFTFM and the improved four-factor model is adding
a new independent variable “sentiment”, the selected stocks remain, and the classification method
does not change as well. Therefore, it merely needs to test the stationary of returns of each portfolio.
The ADF test is suggested to be conducted in Python using the “statsmodels” package, and the results
are delivered in Table 3. As shown in Table 3, the return of eight portfolios all passed the ADF test.
The t-values of them were −5.4165, −5.0223, −4.3714, and −5.2175, respectively, and all p-values were
equal to zero. As the null hypothesis was rejected, there was no unit root in any time series data.
The stationary data could be taken into further research.
-----
_Sustainability 2020, 12, 5170_ 12 of 22
**Table 3. Results of the augmented Dickey–Fuller (ADF) test.**
**Portfolios** **1% Critical Value** **5% Critical Value** **10% Critical Value** **_t-Value_** **_p-Value_**
S/L (S/L including
sentiment) −3.6267 −2.9460 −2.6117 −5.4165 0.000
S/H (S/H including
sentiment) −3.6327 −2.9485 −2.6130 −5.0223 0.000
B/L (B/L including
sentiment) −3.6392 −2.9512 −2.6144 −4.3714 0.000
B/H (B/H including
sentiment) −3.6327 −2.9485 −2.6130 −5.2175 0.000
_5.5. Autocorrelation_
Autocorrelation refers to the correlation between the expected values of random error terms, and it
could harm the effectiveness of the multilinear regression model. So, whether using the traditional
FFTFM or the four-factor model, including the new sentiment parameter, it is necessary to test if
this situation exists. The first detection method was the standard Durbin–Watson (DW) test, and the
results of the test are presented in Table 4. The range of DW was from 0 to 4, and the value of DW
close to 0 indicated that the error terms were a positive autocorrelation while the value close to
4 indicated the negative autocorrelation. If the DW value ranged from dL (lower critical value of d) to
dU (upper critical value of d), it could not judge whether there was autocorrelation. Moreover, if the
DW value was between dU and 4-dU, it could bring greater confidence to conclude the non-existence
of autocorrelation. It is required to refer to a list of DW values to acquire the upper limit value
(dU) and lower limit value (dL) under different situations for checking the autocorrelation accurately.
According to Table 4, there was no autocorrelation in most portfolios. Still, the DW values of portfolio
S/L and B/H in the traditional FFTFM could not confirm whether they passed the test. The S/L portfolio
in the FFTFM was also unable to recognize if there was autocorrelation. Therefore, Breusch–Godfrey
LM (Lagrange multiplier) test was suggested to be considered, and the probabilities of chi2 (Chi-square)
were 0.0818 and 0.0847, respectively. The null hypothesis of the Breusch–Godfrey LM test was that
there was no autocorrelation. The results of the Breusch–Godfrey LM test also emerge in Table 4,
and it can be seen that the probability of all portfolios implied that the null hypothesis was acceptable.
There was no autocorrelation in any multilinear regression models.
**Table 4. Results of the autocorrelation test.**
**Test Variables** **Durbin-Watson** **Durbin-Watson**
**chi2** **Prob > chi2**
**and Focus** **Statistic** **Critical Value (Upper)**
S/L 2.356 1.655 1.342 0.2467
S/H 2.082 1.655 0.152 0.6970
B/L 2.082 1.655 0.152 0.6970
B/H 2.356 1.655 1.342 0.2467
S/L (add sentiment) 2.330 1.723 1.164 0.2807
S/H (add sentiment) 1.996 1.723 0.014 0.9057
B/L (add sentiment) 2.108 1.723 0.180 0.6712
B/H (add sentiment) 2.129 1.723 0.181 0.6703
_5.6. Multicollinearity_
Multicollinearity means that there is a linear correlation among the independent variables.
This situation manifests itself as one independent variable that can be a linear combination of one or
several other independent variables. It hurts the regression model. Perfect multicollinearity could
-----
_Sustainability 2020, 12, 5170_ 13 of 22
result in the non-existence of parameter estimation. Near-extreme multicollinearity allows the estimator
of the ordinary least square model to cease to be effective. Simultaneously, the parameter estimator
and the significance test would not make sense. It could not obtain a reliable prediction under the
multicollinearity. The value of variance inflation factor (VIF) could be used for multicollinearity test.
A greater VIF value means a higher probability of multicollinearity between independent variables.
The results of the multicollinearity test are shown in Table 5. All independent variables in every
regression model could pass the test, and there was no multicollinearity
**Table 5. Variance inflation factor (VIF) results.**
**Collinearity Statistics**
**Three-Factor Model**
**Tolerance** **VIF**
(Constant)
Rm-Rf 0.889 1.124
S/L, S/H, B/L, B/H
SMB 0.95 1.053
HML 0.941 1.063
Collinearity Statistics
Four-factor model
Tolerance VIF
S/L
S/H
B/L
B/H
(Constant)
VAR00002 0.889 1.124
SMB 0.95 1.053
HML 0.941 1.063
S_SL 0.882 1.133
(Constant)
VAR00002 0.951 1.052
SMB 0.946 1.058
HML 0.982 1.018
S_SH 0.973 1.027
(Constant)
VAR00002 0.957 1.045
SMB 0.902 1.109
HML 0.742 1.347
S_BL 0.727 1.375
(Constant)
VAR00002 0.956 1.046
SMB 0.933 1.072
HML 0.976 1.024
S_BH 0.96 1.041
_5.7. Heteroscedasticity_
All error terms have the same variance, which is an essential hypothesis of ordinary least squares
regression that guarantees a reliable result of parameter estimation. If the error terms own a different
variance, it could conclude that heteroscedasticity exists in the linear regression model. There are several
test methods for heteroscedasticity, such as the White test, Park test, Gleiser test, Goldfel—Quandt test,
and a directly subjective judgment is based on the graph. In this paper, the White test was taken into
consideration for heteroscedasticity, and the results are presented in Table 6. The null hypothesis of the
-----
_Sustainability 2020, 12, 5170_ 14 of 22
White test was there is homoscedasticity, and the alternative hypothesis was that there is unrestricted
heteroskedasticity. According to Table 6, the probabilities of all portfolios, be they a three- or four-factor
model, were higher than 0.05 overall. It implies that the null hypothesis was accepted and there was
no heteroscedasticity in any portfolios.
**Table 6. Results of the White test.**
**Test Variables and Focus** **chi2** **Prob > chi2**
S/L 9.70 0.3751
S/H 16.50 0.0572
B/L 16.50 0.0572
B/H 9.70 0.3751
S/L (add sentiment) 12.78 0.5438
S/H (add sentiment) 24.33 0.0518
B/L (add sentiment) 21.67 0.0856
B/H (add sentiment) 17.26 0.2426
After conducting the above tests, we could conclude that there was no autocorrelation,
multicollinearity, and heteroscedasticity. Further regression analysis was allowed to perform.
_5.8. Regression Analysis of the FFTFM_
5.8.1. Goodness of Fit of the FFTFM
The sample data obtained through actual observation used in empirical research were all authentic
reflections of facts. Therefore, after introducing the sample data into the model, it must be able to
describe this part of the objective facts well before the model can be considered meaningful. Therefore,
the model after data processing should be able to describe the fact better. The degree to which
the model approximates the sample is called the “goodness of fit.” In multiple regression analyses,
the determination coefficient R[2] is usually used to determine the goodness of fit of the equation. The R[2]
indicates what percent of the independent variable can explain the dependent variable. The value
of Rˆ2 is between 0 and 1. The closer R[2] is to 1, the better the model fits the sample data. If the R[2] is
close to 0, the model fits the fact badly. The regression could be conducted in Python and the results of
“goodness of fit” are shown below. Durbin–Watson statistics were also included in the table, indicating
no autocorrelation.
The FFTFM model performed relatively well in stocks of China’s blockchain industry, which could
provide a reference for other FinTech companies. The S/L group performed best in terms of explaining
portfolio returns, explaining 77.7% changes in the stock return. Portfolio B/H owned the worst result,
only 45.8%. These results illustrate that more factors could explain that the portfolio return needs to be
included in the mode, as shown in Table 7.
**Table 7. Goodness-of-fit test of traditional FFTFM (Fama-French Three-Factor Model).**
**S/L** **S/H** **B/L** **B/H**
_R[2]_ 0.770 0.509 0.653 0.458
_R2_ 0.749 0.464 0.621 0.409
Durbin-Watson statistic 2.356 2.082 2.082 2.356
-----
_Sustainability 2020, 12, 5170_ 15 of 22
5.8.2. Significance Test of the Model
The goodness of fit can only reflect the results of the FFTFM based on the selected data, but
cannot describe the overall relationship among the factors. Therefore, it was necessary to perform a
significance test on the model to test the degree of approximation of the trend in yields. The universal
test for testing the significance of the whole model is the F-test, as shown in Table 8.
**Table 8. Results of F test.**
**S/L** **S/H** **B/L** **B/H**
F test (3,33) 36.81 11.38 20.68 9.309
Prob (F) 0.0000 0.0000 0.0000 0.000132
Test hypothesis:
**Hypothesis 1 (H1). All coefficients of the regression model are zero, which means that it indicates that the**
_linear relationship of the FFTFM is not significant, and the model is meaningless._
**Hypothesis 2 (H2). At least one of the coefficients are not zero. This shows that the FFTFM has a significant**
_linear relationship, and this model has explanatory power to portfolio returns._
According to Table 8, at a given significance level of 1%, the F statistics of the four portfolios
were all greater than the critical value (F0.05(3, 33) = 2.89). Then the null hypothesis H1 was rejected,
which illustrates that at least one of the regression coefficients was significantly different from 0. It can
be concluded that the linear relationship of the FFTFM was significant. Besides, the probability value
corresponding to the F statistic of each portfolio was equal to 0, which also shows that the overall
linear relationship of the FFTFM was highly significant. In short, the FFTFM can better reflect the
overall characteristics of portfolio returns, which the companies constructed by the companies related
to blockchain technology. It may also provide a baseline for other companies that use blockchain
as FinTech.
5.8.3. Significance Test of Coefficients
In the previous section, the F-test of the FFTFM was performed in this paper. The results showed
that all four portfolios passed the F-test, which indicates that all factors in the FFTFM (Rm-Rf, SMB,
and HML) on stock returns were significant. However, this does not mean that each element in the
model (Rm-Rf, SMB, or HML) had a considerable effect on the yield alone. Therefore, it was necessary
to test the significance of each coefficient in the model. This paper used the t-test to analyze the impact
of a single factor on the stock return in the model.
As an explanatory variable shared by the capital asset pricing model and the FFTFM, testing the
coefficient b of the excess market rate of return (Rm-Rf) can analyze whether market risk factors have a
significant effect on stock returns.
According to Table 9, the coefficients of all portfolios were positive values, which indicates that
the market risk factor was positively correlated with the stock return. Besides, the coefficients of the
portfolios all passed the t-test with a significance level of 1%, and the t-values exceeded the significance
level of the coefficient that represents the scale factor and the coefficient that means book-to-market
ratio factor. This shows that market risk factors had a significant impact on stock returns. However,
it is different from the study of Fama and French [5]. They concluded that market risk factors have
only weak explanatory power.
-----
_Sustainability 2020, 12, 5170_ 16 of 22
**Table 9. Results of t-test.**
**S/L** **S/H** **B/L** **B/H**
βi 1.1939 1.2942 1.2942 1.1939
_t-test_ (−5.009) *** (5.086) *** (5.086) *** (5.009) ***
si 1.0496 0.3943 −0.6057 0.0496
_t-test_ (4.916) *** (1.730) * (−2.657) ** (0.232)
_hi_ −1.1229 −0.0612 −1.0612 −0.1229
_t-test_ (−6.102) *** (−0.312) (−5.402) *** (−0.668)
αi 0.0057 0.0136 0.0136 0.0057
_t-test_ (0.184) (0.409) (0.409) (0.184)
*** Indicates the coefficient passing the t-test at the 1% level. ** Indicates the coefficient passing the t-test at the
5% level. * Indicates the coefficient passing the t-test at the 10% level.
5.8.4. Book-to-Market Ratio Factor
The coefficient of the book-to-market value ratio factor (HML) was conducted to test whether
there was a significant relationship between the HML and portfolio returns. As shown in Table 9, only
the S/L and B/L passed the 1% significance test, and portfolios that were constructed by companies with
a high market-to-book ratio performed worse and did not pass the t-test. Fama and French concluded
that when the stock has a low book-to-market rate, which is called is a growth stock, the HML factor
in the model generally has a negative slope or a decreasing positive slope; when the stock has a
high book-to-market ratio, which is called value stock, the HML factor in the model generally has
an increasing slope [5]. However, the empirical research of the Chinese companies in the blockchain
industry, including which companies take blockchain as FinTech, did not follow this rule, and firms
with high market-to-book ratio could not pass the t-test of HML.
5.8.5. Size Factor
This part examines the linear relationship between the independent variable SMB and portfolio
returns. Analysis of the results in Table 9 showed that the regression coefficients of the company’s size
factor SMB on S/L, S/H, B/L, and B/H were 1.0496, 0.3943, −0.6057, and 0.0496, respectively. The p values
of the t-test were 4.916, 1.730, −2.657, and 0.232, respectively. According to the above results, the SMB
factor of the three portfolios, S/L, S/H, and B/H, had a positive correlation with the excess return of the
portfolio, while the explanatory variable SMB factor of the B/L portfolio had a negative relationship
with its performance. According to the t-test results of the SMB factor, the p values of the S/L and B/L
portfolios were less than 1%, and the p-value of the S/H was less than 5%. Therefore, the correlation
coefficient passed the test significantly, indicating that the SMB factor had a substantially higher
positive correlation for small-scale blockchain stock portfolios. The t-value of the coefficient of portfolio
B/H was less than the critical value, and the confidence level of the p-value was greater than 10%,
which means that the correlation between the scale factor and the return of B/H portfolio was not
significant. Since the samples were companies using blockchain, which are one of the components of
FinTech, it might present a reference for other FinTech firms.
5.8.6. An Improved Four-Factor Model Based on Fama-French Model
After collecting and processing the data about Guba comments on each blockchain company,
the score of the investors’ emotions could be obtained. People generally regard blockchain as an
innovative technology, especially when it comes to FinTech. The score of sentiment is a daily series,
and it should be turned into a monthly sequence. Moreover, the monthly score of the firms that are
in the same portfolio will be average weighted to get the grades of the portfolio’s monthly score.
Subsequently, the four portfolios, S/L, S/H, B/L, and B/H, own their sentiment index each month,
-----
_Sustainability 2020, 12, 5170_ 17 of 22
which could be added into the traditional Fama-French model to construct a new four-factor model.
The equation is shown below:
Ri − Rft = αi + βi(Rmt − Rft) + siSMBt + hiHMLt + sentiSentimentit + εit
5.8.7. Goodness of Fit and F Test
Table 10 illustrates the results of the F test and goodness of fit of the four-factor Fama-French
pricing model.
**Table 10. The results of goodness of fit and F test of four-factor pricing model.**
**Test Focus** **S/L** **S/H** **B/L** **B/H**
_R[2]_ 0.771 0.529 0.689 0.514
_R2_ 0.743 0.47 0.65 0.453
F test(3,32) 27 8.976 17.7 8.447
Prob(F) 0.0000 0.0000 0.0000 0.0000
The F test results for measuring the significance of the regression equation show that the p-value
of F test for the S/L, S/H, B/L, and B/H portfolios was equal to 0 when the number of samples was
150, and the number of explanatory variables was 3. All four portfolios passed the F test at a 99%
confidence level, indicating that the four sets of regression equations were highly significant. From this,
it was concluded that the independent variable market factor (Rm-Rf), the scale factor (SMB), the value
factor (HML), and the sentiment factor (Sentiment) all had significant effects on the dependent variable
portfolio yield.
_R[2]_ can be used to test how well the regression equation fits the sample observations. The closer R[2]
is to 1, the better the regression fit. The results in Table 10 show that the determination coefficient of the
S/L was 0.771, the determination coefficient of the S/H combination was 0.529, the B/L combination was
0.689, and the B/H was 0.514. From the above analysis, it was found that the portfolio with the best fit
was a small-scale and low book-to-market value one, and the worst fit was the portfolio constructed by
the big-scale and high book-to-market value companies. The fit of the four groups was not very good,
so there should be other factors in the market that had higher explanatory power besides three factors.
5.8.8. Parameters’ Analysis
The first is the regression coefficient, and the significance analysis of market returns to the returns
of each portfolio. S market portfolio returns are proxy variables for systemic risk; the beta reflects
the sensitivity of a single asset or portfolio to market changes. Table 11 shows that the market risk
premium coefficient β values of the four portfolios were 1.2245, 1.3174, 1.289, and 1.1852, respectively.
The β coefficients were all greater than 0, indicating that the return on portfolios kept the same moving
direction as the return of the whole market. The p-values of the t-test of the four stock portfolios
were all less than 1%, and the null hypothesis could be rejected at a 99% confidence level. Therefore,
it is believed that the linear relationship between the market factor (Rm-Rf) and the portfolio return
was significant.
Secondly, the linear relationship between independent variable SMB and portfolio return rate was
tested. The results in Table 11 reveal that the regression coefficients of the company size factor (SMB)
on S/L, S/H, B/L, and B/H were 1.06, 0.4275, −0.4984, and 0.1167, respectively. The t-values were 4.878,
1.871, −2.203, and 0.56, and the p-values for the t-test were 0, 0.070, 0.035, and 0.579. It illustrates that
the SMB factors of portfolio S/L, S/H, and B/H had a positive correlation with the excess return, and the
explanatory variable SMB factor of the B/L had a negative correlation with its return. According to the
_t-test result of the SMB factor, the p-value of the S/L portfolio was less than 1%, and the p-value of the_
S/H was less than 10%. Therefore, the correlation coefficient passed the test significantly, indicating
-----
_Sustainability 2020, 12, 5170_ 18 of 22
that the scale factor (SMB) affected small-scale blockchain stocks. The portfolios had a significantly
higher positive correlation.
**Table 11. Results of regression of four-factor pricing model.**
**Test Focus** **S/L** **S/H** **B/L** **B/H**
αi 0.006 0.0126 0.0075 −0.012
_t-test_ 0.189 0.383 0.235 −0.382
βi 1.2245 1.3174 1.289 1.1852
_t-test_ 4.894 *** 5.19 *** 5.267 *** 5.166 ***
si 1.06 0.4275 −0.4984 0.1167
_t-test_ 4.878 *** 1.871 * −2.203 ** 0.56
_hi_ −1.1042 −0.0487 −0.8535 −0.155
_t-test_ −5.793 *** −0.249 −3.922 *** −0.872
_senti_ 0.0351 0.1054 0.2469 0.1493
_t-test_ 0.461 1.172 1.922 * 1.906 *
*** Indicates the coefficient passing the t-test at the 1% level. ** Indicates the coefficient passing the t-test at the 5%
level. * Indicates the coefficient passing the t-test at the 10% level.
The SMB factor in the B/L portfolio also passed the t-test at the 1% confidence level. Still, the t value
of the B/H was less than the critical value, indicating that the correlation between the scale factor of the
B/H portfolio and the portfolio’s return was not significant.
Thirdly, the correlation coefficients that showed the BM effect in the four portfolios were −1.1042,
−0.0487, −0.8535, and −0.155, in turn. The t values were −5.793, −0.249, −3.922, and −0.872, respectively,
with the corresponding p values 0, 0.805, 0, and 0.39. The regression results showed that HML factors of
all portfolios had a negative correlation with their return, which is different from the study of Fama and
French [5]. According to the t-test results, the t-values of value factor (HML) in S/L and B/L portfolio
were greater than the critical value, and these HML factors in the two portfolios passed the t-test at the
1% confidence level. The t-value of HML in portfolio S/H and B/H showed that the HML factors were
not allowed to pass the test, which means that the HML factor in portfolios composed of listed firms
with high market-to-book ratio owned the very weak relationship with the return of the portfolios.
Finally, the fourth-factor, “sentiment”, was suggested to test. The results in Table 11 reflect
that the coefficients “Sentiment” in four portfolios were 0.0351, 0.1054, 0.2469, and 0.1493, in respect,
which implied that the sentiment factors in all portfolios had a positive relationship with the return.
In brief, the more exciting the investors, the higher the returns of stocks. However, when checking the
results from a more detailed perspective, it was not difficult to conclude that merely sentiment factors
in portfolios comprised of big-sized firms could pass the t-test at a 10% confidence level. The t-values
of sentiment factor of S/L and B/L portfolios were 0.0351 and −0.0487, respectively. So it was no
significant relationship between investors’ sentiment and the return of portfolios that were constructed
by small-size firms.
**6. Conclusions**
This paper is relevant to the topic of regression for FinTech, to demonstrate the effective valuation
theory conducted in companies embracing FinTech. It focused on the blockchain companies in China,
including those which treat blockchain as FinTech, and used the FFTFM to find these firms. It also
contained a new sentiment factor collecting from public comments for better explanatory power.
After the above descriptive analysis and regression analysis, it could lead to some conclusions.
-----
_Sustainability 2020, 12, 5170_ 19 of 22
_6.1. Feasibility of FFTFM and an Improved Four-Factor Model_
The FFTFM and the improved Fama-French model that added a new sentiment factor both
passed the F test. The market factor, size factor, and book-to-market ratio factor in FFTFM owned the
explanatory power to describe and review portfolios’ returns. In the improved model, the sentiment
factor could also explain the return of portfolios effectively.
It was noticeable that the explanatory power of four portfolios in FFTFM increased when adding
sentiment factor, rising from 0.77, 0.509, 0.653, and 0.458 to 0.771, 0.529, 0.689, 0.514, respectively.
It revealed that the sentiment factor had a positive effect on the model. All coefficients of sentiment
factors in different portfolios were positive, indicating more optimism brings a higher return.
However, there are still some minor flaws that caused the eight portfolios’ R-square to be not as
good as the expectation: The best goodness-of-fit value was 77.1%, and there could be more explanatory
variables to review the return of portfolios.
Additionally, to guarantee the reliability of the regression results, it was indispensable to test if
there were autocorrelation, multicollinearity, and heteroscedasticity, and we proved that all regressions
were acceptable.
Subsequently, although all portfolios passed the F test, the independent variables in each portfolio
should also be checked for the significance. There were two portfolios (S/L and B/L) under the FFTFM
and one portfolio (B/L) under the four-factor model’s own independent variables that all passed the
significance test.
_6.2. Influence of Market Risk Premium Factor_
The blockchain industry in China, including, but not limited to, the firms that take blockchain
as FinTech, owns a positive relationship with the whole market environment. The coefficients of the
independent variable market risk premium factor of all eight portfolios were more significant than
1, which implied that the investment portfolio could release the nonsystematic risk and contribute
to the return of portfolios. The blockchain industry is an emerging market in China, and several
update companies have begun to brace this technology recently, including the companies which use
blockchain as FinTech. Along with the development of the blockchain industry, numerous investors
are attracted by this industry and plan their investment in this as well. It causes higher volatility in
return than the whole market.
_6.3. The Non-Existence of Size Effect and Book-to-Market Ratio Effect in the Chinese Blockchain Industry_
The size effect is that the return of small listed firms have significantly higher average returns
than the large. Banz first found this effect and Fama and French verified the existence [5,55]. However,
several researchers own opposite views about the size effect. Goyal and Welch concluded that this effect
is caused by the deviation of sample selection rather than the size of the companies [56]. Dimon and
Marsh believe that big companies could achieve a higher return than small firms [57]. Schwert claimed
that the size effect is disappearing gradually. In this empirical research, the conclusion could be drawn
that there is no size effect in the Chinese blockchain industry, which includes the firms using this
technology as FinTech; portfolios with big companies bring a higher return [58].
The BM effect indicates that the return of the stocks has a positive relationship with the company’s
market-to-book ratio. A higher book-to-market rate could bring out a higher stock return. Fama and
French also believe the existence of the BM effect [5]. Chinese researchers drew a different conclusion
about the BM effect in the Chinese securities market. Xu argued that there is a significant BM effect
in the Chinese stock market [59]. Gu and Ding conducted an empirical study of the growth effect of
China’s securities’ market and proved that the BM effect is non-existent [60]. According to the analysis
of the Fama-French model above, the BM effect does not exist in the Chinese blockchain industry.
Portfolios built by the low book-to-market ratio companies earn more returns than others.
-----
_Sustainability 2020, 12, 5170_ 20 of 22
There are various factors that affect the investment in the stock market and investors have been
trying to obtain higher investment returns. Many scholars have also studied the effective factors
in this filed and have proved that FFTFM can be applied to Western developed securities’ markets.
Compared with these mature stock markets, the Chinese stock market has developed late. Therefore,
whether the Chinese market can effectively meet the conditions for using the three-factor model has
been under discussion. In recent years, with the continuous development of technologies, such as big
data, methods for measuring investor sentiment have also advanced. The online forum is an important
window for investors to express their sentiments. This article also took this factor into account to
improve the model’s explanatory power. The empirical analysis of this article showed that there is
no size effect in the Chinese blockchain industry, but there is a BM effect. Companies with more
positive and optimistic concerns can bring higher returns to investors. These can help investors choose
high-return companies in this field. We also need to admit that the method of sentiment analysis is still
relatively simple, and the accuracy of text sentiment measurement needs to be improved. The extent to
which the information in the online forum can affect investors’ decisions needs further research.
**Author Contributions: Conceptualization, Z.J. and V.C.; methodology, Z.J., V.C., and H.L.; software, Z.J.;**
validation, V.C., R.V., and C.-H.R.H.; formal analysis, Z.J. and V.C.; investigation, Z.J. and H.L.; resources, V.C.,
H.L., R.V., and C.-H.R.H.; data curation, Z.J.; writing—original draft preparation, Z.J.; writing—review and
editing, V.C., H.L., R.V., and C.-H.R.H.; visualization, Z.J.; supervision, V.C. and H.L.; project administration, V.C.;
funding acquisition, Z.J., V.C., R.V., and C.-H.R.H. All authors have read and agreed to the published version of
the manuscript.
**Funding: This work was partially supported by VC Research (VCR 0000042) and the National Natural Science**
Foundation of China (Grant No. 61872084).
**Conflicts of Interest: The authors declare no conflict of interest.**
**References**
1. [Nakamoto, S. Bitcoin: A Peer-To-Peer Electronic Cash System. Available online: https://bitcoin.org/bitcoin.pdf](https://bitcoin.org/bitcoin.pdf)
(accessed on 1 September 2019).
2. Mu, J. Opportunities, Challenges and Prospects of the Central Bank to Implement the Legal Digital Currency
DCEP. Economist 2020, 31, 95–105.
3. Mu, C.C.; Di, G.; Lv, Y.; Qian, Y.C.; Qing, S.D. Blockchain Research Group of Digital Currency Research
Institute of PBC. China Financ. 2020, 70, 28–29.
4. Markowitz, H.M. Portfolio selection. J. Financ. 1952, 7, 77–91.
5. [Fama, E.F.; French, K.R. The Cross-Section of Expected Stock Returns. J. Financ. 1992, 47, 427–465. [CrossRef]](http://dx.doi.org/10.1111/j.1540-6261.1992.tb04398.x)
6. Fan, L.Z.; Yu, S.D. A Three-factor Model of Chinese Stock Market. J. Syst. Eng. 2002, 17, 537–546.
7. Yang, X.; Chen, Z.H. Empirical Study on Three-factor Asset Pricing Model in Chinese Stock Market.
_Quant. Tech. Econ. 2003, 20, 137–141._
8. Deng, C.R.; Ma, Y.K. Empirical Study on Three-factor Model in China’s Securities Market. Chin. J. Manag.
**2005, 2, 591–596.**
9. Zhao, P.; Zhou, M. Research on the Applicability of Fama-French Three-Factor Model of Securities Industry
in China’s A-share Market. Times Financ. 2018, 27, 156–157.
10. Liu, T.J.; Zhu, J.Q.; Li, Y. An Empirical Analysis of the Applicability of Fama-French Three-factor Model in
China’s Growth Enterprise Market. J. Jiaozuo Univ. 2013, 27, 64–66.
11. You, D. An Empirical Study of Fama-French Three-Factor Model in China’s A-Share Real Estate Sector.
_Mod. Bus. Trade Ind. 2008, 20, 199–200._
12. Li, Z.H.; Zhao, A.C. Research on the Return Rate Prediction of Electric Power Industry in Chinese Stock
Market. J. Hebei Univ. Econ. Bus. (Compr. Ed.) 2018, 18, 34–39.
13. Guo, Z.X. Fama-French three-factor model and five-factor model empirical test for A-shares steel companies.
_Hebei Enterp. 2019, 30, 33–36._
14. Zhu, M. Empirical Test of Listed Bank Shares in China Based on Fama-French Three-Factor Model. Hebei Financ.
**2019, 26, 17–21.**
15. Liu, W.Q.; Liu, X.X. Individual/institutional investor sentiment and stock returns: Study based on Shanghai
A-share market. J. Manag. Sci. China 2014, 17, 70–87.
-----
_Sustainability 2020, 12, 5170_ 21 of 22
16. Ba, S.S.; Zhu, H. Margin Trading, Short Selling, Investor Sentiment and Stock Market Volatility.
_Stud. Int. Financ. 2016, 32, 82–96._
17. Liu, Y.C.F.; Zhang, Z.C. A Summary of Research on Investor Sentiment in Chinese Stock Market. China Econ.
**2018, 32, 78–79.**
18. Fama, E.F.; French, K.R. Size, Value and Momentum in International Stock Returns. J. Financ. Econ. 2012, 10,
[457–472. [CrossRef]](http://dx.doi.org/10.1016/j.jfineco.2012.05.011)
19. [Carhart, M.M. On persistence in mutual fund performance. J. Financ. 1997, 52, 57–82. [CrossRef]](http://dx.doi.org/10.1111/j.1540-6261.1997.tb03808.x)
20. Xu, H.Y.; Xiong, C. Applicability of Carhart’s four-factor model in China’s stock market. Bus. Cult. 2009,
_2, 128._
21. Xue, S.Z.; Guan, X.Y. An Empirical Study on the Evaluation of Openended Fund Performance Based on
“Four-Factors” Model. Financ. Teach. Res. 2011, 26, 63–65.
22. [Fama, E.F.; French, K.R. A five-factor asset pricing model. J. Financ. Econ. 2015, 116, 1–22. [CrossRef]](http://dx.doi.org/10.1016/j.jfineco.2014.10.010)
23. Racicot, F.E.; Rentz, W.F.; Tessier, D.; Théoret, R. The Conditional Fama-French Model and Endogenous
[Illiquidity: A Robust Instrumental Variables Test. PLoS ONE 2019, 14, 1–26. [CrossRef]](http://dx.doi.org/10.1371/journal.pone.0221599)
24. Racicot, F.E.; Rentz, W.F.; Kahl, A.L.; Mesly, O. Examining the dynamics of illiquidity risks within the phases
[of the business cycle. Borsa Istanb. Rev. 2019, 19, 117–131. [CrossRef]](http://dx.doi.org/10.1016/j.bir.2018.12.001)
25. Sembiring, F.M. Three-Factor and Five-Factor Models: Implementation of Fama and French Model on Market
Overreaction Conditions. J. Financ. Bank. Rev. (JFBR) 2018, 3, 77–83.
26. Cox, S.; Britten, J. The Fama-French five-factor model: Evidence from the Johannesburg Stock Exchange.
_[Invest. Anal. J. 2019, 48, 240–261. [CrossRef]](http://dx.doi.org/10.1080/10293523.2019.1647982)_
27. Bangash, R.; Khan, F.; Jabeen, Z. Size, Value and Momentum in Pakistan Equity Market: Size and Liquidity
[Exposures. Glob. Soc. Sci. Rev. 2018, 3, 376–394. [CrossRef]](http://dx.doi.org/10.31703/gssr.2018(III-I).22)
28. Tian, L.H.; Wang, G.Y.; Zhang, W. Three-factor model pricing: How is China different from the United States?
_Stud. Int. Financ. 2014, 30, 37–45._
29. Yang, J.H.; Fan, L.B. Global Market Integration: Fama-French Three-Factor Model Test in a Global Scenario.
_Technol. Econ. 2017, 36, 109–119._
30. Feng, S.J.; Liu, Z. An Empirical Study of Three-factor Pricing Model in China Stock Market. Econ. Res. Guide
**2018, 20, 145–148.**
31. Yan, X.; Pu, T. An Empirical Study of Fama-French Model in China’s Stock Market: A Case Study of Shanghai
50 Index. Times Financ. 2017, 11, 181–182.
32. Yin, L.Y. An Empirical Study of the Impact of Investor Sentiment on Stock Returns: Based on the Fama-French
Three-Factor Model. Friends Account. 2018, 6, 51–56.
33. Hu, M.A.; Tu, X.P.; Zhu, D.Y. Is the Fama-French five-factor model more explanatory? Co-Oper. Econ. Sci.
**2017, 10, 82–85.**
34. Yuan, J.Y.; Cong, Z. Research on the Stock Return of Hadaqi Listed Companies Based on Fama-French
Three-Factor Model. J. Heilongjiang Inst. Technol. 2015, 29, 51–53.
35. Cheng, S.Y.; Fang, H. Empirical Research on China’s Auto Sector Stock Yield Based on Fama-French
Three-Factor Model. China Price 2019, 1, 71–73.
36. Gou, D.N.; Wang, W.J. Empirical Test of China’s Listed Bank Shares Based on Fama-French Three-Factor
Model. Stat. Decis. 2016, 21, 158–161.
37. Baur, M.; Quintero, S.; Stevens, E. The 1986–88 stock market: Investor sentiment or fundamentals?
_[Manag. Decis. Econ. 1995, 17, 319–329. [CrossRef]](http://dx.doi.org/10.1002/(SICI)1099-1468(199605)17:3<319::AID-MDE776>3.0.CO;2-0)_
38. Mehra, R.; Sah, R. Mood fluctuations, projection bias, and volatility of equity prices. J. Econ. Dyn. Control
**[2002, 26, 869–887. [CrossRef]](http://dx.doi.org/10.1016/S0165-1889(01)00035-5)**
39. [Brown, G.; Cliff, M. Investor Sentiment and Asset Valuation. SSRN Electron. J. 2001. [CrossRef]](http://dx.doi.org/10.2139/ssrn.292139)
40. Cheng, K.; Liu, R. Research on the Interaction between Investor Sentiment and Stock Market.
_Shanghai Econ. Rev. 2005, 11, 88–95._
41. Wang, C.F.; Zhao, W.; Fang, Z.M. Measure of IPO investor sentiment and its relationship with IPO price
behavior. Syst. Eng. 2007, 7, 1–6.
42. Wen, F.H.; Yang, X.; Gong, X.; Huang, C.X.; Yang, X.G. An Infectious Analysis of Sino-US Investor Sentiment
in the Context of the Financial Crisis. Syst. Eng. Theory Pract. 2015, 35, 623–629.
43. Tetlock, P. Giving Content to Investor Sentiment: The Role of Media in the Stock Market. J. Financ. 2007, 62,
[1139–1168. [CrossRef]](http://dx.doi.org/10.1111/j.1540-6261.2007.01232.x)
-----
_Sustainability 2020, 12, 5170_ 22 of 22
44. Chen, X.H.; Peng, W.L.; Tian, M.Y. Research on stock price and trading volume prediction based on investor
sentiment. J. Syst. Sci. Math. Sci. 2016, 36, 2294–2306.
45. Meng, X.J.; Meng, X.L.; Hu, Y. Research on Investor Sentiment Index Based on Text Mining and Baidu Index.
_Macroeconomics 2016, 1, 144–153._
46. Luo, K.; Wang, C.F.; Fang, Z.M. Research on the Regionalization of Capital Asset Pricing under the Influence
of Investors’ Emotion—Based on the Emotional Analysis of Stock Forum Posts. Oper. Res. Manag. Sci. 2017,
_26, 129–136._
47. Xu, T.Y. Research on the Impact of Investor Sentiment on the Stock Market in Social Media.
_Shanghai Manag. Sci. 2018, 40, 67–74._
48. The number of blockchain companies has increased by 7159%. China Econ. Wkly. 2019, 11, 7.
49. Fama, E.F.; French, K.R. Size and book-to-market factors in earnings and returns. J. Financ. 1995, 50, 131–155.
[[CrossRef]](http://dx.doi.org/10.1111/j.1540-6261.1995.tb05169.x)
50. Wu, X.F.; Liu, Y.F.; Lei, M. An Empirical Analysis of Equity Premium in China’s Capital Market. Econ. Probl.
**2007, 336, 85–88.**
51. Qu, Z.X. Research on Value Evaluation Method of Non-tradable Shares of Listed Companies. Business 2012,
_9, 48._
52. Zhang, Z.C.; Liu, Y.C.F. The Construction and Effectiveness Test of Investor Sentiment Index in my country’s
Stock Market. Commer. Times 2018, 7, 156–158.
53. Antweiler, W.; Frank, M.Z. Is all that talk just noise? the information content of internet stock message
[boards. J. Financ. 2004, 59, 1259–1294. [CrossRef]](http://dx.doi.org/10.1111/j.1540-6261.2004.00662.x)
54. Bu, H.; Xie, Z.; Li, J.H.; Wu, J.J. The Impact of Investor Sentiment Based on Stock Evaluation on the Stock
Market. J. Manag. Sci. China 2018, 21, 86–101.
55. Banz, R.W. The Relationship between Return and Market Value of Common Stocks. J. Financ. Econ. 1981, 9,
[3–18. [CrossRef]](http://dx.doi.org/10.1016/0304-405X(81)90018-0)
56. Goyal, A.; Welch, I. Predicting the Equity Premium; manuscript; Yale University: New Haven, CT, USA, 1999.
57. [Dimson, E.; Marsh, P. Murphy’s Law and Market Anomalies. J. Portf. Manag. 1998, 25, 53–69. [CrossRef]](http://dx.doi.org/10.3905/jpm.1999.319734)
58. Schwert, G.W. Anomalies and market efficiency. In Handbook of the Economics of Finance; Elsevier: Amsterdam,
The Netherlands, 2003; pp. 939–974.
59. Xu, Z.H. An Empirical Analysis of China’s Stock Market Scale Effect and Book-to-Market Ratio Effect.
_J. Financ. Dev. Res. 2011, 11, 75–78._
60. Gu, J.; Ding, Y. An Empirical Study on the Growth Effect of China’s Securities Market. Econ. Rev. 2003, 2,
101–105.
© 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
[(CC BY) license (http://creativecommons.org/licenses/by/4.0/).](http://creativecommons.org/licenses/by/4.0/.)
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3390/su12125170?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/su12125170, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "GOLD",
"url": "https://www.mdpi.com/2071-1050/12/12/5170/pdf?version=1593505965"
}
| 2,020
|
[] | true
| 2020-06-24T00:00:00
|
[
{
"paperId": "67f6a970ea6e3c4ee036d62c44131455048601e9",
"title": "The conditional Fama-French model and endogenous illiquidity: A robust instrumental variables test"
},
{
"paperId": "2a932183ee916d7344dc0f89269648e29cb3760b",
"title": "The Fama-French five-factor model: Evidence from the Johannesburg Stock Exchange"
},
{
"paperId": "af456311cae5d703c19bc06c53c3bb6ead2baa2a",
"title": "Examining the dynamics of illiquidity risks within the phases of the business cycle"
},
{
"paperId": "2850a882b61e903d15f32bfcc0d010572cad49fa",
"title": "Three-Factor and Five-Factor Models: Implementation of Fama and French Model on Market Overreaction Conditions"
},
{
"paperId": "b37a7765c4526cac1983952aecf372828e5cac61",
"title": "A Five-Factor Asset Pricing Model"
},
{
"paperId": "be388fb9059c223c97be2bea43d467d598ef4bf0",
"title": "Size, Value, and Momentum in International Stock Returns"
},
{
"paperId": "1198070456cb65844c3766e78dce0b6d2ff9b31f",
"title": "The Role of Investability Restrictions on Size, Value, and Momentum in International Stock Returns"
},
{
"paperId": "c596b215d73617d15528d39a4046560f9dd0c30c",
"title": "NASA technology assessment using real options valuation"
},
{
"paperId": "1fe8a5e84939833f4b9f8c9bce311c70fb0192a9",
"title": "Mood fluctuations, projection bias, and volatility of equity prices"
},
{
"paperId": "52bef0f2fbaead322e7788e9686e837986a9b1ba",
"title": "Investor Sentiment and Asset Valuation"
},
{
"paperId": "747bcf00e45a51a1001fa6fc796502c9aab5974e",
"title": "Is All that Talk Just Noise? The Information Content of Internet Stock Message Boards"
},
{
"paperId": "e20a66dae3fe8496eb76cb94d3f1d48bf85fe804",
"title": "Asset Pricing at the Millennium"
},
{
"paperId": "c6ea69c387da7257429d3411b7ce2c7cb7cb7f67",
"title": "Murphy's Law and Market Anomalies"
},
{
"paperId": "db0a181095577ef68e8e450d69e618c3ae2d25cb",
"title": "The 1986-88 Stock Market: Investor Sentiment or Fundamentals?"
},
{
"paperId": "0065386cafc3ed92ace622e1cd8f9f51583cff33",
"title": "Size and Book-to-Market Factors in Earnings and Returns"
},
{
"paperId": "f299a3b308717f6a17b15ee9509ddd70c3df69dd",
"title": "The Cross‐Section of Expected Stock Returns"
},
{
"paperId": "0c27155b0d5c100db4e86e6dbad1af05b93b6c2f",
"title": "The relationship between return and market value of common stocks"
},
{
"paperId": "433561f47f9416a6500c8350414fdd504acd2e5e",
"title": "Bitcoin Proof of Stake: A Peer-to-Peer Electronic Cash System"
},
{
"paperId": null,
"title": "Blockchain Research Group of Digital Currency Research Institute of PBC"
},
{
"paperId": null,
"title": "Opportunities, Challenges and Prospects of the Central Bank to Implement the Legal Digital Currency DCEP"
},
{
"paperId": "39a0f436cbd7b98830b80b58f7e7b115673ef295",
"title": "Portfolio Selection"
},
{
"paperId": null,
"title": "The number of blockchain companies has increased by 7159%"
},
{
"paperId": null,
"title": "Fama-French three-factor model and five-factor model empirical test for A-shares steel companies"
},
{
"paperId": null,
"title": "Empirical Test of Listed Bank Shares in China Based on Fama-French Three-Factor Model. Hebei Financ"
},
{
"paperId": "83d542f739abbc8aa8f57d3ecc1065ad8e779791",
"title": "Size, Value and Momentum in Pakistan Equity Markets: Size and Liquidity Exposures"
},
{
"paperId": null,
"title": "The Construction and Effectiveness Test of Investor Sentiment Index in my country’s Stock Market"
},
{
"paperId": null,
"title": "An Empirical Study of the Impact of Investor Sentiment on Stock Returns: Based on the Fama-French Three-Factor Model"
},
{
"paperId": null,
"title": "An Empirical Study of Three-factor Pricing Model in China Stock Market"
},
{
"paperId": null,
"title": "Research on the Return Rate Prediction of Electric Power Industry in Chinese Stock Market"
},
{
"paperId": null,
"title": "Research on the Impact of Investor Sentiment on the Stock Market in Social Media"
},
{
"paperId": null,
"title": "Research on the Applicability of Fama-French Three-Factor Model of Securities Industry Sustainability 2020"
},
{
"paperId": null,
"title": "Global Market Integration: Fama-French Three-Factor Model Test in a Global Scenario"
},
{
"paperId": null,
"title": "An Empirical Study of Fama-French Model in China’s Stock Market: A Case Study of Shanghai 50 Index"
},
{
"paperId": null,
"title": "Research on the Regionalization of Capital Asset Pricing under the Influence of Investors’ Emotion—Based on the Emotional Analysis of Stock Forum Posts"
},
{
"paperId": null,
"title": "Empirical Test of China’s Listed Bank Shares"
},
{
"paperId": null,
"title": "Research on stock price and trading volume prediction based on investor sentiment"
},
{
"paperId": null,
"title": "Research on Investor Sentiment Index Based on Text Mining and Baidu Index"
},
{
"paperId": null,
"title": "Margin Trading, Short Selling, Investor Sentiment and Stock Market Volatility"
},
{
"paperId": null,
"title": "An Infectious Analysis of Sino-US Investor Sentiment in the Context of the Financial Crisis"
},
{
"paperId": "00e68210d0125365b8c01e22870b5e3ef76bdd72",
"title": "Individual /institutional investor sentiment and stock returns: Study based on Shanghai A-share market"
},
{
"paperId": null,
"title": "Three-factor model pricing: How is China different from the United States?"
},
{
"paperId": "f5fb24b271bbdad21a1c54a17ce27c69bbd19407",
"title": "Handbook of the Economics of Finance"
},
{
"paperId": null,
"title": "Research on Value Evaluation Method of Non-tradable Shares of Listed Companies"
},
{
"paperId": null,
"title": "An Empirical Analysis of China’s Stock Market Scale Effect and Book-to-Market Ratio Effect"
},
{
"paperId": null,
"title": "An Empirical Study on the Evaluation of Openended Fund Performance Based on “Four-Factors” Model"
},
{
"paperId": "4e9ec92a90c5d571d2f1d496f8df01f0a8f38596",
"title": "Bitcoin: A Peer-to-Peer Electronic Cash System"
},
{
"paperId": null,
"title": "Measure of IPO investor sentiment and its relationship with IPO price behavior"
},
{
"paperId": "ae074faa8712719d5dce6b98b02750abc118c823",
"title": "Giving Content to Investor Sentiment: The Role of Media in the Stock Market"
},
{
"paperId": "69b9a65c6589a3e2827bb0c885132b169f7808ba",
"title": "Chapter 15 Anomalies and market efficiency"
},
{
"paperId": null,
"title": "Predicting the Equity Premium; manuscript; Yale"
},
{
"paperId": null,
"title": "This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license"
}
] | 21,047
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Engineering",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/023de6262322bae283469df6b9c1c056759315c4
|
[
"Computer Science"
] | 0.872681
|
An Improved Genetic Algorithm for Safety and Availability Checking in Cyber-Physical Systems
|
023de6262322bae283469df6b9c1c056759315c4
|
IEEE Access
|
[
{
"authorId": "11032852",
"name": "Z. Wang"
},
{
"authorId": "2122754956",
"name": "Yanan Jin"
},
{
"authorId": "150311209",
"name": "Shasha Yang"
},
{
"authorId": "1864210",
"name": "Jianmin Han"
},
{
"authorId": "49301778",
"name": "Jianfeng Lu"
}
] |
{
"alternate_issns": null,
"alternate_names": null,
"alternate_urls": [
"http://ieeexplore.ieee.org/servlet/opac?punumber=6287639"
],
"id": "2633f5b2-c15c-49fe-80f5-07523e770c26",
"issn": "2169-3536",
"name": "IEEE Access",
"type": "journal",
"url": "http://www.ieee.org/publications_standards/publications/ieee_access.html"
}
|
Cross-IoT infrastructure access frequently occurs when performing tasks in a distributed computing infrastructure of a cyber-physical system (CPS). The access control technology that ensure secure access cross-IoT infrastructure usually automatically establish relationships between user-attribute/role-permission. How to efficiently determine whether an automatic authorization access control state satisfies the safety and availability requirements of a system is a huge challenge. Existing work often focuses on a single aspect of safety or availability, while ignoring the differences between permissions and the differences between users. In this paper, we first propose a fine-grained personalization policy that takes into account the specificity of permissions/users and describes the safety, availability and efficiency requirements of an access control system in CPS. Second, we define a Personalization Policy Checking (PPC) Problem to determine whether a given personalization policy is satisfied in an access control state. We give the computational complexity of the PPC problem in different subcases, and show that it is NP-complete in general. Third, we design a binary genetic search algorithm, whose improvements mainly include continuous update and selection of the best chromosomes in the population for iteration, and exploring and determining the optimal crossover and mutation probabilities, thereby improving the convergence efficiency of the algorithm. Finally, simulation results show the effectiveness of our proposed algorithm, which is especially fit for the case that the computational overhead is even more important than the accuracy in a large-scale CPS system.
|
Received March 2, 2021, accepted April 2, 2021, date of publication April 12, 2021, date of current version April 19, 2021.
_Digital Object Identifier 10.1109/ACCESS.2021.3072635_
# An Improved Genetic Algorithm for Safety and Availability Checking in Cyber-Physical Systems
ZHENG WANG 1,2, YANAN JIN3, SHASHA YANG4, JIANMIN HAN 5, AND JIANFENG LU 5
1Faculty of Science and Engineering, The Open University of China, Beijing 100039, China
2Department of Computer Science, Zhejiang Radio & Television University Haiyan College, Jiaxing 314300, China
3School of Information Management and Statistics, Hubei University of Economics, Wuhan 430205, China
4Xingzhi College, Zhejiang Normal University, Jinhua 321004, China
5Department of Computer Science and Engineering, Zhejiang Normal University, Jinhua 321004, China
Corresponding author: Yanan Jin (jinyanan@yhcrt.com)
This work was supported in part by the National Natural Science Foundation of China under Grant 62072411 and Grant 61872323, in part
by the Zhejiang Provincial Natural Science Foundation of China under Grant LR21F020001, in part by the Zhejiang Provincial Department
of Education under Grant Y202043497, and in part by the Social Development Project of Zhejiang Provincial Public Technology Research
under Grant 2017C33054.
**ABSTRACT** Cross-IoT infrastructure access frequently occurs when performing tasks in a distributed
computing infrastructure of a cyber-physical system (CPS). The access control technology that ensure secure
access cross-IoT infrastructure usually automatically establish relationships between user-attribute/rolepermission. How to efficiently determine whether an automatic authorization access control state satisfies
the safety and availability requirements of a system is a huge challenge. Existing work often focuses on a
single aspect of safety or availability, while ignoring the differences between permissions and the differences
between users. In this paper, we first propose a fine-grained personalization policy that takes into account the
specificity of permissions/users and describes the safety, availability and efficiency requirements of an access
control system in CPS. Second, we define a Personalization Policy Checking (PPC) Problem to determine
whether a given personalization policy is satisfied in an access control state. We give the computational
complexity of the PPC problem in different subcases, and show that it is NP-complete in general. Third,
we design a binary genetic search algorithm, whose improvements mainly include continuous update and
selection of the best chromosomes in the population for iteration, and exploring and determining the optimal
crossover and mutation probabilities, thereby improving the convergence efficiency of the algorithm. Finally,
simulation results show the effectiveness of our proposed algorithm, which is especially fit for the case that
the computational overhead is even more important than the accuracy in a large-scale CPS system.
**INDEX TERMS Access control, personalization policy, genetic algorithm, cyber-physical system.**
**I. INTRODUCTION**
Cyber-physical system is a controllable, credible, scalable
and heterogeneous distributed cyber-physical equipment system. It acquires information based on the IoT perception
environment, and processes the information through deeply
integrated computing, communication and control capabilities to complete a given task [1], [2]. CPS can bring huge
economic benefits and is widely used in digital medical
instruments and systems adopting automatic acquisition and
control technology, distributed energy systems, aerospace
and aircraft control, industrial control, etc [3]–[5]. CPS has
aroused great interest of industry investment and researchers.
The associate editor coordinating the review of this manuscript and
approving it for publication was Po Yang .
In the CPS environment, if users in local nodes or nodes
across IoT infrastructure access sensitive data without authorization, huge losses will occur [2], [6]. For cyber-physical
systems, safety is facing increasing challenges, because illegal access may also come from various networks and physical
interfaces in an increasing number of non-local IoT infrastructures [7]–[9]. Due to the heterogeneity of different IoT
infrastructures, traditional access control are less effective
in protecting sensitive data across IoT infrastructures. In the
field of distributed cyber-physical systems, the research of
access control is becoming more and more important for CPS
designers and users.
The autonomy, heterogeneity and distribution of CPS
nodes make access control mainly focus on multi-entity
access control between different trust domains, while taking
-----
into account geographic location and resource ownership
[10], [11]. The subject and object of access control are highly
dynamic in the CPS environment and there exists a huge
number of terminals and users. Therefore, the authorization
relationship between users and permissions cannot be presented in advance, and the system authorization can only
be performed automatically [11], [12]. However, whether
this automatically authorized access control state satisfies
the safety and availability requirements of the access control
system needs to be determined by corresponding access control policies. Therefore, the study of access control policies
cross-IoT infrastructure in the CPS environment has practical
theoretical significance and application value.
Access control policies restrict the assignment of permissions to ensure the safety and availability of the access control
system [13], [14]. However, there are still shortcomings in
the existing research on access control policies. (1) Access
control policies often focus on security or availability,
and cannot effectively balance these two [11]. Multiple
CPS nodes or even cloud nodes may be involved when
performing tasks. These autonomous nodes have their own
role-permission relationship and may not be able to accurately satisfy task requirements. The redundant permissions
generated by ensuring availability will bring security risks
to the system. If security is strictly ensured, it may lead to
insufficient permissions and affect the smooth execution of
tasks. (2) Existing access control policies consider a large
number of access permissions with negligible impact on
task execution, which will increase the scale of the problem
and reduce the efficiency of access control decision-making.
Ignoring the difference in nature importance between permissions and treating important permissions as ordinary permissions will also bring unpredictable risks to the system.
(3) Determining whether a certain access control policy is
satisfied in a system state is the key issue to efficient access
control decision-making. However, this problem is difficult
to solve, especially for access control systems authorized
across autonomous domains in CPS environments. This is
because the access control subjects and objects involved in the
execution of a task may come from different CPS nodes with
a large number of users and permissions. It is necessary to
determine whether the access control state composed of these
nodes satisfies the goals and constraints of considering the
weight. This greatly increases the computational complexity.
For example, an access control policy may require mutually
disjoint user groups that can perform tasks independently to
satisfy a certain number, and the weight of the permissions
owned by a single user is less than a certain threshold.
It can be seen that existing access control policies are
difficult to effectively ensure the safety and availability in
CPS, and it is intractable to improve the decision-making
efficiency of access control under a large amount of data.
For this reason, this paper introduces the concept of weight
of users and permissions, which expresses the importance of
permissions/users from the attributes of operations, the sensitivity of objects and user attributes. Subsequently, we
propose a refined personalization policy based on weights
to improve the efficiency of access control decision-making
while enhancing the safety and availability of the system.
Then, we analyze the computational complexity of the problem that a given access control state satisfies the requirements
of a personalization policy. To address this problem of general
case, we design an efficient solution based on the idea of
genetic algorithm. Generally, a given access control policy is
the minimum requirement of the system. For example, there
are three groups of mutually disjoint users in an access control
system, and each group has all the permissions to perform
sensitive tasks. But the access control policy requires two
groups of mutually disjoint users to ensure the availability
of the system. Therefore, verifying that the policy is satisfied
only needs to find two sets of mutually disjoint users. It can be
seen that this solution is more effective when the parameters
required by the policy are smaller than the actual parameters
of the system.
Briefly, the main contributions of this paper can be
summarized as follows:
- We propose a personalization policy that considers the
different natural importance of permissions and users.
This policy describes the safety, availability and efficiency requirements of the access control system in a
fine-grained way.
- We give a formal definition of the PPC problem which
determines whether an access control state in CPS environment satisfies a given personalization policy, and
present the computational complexity analysis of PPC
problem in different subcases. In particular, we show
that this problem is NP-complete in the general case.
- We design a Binary Genetic Search (BGS) algorithm,
which first considers the efficiency of solving PPC
problems. This algorithm improves the selection operation and crossover and mutation probability of genetic
algorithm.
- Simulation results further demonstrate the effectiveness
of the BGS algorithm, which is especially fit for the case
that the computational overhead is even more important
than the accuracy in a large-scale CPS system.
The rest of this paper is organized as follows. In Section II,
we start with an overview of previous literature. Section III
presents the formal definition of the personalization policy
and the PPC problem, and studies computational complexity
of its variants subcases. We present an algorithm for the
PPC problem in Section IV. In Section V, we implement the
proposed algorithm. We conclude this paper in Section VI.
**II. RELATED WORK**
The unique technical requirements and constraints of CPS
make the existing research on automatic authorization
of access control focuses on the discovery method of
attribute-permission association in attribute-based access
control (ABAC). And provides flexible control and management through the mapping mechanism of user-role and
role-permission in role-based access control (RBAC) [11].
-----
ABAC regards attributes as the key element of access control, which effectively solves the problem of large-scale,
dynamic and private fine-grained access control in the CPS.
ABAC first establishes the attribute set and describes the
access control policy, and then responds to the access control
request and updates the access control policy during execution [12]. RBAC guarantees flexible control and management
of objects through a dual authority mapping mechanism, and
provides inter-domain role mapping and constraint verification methods in cross-entity access control of CPS [15], [16].
When constructing attribute set and permission mapping,
usually use role engineering or attribute engineering topdown or bottom-up method to mine roles or attributes to further authorize users. However, the automatically authorized
access control state may not necessarily satisfy the safety and
availability requirements of the access control system.
The access control policy which used to restrict permission
assignment to ensure safety and availability in access control
system is a main research for several decades [17]. The
research of access control policy originated from the safety
analysis of access control system, which determines whether
an access control system can reach a state in which an unsafe
access is allowed [18]. In the earliest work, the safety of
the access control system is the focus of consideration, its
purpose is to ensure the safety of the access control system
when performing tasks and prevent abuse of authority. The
separation of duty (SOD) policies is a typical policy used to
ensure safety [19]. It prevent a set of users less than a certain
threshold from being fully authorized to perform sensitive
tasks [15], [16], [20]. Excessive pursuit of system safety may
lead to unavailability of the system. For example, an access
control state that satisfies strict safety requirements does
not have the full permissions to perform tasks. Therefore,
subsequent research also focuses on the availability of the
access control system. Resiliency policy requires that absent
any s users, there is still exist d mutually disjointed set of users
which number is less than t and each set has all permissions in
P to perform tasks to ensure availability of system [21], [22].
The problem of determine whether a certain access control state satisfies a given access control policy in the CPS
environment is difficult to solve. For example, the problem of
checking whether an access control state satisfies a resiliency
policy is intractable (NP-hard) in the general case, and is in
the Polynomial Hierarchy (in coNP[NP]) [21]. In this paper,
although we have comprehensively optimized the description of the policy to ensure that it is easier to solve while
enhancing the safety and available effect. However, the policy
proposed in this paper takes into account the weight of users
and permissions, which obviously increases the difficulty of
analyzing the problem.
The policy checking problem is difficult to solve under
general case. The existing access control policy checking
problem is to reduce the system scale through preprocessing,
and then solve it by a satisfiability problem (SAT) solver [22].
However, due to the massive data scale of the CPS environment, which makes the implementation of this scheme
require a great system overhead. Genetic algorithm has been
proved to be effective in dealing with many problems, especially in dealing with NP-complete problems [23]–[29]. This
is because the fitness value of the optimal solution can be
calculated for this type of problem. The optimization goal
of genetic algorithm is to make the solution set convergence
to the optimal solution with higher efficiency. For example,
the literature [24] proposed multi-granularity genetic algorithm that adopts a multi-granularity space strategy based
on a random tree, which accelerates the searching speed of
the algorithm in the multi-granular space. The literature [25]
optimized crossover and mutation operations were devised
to make the algorithm converge more quickly in solving the
multi-processor scheduling problem in cloud data-centers.
Aiming at the policy checking problem, this paper optimizes
the genetic algorithm in many aspects to achieve the ideal
solution effect.
In summary, the existing access control policy describes
the safety or availability of the access control system, but
it does not give a good balance between these two aspects,
and it is difficult to apply in a distributed CPS environment.
This paper proposes an access control policy applied in the
CPS environment, defines and analyzes the computational
complexity of the weighted policy check problem. Through
the analysis of genetic algorithm, it can be seen that the
algorithm can efficiently obtain the approximate solution of
the problem. Therefore, this paper improves the algorithm to
obtain better efficiency and accuracy.
**III. PROBLEM FORMULATION**
The individuality of every permission/user means that it has
different nature and importance. It is a key topic that should be
introduced to access control policy of CPS environment, but
ignored. In this section, we propose a personalization policy
that takes into account the specificity of permissions/users
and be used to ensure the safety, availability, and efficiency
of the access control system.
_A. PERSONALIZATION POLICY_
The personalization policy considers the particularity of
permissions that have different natural and importance.
In financial institution’s access control systems, for example,
the permission writes asset data is more important than the
permission reads asset data. The weight is a value attached
to a permission/user representing its importance and we
introduce it to personalization policy. Here, we present an
example to motivate the new features of the notation about
the weight of permission/user to optimize the access control
policy. Let us assume that the permission set is p1, p2, p3, p4,
permissions p1 and p2 are assigned to u1, permissions p3 and
_p4 are assigned to u2, permissions p1 and p3 are assigned to_
_u3, permissions p2 and p4 are assigned to u4. It is obvious_
that both {u1, u2} and {u3, u4} are the solutions and each
solution has all permissions to perform tasks. However, it may
not make any sense for choosing {u1, u2}, if the permissions
_p1and p2 are more important resulting weighted u1 beyond_
-----
a certain threshold. This is because that it is easier to put
the system at unpredictable risk if a user has too important
permissions. Furthermore, certain permissions that may be
more critical for system can only be owned by special users,
other users cannot be authorized in the process of performing
sensitive tasks. Safety is an important factor that we consider,
and availability also needs to be considered because it is
related to the smooth execution of the task. For example,
in the previous example, there are two mutually disjoint user
sets to perform sensitive task, this means that even if any one
of the users is absent, the task can still be executed.
There are a lot of resources in the CPS environment.
If the access permissions of these resources are all taken into
account in the access control system, it will bring great system
overhead and affect the efficiency of access control decisionmaking. Therefore, in order to enhance the availability of
the system, we do not consider non-essential permissions
into the access control system. We use weights to indicate
the importance of these resource access permissions to the
system. We set a threshold according to the importance of the
task, and do not add permissions with a weight less than a
certain threshold to the access control policy. This is because
the abuse of these permissions with lower weights has a
tolerable impact on the smooth execution of tasks, and the
deficiencies of these permissions can be resolved through
temporary authorization.
The weight of permissions/users is a value between 0-1
that weighs the importance of permissions/users from the
attributes of operations, the sensitivity of objects, and user
attributes [30]. In this section, formal definition of the weight
of permission and methods of calculating them is not discussion. We assume that the weight of permissions is determined
by the system and the weight of users is the sum of the
weighted user’s permissions. The personalization policy is
defined as follows.
_Definition 1 (Personalization Policy): Given a set U of_
_users, a set P of permissions, the personalization policy sat-_
_isfy the following constraints:_
- Safety constraint: A safety constraints is denoted as
_PP⟨ω, UF_ (pf )⟩, where ω ≥ 0. pf is very important to
_the system and can only be assigned to users in the user_
_set UF_ _.We say that PP⟨U_ _, P, ω, UF_ (pf )⟩ _is satisfied if_
_and only if the following conditions hold:_
- ∃pf ∈ _P(UF_ ) and pf /∈ _P(Un) where Un = U −_
_UF_ _, P(UF_ ) denotes all permissions assigned to the
_users set UF_ _._
- ∃ui ∈ _UP and UP =_ [�]W (uj)<ω _[u][j][, where u][j][ ∈]_ _[U]_
_and W_ (uj) denotes the weight of the uj.
- Efficiency **_constraint:_** _A_ _efficiency_ _constraints_ _is_
_denoted as PP⟨ω0⟩, where 1 ≥_ _ω0 ≥_ 0, We say that
_PP⟨U_ _, P, ω0⟩_ _is satisfied if and only if the following_
_conditions hold:_
- ∃pi ∈ _PP and PP =_ [�](W (pj)>ω0) _[p][j][ −]_ _[P][F]_ _[, where]_
_pj_ _P and PF_ _pf denotes the permissions set_
∈ = [�]
_of all pf ._
- Available constraint: A available constraints is denoted
_as PP⟨U_ _, P, κ⟩, where 0 ≤_ _κ ≤_ _n are positive integer._
_We say that PP⟨κ⟩_ _is satisfied if and only if the following_
_conditions hold:_
- ∃{P(Ui), P(Uj)}, and P(Ui) = P(Uj) = PP, Ui ∩
_Uj = φ, and Ui, Uj ⊆_ _UP. where 0 ≥_ _i, j ≥_ _κ,_
_i_ _j._
̸=
In order to distinguish different types of permissions
and user groups. We define PP, UP, PF _, UF as pivotal per-_
missions, pivotal users, fixed authorized permissions and
fixed authorized users respectively, as shown in definition 1.
We define PN _P/PF as non-fixed authorized permissions._
=
We define the permissions with a weight less than ω0 as
general permission, denoted as PG, and users with a weight
greater than ω0 are dangerous users, denoted as UD.
To specify a subcase of the personalization policy, we combine the three constraints and write it followed by the
list of constraints within a pair of braces. For instance,
_PP⟨P, U_ _, κ, ω0, ω, {Uf 1(pi), . . ., Ufn(pj)}⟩. An access con-_
trol state satisfies such a personalization policy if and only if
fixed authorized permissions {pi, . . ., pj}only belongs fixed
authorized users {Uf 1, . . ., Ufn} respectively, exist at least
_κ mutually disjoint sets of users such that each set has all_
authorized pivotal permissions and total weight of permissions authorized by each users is no more than ω.
Suppose we now give a personalization policy as PP⟨P, U _,_
_κ, ω0, ω, {Mike(Ratify)}⟩. This policy requires that fixed_
authorizations permission ratify only assigned to user Mike.
If κ = 2 and ω0 = 0 is set, the policy requires that overall
permissions except ratify are assigned to at least two mutually
disjointed sets of users. If κ = 2 and ω0 = 0.35 is set,
the permission excepted not only ratify but also permissions
with a weight less than 0.35. If ω = 1.2 is set, this means
that the weight of each user in each mutually disjointed user
groups is no more than 1.2. If we set ω = ∞, this means that
the weight of users is unrestricted.
_Example 1: Given the access control state shown in_
_Figure 1, all permissions in a fund publishing task are P_
=
_input, issue, view, ratify_ _and weighted to 0.7,0.5,0.3 and_
{ }
_0.9, respectively. All users are U_ _Alice, Bob, Ed, Mik,_
= {
_Harry, Jack_ _._
}
As shown in Figure 1, the personalization policy
_PP⟨P, U_ _, 2, 0, 1.2, {Mike(Ratify)}⟩_ is satisfied, because existence of U1 = {alice, ed} and U2 = {bob, jack} have
full pivotal permissions and weighted each user no more
than 1.2. However, PP⟨P, U _, 2, 0, 1, {Mike(Ratify)}⟩_ is not
satisfied, because the weight of U2 alone does not exceed 1.
_PP⟨P, U_ _, 3, 0, ∞, {Mike(Ratify)}⟩_ is not satisfied, because
this access control state has only two mutually disjoint
sets of pivotal users with all pivotal permissions. But
_PP⟨P, U_ _, 3, 0.35, ∞, {Mike(Ratify)}⟩_ is satisfied, because
this access control state has three mutually disjoint sets
of pivotal users has pivotal permissions input and issue,
the weight of permission view is less than 0.35 means that
it’s not importance for the task, so the access control system
is not considered.
-----
**FIGURE 1. An example of access control state.**
The parameters κ requires that existing κ mutually disjoint
sets of users can be perform tasks respectively, mean that any
_κ −_ 1 pivotal users to be absent in emergency situations, there
is still exist one independent team of users to perform tasks.
Such as in the example 1, the access control state satisfies
_κ = 2, mean that the system can be able to tolerate any one_
pivotal user absent. Furthermore, even if absents any number
of pivotal users in κ −1 user sets, the system can still perform
tasks. The parameters ω requires that the weight of a single
user in any user set is no more than ω, which prevents a
single user has more importance permissions to ensure the
system safety. Obviously, if the parameters ω is given, then
the number of users in each sets is no less than ⌈W (PP)/ω⌉,
where W (PP) is weight of all pivotal permissions. Such as
in the Example 1, if given ω = 0.8 then ⌈W (PP)/ω⌉= 2,
it means the number of users in each sets is no less than 2.
_B. PERSONALIZATION POLICY CHECKING PROBLEM_
In access control system, U represents all users and P represents all permissions, assignment relationship between the
user and the permission is represented as UP _U_ _P. How_
⊆ ×
to efficiently determine whether the existing access control
state UP satisfies a given access control policy is the key to
the access control decision. For this reason, we now give a formal definition of the problem and analyze its computational
complexity.
_Definition 2 (Personalization_ _Policy_ _Checking_ _(PPC)_
_Problem): Given a personalization policy PP and an access_
_control state UP, UP satisfies PP is denoted as the satPP(UP)._
_Determining whether satPP(UP) is true is called Personaliza-_
_tion Policy Checking Problem._
In special cases, the parameters of personalization policy PP are not always fully consider. For example, a personalization policy in the subcase PPC⟨κ = 1⟩ has
the form PP⟨P, U _, 1, ω0, ω, {Uf 1(pi), . . ., Ufn(pj)}⟩_ which
means determines whether there exists a set of users have
all pivotal permissions in P and weight of each user no more
than ω. The subcase PPC⟨ω = ∞⟩ determines whether exist
_κ sets of users and each set has all pivotal permissions in P._
The computational complexity results for PPC problem and
it’s various subcases are given as following theorem.
**FIGURE 2. Computational complexity results for PPC problem in various**
subcases.
_Theorem 1: The computational complexity of PPC prob-_
_lem and its subcases is shown in Figure 2._
We study the computational complexity of PPC problem
in various subcases. The following lemma shows that the
_PPC⟨κ = 1⟩, PPC⟨ω = ∞⟩, PPC⟨_ ⟩ are NP-complete.
_Lemma 1: PPC⟨κ = 1⟩_ _is NP-complete_
_Proof: We prove that the PPC⟨κ = 1⟩_ is an NP problem:
given a solution of the PPC⟨κ = 1⟩ problem, it can be
verified in polynomial time whether the solution is correct.
Next, we convert the NP-complete weighted set covering
decision problem [31] to PPC⟨κ = 1⟩ problem in Polynomial
time, and show PPC⟨κ = 1⟩ is NP-complete. In the weighted
set covering problem, given a finite set S, a family F
=
{S1, . . ., Sm} of subsets of S, and a budget B, the goal is to
determine whether the weight of each Si is less than B, where
the union of Si is S. Given an instance of the weighted set
cover problem, we now construct an instance of PPC⟨κ = 1⟩
in the following way: We create permissions p1, . . ., pm for
each element in S, let ω = B, m is the cardinality of the
set S. we create PP⟨P, U _, 1, ω0, ω, {Uf 1(pi), . . ., Ufn(pj)}⟩_
and create an access control state: For each different subset
_Si(1 ≤_ _i ≤_ _m) in F, create a user ui, so that all per-_
missions and their weight values in Si are assigned to ui.
Then whether PP⟨P, U _, 1, ω0, ω, {Uf 1(pi), . . ., Ufn(pj)}⟩_ is
true if and only if there is a union of subsets in F that
covers S, and the weight of any set in the subset is less
than B.
-----
Therefore, the PPC problem when κ = 1 is NP-complete
problem. _Lemma 2: PPC⟨ω = ∞⟩_ _is NP- complete_
_Proof:_ We prove that the PPC⟨ω = ∞⟩ is an NP
problem: given a solution of the PPC⟨ω = ∞⟩ problem,
it can be verified in polynomial time whether the solution is
correct.
Next, we reduce the NP-complete DOMATIC NUMBER
problem [32] to PPC⟨ω = ∞⟩. Given a graph G(V _, E),_
the DOMATIC NUMBER problem asks whether V can be
partitioned into κ mutually disjoint sets V1, V2, . . ., Vk such
that each Vi is a dominating set for G. V[‘]is a dominating set
for G(V _, E) if for every node u in V −_ _V_ [‘], there is a node v
in V[‘] such that (u, v) ∈ _E. An instance of PPC⟨ω = ∞⟩_
asks whether an access control state UP satisfies a policy
_PP⟨P, U_ _, κ, ω0, ∞, {Uf 1(pi), . . ., Ufn(pj)}⟩. Given a graph_
_G = (V_ _, E), we now construct an instance of PPC⟨ω = ∞⟩_
in the following way: We construct an access control state UP
with n users u1, u2, . . ., un for n nodes in G and n permissions
_p1, p2, . . ., pn. v(ui) denotes the node corresponding to user_
_ui. In UP, user ui is authorized for the permission pj if and_
only if either i = j or (v(ui), v(uj)) ∈ _E. Let P denote the set_
{p1, p2, . . ., pn}. A dominating set in G corresponds to a set
of users that together have all permissions in P. UP satisfies
_PP⟨P, U_ _, κ, ω0, ∞, {Uf 1(pi), . . ., Ufn(pj)}⟩_ if and only if V
contains κ mutually disjoint dominating sets.
Therefore, the PPC problem when ω = ∞ is NP-complete
problem. _Lemma 3: PPC_ _is NP-complete_
⟨ ⟩
_Proof: An instance consists of an access control state UP_
and a policy PP⟨P, U _, κ, ω0, ω, {Uf 1(pi), . . ., Ufn(pj)}⟩. UP_
satisfies PP⟨P, U _, κ, ω0, ω, {Uf 1(pi), . . ., Ufn(pj)}⟩_ if and
only if there exist at least κ mutually disjoint sets of users
such that each set has all authorized pivotal permissions and
total weight of permissions authorized by each user is no
more than ω. If these κ sets are given, they can be verified
in polynomial time. Therefore, PPC is in NP, and the
⟨ ⟩
subcase of PPC is NP-complete, then the PPC is
⟨ ⟩ ⟨ ⟩
NP-complete.
**IV. THE BINARY GENETIC SEARCH ALGORITHM FOR PPC**
The fact that PPC problem is intractable, as shown in
Theorem 1, means that there exist difficult problem instances
that take exponential time in the worst case. Therefore,
we propose a Binary Genetic Search (BGS) algorithm to
approximate solve PPC problems, which is inspired by the
idea of the Genetic algorithm.
First, this algorithm performs preprocessing to reduce
the system scale. Second, this algorithm execute optimized
genetic algorithm and search algorithm within T seconds
of system tolerance time. During this time, the number of
mutually disjointed user sets which found in the first half of
the population satisfy the parameters κ of policy, then stop
and output result: true. If not, save the mutually disjoint user
groups, randomly generate new chromosomes, and continue
**TABLE 1. Main notations used in this algorithm.**
to iterate until κ groups are found. If the running time more
than the system tolerance time of T seconds, it is uncertain
whether the policy is satisfied, and the output result: false.
This algorithm has a time complexity of O(lmn), where l,
_m and n denote the number of actually performed iterations,_
the size of population and the number of all available users,
respectively. The main notations used in this paper are shown
in Table 1. Algorithm 1 shows the process of BGS for PPC
problem.
**Algorithm 1 BGS for PPC**
**Data: UP[m][n], W** (pi), PP, Pm, Pc, T
**Result: O[m][n], Str**
**1 Preprocessing();**
**2 while runtime < Tsecond and κ < max do**
**3** OGA( ) ;
**4** Search( );
**5** **if κ ≥** _max then_
**6** Str True ;
=
**7** exit(0);
**8** **end**
**9 end**
**10 if κ < max and runtime ≥** _Tsecond then_
**11** Str False ;
=
**12 end**
**13 return: Str;**
This algorithm is optimized based on the idea of genetic
algorithm, and has the characteristics of rapid convergence
and evolution to the optimal solution. At the same time,
because the PPC problem is an NP-complete problem, it can
be determined in polynomial time whether the obtained solution is optimal.The algorithm is divided into three parts
as shown in Algorithm 1. The first part is preprocessing,
as shown in Algorithm 2; The second part performs optimized
genetic algorithm as shown in Algorithm 3; the third part is to
find mutually disjoint user groups such as algorithm 4 shown.
-----
_A. PREPROCESSING_
We first determine whether the fixed authorization permissions in PP⟨P, U _, κ, ω0, ω, {Uf 1(pi), . . ., Ufn(pj)}⟩_ in the
preprocessing part only belongs to the fixed authorized user,
that is, determine whether {Uf 1(pi), . . ., Ufn(pj)} is true, if it
is false, the policy is not satisfied. Secondly, we perform static
pruning of users and permissions based on PP to reduce the
scale of problem solving, which is of great help to improve the
access control decision-making efficiency of CPS. Finally,
we transform the PPC problem into the chromosome of
genetic algorithm through coding. The preprocessing process
in this section is shown in Algorithm 2.
**Algorithm 2 Preprocessing Function Algorithm for**
PPC Problem
**Data: UP[m][n], W** (pi), PP
**Result: UpPp[m][n], Str**
**1 if Uf 1(pi), . . ., Ufn(pj) == ∅** **then**
**2** Str False ;
=
**3** exit(0);
**4 else**
**5** **foreach pi do**
**6** **if pi** _Pf then_
∈
**7** _P = P/pi;_
**8** **end**
**9** **if W** (pi) < ω0 then
**10** _P = P/pi;_
**11** **end**
**12** **end**
**13** **foreach W** (ui) > ω do
**14** _U = U_ _/ui;_
**15** **end**
**16 end**
**17 return: Str;**
1) STATIC PRUNING
The access permissions of large-scale resources in the CPS
environment are taken into account in the access control
decision system, which causes a large system overhead.
Therefore, this section uses static pruning to delete users and
permissions that do not need to be considered during the
execution of the algorithm to improve the decision-making
effectiveness of the access control system. Users and permissions in the following situations do not need to be considered.
- Fixed authorization permissions: For safety reasons,
fixed authorization permissions can only be owned by
specific users, while other users cannot be authorized,
so we need to exclude these permissions when considering availability.
- Permission with weight less than ω0: The importance of
the permission is less than a certain threshold, so the
permission does not need to be considered to improve
the efficiency of access control decision-making.
During task execution, the lack of such permissions can
be obtained through temporary authorization.
- Users whose weight is greater than ω: If a selected
user’s weight is greater than ω, it does not satisfy the
requirements of the access control policy, so there is no
need to consider it.
2) ENCODING
After static pruning of users and permissions, a sub-state of
the access control state composed of pivotal users and permissions is formed. Next, we optimize the genetic algorithm to
discover the user group containing all the pivotal permissions.
The genetic algorithm coding rules are as follows:
Given an access control state UPPP, UP represents a set of
m pivotal users, and PP represents a set of n pivotal permissions. We use m-bit chromosomes to represent m users. When
the i-th chromosome is 1, it means that user ui is selected.
_B. OPTIMIZATION GENETIC ALGORITHM_
In this section, we introduce the optimized genetic algorithm
(OGA). The core idea of the OGA function is to carry out
genetic iterations according to the optimal crossover and
mutation probabilities determined by experiments, updated
optimal half of the population after Each iteration completes,
and continue iterating with this population. Until the fitness
of the first half of the population is the same and it is equal
to the maximum value of fitness, and the value of relative
fitness is also in a reasonable range. This means that the
user set selected by each chromosome in the first half of
the population covers all pivotal permissions. The execution
steps of the optimized genetic algorithm (OGA) function
are as follows, and Algorithm 3 gives the detailed execution
process.
step i Select a population of m points x1, . . ., xm to represent the users set at random.
step ii Compute fitness: Compute the fitness and relative
fitness of the role set using the evaluation function
respectively.
step iii Replacement: Sort the m points according to the
fitness value from large to small, sort the points with
the same fitness according to the relative fitness, and
then replace the latter half with the front half.
step iv Mutate: For each point xi that m/2 < i ≤ _m in the_
population and for each bit in xi, with probability pm,
alter its value.
step v Crossover: For each yj in the pair points xi and x(i+1)
from the xm/2, . . ., xm, with probability pc, exchange
_xi.yj with x(i+1).yj._
step vi Stop: If the front half of the population has the same
fitness and equal to the maximum fitness, at the
same time, the value of relative fitness is also in a
reasonable range, stop.
_C. FINDING MUTUALLY DISJOINT USER GROUPS_
This section finds whether there are κ groups of mutually
disjoint user groups in the solution E[i] of the optimized
-----
**Algorithm 3 OGA Function Algorithm for PPC**
Problem
**Data: UpPp[m][n], W** (pi), PP, Pm, Pc
**Result: E[i]**
**1 Rand(E[m]);**
**2 while E[0].fit ̸= E[m/2 −** 1].fit ̸= Ps and
_E[m/2 −_ 1].relfit ≥ _W_ (Pi) ∗ 2 do
**3** **foreach i ∈** [0, n) do
**4** _E[i].fit ←_ _Ps −_ _wep −_ 100wmp ;
**5** _E[i].relfit ←_ _W_ (E[i]) ;
**6** **end**
**7** Sort(E[m].fit,E[m].relfit) ;
**8** **foreach i ≤** _m/2 do_
**9** _E[m/2 + i] ←_ _E[i];_
**10** **end**
**11** **foreach i > m/2 do**
**12** **foreach j ∈** [0, n) do
**13** **if rand() < pm then**
**14** _E[i].U_ [j] ← (E[i].¯U [j]);
**15** **end**
**16** **end**
**17** **foreach i%2** 0 do
==
**18** **foreach j ∈** [0, n) do
**19** **if rand() < pc then**
**20** _E[i].U_ [j] ↔ _E[i + 1].U_ [j];
**21** **end**
**22** **end**
**23** **end**
**24** **end**
**25 end**
**26 return:E[i];**
genetic algorithm. This paper proposes two methods. The
first one is to encode the obtained solution E[i] and use the
optimized genetic algorithm to solve it again. The second
method is to find whether there are κ groups mutually disjoint
user groups. This method is an approximate solution method.
The process is: If the obtained solution E[i] belongs to a
subset of any solution in the mutually disjoint solution set
_O[k], replace it. If it is a solution that does not intersect_
with O[k], add into the solution set. For the simulation
experiment, we use the second method, which is shown
in Algorithm 4.
The BGS algorithm is optimized based on the idea of
genetic algorithm, and the improvements are as follows: First,
in the process of execution, the optimal half of the population
obtained by evolution is always updated, and the population
is used to continue iterating. This is because the evolution
based on the best solution has a high probability to get the
better solution. Second, the mutation and crossover probability are determined through experiments. In the experiment,
the value of the mutation probability is an integer multiple
of the reciprocal of the population size. Third, the crossover
operation selects the chromosome with the closest fitness.
**Algorithm 4 Search Function Algorithm for PPC**
Problem
**Data: E[i]**
**Result: k**
**1 foreach i < m/2 and k ≤** _max do_
**2** **if E[i]** _O[k] then_
⊆
**3** _O[k]_ _E[i]_
← ;
**4** **end**
**5** **if E[i] ∩** _O[k] == φ then_
**6** _O[_ _k]_ _E[i]_
+ + ← ;
**7** **end**
**8 end**
**9 return:k ;**
These improvements greatly improve the efficiency of the
algorithm converging to the optimal solution.
**V. IMPLEMENTATION AND EVALUATION**
In order to verify the effectiveness of the BGS algorithm,
we have implemented it and performed several experiments
using randomly generated instances. The implementation of
our algorithm was written in C. Experiments have been carried out on a PC with an Intel(R) Core(TM) i5-8500T CPU
running at 2.11 GHz, and with 4GB memory, running windows 10. In order to get closer to the real access control environment, we add two interference permissions that are not
related to the task. It is assumed that the fixed authorization
permissions satisfy the policy requirements and are pruned to
generate instances. For each instance, 10 randomly generated
test cases are run, the averages time of the test results are used
to generate the runtime graphs, and the number of satisfaction
in ten instances are used to generate another graph.
The evaluation function is used to evaluate the solutions.
The fitness function is defined as Ps − _wep −_ 100wmp, for
more details, please refer to [23]. The relative fitness function
is defined as follows.
_Definition 3 (Evaluation Function of Relative Fitness):_
_Relative fitness of E[i] is defined as:_
_where WPaf (E[i].U_ [j]) represents weight of permissions only
_owned by E[i].U_ [j].
_A. EFFECTIVENESS OF MUTATION AND CROSSOVER_
Figure 3 shows the average CPU times and number of satisfaction under different probability of crossover and mutation
for the two test case (1) Usize = 60, permissons = 12 and
_κ = 3; (2) Usize = 105, permissons = 7 and κ = 5,_
the size of population m 280, and the system tolerance
=
time t 30. The x-axis denotes the probability of mutation,
=
and we fix its value as 1/Usize, . . ., 8/Usize respectively. It can
be clearly seen from Figure3(a) and (c) that the average CPU
times is least when we choose the parameter Pm = 3/Usize
_E[i].relfit =_
_j=Us_
�
_WPaf (E[i].U_ [j]), (ifE[i].U [j] = 1)
_j=0_
-----
**FIGURE 3. The runtime and number of satisfaction for different probability of mutation and crossover.**
with fixed Pc. This means that it’s easier to obtain a solution
quickly by simultaneously mutating 3 bits in a chromosome.
The average CPU times increases with the maximal Pc for
the fixed Pm, because the less Pc will save the CPU times.
As shown in Figure3(b)(d), the number of satisfaction is
maximum when we choose the parameter Pm = 3/Usize
or Pm 5/Usize with fixed Pc, and when we choose the
=
parameter Pc is close to 0.5 with fixed Pm. Together with
the observation, we choose the parameters Pm 3/Usize and
=
_Pc = 0.5 for the remainder experiments._
Figure3(c)(d) shows longer CPU time and higher number of satisfaction than Figure3(a)(b). This is because when
the ratio of users to permissions is large, it is easier to
obtain mutually disjoint user groups, and the CPU time consumed will be reduced. The Figure3(c)(d) is clearer than
Figure3(a)(b) on the curve trend of CPU time and the number
of satisfaction. This is because if the ratio of users to permissions is small, the number of mutually disjoint user groups
in the system is also small. In this case, it is difficult for the
system to obtain a solution that satisfies the policy, and it may
even not have a solution that satisfies the requirements of the
policy. Therefore, if the ratio of users to permissions is small,
the running time and the number of satisfactions of different
random instances are very different.
_B. RUNTIME AND NUMBER OF SATISFACTION_
_FOR BGS ALGORITHM_
Figure4 shows the results of running the experiments for the
four test case (a) Usize : Psize = 5 : 1; (b) Usize : Psize =
10 : 1; (c) Usize : Psize = 15 : 1;(d) Usize : Psize = 20 : 1.
The runtime and number of satisfaction depend on the total
number of the users Usize, pivotal authorized permissions
_Psize, and parameter κ of the personalization policy._
In Figure4, as the parameter κ increases for the fixed Usize,
the number of satisfaction reduces and the overall CPU time
increases. This is because the larger the parameters required
by the policy, the more difficult to satisfy for the system.
As the total number of the users Usize increases for the
fixed parameter κ, the number of satisfaction reduces and the
overall CPU time increases, this change is not obvious when
the value of κ is small. But the change is obvious when the
value of κ becomes large, as shown in Figure4(g)(h). This
is because as the policy parameter κ increases, it is more
difficult for the system to obtain a solution that satisfies the
policy, and the running time of some instances may reach the
system tolerance time. The number of satisfaction increases
also with the maximal Usize _Psize for the fixed Usize and κ._
:
The reason is that the more value of Usize _Psize the more_
:
number of mutually disjoint sets of users. In Figure4(f)(h),
-----
**FIGURE 4. The runtime and number of satisfaction for different users and the parameters κ of policy.**
as the number of Usize increases, the number of satisfaction
reduced when the parameter κ more than 3. The reason is
that the BGS algorithm will stop when CPU times are over
the system tolerance time 30 second. Therefore, if we want
to obtain the better number of satisfaction, we can increase
tolerance time of the system.
Consequently, for the case that the system tolerance time is
more important, we can make the BGS algorithm obtain the
-----
best possible solution within the system tolerance time. The
BGS algorithm is able to solve the PPC problem even though
in a larger scale system.
**VI. CONCLUSION**
In this paper, we have proposed a personalization policy that
has reflected in the particularity of permissions/users and has
described the safety, availability and efficiency requirements
of the access control system in a fine-grained way. We have
introduced the definition of PPC problems and have studied
the computational complexity analysis of various subcases.
We have shown that most instances of PPC problems are
intractable. In particular, we have proposed a BGS algorithm
to solve PPC problems. This algorithm has greatly improved
the efficiency of the algorithm converging to the optimal
solution of the PPC problem within the tolerance time of the
system.
**REFERENCES**
[1] M. Wankhade and S. V. Kottur, ‘‘Security facets of cyber physical system,’’
in Proc. 3rd Int. Conf. Smart Syst. Inventive Technol. (ICSSIT), Tirunelveli,
India, Aug. 2020, pp. 359–363.
[2] Y. Javed, M. Felemban, T. Shawly, J. Kobes, and A. Ghafoor, ‘‘A partitiondriven integrated security architecture for cyberphysical systems,’’ Com_puter, vol. 53, no. 3, pp. 47–56, Mar. 2020._
[3] T. Wang, D. Zhao, S. Cai, W. Jia, and A. Liu, ‘‘Bidirectional predictionbased underwater data collection protocol for end-edge-cloud orchestrated
system,’’ IEEE Trans. Ind. Informat., vol. 16, no. 7, pp. 4791–4799,
Jul. 2020.
[4] T. Wang, H. Luo, X. Zeng, Z. Yu, A. Liu, and A. K. Sangaiah, ‘‘Mobility
based trust evaluation for heterogeneous electric vehicles network in smart
cities,’’ IEEE Trans. Intell. Transp. Syst., vol. 22, no. 3, pp. 1797–1806,
Mar. 2021.
[5] X. Liu, M. S. Obaidat, C. Lin, T. Wang, and A. Liu, ‘‘Movement-based
solutions to energy limitation in wireless sensor networks: State of the art
and future trends,’’ IEEE Netw., vol. 35, no. 2, pp. 188–193, Mar. 2021.
[6] J. Lu, C. Tang, X. Li, and Q. Wu, ‘‘Designing socially-optimal rating
protocols for crowdsourcing contest dilemma,’’ IEEE Trans. Inf. Forensics
_Security, vol. 12, no. 6, pp. 1330–1344, Jun. 2017._
[7] X. Chen, C. Li, D. Wang, S. Wen, J. Zhang, S. Nepal, Y. Xiang, and
K. Ren, ‘‘Android HIV: A study of repackaging malware for evading
machine-learning detection,’’ IEEE Trans. Inf. Forensics Security, vol. 15,
pp. 987–1001, 2019.
[8] T. Wang, Y. Lu, J. Wang, H.-N. Dai, X. Zheng, and W. Jia, ‘‘EIHDP:
Edge-intelligent hierarchical dynamic pricing based on cloud-edge-client
collaboration for IoT systems,’’ IEEE Trans. Comput., early access,
[Feb. 19, 2021, doi: 10.1109/TC.2021.3060484.](http://dx.doi.org/10.1109/TC.2021.3060484)
[9] J. Lu, Y. Xin, Z. Zhang, X. Liu, and K. Li, ‘‘Game-theoretic design
of optimal two-sided rating protocols for service exchange dilemma in
crowdsourcing,’’ IEEE Trans. Inf. Forensics Security, vol. 13, no. 11,
pp. 2801–2815, Nov. 2018.
[10] J. Giraldo, E. Sarkar, A. A. Cardenas, M. Maniatakos, and M. Kantarcioglu,
‘‘Security and privacy in cyber-physical systems: A survey of surveys,’’
_IEEE Des. Test. IEEE Des. Test., vol. 34, no. 4, pp. 7–17, Aug. 2017._
[11] J. H. Huh, R. B. Bobba, T. Markham, D. M. Nicol, J. Hull, A. Chernoguzov,
H. Khurana, K. Staggs, and J. Huang, ‘‘Next-generation access control
for distributed control systems,’’ IEEE Internet Comput., vol. 20, no. 5,
pp. 28–37, Sep./Oct. 2016.
[12] Z. Xu and S. D. Stoller, ‘‘Mining attribute-based access control policies,’’
_IEEE Trans. Dependable Secure Comput., vol. 12, no. 5, pp. 533–545,_
Sep. 2015.
[13] A. Margheri, M. Masi, R. Pugliese, and F. Tiezzi, ‘‘A rigorous framework
for specification, analysis and enforcement of access control policies,’’
_IEEE Trans. Softw. Eng., vol. 45, no. 1, pp. 2–33, Jan. 2019._
[14] P. Yang and L. Xu, ‘‘The Internet of Things (IoT): Informatics methods
for IoT-enabled health care,’’ J. Biomed. Informat., vol. 87, pp. 154–156,
Nov. 2018.
[15] M. U. Aftab, Z. Qin, N. W. Hundera, O. Ariyo, N. T. Son, and T. V. Dinh,
‘‘Permission-based separation of duty in dynamic role-based access control
model,’’ Symmetry, vol. 11, no. 5, p. 669, May 2019.
[16] X. Ma, R. Li, Z. Lu, J. Lu, and M. Dong, ‘‘Specifying and enforcing the principle of least privilege in role-based access control,’’
_Concurrency Comput., Pract. Exper., vol. 23, no. 12, pp. 1313–1331,_
Mar. 2011.
[17] M. Narouei, H. Takabi, and R. Nielsen, ‘‘Automatic extraction of access
control policies from natural language documents,’’ IEEE Trans. Depend_able Secure Comput., vol. 17, no. 3, pp. 506–517, Jun. 2020._
[18] M. A. Harrison, W. L. Ruzzo, and J. D. Ullman, ‘‘Protection in operating
systems,’’ Commun. ACM, vol. 19, no. 8, pp. 461–471, Aug. 1976.
[19] D. D. Clark and D. R. Wilson, ‘‘A comparison of commercial and military
computer security policies,’’ in Proc. IEEE Symp. Secur. Privacy, Oakland,
CA, USA, Apr. 1987, pp. 184–194.
[20] T. Zhu, D. Ye, W. Wang, W. Zhou, and P. Yu, ‘‘More than privacy: Applying differential privacy in key areas of artificial intelligence,’’ IEEE Trans. Knowl. Data Eng., early access, Aug. 4, 2021,
[doi: 10.1109/TKDE.2020.3014246.](http://dx.doi.org/10.1109/TKDE.2020.3014246)
[21] J. Crampton, G. Gutin, and R. Watrigant, ‘‘Resiliency policies in access
control revisited,’’ in Proc. 21st ACM Symp. Access Control Models Tech_nol., Shanghai, China, Jun. 2016, pp. 101–111._
[22] N. Li, Q. Wang, and M. Tripunitara, ‘‘Resiliency policies in access
control,’’ ACM Trans. Inf. Syst. Secur., vol. 12, no. 4, pp. 20:1–20:34,
Apr. 2009.
[23] J. Lu, Z. Wang, D. Xu, C. Tang, and J. Han, ‘‘Towards an efficient approximate solution for the weighted user authorization query
problem,’’ IEICE Trans. Inf. Syst., vol. E100.D, no. 8, pp. 1762–1769,
Aug. 2017.
[24] C. Li, S. Xia, Z. Chen, and G. Wang, ‘‘A multi-granularity genetic algorithm,’’ in Proc. IEEE Int. Conf. Big Knowl. (ICBK), Beijing, China,
Nov. 2019, pp. 145–151.
[25] Y. Xiong, S. Huang, M. Wu, J. She, and K. Jiang, ‘‘A Johnson’s-rule-based
genetic algorithm for two-stage-task scheduling problem in data-centers of
cloud computing,’’ IEEE Trans. Cloud Comput., vol. 7, no. 3, pp. 597–610,
Jul./Sep. 2019.
[26] L. Cui, J. Zhang, L. Yue, Y. Shi, H. Li, and D. Yuan, ‘‘A genetic algorithm
based data replica placement strategy for scientific applications in clouds,’’
_IEEE Trans. Services Comput., vol. 11, no. 4, pp. 727–739, Jul./Aug. 2018._
[27] A. Zhdanov, ‘‘Generation of static YARA-signatures using genetic algorithm,’’ in Proc. IEEE Eur. Symp. Secur. Privacy Workshops (EuroS&PW),
Stockholm, Sweden, Jun. 2019, pp. 220–228.
[28] B. Cao, S. Fan, J. Zhao, P. Yang, K. Muhammad, and M. Tanveer,
‘‘Quantum-enhanced multiobjective large-scale optimization via parallelism,’’ Swarm Evol. Comput., vol. 57, Sep. 2020, Art. no. 100697.
[29] G. Lin, S. Wen, Q.-L. Han, J. Zhang, and Y. Xiang, ‘‘Software vulnerability
detection using deep neural networks: A survey,’’ Proc. IEEE, vol. 108,
no. 10, pp. 1825–1848, Oct. 2020.
[30] X. Ma, R. Li, and Z. Lu, ‘‘Role mining based on weights,’’ in Proc. 15th
_ACM Symp. Access Control Models Technol. (SACMAT), Pittsburgh, PA,_
USA, Jun. 2010, pp. 65–74.
[31] J. Lu, J. B. D. Joshi, L. Jin, and Y. Liu, ‘‘Towards complexity analysis
of user authorization query problem in RBAC,’’ Comput. Secur., vol. 48,
pp. 116–130, Feb. 2015.
[32] G. J. Chang, ‘‘The domatic number problem,’’ Discrete Math., vol. 125,
nos. 1–3, pp. 115–122, Feb. 1994.
ZHENG WANG received the M.S. degree from
the Department of Computer Science and Engineering, Zhejiang Normal University, in 2017.
He is currently a Teaching Assistant with Zhejiang
Radio & Television University Haiyan College.
His current research interests include optimizing
intelligent algorithms and network security.
-----
YANAN JIN received the Ph.D. degree in computer application technology from the Huazhong
University of Science and Technology, in 2011.
In 2012, he was a Visiting Researcher with Concordia University, Montréal, Canada. He is currently an Associate Professor with the School
of Information Management and Statistics, Hubei
University of Economics, Wuhan, China. His
major research interests include web data mining
and recommend systems.
SHASHA YANG received the M.S. degree from
the Department of Computer Science and Engineering, Zhejiang Normal University, in 2020.
She is currently a Teaching Assistant with the
Xingzhi College, Zhejiang Normal University. Her
research interests include mobile crowdsensing,
incentive mechanism, and game theory.
JIANMIN HAN received the B.S. degree from the
Department of Computer Science and Technology,
Daqing Petroleum Institute, in 1992, and the Ph.D.
degree from the Department of Computer Science
and Technology, East China University of Science
and Technology, in 2009. He is currently a Professor with the Department of Computer Science
and Engineering, Zhejiang Normal University. His
research interests include privacy preservation and
game theory.
JIANFENG LU received the Ph.D. degree in computer application technology from the Huazhong
University of Science and Technology, in 2010.
In 2013, he was a Visiting Researcher with the
University of Pittsburgh, Pittsburgh, USA. He is
currently a Professor with the Department of Computer Science and Engineering, Zhejiang Normal
University. His research interests include algorithmic game theory and incentive mechanism with
applications to mobile crowdsensing and federated
learning.
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.1109/ACCESS.2021.3072635?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/ACCESS.2021.3072635, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "GOLD",
"url": "https://ieeexplore.ieee.org/ielx7/6287639/9312710/09400839.pdf"
}
| 2,021
|
[
"JournalArticle"
] | true
| null |
[
{
"paperId": "4f198d0d756cfeb0a96bb4c3610e95a7499ecd69",
"title": "Mobility Based Trust Evaluation for Heterogeneous Electric Vehicles Network in Smart Cities"
},
{
"paperId": "ed6d17b572f55922b0e65754df7793f14e495c0a",
"title": "Movement-Based Solutions to Energy Limitation in Wireless Sensor Networks: State of the Art and Future Trends"
},
{
"paperId": "b5f28dabd7c71f0061054d9083b840d9a2820969",
"title": "EIHDP: Edge-Intelligent Hierarchical Dynamic Pricing Based on Cloud-Edge-Client Collaboration for IoT Systems"
},
{
"paperId": "e60be5a82e8e08dd22f426cdd65e58462b76154c",
"title": "Quantum-enhanced multiobjective large-scale optimization via parallelism"
},
{
"paperId": "7d5b94f9d2cee5823757352155d0234454f09a0e",
"title": "More Than Privacy: Applying Differential Privacy in Key Areas of Artificial Intelligence"
},
{
"paperId": "4739b897bb6508046ec6c2d55699da941e1fafd9",
"title": "Security Facets of Cyber Physical System"
},
{
"paperId": "a8bab484dd26027cdc85cf3c06818f4bab52277d",
"title": "Bidirectional Prediction-Based Underwater Data Collection Protocol for End-Edge-Cloud Orchestrated System"
},
{
"paperId": "24220139c25d5e1191e66ed148382dfdcda1486e",
"title": "Software Vulnerability Detection Using Deep Neural Networks: A Survey"
},
{
"paperId": "a21ab33bb33634ff97dc4466a07afbbcc3e53e8c",
"title": "Automatic Extraction of Access Control Policies from Natural Language Documents"
},
{
"paperId": "9146c9e3dce71db6813cd35a93b7c366fcc69b6a",
"title": "A Multi-granularity Genetic Algorithm"
},
{
"paperId": "6cc5587c8b732a933dfd4a37fa78df8cec6c5652",
"title": "A Johnson's-Rule-Based Genetic Algorithm for Two-Stage-Task Scheduling Problem in Data-Centers of Cloud Computing"
},
{
"paperId": "1c3fc4d76d2034c798d770225b501c1be81029d2",
"title": "Generation of Static YARA-Signatures Using Genetic Algorithm"
},
{
"paperId": "6def59617ec2969f69bccee0e794ff0e63a67901",
"title": "Permission-Based Separation of Duty in Dynamic Role-Based Access Control Model"
},
{
"paperId": "e3e3b01df76c05424aa1e7c0ce6b7049e9eafc2d",
"title": "A Partition-Driven Integrated Security Architecture for Cyberphysical Systems"
},
{
"paperId": "ecabe700cb9b6d48990f8aafa5d1384a9d76bc24",
"title": "The Internet of Things (IoT): Informatics methods for IoT-enabled health care"
},
{
"paperId": "843c1f0d1721f513cf91054a9321978536d5bd5b",
"title": "Android HIV: A Study of Repackaging Malware for Evading Machine-Learning Detection"
},
{
"paperId": "df333a3bcdc4b278a8dd089e6731aa3455f5cf74",
"title": "A Genetic Algorithm Based Data Replica Placement Strategy for Scientific Applications in Clouds"
},
{
"paperId": "45648b4b01cfa81cd3206feaee5b75a39e9cb50d",
"title": "Game-Theoretic Design of Optimal Two-Sided Rating Protocols for Service Exchange Dilemma in Crowdsourcing"
},
{
"paperId": "602cd4722141a97046c691d52308f40cfd7927d2",
"title": "Towards an Efficient Approximate Solution for the Weighted User Authorization Query Problem"
},
{
"paperId": "12f3cbdd93ef47b36594fa567106c5bc0c6ff0f7",
"title": "Designing Socially-Optimal Rating Protocols for Crowdsourcing Contest Dilemma"
},
{
"paperId": "248516d294bf7ee4d83816318ab42a473f2408b6",
"title": "Security and Privacy in Cyber-Physical Systems: A Survey of Surveys"
},
{
"paperId": "d4498c69b91aa4cdfc206abe811db862347fb435",
"title": "A Rigorous Framework for Specification, Analysis and Enforcement of Access Control Policies"
},
{
"paperId": "eff72911641db59b7d299f1c04bff7cf5e6012c5",
"title": "Next-Generation Access Control for Distributed Control Systems"
},
{
"paperId": "508d9edd7c7004fd291ff136ba562c4d3d876958",
"title": "Resiliency Policies in Access Control Revisited"
},
{
"paperId": "edd434f7a4f3e40f734fb8623b34fe4822f35554",
"title": "Towards complexity analysis of User Authorization Query problem in RBAC"
},
{
"paperId": "8fb08b72bba6388b400bafe21dee67600e3f1646",
"title": "Mining Attribute-Based Access Control Policies"
},
{
"paperId": "8f2dfd370068ec5fe03a4aa2d0325bad13c1fbda",
"title": "Specifying and enforcing the principle of least privilege in role‐based access control"
},
{
"paperId": "d6a5c19f70508d808c29cdeb1f7db09d5c111147",
"title": "Role mining based on weights"
},
{
"paperId": "a0e672ba3414c80deefe042e200222fc49c5f1ce",
"title": "Resiliency policies in access control"
},
{
"paperId": "5b33cffc2ee476f7db9b06cd85dbb0ac22da03f5",
"title": "The domatic number problem"
},
{
"paperId": "f97356ffef4cab0adc41e57f7c5b8df53ba481db",
"title": "A Comparison of Commercial and Military Computer Security Policies"
},
{
"paperId": "d5002dcbc414472e38dedb3329b149d3d7d857eb",
"title": "On protection in operating systems"
},
{
"paperId": null,
"title": "He is currently a Teaching Assistant with Zhejiang Radio & Television University Haiyan College. His current research interests include optimizing intelligent algorithms and network security"
},
{
"paperId": "f97561037d53527f6d254110866cd93fa057f192",
"title": "A Comparison of Commercial and Military Computer Security Policies"
},
{
"paperId": null,
"title": "vi Stop"
},
{
"paperId": null,
"title": "step v Crossover: For each y j in the pair points x i and x ( i + 1) from the x m / 2 , . . . , x m , with probability p c , exchange x i . y j with x ( i + 1)"
},
{
"paperId": null,
"title": "step C. FINDING MUTUALLY DISJOINT USER GROUPS This section finds whether there are κ groups of mutually disjoint user groups in the"
},
{
"paperId": null,
"title": "step iv Mutate: For each point x i that m / 2 < i ≤ m in the population and for each bit in x i , with probability p m , alter its value"
},
{
"paperId": null,
"title": "step ii Compute fitness: Compute the fitness and relative fitness of the role set using the evaluation function respectively"
}
] | 14,642
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/023e83b3dce4a29c18934916f08cc3d2d40d535a
|
[
"Computer Science"
] | 0.92317
|
New Directions in Social Authentication
|
023e83b3dce4a29c18934916f08cc3d2d40d535a
|
[
{
"authorId": "3183839",
"name": "Sakshi Jain"
},
{
"authorId": "144516687",
"name": "N. Gong"
},
{
"authorId": "31289371",
"name": "Sreya Basuroy"
},
{
"authorId": "2573527",
"name": "Juan Lang"
},
{
"authorId": "143711382",
"name": "D. Song"
},
{
"authorId": "143615345",
"name": "Prateek Mittal"
}
] |
{
"alternate_issns": null,
"alternate_names": null,
"alternate_urls": null,
"id": null,
"issn": null,
"name": null,
"type": null,
"url": null
}
| null |
# New Directions in Social Authentication
## Sakshi Jain
sjain2@linkedin.com
## Juan Lang
juanlang@google.com
## Neil Zhenqiang Gong
neilz.gong@berkeley.edu
## Dawn Song
dawnsong@cs.berkeley.edu
## Sreya Basuroy
basuroy@princeton.edu
## Prateek Mittal
pmittal@princeton.edu
**_Abstract—Web services are increasingly adopting auxiliary_**
**authentication mechanisms to supplement the security provided**
**by conventional password verification. In the domain of social**
**network based web-services, Facebook has pioneered the use of**
**_social authentication as an auxiliary authentication mechanism._**
**If Facebook detects a user login under suspicious circumstances,**
**then users are asked to verify information about their friends (in**
**addition to verifying their passwords). However, recent work has**
**shown that Facebook’s social authentication is insecure.**
**In this work-in-progress, we propose to rethink the design**
**of social authentication. Our key insight is that online social**
**network (OSN) operators are privy to large amounts of private**
**data generated by users, including information about users’**
**online interactions. Based on this insight, we architect a system**
**for social authentication that asks users to verify information**
**about their social contacts and their interactions. Our system**
**leverages information protected by privacy policies of OSNs to**
**resist attacks, such as questions based on private user interactions**
**including exchanging messages and poking social contacts.**
**We implemented our system prototype as a Facebook ap-**
**plication, and performed a preliminary user study to evaluate**
**feasibility of the approach. Our initial experiments have been**
**encouraging; we find that users have high rates of recall for**
**information generated in the context of OSN interactions. Overall,**
**our work provides a promising new direction for the secure and**
**usable deployment of social authentication.**
I. INTRODUCTION
Web services today such as Facebook rely on user provided
passwords for authentication. However, a critical security issue
in this paradigm is the compromise of passwords [1]. For
example, passwords could be compromised because of password database leakage, phishing attacks, dictionary attacks,
or password reuses across multiple websites. To supplement
the security provided by conventional passwords, websites are
increasingly deploying auxiliary authentication mechanisms.
Auxiliary authentication aims to prevent attackers from taking
over user accounts despite having access to their correct
passwords.
Permission to freely reproduce all or part of this paper for noncommercial
purposes is granted provided that copies bear this notice and the full citation
on the first page. Reproduction for commercial purposes is strictly prohibited
without the prior written consent of the Internet Society, the first-named author
(for reproduction of an entire paper only), and the author’s employer if the
paper was prepared within the scope of employment.
USEC ’15, 8 February 2015, San Diego, CA, USA
Copyright 2015 Internet Society, ISBN 1-891562-40-1
http://dx.doi.org/10.14722/usec.2015.23002
In the domain of social network based web services,
Facebook has pioneered the use of social authentication as
an auxiliary authentication mechanism. Facebook monitors
user accounts for suspicious activity. For instance, if a user
logs into Facebook from very distant locations within a very
short span of time, then in addition to requiring the user
password, Facebook verifies the user by presenting a friend
photo and challenging the user to name the friend [2]. Indeed,
Facebook’s approach has been inspired by similar proposals from the academic community [3]. Interestingly, most
deployed and proposed systems have primarily focused on
the paradigm of users identifying their friends in depicted
photos. A critical vulnerability in this paradigm is the use
of fast improving face recognition algorithms. In fact, recent
works have demonstrated the successful attacks on photo-based
social authentication through theoretical modeling as well as
empirical evaluation [4], [5]. Thus, an open question facing
_our community is whether social authentication in the current_
_form can provide a strong foundation for supplementing the_
_security of password based authentication._
**Our work:** We propose to rethink the design of social
authentication based on the insight that online social network (OSN) operators are privy to large amounts of private
data generated by users. We believe that the space of social
knowledge is much larger than photographs of friends. For
instance, users in online social networks are associated with
rich node attributes such as users’ schools, employments,
faces, and locations. Moreover, users interact with each other
in online social networks. Such interactions include poking
friends and exchanging private messages with friends. In this
work-in-progress, we aim to study how to leverage the rich
space of social knowledge to design mechanisms for social
authentication that are both secure and usable. Towards this
end, we introduce a general architecture and a system for
social authentication that is is able to incorporate the social
knowledge available to OSN operators. Our system challenges
users to verify information that is dynamically generated in
the context of OSN usage, such as information about users’
social contacts and their interactions. Note that our approach
does not rely on users to preselect static “security questions”
and can thus be leveraged on demand.
We propose to group the challenges that can be generated
using social knowledge into three categories: node, pseudo_edge, and edge questions. They are constructed from node_
attributes specific to a single user, common node attributes
of linked users (friends), and attributes of user interactions,
respectively. Under this categorization of social knowledge,
-----
Facebook’s photo-based authentication mechanism is an example of a node question since faces are users’ node attributes.
Moreover, questions based on private user interactions such as
exchanging private messages are examples of edge questions.
To resist attacks against social authentication, our approach
relies on privacy policies applicable on user data that are
enforced by OSN operators.
One of the key challenges in generalizing the concept of
social authentication is usability, i.e., are users able to recall
information that is organically and dynamically generated with
their OSN usage? To study this question, we implemented
a preliminary prototype of our architecture as a Facebook
application. We performed a user study by recruiting 90 participants from Amazon Mechanical Turk to test our prototype.
Our initial results have been encouraging; our study provides
preliminary support to the idea that users have a non-trivial
ability to recall information pertaining to their interactions on
online social networks.
As a part of future work, we plan to (a) conduct a largerscale user study to further our understanding of the usability
of social authentication, (b) develop theoretical models to
quantify the security of the approach, and (c) engage with
OSN operators to impact system design. Overall, our work
opens up promising new directions for research in secure and
usable social authentication mechanisms.
II. MOTIVATION
Facebook designed and implemented an auxiliary authentication mechanism called social authentication [2] for its users
using photos of friends posted on the social network. When
Facebook detects suspicious activity on a user’s account, e.g.,
if a user logged into Facebook from very distant locations
within a small span of time, in addition to the user’s password, it presents photo challenges to the user. In these photo
challenges, Facebook shows 3 tagged photos of a friend with
6 options and the user has to select the correct friend name
that corresponds to the tags in the photos shown. If the user
accurately answers at least 5 out of 7 instances of photo
challenges, he or she is allowed access to the website.
However, recent works [4], [5] have discussed various
security issues with photo-based social authentication. For
instance, Kim et al. [4] pointed out that photo-based social authentication is not secure against the user’s friends who could
also recognize the person in the photo. Polakis [5] designed an
automated attack which exploits face recognition techniques,
to demonstrate the feasibility of carrying out large-scale realworld attack against photo-based social authentication. As a
defense, Polakis et al. [6] recently proposed to transform faces
and show distorted faces in the photos. They showed that these
distorted friend faces, while easy for a user to recognize, are
robust against face recognition attacks and image comparison
attacks where attackers collect publicly available photos to
compare and identify the individuals in displayed photos. In
conclusion, photo-based social authentication constantly finds
itself in arms race with face recognition algorithms, which
are fast improving. In this work, we ask the question, can we
leverage information from a user’s social network other than
the photos?
Fig. 1: Proposed architecture for social authentication systems
Indeed, the space of social knowledge is much larger than
just photos. For instance, users in OSNs usually create profiles
which include diverse information types such as education,
age, employment, and location. Moreover, OSNs offer various
modes of interaction amongst users, for example, users could
poke their friends and exchange private messages on Facebook,
Twitter allows a user to follow another user, Google+ allows
its users to create circles and categorize their connections, and
LinkedIn allows users to write recommendations and endorse
their social contacts for some skills. Can these social data
be leveraged to design social authentication? How difficult or
easy it is to generate challenges based on these data? How
secure and usable would such systems be? Would it be more
secure than photo-based social authentication? Would it have
implications on users’ privacy? Can we categorize the plethora
of information available in social networks in some way in
order to perform a security analysis of them?
We believe that photo-based social authentication is one
aspect of knowledge based social authentication mechanisms
and there lies a large space of social knowledge yet unexplored.
In this work-in-progress work, we lay the basic framework of
exploring the use of other social knowledge and take the first
step towards answering some of the questions asked.
III. ARCHITECTURE OF SOCIAL AUTHENTICATIONS
We denote an OSN as a graph G = (V, E), where each
node corresponds to a user registered on that OSN and an edge
corresponds to two users being friends on the social network.
OSNs store various types of personal information about users
themselves as well as their activities on the website. We divide
these information types into two categories, i.e., node attributes
and edge attributes. Node attributes correspond to details specific to each user independent of their interaction with others.
Some common node attributes across social networks include
user’s name, photo, education, and location. Edge attributes
on the other hand include data corresponding to interactions
among various users. The schema of this information type
largely depends on the various platforms provided by the social
network for user engagement. Some examples of such data
include messages exchanged between users, pokes by friends,
and posts written on a friend’s wall.
**Architecture Overview: A social authentication system com-**
prises of challenges or questions posed to the user. We propose
a schematic architecture for a social authentication system as
follows. The system iterates over k trials to authenticate a
-----
user u. In each trial, a question is selected from the question
database and is displayed to the user via an authentication
_interface. All questions follow a common schema, where the_
user is provided information about an attribute, node or edge,
and is asked to identify the associated friend. The user u inputs
his/her answer (i.e., name of a friend) to the question; and the
_answer matching module checks if the user provided answer_
can be matched to the correct friend.
**Question Database:** The questions in the database are
generated using the node and edge attributes available for the
specific social network. We divide the set of questions into
three main categories.
_Node questions: Questions where the user is provided data_
about some node attribute of a friend and is asked to recognize
the corresponding friend. For instance, “Name your friend in
the displayed photos” or “Name a friend who is currently
studying at UC Berkeley”.
_Pseudo-edge questions: Questions where the user is pro-_
vided information about some node attribute which is common
between the user and a friend. The user is then asked to recognize the friend. For instance, “Who went to the same school
with you?” is a pseudo-edge question because it involves the
school (node attribute) common to the user and his/her friend.
_Edge questions: Questions where the user is provided_
information about some interaction with a friend and the user
is asked to recognize the friend. For instance, “Name a friend
you recently exchanged a message with” is an edge question.
Facebook’s face-recognition challenges fall under node
questions category since faces are node attributes.
**Authentication Interface & Answer Matching:** The authentication interface displays the challenges and receives the
user’s inputs. There could be multiple ways of obtaining
answers from the user, each providing varied usability and
security trade-offs. For example, one way is to show n options
of friend names as radio buttons and the user chooses the
correct one amongst them. Facebook’s current photo-based
social authentication system receives the answers in this way,
where n = 6. Another way is to ask the user to type in the
name of the correct friend by providing just the photos of both
correct and incorrect friends as options. The user in this case
needs to recognize the correct friend from the photos and write
the selected friend’s name in the textbox. The name entered by
the user in this case can be matched to the correct one using
fuzzy matching, to account for spelling mistakes for improved
usability. One can also imagine providing a dropdown menu
of friends’ names to select from, with or without providing
any photo options. Each of the above techniques have their
pros and cons when evaluated against security and usability
metrics. We suspect that the first method is very usable since
it allows the user to click on an option, however, the security of
such method is lower bounded by _n[1]_ [. Although we compromise]
on usability for the second method, its security is strictly
better than providing radio buttons, since the attacker would
have to recognize the correct friend and type in the name.
Quantitatively evaluating the security is however quite tricky
in this case.
**Model Selection and Evaluation: Given the proposed general**
model for a social authentication system, there are multiple
Fig. 2: Example of an edge question from our prototype for
Facebook.
parameters that need analysis. For example, how difficult is it
to come up with the question database for a particular social
network? Is such a model feasible? Would users remember
answers to such questions? How should the answer choices
look like? Do any particular category of questions provide
better security or usability to users? In order to answer some
of these questions and to test the feasibility of such a system,
we build a prototype authentication interface for Facebook
and perform a user study to perform preliminary analysis of
the proposed system. We particularly chose Facebook as our
platform since it is the most popular online social network
(OSN) with more than 1 billion users worldwide [7]. Also
Facebook provides an API to build apps using information
from a user’s social graph.
In the following two sections, we detail our analysis of
the feasibility and usability of the proposed system. We also
briefly discuss the security implications of the various types
of questions in Section V.
IV. USER STUDY DESIGN
_A. Preliminary Study_
We designed a user study to understand the usability of our
new proposed model, to measure how well users perform when
posed with questions about their social network and to help
design a more extensive authentication mechanism model. To
this effect, we recruited 90 participants to take a survey and
performed a quantitative study based on the observations.
_1) Methodology: We invited participants via Amazon Me-_
chanical Turk to take a survey about their Facebook account.
Any participant above 18 years of age owning a Facebook
account was allowed to take the survey. Each participant is
directed to a Facebook application URL and asked to login
with his Facebook credentials. Once logged in, Facebook
takes the participant to our application, called ’Soc-auth’.
Soc-auth requests the following permissions to the user before proceeding: user-groups, user-photos, friends-about-me,
_{_
friends-education-history, friends-photos, read-mailbox . Once
_}_
the participant provides the required permissions, Soc-auth
poses the participant with 4 different questions followed by a
survey about basic personal information and a feedback form.
For each question, client-side Javascript queries Facebook for
-----
TABLE I: Questions used in the Facebook prototype for user study and their corresponding categories
|Question schemas|Description|Category|
|---|---|---|
|Q 1|Type in the complete name of the person with a square box around his/her face in the following picture|Node|
|Q 2|Given the following 5 Facebook friends as options, type the complete name of the friend you went to same school with|Pseudo-edge|
|Q 3|Given the following 5 Facebook friends as options, type the complete name of the friend who poked you on Facebook|Edge|
|Q 4|Given the following 5 Facebook friends as options, type the complete name of the friend with whom you exchanged a message on Facebook|Edge|
appropriate user information and checks the correctness of
the answer provided by the user. We chose to implement
all the logic at the client side to protect the confidentiality
of user information since the above mentioned permissions
provide the app access to sensitive data including inbox.
To protect the privacy of the user, we only store whether
the user answered a question correctly. Each participant was
compensated with $5 paid via Amazon Mechanical Turk. We
recruited 90 participants in total from Amazon Mechanical
Turk over a course of 7 days. These participants had a wide
range of ages (18 - 45+). 42% of the participants fell in the
(18-24) bracket, 39% in the (25-34) bracket, and the remaining
18 % were above 35. We also saw a wide range of educational
background. About 19% had or are pursuing high school
degrees, 57% had or are pursuing bachelor degrees, and 24%
had or are pursuing advanced degrees.
Our goal of this experiment is to understand the feasibility
of a model which uses the user’s social network to generate
authentication questions. To this effect, we chose 4 different
questions to ask each user. Questions were selected based on
most popular sources of activity on Facebook and security
of the question. We first inspected the Facebook Graph API[1]
which is a tool provided by Facebook to represent the nodes
and edges of its social graph. By analyzing a node or user, we
determined the most common interactions or edges they share
with other nodes and designed the questions to ask about these
attributes. Furthermore, according to a survey about people’s
Facebook activity conducted by the Pew Research Center [8],
the top 3 most frequent activity are commenting, liking,
and exchanging messages. While users may post statuses or
comment on friends’ posts frequently, this behavior is easily
viewable by both known and unknown attackers and does not
constitute a secure question. Hence, we ask questions about the
next most frequent set of activities that are not public, such as
private messages and pokes exchanged.
The questions and their corresponding categories are shown
in Table I. Question Q1 presents a user with a photo from his
album and asks the user to type in the name of the tagged
person. This is a node question since answering this question
correctly would require the user to recognize a friend’s face
(a node attribute) correctly. Question Q2 presents a user with
profile photo of five of his friends and asks the user to type in
the name of the friend with whom he went to the same school.
This is a pseudo edge question since the question requires the
knowledge about the node attributes (i.e., school) of both the
user and the correct friend. Questions Q3 and Q4 are edge
questions, each of which presents a user with five options and
1https://developers.facebook.com/docs/graph-api/
TABLE II: 95% confidence intervals of applicability and
reliability of the four question schemas shown in Table I.
Applicability Reliability
_Q1_ 77%±8% 28%±9%
_Q2_ 51%±10% 54%±10%
_Q3_ 48%±10% 71%±9%
_Q4_ 98%±2% 66%±10%
asks the user to type in the correct name. Specifically, Q3
asks the user to choose the friend who poked him recently on
Facebook and Q4 asks who recently exchanged a message with
the user on Facebook. Questions Q3 and Q4 are asked only
when the user has at least one friend who poked/ messaged
him in last one year. This design choice is made to ensure that
the interaction is recent enough for the user to remember the
friend. Figure 2 shows an example of Q4.
To generate options for each question, we randomly choose
one correct option and 4 incorrect options. Note that the user
is not just asked to select the correct friend but to type in the
name of the friend in a text box, thereby increasing security.
To match the answer provided by the user with the correct
friend’s name displayed on Facebook, we adopt DamerauLevenstein edit distance for fuzzy matching. The input answer
is considered correct if the edit distance is no more than 12%,
which roughly means that we tolerate one out of 8 characters
to be removed or replaced or added.
_2) Findings: In order to capture the feasibility of our_
model, we evaluate it using two metrics, applicability and
_reliability. Notice that some or all the four questions might be_
inapplicable to some users. For instance, Q3 is inapplicable
to a user who has not been poked by any friend and Q2 is
not applicable to a user who has not mentioned his school
on Facebook. To quantify this, we define applicability of
each question Qi as the fraction of users to which Qi was
applicable. In order to measure how easy it is for a user to
answer the questions, we define reliability of each question Qi
as the fraction of users for whom this question was applicable
and who correctly answered the question. We use well known
Wilson method to compute 95% confidence interval for both
applicability and reliability of the four questions.
Table II shows the 95% confidence intervals of applicability
and reliability of the four questions obtained from our user
study of 90 participants. We find that the variation in the
applicability of the questions we chose is quite large. Only
about 51% of the participants had a page associated with their
school on Facebook. While about 52% had not been poked
|Col1|Applicability|Reliability|
|---|---|---|
|Q 1|77%±8%|28%±9%|
|Q 2|51%±10%|54%±10%|
|Q 3|48%±10%|71%±9%|
|Q 4|98%±2%|66%±10%|
-----
in last one year, around 98% had exchanged a message with
a friend during the time span of a year. The photo question
has a 77% of applicability, since the photos selected were
chosen from the user’s albums, instead of any and all images
of the individual. While a friend may have many images on
Facebook, not all will have albums.
Similarly large variation is seen in the reliability of our four
questions. We find that the users were able to correctly answer
the two edge questions more easily than the node question,
which fared quite low on reliability ( 28%). We believe that
_∼_
this gap is because an interaction with a friend in the form
of a message or a poke would make it more likely that the
friend is a close friend implying it would be easier for the user
to remember his/her name. On the other hand, a user might
not be familiar with friends or acquaintances (but friends on
Facebook) tagged in some photos,[2] resulting in low reliability
of Q1. Note that from these observations, we cannot firmly
deduce that edge questions are more reliable than node or
pseudo-edge questions since we have used specific examples
for each category of question. It is possible that some instances
of node questions perform better than a poorly chosen instance
of edge question. However, since there is no universal set
of edge, node, and pseudo-node questions, this is difficult to
evaluate at this point.
_B. Next Steps_
Based on the observations from the first study, we are
designing a more extensive and larger scale study to quantitatively evaluate the benefits of the proposed model as a part
of future work. Since the previous study only asked 1 node,
1 pseudo-edge, and 2 edge questions, the results are limited
to the specific question asked within each category. Instead,
we plan to design and analyze a broader set of questions
per category. Examples of node questions other than faceidentification could include asking the user to identify a friend
from his hometown, college, employer, or Facebook groups
that he is a member of. The pseudo-edge category can be
expanded to questions like “Name a friend who attended the
same high school or college as you.”, or “Name a friend who is
going to a given Facebook event with you.” Similarly, the edge
questions can be expanded to more than exchanging pokes
and messages. For example, users can be asked to identify a
friend who sent them a friend request or tagged them in a
photo recently. Each question may have a different memory
recall time and applicability based on the user’s engagement
of Facebook; it would be interesting to examine whether one
particular type of questions are more usable.
Furthermore, we want to quantify the usability and security of the existing face-based authentication model used
by Facebook and compare with our model. The photo test
question in the previous study was similar to the one used
by Facebook, except for the number of images of the friend
displayed in the question and the answering matching mode.
Thus to create a more direct comparison, we plan to design
a separate question to simulate the photo-based challenge as
shown by Facebook. Finally, we’d like to evaluate the ease
of use of various answering methodologies while maintaining
their security properties. We plan to compare the radio button
2These tags could be provided by other users.
option, vanilla text box with no options, and text box with
photos of friends without their names shown as options. Moreover, we plan to construct a formal security model to quantify
the security of different categories of questions and different
answering matching modes, and compare them quantitatively.
V. DISCUSSION
In this section, we briefly discuss the security and privacy
implications of the proposed model.
**Security:** Online social networks often provide users with
fine-grained privacy settings. We assume a user u sets his/her
node attributes (e.g., users’ faces, schools, and employers) to
be accessible to at least his/her friends. The incentives for
users to do so could be to let their friends know who they
are. In fact, Dey et al. [9] showed that 47% of Facebook users
leave their such node attributes publicly accessible by default.
However, we consider edge attributes (e.g., pokes and private
messages exchanged between two users) of an edge (u, f ) are
set to be accessible only to user u and the linked friend f .
Indeed, such edge attributes in Facebook are only accessible
to the two involved users.
Under this privacy setting, the set of users who can access
the attributes that are core to the three types of questions
(i.e., node, pseudo-edge, and edge questions) are different.
Specifically, let u be the user and f be the selected friend about
whom a question q is being asked. If q is a node question, the
node attribute used in q is at least accessible to all the friends of
_f and f_ . If on the other hand q is a pseudo-edge question, the
common node attribute involved in q is only accessible to the
common friends of u and f if they set their node attributes to
be only visible to their friends in their privacy settings. Lastly,
if q is an edge question, the corresponding edge attribute is
accessible only to u and f . The different privacy setting for
node and edge attributes is the fundamental reason why the
three types of questions manifest different levels of security.
We will take the Sybil attack [10] as an example to
further illustrate the security levels. In an Sybil attack, the
attacker creates fake accounts on the social network and tries
to befriend the victim and its friends to get access to their
information. If the authentication challenge is a node question
like the Facebook’s photo based challenge, the attacker has all
the necessary information to solve the challenge once he has
connected himself to the victim’s friends on the social graph.
If the authentication challenge is a pseudo-edge question, the
attacker needs to befriend the victim’s friends and the victim,
which succeeds with a lower probability. Edge questions are
robust to this kind of Sybil attack because interactions are
private to the victim and the friend involved.
We believe edge questions can be significantly more
promising in providing security and worth exploring in the
new versions of social authentication services. Theoretical
modeling of the three types of questions and performing
security experiments on publicly available social graphs is left
for future work.
**Privacy implications:** Social authentication mechanisms
might also raise concerns around leakage of private user
information. For each of the three types of questions, some
information about the node or edge attributes is revealed to be
-----
able to frame the challenge. An example from our prototype
is the message question; the attacker without answering the
question would know that the user exchanged private messages
with one of the friends from the options. Similarly, in the
Facebook’s photo-based questions, user’s friends and their
photos are revealed during the challenge. One can argue
against the privacy leakage since these challenges are only used
when the user has been confirmed via primary authentication
interface (passwords). Moreover, we plan to evaluate users’
privacy concerns in social authentication via user studies.
VI. RELATED WORK
In this section, we review prior work on social authentication mechanisms, which we divide into two categories:
_trustee-based social authentication and knowledge-based social_
authentication.
In trustee-based social authentication [11], [12], [13], [14],
[15], the user or the service provider pre-selects a few friends
of the user as trustees, who aid the user in the authentication
process. Knowledge-based social authentication [3], [2], [4],
[5], [6] utilizes a user’s friends’ information for authentication,
and thus knowledge-based social authentication relies on the
user’s knowledge about their friends. The friends are not
directly involved in knowledge-based social authentication.
Knowledge-based social authentication mechanisms are mainly
used as auxiliary authentication mechanisms while trusteebased social authentication mechanisms are used as backup
authentication service. Our work belongs to knowledge-based
social authentication.
**Trustee-based social authentication:** Brainard et al. [11]
proposed to use somebody you know, i.e., friends of users, in
authentication systems. Originally, Brainard et al. combined
trustee-based social authentications with other authenticators
(e.g., passwords) as a two-factor authentication mechanism.
Later, trustee-based social authentication was adapted to be
used as a backup authenticator [13], [14], [12]. For instance,
Schechter et al. [12] designed and built a prototype of trusteebased social authentication system which was integrated into
Microsoft’s Windows Live ID system. Facebook announced
its trustee-based social authentication system called Trusted
Friends in October 2011 [13], and it was redesigned to be
Trusted Contacts [14] in May 2013. Gong and Wang [15]
proposed a probabilistic security model to quantify the security
of trustee-based social authentication, and their security model
can guide the design of more secure trustee-based social
authentication.
**Knowledge-based social authentication: Yardi et al. [3] were**
the first to propose a photo-based authentication system called
_Lineup, to test if the user belongs to a group (e.g., interest_
groups in Facebook) that he/she tries to access. Specifically,
when a user tries to access a group, Lineup presents a photo
and asks the user to input the names of subjects in the photo
assuming that if the user has the permission to access the
group, he/she should know the subjects. To determine if the
answer given by the user is correct or not, Lineup uses tagged
photos to obtain ground-truth answers. Furthermore, Yardi et
al. discussed Denial of Service (DoS) and network outlier
attacks. In DoS attacks, an attacker could spam the system with
a large number of photos with wrong tags, and thus legitimate
users input “incorrect” answers even if they know the subjects.
The network outlier attacks represent that an attacker can
recognize his/her friends that are in the group and whose
tagged photos are presented. Later, Facebook adopted and
implemented this photo-based authentication mechanism [2]
to verify users when a suspicious user activity is detected.
VII. CONCLUSION AND FUTURE WORK
In this work, we propose to revisit the design space of
social authentication challenges by exploiting the vast amount
of data generated on social networks. Specifically, we present a
general architecture for social authentication that incorporates
a large space of social knowledge and makes it possible to
compare different design strategies under the same framework.
We introduce a categorization of the design space of questions
that can be generated from a social graph, i.e., node, pseudo_edge, and edge questions._
As a proof-of-concept for our proposed model, we implement a prototype as a Facebook application and perform user
study on 90 Amazon Mechanical Turk workers. The results
of the study are encouraging and prove the feasibility and
usability of such a model. Our work thus opens up promising
new directions in knowledge-based social authentication by
exploiting a larger design space.
**Acknowledgement: This work is supported by the NSF under**
Grant No. 1409915 and 1409415.
REFERENCES
[1] D. Balfanz, R. Chow, O. Eisen, M. Jakobsson, S. Kirsch, S. Matsumoto,
J. Molina, and P. van Oorschot, “The future of authentication,” IEEE
_Security & Privacy, 2012._
[2] Facebook’s Knowledge-based Social Authentication.,
“http://blog.facebook.com/blog.php?post=486790652130.”
[3] S. Yardi, N. Feamster, and A. Bruckman, “Photo-based authentication
using social networks,” in WOSN, 2008.
[4] H. Kim, J. Tang, and R. Anderson, “Social authentication: Harder than
it looks,” in FC, 2012.
[5] I. Polakis, M. Lancini, G. Kontaxis, F. Maggi, S. Ioannidis, A. D.
Keromytis, and S. Zanero, “All your face are belong to us: Breaking
facebook’s social authentication,” in ACSAC, 2012.
[6] I. Polakis, P. Ilia, F. Maggi, M. Lancini, G. Kontaxis, S. Zanero,
S. Ioannidis, and A. D. Keromytis, “Faces in the distorting mirror:
Revisiting photo-based social authentication,” in CCS, 2014.
[7] Facebook Company Info, “http://newsroom.fb.com/company-info/.”
[8] Keith Hampton and Lauren Sessions Goulet and Cameron
Marlow and Lee Rainie, “http://www.pewinternet.org/2012/02/03/
part-2-facebook-activity/.”
[9] R. Dey, Z. Jelveh, and K. Ross, “Facebook users have become much
more private: A large-scale study,” in SESOC, 2012.
[10] J. R. Douceur, “The Sybil attack,” in IPTPS, 2002.
[11] J. Brainard, A. Juels, R. L. Rivest, M. Szydlo, and M. Yung, “Fourthfactor authentication: Somebody you know,” in CCS, 2006.
[12] S. Schechter, S. Egelman, and R. W. Reeder, “It’s not what you know,
but who you know,” in CHI, 2009.
[13] Facebook’s Trusted Friends, “https://www.facebook.com/notes/
facebook-security/national-cybersecurity-awareness-month-updates/
10150335022240766.”
[14] Facebook’s Trusted Contacts, “https://www.facebook.com/notes/
facebook-security/introducing-trusted-contacts/10151362774980766.”
[15] N. Z. Gong and D. Wang, “On the security of trustee-based social
authentications,” IEEE Transactions on Information Forensics and Se_curity (TIFS), vol. 9, no. 8, 2014._
-----
|
{
"disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references', 'abstract'}. Paper or abstract available at https://api.unpaywall.org/v2/10.14722/USEC.2015.23002?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.14722/USEC.2015.23002, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "GREEN",
"url": "http://home.engineering.iastate.edu/%7Eneilgong/papers/socauth.pdf"
}
| 2,015
|
[] | true
| null |
[] | 8,489
|
|
en
|
[
{
"category": "Business",
"source": "s2-fos-model"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Economics",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/023fb7dd8bdf0812ffbf351657c08c8920f7d512
|
[] | 0.881474
|
Applicability of Blockchain Technology to The Normal Accounting Cycle
|
023fb7dd8bdf0812ffbf351657c08c8920f7d512
|
Applied Finance and Accounting
|
[
{
"authorId": "115576142",
"name": "W. K. Peprah"
},
{
"authorId": "2155969672",
"name": "Reynaldo P. Abas Jr."
},
{
"authorId": "108686680",
"name": "A. Ampofo"
}
] |
{
"alternate_issns": [
"2374-2429"
],
"alternate_names": [
"Appl Finance Account"
],
"alternate_urls": [
"http://redfame.com/journal/index.php/afa/issue/view/24"
],
"id": "6f9da3f2-7f87-411a-be06-4f8dd8d80439",
"issn": "2374-2410",
"name": "Applied Finance and Accounting",
"type": null,
"url": "http://redfame.com/journal/index.php/afa/index"
}
|
Blockchain technologyis a distributed, unchangeable ledger that makes recording transactions and managing assets in a business network much easier and nowa type of accountingsoftwareconcernedwith the transfer of assetownership and the maintenanceof anaccuratefinancial ledger. Despitethenumerousbenefits ofblockchaintechnology,there is no study on theapplicability of blockchain technologytothenormalaccountingcycle in emerging economies in Africa.Thus,thispaperprovidesgeneralinsightsonhowblockchaintechnologymaybeusedinthenormalaccountingcycle in West Africa.Thestudyadoptedaqualitativeresearchmethodandcontentanalysisresearchdesigntounderstand the extent to which business leaders in West Africa are aware, understand, and utilize blockchain technology in the processing of accounting transactions to the preparation of financial statements.Results indicatethat West African business leaders are well aware, understand and applyblockchaintechnologyapplicationsinthenormalaccountingcycle,anditprovidescostsavings,digitalidentity,andsecurity.Thestudyrecommendsfurtherinvestigationsintohowtoaddressscalabilitywhen dealingwith recurrent and large transactions.
|
Vol. 8, No. 1, August 2022
ISSN 2374-2410 E-ISSN 2374-2429
Published by Redfame Publishing
URL: http://afa.redfame.com
# Applicability of Blockchain Technology to The Normal Accounting Cycle
Williams Kwasi Peprah[1], Reynaldo P. Abas Jr.[2] & Akwasi Ampofo[3]
1Valley View University, School of Business, Oyibi, Accra, Ghana. E-mail: williams.peprah@vvu.edu.gh ORCID
Number: 0000-0002-6802-2586
[2Adventist University of the Philippines, Department of Accountancy, College of Business. E-mail: rpabasjr@aup.edu.ph](mailto:rpabasjr@aup.edu.ph)
3University of Connecticut, School of Business, Accounting Department. E-mail: aaampofo@vt.edu
Received: November 22, 2021 Accepted: December 27, 2022 Available online: February 22, 2022
doi:10.11114/afa.v8i1.5492 URL: https://doi.org/10.11114/afa.v8i1.5492
**Abstract**
Blockchain technology is a distributed, unchangeable ledger that makes recording transactions and managing assets in a
business network much easier and now a type of accounting software concerned with the transfer of asset ownership and
the maintenance of an accurate financial ledger. Despite the numerous benefits of blockchain technology, there is no
study on the applicability of blockchain technology to the normal accounting cycle in emerging economies in Africa.
Thus, this paper provides general insights on how blockchain technology may be used in the normal accounting cycle
in West Africa. The study adopted a qualitative research method and content analysis research design to understand the
extent to which business leaders in West Africa are aware, understand, and utilize blockchain technology in the processing
of accounting transactions to the preparation of financial statements. Results indicate that West African business leaders
are well aware, understand and apply blockchain technology applications in the normal accounting cycle, and it provides
cost savings, digital identity, and security. The study recommends further investigations into how to address scalability
when dealing with recurrent and large transactions.
**Keywords: Blockchain Technology, Normal Accounting Cycle, Accounting**
**1. Introduction**
Blockchain technology is a distributed, unchangeable ledger that makes recording transactions and managing assets in a
business network much easier and now a type of accounting software (Adams et al., 2018; Demirkan et al., 2020)
concerned with the transfer of asset ownership and the maintenance of an accurate financial ledger.
Accounting is primarily concerned with the communication and measurement of accounting transactions and the analysis
of such information (Dai, J., & Vasarhelyi, 2017). A large part of the profession involves determining or quantifying
property rights and obligations to allocate financial resources best. For accountants, blockchain clarifies asset ownership
and the existence of obligations, and the potential for significant efficiency gains.
Despite the numerous literature about blockchain technology and its potential advantages, no study has been conducted
yet on its potential applicability in the normal accounting cycle. Thus, this paper aims to provide general insights on how
blockchain technology may be used in the normal accounting cycle.
According to Supriadi (2020), blockchain technology can improve the accounting profession by lowering the cost of
maintaining and reconciling ledgers and ensuring perfect clarity regarding asset ownership and transaction history.
Blockchain technology has the potential to assist accountants in gaining clarity over their organizations' available
resources and liabilities while also freeing up time to focus on analysis, valuation and planning rather than bookkeeping
(Pimentel, & Boulianne, 2020). Blockchain technology will increase transaction-level accounting being performed but
not by accountants (George & Patatoukas, 2021). Instead, successful accountants analyze the accurate economic
interpretation of blockchain records, reconciling the record with financial reality and valuation.
By removing reconciliations and providing confidence over transaction history, blockchain may also enable accounting
to expand its scope, considering aspects that are currently thought too complex or unreliable to quantify, such as the worth
of a company's data (Zhang et al., 2020).
-----
Blockchain technology can be used to automate bookkeeping and reconciling tasks. This could enhance accountants'
efforts in particular areas while strengthening those focused on generating value elsewhere.
In business, particularly in accounting, the latest accounting software like SAP, Oracle, QuickBooks, Sunplus are just
among the accounting information systems commonly used by most companies to process fast, convenient, and reliable
financial transactions andto generate financial statements. Given the many and successive developments that occur in the
information technology environment and the expansion of its use and utilization in the business environment, and the
direct impact on the practice of accounting data systems need to examine the opportunity to make use of advanced
technologies in the function of banking data systems, which represents the first basis in accounting work in the various
sectors in which it operates (Alsaqa et al., 2019).
Meanwhile, aside from the aforementioned accounting information systems, the potential of using blockchain has become
one of the major trends lately, as evidenced by several research studies (Kwilinski, 2019; Dai & Vasarhelyi, 2017; Kokina
et al., 2017). Blockchain technology has received much press coverage in the last few years (Kokina et al., 2017). Much
has been said recently about Bitcoin, blockchains, and distributed ledger technologies (DLT).
Satoshi Nakamoto, the famed bitcoin developer, is credited with initiating these debates (Appelbaum & Nehmer, 2020).
Smith and Castonguay (2020) stated in their study that the potential for blockchain to be used as an accounting tool
opened the door to more far- reaching implications in the areas of financial reporting, assurance, and corporate governance,
extending the benefits beyond the internal control environment. Accounting and assurance are two professions in which
blockchain technology can make a substantial impact and fundamentally alter current paradigms.
Blockchain technology functions, such as data integrity protection, rapid sharing of pertinent information, and
programmable and automatic process controls, may assist in constructing a new accounting environment (Wei et al. 2020).
This technology could potentially be utilized to give automated assurance, enhancing the agility and precision of the
current auditing paradigm (Dai & Vasarhelyi, 2017). Meanwhile, Orcutt (2018) reports that engineers recognized that
blockchain might be used to track non-monetary assets. In 2013, Vitalik Buterin, then 19 years of age, founded Ethereum,
a cryptocurrency that would follow financial transactions and the status of computer programs known as smart contracts.
**2. Accounting Cycle**
Ballada (2021) described an accounting cycle as a collection of successive processes or procedures to carry out the
accounting process. The cycle's steps and their objectives are as follows:
1) Identifying the events that will be recorded. This tries to collect data on transactions or occurrences in general via
source documents.
2) The journal is used to record transactions. This tries to document the economic impact of transactions on the firm
in a journal, a format that allows transfer to the accounts.
3) The ledger is updated using journal entries. This procedure is intended to transmit data from the journal to the ledger
for classification.
4) Constructing a trial balance. This section contains a listing for verifying the ledger's debits and credits are equal.
5) Worksheet preparation, including correcting entries. This simplifies the process of preparing financial statements.
6) Financial statement preparation. This knowledge is beneficial to decision-makers.
7) Journal entries for adjustments are journalized and posted. This column is used to track accruals, deferrals that have
expired, estimations, and other occurrences from the worksheet.
8) Journal entries for the closing journal are journalized and posted. This results in the closure of temporary accounts
and the transfer of profit to the owner's equity.
9) Preparation of a trial balance post-closing. This validates the debits and credits following the closing entries.
10) Entries in the reversing journal are journalized and posted. This simplifies the subsequent accounting period's
documentation of certain routine transactions.
Throughout the preceding steps, definitions of words may be beneficial for a better understanding of the accounting cycle.
Journalizing is the process of chronologically recording commercial transactions in a journal in terms of debits and credits
(Ballada, 2021). This procedure initiates the process of entering transactions into the books of accounts. Meanwhile,
posting is the process of transferring commercial transactions from journals to ledgers. A ledger is a record that contains
a summary of all journal entries. In other words, journaling comes before posting.
-----
**3. Methodology**
This research was qualitative and applied content analysis techniques in looking at the applicability of blockchain
technology in the normal accounting cycle. By evaluating and coding textual material, the content analysis research design
is utilized to establish replicable and accurate findings (Drisko & Maschi, 2016; Peprah et al., 2018). By assessing texts,
such as documents, oral communication, and visuals, systematically from websites and books (Krippendorff, 2019).
**4. Results and Discussion**
As shown in Table 1 below, and based on the content analysis technique, the researcher summarizes the applicability of
blockchain technology in the normal accounting cycle and its cost savings. The blockchain technology in the normal
accounting cycle provides security to digital identity as per transaction in a way to minimize bookkeeping and
reconciliation of transactions. Financial statement presentation format may be customized and integrated into the
blockchain to give standard accounting reporting.
Table 1. Applicability of Blockchain Technology in the Normal Accounting Cycle
**Potential Benefits of Blockchain Technology** **Normal Accounting Cycle**
**Applicability of Blockchain Technology in**
**the**
**Normal Accounting Cycle**
Blockchain technology can pre-program
transactions to "self-execute" per an agreed
contract, called smart contracts. For
example, a buyer and a seller may decide on
specific criteria for their business
transactions that may be programmed
through blockchain technology. Thus,
minimizing or eliminating the efforts of the
bookkeepers. Moreover, this may also
reduce or eliminate the issuance of formal
accountable forms like invoices, statements
of accounts, which would result in money
savings.
Moreover, the transaction between sellers
and buyers may be guaranteed as authentic
as digital identities are required to use
blockchain technology. A transaction will
not be added to the blockchain unless
approved by all the members of the chain.
Further, blockchain technology assures the
security of the transactions because of the
cryptographic feature, which would be very
difficult to tamper with by unauthorized
persons.
Blockchain technology would help the
bookkeeper to save time on journalizing
recurring transactions.
Thus, they may spend more time on valueadding activities like analyzing transactions,
making special reports to management.
Therefore for those entities, which might be
employing several
bookkeepers/accountants. Blockchain
technology might be a cost-saving solution
to process bulky and recurring transactions
in a given time
Cost Savings/digital identity/security
Cost Savings
Step 1.
Identification of events to be
recorded
Step 2.
Transactions are recorded in
the journal.
-----
Cost Savings
Step 3.
Journal entries are posted to
the ledger.
Step 4.
Cost Savings
Preparation of a trial balance.
See discussions on Step 2 above.
This step may not be necessary anymore as
the transactions are done in Steps 1-3 is
already validated by the members of the
network/chain.
The main purpose of the trial balance is to
check the equality of debits and credits.
While every transaction is done
electronically and according to sequential
criteria, the trial balance may be abolished.
Therefore, blockchain technology may save
time and money again on preparing the trial
balance.
The preparation of the worksheet may be
foregone also as the trial balance in Step 4
above. However, with regard to the adjusting
entries, blockchain technology can handle
this task by following the usual process of
recording and validating through the chain.
This task may be accommodated by
blockchain technology by customizing the
formats of the financial statements required
by the financial reporting framework
applicable to an entity.
See discussions on Step 2 above.
See discussions on Step 2 above.
See discussions on Step 4 above.
Cost Savings
Cost Svaings
Cost Savings
Cost Savings
Cost Savings
Step 5.
Preparation of the worksheet,
including adjusting entries.
Step 6.
Preparation of financial
statements.
Step 7. Adjusting journal
entries are journalized and
posted.
Step 8.
Closing journal entries are
journalized and posted.
Step 9.
Preparation of a post-closing
trial balance.
Step 10. Reversing journal
Cost Savings See discussions on Step 2 above.
entries are journalized
**5. Conclusion and Recommendations**
Organizations may examine the possible benefits of adopting blockchain technology in terms of cost savings, digital
identity, and security, particularly in this age of digitization. The prevalence of cryptocurrencies and other kinds of
electronic evidence may point to the adoption of blockchain technology in the accounting industry, particularly
throughout the typical accounting cycle. The study points to the fact that blockchain technology has come to reduce
the normal accounting cycle required as per the accounting standards in the preparation of financial statement.
Nonetheless, despite the potential benefits of blockchain technology, it is important to evaluate its drawbacks.
For bookkeepers or accountants to use blockchain technology in their current and future jobs, they must understand its
technicalities. To comprehend the basic technological requirements for using blockchain technology, certain skills of
computing must be established. Another issue to overcome is blockchain technology's scalability when dealing with
recurrent and large transactions.
The data's reliability must be maintained at all times and closely monitored by technical personnel who are directly
involved in blockchain technology's upkeep. Finally, more research on the implications of blockchain technology in the
-----
normal accounting cycle, particularly in steps 4-10, is strongly recommended, given that the existing literature and
studies focus solely on the distributed ledger implications of blockchain technology.
**Reference**
Adams, R., Kewell, B., & Parry, G. (2018). Blockchain for good? Digital ledger technology and sustainable development
goals. In Handbook of sustainability and social science research (pp. 127-140). Springer, Cham.
Alsaqa, Z. H., Hussein A. I., and Mahmood, S. M. (2019). The Impact of Blockchain on Accounting Information Systems.
Journal of Information Technology Management. DOI: 10.22059/jitm.2019.74301
Appelbaum, D. & Nehmer, R. A. (2020). Auditing Cloud-Based Blockchain Accounting Systems. Journal of Information
Systems. (34), 2. DOI: 10.2308/isys-52660.
Ballada, W. and Ballada S. (2021). Basic Financial Accounting and Reporting. DomDane Publishers and Made Easy
Books.
Dai J. & Vasarhelyi, M. A. (2017). Toward Blockchain-Based Accounting and
Assurance. Journal of Information Systems, 31(3), 5–21. https://doi.org/10.2308/isys- 51804.
Dai, J., & Vasarhelyi, M. A. (2017). Toward blockchain-based accounting and assurance. Journal of Information Systems,
_31(3), 5-21._
Demirkan, S., Demirkan, I., & McKee, A. (2020). Blockchain technology in the future of business cyber security and
accounting. Journal of Management Analytics, 7(2), 189-208.
Drisko, J. W. & Maschi, T. (2016). Content Analysis. Oxford University Press.
George, K., & Patatoukas, P. N. (2021). The Blockchain Evolution and Revolution of Accounting. In Information for
_Efficient Decision Making: Big Data, Blockchain and Relevance (pp. 157-172)._
Kokina, J., Mancha, R., & Pachamanova, D. (2017). Blockchain: Emergent industry adoption and implications for
accounting. Journal of Emerging Technologies in Accounting, 14(2), 91-100.
Krippendorff, K. (2019). Content Analysis, an introduction to its Methodology. SAGE Publications, Inc.
Kwilinski, A. (2019). Implementation of blockchain technology in accounting sphere. _Academy of Accounting and_
_Financial Studies Journal, 23, 1-6._
OrCutt, M. (2018). Block chain. MIT Technology Review, 121(3), 18–23b.
Peprah, W. K., Afriyie, A. O., Abandoh-Sam, J. A., & Afriyie, E. O. (2018). Dollarization 2.0 a Cryptocurrency: Impact
on Traditional Banks and Fiat Currency. _International Journal of Academic Research in Business and Social_
_Sciences, 8(6), 341-349._
Pimentel, E., & Boulianne, E. (2020). Blockchain in Accounting Research and Practice: Current Trends and Future
Opportunities. Accounting Perspectives, 19(4), 325-361.
Smith, S. S., & Castonguay, J. J. (2020). Blockchain and accounting governance: Emerging issues and considerations for
accounting and assurance professionals. Journal of Emerging Technologies in Accounting, 17(1), 119-131.
Supriadi, I. (2020). The Effect of Applying Blockchain to the Accounting and Auditing. Ilomata International Journal of
_Tax and Accounting, 1(3), 161-169._
Wei, P., Wang, D., Zhao, Y., Tyagi, S. K. S., & Kumar, N. (2020). Blockchain data-based cloud data integrity protection
mechanism. Future Generation Computer Systems, 102, 902-911.
Zhang, Y., Xiong, F., Xie, Y., Fan, X., & Gu, H. (2020). The Impact of Artificial Intelligence and Blockchain on the
Accounting Profession. IEEE Access, 8, 110461-110477.
**Copyrights**
Copyright for this article is retained by the author(s), with first publication rights granted to the journal.
[This is an open-access article distributed under the terms and conditions of the Creative Commons Attribution license](http://creativecommons.org/licenses/by/4.0/)
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.11114/afa.v8i1.5492?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.11114/afa.v8i1.5492, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "GOLD",
"url": "https://redfame.com/journal/index.php/afa/article/download/5492/5693"
}
| 2,022
|
[] | true
| 2022-02-22T00:00:00
|
[
{
"paperId": "fb990d334fd6e2a0a41e6b7ddb1113fe8d1a35ee",
"title": "Blockchain in Accounting Research and Practice: Current Trends and Future Opportunities*"
},
{
"paperId": "95875321d1f95d60f8a73504e2a61c1fa96cc7df",
"title": "The Effect of Applying Blockchain to The Accounting and Auditing"
},
{
"paperId": "b20f2936c062881eb02f4e4e0b9ffd0eb940d201",
"title": "The Impact of Artificial Intelligence and Blockchain on the Accounting Profession"
},
{
"paperId": "206cfd7b124bc80f685a1187784d1ba188bbda9a",
"title": "The Blockchain Evolution and Revolution of Accounting"
},
{
"paperId": "3cc0cd9a47a08a175150c2e3901b6069bda953f5",
"title": "Blockchain technology in the future of business cyber security and accounting"
},
{
"paperId": "7c1dc3b5a1f6f331c782a08277caf26111a4b62d",
"title": "Blockchain data-based cloud data integrity protection mechanism"
},
{
"paperId": "390e429a86a5a88c8bcec5ae33fb38c21284ddc4",
"title": "Blockchain and Accounting Governance: Emerging Issues and Considerations for Accounting and Assurance Professionals"
},
{
"paperId": "24f820574ea153c7a8a2788d54748d96747f5ceb",
"title": "Entries"
},
{
"paperId": "784e3b742b53c448dc88ed5048da87052a8fd591",
"title": "Auditing Cloud-Based Blockchain Accounting Systems"
},
{
"paperId": "bc9e888c876a33b841915fcc1842cd4533d5f82f",
"title": "The Impact of Blockchain on Accounting Information Systems"
},
{
"paperId": "2dd69343163deb265cf3404e781f4b5783467914",
"title": "Dollarization 2.0 a Cryptocurrency: Impact on Traditional Banks and Fiat Currency"
},
{
"paperId": "210eb61e631e244675751e53845d55ff32a096bf",
"title": "Blockchain: Emergent Industry Adoption and Implications for Accounting"
},
{
"paperId": "a6ed0c2467e56f569ce17885e9d5c1dae396b38d",
"title": "Toward Blockchain-Based Accounting and Assurance"
},
{
"paperId": null,
"title": "Basic Financial Accounting and Reporting"
},
{
"paperId": "8290128f83a8bba902898f63338c57634a89dcaa",
"title": "Assurance"
},
{
"paperId": "29dc3f218fffb59992ed545f42f86dbcd95137ff",
"title": "Implementation of Blockchain Technology in Accounting Sphere"
},
{
"paperId": "8f3711a7a4bed59c6002f9a6db2a6625443f2d59",
"title": "Blockchain for Good? Digital Ledger Technology and Sustainable Development Goals"
},
{
"paperId": null,
"title": "Block chain"
},
{
"paperId": "25075e27b0df6f2be5a8c519171bdabd1c3ed817",
"title": "Content Analysis: An Introduction to Its Methodology"
},
{
"paperId": null,
"title": "Worksheet preparation, including correcting entries. This simplifies the process of preparing financial statements"
},
{
"paperId": null,
"title": "Constructing a trial balance"
},
{
"paperId": null,
"title": "Identifying the events that will be recorded"
},
{
"paperId": null,
"title": "The ledger is updated using journal entries. This procedure is intended to transmit data from the journal to the ledger for classification"
},
{
"paperId": null,
"title": "Journal entries for the closing journal are journalized and posted. This results in the closure of temporary accounts and the transfer of profit to the owner's equity"
},
{
"paperId": null,
"title": "Journal entries for adjustments are journalized and posted. This column is used to track accruals, deferrals that have expired, estimations, and other occurrences from the worksheet"
},
{
"paperId": null,
"title": "The journal is used to record transactions. This tries to document the economic impact of transactions on the firm in a journal, a format that allows transfer to the accounts"
}
] | 4,290
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Engineering",
"source": "s2-fos-model"
},
{
"category": "Environmental Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/0240a106144dbee52e5e527f8e2f0cd21fc3126b
|
[
"Computer Science"
] | 0.82466
|
Publicly Verifiable Spatial and Temporal Aggregation Scheme Against Malicious Aggregator in Smart Grid
|
0240a106144dbee52e5e527f8e2f0cd21fc3126b
|
Applied Sciences
|
[
{
"authorId": "2152829737",
"name": "Lei Zhang"
},
{
"authorId": "2155698444",
"name": "Jing Zhang"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"Appl Sci"
],
"alternate_urls": [
"http://www.mathem.pub.ro/apps/",
"https://www.mdpi.com/journal/applsci",
"http://nbn-resolving.de/urn/resolver.pl?urn=urn:nbn:ch:bel-217814"
],
"id": "136edf8d-0f88-4c2c-830f-461c6a9b842e",
"issn": "2076-3417",
"name": "Applied Sciences",
"type": "journal",
"url": "http://www.e-helvetica.nb.admin.ch/directAccess?callnumber=bel-217814"
}
|
We propose a privacy-preserving aggregation scheme under a malicious attacks model, in which the aggregator may forge householders’ billing, or a neighborhood aggregation data, or collude with compromised smart meters to reveal object householders’ fine-grained data. The scheme can generate spatially total consumption in a neighborhood at a timestamp and temporally a householder’s billing in a series of timestamps. The proposed encryption scheme of imposing masking keys from pseudo-random function (PRF) between pairwise nodes on partitioned data ensures the confidentiality of individual fine-grained data, and fends off the power theft of n-2 smart meters at most (n is the group size of smart meters in a neighborhood). Compared with the afore-mentioned methods of public key encryption in most related literatures, the simple and lightweight combination of PRF with modular addition not only is customized to the specific needs of smart grid, but also facilitates any node’s verification for local aggregation or global aggregation with low cost overhead. The publicly verifiable scenarios are very important for self-sufficient, remote places, which can only afford renewable energy and can manage its own energy price according to the energy consumption circumstance in a neighborhood.
|
# applied sciences
_Article_
## Publicly Verifiable Spatial and Temporal Aggregation Scheme Against Malicious Aggregator in Smart Grid
**Lei Zhang** **[1]** **and Jing Zhang** **[2,1,]***
1 College of Computer Science and Technology, Harbin Engineering University, Harbin 150001, China;
lei_power@hrbeu.edu.cn
2 School of Information Science and Engineering, Jinan University, Jinan 250022, China
***** Correspondence: ise_zhangjing@ujn.edu.cn
Received: 16 November 2018; Accepted: 14 January 2019; Published: 31 January 2019
[����������](http://www.mdpi.com/2076-3417/9/3/490?type=check_update&version=1)
**�������**
**Abstract: We propose a privacy-preserving aggregation scheme under a malicious attacks model,**
in which the aggregator may forge householders’ billing, or a neighborhood aggregation data, or
collude with compromised smart meters to reveal object householders’ fine-grained data. The
scheme can generate spatially total consumption in a neighborhood at a timestamp and temporally
a householder’s billing in a series of timestamps. The proposed encryption scheme of imposing
masking keys from pseudo-random function (PRF) between pairwise nodes on partitioned data
ensures the confidentiality of individual fine-grained data, and fends off the power theft of n-2
smart meters at most (n is the group size of smart meters in a neighborhood). Compared with
the afore-mentioned methods of public key encryption in most related literatures, the simple and
lightweight combination of PRF with modular addition not only is customized to the specific needs of
smart grid, but also facilitates any node’s verification for local aggregation or global aggregation with
low cost overhead. The publicly verifiable scenarios are very important for self-sufficient, remote
places, which can only afford renewable energy and can manage its own energy price according to
the energy consumption circumstance in a neighborhood.
**Keywords: smart metering; spatial and temporal aggregation; privacy protection; internal attack;**
pseudo-random function
**1. Introduction**
With the development of Advanced Metering Infrastructure (AMI), Smart Metering as an
important research subject in Smart Grid (SG) plays an increasingly important role and is closely
associated with people’s daily life [1,2]. Aggregating fine-grained metering data attracts householders
and power suppliers. Power suppliers can calculate, forecast, and regulate accurately power
distribution/price of the next period in real time while detecting fraud reports. Based on billing
details and current power price, householders can adjust its appliance consumption module to reduce
the power billing at the peak time; however, accessing householder’s information on metering may
cause security and privacy concerns, such as daily routines, the type of applications, etc. [1,2]. For
this, in SG systems, one of the challenges faced by power big data is how to design one aggregation
mechanism to balance the use of power data and individual privacy protection [2].
Protecting such sensitive private data from individual privacy threats needs to limit the authority
of the utility company employee [2]. Namely, Supplier Billing System (SBS, sub-suppliers) will know
only the total amount of the consumption for each customer, while the Energy Management System
(EMS, demand prediction division) should know only the total consumption of customers in a certain
region for each time period. To achieving the goals, smart metering systems often introduce the Meter
-----
_Appl. Sci. 2019, 9, 490_ 2 of 20
Data Management System (MDMS), which stores the measured values of smart meters (SMs), and
aggregates it before sending the aggregation to the SBS and EMS [2].
With the appearance of MDMS, another concern is upgrading, namely the malicious action of
householders and regional MDMS employees. Unfortunately, a malicious householder may collude
with the regional MDMS employee to report a false consumption to the SBS department; attackers may
steal or forge power usage and consumption information. In addition, a regional MDMS employee
may submit a fraudulent aggregation in a neighborhood. A World Bank report finds that each year
over 6 billion dollars cannot post due to the energy theft and fraud report in the United States, in 2009,
the FBI reported a wide and organized attempt that may have cost up to $400 million loss annually
and power supplier suffered a great monetary loss [3]. To fend off this type of attack, it is desirable
that suppliers or the public should detect the fraud profile from malicious aggregators or dishonest
householders [4].
Privacy-preserving metering protocols have been discussed in lots of literatures [5–24]. They
mainly focused on the studies of homomorphic aggregation [6–17,20,23,24], by which, aggregators
can only obtain the fine-grained aggregation data within a certain region or householders’ billing in a
serial period while protecting individual privacy. However, most of them can only resist against single
external or semi-trustable attack [11,13,14,19], and how to fend off internal attackers (e.g. aggregators or
householders) is an open problem. Internal attackers can legally collect and store power consumption
information of users; therefore, they pose a higher threat than external attackers [18].
Most of existing works [5–10,18,20,21,23,24] about additively homomorphic, multiplicative
homomorphic, and their combination with other cryptography endeavored to address the problem.
Most of them improved Paillier encryption [6–9] by their combination with other cryptography, such
as stream ciphers [5,19,23] and modular addition [7,24], to prevent power suppliers/operators from
intercepting individual user data, and to detect fraudulent from dishonest users. To ensure the integrity
of transmitted messages and fend off attacks such as man-in the middle attack and denial-of-service
attack against SG, signature and authentication methods are proposed in References [8,15–17].
Lu et al. [6] proposed a privacy-preserving, multi-dimensional metering aggregation scheme in a
neighborhood-wide grid with piallier encryption, bilinear pairing and computational Diffie-Hellman
(DH) methods. For resisting against internal attackers possessing private keys, Xiao [8] introduced
a spatial and temporal aggregation and authentication scheme by randomizing Paillier encryption
with Lagrange interpolation. Their protocol requires O(n[2]) bytes of inter-action between the individual
meters as well as relatively expensive cryptography on the meters (public key encryption). Chen [9]
also improved Paillier encryption and proposed a privacy-preserving aggregation scheme resisting at
most t compromised servers in a control center with threshold protocol.
Dimitriou et al. [20] provided a verifiable publicly aggregation scheme against dishonest users
that attempt to provide fraudulent data. Any user node in the community can prove its computation
accuracy by zero-knowledge proof that the two encrypted message with different public keys
corresponding to the same plaintext message. While we can prove our scheme costs lower overhead to
resist fraudulent report from internal nodes.
Erkin et al. [23] adopted a stream cipher (e.g. RC4) to generate pseudo-random keys as masking
keys between nodes to prevent internal nodes from possessing private keys. During the aggregation
within a neighborhood, all masking random keys cancelled out and the aggregation value is revealed
without compromising individual privacy based on the security properties of the Paillier encryption
and stream cipher. We follow its Pseudo-Random Function and combine it with modular addition.
The main difference from ours is they impose the random keys from PRF on the plaintext before
encrypting it with Paillier cryptography, and send the encrypted message to all nodes. We set a
security parameter k to represent the number of communicate nodes in a neighborhood and improve
the encryption method by replacing the costly Paillier encryption with the simple and lightweight
combination. More significantly, we supplement a publicly verifiable property to detect the fraudulent
profile from malicious aggregators or dishonest user nodes.
-----
_Appl. Sci. 2019, 9, 490_ 3 of 20
Castelluccia et al. [19] protected individual data by imposing masking keys from RC4 on the
plaintext data under the multi-level wireless sensors network model, However, the protection protocol
cannot resist malicious aggregators, as the session keys are generated by the sink as the aggregator. We
extend its PRF method into the peer-to-peer system model and propose a privacy-preserving scheme
against maliciously internal attack.
In addition, traditional modular addition was adopted in [7,24] by partitioning individual
plaintext data into n shares and exchanging them between nodes (n is the number size of users
in a neighborhood). Flavio et al. [7] adopted Paillier encryption and modular addition, in which
every user node partitions its meter reading into n shares and transmits the encrypted shares with
different public keys to the aggregator, which aggregates the data with the same public key before
sending the aggregation to the users. Finally, the aggregator collects the plaintext sums to obtain
the final aggregation. The method is privacy-preserving; however, during each spatial aggregation,
three message exchanges are required between every user and the aggregator. Thus, the number
of homomorphic encryption per user increases linearly with n increases, and the communication
overhead is O(n[2]) messages [20]. Jia et al. [24] also generated partitioned data with modular addition
and imposed them on a high-order polynomial coefficient. The values of the polynomial at different
points are transmitted to the aggregator which finds the coefficients of the polynomial with the private
key and gains the aggregation, so the scheme is under the semi-trusted model and the aggregator
is trustable. In addition, the computation overhead is relatively higher when k is increasing. As
every node does the x[k] polynomial operation before the matrix multiplication operation, the scheme
increases greatly the computation overhead.
Ohara K et al. [4] summarized the function requirements during smart metering against internal
attackers: calculating billing and obtaining statistics for energy management. We follow the statistic
function requirements and the spatial and temporal scenarios in References [8,23] against malicious
MDMS/aggregators or dishonest users:
(1) Spatial aggregation. A neighborhood-wide grid corresponds to a group of householders each
equipped with a SG. They submit their encrypted meterings to the MDMS at a timestamp (e.g.,
15 min). The latter aggregates homomorphically them before sending the aggregation to the EMS.
During this aggregation, the individual data are confidential to the MDMS or the EMS.
(2) Temporal aggregation. A single SM submits its power consumption in a series of timestamps
to the MDMS for the billing purpose. In this scenario, SBS charges the householders in
serial timestamps.
Throughout this paper, we refer to the building area network (BAN) region as a neighborhood,
and the regional MDMS as the regional gateway (GW), and the regional SBS as the control center
(CC), respectively.
The main contribution can be summarized as follows:
(1) We design and implement a distributive, temporal and spatial aggregation scheme in the SG, in
which every node sends and receives k encrypted message from k pairwise nodes distributively.
The scheme provides spatial aggregation in a neighborhood at a fine-grained time scale (e.g.
15 min) and an individual temporal aggregation (e.g. monthly) in a series of timestamps for the
billing purpose.
(2) The proposed encryption scheme minimizes the computation and communication overhead by
replacing the costly public key cryptography adopted in most literatures with a combination of
modular addition and PRF.
(3) The novel feature is that the masking keys are imposed on the partitioned data, and the latter are
implemented by traditional modular addition. As the process of modular addition is processed
by the node itself, other nodes cannot gain the true partitioned data, the masking key is only
known to the pairwise nodes, and the combination ensures the confidentiality of individual data
to any node including CC, aggregators, and n-2 nodes at most in a neighborhood.
-----
_Appl. Sci. 2019, 9, 490_ 4 of 20
(4) To detect malicious aggregators or dishonest users, we propose innovatively a publicly verifiable
aggregation method. By this way, any user node in a neighborhood can receive the communication
flow, and verify the accuracy of local aggregation from other nodes or total aggregation from the
aggregator without compromising individual fine-grained data.
(5) The publicly available property for the aggregation also facilities householders regulating in
time its current consumption module and consumption demand in the next time period, as by
comparing their own consumptions with those of other nodes and checking if there is redundant
power, householders can decide to store more energy or to sell excess power to the power supplier
or other nodes. The scenarios are especially very important for self-sufficient, remote places,
particularly, in developing countries, which can only afford renewable energy, such as wind
turbines, solar panels, and carbon-based fuels [23].
The paper is organized as follows: in Section 2, we provide related preliminaries and formalize
the system and attack models. In Section 3, we introduce our proposed aggregation scheme and
correctness analysis. Security notions and proof are given in Section 4, followed by performance
evaluation and comparison in Section 5. The conclusion is drawn in Section 6.
**2. Preliminaries and Models**
For ease of reading, we summarize the main notations in the paper in Table 1.
**Table 1. Notations in the scheme.**
**Symbol** **Meaning**
_HSM/SM_ HAN smart meter/ user/user node/sm
_N_ The number of users in a BAN neighborhood1
_k_ The number of pairwise nodes for every user
_K_ Keystream based on stream cipher
_M_ RSA modular (large prime)
_x([i]_ _j,d)_ _useri partitioned data into userj at timestamp d_
_xd[i]_ _useri’s data at timestamp d_
_r([i]_ _j,d)_ _useri’s pairwise key with userj at timestamp d_
_E([i]_ _j,d)_ The encrypted form of x([i] _j,d)_
_ski_ The secret key between CC and every node
_indi[s](s = 1, · · ·, k)_ _useri’s pairwise nodes set in serial timestamps T_
_LS(j, d)_ _userj’s locally spatial aggregation at timestamp d_
_LT([i]j,d)[(][j][ ∈]_ _[ind][i][[][s][])]_ _userj’s locally temporal aggregation for useri in T_
_AT(i, T)_ _useri’s temporal aggregation in T_
_ASd_ Spatial aggregation in a neighborhood at timestamp d
_2.1. Additively Homomorphic Encryption Based on The Keystream_
Our security property partly comes from the stream cipher. The keystream generated from the
pseudo-random function satisfies the security properties of the additively homomorphic encryption in
the stream cipher. The basic idea [19] is denoted as follows:
Encryption is written as: c = Enck(m + K) mod M where K is randomly generated keystream, m
is the plaintext and m, k [0, M 1].
_∈_ _−_
Decryption is described as: Deck= c−K mod M.
Additively homomorphic property of ciphertext are described as: c1 = EncK1(m1) and c2 =
_EncK2(m2); then, the aggregated ciphertext is expressed as: c = c1 + c2mod M = EncK(m1 + m2),_
where K = K1 + K2 mod M.
_2.2. Pseudo-Random Keystream Generator—RC4_
As a popular PRF generator, with secret keys between communication nodes, RC4 can generate a
keystream. This secret key is pre-computed during the system initialization. As any stream cipher,
the generated keystream can be used for encryption by combining it with the plaintext using bit-wise
-----
_Appl. Sci. 2019, 9, 490_ 5 of 20
Exclusive-Or [19]. However our scheme is to replace the XOR (Exclusive-OR) operation typically
found in stream ciphers with modular addition operation (+). To generate the keystream, RC4 needs
two algorithms, i.e. Key-scheduling algorithm (KSA) and Pseudo-random generation algorithm
(PRGA) [5,14].
KSA: KSA is to initialize a permutation with a variable length key between 40 and 2048 bits
for PRGA.
PRGA: once the permutation initialization of KSA has been completed, the stream of bits is
generated using the PRGA.
**Algorithm 1: Key-scheduling algorithm (KSA)**
**Input:** i = 0;
j = 0 //Two 8-bit index-pointers
S //The initial key keyed with a secret key
**Output: S //A permutation of all 256 possible bytes**
1. for (i = 0; i <= 255; ++ i)
2. S[i] = i;
3. end
4. k = 0;
5. for (i = 0; i <= 255; ++ i)
6. j = (j + s[i] + key[i mod keylength]) mod 256;
7. k = S[i];
8. S[i] = S[j];
9. S[j] = k;
10. end
**Algorithm 2: Pseudo-random generation algorithm (PRGA)**
**Input:** i = 0;
j = 0 //Two 8-bit index-pointers
**Output: Z // Pseudo-random keystream**
1. k = 0;
2. for (i = 0; i <= 255; ++ i)
3. i = (I + 1) mod 256;
4. j = (j + S[i]) mod 256;
5. k = S[i];
6. S[i] = S[j];
7. S[j] = k;
8. Z = S[(S[i] + S[j]) mod 256];
9. end
_2.3. System Model_
In our system model, we consider a typical SG communication architecture [8,9,11,15–17], as
shown in Figure 1. It is based on the SG network model presented from the National Institute of
Standards and Technology (NIST) and consists of six domains, i.e., the power plant, the transmission
domain, the distribution domain and a CC, a residential GW, and the user domain. We mainly focus
on how to report and aggregate the users’ privacy-preserving data into the CC. Hence, the system
model divides especially the BAN into numbers of Household area network (HAN) equipped with a
SG and every BAN includes a GW and numbers of users.
CC: It acts as the SBS and EMS in reality. It needs to monitor the actual data on how much power
is consumed at which timestamp in one BAN (neighborhood), how much power should be reserved
for the next time period, and cumulative consumption for individual billing on a monthly basis, and
how much power is being distributed to a specified neighborhood. In the paper, it is curious about the
individual fine-grained data and may attempt to it as far as possible by all available resources, so it is
assumed a semi-trusted entity.
-----
_Appl. Sci. 2019, 9, 490_ 6 of 20
GW: A powerful entity, acting as the local MDMS, represents a locality (e.g., a region within a
building) is responsible for aggregating real-time spatial data in a neighborhood and individual
temporal data in a series of timestamps, and then transmitting the aggregation to the CC. The
employment of GW relieves CC of aggregation and reducing largely the communication latency.
However, the cost that potentially malicious attacks done to users or power suppliers is unignorable,
as discussed earlier. We assume it is a malicious entity here. A BAN GW represents a locality (e.g., a
region within a neighborhood). For facilitating the communication between BAN GW and CC, WiMax
and other broadband wireless technologies can be adopted. We consider a scenario that one BAN
neighborhood covers a hundred or more HANs, so the longest distance from the BAN GW to a HAN
is more than a hundred miles, so WiMax maybe more suitable for this kind of distance communication.
Household Smart Meter (HSM): A bidirectional communication entity deployed at householders’
premises. The modern SM is given a certain level of autonomy via trusted elements and the ability to
collect, store, aggregate, and encrypt the usage data. Hence it has two interfaces—one interface is for
reading power of householders and the other one acts as a communication GW. Even if we assume SM
_Appl. Sci. 2018, 8, x FOR PEER REVIEW_ 7 of 22
is tamper-resistant, it is not powerful as a GW, so it may be vulnerable to be compromised by the GW
to infer the object users’ data.
**Figure 1. System model under consideration.**
_2.4. Communication Model_
**Figure 1. System model under consideration.**
As can be seen in the Figure 1, all SMs connect each other in a neighborhood by WiFi technique,
_2.5. Data Model_
which constructs public verifiable foundation. Each user would select randomly k pairwise nodes in
one round and can ensure that ifLet _x_ _di_ be the meter reading of the useri choosesi[th] (1 user≤ _i_ ≤j, thenN) user node at the userj chooses userd[th] (1 i and the keys between≤ _d_ ≤ _T) fine-grained_
them are opposite mutually. The valuetimestamp, where N is the number of user in a BAN (a neighborhood-wide grid), and k as a security parameter can take any value from 2 toT is a billing n, and
depend on the specific application circumstance. The higher the value ofperiod. At each fine-grained time index _d, a neighborhood grid (over the entire BAN) spatially k is, the higher the complexity_
is, and vice versa, and the scheme is more vulnerable to be attacked.aggregated utility usage can be expressed as:
_2.5. Data Model_
###### AS d( ) = in=1 xdi ；d = 1, 2,…,T (1)
LetAt the end of a billing period ( xd[i] [be the meter reading of the]d = [ i]T[th (1]), a temporally aggregated utility usage data for the [ ≤] _[i][ ≤]_ _[N][) user node at the][ d][th (1][ ≤]_ _[d][ ≤]_ _[T][) fine-grained]i[th] user_
timestamp, where N is the number of user in a BAN (a neighborhood-wide grid), and T is a billing
is expressed as:
period. At each fine-grained time index d, a neighborhood grid (over the entire BAN) spatially
aggregated utility usage can be expressed as:AT i T(, ) = Td =1 _xdi；i_ = 1 toN (2)
_n_
_AS(d) = ∑i=1_ _[x]d[i]_ [;][ d][ =][ 1, 2, . . .,][ T] (1)
_2.6. Security Requirement and Attack Model_
-----
_Appl. Sci. 2019, 9, 490_ 7 of 20
At the end of a billing period (d = T), a temporally aggregated utility usage data for the ith user is
expressed as:
_T_
_AT(i, T) = ∑d=1_ _[x]d[i]_ [;][ i][ =][ 1 to][ N] (2)
_2.6. Security Requirement and Attack Model_
Within the system model, there are four types of actors involved in the meter data reporting
process: the ith user (self), other users in the same neighborhood (BAN), the GW, and the CC. The CC
requires the spatially aggregated fine-grained neighborhood usage data to optimize power delivery
efficiency and the temporally aggregated user-specific utility usage data for the billing purpose. Hence,
we stipulate the following security/privacy requirements:
**_Requirement R1. Fine-grained, individual utility data are private and should not be disclosed to_**
CC, GW, or other users.
**_Requirement R2. Temporal aggregation for an individual user and spatial aggregation in one_**
neighborhood cannot be tampered by the malicious aggregator or other internal nodes. For this, we
envision a secure and reliable communication model comprising a verifiable publically method, which
is customized to the correctness verification of the aggregation value of SG.
For this, our attack model is based on the malicious aggregator who attempts to tamper the
aggregation value in a neighborhood and the billing value for individual users, or infers fine-grained
meterings of the individual user by colluding with other n-2 compromised nodes at most. Following
the above security requirements, different compositions of the attackers and actions may be grouped
into the following attack types:
(1). External attack
External attackers may compromise the meterings of the object users by eavesdropping the
communication flow between communication nodes through various eavesdropping malware.
(2). Malicious attack
False aggregation report. The aggregator may alter or drop maliciously any individual data, or
tamper the aggregation data to the CC; any malicious user node may provide false local aggregation
to the GW.
Collusion with compromised nodes. The aggregator may collude with compromised users to
attempt to infer the uncompromised users’ data.
(3). Semi-trustable internal attack
The curious CC or any user node can also acquire data through the public communication flow,
such as the message from the user node to the GW or from the GW to the CC. They may infer the
object user’s fine-grained data by the public communication flow.
An attack is an arrangement that enables unauthorized parties to gain access to private data or to
tamper secured data (even by the user itself) without being detected. In this work, we assume the SMs
are tamper-resistant [7,20,23], and can perform the measurement and reporting operations normally,
but do not exclude the possibility of tampering with local aggregation values by itself.
**3. Proposed Scheme**
_3.1. Initialization Phase_
3.1.1. Initializing Pairwise Number k and Session Key
For every billing period, the CC generates randomly the pairwise number for every node in one
neighborhood denoted as k, and broadcasts it to all SMs.
We generate session keys between every node with the computational DH key exchange protocol
as the initial key in RC4 to generate the keystream between pairwise nodes. Once one node joins a
-----
_Appl. Sci. 2019, 9, 490_ 8 of 20
neighborhood size of n, it generates itself one DH public key g[a] (mod M) and remains the secret key
_a, M are DH parameters, and then broadcasts the public key. By this Computational Diffie-Hellman_
CDH exchange key, any two pairwise nodes can identify their session key formed as g[ab].
3.1.2. Modular Addition
The useri partitions its own data xd[i] [into][ k][ partitions denoted as][ x]([i] _indi[j],d)[(][1][ ≤]_ _[j][ ≤]_ _[k][)][ and sends]_
them to every pairwise node. However, the partitioned data can be easily guessed, especially with
brute search, as the consumption value at every timeslot is very small. For this, we impose extra noise
(masking keys) which is only known by pairwise nodes themselves on the partitioned data to further
secure the individual data.
3.1.3. Noise Addition
Masking keys, as extra noise, are generated by pairwise nodes with PRF at every timestamp. The
PRF can be implemented with RC4, the specific process can be referred to the Section 2.2.
_3.2. Encryption and Aggregation_
3.2.1. Data Encryption
(1). Partition of individual data
Each node randomly partitions its individual data into k partitions and sends them to k pairwise
nodes along with the masking keys. The partition form is as follows:
_xd[i]_ [=] ∑ _x([i]_ _j,d)[(][1][ ≤]_ _[s][ ≤]_ _[k][)]_ (3)
_j∈indi[s]_
(2). Generation of pairwise nodes and masking keys
For any node, it chooses randomly any k nodes in one round as its pairing nodes such that if useri
selects userj, then userj also selects useri. With the session key between them, the two pairwise nodes
generate a common key r from RC4; useri adds r([i] _j,d)_ [to][ x]([i] _j,d)[, and][ user][j][ adds][ r]([j]i,d)_ [which satisfies:]
_r([j]i,d)_ [=][ −][r]([i] _j,d)[(][i][ ∈]_ _[ind][j][[][s][]][;][ j][ ∈]_ _[ind][i][[][s][])]_ (4)
For useri, the generated noise set at the timestamp d can be denoted as r([i] _indi[s],d)_ [(][s][ = 1, 2,][ . . .][,][ k][).]
Note that in order to facilitate the temporal aggregation, the pairwise key generated by an SM at the
_T[th]_ timestamp should satisfy the following equation:
_r([j]i,d)_ [=][ −][r]([i] _j,d)[(][i][ ∈]_ _[ind][j][[][s][]][;][ j][ ∈]_ _[ind][i][[][s][])]_ (5)
(3). Encryption process
At the timestamp d, useri adds the pairwise noise to the partitioned data to generate the encrypted
message E[i]
(j,d) [=][ x]([i] _j,d)_ [+][ r]([i] _j,d)[(][j][ ∈]_ _[ind][i][[][s][])][ to][ k][ pairwise nodes separetely as well as receiving the]_
encrypted message they sent. The Figure 2 illustrates an example for spatial and temporal aggregation
among pairwise users in multi-region groups.
-----
_Appl. Sci. 2019, 9, 490_ _r(, )i dj_ = −r(, )ij d [(]i ∈ind _j[s];_ _j_ ∈indi[s]) 9 of 20(5)
_LS_ (1, ) d _LS_ (2, ) d _LS(3, ) d_ _LS i d(, )_ _LS k d(, )_ _LS j d(,_ ) _LS n d(, )_
_user1_ _x12,d_ + _r2,1_ _d_ _xi d1,_ + _ri d1,_ _x1j d,_ + _rj d1,_ _x1n d,_ + _rn d1,_
_user2_ _x1,2d_ + _r1,2d_ _x3,2_ _d_ + _r3,2d_ _xk d2,_ + _rk d2,_ _xn d2,_ + _rn d2,_
_user3_ _x2,3_ _d_ + _r2,3d_ _xi d3,_ + _ri d3,_ _xk d3,_ + _rk d3,_
_useri_ _x1,i_ _d_ + _r1,id_ _x3,i_ _d_ + _r3,i_ _d_ _xij d,_ + _rj di,_
_userk_ _x2,k_ _d_ + _r2,kd_ _x3,k_ _d_ + _r3,kd_ _xkj d,_ + _rj dk,_ _xn dk,_ + _rn dk,_
_userj_ _x1,j_ _d_ + _r1,jd_ _xi dj,_ + _ri d,j_ _xk dj,_ + _rk dj,_ _xn dj,_ + _rn dj,_
_usern_ _x1,nd_ + _r1,nd_ _x2,n_ _d_ + _r2,nd_ _xk dn,_ + _rk dn,_ _xnj d,_ + _rj dn,_
_user1_ _user2_ _user3_ _useri_ _userk_ _userj_ _usern_
_user1_
_user2_
_useri_
_userk_
_userj_
_usern_
_t =_ 1t = 2 _t_ = _d_ _t_ = _T_ −1t = _T_
_LT(, )ij T_
**Figure 2. Example multi-region spatial and temporal aggregation.**
**Figure 2. Example multi-region spatial and temporal aggregation.**
For any SM node j, it will store the encrypted data sent from one of its pairwise node i in a series
of T timestamps in the form of matrix as follows:
|Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|Col11|Col12|Col13|Col14|
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
|||||||||||||||
||||x1 +r1 2,d 2,d||||x1 +r1 i,d i,d||||x1 +r1 j,d j,d||x1 +r1 n,d n,d|
||x2 +r2 1,d 1,d||||x2 +r2 3,d 3,d||||x2 +r2 k,d k,d||||x2 +r2 n,d n,d|
||||x3 +r3 2,d 2,d||||x3 +r3 i,d i,d||x3 +r3 k,d k,d|||||
||xi +ri 1,d 1,d||||xi +ri 3,d 3,d||||||xi +ri j,d j,d|||
||||xk +rk 2,d 2,d||xk +rk 3,d 3,d||||||xk +rk j,d j,d||xk +rk n,d n,d|
||xj +rj 1,d 1,d||||||xj +rj i,d i,d||xj +rj k,d k,d||||xj +rj n,d n,d|
||xn +rn 1,d 1,d||xn +rn 2,d 2,d||||||xn +rn k,d k,d||xn +rn j,d j,d|||
|||||||||||||||
|Col1|Col2|Col3|Col4|Col5|Col6|
|---|---|---|---|---|---|
|||||||
|||||||
_E[1]_ _E[2]_ _E[i]_ _E[n]_
(j,1) (j,1) _· · ·_ (j,1) _· · ·_ (j,1)
_E[1]_ _E[2]_ _E[i]_ _E[n]_
(j,2) (j,2) _· · ·_ (j,2) _· · ·_ (j,2)
... ... ... ...
_E[1]_ _E[2]_ _E[i]_ _E[n]_
(j,T) (j,T) _· · ·_ (j,T) _· · ·_ (j,T)
�
_E([i]_ _j,1)_ [= (][x]([i] _j,d)_ [+][ r]([i] _j,d)[)]_ _j ∈_ _indi[s]_
_E([i]_ _j,d)_ [=][ 0] _j /∈_ _indi[s]_
�
(6)
3.2.2. Storage and Aggregation
(1). Spatial Aggregation
Once receiving encrypted data at timeslot d from all pairwise nodes, useri aggregates them and
generates the local spatial data LS (i, d) as follows:
_n_
_LS(i, d) =_ ∑ (x([j]i,d) [+][ r]([j]i,d)[)][ mod][ M][(][1][ ≤] _[s][ ≤]_ _[k][)]_ (7)
_j∈indi[s]_
Every user sends the local spatial aggregation formed as LS (i, d) to the GW at every timestamp.
Once receiving the locally spatial aggregation LS (i, d) from the pairwise nodes, the GW adds
them up together and the pairwise keys cancel out. The total spatial aggregation is denoted as:
_n_
_ASd =_ ∑ _LS(i, d) mod M_ (8)
_i=1_
(2). Temporal aggregation
Every user node receives the encrypted data from its pairwise nodes and stores it as a matrix of T
rows and n columns formed as Equation (6).
-----
_Appl. Sci. 2019, 9, 490_ 10 of 20
In every billing period T, the user node aggregates every column in the Equation (6) into locally
temporal aggregation after the pairwise keys cancel out. The locally temporal aggregation form is
as follows:
_T_
_LT([i]j,T)_ = ∑ _E([i]_ _j,d)_ [mod][ M][(][j][ ∈] _[ind][i][[][s][])]_
_d=1_ (9)
_T_
= ∑ (x([i] _j,d)_ [+][ r]([i] _j,d)[)][ mod][ M]_
_d=1_
Once the CC issues the temporal aggregation request for useri to the GW, the pairwise nodes of
_useri would report its local temporal aggregation LT([i]j,T)_ [to the GW.]
The GW aggregates them into the temporal aggregation and transmits it to the CC; the aggregation
process is as follows:
_AT(i, T) =_ ∑ _LT([i]j,T)_ [mod][ M][(][1][ ≤] _[s][ ≤]_ _[k][)]_ (10)
_j∈indi[s]_
_Appl. Sci. 2018, 8, x FOR PEER REVIEW_ 11 of 22
We assume j ∈ _indi[s]; i ∈_ _indj[s]._
Figure 3 shows the communication process between the pairwise nodes and GW at the
timestampFigure 3 shows the communication process between the pairwise nodes and GW at the d.
timestamp d.
_SM_ _i_ _SM_ _j_ **_GW_**
Generate _x_ _di_ ;
Select randomly _k pairwise nodes_
and generate k pairwise keys
formed as _r(, )ij d_ ;
Generate its pairwise set S(i,d);
Partition _x_ _di_ into _k_ partitions
formed as _x(, )i_ _j d_ ;
Generate _E(, )ij d_ with the sum of
_x(, )i_ _j d_ and _r(, )ij d_ ;
{ _E(, )ij d_, d } ①
{ _E(, )i dj_, d } ①
Compute local spatial aggregation
_LS(i,d) at every timestamp; and_
local temporal aggregation _LT(, )i Tj_
in every billing period.
_SMj_
_LS(i,d), d }②_
_LT(, )i Tj_
{LS( j,d), d }②
Aggregate all _LS_ (i, _d) (1≤i≤n); generate_
_ASd; aggregate all_ _LT(, )i Tj_ ( _i_ ∈ind sj[ ] ) to
generate AT(j, T).
**Figure 3.Figure 3. Communication process data between the pairwise nodes and GW.Communication process data between the pairwise nodes and GW.**
3.2.3. Decryption Process3.2.3. Decryption Process
In this way, the aggregation process is actually the decryption process, in which the random keysIn this way, the aggregation process is actually the decryption process, in which the random
keys cancel out and individual consumption in a billing period or the spatial aggregation in a
cancel out and individual consumption in a billing period or the spatial aggregation in a neighborhood
neighborhood is revealed. Hence the combination of simple modular addition with noise addition
is revealed. Hence the combination of simple modular addition with noise addition reduces the costly
reduces the costly encryption and decryption operation in public key cryptography.
encryption and decryption operation in public key cryptography.
_3.3. Correctness Analysis_
Now we prove the correctness of our encryption scheme in terms of spatial and temporal
aggregation:
-----
_Appl. Sci. 2019, 9, 490_ 11 of 20
_3.3. Correctness Analysis_
Now we prove the correctness of our encryption scheme in terms of spatial and
temporal aggregation:
3.3.1. Spatial Aggregation
_n_
_ASd_ = ∑ _LS(i, d) modM_
_i=1_
_n_
= ∑ ∑ (x([j]i,d) [+][ r]([j]i,d)[)][ mod][ M][ (][1][ ≤] _[s][ ≤]_ _[k][)]_
_i=1_ _j∈indi[s]_
_n_ _n_
= ( ∑ ∑ _x([j]i,d)_ [+] ∑ ∑ _r([j]i,d)[)][ mod][ M]_
_i=1_ _j∈indi[s]_ _i=1_ _j∈indi[s]_
_n_ _n_ _n_
= ( ∑ ∑ _x([j]i,d)_ [+][ 1]2 [(] ∑ ∑ _r([j]i,d)_ [+] ∑ ∑ _r([j]i,d)[)][ mod][ M]_
_j=1_ _i∈indj[s]_ _i=1_ _j∈indi[s]_ _j=1_ _i∈indj[s]_
_n_ _n_ _n_
= ( ∑ _xd[j]_ [+][ 1]2 [(] ∑ ∑ _r([j]i,d)_ ∑ ∑ _r([j]i,d)[)][ mod][ M]_
_j=1_ _i=1_ _j∈indi[s]_ _[−]_ _i=1_ _j∈indi[s]_
_n_
= ∑ _xd[j]_ [mod][ M]
_j=1_
(11)
We prove the correctness of our spatial aggregation by permuting the row and column of data
matrix formed as Figure 2. Equation (11) shows that the spatial aggregation in a neighborhood equals
to the sum of locally spatial aggregation, i.e., the sum of individual data.
3.3.2. Temporal Aggregation
_AT(i, T)_ = ∑ _LT([i]j,d)_ [mod][ M][(][1][ ≤] _[s][ ≤]_ _[k][)]_
_j∈indi[s]_
_T_
= ∑ ∑ _E([i]_ _j,d)_ [mod][ M]
_j∈indi[s]_ _d=1_
_T_
= ∑ ∑ (x([i] _j,d)_ [+][ r]([i] _j,d)[)][ mod][ M]_
_d=1_ _j∈indi[s]_
_T_
= ∑ _xd[i]_ [mod][ M]
_d=1_
(12)
Equation (12) shows that the temporal aggregation for one user node equals to the sum of
local temporal aggregation from its pairwise nodes, i.e., the sum of its individual data in a series of
timestamps T. It proves further the correctness of our temporal aggregation.
**4. Security Notions**
_4.1. Security Proof_
In this section, we mainly elaborate the security properties of our scheme. In particular, based on
the security requirement and attack model discussed in Section 2.6, we prove our scheme can ensure
the confidentiality of fine-grained meterings for an individual user and the aggregation integrity that
the local aggregation, and total aggregation cannot tampered by malicious individual user nodes or
the aggregator.
We firstly construct the Individual Metering Indistinguishable (IMI) security game to represent
the adversary’s actions.
**Definition 1. (IMI security game).**
**_Setup: the challenger runs the initialization algorithm and first initializes a group of size n, then generates_**
_the system parameter k to the adversary._
-----
_Appl. Sci. 2019, 9, 490_ 12 of 20
**_Queries: the adversary can not only capture meters’ encrypted report but also acquire the encryption and_**
_compromise queries until meeting the constraints._
_Encrypt: The adversary A chooses useri and specifies xd[i]_ _[to ask for the ciphertext. The challenger returns it]_
_the ciphertext E(xd[i]_ [)][.]
_Compromise: The adversary A specifies an integer q ∈{0, 1, · · ·, n}. If q = 0, the challenger returns the_
_adversary the aggregator’ capability, else returns userq’s message._
_Challenge. We denote with {C} the set of the uncompromised users. The adversary selects randomly two_
_meterings x[i][0]_
_d_ _[and][ x]d[i][1]_ [(][i][ ∈{][C][}][)][ at the timestamp d. The challenger flips a random bit][ b][ ∈{][0, 1][}][ uniformly and]
_returns the adversary E(xd[i][b]_ [)][.]
**_Guess: The adversary outputs a guess b[′]_** _∈{0, 1}, and A wins if b = b[’]_ _with unignorable advantage._
**Definition 2. (IMI security)**
_The proposed temporal and spatial aggregation scheme is IMI if no probabilistic polynomial-time adversaries_
_A have more than an ignorable advantage in the IMI security game. The ignorable function for A is as follows:_
_AdvA = |Pr[b = b[′]] −_ [1] (13)
2 _[|][ =][ 0]_
**Theorem 1. The proposed encryption scheme is IMI.**
_The intuition behind the theorem is any adversary cannot distinguish the encrypted individual metering_
_and the scheme cannot leak any individual user consumption at the d[th]_ _timestamp._
**Proof:**
**Setup: The challenger initiates the whole system. The challenger generates a group of scale n and**
pairs number k, and then gives the parameters (n, k) to the adversary.
**Queries:**
(1). Spatial aggregation
Encrypt: A issues the encryption query with (i, d, xd[i] [)][ to the challenger. The challenger generates]
the pairwise key r[i]
(j,d)[(][j][ ∈] _[ind][i][[][s][])][ between the pairwise nodes, and imposes it on the randomly]_
partitioned data x([i] _j,d)[(][j][ ∈]_ _[ind][i][[][s][])][ to generate the encrypted measure formed as][ E][(][x]d[i]_ [) =][ x]([i] _j,d)_ [+]
_r[i]_
(j,d)[mod][ M][(][j][ ∈] _[ind][i][[][s][])][.]_
Compromise: A may compromise the aggregator or up to n-1 users in any pairwise set in order to
acquire more messages for object users. However, the compromise will encounter restriction when
meeting with uncompromised users.
Challenge. For simplifying the proof process and not losing the generalization, we consider the
extreme circumstance that _c_ = 2. If the theorem holds for this circumstance, then it holds for _c_ _> 2._
_|_ _|_ _|_ _|_
We assume the user j is the only uncompromised user in indi[s](1 ≤ _s ≤_ _k). The adversary selects the_
two meterings and gives (i, j, d, xd[i][0][,][ x]d[i][1] [)][ to the challenger, the challenger flips a random bit][ b][ ∈{][0, 1][}]
uniformly and returns the adversary E(xd[i] [)][ when][ b][ = 0, and returns][ E][(][x]d[j] [)][ when][ b][ = 1, and then]
_E(xd[i]_ [)] = ∑ (x([i] _l,d)_ [+][ r]([i] _l,d)[)][ mod][ M][(][1][ ≤]_ _[s][ ≤]_ _[k][)]_
_l∈indi[s]_ (14)
= (x([i] _j,d)[+][r][i](j,d)_ [+] ∑ (x([i] _c,d)[+][r]([i]_ _c,d)[)][ mod][ M]_
_c∈{indi[s]−j}_
_E(xd[j]_ [)] = ∑ (x([j] _l,d)_ [+][ r]([j] _l,d)[)][ mod][ M][(][1][ ≤]_ _[s][ ≤]_ _[k][)]_
_l∈indj[s]_ (15)
= (x([j]i,d)[+][r]([j] _i,d)_ [+] ∑ (x([j] _c,d)[+][r]([j]_ _c,d)[)][ mod][ M]_
_c∈{indj[s]−i}_
-----
_Appl. Sci. 2019, 9, 490_ 13 of 20
In the Equations (14) and (15), the adversary A cannot solve the two equations at the d[th] timestamp
and gain the exact x[i]
(j,d) [and even if he knows][ r]([j]i,d) [=][ −][r]([i] _j,d)[, as the two equations have three unknown]_
variables, so it is more impossible for A to acquire xd[i] [and][ x]d[j] [which ensures the scheme’s security.]
(2). Temporal aggregation
_T_ _T_
_E(_ ∑ _xd[i]_ [)] = ∑ ∑ (x([i] _l,d)_ [+][ r]([i] _l,d)[)][ mod][ M][(][1][ ≤]_ _[s][ ≤]_ _[k][)]_
_d=1_ _d=1_ _l∈indi[s]_
_T_ _T_
= ( ∑ (x([i] _j,d)_ [+][ r]([i] _j,d)[) +]_ ∑ ∑ (x([i] _c,d)_ [+][ r]([i] _c,d)[))][ mod][ M]_
_d=1_ _d=1_ _c∈{indi[s]−j}_
_T−1_ _T_
= ((x([i] _j,T)_ [+][ r]([i] _j,T)[) +]_ ∑ (x([i] _j,d)_ [+][ r]([i] _j,d)[) +]_ ∑ ∑ (x([i] _c,d)_ [+][ r]([i] _c,d)[))][ mod][ M]_
_d=1_ _d=1_ _c∈{indi[s]−j}_
_T−1_ _T_
= (x([i] _j,T)_ [+] ∑ _x([i]_ _j,d)_ [+] ∑ ∑ (x([i] _c,d)_ [+][ r]([i] _c,d)[))][ mod][ M]_
_d=1_ _d=1_ _c∈{indi[s]−j}_
(16)
_T_ _T−1_ _T_
_E(_ ∑ _xd[j]_ [)= (][x]([j]i,T) [+] ∑ _x([j]i,d)_ [+] ∑ ∑ (x([j] _c,d)_ [+][ r]([j] _c,d)[))][ mod][ M]_ (17)
_d=1_ _d=1_ _d=1_ _c∈{indj[s]−i}_
In the Equations (16) and (17), the two equations with four unknown variables make the adversary
_A impossible to acquire x[i]_
(j,d) [or][ x]([j]i,d)[.]
Hence, the encrypted aggregation method can ensure the individual, fine-grained meterings
indistinguishable security as long as there is at least one uncompromised user in its pairwise set. Our
security properties are based on the randomness of modular addition and stream cipher which is used
to blind the individual meterings.
_4.2. Security Analysis_
We can prove that our proposed solution will withstand the other attacks discussed in Section 2.6
and ensure the integrity of the aggregated data, whether total aggregation or local aggregation.
(1). Eavesdropping resistance
Our proposed scheme supports the openness of communication flow. Whether it is the internal
node with access to the communication flow in a community or the external eavesdropper, they can
only get the encrypted individual data (x[i]
(j,d) [+][ r]([i] _j,d)[)][, local aggregation value][ (][LS][(][i][,][ d][)][,][ LT]([i]j,d)[)][ or]_
total aggregation value (AS(d), AT(i, T)) sent by GW to CC. However, all of them can not obtain the
fine-grained data. We have proved that even if all but one node is compromised, object metering still
cannot be leaked. Hence, the proposed encryption method satisfies the security requirement R1.
(2). False command from the GW
The GW attempts to obtain object user’s meterings by issuing false billing commands in the name
of CC, even if he cannot compromise its pairwise nodes. He tries to obtain valuable information from
them at any timestamp. However, even so, he can only get the indistinguishable, individual meterings,
due to the Equations (14)–(17).
We cannot exclude the possibility that all pairwise keys of useri at a timestamp are all compromised
nodes by the malicious aggregator or external attackers. In this case, the object useri’s privacy is
exposed. That is the useri does not select any one honest node, then the probability is 1 − ( _n−k_ 1 [)]n−1−|c|.
Obviously, the larger the value of |c| is, the smaller the value of k is, and the bigger the probability
is. We improve the probability as much as possible and assume n = 1000, k = 30, and |c| = 500 (50%
nodes are compromised), and then the probability is 2.47 × 10[−][7], so much small probability implies it
-----
_Appl. Sci. 2019, 9, 490_ 14 of 20
is almost impossible that one user does not select any one honest node in one timestamp. Even if we fix
a bigger pairs period T = 1 month, then we would have to cost 38.51 years to acquire individual data.
_4.3. Publicly Verifiable Property_
The security requirement R2 given earlier needs to be satisfied with the publicly verifiable property.
We provide the public communication flow between nodes in a neighborhood is to ensure the integrity
of aggregation data. Any internal node in the community can verify publicly the accuracy of the
local aggregation from other nodes and the total aggregation from the GW without compromising the
individual fine-grained data. The special public verification process comprises two parties:
4.3.1. Spatial Verification
Based on the public communication flow, any node in the neighborhood can gain the encrypted
message formed as x[i]
(j,d) [+][ r]([i] _j,d)_ [from the pairwise nodes, and compute its local aggregation formed]
as LS(i,d) and LT[i]
(j,d)[, and thus the total aggregation][ AS][d][ and][ AT][(][i][,][ T][)][ for the neighborhood can be]
computed and compared with the reported result from the GW. If the result is questionable, the user
can report directly to the CC. With such a supervision, the CC can detect the fraudulent profile of the
malicious GW.
4.3.2. Temporal Verification
The public verification method to the spatial aggregation is equally effective to the temporal
verification. For any node, one of its pairwise nodes in the neighborhood gain its encrypted message
formed as x[i]
(j,d) [+][ r]([i] _j,d)_ [in a billing period before computing its local temporal aggregation, and thus its]
total temporal aggregation is computed and verified by summing up local temporal aggregations from
all its pairwise nodes.
Thus, the billing user itself or any user node can verify the accuracy of the billing from the GW
without revealing individual fine-grained data. Hence, they can detect if there is a malicious and
fraudulent profile of the malicious GW and reports it to the CC in time.
**5. Performance Evaluation**
We evaluate the performance of the proposed aggregation scheme to assess the overheads. The
performance metrics used in our empirical evaluation are defined as follows:
(1) Computation overhead: node’s runtime of the proposed scheme in terms of spatial and
temporal aggregation.
(2) Communication overhead: the size of a message transmitted between the nodes and GW (number
of bits).
(3) Security parameter k: we analyze the impact of the different value of k on the two overheads.
We compare these results against several existing works [23,24] using performance metrics based
on Friendly ARM [25] and the library in [17]. By comparison with them, we intend to illustrate our
computing and communication advantages in terms of the combination of PRF and modular addition
methods adopted, respectively, in the scheme [23] and [24]. Each experiment consists of 50 independent
trials and the averaged results of these trials are reported. The computation time required for these
tasks is listed in Table 2.
-----
_Appl. Sci. 2019, 9, 490_ 15 of 20
**Table 2. Average time for functions.**
**Notations** **Descriptions** **Time Cost**
Cadd Addition _≈0.038 ms_
Chash Hash (100 randoms) _≈0.85 ms_
Cmul Multiplication _≈0.013 ms_
Chenc Homomorphic encryption _≈2.7 ms_
Chdec Homomorphic decryption _≈0.61 ms_
Cma Hash/Modular addition _≈0.0023 ms_
Cprf Pseudorandom function _≈0.074 ms_
We fix the number of users at 1 million; the number of C is 10; the number of GW ranges from 1 to
20. Let n denotes a possible number of users in a group, and it ranges from 1 to 5000. We present the
impact of a different number of users in the GW and a different value of k (ranging from 1 to 100) on
the performance. We also assume, for simplicity, that all SMs can be functioning normally.
_5.1. Computation Overhead_
(1). Spatial aggregation
Let Cma and Cprf denote respectively the cost of Modular addition operation and keys generation
operation with PRF, respectively let Cadd and Cmul denote the cost of addition and multiplication
operation respectively, and Cenc and Cdec denote the cost of homomorphic encryption and decryption
operation respectively.
In our spatial aggregation scheme, for every node, partitioning individual data into k partitions
costs one Cma; generating k pairwise keys costs k·Cprf; receiving k encrypted messages and adding
them up cost k·Cadd, then the computation overhead per node is Cma + k·Cprf + k·Cadd and the total
computation overhead per aggregator is (n-1)·Cadd for aggregating data from n nodes.
In Erkin et al.’s scheme [23], at the d[th] time step, every hash function cost is Chash, k masking
random keys cost is k·Cpr f and computing total masking keys cost is 2k·Cadd, and then encrypting
individual data cost is Cenc, so the total computation overhead is Chash + k·Cpr f + 2k·Cadd + Cenc.
In Jia et al.’s scheme [24], at the d[th] time step, the additive secret sharing cost is Css, k hash
functions cost is k·Chash, and then k-order polynomial operation is x[k] and k matrix multiplication
operations cost is (k[2] + 2k)·Cmul, so the total computation overhead is: Css + k·Chash + (k[2] + 2k)·Cmul.
We provide the individual spatial computation overhead comparison in Table 3.
**Table 3. Individual spatial computation overhead comparison (msec).**
**Scheme** **Computation Overhead Per Smart Meter**
Scheme in [23] _Chash + k·Cpr f + 2k·Cadd + Cenc_
Scheme in [24] _Css + k·Chash +_ �k[2] + 2k�·Cmul
Our scheme Cma + k·Cprf + k·Cadd
As described in the related work, the scheme in Reference [23] sets all nodes as communication
nodes instead of selecting a limited number of communication nodes as in ours and [22]; however,
for convenient comparison, we assume that k communication nodes are selected, which is on the
same experiment platform as ours and the scheme in [23]. Even under such relaxation, we can
still prove ours is superior in terms of computation and communication cost through the following
performance evaluation.
The Figure 4 plots the comparison of spatial computation overhead between our scheme and
the schemes in References [23,24] with the value of k increasing. The Figure 4 shows that the three
schemes’ computation overheads all increase with the value of k increasing, the computation overhead
in Reference [23] and ours are lower compared with the scheme in References [24], in which polynomial
-----
p g, p
_Appl. Sci.overhead in Reference [23] and ours are lower compared with the scheme in References [24], in 2019, 9, 490_ 16 of 20
which polynomial operation _x_ _k_ and _k matrix multiplication operations generate too much_
computation overhead with _k growing, it has more cost significantly than ours and Erkin et al.'s_
operation x[k] and k matrix multiplication operations generate too much computation overhead with k
scheme [23], ours is lower slightly than the scheme in [23], and both of them are close to
growing, it has more cost significantly than ours and Erkin et al.’s scheme [23], ours is lower slightly
than the scheme in [O k C( )⋅ _prf_ . 23], and both of them are close to O(k)·Cpr f .
300
250
200
150
100
50
0
|Col1|Col2|Col3|Col4|Col5|[23] ] me|Col7|Col8|Col9|Col10|Col11|
|---|---|---|---|---|---|---|---|---|---|---|
|||Erki Jia e The|n et al. t al.sc propos|scheme heme[24 ed sche|[23] ] me||||||
||||||||||||
||||||||||||
||||||||||||
||||||||||||
||||||||||||
||||||||||||
5 10 15 20 25 30 35 40 45 50
Different k Value
**Figure 4. Variety of spatial computation overhead per node with the value k.**
**Figure 4. Variety of spatial computation overhead per node with the value k.**
(2). Temporal aggregation
(2). Temporal aggregation
In the proposed scheme, each node chooses the same nodes every billing period to satisfy with
In the proposed scheme, each node chooses the same nodes every billing period to satisfy with
the Equation (5), so total temporal computation overhead in T serial time slots for every node is
the Equation (5), so total temporal computation overhead in _T serial time slots for every node is_
_T·T(k⋅(·In Erkin et al.’s scheme [k CC⋅pr fprf ++ ⋅ kk C·Caddadd ++ CCmama)_ ) ++ ⋅T C T23·addC], each node sendsadd. . _T fine-grained utility readings in each of the T time_
In Erkin et al.'s scheme [23], each node sends T fine-grained utility readings in each of the T time
steps, so the overhead per node is T·(Chash + k·Cpr f + 2k·Cadd + Cenc) + T·Cmul. In fact, the temporal
aggregation overhead of the scheme in Reference [steps, so the overhead per node is _T_ ⋅(Chash + ⋅k C23prf] is higher than it, as with the modification of+ 2k C⋅ _add_ + _Cenc)_ +T C⋅ _mul_ . In fact, the temporal
Paillier encryption, spatial and temporal aggregations are not being synchronized. To compensate theaggregation overhead of the scheme in Reference [23] is higher than it, as with the modification of
_r[n]_
lack, every user must add an additional random keyPaillier encryption, spatial and temporal aggregations are not being synchronized. To compensate the R(i,T+1) = _R(i,d)_ at T[th] timestamp, which
∏n _d[T]=1_ _[h]d_
costs much overhead. However, our scheme has no extra cost and the third party’s involvement.lack, every user must add an additional random key _R(,i T_ +1) = _r_ ∏Td =1hdR(, )i d at _T[th ]timestamp, which_
We set the fine-grained reporting interval to be 15 minutes, and billing period T = 2880 (roughly
one month). Figurecosts much overhead. However, our scheme has no extra cost and the third party's involvement. 5 plots the comparison of two schemes in terms of temporal computation overhead
in every billing period forWe set the fine-grained reporting interval to be 15 minutes, and billing period k ranging from 0 to 50. From Figure 5, we can see the temporal computationT = 2880 (roughly
overhead per node grows with the increasing ofone month). Figure 5 plots the comparison of two schemes in terms of temporal computation k value in two schemes; however, our proposed scheme
increases slightly compared with the scheme in References [overhead in every billing period for k ranging from 0 to 50. From Figure 5, we can see the temporal 23], as the latter costs much overhead on
Paillier encryption, while our scheme achieves the same privacy protection effect as the asymmetriccomputation overhead per node grows with the increasing of k value in two schemes; however, our
proposed scheme increases slightly compared with the scheme in References [23], as the latter costs
encryption with simple and low-cost modular addition.
much overhead on Paillier encryption, while our scheme achieves the same privacy protection effect
as the asymmetric encryption with simple and low-cost modular addition.
-----
_Appl. Sci.Appl. Sci. 20192018, 9,, 4908, x FOR PEER REVIEW_ 17 of 2018 of 22
300
250
200
150
100
50
0
|Col1|Col2|Col3|Col4|Col5|[23] me|Col7|Col8|Col9|Col10|Col11|
|---|---|---|---|---|---|---|---|---|---|---|
|||Erki The|n et al. propos|scheme ed sche|[23] me||||||
||||||||||||
||||||||||||
||||||||||||
||||||||||||
||||||||||||
||||||||||||
5 10 15 20 25 30 35 40 45 50
Different k Value
**Figure 5. Variety of temporal computation overhead per node with the value k.**
**Figure 5. Variety of temporal computation overhead per node with the value k.**
_5.2. Communication Overhead_
_5.2. Communication Overhead_
We assume the format of a packet is the same as that in TinyOS [26]. The timestamp occupies 128
We assume the format of a packet is the same as that in TinyOS [26]. The timestamp occupies
bits. The sizes of prime numbers p, and q needed in the Paillier encryption are 512 bits each. The size
128 bits. The sizes of prime numbers p, and q needed in the Paillier encryption are 512 bits each. The
of elements in Zn[∗] [is 1024 bits. We further assume the plaintext data occupies 32 bits, then random from]*
stream cipher occupies the same byte width with the plaintext data, and Paillier encryption occupiessize of elements in _[Z]n[ is 1024 bits. We further assume the plaintext data occupies 32 bits, then ]_
4096 bits, while the hash function with timestamp occupies 256 bits.random from stream cipher occupies the same byte width with the plaintext data, and Paillier
encryption occupies 4096 bits, while the hash function with timestamp occupies 256 bits. For simplicity, we denote {|X|, |R|, |E|, |H|} as the plaintext data size, masking random key (noise)
size, Paillier encryption size, and the size of hash function random.For simplicity, we denote { _X_, _R_, _E_, _H_ } as the plaintext data size, masking random key
(noise) size, Paillier encryption size, and the size of hash function random.
5.2.1. Spatial Communication Overhead Per Node
5.2.1. Spatial Communication Overhead Per Node To generate the spatial aggregation, every node sends the local aggregation to the GW after adding
up the encrypted message from all k pairs. The data sent per node can be denoted as _LS(i, d)_ _t_, the
_{_ _∥_ _}_
To generate the spatial aggregation, every node sends the local aggregation to the GW after
size isadding up the encrypted message from all |X| + k·|R| + 128 bits (a partitioned part size isk pairs. The data sent per node can be denoted as [|][x]k[|] [bits,][ k][ partitions take][ |][x][|][ bits; a noise key]
takes _R_ bits, and then k noise keys take k _R_ bits), so the total packet size is _x_ + k _R_ + 128 bits.
_|_ _|_ _·|_ _|_ _|_ _|_ _·|_ _|_
###### {LS i dFor the scheme in Reference [(, ) }t, the size is X + ⋅k R 23+128], the spatial aggregation packet per node is in the form as bits (a partitioned part size is [x]
k [ bits, ][k][ partitions take ]
_E_ _H_ _R_ _t_, its size is _E_ + _X_ + k _R_ + _H_ + 128 bits.
_{_ _∥_ _∥_ _∥_ _}_ _{|_ _|_ _|_ _|_ _·|_ _|_ _|_ _|_ _}_
###### x bits; a noise key takes R bits, and then k noise keys take k R⋅ bits), so the total packet size is
Every user node in Reference [24] generates k results, the data is in the form of {(y1∥y2∥· · · ∥yk)∥t},
in whichx + ⋅k R yi involves the computation of data sharing and hash random value, so its size is+128 bits. _K·(K·|H| +_
_X_ /k + 128) bits.
_|_ _|_ For the scheme in Reference [23], the spatial aggregation packet per node is in the form
as{E H R tWe provide the individual spatial communication overhead comparison in Table }, its size is { _E_ + _X_ + ⋅k R + _H_ +128} bits. 4.
Every user node in Reference[24] generates Table 4. Individual spatial communication overhead comparison (bits).k results, the data is in the form of
###### {(y y1 2 y ) }k [t], in whichScheme[y]i[ involves the computation of data sharing and hash random value, so ]Computation Overhead Per Smart Meter
|use|er node in Reference[24] generates k results, the data is in Table 4. Individual spatial communication overhead comparison (bits).|
|---|---|
|||
|y ) k|t}, in whichy involves the computation of data sharing and hash ran Schemei Computation Overhead Per Smart Meter|
its size is _K_ ⋅(K H⋅ Scheme in [+ _X_ / _k_ +128)23] bits. _|E| + |X| + k·|R| + |H| + 128_
We provide the individual spatial communication overhead comparison in Table 4. Scheme in [24] _K·(K·|H| + |X|/k + 128)_
Our scheme _|X| + k·|R| + 128_
We plot the individual communication overhead comparison between our scheme and the other
two schemes [23,24] during spatial aggregation in the Figure 6. We can see clearly the three schemes’
individual overhead all grow with the increasing of k value. The packet width per node in the
scheme in Reference [24] grows significantly than the other two schemes, especially when k value is
relatively higher, and communication overhead closes to O(k[2]), due to the x[k] polynomial operation per
|K⋅(K ⋅|HS|c+h|eXm|/e kin+ [12238]) bits. |E| + |X| + k·|R| + |H| + 128|
|---|---|---|---|---|
-----
**Scheme** **Computation Overhead Per Smart Meter**
_Appl. Sci. 2019, 9, 490_ 18 of 20
Scheme in [23] _E_ + _X_ + ⋅k R + _H_ +128
node before the matrix multiplication operation. Our scheme’s growth rate is close to the scheme inScheme in [24] _K_ ⋅(K H⋅ + _X_ / _k_ +128)
Reference [23], which is higher always slightly higher than ours, due to the relatively higher public
key encryption width.Our scheme _X_ + ⋅k R +128
x 104
3
2.5
2
1.5
1
0.5
0
|Col1|Col2|
|---|---|
|Erkin et al.scheme [23] Jia et al.scheme [24]||
|The proposed scheme||
|||
|||
|||
|||
|||
5 10 15 20 25 30 35 40 45 50
Different k Value
**Figure 6. Variety of spatial communication overhead per node with the value k.**
**Figure 6. Variety of spatial communication overhead per node with the value k.**
_Appl. Sci. 2018, 8, x FOR PEER REVIEW_ 20 of 22
5.2.2. Temporal Communication Overhead Per Node
We plot the individual communication overhead comparison between our scheme and the
from PRF saves much computation and communication overhead compared with traditional public Figure 7 shows the comparison result of ours and the scheme [23] in terms of temporal
other two schemes [23,24] during spatial aggregation in the Figure 6. We can see clearly the three
communication overhead per node whenkey encryption without compromising individual privacy. k ranges from 0 to 600, and T ranges from 0 to 6000 mins.
schemes' individual overhead all grow with the increasing of k value. The packet width per node in
the scheme in Reference [24] grows significantly than the other two schemes, especially when _k_
value is relatively higher, and communication overhead closes to O(k[2]), due to the x[k] polynomial
operation per node before the matrix multiplication operation. Our scheme's growth rate is close to 6
x 10
the scheme in Reference [23], which is higher always slightly higher than ours, due to the relatively
10 10000
higher public key encryption width.
8 8000
5.2.2. Temporal Communication Overhead Per Node
6 6000
Figure 7 shows the comparison result of ours and the scheme [23] in terms of temporal
communication overhead per node when 4 _k ranges from 0 to 600, and 4000_ _T ranges from 0 to 6000 mins,_
In Figure 7, our scheme reduces significantly the packet size sent per node to almost three
orders of magnitude than the scheme [23], due to the high overhead of public key encryption. 2 2000
During temporal aggregation, if the process of exchanging random between communication nodes
0 0
is ignorable, then every node sends its serial encrypted packet formed as {E H R t }(1≤≤t _T)_ to
the aggregator, so the packet size is 400 _T_ ⋅( _E6000+_ _X_ + ⋅k R400 + _H_ +128) bits, while in our scheme, one 6000
4000 4000
node's temporal aggregation is computed synchronously200 200before being reported to the aggregator by
2000 2000
_k communication nodes, and they sends the local temporal aggregation packet size of k_ 0 0 T k 0 0 T
###### x + ⋅k R +128 bits to the aggregator every (a) Erkin et al.scheme [23] T timeslot, so aggregating one node's temporal (b) The proposed scheme
consumption in _T serial time slots costs_ _k_ ⋅ ( _x_ + _k R⋅_ +128) bits. Hence, when _k_ _T_, ours
**Figure 7. Variety of temporal communication overhead per node with the values k and T.**
overhead is always lower significantly lower than the scheme in Reference [23]. Just as the Figure 7. Variety of temporal communication overhead per node with the values k and T.
In Figure 7, our scheme reduces significantly the packet size sent per node to almost three orders of
description above, we shorten the number of the communication nodes in Reference [23] into k, and
magnitude than the scheme [6. Conclusions 23], due to the high overhead of public key encryption. During temporal
the performance evaluation shows the proposed collection of modular addition and masking keys
aggregation, if the process of exchanging random between communication nodes is ignorable, then
In the paper, we resolved three issues about privacy-protection aggregation of smart metering
every node sends its serial encrypted packet formed as _E_ _H_ _R_ _t_ (1 _t_ _T) to the aggregator,_
customized to the SG. Firstly, the combination of simple modular addition and PRF we designed { _∥_ _∥_ _∥_ _}_ _≤_ _≤_
so the packet size is T ( _E_ + _X_ + k _R_ + _H_ + 128) bits, while in our scheme, one node’s temporal
serves the same effect as the other most related works with lower overhead, namely fending off · _|_ _|_ _|_ _|_ _·|_ _|_ _|_ _|_
aggregation is computed synchronously before being reported to the aggregator by k communication
maliciously internal attacks without compromising individual fine-grained data. Secondly, we
nodes, and they sends the local temporal aggregation packet size of _x_ + k _R_ + 128 bits to the
proposed innovatively a publicly verifiable platform, by which, every node in a neighborhood can | _|_ _·|_ _|_
verify local aggregation from every node and total aggregation from the GW and detect the
-----
_Appl. Sci. 2019, 9, 490_ 19 of 20
aggregator every T timeslot, so aggregating one node’s temporal consumption in T serial time slots
costs k ( _x_ + k _R_ + 128) bits. Hence, when k _T, ours overhead is always lower significantly_
_·_ _|_ _|_ _·|_ _|_ _≪_
lower than the scheme in Reference [23]. Just as the description above, we shorten the number
of the communication nodes in Reference [23] into k, and the performance evaluation shows the
proposed collection of modular addition and masking keys from PRF saves much computation and
communication overhead compared with traditional public key encryption without compromising
individual privacy.
**6. Conclusions**
In the paper, we resolved three issues about privacy-protection aggregation of smart metering
customized to the SG. Firstly, the combination of simple modular addition and PRF we designed
serves the same effect as the other most related works with lower overhead, namely fending off
maliciously internal attacks without compromising individual fine-grained data. Secondly, we
proposed innovatively a publicly verifiable platform, by which, every node in a neighborhood
can verify local aggregation from every node and total aggregation from the GW and detect the
fraudulent profiles from maliciously internal nodes or dishonest user nodes. Thirdly, every node
chooses randomly k nodes rather than all nodes as pairwise nodes to communicate, which saves
significantly communication and computation overhead, and the independence of the number of users
provides scalability and high efficiency under the circumstance of SG big data. From the performance
evaluation shows that the proposed scheme is applicable for the security and privacy protection of SG
and has practical significance.
**Author Contributions: L.Z. and J.Z. designed the hierarchical architecture model, attack models, communication**
models, and encryption methods together; J.Z. optimized the communication models, and L.Z. wrote the paper.
**Funding: This research was funded by National Natural Science Foundation of China (NSFC) (2017–2020,**
No. 51679058).
**Conflicts of Interest: The authors declare no conflicts of interest.**
**References**
1. Wang, W.; Lu, Z. Cyber security in the Smart Grid: Survey and challenges. Comput. Netw. 2013, 57, 1344–1371.
[[CrossRef]](http://dx.doi.org/10.1016/j.comnet.2012.12.017)
2. Ambrosin, M.; Hosseini, H.; Mandal, K. Despicable meter: Anonymous and fine-grained metering data
reporting with dishonest meters. In Proceedings of the 2016 IEEE Conference on Communications and
Network Security (CNS 2016), Philadelphia, PA, USA, 17–19 October 2016; pp. 163–171.
3. [Krebs, B. FBI: Smart Meter Hacks Likely to Spread. 2013. Available online: http://krebsonsecurity.com/](http://krebsonsecurity.com/2012/04/fbi-smart-meter-hackslikely-to-spread/)
[2012/04/fbi-smart-meter-hackslikely-to-spread/ (accessed on 07 April 2012).](http://krebsonsecurity.com/2012/04/fbi-smart-meter-hackslikely-to-spread/)
4. Ohara, K.; Sakai, Y.; Yoshida, F.; Iwamoto, M.; Ohta, K. Privacy-preserving smart metering with verifiability
for both billing and energy management. In Proceedings of the 2nd ACM Workshop on ASIA Public-Key
Cryptography (ASIAPKC’14), Kyoto, Japan, 3–6 June 2014; pp. 23–32.
5. Lincoln, K.; Philip, K.; Christopher, M. The use of RC4 encryption for smart meters. In Proceedings of
the 2014 International Conference on Sustainable Research and Innovation, Nairobi, Kenya, 7–9 May 2014;
pp. 58–62.
6. Lu, R.; Liang, X.; Li, X.; Lin, X.; Shen, X. EPPA: An efficient and privacy-preserving aggregation scheme for
[secure smart grid communications. IEEE Trans. Parallel Distrib. Syst. 2012, 23, 1621–1631. [CrossRef]](http://dx.doi.org/10.1109/TPDS.2012.86)
7. Garcia, F.D.; Jacobs, B. Privacy-friendly energy-metering via homomorphic encryption. _IEEE Trans._
_[Parallel Distrib. 2010, 6710, 226–238. [CrossRef]](http://dx.doi.org/10.1007/978-3-642-22444-7_15)_
8. Wang, X.; Mu, Y.; Chen, R. An efficient privacy-preserving aggregation and billing protocol for smart grid.
_[Secur. Commun. Netw. 2016, 9, 4536–4547. [CrossRef]](http://dx.doi.org/10.1002/sec.1645)_
9. Chen, L.; Lu, R.; Cao, Z. PDAFT: A privacy-preserving data aggregation scheme with fault tolerance for
[smart grid communications. Peer-to-Peer Netw. Appl. 2015, 8, 1122–1132. [CrossRef]](http://dx.doi.org/10.1007/s12083-014-0255-5)
-----
_Appl. Sci. 2019, 9, 490_ 20 of 20
10. He, D.; Kumar, N.; Zeadally, S. Efficient and Privacy-Preserving Data Aggregation Scheme for Smart Grid
[against Internal Adversaries. IEEE Trans. Smart Grid 2017, 8, 2411–2419. [CrossRef]](http://dx.doi.org/10.1109/TSG.2017.2720159)
11. Bao, H.; Lu, R. A New Differentially Private Data Aggregation with Fault Tolerance for Smart Grid
[Communications. IEEE Internet Things J. 2015, 2, 248–258. [CrossRef]](http://dx.doi.org/10.1109/JIOT.2015.2412552)
12. Kursawe, K.; Danezis, G.; Kohlweiss, M. Privacy-friendly aggregation for the smart-grid. In Proceedings
of the International Symposium on Privacy Enhancing Technologies Symposium, Cambridge, UK,
27–29 July 2011; pp. 175–191.
13. Shi, Z.; Sun, R.; Lu RChen, L.; Chen, J.; Shen, X.S. Diverse grouping-based aggregation protocol with error
[detection for smart grid communications. IEEE Trans. Smart Grid 2015, 6, 2856–2868. [CrossRef]](http://dx.doi.org/10.1109/TSG.2015.2443011)
14. Gupta, S.S.; Maitra, S.; Paul, G.; Sarkar, S. (Non-)Random Sequences from (Non-)Random
[Permutations—Analysis of RC4 Stream Cipher. J. Cryptol. 2014, 27, 67–108. [CrossRef]](http://dx.doi.org/10.1007/s00145-012-9138-1)
15. Mahmood, K.; Chaudhry, S.A.; Naqvi, H.; Shon, T.; Ahmad, H.F. A lightweight message authentication
[scheme for smart grid communications in power sector. Comput. Electr. Eng. 2016, 52, 114–124. [CrossRef]](http://dx.doi.org/10.1016/j.compeleceng.2016.02.017)
16. Li, H.; Lu, R.; Zhou, L.; Yang, B.; Shen, X. An Efficient Merkle-Tree-Based Authentication Scheme for Smart
[Grid. IEEE Syst. J. 2014, 8, 655–663. [CrossRef]](http://dx.doi.org/10.1109/JSYST.2013.2271537)
17. Chim, T.W.; Yiu, S.M.; Li, V.O.K. PRGA: Privacy-Preserving Recording Gateway-Assisted Authentication of
[Power Usage Information for Smart Grid. IEEE Trans. Depend. Secur. Comput. 2015, 12, 85–97. [CrossRef]](http://dx.doi.org/10.1109/TDSC.2014.2313861)
18. Fan, C.I.; Huang, S.Y.; Lai, Y.L. Privacy-Enhanced Data Aggregation Scheme Against Internal Attackers in
[Smart Grid. IEEE Trans. Ind. Inform. 2013, 10, 666–675. [CrossRef]](http://dx.doi.org/10.1109/TII.2013.2277938)
19. Castelluccia, C.; Mykletun, E.; Tsudik, G. Efficient Aggregation of encrypted data in Wireless Sensor
Networks. In Proceedings of the 2th International Conference on Mobile and Ubiquitous Systems:
Networking and Services (MOBIQUITOUS’05), San Diego, CA, USA, 17–21 July 2005; pp. 109–117.
20. Dimitriou, T.; Awad, M.K. Secure and scalable aggregation in the smart grid resilient against malicious
[entities. Ad Hoc Netw. 2016, 50, 58–67. [CrossRef]](http://dx.doi.org/10.1016/j.adhoc.2016.06.014)
21. Rahman, M.A.; Manshaei, M.H.; Al-Shaer, E.; Shehab, M. Secure and Private Data Aggregation for Energy
[Consumption Scheduling in Smart Grids. IEEE Trans. Depend. Secur. Comput. 2017, 14, 221–234. [CrossRef]](http://dx.doi.org/10.1109/TDSC.2015.2446492)
22. [Shamir, A. How to Share a Secret. Commun. ACM 1979, 22, 612–613. [CrossRef]](http://dx.doi.org/10.1145/359168.359176)
23. Erkin, Z.; Tsudik, G. Private computation of spatial and temporal power consumption with smart meters.
In Proceedings of the 10th International Conference on Applied Cryptography and Network Security
(ACNS’12), Singapore, 26–29 June 2012; pp. 561–577.
24. Jia, W.; Zhu, H.; Cao, Z.; Dong, X.; Xiao, C. Human-factor-aware privacy-preserving aggregation in smart
[grid. IEEE Syst. J. 2017, 18, 598–607. [CrossRef]](http://dx.doi.org/10.1109/JSYST.2013.2260937)
25. [FriendlyARM. 2011. Available online: http://www.friendlyarm.net/ (accessed on 17 August 2011).](http://www.friendlyarm.net/)
26. Ahlswede, R.; Csiszar, I. Common randomness in information theory and cryptography I. Secret sharing.
_[IEEE Trans. Inform. Theory 1993, 39, 1121–1132. [CrossRef]](http://dx.doi.org/10.1109/18.243431)_
© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
[(CC BY) license (http://creativecommons.org/licenses/by/4.0/).](http://creativecommons.org/licenses/by/4.0/.)
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3390/APP9030490?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/APP9030490, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "GOLD",
"url": "https://www.mdpi.com/2076-3417/9/3/490/pdf?version=1548927505"
}
| 2,019
|
[] | true
| 2019-01-31T00:00:00
|
[
{
"paperId": "c93662c31635020079f18e4bc41b9eac7b19549b",
"title": "Efficient and Privacy-Preserving Data Aggregation Scheme for Smart Grid Against Internal Adversaries"
},
{
"paperId": "2789fde75cddbfd59dc9582add23d50bbd95d88b",
"title": "Secure and Private Data Aggregation for Energy Consumption Scheduling in Smart Grids"
},
{
"paperId": "a921c477112716f5f0749fb4907dd76286ec1600",
"title": "An efficient privacy-preserving aggregation and billing protocol for smart grid"
},
{
"paperId": "17a14af6522d344a044c2567658d1754e005a29a",
"title": "Secure and scalable aggregation in the smart grid resilient against malicious entities"
},
{
"paperId": "f6370a080c9463b0413e62ddce8ab458801edfdc",
"title": "Despicable me(ter): Anonymous and fine-grained metering data reporting with dishonest meters"
},
{
"paperId": "4862b22ffcca2c7e578b6a94c1251ffd4c752679",
"title": "A lightweight message authentication scheme for Smart Grid communications in power sector"
},
{
"paperId": "e08b1f7a7f9a933ea03a4b6063ec7425de77c655",
"title": "Diverse Grouping-Based Aggregation Protocol With Error Detection for Smart Grid Communications"
},
{
"paperId": "a55399ff53108b10db3265d32d2e8b1f08841b10",
"title": "A New Differentially Private Data Aggregation With Fault Tolerance for Smart Grid Communications"
},
{
"paperId": "e9fe07ff6b4fbe7d9113d7909578de71333269b5",
"title": "Privacy-preserving smart metering with verifiability for both billing and energy management"
},
{
"paperId": "0013d84b59ab11392862d47a270723e46dc455cb",
"title": "Human-Factor-Aware Privacy-Preserving Aggregation in Smart Grid"
},
{
"paperId": "f77a386ce173edde475d1f98818f560ff87d69a8",
"title": "An Efficient Merkle-Tree-Based Authentication Scheme for Smart Grid"
},
{
"paperId": "bd5bc5eaaf5686e8525c36df8bba9d7e190d1aad",
"title": "The use of RC4 Encryption for Smart Meters"
},
{
"paperId": "b33495fe50e37034b731cdacd76059b65344f70d",
"title": "PDAFT: A privacy-preserving data aggregation scheme with fault tolerance for smart grid communications"
},
{
"paperId": "f01c9a4010f3b4d03360834399d385d8bee52a27",
"title": "Privacy-Enhanced Data Aggregation Scheme Against Internal Attackers in Smart Grid"
},
{
"paperId": "0b458ce6c0d6d7fd20499e5b64a46132d7c380f2",
"title": "Cyber security in the Smart Grid: Survey and challenges"
},
{
"paperId": "d7b856e16f1f5ad328d2d3e4602190937b7c46ab",
"title": "(Non-)Random Sequences from (Non-)Random Permutations—Analysis of RC4 Stream Cipher"
},
{
"paperId": "a996987a529f6c436f1c348af2ff4c1983b18b23",
"title": "EPPA: An Efficient and Privacy-Preserving Aggregation Scheme for Secure Smart Grid Communications"
},
{
"paperId": "528d3b56227b2a91c55116db36d00598ca8a6511",
"title": "Private Computation of Spatial and Temporal Power Consumption with Smart Meters"
},
{
"paperId": "8104944dab19c8c16656238686f0e90e9bb461a8",
"title": "Privacy-Friendly Aggregation for the Smart-Grid"
},
{
"paperId": "1bbc6fae2cf1d6dba0954384b7393651fcc61d8a",
"title": "Privacy-Friendly Energy-Metering via Homomorphic Encryption"
},
{
"paperId": "6f5ba057d6fb21b474c3d626e91807d42befe51c",
"title": "Efficient aggregation of encrypted data in wireless sensor networks"
},
{
"paperId": "5553970601d94c090eb2d3431b653106bf271c4c",
"title": "Common randomness in information theory and cryptography - I: Secret sharing"
},
{
"paperId": "88abb2cda4f2a57499a717966ac4fbe9a993027a",
"title": "How to share a secret"
},
{
"paperId": "1c339243e44c835b2ceeb8fb2be183e1bdd45567",
"title": "PRGA: Privacy-Preserving Recording & Gateway-Assisted Authentication of Power Usage Information for Smart Grid"
},
{
"paperId": null,
"title": "FBI : Smart Meter Hacks Likely to Spread"
},
{
"paperId": null,
"title": "FriendlyARM"
},
{
"paperId": null,
"title": "Smart Meter Hacks Likely to Spread"
}
] | 21,608
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Business",
"source": "s2-fos-model"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Economics",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/0241483aa2623ab398de10982b958d01e6f5e453
|
[
"Computer Science"
] | 0.831905
|
Incentive-Driven Information Sharing in Leasing Based on a Consortium Blockchain and Evolutionary Game
|
0241483aa2623ab398de10982b958d01e6f5e453
|
Journal of Theoretical and Applied Electronic Commerce Research
|
[
{
"authorId": "1390607004",
"name": "Hanlei Cheng"
},
{
"authorId": "48514961",
"name": "J. Li"
},
{
"authorId": "2115405226",
"name": "Jing Lu"
},
{
"authorId": "1778780",
"name": "Sio-Long Lo"
},
{
"authorId": "2057445669",
"name": "Zhiyu Xiang"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"J Theor Appl Electron Commer Res"
],
"alternate_urls": [
"http://www.jtaer.com/"
],
"id": "890beb40-ba59-4681-9bb6-88ed97b7decb",
"issn": "0718-1876",
"name": "Journal of Theoretical and Applied Electronic Commerce Research",
"type": "journal",
"url": "http://www.scielo.cl/scielo.php?lng=en&pid=0718-1876&script=sci_serial"
}
|
Blockchain technology (BCT) provides a new way to mitigate the default risks of lease contracts resulting from the information asymmetry in leasing. The conceptual architecture of a consortium blockchain-based leasing platform (CBLP) is first proposed to facilitate information sharing between small and medium-sized enterprises (SMEs, the “lessees”) and leasing firms (LFs, the “lessors”). Then, based on evolutionary game theory (EGT), this study builds a two-party game model and analyzes the influences of four types of factors (i.e., information sharing, credit, incentive–penalty, and risk) on SMEs’ contract compliance or default behaviors with/without blockchain empowerment. The primary findings of this study are as follows: (1) SMEs and LFs eventually evolve to implement the ideal “win–win” strategies of complying with the contract and adopting BCT. (2) The large residual value of the leased asset can tempt SMEs to conduct a default action of unauthorized asset disposal, while leading LFs to access the CBLP to utilize information shared on-chain. (3) When the maintenance service is outsourced instead of being provided by lessors, the maintenance fee is not a core determinant affecting the equilibrium state. (4) There is a critical value concerning the default penalty on-chain to incentivize the involved parties to keep their commitments. (5) The capability of utilizing information, storage overhead, and security risk should all be taken into consideration when deciding on the optimal strategies for SMEs and LFs. This study provides comprehensive insights for designing an incentive mechanism to encourage lessees and lessors to cooperatively construct a sustainable and trustworthy leasing environment.
|
_Article_
# Incentive-Driven Information Sharing in Leasing Based on a Consortium Blockchain and Evolutionary Game
**Hanlei Cheng** **[1]** **, Jian Li** **[1,2]** **, Jing Lu** **[3,4], Sio-Long Lo** **[1,]* and Zhiyu Xiang** **[4]**
1 Faculty of Innovation Engineering, Macau University of Science and Technology, Macau 999078, China
2 School of Advanced Manufacturing, Guangdong University of Technology, Jieyang 522000, China
3 Department of Computer Science and Technology, Hubei University of Education, Wuhan 430205, China
4 Blockchain Laboratory, YGSoft Incorporation, Zhuhai 519085, China
***** Correspondence: sllo@must.edu.mo
**Citation: Cheng, H.; Li, J.; Lu, J.; Lo,**
S.-L.; Xiang, Z. Incentive-Driven
Information Sharing in Leasing Based
on a Consortium Blockchain and
Evolutionary Game. J. Theor. Appl.
_Electron. Commer. Res. 2023, 18,_
[206–236. https://doi.org/10.3390/](https://doi.org/10.3390/jtaer18010012)
[jtaer18010012](https://doi.org/10.3390/jtaer18010012)
Academic Editor: Jani Merikivi
Received: 28 September 2022
Revised: 16 December 2022
Accepted: 24 January 2023
Published: 29 January 2023
**Copyright:** © 2023 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
[Attribution (CC BY) license (https://](https://creativecommons.org/licenses/by/4.0/)
[creativecommons.org/licenses/by/](https://creativecommons.org/licenses/by/4.0/)
4.0/).
**Abstract: Blockchain technology (BCT) provides a new way to mitigate the default risks of lease**
contracts resulting from the information asymmetry in leasing. The conceptual architecture of a
consortium blockchain-based leasing platform (CBLP) is first proposed to facilitate information
sharing between small and medium-sized enterprises (SMEs, the “lessees”) and leasing firms (LFs,
the “lessors”). Then, based on evolutionary game theory (EGT), this study builds a two-party game
model and analyzes the influences of four types of factors (i.e., information sharing, credit, incentive–
penalty, and risk) on SMEs’ contract compliance or default behaviors with/without blockchain
empowerment. The primary findings of this study are as follows: (1) SMEs and LFs eventually
evolve to implement the ideal “win–win” strategies of complying with the contract and adopting
BCT. (2) The large residual value of the leased asset can tempt SMEs to conduct a default action
of unauthorized asset disposal, while leading LFs to access the CBLP to utilize information shared
on-chain. (3) When the maintenance service is outsourced instead of being provided by lessors, the
maintenance fee is not a core determinant affecting the equilibrium state. (4) There is a critical value
concerning the default penalty on-chain to incentivize the involved parties to keep their commitments.
(5) The capability of utilizing information, storage overhead, and security risk should all be taken
into consideration when deciding on the optimal strategies for SMEs and LFs. This study provides
comprehensive insights for designing an incentive mechanism to encourage lessees and lessors to
cooperatively construct a sustainable and trustworthy leasing environment.
**Keywords: small and medium-sized enterprises; leasing; blockchain technology; evolutionary game**
theory; information sharing
**1. Introduction**
Small and medium-sized enterprises (SMEs) typically encounter capital constraints
when buying heavy machinery and industrial equipment for manufacturing, such as
forklifts, trucks, hoists, etc. [1]. To cut back on the capital expense, leasing an asset from
the Original Equipment Manufacturer (OEM) or a Leasing Firm (LF) is a common and
economical option to meet the demand for equipment [2]. Leasing has become a popular
financing instrument [3].
In general, a lessee (e.g., SMEs) selects the required equipment, and then a lessor
can directly lease the asset they manufacture (if the lessor is the OEM) or purchase the
requested asset from the OEM for leasing it out (if the lessor is an LF, such as a financial
institution or a firm that specializes in leasing), with the lessee paying rent to the lessor in
exchange for using the asset [4]. When the leasing service period expires, the lessee may
opt to retain, renew, or return the leased equipment depending on the lease’s contractual
provisions. The leasing business emphasizes the separation of ownership and uses the
rights of the leased asset [5].
-----
_J. Theor. Appl. Electron. Commer. Res. 2023, 18_ 207
Compared with purchasing, although leasing is more flexible and cost-efficient for
a lessee, the current leasing business could encounter some challenges, such as lacking
knowledge about the lessee’s credit history, being unable to fully track assets in real time,
and failing to discover any default behavior arising from the information asymmetry.
Blockchain technology (BCT) stores data in a tamper-proof ledger that can be shared P2P
among many nodes without the aid of a reliable third party [6], which can help to transmit
data efficiently and accurately among multiple organizations, particularly in the area
of equipment leasing (asset management). Hence, in general, the lessee and lessor are
encouraged to participate in information sharing on the blockchain.
Nonetheless, it is challenging for lessees or lessors (particularly for SMEs) to develop
or participate in a blockchain-based application system (especially for the consortium permissioned blockchain system) due to the budget constraints of the BCT membership fee,
heavy data storage (computation) charges, and other barriers [7]. Meanwhile, considering
that stakeholders may maliciously handle sensitive data, the participants might be reluctant
to share critical information with other parties [8] unless sufficient incentives nudge the lessee.
However, previous researchers have not performed a quantitative exploration of stakeholders’
BCT-participating behaviors in the context of leasing, where there is a trade-off between the
factors of information sharing, credit, incentive–penalty, and risk. On the other hand, although
there exists some research relating to the blockchain applied in leasing, the interactive behaviors among lessees and lessors rarely receive adequate attention. Hence, our study aims to
bridge this gap by addressing the following research questions.
(1) How can the blockchain technically drive information sharing and storage between
the SME (the “lessee”) and LF (the “lessor”)?
(2) How to incentivize excellent lessees to share more information while expecting that
rational lessees and lessors can both maximally benefit from the leasing business
empowered by BCT.
(3) How can the lessee and lessor adjust their behavior strategies to ensure that all parties’
payoffs reach equilibrium through continuous trial-and-error learning?
To address these questions, we need to accomplish the following research objectives.
First, a conceptual architecture of a consortium blockchain-based leasing platform (CBLP)
is devised, suitable for information sharing among SMEs and LFs in a P2P distributed
network. Secondly, we employ evolutionary game theory (EGT) to formulate a game model,
taking fully into consideration the four kinds of factors (i.e., information sharing, credit,
incentive–penalty, and risk) that affect the leasing strategies involving the two game parties
(lessee and lessor). Finally, the main impacting results are discussed in-depth, and policy
implications are provided on the ground.
In addition, compared with other similar works investigating BCT strategies using the
evolutionary game, the novelty and contributions of this study can be summarized as follows:
(1) Our evolutionary game model is developed on the blockchain-based leasing business
(specifically the operating lease) in manufacturing, which pays more attention to the
SME’s leasing behavior (i.e., making the rental payment, reverting the leased asset,
maintenance responsibility, and asset monitoring) dynamically changes with the BCT
adoption/non-adoption strategy. This study can mitigate the shortcomings of today’s
leasing management.
(2) We provide a more comprehensive analysis demonstrating that the four factors of
“information sharing, credit, incentive–penalty, and risk” dynamically impact the lessee’s
complying performance on the LC and the lessor’s decision-making on BCT adoption.
More importantly, we carefully consider technical barriers faced by the organizational
players when implementing BCT, such as on-chain and off-chain storage overheads,
leasing transaction verification overheads, and credit assessment in BCT.
(3) Based on the game analysis, our experimental results can support LFs (the “lessor”) in
comprehensively understanding how SMEs (the “lessee”) meet the obligations in the
-----
_J. Theor. Appl. Electron. Commer. Res. 2023, 18_ 208
LC and give some implications to policymakers when designing a proper incentive
mechanism on the lease.
The remainder of this paper is structured as follows: Section 2 provides a theoretical
background of the leasing business, BCT, leasing empowered by BCT, and the evolutionary game integrated with BCT. Section 3 exhibits the CBLP composed of SMEs and LFs.
Section 4 states the problem description. Section 5 builds the game model. Section 6
provides a mathematical analysis of the model stability. Section 7 conducts a numerical
simulation. Some recommendations and policy implications are provided in Section 8.
**2. Literature Review**
The following subsections will give an overview of crucial terminologies (i.e., leasing
and BCT), the state of the art of BCT-based leasing, and BCT strategies using evolutionary
game theory, which serve as a solid theoretical background for this study. After that, the
main challenges in conventional leasing are highlighted.
_2.1. Definition of Leasing_
SMEs typically need to decide between leasing and buying an expensive heavy asset
(such as real estate, transportation equipment, industrial equipment, etc.). In recent years,
leasing assets has become a popular financing tool for SMEs to solve capital problems in the
supply chain [9]. According to the Accounting Standard IAS 17, “a lease is an agreement
whereby the lessor conveys to the lessee in return for payment or series of payment
the right to use an asset for an agreed period of time” (see, e.g., European Commission,
2012) [10]. From an accounting perspective, there are two types of lease, a capital lease and
an operating lease [11]. In a capital lease, the lessor transfers the ownership of the asset to
the lessee at the end of the lease. In contrast, an operating lease only allows the lessee to
have the right to use the assets. Still, it requires the asset to be reverted to the lessor, such
that the lessor will either re-lease the asset in another LC or sell it to release the residual
value. At present, the operating lease dominates the leasing market.
Concerning the determinants in default actions of the LC, Kaposty et al. [12] defined
an LC as having defaulted when the lessee becomes insolvent or the lessor terminates
the contract due to an overdue payment owed by the lessee. The latter case is considered in this study. Difficulty in repossessing the leased asset is also one of the results of
defaulting [13]. Altman et al. [14] discovered that a lessee with poor creditworthiness
defaults more easily, resulting in higher default losses. On the other hand, Kysucky and
Norden [15] revealed that reducing information asymmetry between the lessee and lessor
could motivate the lessee to maintain its reputation to obtain future leases. In addition,
an exhaustive inspection of asset maintenance and disposals plays an essential role in
contract defaults [12]. However, the current leasing system lacks the ability to reliably
record real-time information (including the documents) about the leased asset’s operational
activity, which hampers the lessor’s ability to ensure the lessee’s compliance with the LC.
_2.2. Blockchain Technology (BCT)_
Blockchain Technology (BCT) was initially proposed by Satoshi Nakamoto, and it enables
transactions to be encapsulated in data blocks and appended to a ledger as a chain structure [16]. It allows distributed and mutually distrustful nodes validate transactions through
consensus mechanisms while utilizing cryptography (i.e., public–private key encryption and
hash functions) to ensure data integrity. By its nature, blockchain technology makes transactions synchronous, non-reversible, immutable, and traceable in distributed databases, enabling
organizations to store and share reliable data without double-checking [17]. Blockchain technology can effectively solve the problem of information asymmetry [18] and monitor the
asset’s operation in real time, which helps the lessor to alleviate the default risks caused by a
low-credit lessee [19]. Therefore, it is beneficial to encourage SMEs (the “lessee”) to use BCT
and energetically participate in information sharing.
-----
_J. Theor. Appl. Electron. Commer. Res. 2023, 18_ 209
BCT is generally classified as the permissionless blockchain and the permissioned
blockchain [20], depending on whether or not nodes are granted to participants in a
blockchain network [14,17]. The permissionless blockchain, also called public blockchain,
is open access, allowing any node to participate in the consensus procedure, such as
Bitcoin and Ethereum. The permissioned blockchain can be further categorized into private
permissioned blockchain, in which whitelisted participants in one organization are selected
in advance to join the invitation-only network, and a consortium permissioned blockchain,
which is operated under the control of several authorized organizations allowing the
identifiable participants to execute certain on-chain actions, such as Quorum, Hyperledger
Fabric, and Corda. The consortium blockchain is becoming popular with enterprises, where
a group of companies collaboratively use the blockchain to improve business processes. Any
organization can apply to join the blockchain network, but only authorized organizations
granted membership are allowed to write or read information on-chain [21]. This study
is dedicated to introducing a consortium permissioned blockchain jointly maintained by
the OEM, SMEs, LFs, third-party maintenance centers (MCs), and regulators. At present,
blockchain has been widely used in many industries for sharing information, such as supply
chain [22], energy [23], healthcare [24], industrial manufacturing [25], smart cities [26], and
online education [27].
_2.3. BCT Application in Leasing_
Currently, there is relatively limited research on the application of BCT in the leasing
business. Most research focuses on how BCT fosters information exchange through the
lifecycle of the leasing process and improves the efficiency and transparency of leased
asset management. For instance, to address the issues of lengthy negotiation cycles and
cumbersome financing procedures caused by information asymmetry, IBM proposed a
crane leasing model based on the IBM Blockchain Platform that requires the identity of the
leased crane to be registered on the blockchain and leasing transactions to be recorded on the
leasing blockchain [28]. Leased physical aircraft can be tokenized via blockchain, facilitating
asset management [13]. In addition, several researchers are particularly interested in BCTbased car-leasing. Auer et al. [29] developed a prototype blockchain-IoT-based car-leasing
platform, demonstrating that the blockchain can facilitate collaboration among stakeholders
to some extent while relying on the appropriate balance among factors such as security,
authenticity, traceability, scalability, etc. It also emphasizes considering storing car-renting
events on- or off-chain to support scalability, as agreed by Faber et al. [30]. To address
the problem of inefficiency in delivering and searching records, Agyekum et al. [31] used
Ethereum to construct a car-leasing platform that enables the transfer of ownership of
a leased car by invoking a transaction on the blockchain, hence helping the regulator to
clearly monitor every leasing transaction.
The aforementioned cases imply the following potential benefits of the blockchain
applied in leasing: (1) Stakeholders (such as lessors) spend less time verifying the leasing
information’s authenticity on-chain, since the blockchain can record lease contracts and
financial transactions in a non-editable way, which reduces the credit investigation cost [32].
(2) Smart contracts deployed on a distributed ledger can help automate some lease payments or ownership transfers, speeding up the processing of rental transactions [33]. (3) All
historical events associated with the leased assets’ operation and maintenance and the
provenance-related financing activities are objectively recorded by multiple nodes on-chain,
which could guarantee asset traceability and data credibility in the leasing business [34].
Hence, BCT is conducive for SMEs and LFs to effectively choose the suitable potential
lessor/lessee to sign the LC.
_2.4. Evolutionary Game Theory (EGT)_
Evolutionary game theory (EGT) is derived to explore the behavior of the large population of boundedly rational agents who repeatedly engage in strategic interactions under
incomplete information circumstances [35]. In contrast with classic game theory, EGT
-----
_J. Theor. Appl. Electron. Commer. Res. 2023, 18_ 210
has the advantage of analyzing how the game player would dynamically change their
own strategic decisions over time through learning and adapting to the other’s strategic
decisions [36]. There is a core concept in EGT named evolutionary stable strategy (ESS),
which, if adopted by all players, means that the game cannot be invaded by alternative
strategies [37]. Several researchers have applied EGT to leasing issues. For example, to
study the safety supervision of town crane operation, Chen et al. [38] established a tripartite
evolutionary game model, which reveals that when the penalty amount resulting from
poor crane supervision is greater than the total safety investment cost, the stakeholders will
apply strict supervision strategies to the asset.
In addition, some scholars have conducted similar works deciding whether to adopt
BCT using evolutionary game theory. However, most of them focus on the supply chain
(finance). For instance, Su et al. [39] constructed a tripartite game model to explore BCT in
relation to the evolutionary stability strategies among CEs, SMEs, and FIs, discovering that
relatively large default losses can help SMEs to repay receivables on time. Tang et al. [40]
used an evolutionary game to demonstrate that BCT can effectively facilitate information sharing in supply chain collaborations. Sun et al. [41] established an evolutionary
game model to reveal that BCT impacts credit risk, which plays a vital role in deciding whether financial institutions accept factoring applications in supply chain finance.
Song et al. [42] analyzed a tripartite game model of an agricultural supply chain, discovering that blockchain operation costs significantly affect the behaviors of governments and
agricultural enterprises.
Based on the above literature analysis, it can be found that most research mainly
focuses on proposing a blockchain-based leasing scheme or using EGT to study the SME’s
BCT strategies in the supply chain. Nevertheless, our work not only provides a consortium
blockchain-based leasing platform (CBLP) to streamline the information sharing between
lessees and lessors, but also contributes to establishing a two-player evolutionary game
model analyzing the lessee’s leasing behaviors (i.e., complying with or defaulting on the
LC) considering whether to adopt BCT for information sharing. This study will shed light
on the long-term development of the leasing industry.
**3. Description of Consortium Blockchain-Based Leasing Platform (CBLP)**
Since the evolutionary model that this study will construct is of significant relevance
to information storage (i.e., on- and off-chain) and consensus mechanisms (i.e., Raft with
credit evaluation and transaction verification), in this section, it is necessary first to present
the conceptual architecture of the proposed consortium blockchain-based leasing platform
(CBLP), and then the transaction verification process will be concisely explained.
_3.1. Conceptual Architecture of CBLP_
This research first provides a consortium blockchain-based leasing platform integrated
with RFID devices [43], which incorporates stakeholders such as OEM, SMEs, LFs, MCs,
and regulators. The platform is built by using Hyperledger Fabric (HLF) [44], which is an
enterprise-grade permissioned blockchain platform facilitating information sharing among
multiple organizations [45,46]. In general, an authorized node is required to pay fees to
access the consortium blockchain [47]. In the distributed ledger, only organizations with
valid IDs can process transactions. More specifically, all authorized lessees and lessors
have their PKI-CA certificate and unique Decentralized Identifiers (DIDs) registered on a
blockchain with restricted access [48,49]. By scanning the RFID sensor tags encapsulated in
the DID, the running condition of the leased asset is recorded as a transaction, which is then
turned into a “block” and appended to the ledger. This means that the lessee cannot refute
or alter the historical logs of equipment operation and maintenance. This data reliability
is conducive to efficiently managing the whole life cycle of the physical leased asset [50],
which can help to reduce the inspection costs during the leasing period. It is also critical for
each participant involved to be certain about the asset traceability in case of fraud, damages,
dispossession, or misdisposition.
-----
_J. Theor. Appl. Electron. Commer. Res.case of fraud, damages, dispossession, or misdisposition. 2023, 18_ 211
#### On the other hand, due to the large data sets shared by many various stakeh there is a need to consider data storage and scalability challenges when assessing t extent SMEs are willing to adopt BCT [51]. Our work leverages a hybrid on-chain aOn the other hand, due to the large data sets shared by many various stakeholders,
there is a need to consider data storage and scalability challenges when assessing to what
#### chain storage mechanism to store and access information (especially complexity d
extent SMEs are willing to adopt BCT [51]. Our work leverages a hybrid on-chain and
#### the CBLP [52,53]. Specifically, the encrypted raw data (e.g., file) are stored on an of
off-chain storage mechanism to store and access information (especially complexity data)
#### cloud storage provider (CSP) or distributed storage system (e.g., InterPlanetary Fion the CBLP [52,53]. Specifically, the encrypted raw data (e.g., file) are stored on an off- tem, IPFS) [54]. The off-chain data are linked with the specific metadata via a hash pchain cloud storage provider (CSP) or distributed storage system (e.g., InterPlanetary File
System, IPFS) [54]. The off-chain data are linked with the specific metadata via a hash
#### which is stored as a transaction validated to the ledger and can be used to audit t
pointer, which is stored as a transaction validated to the ledger and can be used to audit the
#### chain data that were not modified [55]. The architecture of CBLP with on- and of
off-chain data that were not modified [55]. The architecture of CBLP with on- and off-chain
#### information storage mechanisms is shown in Figure 1. information storage mechanisms is shown in Figure 1.
##### Figure 1. Figure 1. The architecture of CBLP with an on- and off-chain information storage mechanism.The architecture of CBLP with an on- and off-chain information storage mechanism
_3.2. Raft Consensus Based on Credit_
#### 3.2. Raft Consensus Based on Credit
Since various transactions are executed through triggering relevant smart contracts
(e.g., lease contract, asset ownership transfer contract, lease payment contract, data sharingSince various transactions are executed through triggering relevant smart co
#### (e.g., lease contract, asset ownership transfer contract, lease payment contract, datcontract) using multiple organizational nodes, a consensus protocol is used by the CBLP. It
plays a crucial role in ensuring that the transactions are recorded in an agreed order on-chain.
#### ing contract) using multiple organizational nodes, a consensus protocol is used
Meanwhile, to avoid the untrusted consortium stakeholders uploading false information
#### CBLP. It plays a crucial role in ensuring that the transactions are recorded in an
-----
_J. Theor. Appl. Electron. Commer. Res. 2023, 18_ 212
### false information or changing the private data content on-chain to comply with the tract, we use a Raft consensus algorithm combined with “credit” incentives to mai the blockchain. That is, an SME (or changing the private data content on-chain to comply with the contract, we use a RaftOrg-lessee node A) and an LF (Org-lessor node B) adop
consensus algorithm combined with “credit” incentives to maintain the blockchain. That is,
### flow of “execute–order–validate” to record the transactions on-chain [45]. This is the
an SME (Org-lessee node A) and an LF (Org-lessor node B) adopt the flow of “execute–order–
### fundamental transaction process of Hyperledger Fabric at present and will not be e
validate” to record the transactions on-chain [45]. This is the most fundamental transaction
### rated on further due to the limitations of this paper’s length, while it is depicted in Fprocess of Hyperledger Fabric at present and will not be elaborated on further due to the 2. limitations of this paper’s length, while it is depicted in Figure 2.
##### Figure 2. Figure 2.Transaction flow in Hyperledger Fabric with Raft consensus algorithm. Transaction flow in Hyperledger Fabric with Raft consensus algorithm.
Notably, the Raft consensus protocol follows the “leader and follower” architecture [56,57]
### Notably, the Raft consensus protocol follows the “leader and follower” archite
to implement the ordering service, where the leader nodes are dynamically elected from the
### [56,57] to implement the ordering service, where the leader nodes are dynamically elConsenter Set and a node’s credit value determines whether it can join the Consenter Set [58]. from the When the credit value exceeds a predetermined threshold, the node can join theConsenter Set and a node’s credit value determines whether it can join the Consenter senter SetSet as a consensus node. On the contrary, when the credit is lower than the minimum [58]. When the credit value exceeds a predetermined threshold, the nod
threshold, the node will face a penalty imposed by the CBLP. In addition, we propose
### join the Consenter Set as a consensus node. On the contrary, when the credit is lower
that the credit value is increased or decreased according to the contract compliance or
### the minimum threshold, the node will face a penalty imposed by the CBLP. In adddefault behavior of the (SME) Org-lessee node in the leasing business; the more frequently we propose that the credit value is increased or decreased according to the contract it conforms to the lease contract, the higher its credit value and the greater the likelihood is
that the node will be selected as the “Leader” in the Consenter Set to package transactions
### pliance or default behavior of the (SME) Org-lessee node in the leasing business; the
into a new block and finalize it. Credit assessed for the SME cannot be treated as an
### frequently it conforms to the lease contract, the higher its credit value and the greate
indicator of the lessee’s reputation and contribution to the leasing business. Still, it can help
### likelihood is that the node will be selected as the “Leader” in the the enterprise earn more recognition and achieve more lease financing opportunities fromConsenter Set to pac transactions into a new block and finalize it. Credit assessed for the SME cannot be trlessors, encouraging each lessee to share information on the CBLP [58].
In summary, based on the right balance of the above on- and off-chain storage costs
### as an indicator of the lessee’s reputation and contribution to the leasing business. S
and credit incentive mechanisms, the nodes will choose to actively comply with the LC
### can help the enterprise earn more recognition and achieve more lease financing oppand upload authentic information on-chain by joining the consortium blockchain for larger nities from lessors, encouraging each lessee to share information on the CBLP [58]. gains, resulting in the eventual emergence of a Nash equilibrium. In summary, based on the right balance of the above on- and off-chain storage and credit incentive mechanisms, the nodes will choose to actively comply with th
-----
_J. Theor. Appl. Electron. Commer. Res. 2023, 18_ 213
**4. Problem Description**
In this section, we will first describe the problem we studied. The basic lease scenario
explored in this study is provided to better understand the game model. Afterward, the
critical parameters are elucidated and presented in Table 1.
**Table 1. Parameters. Explanation under the conventional/blockchain-based lease mode.**
**Mode** **Party** **Notation** **Definition**
_R_ Total rental payments to the LF under the terms of the lease
_r1_ Return rate of the SME on the lease
_r3_ Reinvestment rate of the SME after the contract’s default
_f_ Maintenance fee for the leased asset during the lease period
_p1_ Default penalty of the SME under the conventional lease
_σ_ Incentives of the SME given by the LFs due to LC compliance
_r2_ Return rate of the LF on the lease
_Ct_ Marginal credit investigation costs of the LF
_C0_ Original acquisition cost of the leased asset
_C1_ Monitoring cost of asset’s operation under the conventional lease
_vs_ Residual value of the leased asset at the end of the lease
_ε_ Loss rate of the LF caused by the contract default
_Cb_ Membership cost of the SME joining the consortium blockchain
∆νc Increased credit value of the SME due to LC compliance on-chain
_I_ Fixed reward when mining a block on-chain
_p2_ Default penalties of the SME on-chain
_ZA_ Quantity of information shared by the SME on-chain
_uA_ Relative computing power provided by the SME on-chain
_g_ Synergy gain on the lease business empowered by the blockchain
_C2_ Monitoring cost of asset’s operation under the blockchain-based lease
_ZB_ Quantity of information shared by the LF on-chain
_uB_ Relative computing power provided by the LF on-chain
_ϕ_ Coefficient of information transmission efficiency on-chain
_ω_ Validation cost coefficient of confirming transaction on-chain
_λ_ Storage cost coefficient of information stored off-chain CSP/IPFS
_η_ Security risk coefficient of sharing information on-chain
Under the
conventional
lease mode
Under the
blockchain-based
lease mode
SME
(Org-lessee node A)
LF
(Org-lessor node B)
SME
(Org-lessee node A)
LF
(Org-lessor node B)
SME and LF
_4.1. Description of Problem_
The game model involves two types of players in a lease: the lessee (i.e., an SME) and
lessor (i.e., an LF). Under a conventional lease, the SME is responsible for paying fees for the
right to use an asset leased from the LF, and generally, the SME as a lessee must maintain
the asset to ensure that it remains in an operational condition [59]. The LF will pay the
credit investigation cost to evaluate whether an SME can pay its rent on time. Meanwhile, it
is difficult for the parties to immediately share information (including the historical default
records) and for the LF to monitor the leased equipment/assets in real time. Applying BCT
can solve the aforementioned issues [29]. If the LF requires the SME to join the consortium
so as to upload information (such as historical asset operation documents or payment
performance, etc.) on the CBLP, not only can a credit review be instantly conducted, but
the ownership and provenance of the leased asset can also be tracked in real time through
-----
_J. Theor. Appl. Electron. Commer. Res. 2023, 18_ 214
the asset’s operation. Moreover, a smart contract can be automatically executed to make
the lease payment as agreed in the lease contract (LC), which results in the synergetic gains
of the leasing process and improves the efficiency of asset management empowered by the
blockchain [60]. If stakeholders can share and process a greater quantity of information
on-chain, thereby significantly improving the precision of the decisions made by each party,
it is essential to take into consideration the corresponding data storage overheads, security
exposure, and transaction verification costs when practically using BCT. Once the SME
and LF adopt the CBLP, they need to be subject to harsher punishments resulting from
defaulting behavior, which undermines the enterprise’s reputation on the whole network.
In addition, a blockchain-based credit evaluation mechanism can enhance the effective
management of the leased asset, which is also conducive to mining reliable data blocks in
the distributed ledger.
Therefore, this study intends to model the problem that considers the SME and
LFs individual decision-makers concerning “complying with or defaulting on the LC”
and “accessing or not accessing the CBLP” and to comprehensively analyze the dynamic
influence of the four factors (i.e., information sharing, credit, incentive–penalty, and risk)
on the choice of strategy by using evolutionary games.
_4.2. Basic Lease Scenarios_
The game model is constructed based on an operating lease with respect to heavy
equipment (e.g., forklifts, trucks, and hoists) in the manufacturing supply chain. Generally,
the equipment is relatively expensive to purchase, and leasing it is a better option for SMEs.
The LF (the “lessor”) first acquires an asset from an OEM and lends the asset to the SME
(the “lessee”) for a specific term in exchange for periodic rental payments. Once the LC is
signed, the SME has the right to use the asset, whereas the ownership of the leased asset
remains with the LF. Therefore, the SME must comply with the contract, not only paying
the full rent on time but also reverting the asset to the LF at the maturity date of the lease;
otherwise, the SME will pay the penalty for their default. In addition, the LC specifies that
the lessee takes responsibility for the maintenance and outsources it to the OEM or MCs
other than the lessor (LF in this case).
_4.3. Model Parameters_
Rental Payments (R): Rental payments refer to the monthly/quarterly rent that the
SME (the “lessee”) pays to the LF (the “lessor”).
(1) Return rate (r1, r2): Return rate refers to the yield that can be earned when completing
the investment activity on the lease.
(2) Reinvestment rate (r3): Reinvestment rate refers to the yield that the lessee expects to
earn when it does not pay or defers the full rental price, which can be put into other
investments for extra gains.
(3) Maintenance fee ( _f_ ): Maintenance fee refers to the cost of carrying out maintenance
actions to ensure that the leased asset is in a proper operating condition. In this
study, the LC states that the maintenance service must be provided by MCs and
completed until the lease termination—the maintenance fee is not embedded in the
rental payment.
(4) Loss rate (ε): Loss rate refers to the loss that could result from the lessee’s defaulting
behavior—for instance, if the lessee defaults by not returning the leased asset at the
end of the lease, which cannot be re-leased to the next lessee upon termination of the
previous LC.
The relevant parameters’ notations and definitions are shown in Table 1.
**5. Model Formulation**
In this section, an evolutionary game model between the SME (the “lessee”) and the
LF (the “lessor”) is developed. Before the mathematical payoff matrix is constructed, some
assumptions are first provided.
-----
_J. Theor. Appl. Electron. Commer. Res. 2023, 18_ 215
_5.1. Basic Assumptions_
**Assumption 1 (A1): Rational Participants Assumption.**
All participants in the game are boundedly rational [61]. Nodes with constrained
computing power may not be able to perfectly utilize all of the information on-chain owing
to hardware faults or network congestion. Under asymmetric information, each node will
constantly select the optimal strategy to maximize its interests while being affected by
multiple factors and will eventually reach a state of equilibrium [62,63].
**Assumption 2 (A2): Strategy Selection Assumption.**
Assume that the SME (the “lessee”) chooses to “comply with or default on the LC”
with the respective probability of x or 1 _x, x_ [0, 1]. The LF (the “lessor”) chooses to
_−_ _∈_
“access or not access the CBLP”, utilizing the information shared by stakeholders with the
respective probability of y or 1 _y, y_ [0, 1].
_−_ _∈_
**Assumption 3 (A3): Default Behavior Assumption.**
For the SME’s unilateral default behavior, assume that it consists of two primary
actions: one is not making the full rental payment by the due date, and the other is failing to
return the leased asset to the lessor when the LC expires (for simplicity, this study considers
that the two actions simultaneously occur when modeling).
**Assumption 4 (A4): Information Sharing Assumption.**
The quantity of information shared by the player is Zi, and the ability of each player
to process and utilize the information on-chain [40] is uj, which depends on the computing
power. Hence, the amount of effective information on-chain that the SME or the LF obtains
from each other is, respectively, uAZB and uBZA.
**Assumption 5 (A5): Credit Assumption.**
Assume that each SME will be assigned a credit value ∆vc, which increases with the
LC compliance performance. The higher the credit value owned by the SME, the greater
the possibility of the enterprise becoming the “Leader” in the Raft Consensus Protocol
(Section 3.2) to validate the lease transaction on each consensus round, so improving the
lessee’s reputation and recognition on-chain.
**Assumption 6 (A6): Incentive–Penalty Assumption.**
After the SME participates in information sharing, if the enterprise breaches the LC or
tries to tamper with the existing LC to legalize the default behavior on-chain, this has a
profoundly negative effect on the leasing business, hence making the SME’s default penalty
intensity larger under the CBLP—that is, p2 > p1. The penalty can be deducted from the
lessee’s token deposit by executing a smart contract of transferring transactions [64].
**Assumption 7 (A7): (Technology) Risk Assumption.**
BCT can improve the data transmission efficiency (ϕ) end-to-end, but meanwhile, each
player has to bear the data validation on-chain (ω) and storage costs off-chain (λ) due to
the blockchain’s storage limitations. The player also may suffer security risks (η), such as
data leakage risks and network attack risks [65]. This has practical implications, in that the
benefits of sharing information outweigh the relevant costs after joining the blockchain
network, ϕ > ω + λ + η.
**Assumption 8 (A8): Other Cost Assumption.**
-----
_J. Theor. Appl. Electron. Commer. Res. 2023, 18_ 216
**A8.1: The blockchain can record the SME’s credit history as tamper-resistant and traceable data [66],**
_and if the SME joins the blockchain network, the marginal credit investigation cost (Ct) will shrink_
_and asymptotically approach zero._
**A8.2: In leased asset management empowered by BCT, the LF can continuously monitor the**
_condition of the leased asset at lower costs [67]—as such, C1 > C2._
_5.2. Payoff Matrix_
In this model, there are four different strategies. Considering that the model is complicated to understand, based on the above lease scenario and assumptions, this subsection
will thoroughly explain the players’ payoff under each strategy.
(1) Strategy I: S1 = {Comply, Access}
Since the SME (the “lessee”) fully abides by the LC, and the LF accesses the CBLP,
taking advantage of information sharing on-chain, this results in the SME obtaining the rewards of successfully mining the block _I,_ effective information utilization
[(ϕ − _ω −_ _λ −_ _η)ZA + uAZB], the return on the lease Rr1, the incentive σ given by the_
LF (the “lessor”), plus the credit value; however, the SME has to make the payments of
rental R, maintenance fee f, and the consortium membership cost Cb.
On the other hand, the LF obtains rent R, the return on the lease Rr2, effective information utilization [(ϕ − _ω −_ _λ −_ _η)ZB + uBZA], synergy gain g on the lease empowered by_
BCT, plus the residual value of the leased asset vs after receiving the leased asset returned
from the SME, while bearing the cost of purchasing the asset from the OEM at price C0,
monitoring cost C2, and the consortium membership cost Cb.
Therefore, in Strategy I, the payoffs of the SME and LF are formulated as in
Equations (1) and (2), respectively.
_PA[S][1]_ [=][ I][ + [(][ϕ][ −] _[ω][ −]_ _[λ][ −]_ _[η][)][Z][A][ +][ u][A][Z][B][] +][ R][(][r][1][ −]_ [1][) +][ σ][ +][ ∆][v][c][ −] _[f][ −]_ _[C][b]_ (1)
_PB[S][1]_ [=][ R][(][r][2][ +][ 1][) +][ v][s][ + [(][ϕ][ −] _[ω][ −]_ _[λ][ −]_ _[η][)][Z][B][ +][ u][B][Z][A][] +][ g][ −]_ _[C][0][ −]_ _[C][2][ −]_ _[C][b]_ (2)
(2) Strategy II: S2 = {Default, Access}
Due to default actions (Section 4.2.), the SME uses the rent to perform re-investment
and dispose of the leased asset that has been exhaustively used for manufacturing at the
end of the lease. Therefore, it gives the SME the chances to earn extra re-reinvestment
return Rr3 and sell the leased asset at the market value of the residual vs. Adopting the
BCT provides the SME with effective information utilization [(ϕ − _ω −_ _λ −_ _η)ZA + uAZB]._
However, to continuously keep the leased asset effectively operating without impacting
production, the SME still needs to pay the maintenance fee to the MC (instead of the LF)
and will be punished in p2 resulting from the default actions.
Meanwhile, although the LF can obtain the effective information utilization empowered by the BCT, the default behavior by the SME not only causes the LF to be unable
to receive the rental payment R, but also leads to it losing the further earnings Rε from
re-leasing to other lessees due to the out-of-control of the leased asset when the rental
period is complete. The asset acquiring cost C0, monitoring cost C2, and the consortium
membership cost Cb occur.
Therefore, in Strategy II, the payoffs of the SME and LF are formulated as in
Equations (3) and (4), respectively.
_PA[S][2]_ [=][ Rr][3][ +][ v][s][ + [(][ϕ][ −] _[ω][ −]_ _[λ][ −]_ _[η][)][Z][A][ +][ u][A][Z][B][]][ −]_ _[f][ −]_ _[p][2][ −]_ _[C][b]_ (3)
_PB[S][2]_ [= [(][ϕ][ −] _[ω][ −]_ _[λ][ −]_ _[η][)][Z][B][ +][ u][B][Z][A][]][ −]_ _[R][ε][ −]_ _[C][0][ −]_ _[C][2][ −]_ _[C][b]_ (4)
-----
_J. Theor. Appl. Electron. Commer. Res. 2023, 18_ 217
(3) Strategy III: S3 = {Comply, Not-access}
When the SME actively keeps to the stipulations of the LC, the SME earns the investment return Rr1 on the lease and is rewarded by the LF with the incentive σ. However, the
SME needs to pay the maintenance fee f to ensure that the leased asset is in good condition.
For the lessor, the LF not only obtains benefit Rr2 from the leasing activity, but it also
retains the value of the leased asset vs at the end of the LC. Nonetheless, the insufficient
credit record integrity of the SME forces the LF to incur credit audit expenses Ct before
making a decision on the lease. The costs (C0, C1) of acquiring and monitoring the leased
assets are ineluctable.
Therefore, in Strategy III, the payoffs of the SME and LF are formulated as in
Equations (5) and (6), respectively.
_PA[S][3]_ [=][ R][(][r][1][ −] [1][) +][ σ][ −] _[f]_ (5)
_PB[S][3]_ [=][ R][(][r][2][ +][ 1][) +][ v][s][ −] _[C][0][ −]_ _[C][t][ −]_ _[C][1]_ (6)
(4) Strategy IV: S4 = {Default, Not-access}
Based on the above Strategies II and III, the SME will always earn the reinvestment
return Rr3 and residual value vs, but it may also suffer from the default punishment p1. In
addition, although the SME will resell the leased asset by defaulting, the enterprise has
to take responsibility for maintaining it f to ensure that the leased asset remains in an
operational condition for manufacturing.
If the SME breaches the LC without joining the blockchain network, the LF does not get
any returns, and may even be charged with the costs in credit auditing Ct, asset acquiring
_C0, and monitoring C1._
Therefore, in Strategy IV, the payoffs of the SME and LF are formulated as in
Equations (7) and (8), respectively.
_PA[S][4]_ [=][ Rr][3][ +][ v][s][ −] _[f][ −]_ _[p][1]_ (7)
_PB[S][4]_ [=][ −][R][ε][ −] _[C][0][ −]_ _[C][t][ −]_ _[C][1]_ (8)
Consequently, the profit matrix of the two-party game is shown in Table 2.
**Table 2. Evolutionary game payoff matrix of the SME and LF.**
**LF**
**Strategy** **(Org-Lessor Node B)**
**Access** **Not-Access**
SME
(Org-lessee node A)
Comply _I + [(ϕ −_ _ω −_ _λ −_ _η∆)ZvcA − +f u −AZCbB] + R(r1 −_ 1) + σ + _R(r1 −_ 1) + σ − _f_
_R(r2 + 1) + vs + [(ϕC −0 −ω −C2λ − −Cηb)ZA + uBZA] + g −_ _R(r2 + 1) + vs −_ _C0 −_ _Ct −_ _C1_
Default _Rr3 + vs + [(ϕ −_ _ω −_ _λ −_ _η)ZA + uAZB] −_ _f −_ _p2 −_ _Cb_ _Rr3 + vs −_ _f −_ _p1_
[(ϕ − _ω −_ _λ −_ _η)ZB + uBZA] −_ _Rε −_ _C0 −_ _C2 −_ _Cb_ _−Rε −_ _C0 −_ _Ct −_ _C1_
**6. Model Stability Analysis**
This section will first construct the replicator dynamic equations between the SME and
LF and then will discuss in depth how the two rational players reach an equilibrium state
through iteratively changing strategies. A mathematical sensitivity analysis on each type of
factor (i.e., information sharing, credit, incentive–penalty, risk) will ultimately be provided.
-----
_JTAERJ. Theor. Appl. Electron. Commer. Res. 2023, 18, FOR PEER REVIEW 2023, 18_ 21813
_6.1. Replicator Dynamic SystemBased on the above evolutionary game payoff matrix, we can calculate the expected_
returns of the SME (the “lessee”) and LF (the “lessor”) when they choose different strate-Based on the above evolutionary game payoff matrix, we can calculate the expected
gies, and then construct the replicator dynamic equations for each subject. returns of the SME (the “lessee”) and LF (the “lessor”) when they choose different strategies,
and then construct the replicator dynamic equations for each subject.
6.1.1. Replication Dynamic Equation of the SME
6.1.1. Replication Dynamic Equation of the SME
Assuming the expected return of the SME’s compliance with and, defaulting on the
LC, the average returns are Assuming the expected return of the SME’s compliance with and, defaulting on the𝐸�, 𝐸���, and ���𝐸�, respectively. Then:
LC, the average returns are𝐸� = 𝑦�𝐼+ �(𝜑−𝜔−𝜆−𝜂)𝓏 Ex, E1−x, and� + 𝑢 E�x𝓏, respectively. Then:��+ 𝑅(𝑟� −1) + 𝜎+ ∆𝑣� −𝑓−𝐶�� (9)
_Ex = y[I + [(+ (1 −𝑦)�𝑅(𝑟ϕ −_ _ω −_ _λ −_ �η−1) + 𝜎−𝑓�)ZA + uAZB] + R(r1 − 1) + σ + ∆vc − _f −_ _Cb]_ (9)
𝐸��� = 𝑦�𝑅𝑟� + 𝑣� + �(𝜑−𝜔−𝜆−𝜂)𝓏+(1 − _y)[R(r1 −_ 1) +� _σ+ 𝑢 −�𝓏f ]��−𝑓−𝑝�_ −𝐶�� (10)
_E1−x = y+ (1 −𝑦)(𝑅𝑟[Rr3 + vs + [(�_ + 𝑣ϕ −� _ω−𝑓−𝑝 −_ _λ −�η) )ZA + uAZB] −_ _f −_ _p2 −_ _Cb]_ (10)
+(���= 𝑥𝐸𝐸1� − _y)(� Rr+ (1 −𝑥)𝐸3 + vs −_ _f��� −_ _p1)_ (11)
_Ex = xEx + (1 −_ _x)E1−x_ (11)
The replication dynamics equation (RDE) [68] of the SME is denoted as follows:
The replication dynamics equation (RDE) [68] of the SME is denoted as follows:
𝐹(𝑥) = [𝑑𝑥]
𝑑𝑡
= 𝑥(𝐸� −𝐸F���)(�x) = _[dx]dt_
= 𝑥(1 −𝑥)(𝐸= x��E−𝐸x −���Ex)� (12) (12)
= 𝑥(1 −𝑥)�𝑦(𝐼+ ∆𝑣= x(1 − _x)(�E+ 𝑝x −�_ _E−𝑝1−�x) + (𝑟)_ � −𝑟� −1) 𝑅+ 𝜎+ 𝑝� −𝑣��
= x(1 − _x)[y(I + ∆vc + p2 −_ _p1) + (r1 −_ _r3 −_ 1) R + σ + p1 − _vs]_
Let 𝐹(𝑥) = 0, and we obtain the stationary point of the differential equation as follows: Let F(x) = 0, and we obtain the stationary point of the differential equation as follows:
_x𝑥1[∗]�∗[=]= 0, 𝑥[ 0,][ x]2[∗]�∗_ [=]= 1[ 1] (13) (13)
_y[∗]𝑦[∗]= = [(][(1 + 𝑟][1][ +][ r][3][�][ −][−𝑟][r][1][�][)][)𝑅+ 𝑣][R][ +][ v][s][�][ −][−𝜎−𝑝][σ][ −]_ _[p][1][�]_ (14) (14)
_I +𝐼+ ∆𝑣 ∆vc� ++ 𝑝 p2� −−𝑝p1�_
Based on Equation (14), we can discover that, as shown in Figure 3:
Based on Equation (14), we can discover that, as shown in Figure 3:
(a) (b)
**Figure 3. (a) The dynamic trend of the SME’s strategy in the case of** 𝑦> 𝑦[∗]; (b) The dynamic trend
**Figure 3. (a) The dynamic trend of the SME’s strategy in the case of y > y[∗]; (b) The dynamic trend of**
of the SME’s strategy in the case of 𝑦< 𝑦[∗].
the SME’s strategy in the case of y < y[∗].
When When y𝑦= 𝑦 = y[∗], the LF can access the CBLP to use the information on-chain with a [∗], the LF can access the CBLP to use the information on-chain with a
��(�)
probability of probability of y𝑦[∗][∗], and, and _[∂][F]∂��[(]x[x][)]_ = 0= 0 always holds. That is, the state is always stable, regardless always holds. That is, the state is always stable, regard
less of the value of of the value of x. Moreover, any change in other exogenous variables will not alter the𝑥. Moreover, any change in other exogenous variables will not alter
the stability of the state. stability of the state. x is the equilibrium point, and all states are stable.𝑥 is the equilibrium point, and all states are stable.
WhenWhen𝑦�𝑦 y ̸= y[∗], the system needs to satisfy two requirements to obtain evolutionary [∗], the system needs to satisfy two requirements to obtain evolutionary
stability, i.e., stability, i.e., F𝐹(𝑥(x[∗][∗]) = 0) = 0 and and F𝐹′(𝑥[′](x[∗][∗])) < 0 < 0. Then:. Then:
- In the case of 𝑦> 𝑦[∗], 𝐹[�](1) < 0. 𝑥= 𝑥�∗ = 1 is an evolutionary stable strategy (ESS).
When the probability of the LF accessing the CBLP is larger than 𝑦[∗], the SME will
-----
_J. Theor. Appl. Electron. Commer. Res. 2023, 18_ 219
_•_ In the case of y > y[∗], F[′](1) < 0. x = x2[∗] [=][ 1 is an evolutionary stable strategy (ESS).]
When the probability of the LF accessing the CBLP is larger than y[∗], the SME will
converge with the equilibrium strategy of “comply with the LC”. The number of SMEs
who abide by the contract will gradually increase.
_•_ In the case of y < y[∗], F[′](0) < 0. x = x1[∗] [=][ 0 is an evolutionary stable strategy (ESS). It]
implies that more SMEs will eventually evolve into a stable state of defaulting on the
LC, since LFs struggle to distinguish the forgery of credit records without BCT [69].
According to Equation (14), we can find that the probability of the LF choosing to
access the CBLP, requiring the SME to join the consortium, is small, and the SME tends to
breach the LC. Moreover, the “comply with or default on the LC” decision of the SME has
nothing to do with the information sharing (Zi, uj), the asset maintenance fee ( _f_ ), or the
consortium membership fee (Cb). In contrast, the determinant of the decision is the size of
the gap between the lease reinvestment earnings (i.e., (1 + r3 − _r1)R) that the SME would_
gain for the default and the rewards (i.e., σ) it would receive for its compliance.
In addition, Equation (14) gives some further insights that the residual value of the
leased asset (vs) is positively correlated with the probability y that the LF chooses to access
the CBLP. This is because if the LC provisions are that the lessors (i.e., LF) ensure that
the residual value of the leased asset is immutably recorded on-chain, it mitigates the
uncertainty of residual value risk, aiding the lessor to retain ownership of the asset at the
end of the lease. Meanwhile, the default margin penalty ( _[p][2]p[−]1[p][1]_ ) imposed on the SME is
negatively correlated with y. When _[p][2]p[−]1[p][1]_ decreases, the probability of the LF choosing to
access the CBLP increases, the main reason for which is that the relatively small penalty
(p2) set up on-chain can effectively reduce the default risk of the SME, which makes the
LF is more willing to access the CBLP. In addition, owing to the compliance behavior, the
higher credit (∆vc) achieved by the SME will stimulate the LF to stick with the conventional
leasing mode.
6.1.2. Replication Dynamic Equation of the LF
Assuming the expected return of the LF accessing and not accessing the CBLP to utilize
the information shared on-chain, the average returns are Ey, E1−y, and Ey, respectively. Then:
_Ey_ = x[R(r2 + 1) + vs + [(ϕ − _ω −_ _λ −_ _η)ZB + uBZA] + g −_ _C0 −_ _C2 −_ _Cb]_ (15)
+(1 − _x)[[(ϕ −_ _ω −_ _λ −_ _η)ZB + uBZA] −_ _Rε −_ _C0 −_ _C2 −_ _Cb]_
_E1−y = x[R(r2 + 1) + vs −_ _C0 −_ _Ct −_ _C1] + (1 −_ _x)[−Rε −_ _C0 −_ _Ct −_ _C1]_ (16)
_Ey_ = yEy + (1 − _y)E1−y_ (17)
The replication dynamics equation (RDE) [68] of the LF is denoted as follows:
_F(y) =_ _[dy]dt_
= y�Ey − _Ey�_
= y(1 − _y)�Ey −_ _E1−y�_
= y(1 − _y)[gx + ((ϕ −_ _ω −_ _λ −_ _η)ZB + uBZA + Ct + C1 −_ _C2 −_ _Cb)]_
(18)
Let F(y) = 0, and we obtain the stationary point of the differential equation as follows:
_y1[∗]_ [=][ 0,][ y]2[∗] [=][ 1,] (19)
_x[∗]_ = _[C][b][ +][ C][2][ −]_ [(][ϕ][ −] _[ω][ −]_ _[λ][ −]_ _[η][)][Z][B][ −]_ _[u][B][Z][A][ −]_ _[C][t][ −]_ _[C][1]_ (20)
_g_
Based on Equation (20), we can discover that, as shown in Figure 4:
-----
_JTAERJ. Theor. Appl. Electron. Commer. Res. 2023, 18, FOR PEER REVIEW 2023, 18_ 15 220
(a) (b)
**Figure 4. (a) Figure 4. (aThe dynamic trend of the LF’s strategy in the case of ) The dynamic trend of the LF’s strategy in the case of𝑥> 𝑥 x > x[∗]; [∗](b); (b The dynamic trend of ) The dynamic trend of**
the LF’s strategy in the case of 𝑥< 𝑥[∗].
the LF’s strategy in the case of x < x[∗].
��(�)
When When𝑥= 𝑥 x = x[∗][∗], the SME complies with the LC with a probability of, the SME complies with the LC with a probability of x𝑥[∗][∗], and, and [∂][F]∂[(]y��[y][)] == 0 0 is
is always established. The state is always stable no matter how the value of always established. The state is always stable no matter how the value of y changes.𝑦 changes. In this
In case,this case, y is the equilibrium point, and all states are stable. 𝑦 is the equilibrium point, and all states are stable.
WhenWhen𝑥�𝑥 x ̸= x[∗], the system needs to satisfy two requirements to obtain evolutionary [∗], the system needs to satisfy two requirements to obtain evolutionary
stability, i.e., stability, i.e., F𝐹(𝑦(y[∗][∗]) = 0) = 0 and and F𝐹′(𝑦[′](y[∗][∗])) < 0 < 0. Then:. Then:
- _•_ In the case of In the case of𝑥> 𝑥 x > x[∗], [∗],𝐹 F[�](0) < 0[′](0) < 0.. 𝑦= 𝑦 y = y�∗1[∗]= 0[=][ 0 is an evolutionary stable strategy (ESS).] is an evolutionary stable strategy (ESS).
When the probability of SME compliance is larger than When the probability of SME compliance is larger than𝑥 x[∗], the LF converges to the [∗], the LF converges to the
equilibrium strategy of “not accessing the CBLP ”, and thereby the SME does not need
equilibrium strategy of “not accessing the CBLP ”, and thereby the SME does not
to join the consortium blockchain to share information.
need to join the consortium blockchain to share information.
- _•_ In the case of In the case of𝑥< 𝑥 x < x[∗], [∗],𝐹 F[�](1) < 0[′](1) < 0.. 𝑦= 𝑦 y = y�∗2[∗]= 1[=][ 1 is an evolutionary stable strategy (ESS).] is an evolutionary stable strategy (ESS).
When the probability of SME compliance is less than x[∗], the LF will converge with
When the probability of SME compliance is less than 𝑥[∗], the LF will converge with
the equilibrium strategy of “access the CBLP ” to participate in information sharing
the equilibrium strategy of “access the CBLP ” to participate in information sharing
on-chain to complete the lease.
on-chain to complete the lease.
According to Equation (20), we can find that considering the long-term cooperation,
According to Equation (20), we can find that considering the long-term cooperation,
when the SME is more likely to abide by the LC, the LF will decide not to access the CBLP
when the SME is more likely to abide by the LC, the LF will decide not to access the CBLP
due to the limited synergy and information utilization benefits obtained.
due to the limited synergy and information utilization benefits obtained.
In addition, Equation (20) gives some further insights that the asset maintenance
In addition, Equation (20) gives some further insights that the asset maintenance cost
cost ( _f_ ) will not affect the SME’s decision to comply or default, since once an outsourced
(𝑓) will not affect the SME’s decision to comply or default, since once an outsourced
maintenance action begins, it will not be interrupted until the lease expires. Both the
maintenance action begins, it will not be interrupted until the lease expires. Both the consortium membership fee (consortium membership fee (correlated with the x. When the CBLP sets up higher costs for stakeholders to join the𝐶�) and asset monitoring cost on-chain (Cb) and asset monitoring cost on-chain (𝐶�) are positively corre-C2) are positively
lated with the 𝑥. When the CBLP sets up higher costs for stakeholders to join the consor
consortium, to improve the lease willingness of the LF, the SME is more inclined to comply
tium, to improve the lease willingness of the LF, the SME is more inclined to comply with
with the LC and provide genuine lease information. When the leased asset is fully inspected
the LC and provide genuine lease information. When the leased asset is fully inspected
under the CBLP, the SME will not easily default by deferring the lease payment or not
under the CBLP, the SME will not easily default by deferring the lease payment or not
returning the leased asset after signing the LC on-chain. Notably, the LF has a higher
returning the leased asset after signing the LC on-chain. Notably, the LF has a higher abil
ability to absorb more high-quality information that the SME shares on-chain, which can
ity to absorb more high-quality information that the SME shares on-chain, which can ex
expedite the SME’s default behavior. It seems to be a paradox that is contrary to the LF’s
pedite the SME’s default behavior. It seems to be a paradox that is contrary to the LF’s
decision-making in terms of accessing or not accessing the CBLP. This is because when
decision-making in terms of accessing or not accessing the CBLP. This is because when BCT empowers more information synergy for the LF, the LF is more willing to access the
BCT empowers more information synergy for the LF, the LF is more willing to access the CBLP, which compels the SME to bear the consortium membership cost, data storage, and
CBLP, which compels the SME to bear the consortium membership cost, data storage, and verification overhead. To compensate for the potential losses that may be suffered, the
verification overhead. To compensate for the potential losses that may be suffered, the SME will take risks, opting for default, decreasing the likelihood of compliance. The whole
SME will take risks, opting for default, decreasing the likelihood of compliance. The whole process will finally be formed as an unstable circle.
process will finally be formed as an unstable circle.
_6.2. Analysis of Equilibrium Stability and ESS_
_6.2. Analysis of Equilibrium Stability and ESS Based on the above analysis, the game system has five local equilibrium points:_
(0, 0Based on the above analysis, the game system has five local equilibrium points: ), (1, 0), (0, 1), (1, 1), and (x[∗], y[∗]).
(0, 0), (1, 0), (0, 1), (1, 1), and (𝑥[∗], 𝑦[∗]).
-----
_J. Theor. Appl. Electron. Commer. Res. 2023, 18_ 221
To find the evolutionary stability strategy (ESS), the local stability analysis of the
Jacobian matrix is employed [70,71], and thereby we take the first-order derivatives of
Equations (12) and (18), respectively, achieving the following Jacobian matrix J.
(21)
**_J =_**
_∂F(x)_ _∂F(x)_
_∂x_ _∂y_
_∂F(y)_ _∂F(y)_
_∂x_ _∂y_
where:
_∂F(x)_
_∂x_ = (1 − 2x)[y(I + ∆vc + p2 − _p1) + (r1 −_ _r3 −_ 1) R + σ + p1 − _vs]_ (22)
_∂F(x)_
_∂y_ = x(1 − _x)(I + ∆vc + p2 −_ _p1)_ (23)
_∂F(y)_
= y(1 _y)g_ (24)
_−_
_∂x_
_∂F(y)_
_∂y_ = (1 − 2y)[gx + ((ϕ − _ω −_ _λ −_ _η)ZB + uBZA + Ct + C1 −_ _C2 −_ _Cb)]_ (25)
Next, we can calculate the trace value trJ and determinant value detJ of the Jacobian
matrix J.
_trJ =_ _[δ][F][(][x][)]_ + _[δ][F][(][y][)]_ (26)
_δx_ _δy_
� _δF(x)_
_detJ =_
_δx_
�� _δF(y)_
_δy_
� � _δF(x)_
_−_
_δy_
�� _δF(y)_
_δx_
�
(27)
Thus, the trJ and detJ to the equilibrium point are shown in Table 3.
**Table 3. The analysis table for judging the stability of equilibrium points.**
**Equilibrium Point** **trJ** **detJ**
_E1(0, 0)_ [(r1 − _r3 −_ 1) R ++[(σ +ϕ − p1 −ω −vsλ] _−_ _η)ZB + uBZA_ [(r1 − _r3 −_ 1) R +∗[(σ +ϕ − p1ω − −vλs] − _η)ZB + uBZA_
+Ct + C1 − _C2 −_ _Cb]_ +Ct + C1 − _C2 −_ _Cb]_
_E2(0, 1)_ [(r1 − _r3 −_ 1) R+ (+[IC +b + ∆ Cvc2 + − p(ϕ2 + − σω − −vλs −)] _η)ZB_ [(r1 − _r3 −_ 1) R +(∗[CI +b + ∆ Cv2c − + p(ϕ2 + − σω − −vλs −)] _η)ZB_
_−uBZA −_ _Ct −_ _C1]_ _−uBZA −_ _Ct −_ _C1]_
_E3(1, 0)_ [(1 + r3 − _r1)R +(+[vgs − + (σϕ − −pω1)] −_ _λ −_ _η)ZB_ [(1 + r3 − _r1)R +(∗[vgs + ( −_ _σϕ − −pω1 −)]_ _λ −_ _η)ZB_
+uBZA + Ct + C1 − _C2 −_ _Cb]_ +uBZA + Ct + C1 − _C2 −_ _Cb]_
_E4(1, 1)_ [−(I + ∆vc + p2 ++[σ− −(gv + (s + (ϕ −r1 −ω −r3 −λ −1)ηR))]ZB [−(I + ∆vc + p2 +∗[−σ −(g + (vs + (ϕ −r1ω − −r3λ − −1η)R)Z)]B
+uBZA + Ct + C1 − _C2 −_ _Cb)]_ +uBZA + Ct + C1 − _C2 −_ _Cb)]_
_E5(x[∗], y[∗])_ 0 _H_ [1]
1 H = −x∗(1 − _x∗)(I + vc + p2 −_ _p1) ∗_ _y∗(1 −_ _y∗)_
The local stability analysis of the five equilibria was performed to investigate the relationship between positive and negative trJ and detJ and evolutionary stability at the five
equilibrium points. When a local equilibrium point satisfies the conditions that trace trJ < 0
and the determinant detJ > 0 of the Jacobian matrix J, it is an evolutionary stable strategy
(ESS) [72]. If the trJ > 0 and detJ > 0, the equilibrium point is unstable or a saddle point.
Nonetheless, it is obvious that the stationary point (x[∗], y[∗]) should meet
0 _x[∗]_ 1, 0 _y[∗]_ 1, which is meaningful. Considering g > 0 and
_≤_ _≤_ _≤_ _≤_
-----
_J. Theor. Appl. Electron. Commer. Res. 2023, 18_ 222
Nonetheless, it is obvious that the stationary point (𝑥[∗], 𝑦[∗]) should meet 0 �𝑥[∗] �
1, 0 �𝑦[∗] �1, which is meaningful. Considering 𝑔> 0 and
0 �0 ≤ �[C]�[b]��[+][C]��(�������)𝓏[2][−][(][ϕ][−][ω][−][λ][−][η]g�[)]��[Z][B]�[−]𝓏[u]�[B]��[Z]�[A]��[−][C]� _[t]�1[−][C][1], demonstrating that: ≤_ 1, demonstrating that:
�
𝐶� + 𝐶Cb� +−(𝜑−𝜔−𝜆−𝜂)𝓏 C2 − (ϕ − _ω −_ _λ −�_ −𝑢η)Z�B𝓏 −� −𝐶uBZ� −𝐶A −� _C> 0t −_ _C1 > 0_
Condition 1: Condition 1 :� 𝑔> 0 _g > 0_ (28) (28)
(𝜑−𝜔−𝜆−𝜂)𝓏(ϕ − _ω −_ _λ −�_ _η+ 𝑢)Z�B𝓏 +�_ _u+ 𝐶BZ�_ _A+ 𝐶 +� C+ 𝑔−𝐶t + C1 +�_ _g−𝐶 −�C> 02 −_ _Cb > 0_
nominator of nominator ofSimilarly, according to Assumption 3, we know that Similarly, according to Assumption 3, we know that y𝑦[∗][∗],, I𝐼+ ∆𝑣 + ∆� _v+ 𝑝c +� p−𝑝2 −�_ - 0p1, then: > 0, then:0 � (�����∆��0�� p�𝑝�≤)���2��� >> 𝑝���� p(����1+��1, which indicates the de-, which indicates the de-rI3+−�∆r �11v)cR++p, demonstrating 2v−s−pσ1 _−p1_ _≤_ 1,
that: demonstrating that:
Condition 2: � (1 + 𝑟𝐼+ ∆𝑣� −𝑟�)𝑅+ 𝑣�(+ 𝑝1 +� r�−𝑝3 −−𝜎−𝑝�r1> 0)R +� - 0 vs − _σ −_ _p1 > 0_ (29)
Condition 2 : _I + ∆vc + p2_ _p1 > 0_ (29)
𝐼+ ∆𝑣� + 𝑝�I+ 𝜎−𝑣 + ∆vc +� −(1 + 𝑟 p2 + σ −� −𝑟vs −� −)𝑅> 0(1 + r3 − _r1)R > 0_
Based on Condition 1 and Condition 2, we can use the signs of tr𝑱 and det𝑱 to judge
Based on Condition 1 and Condition 2, we can use the signs of trJ and detJ to judge
the stability of the equilibrium point of the evolutionary game. The results are shown in
the stability of the equilibrium point of the evolutionary game. The results are shown in
Table 4.
Table 4.
**Table 4. The analysis of the evolutionary stability of the system equilibrium point.**
**Table 4. The analysis of the evolutionary stability of the system equilibrium point.**
**Equilibrium Point** 𝑬𝒊 𝐒𝐲𝐦𝐛𝐨𝐥 𝐨𝐟 𝐭𝐫𝑱 𝐒𝐲𝐦𝐛𝐨𝐥 𝐨𝐟 𝐝𝐞𝐭𝑱 **Judgment**
**Equilibrium Point𝐸�(0,0)** **_Ei_** **Symbol of tr<0** **_J_** **Symbol of det>0** **_J_** **JudgmentESS**
𝐸EE�12(0,1)((0, 00, 1)) >0 <0>0 >0 >0>0 Unstable point Unstable pointESS
𝐸E�3(1,0)(1, 0) >0 >0 >0 >0 Unstable point Unstable point
𝐸E�4(1,1)(1, 1) <0 <0 >0 >0 ESS ESS
𝐸E�5(𝑥(x[∗][∗], 𝑦, y[∗][∗])) 0 0 +/-+/- Saddle point Saddle point
According to Table 4, we can find that According to Table 4, we can find that E𝐸�2(0,1)(0, 1) and and E𝐸�3(1,0)(1, 0) are unstable points. are unstable points.
𝐸E�5(𝑥(x[∗][∗], 𝑦, y[∗][∗])) is a saddle point revealing that evolutionary stability is affected by the values is a saddle point revealing that evolutionary stability is affected by the values
of of x𝑥[∗] [∗]andand y[∗]. The game system has two ESS equilibrium points:𝑦[∗] . The game system has two ESS equilibrium points: E1(0, 0) and E𝐸�4((0,0)1, 1) and . This
𝐸indicates that the game’s ultimate evolutionary strategies are “Strategy I:�(1,1). This indicates that the game’s ultimate evolutionary strategies are “Strategy I: S1 = {Comply,
SAccess� = {Comply, Access}” and “Strategy IV:}” and “Strategy IV: S4 = {Default, Not-accessS� = {Default, Not-access}”, meaning that both SMEs and LFs}”, meaning that both
SMEs and LFs converge at the locations converge at the locations E1 and E4 in Figure𝐸� and 5. 𝐸� in Figure 5.
**Figure 5. Figure 5. Dynamics evolution schematic diagram of the SME and the LF.Dynamics evolution schematic diagram of the SME and the LF.**
That is, when both parties’ decisions are in the region E1E2E5E3, the game evolves to
point E1(0, 0), i.e., the SME breaches the LC, and the LF does not access the CBLP, requiring
the SME to access the consortium blockchain. When both parties’ decisions are located
in region E2E4E3E5, the game evolves into the ideal stable state E4(1, 1), i.e., the SME
-----
_J. Theor. Appl. Electron. Commer. Res. 2023, 18_ 223
abides by the LC and the LF requires the SME to access the consortium blockchain. The
probability of the evolutionary outcome between the game subjects can be represented in
terms of the area of the regions E1E2E5E3 and E2E4E3E5 [73], the size of which depends on
the coordinates of the point E5 (the saddle point (x[∗], y[∗])), where: SE1E2E5E3 = 2[1] [(][x][∗] [+][ y][∗][)][,]
_SE2E4E3E5 =_ 2[1] [[(][1][ −] _[x][∗][) + (][1][ −]_ _[y][∗][)]][. The possibility that the SME will conform to the contract]_
and be required to access the CBLP increases as the region E2E4E3E5 expands.
_6.3. Sensitivity Analysis in the Evolutionary Game_
The choices made by the game subjects are influenced by the exogenous variables in
the model. By taking derivatives of SE2E4E3E5 (abbreviated hereafter as ‘S’) while holding
the other parameters constant, it is possible to determine how each parameter affects the
game’s evolutionary results (i.e., SE2E4E3E5 ).
_x[∗]_ = _[C][b][ +][ C][2][ −]_ [(][ϕ][ −] _[ω][ −]_ _[λ][ −]_ _[η][)][Z][B][ −]_ _[u][B][Z][A][ −]_ _[C][t][ −]_ _[C][1]_ (30)
_g_
_y[∗]_ = [(][1][ +][ r][3][ −] _[r][1][)][R][ +][ v][s][ −]_ _[σ][ −]_ _[p][1]_ (31)
_I + ∆vc + p2_ _p1_
_−_
_SE2E4E3E5 = S =_ 21 ��1 − _[C][b][+][C][2][−][(][ϕ][−][ω][−][λ][−][η]g[)][Z][B][−][u][B][Z][A][−][C][t][−][C][1]_ �
(32)
� ��
+ 1 − [(][1][+][r]I[3]+[−]∆[r][1]v[)]c[R]+[+]p2[v]−[s][−]p[σ]1 _[−][p][1]_
The evolutionary game results are primarily related to the four main determinants:
information sharing, credit, incentive–penalty, and risk. The sensitivity analysis of the
influence of the four factors on SE2E4E3E5 is described below.
6.3.1. Impact of Information Sharing on S
Taking the derivatives of Equation (32) corresponding to ZA, ZB, and uB,
_∂S_
= _[u][B]_ (33)
_∂ZA_ 2g _[>][ 0]_
_∂S_
= _[ϕ][ −]_ _[ω][ −]_ _[λ][ −]_ _[η]_ _> 0_ (34)
_∂ZB_ 2g
_∂S_
= (35)
_[Z][A]_
_∂uB_ 2g _[>][ 0]_
_S is an increasing function of ZA and ZB, since Equation (34) is true under Assumption_
7 (A7). That is, as the amount of information shared on-chain (Zi) increases, S will gradually
increase, indicating the possibility of evolution to the stable state E4(1, 1) as the quantity of
information shared on-chain increases. It further means that this parameter has a favorable
impact on the probability of SMEs’ decisions to comply with the contract and being required
to join the consortium. The more accurate the information that is shared on the blockchain
network, the easier it will be to create a transparent and reliable environment for leasing,
and the more SMEs will actively disclose high-quality information to ensure that leasing
transactions are executed smoothly.
_S is an increasing function of uB. The more the LF can use the effective data on-chain,_
the more the LF is likely to access the CBLP to clearly monitor the SME’s compliance with
the LC, thus increasing the likelihood of rental payment on time.
6.3.2. Impact of Credit on S
Taking the derivatives of Equation (32) corresponding to ∆vc,
_∂S_
= [(][1][ +][ r][3][ −] _[r][1][)][R][ +][ v][s][ −]_ _[σ][ −]_ _[p][1]_ _> 0_ (36)
_∂∆vc_ 2(I + ∆vc + p2 _p1)[2]_
_−_
-----
_J. Theor. Appl. Electron. Commer. Res. 2023, 18_ 224
_S is an increasing function of ∆vc. Therefore, as ∆vc increases, the SME has a higher_
probability of conforming to the LC. Then, a higher credit value is assigned to the SME,
resulting in the SME’s nodes having a higher probability of being selected as a leader node
when performing the Raft consensus protocol. Conversely, a node will be removed from the
_Consenter Set if its total credit value falls below the minimum threshold because of multiple_
defaults, escalating the penalty and helping to establish a high-credit leasing environment.
6.3.3. Impact of Incentive–Penalty on S
Taking the derivative of Equation (32) corresponding to σ, I, g and p2,
_∂S_ 1
(37)
_∂σ_ [=] 2(I + ∆vc + p2 _p1)_ _[>][ 0]_
_−_
_∂S_
_> 0_ (38)
_∂I_ [= (][1][ +]2([ r]I[3] +[ −] ∆[r][1]v[)]c[R] +[ +] p[ v]2 _[s][ −]p[σ]1)[ −][2]_ _[p][1]_
_−_
_∂S_
_> 0_ (39)
_∂g_ [=][ C][b][ +][ C][2][ −] [(][ϕ][ −] _[ω][ −]_ _[λ][ −]2g[η][2][)][Z][B][ −]_ _[u][B][Z][A][ −]_ _[C][t][ −]_ _[C][1]_
_∂S_
= [(][1][ +][ r][3][ −] _[r][1][)][R][ +][ v][s][ −]_ _[σ][ −]_ _[p][1]_ _> 0_ (40)
_∂p2_ 2(I + ∆vc + p2 _p1)[2]_
_−_
_S is an increasing function of σ, I, g, and p2. The likelihood that previously defaulting_
SMEs will start to keep their contracts and that the SME is motivated to access the blockchain
increases with the incentives the LF provides to them. To encourage a node to choose the
on-chain strategy and to encourage more SMEs to become consortium nodes, the LF can
appropriately boost the compliance reward of the SMEs when creating the incentive strategy.
No matter whether the SME defers the rental payment or refuses to return the leased asset,
both default behaviors will lead to the lessee being charged a penalty, which irrevocably
damages the SME’s reputation on-chain. Hence, once the SME joins the consortium, the
likelihood of compliance rises as the default penalties rise. The LF will also keep using the
on-chain strategy to observe how the SMEs choose their payment strategies. Therefore, the
penalties for SMEs should be suitably enhanced to guarantee a prompt rental payment.
6.3.4. Impact of Risk on S
Taking the derivatives of Equation (32) corresponding to ϕ, ω, λ, and η,
_∂S_
(41)
_∂ϕ_ [=][ Z]2g[B] _[>][ 0]_
_∂S_
_< 0_ (42)
_∂ω_ [=][ −Z]2g[B]
_∂S_
_< 0_ (43)
_∂λ_ [=][ −Z]2g[B]
_∂S_
_< 0_ (44)
_∂η_ [=][ −Z]2g[B]
_S is an increasing function of ϕ. When more high-quality data are effectively dis-_
tributed and shared by each subject on-chain, more subjects will join the consortium to
complete the leasing transactions as more return is generated.
In addition, S is a decreasing function of ω, λ, and η. If the participants bear more
consensus verification and storage costs, and take on greater security risks, they will be
more reluctant to join the consortium and the probability of default will increase.
To sum up, different types of factors will have different effects on decision-making.
-----
_J. Theor. Appl. Electron. Commer. Res. 2023, 18_ 225
**7. Numerical Experiments and Implications**
This section will present some simulation results. We first use VENSIM PLE to build a
system dynamics (SD) model to analyze the causal relationships among the variables and
strategies. Then, MATLAB_R2021b is employed to examine the efficacy of the evolutionary
stable strategies (ESSs) and to demonstrate the previous mathematical sensitivity analysis
of each factor. Lastly, some implications of the results are given.
The game system has intermediate variables (including Ex, E1−x, Ey, _E1−y) and_
a range of exogenous variables (as presented in Table 1). We set initial values for the
exogenous variables involved in the model, as shown in Table 5. In this study, it is assumed
that all exogenous variables are positive, and the return of each strategy of each game
subject is guaranteed to be positive.
**Table 5. Initial value of simulation parameters.**
**_R_** **_r1_** **_r2_** **_r3_** **_f_** **_vs_** **_p1_** **_p2_** **_Cb_** **_Ct_** **_C1_** **_C2_** **∆vc**
8 0.2 0.25 0.3 0.05 0.8 3 6 4 0.08 0.6 0.4 1
_I_ _ZA_ _ZB_ _uA_ _uB_ _ε_ _σ_ _g_ _ϕ_ _ω_ _λ_ _η_ /
1.2 5 3 0.5 0.5 0.15 5 1.5 0.6 0.2 0.2 0.2 /
_7.1. System Dynamics Model Experiment_
We establish the SD model of the two-party evolutionary game system as depicted in
Figure 6. The arrow tails in Table 5 are connected to the independent variables in the associated
equation, and the arrowheads are connected to the dependent variables. We set the simulation
parameters of INITIAL TIME = 0, FINAL TIME = 10, and TIME STEP = 0.0078125.
It can be seen from Figure 7a,b that when the initial states of both sides are pure
strategies (i.e., (0, 0), (0, 1), (1, 0), and (1, 1)), no party in the system is willing to change
the current state to break the equilibrium. For instance, the initial state of (x = 0, y = 0) or
(x = 1, y = 1) will be unchanged if there is no interruption during the evolution. However,
this does not mean that these equilibrium states are stable, and once one or both parties
take the initiative to make a small change, the equilibrium state will be broken. Although
the SME’s compliance probability x and the LF’s CBLP access probability y (for 0.0001)
evolve with small mutations, they quickly shift to a new strategy once they find that doing
this will yield a higher expected return, thus adjusting the strategy through a mutation of
parties to bring the system into a new equilibrium. In addition, through simulating the
model, we also discover that the ultimate equilibrium state is (1, 1) when the initial state is
_x = 0.5, y = 0.5, as shown in Figure 7c._
In fact, when x = 0, no matter how y ranges from 0 to 1, the system will reach an
equilibrium state (0, 0). When x = 1, no matter how y ranges from 0 to 1, the system will
reach an equilibrium state (1, 1). Similarly, when y = 0, no matter how x ranges from 0
to 1, the system will reach an equilibrium state (0, 0), and when y = 1, no matter how x
ranges from 0 to 1, the system will reach an equilibrium state (1, 1).
_7.2. Effect of Parameter Changes on Evolutionary Stable Strategies_
We initiate the probabilities of x and y with values ranging from 0 to 1 in steps of 0.1.
It can be seen from Figure 8 that almost all curves converge at (0, 0) and (1, 1), which is
consistent with the preceding discussion in Section 6.2.
The following subsection further discusses the impacts of the four factors on evolution.
Here, we assume the initial strategy probability for each participant is 0.5.
-----
##### E IE
_J. Theor. Appl. Electron. Commer. Res. 2023, 18_ 226
### Figure 6. System dynamics model for the consortium blockchain-based leasing strategies.Figure 6. System dynamics model for the consortium blockchain-based leasing strategies.
## It can be seen from Figure 7a,b that when the initial states of both sides are pure strategies (i.e., (0, 0), (0, 1), (1, 0), and (1, 1)), no party in the system is willing to change the current state to break the equilibrium. For instance, the initial state of (𝑥= 0, 𝑦= 0 or (𝑥= 1, 𝑦= 1) will be unchanged if there is no interruption during the evolution. How ever, this does not mean that these equilibrium states are stable, and once one or both parties take the initiative to make a small change, the equilibrium state will be broken Although the SME’s compliance probability 𝑥 and the LF’s CBLP access probability 𝑦 (fo
-----
, 18, FOR PEER REVIEW J. Theor. Appl. Electron. Commer. Res. 2023, 18 22 227
##### (a)
(b)
(c)
**Figure 7. (a) The dynamic diagram to strategy (0,0); (Figure 7. (a) The dynamic diagram to strategy (0,0); (b) the dynamic diagram to strategy (1,1); (b) the dynamic diagram to strategy (1,1); (c)** **c) the**
the dynamic diagram to strategy of dynamic diagram to strategy of𝑥= 0, 𝑦= {0.1,0.3,0.5,0.7,0.9} x = 0, y = {0.1, 0.3, 0.5, 0.7, 0.9. * Considering that }. * Considering that𝑑𝑥/𝑑𝑡 and dx/dt and
𝑑𝑦/𝑑𝑡 have to be explanatory, when performing the simulation, we take the initial state (dy/dt have to be explanatory, when performing the simulation, we take the initial state (x=0.0001, x = 0.0001,
_y=0.0001), which is close to 0. Similarly, the initial state (y = 0.0001), which is close to 0. Similarly, the initial state (x =0.9999, y=0.9999x = 0.9999,) is set as 1. y = 0.9999) is set as 1._
##### In fact, when 𝑥= 0, no matter how 𝑦 ranges from 0 to 1, the system will reach an ilib i (0 0) Wh 1 h f 0 1 h
-----
#### p 𝑦 g g
_J. Theor. Appl. Electron. Commer. Res. 20230.1. It can be seen from Figure 8 that almost all curves converge at, 18_ (0, 0) and 228 (1, 1)
#### is consistent with the preceding discussion in Section 6.2.
##### Figure 8. The dynamic diagram of SMEs and LFs.
#### The following subsection further discusses the impacts of the four factors on evolu- tion. Here, we assume the initial strategy probability for each participant is 0.5.
#### 7.2.1. Evolution Impacted by Information Sharing
**Figure 8. The dynamic diagram of SMEs and LFs.**
##### Figure 8. The dynamic diagram of SMEs and LFs.
#### The initial quantity of information sharing (𝓏�) for each participant was set to 1, 3, 5,
7.2.1. Evolution Impacted by Information Sharing
#### 7, and 9. As shown in Figure 9a. there are two critical values. When 𝓏� is greater than 5
The initial quantity of information sharing (The following subsection further discusses the impacts of the four factors onZi) for each participant was set to 1, 3, 5,
#### and less than 3, the probability is that all parties will converge at 1 and 0, respectively. The
tion. Here, we assume the initial strategy probability for each participant is 0.5. 7, and 9. As shown in Figure 9a. there are two critical values. When Zi is greater than 5
system evolves to the states (1, 1) and (0, 0), accordingly. When the computing power and less than 3, the probability is that all parties will converge at 1 and 0, respectively. The𝑢� of each organization is less than 0.3, system evolves to the states (1, 1) and (0, 0), accordingly. When the computing power𝑥 and 𝑦 both converge at 0, and the system evolves uj
7.2.1. Evolution Impacted by Information Sharing
to the state (0, 0of each organization is less than 0.3,) (see Figure 9b). When 𝑢� is greater than 0.5, it results in an evolutionary x and y both converge at 0, and the system evolves state (1, 1), indicating that the parties with higher computation are more willing to be to the state (0, 0) (see FigureThe initial quantity of information sharing ( 9b). When uj is greater than 0.5, it results in an evolutionary𝓏𝑖) for each participant was set to incentivized to join the consortium blockchain to share information. 7, and 9. As shown in Figure 9a. there are two critical values. When stateincentivized to join the consortium blockchain to share information. (1, 1), indicating that the parties with higher computation are more willing to be𝓏𝑖 is greate
and less than 3, the probability is that all parties will converge at 1 and 0, respectiv system evolves to the states (1, 1) and (0, 0), accordingly. When the computing po1 of each organization is less than 0.3, 𝑥 and 𝑦 both converge at x: zi = 1 0, and the system
0.9 _y: z_
#### to the state (0, 0) (see Figure 9b). When 𝑢𝑗 is greater than 0.5, it results in an evolui = 1
_x: z_
#### state 0.8 (1, 1), indicating that the parties with higher computation are more willini = 3
_y: z_
_i = 3_
#### incentivized to join the consortium blockchain to share information. 0.7 x: zi = 5
_y: z_
_i = 5_
0.6 _x: zi = 7_
##### The dynamic diagram of SMEs and LFs.
#### The following subsection further discusses the impacts of the four factors on evolu- tion. Here, we assume the initial strategy probability for each participant is 0.5.
0.5
0.4
0.3
0.2
0.1
0
0 2 4 6 8 10 12 14 16 18 20
t
#### (a)
**Figure 9. Cont.**
-----
FOR PEER REVIEW 24
_J. Theor. Appl. Electron. Commer. Res. 2023, 18_ 229
1
1 _x: uj = 0.1_
0.9 _x: uj = 0.1_ _y: uj = 0.1_
0.9 0.8 _y: uj = 0.1_ _x: uj = 0.3_
_y: u_
_x: u_ _j = 0.3_
0.8 _j = 0.3_ _x: u_
0.7 _y: u_ _j = 0.5_
_j = 0.3_
_y: u_
_x: u_ _j = 0.5_
0.7 _j = 0.5_
0.6 _y: u_ _x: uj = 0.7_
_j = 0.5_ _y: u_
0.6 0.5 _x: uj = 0.7_ _x: uj = 0.7_
_y: u_ _j = 0.9_
_j = 0.7_ _y: u_
0.5 0.4 _x: u_ _j = 0.9_
_j = 0.9_
_y: u_
0.4 0.3 _j = 0.9_
0.3 0.2
0.2 0.1
0.1
0
0 1 2 3 4 5 6 7 8 9 10
t
0 1 2 3 4 5 6 7 8 9 10
#### (b)
t
**Figure 9. System evolution of** 𝓏(�b and ) 𝑢�: (a) system evolution of 𝓏� = {1, 3, 5, 7, 9}; (b) system
lution of 𝑢� = {0.1, 0.3, 0.5, 0.7, 0.9}.
##### Figure 9. System evolution of Figure 9. System evolution of𝓏� and 𝑢�: ( Za) system evolution of i and uj: (a) system evolution of𝓏� = {1, 3, 5, 7, 9} Zi = {1, 3, 5, 7, 9; (b) system evo-}; (b) system lution of 𝑢� = {0.1, 0.3, 0.5, 0.7, 0.9}evolution of7.2.2. Evolution Impacted by Credit uj = {0.1, 0.3, 0.5, 0.7, 0.9. }.
7.2.2. Evolution Impacted by CreditIn order to further study how different levels of credit affect the decision-maki
#### 7.2.2. Evolution Impacted by Credit
SMEs and LFs, we simulate the factor “credit value” in the range from 1 to 10, with aIn order to further study how different levels of credit affect the decision-making of
In order to further study how different levels of credit affect the decision-making of SMEs and LFs, we simulate the factor “credit value” in the range from 1 to 10, with a stepsize of 2, while keeping the other parameters at their initial values. Figure 10 indicates
size of 2, while keeping the other parameters at their initial values. Figure 10 indicates
#### SMEs and LFs, we simulate the factor “credit value” in the range from 1 to 10, with a step all curves gradually converged to 𝑥= 1, 𝑦= 1, indicating that higher credit helps to
that all curves gradually converged to x = 1, y = 1, indicating that higher credit helps to
#### size of 2, while keeping the other parameters at their initial values. Figure 10 indicates that tivate the SME to fulfill the contract and join the blockchain to share their informa
motivate the SME to fulfill the contract and join the blockchain to share their information.
#### all curves gradually converged to However, it also can be found that the SME’s strategy of choosing to keep the contract isHowever, it also can be found that the SME’s strategy of choosing to keep the contr𝑥= 1, 𝑦= 1, indicating that higher credit helps to mo- tivate the SME to fulfill the contract and join the blockchain to share their information. more influenced and motivated by “credit” than the strategy of joining the blockchain.more influenced and motivated by “credit” than the strategy of joining the blockcha However, it also can be found that the SME’s strategy of choosing to keep the contract is more influenced and motivated by “credit” than the strategy of joining the blockchain. 1
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0 2 4 6 8 10 12 14 16 18 20
t
**Figure 10.Figure 10. System evolution ofSystem evolution of ∆vc =∆𝑣 {1, 3, 5, 7, 9�** = {1, 3, 5, 7, 9}}. .
0 2 4 6 8 10 12 14 16 18 20
-----
_J. Theor. Appl. Electron. Commer. Res. 2023, 18_ 230
7.2.3. Evolution Impacted by Incentive-Penalty
_18, FOR PEER REVIEW_ 25
We award an incentive I (e.g., a kind of “gas” fee with blockchain) to the SMEs who
publish a valid block during the financing transaction confirmation process. When we
dynamically adjust the incentive value I from 1.2 to 9.6, we are surprised to find that the
##### matter how many block rewards are paid to the SME for being a block verifier/miner, the system always evolves to the equilibrium point (1, 1) in Figure 11a. This means that no SME resolutely adheres to conforming to the contract and enters the blockchain, whereas matter how many block rewards are paid to the SME for being a block verifier/miner, the with a higher 𝐼SME resolutely adheres to conforming to the contract and enters the blockchain, whereas, the SME is more proactive in participating in information sharing on- chain. with a higher I, the SME is more proactive in participating in information sharing on-chain.
1
_x: I = 1.2_
0.9 _y: I = 1.2_
_x: I = 2.4_
_y: I = 2.4_
0.8
_x: I = 4.8_
_y: I = 4.8_
0.7 _x: I = 7.2_
_y: I = 7.2_
0.6 _x: I = 9.6_
_y: I = 9.6_
0.5
0.4
0.3
0.2
0.1
0
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0 2 4 6 8 10 12 14 16 18 20
t
##### (a)
_x: p2 = 2_
_y: p2 = 2_
_x: p2 = 4_
_y: p2 = 4_
_x: p2 = 6_
_y: p2 = 6_
_x: p2 = 8_
_y: p2 = 8_
_x: p2 = 10_
_y: p2 = 10_
0 2 4 6 8 10 12 14 16 18 20
t
##### (b)
**Figure 11.** (a) System evolution of Figure 11. (a) System evolution of𝐼= {1.2, 2.4, 4.8, 7.2, 9.6} I = _{1.2, 2.4, 4.8, 7.2, 9.6 ; (b) system evolution of }; (b) system evolution of𝑝�_ =
{2, 4, 6, 8, 10}. _p2 = {2, 4, 6, 8, 10}._
##### We set the penalty on-chain (We set the penalty on-chain (𝑝�) to 2, 4, 6, 8, and 10, revealing the evolution curves of p2) to 2, 4, 6, 8, and 10, revealing the evolution curves 𝑥(𝑡) and 𝑦(𝑡) following the change in of x(t) and y(t) following the change in𝑝�, as shown in Figure 11b. The figure shows that p2, as shown in Figure 11b. The figure shows
that when the punishment intensity is relatively small, for instance p2 = 2, the SME tends
##### when the punishment intensity is relatively small, for instance 𝑝� = 2, the SME tends to breach the contract and when the punishment intensity increases to 𝑝= 8 the SME
-----
_J. Theor. Appl. Electron. Commer. Res. 2023, 18_ 231
to breach the contract, and when the punishment intensity increases to p2 = 8, the SME
tends to actively comply with the contract. In other words, the penalty has a threshold that
affects the SME’s strategy selection of SMEs joining the blockchain, which is outside of
_JTAER 2023, 18, FOR PEER REVIEW_
the initial expectations—for example, a high penalty erodes the incentive to participate in
information sharing.
7.2.4. Evolution Impacted by Risk
#### We set the risk cost of the consensus on-chain (𝜔) and storage off-chain (𝜇) to 0.2 0.6, 0.8, and 1.0, and the corresponding impacts on the two parties’ strategies were We set the risk cost of the consensus on-chain (ω) and storage off-chain (µ) to 0.2,
0.4, 0.6, 0.8, and 1.0, and the corresponding impacts on the two parties’ strategies were
#### lyzed. As shown in Figure 12, the critical value of the initial risk cost is between 0.2
analyzed. As shown in Figure 12, the critical value of the initial risk cost is between 0.2 and
#### 0.4. When the risk is less than 0.2, the SME’s and LF’s probabilities 𝑥 and 𝑦 both
0.4. When the risk is less than 0.2, the SME’s and LF’s probabilities x and y both converged
at 1. Vice versa, when the risk level is greater than 0.4, the system evolves to pointverged at 1. Vice versa, when the risk level is greater than 0.4, the system evolves to p (1, 1).
Similar results can be inferred from the influencing security risk ((1, 1). Similar results can be inferred from the influencing security risk (η). 𝜂).
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0 2 4 6 8 10 12 14 16 18 20
t
**Figure 12.Figure 12. System evolution ofSystem evolution of ω = {𝜔= {0.2, 0.4, 0.6, 0.8, 1}0.2, 0.4, 0.6, 0.8, 1}.** .
_7.3. Implications of the Results_
#### 7.3. Implications of the Results
Based on the above replicator dynamic analysis and simulation results, this study
provides some implications:Based on the above replicator dynamic analysis and simulation results, this s
(1)provides some implications: The results reveal that the residual value of the leased asset is a decisive factor
supporting the lessor’s access strategy. Before signing the LC, it is necessary to
#### (1) The results reveal that the residual value of the leased asset is a decisive factor
estimate the asset residual value; if the value is relatively large at the termination of
#### porting the lessor’s access strategy. Before signing the LC, it is necessary to esti
the lease, LFs (lessors) have a high probability of actively adopting BCT to efficiently
prove their ownership of the leased asset on-chain. Thus, from the perspective ofthe asset residual value; if the value is relatively large at the termination of the l
reducing risks of leased asset default, a blockchain-based leasing service provided byLFs (lessors) have a high probability of actively adopting BCT to efficiently p
the lessor is more beneficial for an operating lease than a capital lease.their ownership of the leased asset on-chain. Thus, from the perspective of redu
(2) Most leasing businesses tend to treat maintenance as a non-core activity and com
#### risks of leased asset default, a blockchain-based leasing service provided by the le
monly outsource it to a third-party MC [10], as assumed in this study (Section 4). The
#### is more beneficial for an operating lease than a capital lease.
results indicate that when the maintenance fee is not embedded in the rental payment,
#### (2) Most leasing businesses tend to treat maintenance as a non-core activity and
the maintenance charge is not a determinant impacting the lessee’s decisions regarding compliance with/defaulting on the LC. Hence, before the lessor decides whethermonly outsource it to a third-party MC [10], as assumed in this study (Section 4)
to adopt BCT, it is necessary to take into consideration the in-house or outsourcedresults indicate that when the maintenance fee is not embedded in the rental
maintenance problem.ment, the maintenance charge is not a determinant impacting the lessee’s decis
(3) To encourage lessees and lessors to evolve to the ideal equilibrium state, an incentive
#### regarding compliance with/defaulting on the LC. Hence, before the lessor dec
mechanism should be designed to motivate all parties to cooperatively construct a
#### whether to adopt BCT, it is necessary to take into consideration the in-house or
-----
_J. Theor. Appl. Electron. Commer. Res. 2023, 18_ 232
sustainable and more trustworthy leasing environment. More high-quality information should be shared on-chain, and stakeholders should also improve the capability
to effectively utilize the data on- and off-chain [74]. In contrast to the fixed rewards
resulting from block mining, the incentive associated with incremental or deductible
credit value for consensus action tends to inspire lessees’ willingness to comply with
the contract under the BCT-based leasing business. An appropriate default penalty
should be set up on-chain that can deter the lessee from defaulting and encourage it to
make rental payments on time and return the leased asset as agreed in the LC. When
making strategic decisions to join the consortium to share information, participants
(particularly lessees) are more sensitive to the technology risk factor to which they
are subject. To reduce the cost of building and maintaining the blockchain system to
support the leasing business (e.g., on-chain and off-chain storage costs, verification
costs, etc.), it is advised and helpful to embed blockchain-as-a-service (BaaS) in our
CBLP in the future [75], which will also enhance SMEs’ willingness to share more
valuable information on-chain, achieving a win–win outcome in the leasing business.
**8. Conclusions and Future Works**
_8.1. Conclusions_
BCT provides a new idea for leasing to address the challenges of the information
asymmetry and traceability of leased assets to some degree. Hence, there exists great
significance in designing an incentive mechanism to encourage lessees and lessors to
join the consortium blockchain and actively share information on-chain. This study first
proposes a conceptual architecture of the consortium blockchain-based leasing platform
(CBLP), then constructs a dynamical evolutionary game model between the SME (the
“lessee”) and LF (the “lessor”). Our primary findings are as follows:
(1) With long-term cooperation, the two parties (lessee and lessor) eventually evolve to
adopt strategies in which the lessee is more inclined to conform to the LC and the
lessor becomes more proactive in accessing the CBLP as a consortium node to share
information on-chain.
(2) According to previous basic lease scenarios that we assumed, two default actions are
explored: (i) overdue rental payment; (ii) asset disposal against the LC. For the former
default action, we found that the larger proceeds gained resulting from reinvesting the
rental payment will cause the lessee to default, and at this time, the lessor will tend to
adopt BCT to mitigate the overdue-payment default risk. In addition, the residual value
of the leased asset has a positive impact on the exposure at default, and the lessee will be
more likely to default by not returning the leased asset to the lessor due to the temptation
of the high profit achieved from asset disposal at the end of the lease. Meanwhile, the
lessee’s default on asset disposals will result in the lessor being more inclined to adopt
BCT to ensure a timely claim of repossession of the leased asset.
(3) Although blockchain can guarantee data reliability (e.g., maintenance events) [76],
maintenance cost is not a determinant of the equilibrium state once the maintenance
service is outsourced. On the contrary, in-house maintenance provided by the lessor
may affect the two parties’ strategic decisions.
(4) When the lessee and lessor have incentives to participate in sharing or utilizing
more information on-chain, the lessee will eventually evolve to conform to the LC,
which will benefit the lessor and leasing industry. Setting up a changeable credit
associated with the lessee’s LC performance to compete for a block accounting right
via a consensus mechanism [77] is an effective way to incentivize the lessee to comply
with the LC, while this method does not work much to incentive the lessor to adopt
BCT. In addition, only when the default penalty on-chain exceeds a critical value can
it work to incentivize lessees to correctly fulfill their obligations in the LC [78], once
the penalty is lower than a critical value, which will in return increase the default risk.
The technology risks and relevant costs concerning CBLP deployment play a vital role
-----
_J. Theor. Appl. Electron. Commer. Res. 2023, 18_ 233
in encouraging the consortium to participate in information sharing on-chain, which
is consistent with what we expected in reality.
In summary, this study enables lessees and lessors to build a trustworthy cooperative
relationship on the consortium blockchain-based leasing platform, while also assisting
lessors or regulators in taking effective measures to incentivize lessees to comply with the
lease contract and share more information on-chain to enhance the management of default
risks in the leasing industry.
_8.2. Limitations and Future Directions_
Considering that the practical application of BCT in the leasing industry is rare, it
is difficult to obtain real data. Thus, this study focuses on mathematical modeling and
numerical simulation. The conclusions of this study can be further demonstrated and
enriched via empirical analysis of specific cases. Meanwhile, there are still some avenues
to be explored in the future. For example, it is meaningful to explore “tripartite–win”
strategies among lessees, lessors, and OEMs (or third-party MCs). Additionally, blockchain
smart contracts play an essential role in financing transactions [79]. Further research can
offer new insights into the cost reduction and value transfer of using smart contracts [80] to
motivate the related parties to share information.
**Author Contributions: Conceptualization, H.C.; methodology, H.C. and J.L. (Jing Lu); software, H.C.**
and J.L. (Jian Li); validation, H.C., J.L. (Jing Lu) and Z.X.; formal analysis, H.C. and J.L. (Jian Li);
data curation, H.C. and J.L. (Jian Li); writing-original draft, H.C.; writing-review and editing, S.-L.L.;
visualization, H.C. and Z.X; supervision, S.-L.L.; All authors have read and agreed to the published
version of the manuscript.
**Funding: This research was partly funded by the Department of Science and Technology of Guang-**
dong Province [Grant number 2020A0505090004] and Macau Science and Technology Development
Funds [Grant number 0061/2020/A2, 0158/2019/A3].
**Institutional Review Board Statement: Not applicable.**
**Informed Consent Statement: Not applicable.**
**Data Availability Statement: Not applicable.**
**Acknowledgments: The authors are grateful to the editors and the anonymous referees for their**
constructive and thorough comments, which helped to improve our paper.
**Conflicts of Interest: The authors declare no conflict of interest.**
**Abbreviations**
OEM Original Equipment Manufacturer
SMEs Small and Medium-Sized Enterprises
LFs Leasing Firms
MCs Maintenance Centers
LC Lease Contract
CPL Capital Lease
OPL Operating Lease
EGT Evolutionary Game Theory
ESS Evolutionary Stable Strategy
BCT Blockchain Technology
RDE Replication Dynamics Equation
CBLP Consortium Blockchain-Based Leasing Platform
HLF Hyperledger Fabric
BaaS Blockchain as a Service
CSP Cloud Storage Provider
IPFS InterPlanetary File System
-----
_J. Theor. Appl. Electron. Commer. Res. 2023, 18_ 234
**References**
1. Mol-Gómez-Vázquez, A.; Hernández-Cánovas, G.; Köeter-Kant, J. Economic and Institutional Determinants of Lease Financing
[for European SMEs: An Analysis across Developing and Developed Countries. J. Small Bus. Manag. 2020, 1–22. [CrossRef]](http://doi.org/10.1080/00472778.2020.1800352)
2. Li, J.; Wang, H.; Deng, Z.; Zhang, W.; Zhang, G. Leasing or Selling? The Channel Choice of Durable Goods Manufacturer
[Considering Consumers’ Capital Constraint. Flex. Serv. Manuf. J. 2022, 34, 317–350. [CrossRef]](http://doi.org/10.1007/s10696-021-09429-4)
3. [Eisfeldt, A.L.; Rampini, A.A. Leasing, Ability to Repossess, and Debt Capacity. Rev. Financ. Stud. 2009, 22, 1621–1657. [CrossRef]](http://doi.org/10.1093/rfs/hhn026)
4. Van Loon, P.; Delagarde, C.; Van Wassenhove, L.N.; Miheliˇc, A. Leasing or Buying White Goods: Comparing Manufacturer
[Profitability versus Cost to Consumer. Int. J. Prod. Res. 2020, 58, 1092–1106. [CrossRef]](http://doi.org/10.1080/00207543.2019.1612962)
5. [Smith, C.W.; Wakeman, L.M. Determinants of Corporate Leasing Policy. J. Financ. 1985, 40, 895–908. [CrossRef]](http://doi.org/10.1111/j.1540-6261.1985.tb05016.x)
6. Zheng, B.-K.; Zhu, L.-H.; Shen, M.; Gao, F.; Zhang, C.; Li, Y.-D.; Yang, J. Scalable and Privacy-Preserving Data Sharing Based on
[Blockchain. J. Comput. Sci. Technol. 2018, 33, 557–567. [CrossRef]](http://doi.org/10.1007/s11390-018-1840-5)
7. Dedeoglu, V.; Jurdak, R.; Dorri, A.; Lunardi, R.C.; Michelin, R.A.; Zorzo, A.F.; Kanhere, S.S. Blockchain Technologies for IoT; Springer:
Berlin/Heidelberg, Germany, 2020; pp. 55–89. ISBN 9789811387753.
8. Saberi, S.; Kouhizadeh, M.; Sarkis, J.; Shen, L. Blockchain Technology and Its Relationships to Sustainable Supply Chain
[Management. Int. J. Prod. Res. 2019, 57, 2117–2135. [CrossRef]](http://doi.org/10.1080/00207543.2018.1533261)
9. Wang, W.; Feng, L.; Li, Y.; Xu, F.; Deng, Q. Role of Financial Leasing in a Capital-Constrained Service Supply Chain. Transp. Res.
_[Part E Logist. Transp. Rev. 2020, 143, 102097. [CrossRef]](http://doi.org/10.1016/j.tre.2020.102097)_
10. Lease Accounting Framework and the Development of International Accounting Standards|SpringerLink. Available online:
[https://link.springer.com/chapter/10.1007/978-3-030-71633-2_2 (accessed on 29 November 2022).](https://link.springer.com/chapter/10.1007/978-3-030-71633-2_2)
11. Murthy, D.N.P.; Jack, N. Extended Warranties, Maintenance Service and Lease Contracts; Springer Series in Reliability Engineering;
Springer: London, UK, 2014; ISBN 978-1-4471-6439-5.
12. Kaposty, F.; Klein, P.; Löderbusch, M.; Pfingsten, A. Loss given default in SME leasing. Rev. Manag. Sci. 2022, 16, 1561–1597.
[[CrossRef]](http://doi.org/10.1007/s11846-021-00486-5)
13. Kuhle, P. Building A Blockchain-Based Decentralized Digital Asset Management System for Commercial Aircraft Leasing. Comput.
_[Ind. 2021, 21, 103393. [CrossRef]](http://doi.org/10.1016/j.compind.2020.103393)_
14. Altman, E.I.; Brady, B.; Resti, A.; Sironi, A. The Link between Default and Recovery Rates: Theory, Empirical Evidence, and
[Implications. J. Bus. 2005, 78, 2203–2228. [CrossRef]](http://doi.org/10.1086/497044)
15. Kysucky, V.; Norden, L. The Benefits of Relationship Lending in a Cross-Country Context: A Meta-Analysis. Manag. Sci. 2016, 62,
[90–110. [CrossRef]](http://doi.org/10.1287/mnsc.2014.2088)
16. [Nakamoto, S. Bitcoin: A Peer-to-Peer Electronic Cash System. Decentralized Bus. Rev. 2008, 21260. Available online: https:](https://bitcoin.org/bitcoin.pdf)
[//bitcoin.org/bitcoin.pdf (accessed on 27 January 2023).](https://bitcoin.org/bitcoin.pdf)
17. Politou, E.; Casino, F.; Alepis, E.; Patsakis, C. Blockchain Mutability: Challenges and Proposed Solutions. IEEE Trans. Emerg. Top.
_[Comput. 2021, 9, 1972–1986. [CrossRef]](http://doi.org/10.1109/TETC.2019.2949510)_
18. Viriyasitavat, W.; Hoonsopon, D. Blockchain Characteristics and Consensus in Modern Business Processes. J. Ind. Inf. Integr. 2019,
_[13, 32–39. [CrossRef]](http://doi.org/10.1016/j.jii.2018.07.004)_
19. Yu, Y.; Huang, G.; Guo, X. Financing Strategy Analysis for a Multi-Sided Platform with Blockchain Technology. Int. J. Prod. Res.
**[2021, 59, 4513–4532. [CrossRef]](http://doi.org/10.1080/00207543.2020.1766718)**
20. Belotti, M.; Boži´c, N.; Pujolle, G.; Secci, S. A Vademecum on Blockchain Technologies: When, Which, and How. IEEE Commun.
_[Surv. Tutor. 2019, 21, 3796–3838. [CrossRef]](http://doi.org/10.1109/COMST.2019.2928178)_
21. Singh, S.K.; Jenamani, M.; Dasgupta, D.; Das, S. A Conceptual Model for Indian Public Distribution System Using Consortium
[Blockchain with On-Chain and off-Chain Trusted Data. Inf. Technol. Dev. 2021, 27, 499–523. [CrossRef]](http://doi.org/10.1080/02681102.2020.1847024)
22. Dutta, P.; Choi, T.-M.; Somani, S.; Butala, R. Blockchain Technology in Supply Chain Operations: Applications, Challenges and
[Research Opportunities. Transp. Res. Part E Logist. Transp. Rev. 2020, 142, 102067. [CrossRef]](http://doi.org/10.1016/j.tre.2020.102067)
23. Andoni, M.; Robu, V.; Flynn, D.; Abram, S.; Geach, D.; Jenkins, D.; McCallum, P.; Peacock, A. Blockchain Technology in the
[Energy Sector: A Systematic Review of Challenges and Opportunities. Renew. Sustain. Energy Rev. 2019, 100, 143–174. [CrossRef]](http://doi.org/10.1016/j.rser.2018.10.014)
24. Tandon, A.; Dhir, A.; Islam, A.K.M.N.; Mäntymäki, M. Blockchain in Healthcare: A Systematic Literature Review, Synthesizing
[Framework and Future Research Agenda. Comput. Ind. 2020, 122, 103290. [CrossRef]](http://doi.org/10.1016/j.compind.2020.103290)
25. Leng, J.; Ruan, G.; Jiang, P.; Xu, K.; Liu, Q.; Zhou, X.; Liu, C. Blockchain-Empowered Sustainable Manufacturing and Product
[Lifecycle Management in Industry 4.0: A Survey. Renew. Sustain. Energy Rev. 2020, 132, 110112. [CrossRef]](http://doi.org/10.1016/j.rser.2020.110112)
26. Xie, J.; Tang, H.; Huang, T.; Yu, F.R.; Xie, R.; Liu, J.; Liu, Y. A Survey of Blockchain Technology Applied to Smart Cities: Research
[Issues and Challenges. IEEE Commun. Surv. Tutor. 2019, 21, 2794–2830. [CrossRef]](http://doi.org/10.1109/COMST.2019.2899617)
27. Loukil, F.; Abed, M.; Boukadi, K. Blockchain Adoption in Education: A Systematic Literature Review. Educ. Inf. Technol. 2021, 26,
[5779–5797. [CrossRef]](http://doi.org/10.1007/s10639-021-10481-8)
28. Wang, J.; Wu, P.; Wang, X.; Shou, W. The Outlook of Blockchain Technology for Construction Engineering Management. Front.
_[Eng. Manag. 2017, 67–75. [CrossRef]](http://doi.org/10.15302/J-FEM-2017006)_
29. Auer, S.; Nagler, S.; Mazumdar, S.; Mukkamala, R.R. Towards Blockchain-IoT Based Shared Mobility: Car-Sharing and Leasing as
[a Case Study. J. Netw. Comput. Appl. 2022, 200, 103316. [CrossRef]](http://doi.org/10.1016/j.jnca.2021.103316)
-----
_J. Theor. Appl. Electron. Commer. Res. 2023, 18_ 235
30. Faber, B.; Michelet, G.C.; Weidmann, N.; Mukkamala, R.R.; Vatrapu, R. BPDIMS: A Blockchain-Based Personal Data and Identity
Management System. In Proceedings of the 52nd Hawaii International Conference on System Sciences, Maui, HI, USA, 8–11
January 2019; pp. 6855–6864.
31. Obour Agyekum, K.O.-B.; Xia, Q.; Boateng Sifah, E.; Amofa, S.; Nketia Acheampong, K.; Gao, J.; Chen, R.; Xia, H.; Gee, J.C.;
Du, X.; et al. V-Chain: A Blockchain-Based Car Lease Platform. In Proceedings of the 2018 IEEE International Conference on
Internet of Things (iThings) and IEEE Green Computing and Communications (GreenCom) and IEEE Cyber, Physical and Social
Computing (CPSCom) and IEEE Smart Data (SmartData), HICSS, Halifax, NS, Canada, 30 July–3 August 2018; pp. 1317–1325.
32. Zheng, K.; Zheng, L.J.; Gauthier, J.; Zhou, L.; Xu, Y.; Behl, A.; Zhang, J.Z. Blockchain Technology for Enterprise Credit Information
[Sharing in Supply Chain Finance. J. Innov. Knowl. 2022, 7, 100256. [CrossRef]](http://doi.org/10.1016/j.jik.2022.100256)
33. Fiorentino, S.; Bartolucci, S. Blockchain-Based Smart Contracts as New Governance Tools for the Sharing Economy. Cities 2021,
_[117, 103325. [CrossRef]](http://doi.org/10.1016/j.cities.2021.103325)_
34. Delle Foglie, A.; Panetta, I.C.; Boukrami, E.; Vento, G. The Impact of the Blockchain Technology on the Global Sukuk Industry:
[Smart Contracts and Asset Tokenisation. Technol. Anal. Strateg. Manag. 2021, 1–15. [CrossRef]](http://doi.org/10.1080/09537325.2021.1939000)
35. Wang, J.; Zhou, Z.; Botterud, A. An Evolutionary Game Approach to Analyzing Bidding Strategies in Electricity Markets with
[Elastic Demand. Energy 2011, 36, 3459–3467. [CrossRef]](http://doi.org/10.1016/j.energy.2011.03.050)
36. Wang, J.; Peng, X.; Du, Y.; Wang, F. A Tripartite Evolutionary Game Research on Information Sharing of the Subjects of Agricultural
[Product Supply Chain with a Farmer Cooperative as the Core Enterprise. Manag. Decis. Econ. 2022, 43, 159–177. [CrossRef]](http://doi.org/10.1002/mde.3365)
37. Apaloo, J.; Brown, J.S.; Vincent, T.L. Evolutionary Game Theory: ESS, Convergence Stability, and NIS. Evol. Ecol. Res. 2009, 11,
489–515.
38. Chen, Y.; Zeng, Q.; Zheng, X.; Shao, B.; Jin, L. Safety Supervision of Tower Crane Operation on Construction Sites: An Evolutionary
[Game Analysis. Saf. Sci. 2022, 152, 105578. [CrossRef]](http://doi.org/10.1016/j.ssci.2021.105578)
39. Su, L.; Cao, Y.; Li, H.; Tan, J. Blockchain-Driven Optimal Strategies for Supply Chain Finance Based on a Tripartite Game Model.
_[J. Theor. Appl. Electron. Commer. Res. 2022, 17, 67. [CrossRef]](http://doi.org/10.3390/jtaer17040067)_
40. Tang, Q.; Zhang, Z.; Yuan, Z.; Li, Z. The Game Analysis of Information Sharing for Supply Chain Enterprises in the Blockchain.
_[Supply Chain 2022, 2, 13. [CrossRef]](http://doi.org/10.3389/fmtec.2022.880332)_
41. Sun, R.; He, D.; Su, H. Evolutionary Game Analysis of Blockchain Technology Preventing Supply Chain Financial Risks. J. Theor.
_[Appl. Electron. Commer. Res. 2021, 16, 155. [CrossRef]](http://doi.org/10.3390/jtaer16070155)_
42. Song, L.; Luo, Y.; Chang, Z.; Jin, C.; Nicolas, M. Blockchain Adoption in Agricultural Supply Chain for Better Sustainability: A
[Game Theory Perspective. Sustainability 2022, 14, 1470. [CrossRef]](http://doi.org/10.3390/su14031470)
43. Figueroa-Lorenzo, S.; Añorga, J.; Arrizabalaga, S. Methodological Performance Analysis Applied to a Novel IIoT Access Control
[System Based on Permissioned Blockchain. Inf. Process. Manag. 2021, 58, 102558. [CrossRef]](http://doi.org/10.1016/j.ipm.2021.102558)
44. Xu, X.; Sun, G.; Luo, L.; Cao, H.; Yu, H.; Vasilakos, A.V. Latency Performance Modeling and Analysis for Hyperledger Fabric
[Blockchain Network. Inf. Process. Manag. 2021, 58, 102436. [CrossRef]](http://doi.org/10.1016/j.ipm.2020.102436)
45. Androulaki, E.; Barger, A.; Bortnikov, V.; Cachin, C.; Christidis, K.; De Caro, A.; Enyeart, D.; Ferris, C.; Laventman, G.; Manevich,
Y. Hyperledger Fabric: A Distributed Operating System for Permissioned Blockchains. arXiv 2018, arXiv:1801.10228.
46. Rizzardi, A.; Sicari, S.; Miorandi, D.; Coen-Porisini, A. Securing the Access Control Policies to the Internet of Things Resources
[through Permissioned Blockchain. Concurr. Comput. Pract. Exp. 2022, 34, e6934. [CrossRef]](http://doi.org/10.1002/cpe.6934)
47. Steinhoff, S.; Stathakopoulou, C.; Pavlovic, M.; Vukoli´c, M. BMS: Secure Decentralized Reconfiguration for Blockchain and BFT
Systems. arXiv 2021, arXiv:210903913.
48. Liu, Y.; Lu, Q.; Paik, H.-Y.; Xu, X.; Chen, S.; Zhu, L. Design Pattern as a Service for Blockchain-Based Self-Sovereign Identity. IEEE
_[Softw. 2020, 37, 30–36. [CrossRef]](http://doi.org/10.1109/MS.2020.2992783)_
49. Viriyasitavat, W.; Xu, L.D.; Sapsomboon, A.; Dhiman, G.; Hoonsopon, D. Building Trust of Blockchain-Based Internet-of-Thing
[Services Using Public Key Infrastructure. Enterp. Inf. Syst. 2022, 16, 2037162. [CrossRef]](http://doi.org/10.1080/17517575.2022.2037162)
50. Kuhn, M.; Funk, F.; Zhang, G.; Franke, J. Blockchain-Based Application for the Traceability of Complex Assembly Structures.
_[J. Manuf. Syst. 2021, 59, 617–630. [CrossRef]](http://doi.org/10.1016/j.jmsy.2021.04.013)_
51. Herrera-Joancomartí, J.; Pérez-Solà, C. Privacy in Bitcoin Transactions: New Challenges from Blockchain Scalability Solutions. In
Proceedings of the Modeling Decisions for Artificial Intelligence; Springer International Publishing: Cham, Switzerland, 2016; pp.
26–44.
52. Miyachi, K.; Mackey, T.K. HOCBS: A Privacy-Preserving Blockchain Framework for Healthcare Data Leveraging an on-Chain
[and off-Chain System Design. Inf. Process. Manag. 2021, 58, 102535. [CrossRef]](http://doi.org/10.1016/j.ipm.2021.102535)
53. Solaiman, E.; Wike, T.; Sfyrakis, I. Implementation and Evaluation of Smart Contracts Using a Hybrid On-and Off-blockchain
[Architecture. Concurr. Comput. Pract. Exp. 2021, 33, e5811. [CrossRef]](http://doi.org/10.1002/cpe.5811)
54. Wu, H.; Peng, Z.; Guo, S.; Yang, Y.; Xiao, B. VQL: Efficient and Verifiable Cloud Query Services for Blockchain Systems. IEEE
_[Trans. Parallel Distrib. Syst. 2021, 33, 1393–1406. [CrossRef]](http://doi.org/10.1109/TPDS.2021.3113873)_
55. Udokwu, C.; Norta, A. Deriving and Formalizing Requirements of Decentralized Applications for Inter-Organizational Collabo[rations on Blockchain. Arab. J. Sci. Eng. 2021, 46, 8397–8414. [CrossRef]](http://doi.org/10.1007/s13369-020-05245-4)
56. Surjandari, I.; Yusuf, H.; Laoh, E.; Maulida, R. Designing a Permissioned Blockchain Network for the Halal Industry Using
[Hyperledger Fabric with Multiple Channels and the Raft Consensus Mechanism. J. Big Data. 2021, 8, 10. [CrossRef]](http://doi.org/10.1186/s40537-020-00405-7)
-----
_J. Theor. Appl. Electron. Commer. Res. 2023, 18_ 236
57. Fu, W.; Wei, X.; Tong, S. An Improved Blockchain Consensus Algorithm Based on Raft. Arab. J. Sci. Eng. 2021, 46, 8137–8149.
[[CrossRef]](http://doi.org/10.1007/s13369-021-05427-8)
58. Zhang, Y.; Zhang, L.; Liu, Y.; Luo, X. Proof of Service Power: A Blockchain Consensus for Cloud Manufacturing. J. Manuf. Syst.
**[2021, 59, 1–11. [CrossRef]](http://doi.org/10.1016/j.jmsy.2021.01.006)**
59. Nechaev, A.S.; Zakharov, S.V.; Barykina, Y.N.; Vel’m, M.V.; Kuznetsova, O.N. Forming Methodologies to Improving the Efficiency
[of Innovative Companies Based on Leasing Tools. J. Sustain. Financ. Invest. 2022, 12, 536–553. [CrossRef]](http://doi.org/10.1080/20430795.2020.1784681)
60. Li, Z.; Zhong, R.Y.; Tian, Z.-G.; Dai, H.-N.; Barenji, A.V.; Huang, G.Q. Industrial Blockchain: A State-of-the-Art Survey. Robot.
_[Comput.-Integr. Manuf. 2021, 70, 102124. [CrossRef]](http://doi.org/10.1016/j.rcim.2021.102124)_
61. Li, B.; Li, H.; Sun, Q.; Chen, X. Evolutionary Game Analysis between Businesses and Consumers under the Background of
[Internet Rumors. Concurr. Comput. Pract. Exp. 2022, 34, e5897. [CrossRef]](http://doi.org/10.1002/cpe.5897)
62. Weibull, J.W. Evolutionary Game Theory; MIT Press: Cambridge, MA, USA, 1997; ISBN 0-262-73121-5.
63. Xu, B.; Li, Q. Research on Supervision Mechanism of Big Data Discriminatory Pricing on the Asymmetric Service Platform—Based
[on SD Evolutionary Game Model. J. Theor. Appl. Electron. Commer. Res. 2022, 17, 63. [CrossRef]](http://doi.org/10.3390/jtaer17040063)
64. Kang, H.; Dai, T.; Jean-Louis, N.; Tao, S.; Gu, X. Fabzk: Supporting Privacy-Preserving, Auditable Smart Contracts in Hyperledger Fabric;
IEEE: Piscataway, NJ, USA, 2019; pp. 543–555.
65. Bhushan, B.; Sinha, P.; Sagayam, K.M.; Andrew, J. Untangling Blockchain Technology: A Survey on State of the Art, Security
[Threats, Privacy Services, Applications and Future Research Directions. Comput. Electr. Eng. 2021, 90, 106897. [CrossRef]](http://doi.org/10.1016/j.compeleceng.2020.106897)
66. Liu, J.; Zhang, H.; Zhen, L. Blockchain Technology in Maritime Supply Chains: Applications, Architecture and Challenges. Int. J.
_[Prod. Res. 2021, 1–17. [CrossRef]](http://doi.org/10.1080/00207543.2021.1930239)_
67. Chang, V.; Baudier, P.; Zhang, H.; Xu, Q.; Zhang, J.; Arami, M. How Blockchain Can Impact Financial Services—The Overview,
[Challenges and Recommendations from Expert Interviewees. Technol. Forecast. Soc. Chang. 2020, 158, 120166. [CrossRef]](http://doi.org/10.1016/j.techfore.2020.120166)
68. Li, Q.; Wang, Y.; Li, K.; Chen, L.; Wei, Z. Evolutionary Dynamics of the Last Mile Travel Choice. Phys. Stat. Mech. Appl. 2019, 536,
[122555. [CrossRef]](http://doi.org/10.1016/j.physa.2019.122555)
69. Qiao, Y.; Lan, Q.; Zhou, Z.; Ma, C. Privacy-Preserving Credit Evaluation System Based on Blockchain. Expert Syst. Appl. 2022, 188,
[115989. [CrossRef]](http://doi.org/10.1016/j.eswa.2021.115989)
70. [Friedman, D. On Economic Applications of Evolutionary Game Theory. J. Evol. Econ. 1998, 8, 15–43. [CrossRef]](http://doi.org/10.1007/s001910050054)
71. [Nowak, M.A.; Sigmund, K. Evolutionary Dynamics of Biological Games. Science 2004, 303, 793–799. [CrossRef] [PubMed]](http://doi.org/10.1126/science.1093411)
72. Ritzberger, K.; Weibull, J.W. Evolutionary Selection in Normal-Form Games. Econom. J. Econom. Soc. 1995, 63, 1371–1399.
[[CrossRef]](http://doi.org/10.2307/2171774)
73. Kang, K.; Zhao, Y.; Zhang, J.; Qiang, C. Evolutionary Game Theoretic Analysis on Low-Carbon Strategy for Supply Chain
[Enterprises. J. Clean. Prod. 2019, 230, 981–994. [CrossRef]](http://doi.org/10.1016/j.jclepro.2019.05.118)
74. Yadav, S.; Singh, S.P. Blockchain Critical Success Factors for Sustainable Supply Chain. Resour. Conserv. Recycl. 2020, 152, 104505.
[[CrossRef]](http://doi.org/10.1016/j.resconrec.2019.104505)
75. Jie, S.; Zhang, P.; Alkubati, M.; Yubin, B.; Ge, Y. Research Advances on Blockchain-as-a-Service: Architectures, Applications and
Challenges. Digit. Commun. Netw. 2021, 8, 466–475.
76. Mohril, R.S.; Solanki, B.S.; Lad, B.K.; Kulkarni, M.S. Blockchain Enabled Maintenance Management Framework for Military
[Equipment. IEEE Trans. Eng. Manag. 2021, 69, 3938–3951. [CrossRef]](http://doi.org/10.1109/TEM.2021.3099437)
77. Zhu, S.; Cai, Z.; Hu, H.; Li, Y.; Li, W. ZkCrowd: A Hybrid Blockchain-Based Crowdsourcing Platform. IEEE Trans. Ind. Inform.
**[2019, 16, 4196–4205. [CrossRef]](http://doi.org/10.1109/TII.2019.2941735)**
78. Holden, R.; Malani, A. Can Blockchain Solve the Hold-up Problem in Contracts? Cambridge University Press: Cambridge, UK, 2021;
ISBN 1-00-900479-4.
79. Natanelov, V.; Cao, S.; Foth, M.; Dulleck, U. Blockchain Smart Contracts for Supply Chain Finance: Mapping the Innovation
[Potential in Australia-China Beef Supply Chains. J. Ind. Inf. Integr. 2022, 30, 100389. [CrossRef]](http://doi.org/10.1016/j.jii.2022.100389)
80. Oh, J.; Choi, Y.; In, J. A Conceptual Framework for Designing Blockchain Technology Enabled Supply Chains. Int. J. Logist. Res.
_[Appl. 2022, 1–19. [CrossRef]](http://doi.org/10.1080/13675567.2022.2052824)_
**Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual**
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3390/jtaer18010012?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/jtaer18010012, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "GOLD",
"url": "https://www.mdpi.com/0718-1876/18/1/12/pdf?version=1674979153"
}
| 2,023
|
[
"JournalArticle"
] | true
| 2023-01-29T00:00:00
|
[
{
"paperId": "2a5e3cd25d4a5d10c75537a9b69cb01e3912547a",
"title": "Blockchain-Driven Optimal Strategies for Supply Chain Finance Based on a Tripartite Game Model"
},
{
"paperId": "fc23a434a7472082ef4920938b6d0c1262f4fdc6",
"title": "Blockchain technology for enterprise credit information sharing in supply chain finance"
},
{
"paperId": "b445bced95aab4b73afdfe0243c4fbafadec7e5c",
"title": "Research on Supervision Mechanism of Big Data Discriminatory Pricing on the Asymmetric Service Platform - Based on SD Evolutionary Game Model"
},
{
"paperId": "3d34e0af4358de6a70f96b0b6133135364661942",
"title": "Blockchain smart contracts for supply chain finance: Mapping the innovation potential in Australia-China beef supply chains"
},
{
"paperId": "be8836dab5f92d5a63bc7ad72ddc91d476fa94b9",
"title": "The Game Analysis of Information Sharing for Supply Chain Enterprises in the Blockchain"
},
{
"paperId": "851ab39c0b1bc14bc905fb6f48d027955e3a5183",
"title": "VQL: Efficient and Verifiable Cloud Query Services for Blockchain Systems"
},
{
"paperId": "74067550d9ae597bab0046f098a491db1425f3ec",
"title": "A conceptual framework for designing blockchain technology enabled supply chains"
},
{
"paperId": "e24e9d576448ba97fcb752dda8f09b5c1d2d6ac7",
"title": "Securing the access control policies to the Internet of Things resources through permissioned blockchain"
},
{
"paperId": "a3846bafe94156d63f5ee9f7ee492d08c674a38f",
"title": "Building trust of Blockchain-based Internet-of-Thing services using public key infrastructure"
},
{
"paperId": "c8bfd7a7c7500addb70cf9c3661569e8f2690dc0",
"title": "Privacy-preserving credit evaluation system based on blockchain"
},
{
"paperId": "666d5c4bafc6b17288146277bccb85206f57596f",
"title": "Blockchain Adoption in Agricultural Supply Chain for Better Sustainability: A Game Theory Perspective"
},
{
"paperId": "ff83766d73b62f34b96c2caf243312ac518e57bd",
"title": "Towards blockchain-IoT based shared mobility: Car-sharing and leasing as a case study"
},
{
"paperId": "1c1486ac0f89baea07c66d7fa58223fb02553e84",
"title": "Safety supervision of tower crane operation on construction sites: An evolutionary game analysis"
},
{
"paperId": "4c918e373c8fe9f2a293d7d0de3488caf427acfe",
"title": "Evolutionary Game Analysis of Blockchain Technology Preventing Supply Chain Financial Risks"
},
{
"paperId": "856aad8a001944deb982be3b9d1e250d763c79e5",
"title": "Leasing or selling? The channel choice of durable goods manufacturer considering consumers’ capital constraint"
},
{
"paperId": "7edae9e82a5d05d4d9d7a0b43027a415910459bd",
"title": "Blockchain-based smart contracts as new governance tools for the sharing economy"
},
{
"paperId": "72c849e9b3cdf44bfd3ebb6a38d7f029b3478b6a",
"title": "Loss given default in SME leasing"
},
{
"paperId": "807735c594c6f5140b35319da23528ddb3abf77f",
"title": "Blockchain Enabled Maintenance Management Framework for Military Equipment"
},
{
"paperId": "845596c2ca836ab7c0589de32d55f12884bd611b",
"title": "Industrial Blockchain: A state-of-the-art Survey"
},
{
"paperId": "c29b43e0e3cd15e6c6131b864986b9bfcfc07d22",
"title": "Methodological performance analysis applied to a novel IIoT access control system based on permissioned blockchain"
},
{
"paperId": "50836ce628def307a7f3dc0e76571d6e14778611",
"title": "The impact of the Blockchain technology on the global Sukuk industry: smart contracts and asset tokenisation"
},
{
"paperId": "1a893681c3ae37cf4f77c125679d73f10c8e57aa",
"title": "A tripartite evolutionary game research on information sharing of the subjects of agricultural product supply chain with a farmer cooperative as the core enterprise"
},
{
"paperId": "134bafb2f0e5e0f051f1e242fa60c07f7f57a67d",
"title": "Blockchain technology in maritime supply chains: applications, architecture and challenges"
},
{
"paperId": "c6475211b6b36bfd41cff1a31f4ae7ab99d014d5",
"title": "Blockchain adoption in education: a systematic literature review"
},
{
"paperId": "40a1b5f940fbfc58b4adecec08977850729141ac",
"title": "Blockchain-based application for the traceability of complex assembly structures"
},
{
"paperId": "4c4b2fe43c10e129a021a02712e3eecb100c500f",
"title": "Proof of service power: A blockchain consensus for cloud manufacturing"
},
{
"paperId": "0d9e0cbef607991de72dd1d8c567d242a8bf4a86",
"title": "Building A blockchain-based decentralized digital asset management system for commercial aircraft leasing"
},
{
"paperId": "4b014c993c0ba68c5f3c9000c514c4baa043a8d0",
"title": "Deriving and Formalizing Requirements of Decentralized Applications for Inter-Organizational Collaborations on Blockchain"
},
{
"paperId": "528b5a671ddbb2a17f652be43eefd2a83a255c46",
"title": "An Improved Blockchain Consensus Algorithm Based on Raft"
},
{
"paperId": "f956e4ed93e5b3eebed6568885deee873f684fa9",
"title": "Research advances on blockchain-as-a-service: architectures, applications and challenges"
},
{
"paperId": "2e5af24126293387fdf9e9ae96bc9aab3871908b",
"title": "A conceptual model for Indian public distribution system using consortium blockchain with on-chain and off-chain trusted data"
},
{
"paperId": "5ef707d794c63527137e177e1800e27fac61f6bf",
"title": "Untangling blockchain technology: A survey on state of the art, security threats, privacy services, applications and future research directions"
},
{
"paperId": "bcd8b89a9e7a1fe4ab6ea1486f2aae5ccebbd071",
"title": "Role of financial leasing in a capital-constrained service supply chain"
},
{
"paperId": "5e04786ee99ba2dd4bcdbdbc18f0189ccac0b065",
"title": "Blockchain-empowered sustainable manufacturing and product lifecycle management in industry 4.0: A survey"
},
{
"paperId": "a9e212ef2e0abaca7878194ab831e23d5b5296df",
"title": "Blockchain technology in supply chain operations: Applications, challenges and research opportunities"
},
{
"paperId": "36097e3078e33c51c30f06df508b026a3fcb2b65",
"title": "Designing a Permissioned Blockchain Network for the Halal Industry using Hyperledger Fabric with multiple channels and the raft consensus mechanism"
},
{
"paperId": "5baec0e8f6de64ddb33db85fcf4ea38ac10a8527",
"title": "Economic and institutional determinants of lease financing for European SMEs: An analysis across developing and developed countries"
},
{
"paperId": "64aa9e9bbd1a37e67fc5dc1d0d3ac796bcc854be",
"title": "Evolutionary game analysis between businesses and consumers under the background of Internet rumors"
},
{
"paperId": "b4ad7ce1b6a0fefcc2eced1ff0e4943caa613f95",
"title": "Forming methodologies to improving the efficiency of innovative companies based on leasing tools"
},
{
"paperId": "44ff32b8bbfa0b4bdda40162afcbe12af6e5b1b5",
"title": "zkCrowd: A Hybrid Blockchain-Based Crowdsourcing Platform"
},
{
"paperId": "1d7a9791c06a97631b04518266e0eeddc59469f9",
"title": "Financing strategy analysis for a multi-sided platform with blockchain technology"
},
{
"paperId": "441dab9d65a7b1c4658dff33c094330a733210d6",
"title": "Implementation and evaluation of smart contracts using a hybrid on‐ and off‐blockchain architecture"
},
{
"paperId": "a7a0e0e85625485218ff382ebcbf76ca20bd9960",
"title": "Design Pattern as a Service for Blockchain-Based Self-Sovereign Identity"
},
{
"paperId": "69d3bca2a2ec91801c6542cc63476e0660151b36",
"title": "Evolutionary dynamics of the last mile travel choice"
},
{
"paperId": "84f228f6ab0b5710ed88deb3af2e1986a32af59f",
"title": "Evolutionary game theoretic analysis on low-carbon strategy for supply chain enterprises"
},
{
"paperId": "c4ed360660a44a7623154fc18266ce013a5f0c44",
"title": "Blockchain Mutability: Challenges and Proposed Solutions"
},
{
"paperId": "ea8a68d7f356c9e0a6f74706e429e8b41250b784",
"title": "A Vademecum on Blockchain Technologies: When, Which, and How"
},
{
"paperId": "2aa44a41ae725f07952f240d5cb6eefe6daff385",
"title": "Leasing or buying white goods: comparing manufacturer profitability versus cost to consumer"
},
{
"paperId": "cf58f0c53a2bd2c0a1ee7bec77920c3f80036ac2",
"title": "Blockchain characteristics and consensus in modern business processes"
},
{
"paperId": "7117fdbfc3f3a4561ddc7843a54eec31b7373f18",
"title": "A Survey of Blockchain Technology Applied to Smart Cities: Research Issues and Challenges"
},
{
"paperId": "60be2610dba19761d6458bbac27527b744b0109e",
"title": "Blockchain technology in the energy sector: A systematic review of challenges and opportunities"
},
{
"paperId": "2e82b8539af92b4af1f5c1c59dcce9d31dcefccc",
"title": "Blockchain technology and its relationships to sustainable supply chain management"
},
{
"paperId": "ec6cbabde4283b20bf65fac9b1b04bfda468566b",
"title": "Scalable and Privacy-Preserving Data Sharing Based on Blockchain"
},
{
"paperId": "1b4c39ba447e8098291e47fd5da32ca650f41181",
"title": "Hyperledger fabric: a distributed operating system for permissioned blockchains"
},
{
"paperId": "64d67a34a2fb537f56bb8ea43998553f3de92701",
"title": "The Benefits of Relationship Lending in a Cross-Country Context: A Meta-Analysis"
},
{
"paperId": "dfa03b4ffe5dfb9cb0b419c12dbd9faa609ff739",
"title": "Extended Warranties, Maintenance Service and Lease Contracts: Modeling and Analysis for Decision-Making"
},
{
"paperId": "cf766eb8fbe8bd9d5f9716593f8ce3eb6b4e805c",
"title": "An evolutionary game approach to analyzing bidding strategies in electricity markets with elastic de"
},
{
"paperId": "f9c1aaab91aaf6bd27649539417cc6b8fddc6233",
"title": "Leasing, Ability to Repossess, and Debt Capacity"
},
{
"paperId": "2feedd23395f882e40dc46f864f2a536c0fe0234",
"title": "Evolutionary Dynamics of Biological Games"
},
{
"paperId": "5a435a7774f8d73836dd86546808abcc876ba3b6",
"title": "The Link between Default and Recovery Rates: Theory, Empirical Evidence and Implications"
},
{
"paperId": "42ef29fa9b184295459b51b5854e56ed27df3496",
"title": "On economic applications of evolutionary game theory"
},
{
"paperId": "fc11bbcaea9eaaf4b5a00e3c7518867f832ed849",
"title": "Evolutionary Selection in Normal-Form Games"
},
{
"paperId": "403d78d0193442d6c5f56d3f29521ab583a862b7",
"title": "Determinants of Corporate Leasing Policy"
},
{
"paperId": "28ff211bf98028a4acf3d1495c11efc7a7c50801",
"title": "hOCBS: A privacy-preserving blockchain framework for healthcare data leveraging an on-chain and off-chain system design"
},
{
"paperId": "961a69342dbac1523d0eb09557d1a00de79d6c57",
"title": "Latency performance modeling and analysis for hyperledger fabric blockchain network"
},
{
"paperId": null,
"title": "Can Blockchain Solve the Hold-up Problem in Contracts?"
},
{
"paperId": "bce3e557e24247f99b42370c5c718b44a3107307",
"title": "Blockchain critical success factors for sustainable supply chain"
},
{
"paperId": "466c6f989490a3819075908336b27ca722f6337f",
"title": "Technological Forecasting & Social Change"
},
{
"paperId": "cdb46e45d8b96cef76ed1c081275165e2f8cc726",
"title": "Evolutionary game theory: ESS, convergence stability, and NIS"
},
{
"paperId": "9e84bd933c267b09422d9f25d53fed92846e18f2",
"title": "Computers in Industry"
}
] | 36,384
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Mathematics",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/024384b677b58c1f48697245b748ec291812b563
|
[
"Computer Science"
] | 0.880916
|
Universal Re-encryption for Mixnets
|
024384b677b58c1f48697245b748ec291812b563
|
The Cryptographer's Track at RSA Conference
|
[
{
"authorId": "2779068",
"name": "P. Golle"
},
{
"authorId": "2836467",
"name": "M. Jakobsson"
},
{
"authorId": "1687161",
"name": "A. Juels"
},
{
"authorId": "3213341",
"name": "P. Syverson"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"Cryptogr Track RSA Conf",
"The Cryptographers’ Track at the RSA Conference",
"CT-RSA"
],
"alternate_urls": null,
"id": "7d878997-4b28-4d42-97e4-1146b7c090bc",
"issn": null,
"name": "The Cryptographer's Track at RSA Conference",
"type": "conference",
"url": "http://www.wikicfp.com/cfp/program?id=620"
}
| null |
# Universal Re-encryption for Mixnets
**Abstract. We introduce a new cryptographic technique that we call universal re-encryption. A**
conventional cryptosystem that permits re-encryption, such as ElGamal, does so only for a player
with knowledge of the public key corresponding to a given ciphertext. In contrast, universal reencryption may be performed without knowledge of public keys. We demonstrate an asymmetric
cryptosystem with universal re-encryption that is half as efficient as standard ElGamal in terms
of both computation and storage.
While technically and conceptually simple, universal re-encryption leads to new types of functionality in mixnet architectures. Conventional mixnets are often called upon to enable players to
communicate with one another through channels that are externally anonymous, i.e., that hide
information permitting traffic-analysis. Universal re-encryption permits a mixnet of this kind to
be constructed in which servers hold no public or private keying material, and may therefore
dispense with the cumbersome requirements of key generation, key distribution, and private-key
management. We describe two practical mixnet constructions, one involving asymmetric input
ciphertexts, and another with hybrid-ciphertext inputs.
**Key words: anonymity, mix networks, private channels, public-key cryptography, universal**
re-encryption
## 1 Introduction
A mix network or mixnet is a cryptographic construction that invokes a set of servers to
establish private communication channels [5]. One type of mix network accepts as input a
collection of ciphertexts, and outputs the corresponding plaintexts in a randomly permuted
order. The main privacy property desired of such a mixnet is that the permutation matching
inputs to outputs should be known only to the mixnet, and no one else. In particular, an
adversary should be unable to guess which input ciphertext corresponds to an output plaintext
any more effectively than by guessing at random.
One common variety of mixnet known as a re-encryption mixnet relies on a public-key
encryption scheme, such as ElGamal [11], that allows for re-encryption of ciphertexts. For a
given public key, a ciphertext C[′] is said to represent a re-encryption of C if both ciphertexts
decrypt to the same plaintext. In a re-encryption mixnet, the inputs are submitted encrypted
under the public-key of the mixnet. (The corresponding private key is held in distributed
form among the servers.) The batch of input ciphertexts is processed sequentially by each mix
server. The first server takes the set of input ciphertexts, re-encrypts them, and outputs the
re-encrypted ciphertexts in a random order. Each server in turn takes the set of ciphertexts
output by the previous server, and re-encrypts and mixes them. The set of ciphertexts produced
by the last server may be decrypted by a quorum of mix servers to yield plaintext outputs.
Privacy in this mixnet construction derives from the fact that the ciphertext pair (C, C[′])
is indistinguishable from a pair (C, R) for a random ciphertext R to any adversary without
knowledge of the private key.
In this paper, we propose a new type of public-key cryptosystem that permits universal re_encryption of ciphertexts. We introduce the term universal encryption to mean re-encryption_
without knowledge of the public key under which a ciphertext was computed. Like standard
re-encryption, universal re-encryption transforms a ciphertext C into a new ciphertext C[′] with
-----
same corresponding plaintext. The novelty in our proposal is that re-encryption neither requires
nor yields knowledge of the public key under which a ciphertext was computed[1].
When applied to mix networks, our universal re-encryption technique offers new and interesting functionality. Most importantly, mix networks based on universal re-encryption dispense
with the cumbersome protocols that traditional mixnets require in order to establish and maintain a shared private key. We discuss more benefits and applications of universal mixnets in
the next section. It is possible to construct a universal mixnet based on universal re-encryption
roughly as follows. Every input to the mixnet is encrypted under the public key of the recipient
for whom it is intended. Thus, unlike standard re-encryption mixnets, universal mixnets accept
ciphertexts encrypted under the individual public keys of receivers, rather than encrypted under the unique public key of the mix network. These ciphertexts are universally re-encrypted
and mixed by each server. The output of a universal mixnet is a set of ciphertexts. Recipients
can retrieve from the set of output ciphertexts those addressed to them, and decrypt them.
**Organization**
The rest of the paper is organized as follows. In the next section, we give an overview of
the main properties that distinguish universal mixnets from standard mixnets, and give one
example of a new application made possible by universal mixnets. This is followed in section 3
by a formal definition of semantic security for universal re-encryption, as well as a proposal
for creating a public-key cryptosystem with universal re-encryption based on ElGamal. In
section 4, we describe our construction for an asymmetric universal mixnet. We define and
prove the security properties of our system in section 5. In section 6, we propose a hybrid variant
of our universal mixnet construction that combines public-key and symmetric encryption to
handle long messages efficiently. We conclude in section 7.
## 2 Universal Mixnets: Properties and Applications
To motivate the constructions of this paper, we list here some of the main properties that set
apart universal mixnets from traditional re-encryption mixnets. We also give one example of
a new application made possible by universal mixnets: Anonymization of RFID tags.
**Universal mixnets hold no keying material. A universal mixnet operates without a mono-**
lithic public key and thus dispenses at the server level with the complexities of key generation,
key distribution, and key maintenance. This allows a universal mixnet to be set up more efficiently and with greater flexibility than a traditional re-encryption mixnet. A universal mixnet
can be rapidly re-configured: Servers can enter and leave arbitrarily, even in the middle of
a round of processing, without going through any setup. A mix server that crashes or otherwise disappears in the midst of the mixing process can thus be easily replaced by another server.
**Universal mixnets guarantee forward anonymity. The absence of shared keys means that**
universal mixnets offer perfect forward-anonymity. Even if all mix servers become corrupted,
the anonymity of previously mixed batches is preserved (provided that servers do not store
the permutations or re-encryption factors they used to process their inputs). In contrast, if the
keying material of a standard mix is revealed, an adversary with transcripts from previous mix
sessions can compromise the privacy of users.
1 We note that universal re-encryption has been independently devised by Danezis [7], although with a somewhat different application than we consider here.
-----
**Universal mixnets do not support escrow capability. The flip-side of perfect forward-**
anonymity is that is that it is not possible to escrow the privacy offered by a universal mixnet
in a straightforward fashion. Escrow is only achievable in a universal mix as long as every server
involved in the mixing remembers how it permuted its inputs and is willing to reveal that permutation. This may be a drawback from the perspective of law enforcement. In comparison,
escrow is possible in a traditional mix, provided that the shared key can be reconstructed. This
requires the participation of only a quorum of servers, not all of them.
**Efficiency. We present in this paper a public-key cryptosystem with universal re-encryption**
that is half as efficient as standard ElGamal: It requires exactly twice as much storage, and
also twice as much computation for encryption, re-encryption, and decryption. In this regard,
the universal mixnet constructions we propose in this paper are practical. The drawback of a
universal mixnet, as we discuss in detail below, is that receivers must attempt to decrypt all
output items in order to identify the messages intended for them.
**2.1** **Anonymizing RFID tags**
An interesting new application made possible by universal mixnets is the anonymization of
radio-frequency identification (RFID) tags. An RFID tag is a small device that is used to
locate and identify physical objects. RFID tags have very limited processing ability (insufficient
to perform any re-encryption of data), but they allow devices to read and write to their
memory [20, 21]. Communication with RFID tags is performed by means of radio, and the
tags themselves often obtain power by induction. Examples of uses of RFID tags include the
theft-detection tags attached to consumer items in stores and the plaques mounted on car
windshields for automated toll payment. Due to the projected decrease in the cost of RFID
tags, their use is likely to extend in the near future to a wide range of general consumer items,
including possibly even banknotes [26, 16].
This raises concerns of an emerging privacy threat. Most RFID tags emit static identifiers.
Thus, an adversary with control of a large base of readers for RFID tags may be able to track the
movement of any object in which an RFID tag is embedded, and hence learn the whereabouts
of the owner of that object. In order to prevent tracking of RFID tags, one could let some set of
(honest-but-curious) servers perform re-encryption of the information that is publicly readable
from RFID tags. The resulting system is surprisingly similar to a mix network, in which the
permutation of ciphertexts is replaced by the movement of the RFID tags.
A traditional mix network, however, only partially solves the problem of tracking. The
difficulty lies in the fact that the data contained in different RFID tags may be encrypted
under different public keys, depending on who possesses the authority to access that data. For
example, while the data contained in tags used for automated toll payment may be encrypted
under the public key of the transit agency, the data contained in tags attached to merchandise
in a department store may be encrypted under the public key of that department store. To
re-encrypt RFID tag data, a traditional mix network would need knowledge of the key under
which that data was encrypted. The public key associated with an RFID tag could be made
readable, but then the public key itself becomes an identifier permitting a certain degree of
tracking. This is particularly the case if a user carries a collection of tags, and may therefore
be identified by means of a constellation of public keys.
Universal mixnets offer a means of addressing the problem of RFID-tag privacy. If the data
contained in RFID tags is encrypted with a cryptosystem that permits universal re-encryption,
then this data can be re-encrypted without knowledge of the public-key. Thus universal re
-----
encryption may offer heightened privacy in this setting by permitting agents to perform reencryption without knowledge of public keys. While there have been previous designs using
mixes for the purposes of privacy protection for low-power devices (e.g., [19]), universal reencryption permits significant protocol and management simplification.
## 3 Universal Re-encryption
A conventional randomized public-key cryptosystem comprises a triple of algorithms, CS =
(KG, E, D), for key generation, encryption, and decryption respectively. We assume, as is often
the case for discrete-log-based cryptosystems, that system parameters and underlying algebraic
structures for CS are published in advance by a trusted party. These are generated according
to a common security parameter k. System parameters include or imply specifications of M,
**C, and R – respectively a message space, ciphertext space, and set of encryption factors. In**
more detail:
**– The key-generation algorithm (PK, SK)** KG outputs a random key pair.
_←_
**– The encryption algorithm C** E(m, r, PK) is a deterministic algorithm that takes as
_←_
input a message m **M, an encryption factor r** **R and a public key PK, and outputs a**
_∈_ _∈_
ciphertext C **C.**
_∈_
**– The decryption algorithm m** D(SK, C) takes as input a private key SK and ciphertext
_←_
_C_ **C and outputs the corresponding plaintext.**
_∈_
A critical security property for providing privacy in a mix network is that of semantic
_security. Loosely speaking, this property stipulates the infeasibility of learning any informa-_
tion at all about a plaintext from a corresponding ciphertext [12]. For a more formal definition, we consider an adversary that is given a public key PK, where (PK, SK) KG.
_←_
This adversary chooses a pair (m0, m1) of plaintexts. Corresponding ciphertexts (C0, C1) =
(E(m0, r0, PK), E(m1, r1, PK)) for r0, r1 ∈U R are computed, where ∈U denotes uniform, random selection. For a random bit b, the adversary is given the pair (Cb, C1−b), and tries to guess
_b. The cryptosystem CS is said to be semantically secure if the adversary can guess b with_
_advantage at most negligible in k, i.e. with probability at most negligibly larger than 1/2._
For a re-encryption mix network, an additional component known as a re-encryption algorithm, denoted by Re, is required in CS. This algorithm re-randomizes the encryption
factor in a ciphertext. In a standard cryptosystem, this means that C[′] Re(C, r, PK) for
_←_
_C, C[′]_ **C, r** **R, and a public key PK. Observe that re-encryption, in contrast to encryption,**
_∈_ _∈_
may be executed without knowledge of a plaintext. The notion of semantic security may be
naturally extended to apply to the re-encryption operation by considering an adversary that
chooses ciphertexts (C0, C1) under PK. The property of semantic security under re-encryption,
then, means the following: Given respective re-encryptions (Cb[′][, C]1[′] _−b[) in a random order, the]_
adversary cannot guess b with non-negligible advantage in k. Provided that Re yields the same
distribution of ciphertexts as E (given r ∈U R) or that the two distributions are indistinguishable, it may be seen that basic semantic security implies semantic security under re-encryption.
Bellare et al. [3] define another useful property possessed by the El Gamal cryptosystem.
Known as “key-privacy,” this property may be loosely stated as follows. Given a ciphertext
encrypted under a public key randomly selected from a published pair (PK0, PK1), an adversary cannot determine which key corresponds to the ciphertext with non-negligible advantage.
Key-privacy is one feature of the security property we develop in this paper for universal
re-encryption.
-----
As already explained, a universal cryptosystem permits re-encryption without knowledge
of the public key corresponding to a given ciphertext. Let us denote such a cryptosystem by
_UCS = (UKG, UE, URe, UD), where UKG, UE, and UD are key generation, encryption, and de-_
cryption algorithms. These are defined as in a standard cryptosystem. The difference between
a universal cryptosystem UCS and a standard cryptosystem resides in the re-encryption algorithm URe. The algorithm URe takes as input a ciphertext C and re-encryption factor r, but
no public key PK. Thus, we have C[′] URe(C, r) for C, C[′] **C, r** **R.**
_←_ _∈_ _∈_
To define universal semantic security under re-encryption, i.e., with respect to URe, it is
necessary to consider an adversarial experiment that is a variant on the standard one for semantic security. We define an experiment uss as follows for a (stateful) adversarial algorithm
. This experiment terminates on issuing an output bit. As above, we assume an appropriate
_A_
implicit parameterization of UCS under security parameter k. The idea behind the experiment is as follows. The adversary is permitted to construct universal ciphertexts under two
randomly generated keys, PK0 and PK1. These ciphertexts are then re-encrypted. The aim
of the adversary is to distinguish between the two re-encryptions. The adversary should be
unable to do so with non-negligible advantage.
Experiment Exp[uss]A [(][UCS, k][)]
_PK0 ←_ UKG; PK1 ← UKG;
(m0, m1, r0, r1) ←A(PK0, PK1, “specify ciphertexts”);
if m0, m1 ̸∈ **M or r0, r1 ̸∈** **R then**
output ‘0’;
_C0 ←_ UE(m0, r0, PK0); C1 ← UE(m1, r1, PK1);
_r0[′]_ _[, r]1[′]_ _[∈][U]_ **[R][;]**
_C0[′]_ 0[);][ C]1[′] 1[);]
_[←]_ [URe][(][C][0][, r][′] _[←]_ [URe][(][C][1][, r][′]
_b ∈U {0, 1};_
_b[′]_ _←A(Cb[′][, C]1[′]_ _−b[,][ “guess”);]_
if b = b[′] then
output ‘1’;
else
output ‘0’;
We say that UCS is semantically secure under re-encryption if for any adversary with
_A_
resources polynomial in K, the probability pr[Exp[uss]A [(][UCS, k][) = ‘1’]][ −] [1][/][2 is negligible in][ k][.]
The experiment uss captures the idea that the keys associated with ciphertexts are concealed by the re-encryption process in UCS. Thus, even an adversary with the opportunity
to compose the ciphertexts undergoing re-encryption cannot make use of differences in public
keys in order to defeat the semantic security of the cryptosystem.
**3.1** **Universal re-encryption based on ElGamal.**
We present a public-key cryptosystem with universal re-encryption that may be based on the
ElGamal cryptosystem implemented over any suitable algebraic group. The basic idea is simple:
We append to a standard ElGamal ciphertext a second ciphertext on the identity element. By
exploiting the algebraic homomorphism of ElGamal, we can use the second ciphertext to alter
the encryption factor in the first ciphertext. As a result, we can dispense with knowledge of
the public key in the re-encryption operation. As already noted, this construction is half as
efficient as standard ElGamal.
-----
Let E[m] loosely denote ElGamal encryption a plaintext m (under some key). In a universal
cryptosystem, a ciphertexts on message m consists of a pair [E[m]; E[1]]. ElGamal possesses
a homomorphic property, namely that E[a] _E[b] = E[ab] for group operator_ . Thanks to
_×_ _×_
this property, the second component can be used to re-encrypt the first without knowledge of
the associated public key. To provide more detail, let denote the underlying group for the
_G_
ElGamal cryptosystem; let q denote the order of . (Here the security parameter k is implicit
_G_
in the choice of .) Let g be a published generator for . The universal cryptosystem is as
_G_ _G_
follows. Note that we assume random selection of encryption and re-encryption factors in this
description.
**– Key generation (UKG): Output (PK, SK) = (y = g[x], x) for x ∈U Zq.**
**– Encryption (UE): Input comprises a message m, a public key y, and a random en-**
cryption factor r = (k0, k1) ∈ _Zq[2][. The output is a ciphertext][ C][ = [(][α][0][, β][0][); (][α][1][, β][1][)] =]_
[(my[k][0], g[k][0]); (y[k][1], g[k][1])]. We write C = UEPK(m, r) or C = UEPK(m) for brevity.
**– Decryption (UD): Input is a ciphertext C = [(α0, β0); (α1, β1)] under public key y. Verify**
_α0, β0, α1, β1 ∈G; if not, the decryption fails, and a special symbol ⊥_ is output. Compute
_m0 = α0/β0[x]_ [and][ m][1][ =][ α][1][/β]1[x][. If][ m][1][ = 1, then the output is][ m][ =][ m][0][. Otherwise, the]
decryption fails, and a special symbol is output. Note that this ensures a binding between
_⊥_
ciphertexts and keys: a given ciphertext can be decrypted only under one given key.
**– Re-encryption (URe): Input is a ciphertext C = [(α0, β0); (α1, β1)] with a random re-**
encryption factor r[′] = (k0[′] _[, k]1[′]_ [)][ ∈] _[Z]q[2][. Output is a ciphertext][ C][′][ = [(][α]0[′]_ _[, β]0[′]_ [); (][α]1[′] _[, β]1[′]_ [)] =]
[(α0α1k0[′] _[, β][0][β]1k0[′]_ [); (][α]1k1[′] _[, β]1k1[′]_ [)], where][ k]0[′] _[, k]1[′]_ _[∈][U]_ _[Z][q][.]_
Observe that the ciphertext size and the computational costs for all algorithms are exactly twice
those of the basic ElGamal cryptosystem. The properties of standard semantic security and
also universal semantic security under re-encryption (as characterized by experiment uss) may
be shown straightforwardly to be reducible to the Decision Diffie-Hellman (DDH) assumption
[4] over the group, in much the same way as the semantic security of ElGamal [25]. Thus,
_G_
one possible choice of G is the subgroup of order q of Zp[∗][, where][ p][ and][ q][ are primes such that]
_q_ _p_ 1. An alternative, with the advantage of more compact ciphertext representation, is a
_|_ _−_
group of prime order q defined over an appropriately selected elliptic curve such that the DDH
assumption is believed to be hard. Throughout the remainder of the paper, we work with the
ElGamal implementation of universal re-encryption, and let g denote a published generator for
the choice of underlying group .
_G_
## 4 Universal Mix Network Construction
We use the following scenario to introduce our universal mixnet construction. We consider a
number of senders who wish to send messages to recipients in such a way that the communication is concealed from everyone but the sender and recipient themselves. In other words, we
wish to establish channels between senders and receivers that are externally anonymous. We
assume that every recipient has an ElGamal private/public key pair (x, y = g[x]) in some published group . We also assume that every sender knows the public key of all the receivers with
_G_
whom she intends to communicate. (Alternatively, the sender may have a “blank” ciphertext
for this party. By this we mean an encryption using UE of the identity element in under the
_G_
public key of the recipient. A “blank” may be filled in without knowledge of the corresponding
public key through exploitation of the underlying algebraic homomorphism in ElGamal.) The
communication protocol proceeds as follows:
-----
1. Submission of inputs. Senders post to a bulletin board messages that are universally
encrypted under the public key of the recipient for whom they are intended. Every entry
on the bulletin board thus consists of a pair of ElGamal ciphertexts (E[m]; E[1]) under
the public key of the recipient. Recall that the semantic security of ElGamal ensures the
concealment of plaintexts. In other words, for plaintexts m and m[′], a universal ciphertext
(E[m]; E[1]) is indistinguishable from another (E[m[′]]; E[1]) to any entity without knowledge
of the corresponding private key.
2. Universal mixing. Any server can be called upon to mix the contents of the bulletin board.
This involves two operations: (1) The server re-encrypts all the universal ciphertexts on
the bulletin board using URe, and (2) The server writes the resulting new ciphertexts back
to the bulletin board in random order, overwriting the old ones. It is also desirable that a
server that mixes the inputs be able to prove that it operated correctly. This can be done
using a number of existing mixing schemes, e.g. [1, 2, 10, 13, 15, 17], and will be discussed
in greater detail below.
3. Retrieval of the outputs. Potential recipients must try to decrypt every encrypted
message output by the universal mixnet. Successful decryptions correspond to messages that
were intended for that recipient. The others (corresponding to decryption output ‘ ’) are
_⊥_
discarded by the party attempting to perform the decryption. Recall that our construction
of universal encryption based on El Gamal ensures a binding between ciphertexts and keys,
so that a given ciphertext can be decrypted only under one given key.
**Properties of the basic protocol:**
1. The universal mixnet holds no keying information. Public and private keys are managed
exclusively by the players providing input ciphertexts and receiving outputs from the mix.
2. The universal mixnet guarantees only external anonymity. It does not provide anonymity
for senders with respect to receivers. Indeed a receiver can trace a message intended for
her throughout the mixing process, since that message is encrypted under her public key.
If ciphertexts are not posted anonymously, this means that the receiver can identify the
players who have posted messages for her. This restriction to external anonymity is of little
consequence for the applications we focus on, namely protection against traffic analysis,
but should be borne in mind for other applications.
3. The chief drawback of universal mixnets is the overhead that they impose on receivers.
Because the public keys corresponding to individual output ciphertexts are unknown, it
may be necessary for a receiver to attempt to decrypt each output ciphertext in order to
find the right one, i.e., the ciphertext corresponding to her private key. Thus, a universal
mixnet imposes an overhead on receivers that is linear in the input batch size. (We discuss
ways below and in section 6 to reduce this overhead somewhat.)
**Low-volume anonymous messaging: anonymizing bulletin boards.**
For simplicity, we have described above the operation of a universal mixnet in which inputs are
submitted, mixed and finally retrieved. This sequence of events is characteristic of all mixes.
Unlike regular mixes however, universal mixes allow for repeated interleaving of the submission,
mixing and retrieval steps. What makes this possible is that the decryption is performed by
the recipients of the message rather than by the mixnet, so that existing messages posted to
the bulletin board are at all times indistinguishable from new messages. New inputs may be
constantly added to the existing content of the bulletin board, and outputs retrieved, provided
there is at least one round of mixing between every submission and retrieval to ensure privacy.
This suggests a generalization of the private communication protocol described above, in
which the bulletin board maintains at all times a pool of unclaimed messages. In other words,
-----
universal mixing lends itself naturally to the construction of an anonymizing bulletin board.
Senders may add messages and receivers retrieve them at any time, provided there is always
at least one round of mixing between each posting and retrieval. This protocol appears well
suited to guarantee anonymity from external observers in a system in which few messages are
exchanged. The privacy of the protocol relies on the existence of a steady pool of undelivered
messages rather than on a constant flow of new messages. The former condition appears much
easier to satisfy than the latter in cases when the total number of exchanged messages is
small. This pooling of messages affords good anonymity protection, without the usual lack of
verifiability of correct performance that vexes such schemes.[2]
A potential drawback of a bulletin board based on universal mixing is that one must download the full contents in order to be assured of obtaining all of the messages addressed to
oneself. This becomes problematic if the number of messages on the bulletin board is permitted to grow indefinitely. To mitigate this problem, it is possible to have recipients remove the
messages they have received.[3] An anonymizing bulletin board based on universal mixing has
the important privacy-protecting feature that removal of a particular message does not reveal
which entity posted that message. Another important observation, as described in the next
section, is that only a portion of each message on a bulletin board need be downloaded to
allow a recipient to determine which messages are intended for her. This further restricts the
work required by a receiver.
**RFID-tag privacy.**
Universal re-encryption may be used to enhance the privacy of RFID tags. The idea is to permit
powerful computing agents external to RFID tags to universally re-encrypt the tag data (recall
that the tags lack the computing power necessary to do the re-encryption themselves). Thus,
for example, a consumer walking home with a bag of groceries containing RFID tags might
have the ciphertexts on these tags re-encrypted by computing agents that are provided as a
public service by shops and banks along the way. In this case, the tags in the bag of groceries
will periodically change appearance, helping to defeat any tracking attempt.
Application of universal mixnets to RFID-tag privacy is different in some important respects
from realization of an anonymous bulletin board. As re-encryption naturally occurs for RFID
tags on an individual basis, re-encryption in this setting may be regarded as realizing an
_asynchronous mixnet. There is also a special security consideration in this setting. Suppose_
that the ciphertext on an RFID tag is of the form (α, β); (1, 1) (where ’1’ represents the identity
element for ). Then the ciphertext on the tag will not change upon re-encryption. Thus, it is
_G_
important to prevent an active adversary from inserting such a ciphertext onto an RFID tag so
as to be able to trace it and undermine the privacy of the possessor. In particular, on processing
ciphertexts, re-encryption agents should check that they do not possess this degenerate form.
Of course, an adversary in this environment can always corrupt ciphertexts. Note, however,
that even a corrupted ciphertext (α[′], β[′]); (γ, δ) will be rendered unrecognizable to an adversary
provided that γ, δ = 1.
_̸_
2 So-called pool mixes typically use processing delays in asynchronous settings to hide timing information.
They were first described by Lance Cottrell in the nineties [6]. See [23] for a further discussion of pool mixes,
and [9] for an approach to verifying correct functioning of pool mixes.
3 To ensure that messages are only removed by the intended recipient, a proof of knowledge of the corresponding
decryption key is required. Note that such a proof can be performed without disclosing the public key
associated with the required decryption key. For ciphertext C = [(α0, β0); (α1, β1)], this may take the form
of a non-interactive zero-knowledge proof of knowledge of an exponent x such that α1 = β1[x] [– essentially a]
Schnorr signature [22].
-----
## 5 Security
In this section, we define two security properties of universal mixnets:
**– Correctness: The mixnet is correct if the set of output it produces is a permutation of**
the set of inputs.
**– Communication privacy: The mixnet guarantees communication privacy if, when Alice**
sends a message to Bob and Cathy sends a message to Dario, an observer can not tell
whether Alice (resp. Cathy) sent a message to Bob or Dario.
**Correctness. Correctness for universal mixnets follows directly from the definition of cor-**
rectness for standard mixnets. Like standard mix servers, universal servers must prove that
they have performed the mixing operation correctly. For this, it is possible to draw on essentially any of the proof techniques presented in the literature on mixnets, as nearly all apply
to ElGamal ciphertexts. For example, to achieve universal verifiability, it is possible to employ
the proof techniques in [10, 17, 15]. A small technical consideration, which may be dealt with
straightforwardly, is the form of input ciphertexts. Input ciphertexts in most mix network constructions consist of a single ElGamal ciphertext, while in our construction, an input consists
of a universal ciphertext, and thus two related ElGamal ciphertexts.
**Communication privacy. We define next the property of communication privacy. In order**
to state this definition formally, we abstract away some of the operations of the mixnet by
defining them in terms of oracle operations. We do this so as to focus our exposition on our
universal construction, rather than underlying primitives, particularly as our construction can
make use of a broad range of choices of such primitives. We define three oracles:
**– An oracle MIX It universally re-encrypts all ciphertexts on the bulletin board BB and**
outputs back to BB the new set of ciphertexts in a randomly permuted order. In practice,
any mix network with public verifiability may be substituted for our oracle MIX.
**– An oracle POST that permits message posting. This oracle requires a poster to submit a**
message, encryption factors and ciphertext. The oracle verifies that the message, encryption
factors and ciphertext are elements of the appropriate groups. The oracle permits posting if
the ciphertext is a valid encryption of the message with the given encryption factors. Note
that the oracle POST may be regarded as simulating a proof of knowledge of the plaintext
and the encryption factor and a verification thereof. In practice, it could be instantiated
with standard discrete-log-based proofs of knowledge, e.g., [8], in either their interactive or
non-interactive forms.
**– An oracle RETRIEVE that permits message retrieval. The oracle takes a private key and**
ciphertext from a user. The oracle verifies that the private key and ciphertext are elements
of the appropriate groups. The user is allowed to remove the ciphertext if it is encrypted
under the private key. Recall that our construction of universal encryption based on El
Gamal ensures a binding between ciphertexts and keys, so that a given ciphertext can be
decrypted only under one given key. The oracle RETRIEVE, like POST, abstracts away a proof
of knowledge of the plaintext.
We define communication privacy in terms of an experiment Exp[comm][−][priv] defined as
follows. The adversary may make an arbitrary number of calls to any of the oracles RETRIEVE,
```
MIX, or POST and may order these calls as desired. We enumerate the first several steps here
```
for reference in our proof.
-----
Experiment Exp[comm][−][priv](UCS, k)
_A_
1. PK0 ← UKG; PK1 ← UKG;
2. (m0, m1) ←A(PK0, PK1, “specify plaintexts”);
3. b ∈U {0, 1};
4. C0[′] [=][ UE][PK]b[(][m][b][) and][ C]1[′] [=][ UE][PK]1−b[(][m][1][−][b][) appended to][ BB][;]
5. MIX invoked;
6. (BB);
_A_
7. L ←{C ∈ _BB s.t. C is a valid ciphertext under PK0};_
8. b[′] (L, “guess b”);
_←A_
if b = b[′] then
output ‘1’;
else
output ‘0’;
An intuitive description of this experiment is as follows. Alice and Bob wish each to transmit
a single message to one of Cathy and Dario, who possess public keys PK0 and PK1 respectively.
Our aim is to ensure that the adversary cannot tell whether Alice is sending a message to Cathy
or Dario – and likewise to whom Bob is transmitting. The adversary is given the special (strong)
power of determining which plaintexts, m0 and m1, are to be received by Cathy and Dario.
The adversary observes Alice posting ciphertext C0[′] [and Bob posting ciphertext][ C]1[′] [, but does]
not know which ciphertext is for Cathy and which is for Dario. The bulletin board is then
subjected to a mixing operation so as to conceal the communication pattern. The adversary
may subsequently control when and how the mix network is invoked, and may place its own
ciphertexts on the bulletin board. Finally, at the end of the experiment, the adversary is given
a list L of all ciphertexts encrypted under PK0, i.e., all the messages that Cathy retrieves. This
list L will include the one such message posted by Alice or Bob in addition to all messages
encrypted under PK0 and posted by the adversary. The task of the adversary is to guess
whether it was Alice who sent a message to Cathy (case b = 0) or Bob (case b = 1).
**Definition 1. (Communication privacy) We say that a universal mixnet for UCS pos-**
_sesses communication privacy if for any adversary_ _that is polynomial time in k, we have_
_A_
pr[Exp[comm][−][priv](UCS, k) = 1] 1/2 is negligible in k.
_A_ _−_
**Theorem 1. Our universal mixnet possesses communication privacy provided that UCS has**
_universal semantic security under re-encryption. For our described construction involving El-_
_Gamal, privacy may consequently be reduced to the DDH assumption over_ _._
_G_
**Proof: Assume that we have an adversary** for which pr[Exp[comm][−][priv](UCS, k) = 1] 1/2 is
_A_ _A_ _−_
non-negligible in k. We build a new adversary which uses as a subroutine and for which
_A[′]_ _A_
pr[Exp[uss]
_A[′][ (][UCS, k][) = ‘1’]][ −]_ [1][/][2 is non-negligible in][ k][ (i.e.][ A][′][ breaks the universal semantic]
security of the underlying encryption scheme). operates as follows:
_A[′]_
**– At the beginning of the experiment Exp[uss], A[′]** is given two public keys PK0 and PK1. A[′]
gives these two keys to . This simulates step 1 of Exp[comm][−][priv].
_A_
**– When** calls one of the oracles POST, MIX or RETRIEVE, can trivially simulate the oracle
_A_ _A[′]_
for the requested operation for .
_A_
**– In step 2 of experiment Exp[comm][−][priv], A specifies plaintexts m0 and m1. A[′]** selects random
encryption factors r0 and r1 and computes C0 = UEPK0(m0, r0) and C1 = UEPK1(m1, r1).
submits these in the second step of experiment Exp[uss]. then receives as input from
_A[′]_ _A[′]_
experiment Exp[uss] two new ciphertexts C0[′] [and][ C]1[′] [.]
-----
**– In step 4 of Exp[comm][−][priv], A[′]** posts C0[′] [and][ C]1[′] [to the bulletin board.]
**– In step 7 of Exp[comm][−][priv], A[′]** must identify the set of outputs encrypted under PK0.
Note that can easily identify among the outputs that correspond to inputs originally
_A[′]_
submitted by A those encrypted under PK0, since it controls the oracle POST and MIX. The
only difficulty is for A[′] to decide which of C0[′] [and][ C]1[′] [is encrypted under][ PK][0][ and which]
under PK1. Since A[′] doesn’t know that, it arbitrarily assigns C0[′] [to the list][ L][ of ciphertexts]
encrypted under PK0.
In the last step of the simulation, A[′] assigns C0[′] [arbitrarily to][ L][. We claim that if][ A][ can]
distinguish between the case where this assignment to L is correct and the case where it is
incorrect, then can be used to break universal semantic security in Exp[uss]. This may be
_A_
achieved with a small modification of our simulation as follows: (1) A[′] lets C0[′] [=][ C][0][ and]
_C1[′]_ [=][ C][1][, but invokes][ Exp][uss][ on the pair (][C]0[′] _[, C]1[′]_ [) during the mixing operation in step 5 and]
(2) A[′] submits to Exp[uss] the bit b[′] yielded by A at the end of the experiment. Let us assume,
therefore, that the assignment to L is correct.
Given this, when A outputs its guess b[′], A[′] then outputs the same bit b[′] as its guess for the
experiment Exp[uss]. It is clear now that when guesses correctly, so does . This concludes
_A_ _A[′]_
our proof.
_⊓⊔_
**Security of UCS and chosen-ciphertext attacks.**
The cryptosystem UCS we employ here inherits the semantic security property of the underlying El Gamal cipher under the DDH assumption. This property is critical to our definition
of communications privacy. Our model for communications privacy makes one simplifying assumption that must be noted, though: We assume that the adversary does not learn any
information about plaintexts. For this reason, we do not require adaptive-chosen ciphertext
(CCA) security of our cryptosystem. In fact, we cannot achieve CCA security in the strictest
sense in our system: In order to permit re-encryption, ciphertext must be malleable. Note,
however, that because of the need to demonstrate knowledge of the plaintext and encryption
factors in the POST operation, it is infeasible for an adversary to re-post a message or to post
a new message with a related plaintext.
On the other hand, there may be circumstances in which an adversary may indeed learn
information about plaintexts in our system. To show this in a formal sense, however, it would
be necessary to modify our universal cryptosystem so as to achieve CCA security with benign
_malleability, as defined by Shoup [24]. In Shoup’s terminology, we would need to require an_
induced compatible relation of plaintext equivalence by formatting plaintexts with appropriate
padding. We omit detailed discussion of this topic, however, in this paper. An adversary that
can gain significant information about received messages can, after all, break the basic privacy
guarantees of the system.
## 6 Hybrid universal mixing
We describe next a variant mixnet called a hybrid universal mixnet. This type of mixnet combines symmetric and public-key encryption to accommodate potentially very long messages (all
of the same size) in an efficient manner. We refer the interested reader to [18, 14] for definitions
and examples of hybrid mixnets. Our definition of a universal hybrid mix considers a weaker
threat model than above with respect to correctness. Our universal hybrid mix cannot be verified to correctly execute the protocol because of the use we make of symmetric encryption.
-----
Thus, we restrict our security model to mix servers subject only to passive adversarial corruption. Such servers are also known as honest-but-curious. They follow the protocol correctly but
try to learn as much information as possible from its execution.
For efficiency, inputs m are submitted to a hybrid mix encrypted under an initial symmetric
(rather than public) key. We denote by ϵk[m] the symmetric-key encryption of m under key
_k. Each mix server Si consecutively re-encrypts the output of the previous mix under a new_
random symmetric key ki. If there are k mix servers, the final output of the mix is therefore
_ϵkn[ϵkn−1[. . . ϵk1[ϵk[m]] . . .]. The symmetric keys k, k1, . . ., kn must be conveyed alongside the_
encrypted message to enable decryption by the final recipient. These keys are themselves
encrypted as universal ciphertexts under the public key of the recipient. Universal encryption
provides a very efficient way of transmitting encryptions of the symmetric keys in a way that
does not compromise privacy.
Let us now give a more detailed definition of our hybrid universal mixnet. Our construction
imposes an upper bound n on the maximum number of times that the mixing operation is
performed by the mixnet on any given ciphertext. The protocol consists of the following steps:
1. Submission of inputs. An input ciphertext takes the form
_ϵk0[m], E[1], (E[k0], E[1] . . . E[1])_
where ϵk0[m] denotes symmetric-key encryption of m under key k0. This is followed by an
encryption of 1, and by a vector of ciphertexts on keys, where only the first element is filled
in (with k0), leaving the remaining n − 1 elements as encryptions of 1.
2. Universal mixing. The i[th] server to perform the mixing operation does the following for
each of the ciphertexts on the bulletin board:
**– Generates a random symmetric key ki;**
**– Adds a new layer of symmetric encryption to m under key ki;**
**– Uses the second element, E[1], to compute an encryption of ki – call this E[ki];**
**– Rotates the elements of the vector one step leftwards, then substituting the first element**
with E[ki]; and
**– Re-encrypts the second element and each element of the vector.**
When it has thus processed all its inputs in this manner, the server outputs them back to
the bulletin board in a random order.
3. Retrieval of the outputs. At the end of d _n mixing operations, the final output of the_
_≤_
mixnet assumes the form:
_ϵkd[ϵkd−1[. . . ϵk0[m]] . . .], E[1], ({E[1]}[n][−][d], E[k0] . . . E[kd]),_
where _E[1]_ denotes n _d ElGamal ciphertexts on the identity element. As before,_
_{_ _}[n][−][d]_ _−_
recipients try to decrypt every output of the mixnet and discard those outputs for which
the decryption fails. Only the second element, E[1], however, has to be decrypted in order
for a party to determine whether the ciphertext is intended for her.
**Remark: In principle, it is possible to use the “blank” ciphertext E[1] to append ciphertexts on**
as many symmetric keys as desired, and thus re-encrypt indefinitely. The reason for restricting
the number of “blank” ciphertexts to exactly n is to preserve a uniform length, without which
an adversary can distinguish among ciphertexts that have undergone differring numbers of
re-encryptions. A drawback of this approach is that a ciphertext re-encrypted more than n
times will become undecipherable by the receiver. Given enough messages, it is alternatively
possible to permit messages to grow in sizes according to their “ages”, i.e., the number of
re-encryptions they have undergone, and to pool them accordingly.
-----
## 7 Conclusion
Universal re-encryption represents a simple modification to the basic El Gamal cryptosystem
that permits re-randomization of ciphertexts without knowledge of the corresponding private
key. This provides a valuable tool, as we show, for the construction of privacy-preserving
architectures that dispense with the complications and risks of distributed key setup and
management. The costs for the basic universal cryptosystem are only twice those of ordinary
El Gamal. On the other hand, the problem of receiver costs in a universal mixnet presents
a compelling line of further research. In the construction we have proposed, a receiver must
perform a linear number of decryptions to identify messages intended for her. A method for
reducing this cost would be appealing from both a technical and practical standpoint.
## References
1. M. Abe. Mix-networks on permutation networks. In K-Y. Lam, E. Okamoto, and C. Xing, editors, ASI_ACRYPT ’99, volume 1716 of Lecture Notes in Computer Science, pages 258–273. Springer-Verlag, 1999._
2. M. Abe and F. Hoshino. Remarks on mix-networks based on permutation networks. In PKC ’01, pages
317–324. Springer-Verlag, 2001. LNCS no. 1992.
3. M. Bellare, A. Boldreva, A. Desai, and D. Pointcheval. Key-privacy in public-key encryption. In C. Boyd,
editor, ASIACRYPT ’01, pages 566–582, 2001. LNCS no. 2248.
4. D. Boneh. The Decision Diffie-Hellman problem. In ANTS ’98, pages 48–63. Springer-Verlag, 1998. LNCS
no. 1423.
5. D. Chaum. Untraceable electronic mail, return addresses, and digital pseudonyms. Communications of the
_ACM, 24(2):84–88, 1981._
6. L. Cottrell. Mixmaster & remailer attacks, 1995. http://www.obscura.com/ loki/ remailer/remaileressay.html.
7. G. Danezis, 2002. Personal communication.
8. A. de Santis, G. di Crescenzo, G. Persiano, and M. Yung. On monotone formula closure of SZK. In FOCS
_’94, pages 454–465. IEEE Press, 1994._
9. E. Franz, A. Graubner, A. Jerichow, and A. Pfitzmann. Comparison of commitment schemes used in mixmediated anonymous communication for preventing pool-mode attacks. In C. Boyd and E. Dawson, editors,
_ACISP ’98, pages 111–122. Springer-Verlag, 1998. LNCS no. 1438._
10. J. Furukawa and K. Sako. An efficient scheme for proving a shuffle. In J. Kilian, editor, CRYPTO ’01,
volume 2139 of Lecture Notes in Computer Science, pages 368–387. Springer-Verlag, 2001.
11. T. El Gamal. A public key cryptosystem and a signature scheme based on discrete logarithms. _IEEE_
_Transactions on Information Theory, 31:469–472, 1985._
12. S. Goldwasser and S. Micali. Probabilistic encryption. J. Comp. Sys. Sci, 28(1):270–299, 1984.
13. M. Jakobsson and A. Juels. Millimix: Mixing in small batches, June 1999. DIMACS Technical Report
99-33.
14. M. Jakobsson and A. Juels. An optimally robust hybrid mix network. In PODC ’01, pages 284–292. ACM
Press, 2001.
15. M. Jakobsson, A. Juels, and R. Rivest. Making mix nets robust for electronic voting by randomized partial
checking. In D. Boneh, editor, USENIX ’02, pages 339–353, 2002.
16. A. Juels and R. Pappu. Squealing euros: Privacy protection in rfid-enabled banknotes. In R. Wright, editor,
_Financial Cryptography 2003, 2003. To appear._
17. A. Neff. A verifiable secret shuffle and its application to e-voting. In P. Samarati, editor, ACM CCS ’01,
pages 116–125. ACM Press, 2001.
18. M. Ohkubo and M. Abe. A length-invariant hybrid mix. In T. Okamoto, editor, ASIACRYPT ’00, volume
1976 of Lecture Notes in Computer Science, pages 178–191. Springer-Verlag, 2000.
19. M. Reed, P. Syverson, and D. Goldschlag. Protocols using anonymous connections: mobile applications. In
_Security Protocols ’97, pages 13–23. Springer-Verlag, 1997. LNCS 1361._
20. S. Sarma. Towards the five-cent tag. Technical Report MIT-AUTOID-WH-006, MIT Auto ID Center, 2001.
Available from http://www.autoidcenter.org/.
21. S. Sarma. Radio-frequency identification systems. In B. Kaliski, editor, CHES ’02. Springer-Verlag, 2002.
To appear.
22. C.-P. Schnorr. Efficient signature generation by smart cards. Journal of Cryptology, 4(3):161–174, 1991.
-----
23. A. Serjantov, R. Dingledine, and P. Syverson. From a trickle to a flood: active attacks on several mix types.
In Information Hiding ’02, pages 36–52. Springer-Verlag, 2002. LNCS no. 2578.
24. V. Shoup. A proposal for an iso standard for public key encryption (version 2.1), 20 December 2001.
Manuscript.
25. Y. Tsiounis and M. Yung. On the security of ElGamal-based encryption. In Workshop on Practice and
_Theory in Public Key Cryptography (PKC ’98), pages 117–134. Springer, 1998. LNCS no. 1431._
26. J. Yoshida. Euro bank notes to embed RFID chips by 2005. EE Times, 19 December 2001. Available at
http://www.eetimes.com/story/OEG20011219S0016.
-----
|
{
"disclaimer": "Notice: The following paper fields have been elided by the publisher: {'abstract', 'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1007/978-3-540-24660-2_14?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1007/978-3-540-24660-2_14, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "CLOSED",
"url": ""
}
| 2,004
|
[
"JournalArticle"
] | false
| 2004-02-23T00:00:00
|
[] | 12,719
|
en
|
[
{
"category": "Engineering",
"source": "external"
},
{
"category": "Engineering",
"source": "s2-fos-model"
},
{
"category": "Environmental Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/0244a3e459fe7b78debc2ee71bbf73514ca013bd
|
[
"Engineering"
] | 0.820646
|
Distributed control and energy storage requirements of networked Dc microgrids
|
0244a3e459fe7b78debc2ee71bbf73514ca013bd
|
[
{
"authorId": "145219251",
"name": "W. Weaver"
},
{
"authorId": "143767900",
"name": "R. Robinett"
},
{
"authorId": "3129393",
"name": "G. Parker"
},
{
"authorId": "40213827",
"name": "D. Wilson"
}
] |
{
"alternate_issns": null,
"alternate_names": null,
"alternate_urls": null,
"id": null,
"issn": null,
"name": null,
"type": null,
"url": null
}
| null |
# Distributed Control and Energy Storage Requirements of Networked Dc Microgrids
Wayne W. Weaver[a], Rush D. Robinett, III[a], Gordon G. Parker[a], David G.
Wilson[b]
_aMichigan Technological University, 1400 Townsend Dr., Houghton, Michigan, 49931,_
_USA_
_bSandia National Laboratories, P.O. Box 5800, Albuquerque, New Mexico, 87185, USA_
**Abstract**
Microgrids are a key technology to help improve the reliability of electric
power systems and increase the integration of renewable energy sources. In
terconnection and networking of smaller microgrids into larger systems have
potential for even further improvements. This paper presents a novel ap
proach to a distributed droop control and energy storage in networked dc
microgrids. Distributed control is necessary to prevent single points of fail
ure along with flexibility and adaptability to changing energy resources. The
results show that systems with random sources and fast update rates, a
networked microgrid structure can minimize required energy storage require
ments.
_Keywords:_ microgrid, distributed control, energy storage, optimization,
power electronics
_Email addresses: wwweaver@mtu.edu (Wayne W. Weaver), rdrobine@mtu.edu_
(Rush D. Robinett, III), ggparker@mtu.edu (Gordon G. Parker), dwilso@sandia.gov
(David G. Wilson)
_Preprint submitted to Control Engineering Practice_ _July 6, 2015_
-----
**1. Introduction**
A microgrid is a collection of energy resources on a common network.
These resources include generation, conversion, loads and storage devices
[1]. The model of centralized generation is gradually being replaced by a
distributed generation model [2]. The emerging technologies in renewable
and distributed generation can have lower emissions and cost. The micro
grid concept gives a solution for integration of a large number of distributed
generations without causing disruption in the utility network. Microgrids
also allow for local control of the distributed generation units and attests
to the flexibility to operate autonomously during disturbances in the util
ity network to increase reliability [3, 4, 5]. In addition, the interconnection
and networking of groups of microgrids can reduce the energy storage re
quirements. However, the interconnections and power flow control between
microgrids increase costs, complexity and failure modes [6].
One of the main challenges for microgrid design and control is that gener
ation capacity is very close to load demand. In addition, with the stochastic
nature of most renewable energy sources there is a need for energy storage
[7, 8, 9]. Energy storage can mitigate both long-term and short-term system
transients. For example, a long term transient would be the generation vari
ations over hours and days from a wind turbine or photo-voltaic array due to
weather patterns. Short-term transients could include step changes in load
or faults in the system where the response is on the order of seconds or frac
tions of a second. Therefore, a proper energy storage strategy will include
devices that can respond at the proper bandwidth of system transients.
Within micorgrids, there are many approaches to the control and opti
2
-----
mization of each element. A centralized approach is able to reach higher levels
of performance at the cost of single points of failure and lack of flexibility.
A distributed and de-centralized approach allows a very flexible system that
can adapt to changing system structures and situations. A typical approach
to distributed control is droop control [10, 11]. In dc microgrids, droop con
trol is equivalent to creating a virtual impedance between the source and the
bus such that the total load current is distributed to the sources based on
the weighted sum of droop settings. The standard way to implement droop
control is through duty cycle control of the dc/dc converter interface to the
bus.
Since many renewable sources are dc, such as photovoltaics, they require
additional power conversion to connect to an ac system. In addition, most
electronic loads require a dc power conversion step and many energy storage
technologies, such as batteries or super-capacitors are also dc. Therefore, a
dc system is a viable option for power distribution in microgrids [12, 13].
However, a dc microgrid with a high penetration of renewable sources can
require large energy storage capacity to maintain the system and to mitigate
variations in the sources [14].
This paper presents an alternative approach that uses the local energy
storage device at the source to actuate the droop control in local and net
worked microgrids. The duty cycles of the converters are updated on a
periodic interval to only match the source voltage to the bus. This approach
allows the requirements for energy storage in capacity and bandwidth to
be studied and designed with variations in the renewable energy sources and
load. The novelty of this approach is that the system is not actuated through
3
-----
the typical approach of feed-back control of duty cycle of the dc-dc converter,
but through feed-back control of the storage devices in the system. Further,
the feed-back process of the energy storage actuation is implemented through
a distributed droop control. The duty cycles of the dc-dc converter are only
updated on periodic intervals through a feed-forward process.
The paper will first present the electrical system model of the dc-dc boost
coverer, energy storage devices and microgrid structure. Next, the controls
are developed for the feed-forward control of the duty cycles and the feed
back control of the energy storage devices. Then, the distributed droop
control is shown. Finally the feed-forwad, feed-back and distributed droop
are demonstrated through simulation of several operational senarios.
**2. Dc-dc converter and microgrid model with energy storage**
In dc microgrids the interface to the distribution bus is through dc to
dc converters. If the bus voltage is higher than the sources, the interface
converter will be in the form of a boost converter. In this paper, the sources
will be paired with an energy storage device and therefore the boost converter
will be bi-directional.
_2.1. Dc-dc Converter Model_
A bidirectional dc-dc converter circuit is shown in Fig. 1. The converter
is implemented with a power pole of two power MOSFETs, which enable
forward blocking and reverse current [15]. In this configuration the top switch
state is defined as q while the bottom switch is 1 _q, then the dynamic_
_−_
4
-----
equations for the converter are
_L_ _[di][L]_ (1)
_dt_ [=][ v][s][ −] _[i][L][R][L][ −]_ _[q][(][t][)][v][C]_
_dvC_
_CB_ (2)
_dt_ [=][ q][(][t][)][i][L][ −] _[i][Load][.]_
Formally, when iL > 0 this converter is a boost converter and when iL < 0
it is a buck converter. However, in this paper we will refer to the circuit as
a boost, even though it can have bidirectional current. The time average of
the switch state is found from
1
_λ =_
_Tsw_
_t_
�
_t−Tsw_
_q(t)dt_ (3)
where Tsw is the switching period. The dynamic average model of the con
verter [16] is
_L_ _[di][L]_ (4)
_dt_ [=][ v][s][ −] _[i][L][R][L][ −]_ _[λv][C]_
_dvC_
_CB_ (5)
_dt_ [=][ λi][L][ −] _[i][Load][.]_
This model of the converter ensures two quadrant operation because of the
choice of the switches in Fig. 1 as, vC > vs and bidirectional current iL.
_2.2. Source and storage model_
Consider the bus interface boost converter shown in Fig. 2 which has
two voltage sources. The voltage source vv represents an energy source such
as a generator or photo-voltaic panel. The source vu represents an energy
storage device such as a battery or capacitor. Both voltage sources include a
5
-----
_s_ Source
Figure 1: A bi-directional dc-dc converter.
series equivalent resistance Rv. Both the source and storage model in Fig. 2
represent a Thevenin equivalent of a source and a second stage converter
such that the output terminal voltages are controllable. Both the source and
storage converters will contribute current to the inductor of the bus interface
boost converter. The total inductor current is
_iL =_ _[v][u][ −]_ _[v][L]_ + _[v][v][ −]_ _[v][L]_ (6)
_Ru_ _Rv_
|Load L Load C Source|Col2|
|---|---|
|Source||
|||
and the node voltage vL is
_Ru_ _Rv_ _RuRv_
_vL = vv_ + vu _−_ _iL_ _._ (7)
_Ru + Rv_ _Ru + Rv_ _Ru + Rv_
It is seen in (7) that the total voltage is a sum of two series sources and a
line impedance. The voltages and resistances (7) can be lumped into the new
6
-----
_L_ _B_
_v_ _u_
Figure 2: A boost converter model with two parrallel voltage sources, where vv represents
an energy source and vu represents an energy storage device.
variables
_Ru_
_v = vv_ (8)
_Ru + Rv_
_Rv_
_u = vu_ (9)
_Ru + Rv_
_RuRv_
_RL =_ _._ (10)
_Ru + Rv_
The boost converter with energy source and storage devices can be modeled
as a series combination in the new variables of voltage sources and Thevenin
resistance as shown in boost converter models of Fig. 3 with the average
mode [16] dynamic model
|Source|Col2|Storage|
|---|---|---|
|v||u|
||||
_L_ _[di][L]_ (11)
_dt_ [=][ −][λv][B][ −] _[i][L][R][L][ +][ u][ +][ v]_
_dvB_ 1
_CB_ (12)
_dt_ [=][ λi][L][ −] _[v][B]_ _RB_
where λ represents duty cycle of the converter.
_2.3. Microgrid model_
A simple dc microgrid model is shown in Fig. 3 where N boost converters
have the series model derived in section 2.2 for the source and local energy
7
-----
|1 L,1 1 B 1 i i i i L,i i i|1|B|Col4|
|---|---|---|---|
|||||
Figure 3: Microgrid of equivalent series boost converters.
storage devices. The model of the microgrid in Fig. 3 has the dynamic
equations
_diL,i_
_Li_ _dt_ = −λivB − _iL,iRL,i + ui + vi,_ _i = 1..N_ (13)
_N_
_dvB_ � 1
_CB_ _dt_ [=] _λiiL,i −_ _vB_ _RB_ + uB (14)
_i=1_
where uB represents a centralized bus energy storage device. In implemen
tation, uB would be an energy storage device, such as a battery or ultra
capacitor, interfaced to the bus through a dc to dc power converter. How
ever, for this study it is important to represent the energy storage devices
as ideal sources that are controlled to respond to system dynamics. In this
way the storage requirements, such as total energy delivered or absorbed and
response bandwidth can be determined. The microgrid load is modeled as
a simple bus resistor and capacitor RB and CB, so each microgrid will have
_N + 1 states. The next challenge is to control the boost converters in such_
way as the load current is shared between the sources. In addition, because
8
-----
of highly variable sources vi, which is indicative of renewable energy sources,
storage at the local converters and central bus is required to maintain a nom
inal operating point of the bus voltage. In general, the optimal distribution,
capacity and bandwidth of the energy storage devices is not well understood,
especially for a distributed control architecture.
_2.4. Networked microgrid model_
Each microgrid as depicted in Fig. 3 is a self contained power system.
However, due to the lack of diversity and inertia in the microgrid, its stability,
reliability and flexibility may not be optimal. If multiple microgrids are
interconnected then it may be possible to improve the reliability, flexibility
and stability of the overall system. However energy storage remains a crucial
element in the system. In addition, as the network of microgrids becomes
larger, a distributed control architecture becomes even more attractive given
that there is no need for a communication infrastructure or central controller.
However, it is important to understand the stability margins and energy
storage requirements in such a system.
Given a known set of sources and loads, there are numerous permutations
of networked microgrids. In general, a networked microgrid will have a bus
bar voltage vector V, and an injected current vector I and an interconnection
admittance matrix Y. Then, the following matrix relationship can be written
**I = YV** (15)
where all injected current and voltage vectors are k 1, with k representing
_×_
the number of bus bars. The bus admittance matrix Y is symmetric in most
9
-----
cases and is a function of line admittances and shunt load resistances of the
microgrid. The solution of the nodal equations in (15) follows the method
ology discussed in [17]. In this paper, a simple common bus interconnected
microgrid structure will be used.
Consider the networked microgrid configuration shown in Fig. 4 where
there are N sub-microgrids connected to a common bus. Each microgrid has
a bus storage element uB,k, an RC load, and M source converters. Power
flow for interconnection of the buses is achieved through an interconnection
converter from bus k to network n denoted k _n with local energy storage_
_→_
_uk→net. In the configuration shown in Fig. 4, the network bus net must be_
at a higher voltage than any of the lower microgrids 1..N . In this configura
tion, power can directly flow from one microgrid to another. The networked
microgrid Fig. 4 will have the system of equations
_diL,k,m_
_Lk,m_ _dt_ = −λk,mvB,k − _iL,k,mRL,k,m + uk,m + vk,m,_
_k = 1..N,_ _m = 1..M_ (16)
_M_
�
_λk,miL,k,m −_ _iL,k→net −_ _R[v][B,k]B,k_ + uB,k,
_m=1_
_CB,k_
_dvB,k_
=
_dt_
_k = 1..N_ (17)
10
-----
_uB,net_
_uk,1iL,k,1_ _uk n_ CB,net RB,net
vk,1
_uB,k_
RL,k,m Lk,m,k,m
_uk,miL,k,m_ CB,k RB,k
vk,m
Figure 4: Network of k microgrids connected to a common bus. Microgrid 1 has i source
converters and microgrid k has m source converters .
_Lk→net_ _diL,kdt→net_ = −λk→netvB,net − _iL,k→netRL,k→net_
+ uk→net + vB,k,
_k = 1..N_ (18)
_N_
_dvB,net_ �
_CB,net_ _dt_ = _λi→netiL,i→net_
_i=1_
_−_ _[v][B,net]_ + uB,net. (19)
_RB,net_
11
-----
**3. Feed-forward duty cycle and feed-back energy stortage control**
For the control development consider the microgrid structure in Fig. 4
with 2 sub-microgrids (N = 2) and both microgrids have 2 boost converter
sources (M = 2). Then the dynamic system equations (16) - (19) will have
_N_ (2 + M ) + 1 = 9 states. In this paper, the duty cycles, λ, will only be
updated on the interval ∆tλ, and held constant otherwise. This zero-order
hold approach will mimic the effect of a discrete digital control system that
may have limited computational power. Then, by using the zero-order hold
approach, the effects of a digital control and information flow can be studied
on the performance of the system. In this paper, it is desired to study a
distributed control that will only require local information at the converter
to determine the control action. However, a centralized approach based on
system wide information can also be used [3, 9, 6]. The updated duty cycle
commands are obtained from the steady-state solution of (16) and (18), with
_u = 0, such that_
1
_λk,m =_ (vk,m + iL,k,mRL,k,m),
_vB,k_
_k = 1..N,_ _m = 1..M_ (20)
1
_λk→net =_ _vB,net_ (vB,k + iL,k→netRL,k→net),
_k = 1..N._ (21)
The duty cycle update strategy for the source converters in (20) and the inter
face converter in (21) match the high side terminal voltage of the converter,
the bus voltage, to the low side source voltage minus the energy storage
12
-----
voltage. This forces the energy storage device back to a zero output condi
tion. The feed-forward control in (20) and (21) were chosen to remove the
current load from the converters energy storage device. Between updates of
the feed-forward duty cycle command the energy storage device actuates and
enforces the reference command values. Since the boost converter duty cycle
held constant between feed-forward updates, and not through feed-back, the
problem of non-minimum phase in boost converters is eliminated [18]. To
model the system in more general terms, a change of notation for the states
is
(x11, x12, x21, x22) = (iL,1,1, iL,1,2, iL,2,1, iL,2,2)
(x13, x23) = (iL,1→net, iL,2→net)
(x1, x2, x3) = (vB,1, vB,2, vB,net).
Then the state equations can be written in the compact form
(22)
�
**M˙x = Rx + u + v =**
�
**R +** **R[˜]** **x + u + v** (23)
where
**x = [x11, x12, x21, x22, x13, x23, x1, x2, x3][T]** (24)
**v = [v11, v12, v21, v22, 0, 0, 0, 0, 0][T]** (25)
**u = [u11, u12, u21, u22, u1→3, u2→3, uB1, uB2, uB3][T]** _._ (26)
The matrices from (23) for the example in Fig. 4 are
13
-----
_−RL,11_ 0 0 0 0 0 0 0 0
0 _−RL,12_ 0 0 0 0 0 0 0
0 0 _−RL,21_ 0 0 0 0 0 0
0 0 0 _−RL,22_ 0 0 0 0 0
0 0 0 0 _−RL,13_ 0 0 0 0
0 0 0 0 0 _−RL,23_ 0 0 0
0 0 0 0 0 0 1 0 0
_−_ _RB1_
0 0 0 0 0 0 0 1 0
_−_ _RB2_
0 0 0 0 0 0 0 0 1
_−_ _RB3_
(28)
**R =**
(29)
(27)
0 0 0 0 0 0 _−λ11_ 0 0
0 0 0 0 0 0 _−λ12_ 0 0
0 0 0 0 0 0 0 _−λ21_ 0
0 0 0 0 0 0 0 _−λ22_ 0
0 0 0 0 0 0 0 0 _−λ13_
0 0 0 0 0 0 0 0 _−λ23_
_λ11_ _λ12_ 0 0 0 0 0 0 0
0 0 _λ21_ _λ22_ 0 0 0 0 0
0 0 0 0 _λ13_ _λ23_ 0 0 0
**˜R =**
_L11_
_L12_
_L21_ 0
_L22_
_L13_
_L23_
0 _CB1_
_CB2_
_CB3_
**M =**
and the resistive matrices are shown in (28) and (29).
14
-----
_3.1. Hamiltonian surface shaping power flow control_
A key element of the proposed distributed droop control is a PI con
trol strategy based on a Hamiltonian Surface Shaping Power Flow Control
(HSSPFC) approach [19]. The first step is to define an error state for (23)
**˜x = xref −** **x** (30)
where the reference state and control vectors are defined by
�
**M˙xref =**
�
**R +** **R[˜]** **xref + uref + v.** (31)
The reference vector xref is the set of nominal state variables. Some of the
nominal values, such as the nominal bus voltage is set by the system designer.
The other reference states, such as the converter currents, are set through
the droop control calculations as will be discussed in section 4. The next
step is to define the Hamiltonian as
**H = [1]**
2 **[˜x][T]** **[M˜x][ + 1]2**
�� _t_ �
**˜x[T]** _dτ_ **KI**
0
15
�� _t_ �
**˜xdτ** (32)
0
-----
�
which is positive definite about **˜x = 0 for M and K positive definite and**
is the static stability condition. The time derivative of (32) is
**˙H = ˜x[T]** **M˜˙x + ˜x[T]** **KI**
� _t_
**˜xdτ**
0
� _t_
= ˜x[T] [M˙xref − **M˙x] + ˜x[T]** **KI** **˜xdτ**
0
� � � �
= ˜x[T][ ��]R + **R[˜]** **xref + uref + v −** **R +** **R[˜]** **x −** **u −** **v**
+ ˜x[T] **KI**
� _t_
**˜xdτ**
0
� _t_
�
= ˜x[T][ �]R + **R[˜]** **˜x + ˜x[T]** (uref − **u) + ˜x[T]** **KI** **˜xdτ**
0
� _t_
= ˜x[T] **R˜x + ˜x[T]** **∆u + ˜x[T]** **KI** **˜xdτ** (33)
0
since ˜x[T][ ˜]R˜x = 0.
Now, select a proportional-integral (PI) controller as
**∆u = −KP˜x −** **KI**
� _t_
**˜xdτ** (34)
0
which gives
and
**u = uref −** **∆u** (35)
�
**˙H = −˜x[T][ �]KP −** **R** **˜x < 0** (36)
where KP and KI are positive definite controller gain matrices. Equation
(36) enables a guideline for picking the gains of the PI controller to maintain
16
-----
stability and performance. For (36), only **H[˙]** (˜x = 0) = 0. However, this
only proves stability for the state variables ˜x. Since the control dynamics
� _t_
of
0 **[˜x][dτ][ are not included, further analysis is needed to prove asymptotic]**
stability.
_3.2. Feed-back control dynamics stability_
The stability of the feed-back PI control dynamics can be found from the
higher order derivatives of the Hamiltonian in (32) [20].
**Theorem Assume there exists a Lyapunov function V (x) of the dynamical**
_system ˙x = f_ (x). Let Ω _be a non-empty set of the state vectors such that_
**x** Ω _V (x) = 0._ (37)
_∈_ _⇒_ [˙]
_If the first k_ 1 derivatives of V (x), evaluated on the set Ω, are zero
_−_
_d[i]v(x)_
= 0 **x** Ω _i = 1, 2, ..., k_ 1 (38)
_∀_ _∈_ _−_
_dx[i]_
_and the k-th derivative is negative definite on the set Ω_
_dV (x)_
_< 0_ **x** Ω (39)
_∀_ _∈_
_dx_
_then the system x(t) is asymptotically stable, if k is an odd number._
The feed-forward and feed-back PI control law is
**u = uref −** **∆u = M˙xref −** **Rxref −** **v + KP˜x + KI**
17
� _t_
**˜xdτ** (40)
0
-----
where uref is found from the solution of (31). Then the system state trajec
tories are
**M˙x = Rx + v + u**
= Rx + v + M˙xref − **Rxref −** **v + KP˜x + KI**
The deviation in the state trajectories are
� _t_
**˜xdτ.** (41)
0
**M˜x[˙]** = [R − **KP] ˜x −** **KI**
� _t_
**˜x[T]** _dτ_ (42)
0
then
�
**˙˜x = M[−][1]** (R − **KP) ˜x −** **KI**
� _t_ �
**˜xdτ** _._ (43)
0
The second time derivative of the Hamiltonian is
�
**¨H = −2˜x[T][ �]KP −** **R** _x˙˜_
�
�
= −2˜x[T][ �]KP − **R** **M[−][1]** (R − **KP) ˜x −** **KI**
� _t_ �
**˜xdτ**
0
= 0 _for_ **˜x = 0.** (44)
18
-----
The third oder time derivative of the Hamiltonian from (32) is
... _T_ � � _t_ �
**H = −2˜x[˙]** [�]KP − **R�** **M[−][1]** (R − **KP) ˜x −** **KI** **˜xdτ**
0
�
�
_−_ 2˜x[T][ �]KP − **R** **M[−][1][ �](R −** **KP)** **˜x[˙]** _−_ **KI˜x**
� � � _t_ ��T
� �
= −2 **M[−][1]** (R − **KP) ˜x −** **KI** **˜xdτ** **KP −** **R** **M[−][1]**
0
� � _t_ �
(R − **KP) ˜x −** **KI** **˜xdτ**
0
�
_−_ 2˜x[T][ �]KP − **R** **M[−][1]**
� �
(R − **KP) M[−][1]** (R − **KP) ˜x −** **KI**
� _t_ � �
**˜xdτ** _−_ **KI˜x**
0
�
= −2 **M[−][1]KI**
� _t_ �T � � [�]
**˜xdτ** **KP −** **R** **M[−][1]KI**
0
� _t_ �
**˜xdτ** _< 0_ (45)
0
and
... � _t_
**H = 0** _for_ **˜xdτ = 0.** (46)
0
Therefore the control dynamics are asymptotically stable. Furthermore proof
of control dynamic stability can also be found in [21, 20].
**4. Droop control**
When two or more sources inject power into a common bus or grid a load
sharing control scheme is required. If each source tries to control the bus volt
age, or frequency in ac systems, then large circulation currents can result. A
load sharing scheme can be centralized if an interconnected communication
system is present and signals can be passed between sources to balance the
19
-----
load current. However, this can create a single point of failure. Droop control
is a common technique for distributed control of electrical sources in a mi
crogrid where the control implements a virtual impedance such that the load
current is distributed between the sources proportional to the droop settings
_Vref,i and Rd,i [14]. The equivalent boost converter under droop control is_
shown in Fig. 5, where Vref,i and Rd,i are parameter settings for the control.
To implement the control in a source, and to control the bus voltages
through the actuation of the energy storage devices, the reference for the bus
current injection in the i[th] converter in Fig. 3, is defined as
_iref,i =_ _[V][ref,i][ −]_ _[v][B]_ _._ (47)
_Rd,i_
An error signal is created from the reference inductor current as
_ei = iref,i_ _ii._ (48)
_−_
A control law to drive the error of (48) to zero is needed. However, it is
important to point out that the control inputs to this model are the energy
storage devices ui for the source converters and uB,k for the bus storage
elements. This droop controller implements a decentralized version of the
centralized feed-forward control of section 3.1 and utilizes the decoupled feed
back control. The droop control will be actuated by the storage elements
while the duty cycles λi will be considered as a constant during a discrete
epoch. The duty cycles will then be updated at set intervals as discussed in
section 3.
20
-----
For the source converter the current injected into the bus is
_ii = iL,iλi_ (49)
with the error signal for the boost converters as
_ei = iref,i_ _iL,iλi._ (50)
_−_
A proportional-integral control law for the droop control actuated by the
energy storage device is
�
_uk,m = RL,k,miref,k,m + λivB,k_ _vk,m +_ _Kp,iei + Ki,i_
_−_
� �
_eidt_ (51)
where uk,m ∈ **u from (40). The droop control law for the bus storage devices**
is
_uB,j =_ _[V][ref,i][ −]_ _[v][B,k]_ _._ (52)
_Rd,i_
The power from the boost storage devices
_pk,m = uk,miL,k,m._ (53)
The power from the bus storage devices
_pB,j = uB,jvB,k._ (54)
21
-----
Figure 5: Equivalent terminal characteristics of a boost converter in a dc microgrid under
voltage droop control.
Figure 6: Equivalent terminal characteristics the interconnected bus boost converter.
The energy supplied from the storage device is
�
_wu =_ _pust._ (55)
The boost converters that interconnect the buses in the microgrid also
need a distributed control law to maintain bus voltages. Droop control can
also be applied to these converter, however, the voltage must be references
to the lower voltage bus. The terminal characteristics of the interconnecting
converter are shown in Fig. 6 where vB,n > vB,k.
22
-----
**5. Simulation examples**
To demonstrate the distributed droop control approach to networked dc
microgrids, a model was built and simulated in Wolfram Mathematica, Wol
fram SystemModeler and Modelica [22]. The system shown in Fig. 4 has 2
sub-microgrids (N = 2) and both microgrids have 2 boost converter sources
(M = 2). The sub-microgrids 1 and 2 have a nominal bus voltage of 100 V
and the network bus 3 has a nominal voltage of 200 V . All voltage sources
(v11, v12, v21, v22) have an average dc voltage of 48 V. However, all these
voltages also have uniform random white noise sampled on 1 second intervals
superimposed on the dc voltage. This random noise input tests the perfor
mance of the proposed controller under non-ideal conditions that would be
an extreme worst-case scenario for a field deployed microgrid. In microgrid
applications the renewable sources are stochastic, but would be better be
haved with lower bandwidth and magnitude variations in the transients than
what is tested in this paper [23]. Therefore, the simulations shown in this sec
tion demonstrate the proposed control is a viable solution for field deployed
microgrids.
The sources in microgrid 1 represent well behaved devices and are rep
resentative of dispachable sources such as diesel generators with very little
voltage variation, in this example less than 2 V. The sources in microgrid 2
represent highly variable sources and are representative of renewable sources
such as photovoltaic and wind with large voltage variation, in this example
up to 45 V. The resistive loads on each sub-microgrid are RB,1 = RB,2 = 10 Ω
and the load on the network bus is RB,3 = 100 Ω. The droop voltage settings
_Vref,i have been set to the respective bus nominal voltages. The boost con-_
23
-----
verter droop resistances are set to Rd,i = 0.5 Ω, and the bus storage droop
resistances are 2 Ω. The PI control gains were chosen to meet the require
ments defined in section 3 and were picked to be KP = 50, KI = 200 for all
controllers in the system, which were chosen to satisfy (36).
The structure and parameters of this example system were chosen to be
indicative of many applications of networked microgrids. Microgrid applica
tions such as military forward operating bases and electric ships have a zone
based architecture [24]. Each zone of a microgrid can be a self contained
power distribution microgrid of sources, loads and storage. Another way to
view the zonal microgrid is as a network of smaller microgrids [25]. However,
by interconnecting the zones or networked microgrids, greater reliability and
potentially lower energy storage requirements can be achieved.
_5.1. Constant load example_
A short 10 s simulation example was run with constant loads and random
voltage sources. The results from microgrid 2 are shown in Figs. 7-12. The
random white noise voltage sources are shown in Fig. 7. The boost energy
storage device voltages are shown in Fig. 8 with the duty cycles shown in
Fig. 9. The boost energy storage device current and powers are shown in
Fig. 10 and Fig. 11 respectively. The resulting bus 2 voltage is seen in Fig. 12.
The rate at which all duty cycles are updated in this simulation example
is 0.1 s. As can be seen in Figs. 7-12, the bus voltages quickly return to
their reference values set through the droop control when the source voltage
variations occur.
It is seen in Fig. 12 that the bus voltage has an approximate average of
97.487 V which is less than the nominal 100 V due to droop control. It is also
24
-----
90
80
70
60
50
40 v21
v22
30
0 2 4 6 8 10
|Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|
|---|---|---|---|---|---|---|---|---|---|
|||||||||||
|||||||||||
|||||||||||
|||||||||||
|||||||||||
|v 21||||||||||
|v 22||||||||||
_t_ s
H L
Figure 7: Boost converter source voltages.
0.5
0.0
-0.5
-1.0
|Col1|Col2|Col3|Col4|Col5|
|---|---|---|---|---|
||||||
|u 21|||||
|u 22|||||
0 2 4 6 8 10
_t_ s
H L
Figure 8: Boost converter storage device voltages.
seen that the bus voltage varies little when the source voltages vary because
the boost converter energy storage devices act to maintain the desired droop
characteristics. When the duty cycle is updated every 0.1 s, the energy
storage device voltage returns to zero along with the power output.
_5.2. Step change in load example_
To demonstrate the distributed droop control in networked microgrids,
the same model from the previous section was modified such that the resistive
25
-----
0.9
0.8
0.7
0.6
0.5
0.4
0.3
|Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|
|---|---|---|---|---|---|---|---|---|---|
|||||||||||
|||||||||||
|||||||||||
|||||||||||
|l 21||||||||||
|l 22||||||||||
|||||||||||
0 2 4 6 8 10
_t_ s
H L
Figure 9: Boost converter duty cycles.
load on bus 2 is
10 Ω, if t < 10 s
_._ (56)
1 Ω, if t 10 s
_≥_
_RB,2(t) =_
The step change in load from (56) causes a transient in the system which
draws more current from the sources. Through the droop control, the extra
load current is shared between the boost converters and energy bus storage
devices. The amount of current each device contributes is negotiated through
the bus voltages by means of the droop control strategy. The following
examples will show the performance of the system when each bus is isolated
versus when they are interconnected.
_5.2.1. Isolated microgrids_
In the first simulation with the step change in load at bus 2, each bus is
isolated from each other. This is accomplished through a very large droop
setting of the interconnecting converters. In this case, no energy is shared
between buses. The resulting sub-microgrid bus voltages are seen in Fig. 13,
where the bus 2 voltage droops from 97.5 V down to 81.6 V, while the bus 1
26
-----
voltage remains constant.
The energy from the local energy storage devices in microgrid 1 are shown
in Fig. 14. The energy from the local energy storage devices in microgrid 2
are shown in Fig. 15. In addition, Fig. 16 shows that bus storage device 2
must output energy when the load steps. It is also seen in Fig. 16 that the
bus storage device 3 is always providing energy to bus 3, since in the isolated
configuration, it is the only source of energy.
_5.2.2. Networked microgrids_
Next, the interconnecting converters were controlled to maintain droop
resistances of 0.5 Ω, and thus allowing energy exchange between buses. It is
seen in Fig. 17 and Fig. 18 that all three bus voltages droop in response to
the increased load. It is important to point out that the change in the bus 2
voltage is now only approximately 85.6 V, or 4 V less than in the previous
isolated case.
The energy from the local boost energy storage devices in microgrid 1 and
2 are shown in Fig. 19 and Fig. 20 respectively. The energy from the bus stor
age devices is shown in Fig. 21. The energy from the interconnecting boost
converters is shown in Fig. 22. It is seen in Fig. 22 that the storage devices
in the interconnection converters contribute nor draw any significant amount
energy, at least in comparison to the source and bus storage devices. This
indicates that in the interconnection storage device may not be necessary in
future iterations of a network microgrid design.
The performance and energy storage requirements of the system are
greatly affected by the networked interconnection. Table 1 provides the a
comparison of the total energy used as well as the maximum bus voltage
27
-----
Table 1: Total Energy and Voltage Droop Comparison
**Networked** **Isolated**
� _Wu (kJ)_ 79.75 119.68
max ∆VB 12.36 16.42
avg ∆VB 6.35 5.56
droop and average voltage droop from the networked and isolated simula
tion examples provided. Table 1 shows that in the networked configuration
a total of only 79.75 kJ of energy was needed, while in the isolated case
119.68 kJ was needed. It is also seen in Table 1 that the average voltage
droop changes very little between cases. These results lead to the conclusion
if this system were to be built, smaller energy storage devices can be used,
if the sub-microgrids can be networked together.
_5.2.3. Duty cycle feed-forward update rates_
Lastly, a series of simulations were performed to determine the effect of
duty cycle update rate on the total energy required. In these simulations the
models from section 5.2.1 and 5.2.2 was swept with duty cycle update rates
of ∆tλ = 0.1 s to 1 s at steps of 0.001 s. The total energy supplied by all the
energy storage devices in the system for the networked and isolated cases are
shown in Fig. 23. The results in Fig. 23 show that for small update rates,
the energy required by the isolated case are greater than the networked case.
As the time between updates increases, the isolated grid storage utilization
approaches that of the networked grid. A possible explanation is that as the
update period increases, the advantages of the interconnectivity degrade and
the grid operates as a set of isolated microgrids. It’s interesting to note that
for both the interconnected and isolated cases the storage usage decreases as
28
-----
the duty cycle update increases. Based on the results of section 5.2.1 and
5.2.2 the bus voltage droop for the interconnected topology was lower than
the grid’s isolated operation. Extending that result to the update period
study, the decrease in storage usage for longer update periods comes with a
reduction in bus voltage performance.
**6. Conclusions**
This paper has presented a novel approach to droop control actuated by
the local energy storage device. The novelty of this approach lies in the
actuation of the system through the energy storage devices. The duty cycles
are updated at periodic intervals through feed-forward control to match the
high side and low side source voltages. This approach allows the storage
needs for capacity and bandwidth response to be identified for the given
distributed droop control approach. The results show that the networked
microgrids need overall less energy storage in response to transients at certain
duty cycle update rates. However, as the update rates decrease, the storage
requirements lessen and become the same for a networked system versus
isolated systems.
**7. Acknowledgment**
Sandia National Laboratories is a multiprogram laboratory operated by
Sandia Corporation, a Lockheed Martin Company, for the U.S. Department
of Energy’s National Nuclear Security Administration under contract DE
AC04-94AL85000.
29
-----
**References**
[1] R. H. Lasseter, Microgrids, in: IEEE Power Engineering Society Winter
Meeting, Vol. 1, IEEE, 2002, pp. 305–308.
[2] K. A. Nigim, W.-J. Lee, Micro grid integration opportunities and chal
lenges, in: IEEE Power Engineering Society General Meeting, IEEE,
2007, pp. 1–6.
[3] E. Trinklein, G. Parker, W. Weaver, R. Robinett, L. Babe, C.-W. Ten,
W. Bower, S. Glover, S. Bukowski, Scoping study: Networked micro
grids, Tech. Rep. Sandia Report SAND2014-17718, Sandia National
Labs (2014).
[4] V. Ramirez, R. Ortega, O. Bethoux, A. S´anchez-Squella, A dynamic
router for microgrid applications: Theory and experimental results, Con
trol Engineering Practice 27 (0) (2014) 23 – 31.
[5] M. Erol-Kantarci, B. Kantarci, H. Mouftah, Reliable overlay topology
design for the smart microgrid network, IEEE Network 25 (5) (2011)
38–43.
[6] R. D. Robinett, D. G. Wilson, S. Y. Goldsmith, Collective control of
networked microgrids with high penetration of variable resources part
i: Theory, in: IEEE International Conference on Cyber Technology in
Automation, Control, and Intelligent Systems, 2012, pp. 1–4.
[7] A. Chaouach, R. M. Kamel, R. Andoulsi, K. Nagasaka, Multiobjective
intelligent energy management for a microgrid, IEEE Transactions on
Industrial Electronics 60 (2013) 1688–1699.
30
-----
[8] M. H. Nazari, M. Ilic, J. Lopes, Small-signal stability and decentralized
control design for electric energy systems with a large penetration of
distributed generators, Control Engineering Practice 20 (9) (2012) 823–
831.
[9] D. G. Wilson, R. D. Robinett, S. Y. Goldsmith, Renewable energy mi
crogrid control with energy storage integration, in: International Sympo
sium on Power Electronics, Electrical Drives, Automation and Motion,
2012, pp. 158–163.
[10] B. K. Johnson, R. H. Lasseter, F. L. Alvarado, R. Adapa, Expandable
multiterminal dc systems based on voltage droop, IEEE Transactions
on Power Delivery 8 (4) (1993) 1926–1932.
[11] A. Nagliero, R. Mastromauro, D. Ricchiuto, M. Liserre, M. Nitti, Gain
scheduling-based droop control for universal operation of small wind
turbine systems, in: IEEE International Symposium on Industrial Elec
tronics, IEEE, 2011, pp. 1459–1464.
[12] J. Guerrero, P. C. Loh, T.-L. Lee, M. Chandorkar, Advanced control
architectures for intelligent microgrids - part ii: Power quality, energy
storage, and ac/dc microgrids, IEEE Transactions on Industrial Elec
tronics 60 (4) (2013) 1263–1270.
[13] A. M. Dizqah, A. Maheri, K. Busawon, P. Fritzson, Standalone dc mi
crogrids as complementarity dynamical systems: Modeling and applica
tions, Control Engineering Practice 35 (0) (2015) 102 – 112.
31
-----
[14] W. W. Weaver, R. D. Robinett, G. G. Parker, D. G. Wilson, Energy
storage requirements of dc microgrids with high penetration renewables
under droop control, International Journal of Electrical Power & Energy
Systems 68 (0) (2015) 203 – 209.
[15] N. Mohan, T. Undeland, Power Electronics: Converters, Applications,
and Design, John Wiley & Sons, 2007.
[16] P. T. Krein, J. Bentsman, R. M. Bass, B. L. Lesieutre, On the use of av
eraging for the analysis of power electronic systems, IEEE Transactions
on Power Electronics 5 (2) (1990) 182–190.
[17] W. W. Weaver, P. T. Krein, Game-theoretic control of small-scale power
systems, IEEE Transactions on Power Delivery 24 (3) (2009) 1560–1567.
[18] Z. Chen, W. Gao, J. Hu, X. Ye, Closed-loop analysis and cascade control
of a nonminimum phase boost converter, IEEE Transactions on Power
Electronics 26 (4) (2011) 1237–1252.
[19] R. D. Robinett III, D. G. Wilson, Nonlinear Power Flow Control Design:
Utilizing Exergy, Entropy, Static and Dynamic Stability, and Lyapunov
Analysis, Springer, 2011.
[20] H. Schaub, J. Junkins, Analytical mechanics of space systems, Aiaa,
2003.
[21] R. Robinett, G.G., Parker, H. Schaub, J. Junkins, Lyapunov optimal
saturated control for nonlinear systems, Journal of Guidance, Control
and Dynamics 20 (6) (1997) 1083–1088.
32
-----
[22] P. Fritzson, Introduction to modeling and simulation of technical and
physical systems with Modelica, Wiley. com, 2011.
[23] A possibilistic probabilistic tool for evaluating the impact of stochastic
renewable and controllable power generation on energy losses in distribu
tion networks, a case study, Renewable and Sustainable Energy Reviews
15 (1) (2011) 794 – 800.
[24] D. Wilson, J. Neely, M. Cook, S. Glover, J. Young, R. Robinett, Hamil
tonian control design for dc microgrids with stochastic sources and loads
with applications, in: IEEE International Symposium on Power Elec
tronics, Electrical Drives, Automation and Motion, 2014, pp. 1264–1271.
[25] R. Robinett, D. Wilson, S. Goldsmith, Collective control of networked
microgrids with high penetration of variable resources part i: Theory,
in: IEEE International Conference on Cyber Technology in Automation,
Control, and Intelligent Systems, 2012, pp. 1–4.
33
-----
20
18
16
14
12
10
8
6
|Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|Col11|
|---|---|---|---|---|---|---|---|---|---|---|
||||||||||||
||||||||||||
||u 21||||||||||
|u 22|||||||||||
||||||||||||
||||||||||||
||||||||||||
||||||||||||
0 2 4 6 8 10
_t_ s
H L
Figure 10: Boost converter storage device currents.
10
5
0
-5
-10
-15
|Col1|Col2|Col3|Col4|Col5|Col6|
|---|---|---|---|---|---|
|||||||
|||||||
|||||||
|u 2|1|||||
|u 2|2|||||
|||||||
0 2 4 6 8 10
_t_ s
H L
Figure 11: Boost converter storage device power.
97.500
97.495
97.490
97.485
97.480
0 2 4 6 8 10
_t_ s
H L
Figure 12: Bus 2 voltage.
34
-----
100
95
90
85
80
|Col1|Col2|Col3|Col4|Col5|Col6|
|---|---|---|---|---|---|
|v v|B1 B2|||||
|||||||
|||||||
0 5 10 15 20 25 30
_t_ s
� �
Figure 13: Bus 1 and 2 voltages with step change in load on bus 2 in isolated microgrid.
0.006
0.004
0.002
0.000
�0.002
�0.004
0 5 10 15 20 25 30
|Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|Col11|Col12|Col13|Col14|Col15|Col16|Col17|Col18|Col19|Col20|
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
|u|||11|||||||||||||||||
|u|||12|||||||||||||||||
|||||||||||||||||||||
|||||||||||||||||||||
|||||||||||||||||||||
|||||||||||||||||||||
_t_ s
� �
Figure 14: Energy supplied in boost devices storage u1,j in isolated microgrid.
80000
60000
40000
20000
0
0 5 10 15 20 25 30
|Col1|Col2|Col3|Col4|Col5|Col6|
|---|---|---|---|---|---|
|u|21|||||
|u|22|||||
|||||||
|||||||
_t_ s
� �
Figure 15: Energy supplied in boost devices storage u2,j in isolated microgrid.
35
-----
12000
10000
8000
6000
4000
2000
0
|u|Col2|Col3|Col4|Col5|Col6|
|---|---|---|---|---|---|
|u|1 2|||||
|u|3|||||
|||||||
|||||||
|||||||
|||||||
0 5 10 15 20 25 30
_t_ s
� �
Figure 16: Energy from the bus storage devices in isolated microgrid.
197
196
195
194
193
192
191
0 5 10 15 20 25 30
_t_ s
� �
Figure 17: Bus 3 voltage with step change in load on bus 2 in networked microgrid.
100
95
90
85
80
|Col1|Col2|Col3|Col4|Col5|Col6|
|---|---|---|---|---|---|
|v v|B1 B2|||||
|||||||
|||||||
0 5 10 15 20 25 30
_t_ s
� �
Figure 18: Bus 1 and 2 voltages with step change in load on bus 2 in networked microgrid.
36
-----
0.12
0.10
0.08
0.06
0.04
0.02
0.00
�0.02
0 5 10 15 20 25 30
|Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|Col11|Col12|Col13|
|---|---|---|---|---|---|---|---|---|---|---|---|---|
||u 11||||||||||||
||u 12||||||||||||
||||||||||||||
||||||||||||||
||||||||||||||
||||||||||||||
||||||||||||||
_t_ s
� �
Figure 19: Energy supplied in boost devices storage u1,j in networked microgrid.
40000
30000
20000
10000
0
|Col1|Col2|Col3|Col4|Col5|Col6|Col7|
|---|---|---|---|---|---|---|
||u 21||||||
||u 21||||||
||u 22||||||
||||||||
||||||||
0 5 10 15 20 25 30
_t_ s
� �
Figure 20: Energy supplied in boost devices storage u2,j in networked microgrid.
12000
10000
8000
6000
4000
2000
0
|Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|
|---|---|---|---|---|---|---|---|
|||u 1 u||||||
|||2 u 3||||||
|||||||||
|||||||||
|||||||||
|||||||||
0 5 10 15 20 25 30
_t_ s
� �
Figure 21: Energy from the bus storage devices in networked microgrid.
37
-----
0
�50
�100
�150
0 5 10 15 20 25 30
_t_ s
� �
Figure 22: Energy bus interface converters storage u1→3 and u2→3 in networked microgrid.
100
80
60
40
20
0
0.0 0.2 0.4 0.6 0.8 1.0
Dtl HsL
Figure 23: Total energy supplied by enegy storage devices as a function of duty cycle
feed-forward update rate.
38
-----
|
{
"disclaimer": "Notice: The following paper fields have been elided by the publisher: {'abstract', 'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1016/J.CONENGPRAC.2015.06.008?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1016/J.CONENGPRAC.2015.06.008, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "publisher-specific-oa",
"status": "BRONZE",
"url": "https://www.sciencedirect.com/science/article/am/pii/S0967066115001185"
}
| 2,015
|
[] | true
| 2015-11-01T00:00:00
|
[] | 14,721
|
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Business",
"source": "s2-fos-model"
},
{
"category": "Environmental Science",
"source": "s2-fos-model"
},
{
"category": "Engineering",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/0246366d638b26f74404861edb9f4071bee082bb
|
[
"Computer Science"
] | 0.882219
|
Moving from 5G in Verticals to Sustainable 6G: Business, Regulatory and Technical Research Prospects
|
0246366d638b26f74404861edb9f4071bee082bb
|
International Conference on Cognitive Radio Oriented Wireless Networks and Communications
|
[
{
"authorId": "1381676255",
"name": "Marja Matinmikko-Blue"
},
{
"authorId": "2741612",
"name": "Seppo Yrjölä"
},
{
"authorId": "2120804",
"name": "Petri Ahokangas"
}
] |
{
"alternate_issns": [
"2166-5370"
],
"alternate_names": [
"CROWNCOM",
"Int Conf Cogn Radio Oriented Wirel Netw Commun",
"CROWNC"
],
"alternate_urls": null,
"id": "464a5ecb-f65b-4e71-a90d-3310a7d6ea8e",
"issn": "2166-5419",
"name": "International Conference on Cognitive Radio Oriented Wireless Networks and Communications",
"type": "conference",
"url": "http://ieeexplore.ieee.org/xpl/conhome.jsp?punumber=1001755"
}
| null |
# Moving from 5G in Verticals to Sustainable 6G: Business,
Regulatory and Technical Research Prospects
Marja Matinmikko-Blue[1[0000-0002-0094-6344]], Seppo Yrjölä[2[0000-0003-2053-9700]], and Petri
Ahokangas[3[0000-0002-2351-8473]]
1 Centre for Wireless Communications, University of Oulu, Finland
2 University of Oulu, Finland and Nokia, Finland
3 Oulu Business School, Martti Ahtisaari Institute, University of Oulu, Finland
```
marja.matinmikko@oulu.fi
```
**Abstract. Mobile communication research is increasingly addressing the use of**
5G in verticals, which has led to the emergence of local and often private 5G
networks. At the same time, research on 6G has started, with a bold goal of building a strong linkage between 6G and the United Nations Sustainable Development Goals (UN SDGs). Both of these developments call for a highly multi-disciplinary approach covering the inter-related perspectives of business, regulation
and technology. This paper summarizes recent advances in using 5G to serve
vertical sectors’ needs and describes a path towards sustainable 6G considering
business, regulation and technology viewpoints. By focusing on key trends, the
research summarizes four alternative scenarios for the futures business of 6G and
considers related regulatory and technology aspects. Our findings highlight the
importance of understanding the complex relations of business, regulation and
technology perspectives and the role of ecosystems in both 5G in verticals and
ultimately in the development of sustainable 6G to bring together stakeholders to
solve long-term sustainability problems.
**Keywords: Business Strategy, Regulation, Scenario Planning, Sustainability,**
5G, 6G.
## 1 Introduction
5G deployments are underway in the global scale with the first applications focusing
on offering high capacity mobile broadband services. The promise of 5G to boost the
digitalization of various vertical industries is gradually gaining increasing attention and
the emergence of local 5G networks [Matinmikko et al. 2017; Matinmikko et al. 2018]
is starting to take place in some countries. Local 5G networks allow different stakeholders to use their own local connectivity platforms without having to rely on mobile
network operators. These developments are occurring in complex multi-stakeholder
ecosystems where regulatory, business, and technical perspectives are highly intertwined. The emergence of using 5G in the various verticals brings together the ICT
sector and the vertical sector in question with their own structures and rules for operations, calling for an ecosystem-level focus. Especially, the availability of spectrum for
-----
2
local networks fully depends on the country of operations, emphasizing the importance
of regulatory decisions.
At the same time, research on the sixth generation (6G) of mobile communication
networks has started globally aiming at first deployments in the 2030s. The first 6G
White Paper published in 2019 presented a joint 6G research vision as a group work of
70 experts globally [Latva-aho & Leppänen, 2019]. The paper depicted the future 6G
networks as an intelligent system of systems that combines the communication services
with a set of other services including imaging, sensing, and locationing services, opening a myriad of new application areas. A set of continuation 6G White Papers published
in 2020 [6G Flagship White Papers 2020] prepared in collaboration with 250 international experts went more into details and presented e.g. alternative future scenarios for
the business of 6G [Yrjölä et al., 2020], and developed a tight linking between 6G and
the United Nations Sustainable Development Goals (UN SDGs) [Matinmikko-Blue et
al., 2020a]. Some of the developed future 6G business scenarios have taken sustainability as the starting point, stressing that the whole development of the future mobile
communication networks should aim at helping society at large in its attempts to meet
the sustainable development goals [Latva-aho & Leppänen 2019; Yrjölä et al. 2020a;
Yrjölä et al. 2020b; Matinmikko-Blue et al. 2020a].
To make sense of moving from 5G in verticals towards 6G, we must envision future
6G systems targeting 2030 holistically from the perspective of the interaction between
business, regulation and technology perspectives in envisioning future research prospects. The alternative futures of 6G will be shaped by growing societal requirements
like inclusivity, sustainability, resilience, and transparency – a highly complex area that
will call for major changes in industrialized societies in the long run, see [Latva-aho et
al. 2019; Matinmikko-Blue et al. 2020a]. The business perspective specifically needs
to consider sustainability [Kuhlman & Farrington, 2010; Evans et al. 2017] in a way
that combines the economic (e.g., profit, business stability, financial resilience, viability), societal (e.g., individuals’, communities’, regulative values) and environmental
(e.g., renewables, low emissions, low waste, biodiversity, pollution prevention) perspectives. As an emerging field, 6G business scenarios and strategies have not been
widely discussed in the literature to date. However, vision papers on future communication needs, enabling technologies, the role of artificial intelligence (AI), and emerging applications have recently been published [Viswanathan & Mogensen, 2020; Saad,
Bennis & Chen, 2019; Letaief et al. 2019]. Furthermore, discussion has latterly expanded to 6G indicators of value and performance [Ziegler & Yrjölä, 2020], the role of
regulation and spectrum sharing [Matinmikko-Blue et al., 2020a], the antecedents of
multi-sided transactional platforms [Yrjölä, 2020], antecedents of the 6G ecosystem
[Ahokangas et al. 2020a] and the exploratory scenarios of 6G business [Yrjölä et al.,
2020].
Building on the above discussion, this paper provides an overview of 5G in verticals
towards sustainable 6G from business, regulation and technology perspectives and presents related research prospects. The paper summarizes future scenarios for sustainable
6G business strategies in the timeframe 2030-2035, originally documented in [Yrjölä
et al., 2020], and related strategic options. The rest of this paper is organized as follows.
Chapter 2 summarizes the state of the art of 5G in verticals from business, regulation
-----
3
and technology perspectives. Chapter 3 presents an overview of sustainable 6G. Future
business scenarios for sustainable 6G and related strategic options are presented in
Chapter 4. Finally, future outlook and conclusions are provided in Chapter 5.
## 2 State of the Art of 5G in Verticals
5G has been set high in national agendas to speed up digitalization of various sectors
of society in many countries. This chapter presents recent developments in the use of
5G networks to serve the needs of different vertical sectors, such as industry, energy,
and health, and their public sector counterparts, from the interrelated business, technology and regulation perspectives.
**2.1** **Business Perspective**
Business perspective plays an important role in understanding the opportunities that a
new technology can offer. The identification of the opportunity space for 5G business
in verticals requires discussing four inter-related key themes: 1) the convergence of
connectivity and data platforms and related ecosystems, 2) enablers, barriers and limitations to scalability and replicability of 5G solutions and business models, 3) legitimation of the new roles and business models within the verticals, and the 4) economic,
societal and environmental sustainability of 5G solutions and business models. As vertical 5G networks are often considered as local networks, the platform-based business
models utilized by different stakeholders face several challenges related to the aforementioned themes.
Mobile communication networks have for long been seen as platforms [Pujol et al,
2016] or ecosystems [Basole and Karla, 2011]. However, with the deployment of 5G
networks, the mobile connectivity platforms operated by mobile network operators
(MNOs) are increasingly becoming converged with the data platforms of various cloud
service providers, giving rise to novel kinds of platform ecosystems. In industry verticals also the Industry 4.0 platforms as a specific type of data platforms play an important role. Extant literature identifies centralized, hybrid and fragmented types of
converged connectivity and data platforms for industry verticals [Ahokangas et al.,
2020b]. In this kind of vertical context, a key feature of the converged platforms is the
degree of openness achieved for different stakeholders of the ecosystem. Related to
openness, the complexity, complementarity and interdependence of the converged connectivity and data platforms can be clarified by looking at the various components, interfaces, data and algorithms utilized in these platforms [Yrjölä et al., 2019] in connection to the connectivity (5G or other), content (e.g., information or data), context (location- or use-case specific data) or commerce (offering made available via a platform)
business models utilized [Iivari, et al., 2020]. The vertical business model for local 5G
operators presented by [Ahokangas et al., 2019] builds specifically on the idea to provide tailored end-to-end services in restricted geographical areas, such as industry sites,
to the users locally. Vertical business models form a vertically structured ecosystem
around the activity. The presented oblique business model and corresponding oblique
-----
4
ecosystem in turn builds on mass-tailored end-to-end services with stricter requirement
for segmentation [Ahokangas et al., 2019].
The different types of converged connectivity and data platforms and the business
models identified for them have varying potential for scalability and replicability. A
scalable business model is agile and provides exponentially increasing returns to scale
in terms of growth from additional resources applied [Nielsen and Lund, 2018],
whereas a replicable business model can be copied to several markets simultaneously
with minimum variations [Aspara et al., 2010]. For a firm running a vertical business
model, scalability is based on the firm’s capability to understand customer-specific
needs and fulfill them, but limited on the size of the cases, their volume and timeline.
For a firm running an oblique business model, scalability is based on the volume of
unmet local needs and limited by access and availability of local infrastructures needed
for providing the service [Ahokangas et al., 2019].
Within converged connectivity and data platform ecosystems, different stakeholders
have varying roles and can act as service providers. This raises the issue of legitimacy,
meaning that the activities of the stakeholder providing the service is legal and fits with
the institutionalized practices within the industry in question [Marano et al., 2020].
Achieving legitimacy for local vertical-specific 5G services and service providers
through the deployment of local 5G networks is, however, an open question in many
countries. Indeed, disruptive innovations such as 5G have been found to cause regulatory, incumbent and social “pushbacks” and they can be expected also for vertical 5G
services, as legitimacy is a precondition for successful value creation and capture on a
technology [Biloslavo et al., 2020].
The above discussion points out several challenges for reaching sustainable business
models in 5G verticals. “A business model for sustainability helps describing, analyzing, managing and communicating (i) a company’s sustainable value proposition to its
customers, and all other stakeholders, (ii) how it creates and delivers this value, (iii)
and how it captures economic value while maintaining or regenerating natural, social,
and economic capital beyond its organizational boundaries” [Schaltegger et al., 2016,
p. 6]. Building vertical 5G business opportunities calls thus for filling in the requirements of scalability, replicability, and sustainability in a legitimate way in a platform
ecosystem comprising connectivity and data services.
**2.2** **Regulation Perspective**
The serving of the different verticals with 5G networks is not only addressed by the
current MNOs but increasing attention is being paid to local and often private 5G networks [Matinmikko et al. 2017; Matinmikko et al., 2018] that can be operated independent of the MNOs. Their emergence is highly dependent on the regulations that
govern both the electronic communications market as well as the specific verticals,
leading to a complex environment to operate. Regulations at national, regional and international levels define the operational conditions and there is wide variation between
the national approaches but also some level of harmonization such as on the spectrum
for 5G.
-----
5
Prior work on regulatory developments on local 5G networks [Matinmikko et al.,
2018; Vuojala et al., 2019; Lemstra, 2018; Ahokangas et al., 2020b] have considered
access regulation, pricing regulation, competition regulation, privacy and data protection, and authorization of networks and services. Especially, the authorization of networks and services defining the ways how rights to use radio frequencies are granted is
critical for the establishment of local private 5G networks. Without the timely availability of sufficient amount of spectrum suitable for operations in the given environment,
it is not possible to deploy the local networks. Specific spectrum options for local 5G
networks are analyzed in detail in [Vuojala et al., 2019] including unlicensed access,
secondary licensing, spectrum trading/leasing, virtual network or local licensing. Local
licensing has emerged as a new spectrum access model in 5G to allow different stakeholders to deploy local networks in addition to the MNOs. A study on the recent 5G
spectrum awards decisions in the 3.5 GHz band presented in [Matinmikko-Blue et al.,
2019] shows that there is a big divergence in the spectrum awards by different countries
taken by the regulators globally.
5G regulatory situation in Europe is discussed in [Lemstra, 2018] where two con
trasting scenarios for the future telecommunication market are presented including evolutionary and revolutionary scenarios. Evolutionary scenario continues the MNO market dominance which is likely to occur under the current European regulatory framework. The revolutionary scenario introduces new virtual MNOs that serve specific industry sectors which calls for additional policy and regulatory measures. The mobile
communication market is in a turning point with the emergence of locally operated 5G
networks by different stakeholders, especially aiming at serving the verticals’ specialized local needs.
**2.3** **Technology Perspective**
Previous generation mobile technologies have been largely deployed by national (or
multi-national) incumbent MNOs for public use, given the high levels of investments
required for the infrastructure, and to acquire exclusive radio spectrum. Furthermore,
management and operational costs of the networks have been significant, and mobile
technologies have required large and complex system integration from global infrastructure vendors with specialized capabilities. In addition to improved performance
characteristics in capacity, speed and latency, novel 5G architecture is bringing additional flexibility for traditional MNOs as well as local operators in system deployments.
Key technologies expected to transform 5G for verticals include localization and decomposition of network functions, software defined networking and network virtualization among others [Morgado et al., 2018].
A critical aspect of the local private industrial 5G networks is the ability to create
customized network slices, where instances of virtual network resources and applications can be delivered to a new breed of services tailored to specific customer or tenant
needs with service level agreed performance on demand. Furthermore, the softwarebased network architecture enables efficient sharing of common network infrastructure
and resource by different tenants. Abstracting the slice functionality through open interfaces exposure to third party service provisioning enables service-dominant model
-----
6
for the connectivity and underlying network resources, e.g., computing, data and intelligence. The evolution towards the cloud-native infrastructure abstraction both on core
and radio access empowers technology vendors and service providers to deploy and
operate flexible and portable processes and applications in dynamic multi-vendor cloud
environments. The cloud embedded in the edge of the network provides tools for optimized performance and economics for both the virtualized network functions and any
other performance critical enterprise or vertical service and can become a control point
of the local connectivity and intelligence. Edge cloud use cases considered in 5G are
e.g., cloud radio access network (Open RAN, Virtual RAN), edge security, network
and service automation enhancing the network itself, and industrial automation, massive scale Internet of Things (IoT), and augmented intelligence with augmented reality
(AR)/virtual reality (VR). Another critical aspect is the spectrum. Operations in higher
carrier frequencies represent a challenge in terms of deployment. The availability of
suitable spectrum for serving the verticals cannot be based on dedicated spectrum paradigm but requires sharing in different domains.
Figure 1 summarizes the presented business, regulation and technology perspectives
for 5G in verticals.
**Fig. 1. Business, regulation and technology perspectives for 5G in verticals.**
## 3 Towards Sustainable 6G
In parallel with the on-going development and deployment of 5G in verticals, research
on the next generation, namely 6G, systems has already started in different parts of the
world, see [Latva-aho & Leppänen, 2019] and [6G Flagship White Papers, 2020]. The
research on 6G [Latva-aho & Leppänen, 2019; Matinmikko-Blue et al., 2020a] has
-----
7
identified sustainability stemming from the UN SDGs as the starting point, and it needs
to address the technical, business and regulation perspectives, which are discussed next.
**3.1** **Role of UN SDGs in 6G**
Future 6G networks are aiming at first deployments around the year 2030 which is
also the target year for the achievement of the UN SDGs. While 6G communications is
expected to boost global growth and productivity, create new business models and
transform many aspects of society, its linking with the UN SDGs needs to be clearly
formulated. The starting point of 6G research vision presented in [Latva-aho &
Leppänen 2019] is that the development of 6G should be fully aligned with the UN
SDGs [United Nations, 2018]. In a follow-up white paper, Matinmikko-Blue et al.
[2020a] have developed a novel linking between 6G and the UN SDGs through the
indicators of the UN SDG framework.
In [Matinmikko-Blue et al., 2020a] a three-fold role is foreseen for 6G as 1) provider
of services to help reaching the UN SDGs, 2) enabler of measuring tools for data collection to help with the reporting of indicators, and 3) reinforcer of a new ecosystem to
be developed in line with the UN SDGs. The white paper further details the linking
between 6G and UN SDGs trough the existing indicators of the UN SDG framework
where only 7 out of the 231 individual indicators are identified as being related to ICT.
In reality, the ICT sector can influence many of the indicators, if not all. The white
paper [Matinmikko-Blue et al., 2020a] analyses what 6G can do to contribute to the
different UN targets within the SDG framework via the existing UN SDG indicators.
The white paper proceeds to stating the need for a new set of indicators for 6G, characterizing the three-fold role of 6G. Additionally, a preliminary action plan is presented,
calling for research and educational organizations, governments, standards developers,
users, MNOs, network equipment manufacturers, application and service providers and
verticals to think out-of-the box and create new technology solutions and collaborative
business models to develop new operational models that support the achievement of the
SDGs which may need changes to the existing regulations.
**3.2** **Business, Regulation and Technology Perspectives**
The discussion on 5G business perspective for deployment in verticals presented in
Section 2.1 proposed to focus on business models as a way of thinking future 6G ecosystem stakeholders’ choices regarding opportunities, value-add and capabilities, and
their expected consequences as scalability, replicability, and sustainability. With the
right business choices, opportunities will be identified related to novel and unmet needs,
new types of customer and service provider, as well as the interfacing of humans with
machines in 6G. New value-add is seen to come from real-time and trustworthy communications, the use of local data and intelligence, and the commoditization of 6G resources as its competitive advantages, including extreme capacity and security, transaction and innovation platformization, and ubiquitous access. The expected business
consequences of scalability may be related to the long tail of services, dataflow architecture, automation, and open collaboration between stakeholders; in terms of replicability, to deliberately design modularity and complementarity within platforms; and in
-----
8
terms of sustainability, to empower users and communities, and the utilization of sharing economic mechanisms in the markets.
Overall, governments and industries are under high pressure from the sustainability
targets arising from the UN SDGs to renew their operations and the achievement of the
goals provides new business opportunities especially for ICT solutions. These data and
connectivity solutions can significantly contribute to industries to improve their resource efficiency and reduce waste but the solutions themselves need to be developed
in alignment with the sustainability goals as well.
Digital convergence across industries and multi-level 6G platforms and ecosystems
are creating a complex strategic environment that can lead to incomparable and distinct
opportunities, as well as emergent problems. The regulations governing the use of future telecommunication systems and the relevant industry specific regulations together
create a complex environment, especially around the use of data and connectivity platforms for different purposes. In particular, unanswered questions remain about ecosystemic business models in the context of sustainability. According to our recent findings
[Yrjölä et al., 2020a; Yrjölä et al., 2020b], business ecosystems that aim to bring together stakeholders to solve systemic sustainability problems will require open ecosystem-focused value configuration and decentralized power configuration, where traditional stakeholder roles change, and new roles emerge. The focus needs to be on the
long tail of specialized user requirements that crosses a variety of industries where related needs can be met with different resource configurations.
Spectrum continues to be the key resource for 6G systems as for any wireless net
works throughout the times, and the availability of suitable spectrum continues to be
significantly restricted due to the existing incumbent spectrum usage, see [MatinmikkoBlue et al. 2020b]. Spectrum availability is a good example of the complex relations of
business, regulation and technology perspectives. The availability of spectrum is a regulation decision, which defines the business opportunities and yet is restricted with
technical aspects. Potential operations of future 6G systems in the new higher frequency
bands at upper millimeterwaves (mmW) and terahertz (THz) regions pose significant
technical, regulatory and deployment related challenges. Therefore, future 6G is not
restricted only to higher frequency bands but can also be used in the existing bands for
mobile communications. What are the economically feasible operational models, how
to protect existing incumbent users of the feasible bands and how to implement THz
radio links continue to be open topics for 6G.
The technology vision work in the global scale for systems towards 2030 and beyond
has started at the International Telecommunication Union Radiocommunication sector
(ITU-R) with the development of a report on future technology trends. The need for
new indicators to characterize the performance of future 6G networks is evident [Latvaaho & Leppänen, 2019; Matinmikko-Blue et al., 2020a; Pouttu et al., 2020], especially
for defining and measuring resource efficiency and particularly energy efficiency. Also,
the network architecture of 6G needs to be re-thought from prior generations of networks, see [Taleb et al., 2020]. Figure 2 provides a summary emphasizing the need to
develop sustainable 6G in line with the UN SDGs from business, regulation and technology perspectives.
-----
9
**Fig. 2. Business, regulation and technology perspectives for sustainable development of 6G.**
## 4 Business Scenarios and Strategic Options for 6G
Next, we proceed to new business scenarios developed for 6G and related strategic options developed through a set of virtual future-oriented white paper expert group workshops organized by 6G Flagship at the University of Oulu in 2020 and documented and
analyzed in [Yrjölä et al., 2020a; Yrjölä et al., 2020b].
**4.1** **Methodology**
The alternative scenarios for the future business of 6G summarized in this paper were
created using anticipatory action learning (AAL) research method [Stevenson, 2012]
within 6G Flagship’s white paper preparation [6G Flagship 2020]. The process involved a series of online workshops in January-April 2020 where a group of experts
from research, standardization and development, telecommunication industry, government, and verticals joined to collaboratively create future business scenarios for 6G.
-----
10
First, the key change drivers for future 6G business were identified resulting in 153
forces [Yrjölä et al, 2020a]. Using these drivers, a set of dimensions and endpoints were
selected to form the basis for the scenario development as shown in Figure 3. Value
creation and value configuration were selected as the main business dimensions with
different end points emphasizing closed and open alternatives.
**_Business scenario theme_**
_Incumbent customer lock-in_ Value creation _Novel service providers_
_Supply-driven, Proprietary_ Value configuration _Open Ecosystem driven_
**End 1** **Dimension** **End 2**
**Fig. 3. Selected business scenario logic and dimensions.**
We also used a simple rules strategy framework presented in [Eisenhardt & Sull,
2001], which is a strategic management tool to develop strategies around identified
business opportunities and describing the main processes. It provides a highly practical
approach with guidelines in the following six rule categories introduced in [Eisenhardt
& Sull, 2001] and applied in the mobile communication market in [Ahokangas et al.,
2013]: 1) Nature of opportunity rules, 2) How to conduct business and processes in a
unique way, 3) Boundary rules to decide, which opportunities to pursue, 4) Priority
rules to identify and rank the opportunities, 5) Timing rules to synchronize emerging
opportunities and other parts of the company, and 6) Exit rules to selecting things to be
ended.
Next, we introduce the four developed business scenarios using the dimensions of
Figure 3 including Sustainable edge, Telco brokers, MNO6.0 and Over-the-top, as summarized in Figure 4, and presented in [Yrjölä et al., 2020a]. We also briefly summarize
strategies as simple rules that were created for the most plausible MNO6.0 scenario and
the most preferred Sustainable edge scenario.
**4.2** **6G Business Scenarios**
A set of business scenarios were developed in 6G Flagship’s white paper process in
2020, documented in [Yrjölä et al. 2020a] and summarized in the follows. Figure 4
summarizes the developed four business scenarios following the scenario logic of Figure 3.
In the first scenario, the Sustainable Edge Value Creation, scenario in the upper-right
corner of Figure 4, the value creation is customer attraction-driven, and the value configuration is open ecosystem-focused. This scenario is built on decentralized open value
configuration and ecosystem-driven business models where novel stakeholders take
over customer ownership and networks. Changing stakeholder roles include webscales,
over the top (OTT) companies and device vendors being responsible for business to
consumer (B2C) customers and local private cloud native networks serve business to
business (B2B) customers. The role of traditional MNOs has changed into a wholesale
connectivity service provider. Open source principles have become widely spread
-----
11
leading to technology and innovation ownership beyond traditional technology providers through open application programming interfaces (API) and novel resource brokerage. This scenario includes new stakeholder roles also in the form of local communities
and special interest groups operating various edge resources in specific locations, such
as campuses and remote areas to promote local innovation. New applications come with
6G technology that act as digital value platforms expanding our experiences towards
digital computer-generated virtual worlds. The current focus on global-scale solutions
changes towards local solutions that balance local demand with local supply and support circular economies. Especially the manufacturing vertical will move towards local
decentralized manufacturing supporting a new crowdsourcing-based production ecosystem.
**Empowered world**
**Competitive world**
**_REVOLUTIONARY DEVELOPMENT_**
Value creation - Customer attraction
_Novel service providers_
- Localized services Most probable
Most plausible
MNO6.0 Telco Brokers Most preferable
- Telcos drive technology - Telcos have primary customer
innovation and e2e value relationship, own data & run
chain service platform ecosystem
- Telcos’ own B2C & B2B - Tech providers drive
customer relationship technology ecosystem and run
- Platform as broker between NW infrastructure platform
customers and OTTs - Platform-based ecosystemic
- Innovation “engineering” business models
platform - Industry 5.0
- Resilient smart cities
**Protective world** _Incumbents_ **Networked world**
**_EVOLUTIONARY DEVELOPMENT_**
**Fig. 4. Summary of developed 6G business scenarios.**
In the second scenario, the Telco Broker Value Creation by Incumbents and Open
Ecosystem Value Configurations scenario shown in lower-right corner of Figure 4, the
main drivers for value creation remain the existing MNOs while value configuration is
based on open ecosystem-focus. The MNOs are in charge of customer relationships and
use service platform ecosystem to capture value. Technology providers’ role is to develop the required technologies and provide network infrastructure via platform-based
ecosystemic business models. Innovation ecosystem is broadened by the decoupling of
technology platforms. Industry 5.0 (I5.0) has emerged as a key vertical for collaborative
human machine interaction with robotization across services and industries. Real-time
data and high level of digital automation allow the industries to focus on servitization
of products. The speed of operations gets more and more rapid within the increasingly
-----
12
reprogrammable and reconfigurable world where design focus gets more and more
short-term.
In the third scenario, the MNO6.0 Value Creation scenario shown in the lower-left
corner of Figure 4, value creation is driven by the incumbent MNOs, and value configuration is closed supply-focused. The role of MNOs is strong, and they drive technological innovation and own the customer relationships. The existing dominant MNO
market position with strong customer base acts as the opportunity for businesses is and
the focus is on how to cost-efficiently increase the capacity to meet the growing demand. Technology developments on dynamic networks slicing allowing increasing
flexibility, shorter time-to-market, and cost optimization. With the MNO market dominance, the use of 6G in verticals is heavily dependent on MNOs’ business decisions.
Key technology developments in the form of automated network slicing and operations
in higher frequency bands and new machine learning inspired tools will be used to optimize network operations in a predictive manner allowing new applications. These networks will have been assembled with a public-private-partnership funding model, with
a view to resiliency and sustainability.
In the fourth scenario, the Over-the-Top Value Creation scenario shown in upper
left corner of Figure 4, value creation is customer attraction- and lock-in-driven, and
value configuration is closed supply-focused. The MNO dominance is replaced by
OTTs that have taken over the customer relationships with the help of their access to
customer data. The role of operators is to control the standardized and commoditized
connectivity technologies and manage the value chains. The role of edge computing is
to act as a new control point for serving of the verticals. Networks are programmable
and make use of digital twins that represent replicas of complex physical systems to
help in optimizing these systems. The ecosystem gets increasingly complicated with
different resources and assets needed to meet the versatile needs are brought together
by a set of stakeholders including physical infrastructure providers, equipment providers, and data providers under a complex regulatory framework defined by policymakers. Countries with more permitting rules act as resource pools and offer cheap labor,
natural resources, and data.
The four developed scenarios were then assessed in terms of their probability, plau
sibility and preferability. Both the most probable scenario was the Over-the-top scenario while the most plausible scenario was the MNO6.0 scenario. The most preferable
scenarios was the Sustainable edge scenario that can be seen to take a bold step towards
achievement of the UN SDGS, representing revolutionary and demand-driven transformations.
The developed business scenarios for 6G indicate that from economic perspective,
user experiences will be increasingly local and customized, delivered by local supply
models supporting spatial circular economies. New societal service delivery models
will appear through community-driven networks and public private partnerships and
the role of 6G will be substantial in vertical industries. The developed scenarios revealed interesting societal observations including increasing tensions between competitive, protective, networked and empowered worldviews. The role of power configurations keeps increasing and may shift from a multi-polarized world to a poly-nodal
world.
-----
13
The pressure on companies and governments to meet the UN SDGs is evident in the
business scenarios for 6G and the role of 6G as a provider of services towards environmental impact will be important. 6G with a set of new technologies will help in the
monitoring and steering of circular economy to promote a truly sustainable data economy. The developed scenarios also show that 6G development faces privacy and security issues related to business and regulation including different aims of governance
either stemming from governmental, company or end user perspectives. There the ecosystem-level configurations related to users, decentralized and community-driven business models and platforms and related user empowerment become increasingly important to support the role of local 6G services.
**4.3** **Strategic Options for 6G as Simple Rules**
Next, we summarize the developed strategic options for selected two scenarios using
the simple rules framework from [Eisenhardt & Sull, 2001] that was applied to characterize MNOs’ strategic choices in [Ahokangas et al., 2013]. For the most plausible
MNO 6.0 scenario, the baseline for building the simple rules is in the use of MNOs’
wide existing customer base that has growing capacity needs through investments to
strengthen customer lock-in and dominant market position in connectivity, enhanced
with customer data and holding on to spectrum. The goal is to maintain dominant market position through gaining access to a new wideband spectrum. Automation of network operations and the ability to dynamically create large numbers of networks slices
on-demand will help to increase flexibility, shorten time-to-market, and optimize costs.
Resources and services will be traded in automated marketplaces. The MNOs could
become a wholesale platform provider for other operators which would further
strengthen their market position. Regulations plays a key role in maintaining the MNO
market dominance which calls for close contact with the regulator. In the MNO6.0 scenario, the MNOs would never give up on their spectrum and customer data.
For the most preferred Sustainable Edge Scenario, the simple rules are built on the
use of new, local, and specialized demand, challenging incumbent MNOs in narrow
business segments specializing in governmental, municipal, vertical, or enterprise customers and vertical differentiation with increasing requirements for sustainability in
specific industry segments like education, healthcare, and manufacturing. These challenger operators think and act locally, close to the customer and promote resource sharing in different format such as spectrum and virtualized cloud infrastructures. Sustainability requirements in verticals are a major business opportunity through providing
vertical differentiation in specific segments like education, healthcare, manufacturing,
energy, and media and entertainment. The sustainable edge service provider supports
circular economy and promotes sharing economy principles in network deployment.
These locally operated networks have opportunities to scale up from local operations
to a multi-locality business. Local and private networks provide several benefits in
terms of security and data control, separation from public networks, access to highquality services in specific locations, increased flexibility, scalability and customization, and trustworthy reliabilities and latencies. Furthermore, networks can be deployed
as standalone sub-networks or integrated with MNO networks. This requires the
-----
14
establishment of multi-sided platforms -based regulations to govern privacy and security of users.
## 5 Future Outlook and Conclusions
Mobile communication research is increasingly addressing the use of 5G in verticals,
which has led to the emergence of local 5G network deployment models. Research on
6G has also started, with a bold goal of building a strong linkage with the United Nations Sustainable Development Goals (UN SDGs). These developments call for a
highly multi-disciplinary approach covering business, regulation and technology perspectives and our research is addressing these interrelated themes. This paper has provided an overview of the recent developments in 5G in verticals towards the development of sustainable 6G. We have highlighted the importance of the triangle of business
– regulation – technology perspectives in the development of new wireless technologies
and their deployments and summarized the advancements with a focus on local 5G
networks for serving the verticals’ needs towards meeting the sustainable development
goals.
From the business perspective, a business model for sustainability can help in de
scribing, analyzing, managing and communicating 1) a company’s sustainable value
proposition to its customers, and other stakeholders, 2) how it creates and delivers this
value, 3) and how it captures economic value while maintaining or regenerating natural,
social, and economic capital beyond its organizational boundaries. The development of
new vertical-specific 5G business opportunities calls for filling in the requirements of
scalability, replicability, and sustainability in a legitimate way in a platform ecosystem
of connectivity and data services. Digital convergence across industries and multi-level
6G platforms and ecosystems will create a complex environment where ecosystemic
business models for sustainability and the evolution of related regulations become important. Business ecosystems that aim to bring together stakeholders to solve systemic
sustainability problems will require open ecosystem-focused value configuration and
decentralized power configuration, focusing on the long tail of specialized user requirements that crosses a variety of industries. Future research prospects are particularly
related to the new business ecosystems, ecosystemic business models and changing
stakeholder roles that support sustainability.
From the regulation perspective, the serving of different verticals with 5G and future
6G networks introduces local and often private wireless networks to complement the
current mobile network operators (MNOs). The regulatory environment for 5G in verticals is very complex encompassing rules from both the electronic communications
market as well as specific verticals. Especially, the ways how rights to use radio frequencies are granted is critical for the establishment of local 5G and 6G networks. The
divergence in spectrum awards between countries is increasing with 5G, directly influencing the business opportunities in those countries. There are research prospects in
finding the best practices from the decisions by analyzing their impact.
For the technology perspective, 5G and future 6G architecture is expected to bring
additional modularity and flexibility for traditional MNOs as well as for new local
-----
15
operators in system deployments. Key technologies to enable open general-purpose 6G
architecture include distributed heterogenous cloud-native architecture, localization
and decomposition of network functions, software defined networking and network virtualization, among others. A critical aspect for the local private industrial networks is
their ability to create customized network slices that allow the delivery of services tailored to specific customer needs with service level agreed performance on demand. The
availability of spectrum for serving the verticals and operations in higher carrier frequencies present a major technical deployment challenge. The availability of spectrum
for serving the verticals on shared basis is important. New research prospects are especially in the 6G domain in order to find new indicators for 6G that take sustainability
in to account as well as the new network architecture for 6G needs.
This study has identified a further need for foresight research that explores the inter
related business – regulation – technology perspectives in the context of 5G in verticals
and on the road to sustainable 6G, with a special focus on how can 6G become a truly
general-purpose technology instead of simply an enabling technology, to support countries and organizations in the journey towards the achievement of the UN SDGs. Especially, the verticals burdened by increasing requirements for sustainability will be in
the key position in to realize the benefits of using the new technologies.
## References
1. 6G Flagship. White Papers. Available online: https://www.6gchannel.com/6g-white-papers/
(accessed on 24072020) (2020).
2. Ahokangas, P., Matinmikko, M., Yrjölä, S., Okkonen, H., Casey T. "Simple rules" for mo
bile network operators' strategic choices in future cognitive spectrum sharing networks.
IEEE Wireless Communications 20(2), 20-26 (2013).
3. Ahokangas, P., Matinmikko-Blue, M., Yrjölä, S., Seppänen, V., Hämmäinen, H., Jurva, R.,
& Latva-aho, M. Business models for local 5G micro operators. IEEE Transactions on Cognitive Communications and Networking, 5(3), 730-740 (2019).
4. Ahokangas, P., Yrjölä, S., Matinmikko-Blue, M., Seppänen, V. Transformation towards 6G
ecosystem. In: Proceedings of 2nd 6G Wireless Summit, Levi, Finland, (2020a).
5. Ahokangas, P., Matinmikko-Blue, M., Yrjölä, S. & Hämmäinen, H. Future vertical 5G plat
form ecosystems: Case study of a 5G enabled digitalized port stakeholders’ new interactions
and value configurations. In: Proceedings of International Telecommunications Society
online conference, Gothenburg, Sweden (2020b).
6. Aspara, J., Hietanen, J., Tikkanen, H. Business model innovation vs replication: financial
performance implications of strategic emphases. Journal of Strategic Marketing, 18(1), 3956 (2010).
7. Basole, R. C., Karla, J. On the evolution of mobile platform ecosystem structure and strat
egy. Business & Information Systems Engineering, 3(5), 313 (2011).
8. Biloslavo, R., Bagnoli, C., Massaro, M., Cosentino, A. Business model transformation to
ward sustainability: the impact of legitimation. Management Decision (2020).
9. Dreborg, K.H. Essence of backcasting. Futures 28(9), 813-828 (1996).
10. Eisenhardt K.M., Sull D.N. Strategy as simple rules. Harvard Business Review 79(1), 107
116 (2001).
-----
16
11. Evans, S., Vladimirova, D., Holgado, M., van Fossen, K., Yang, M., Silva, E.A., Barlow,
C.Y. Business Model Innovation for Sustainability: Towards a Unified Perspective for Creation of Sustainable Business Models. Bus. Strateg. Environ. 26, 597-608 (2017).
12. Iivari, M., Ahokangas, P., Matinmikko-Blue. M., Yrjölä, S. Opening closed business eco
systems boundaries with digital platforms: empirical case of a port. In: Ziouvelou, X. &
McGroarty, F. (Eds.), Emerging Ecosystem-Centric Business Models for Sustainable Value.
IGI Global (2020).
13. Kuhlman, T., Farrington, J. What is sustainability? Sustainability 2(11), 3436-3448 (2010).
14. Latva-aho, M., Leppänen, K. (Eds). Key drivers and research challenges for 6G ubiquitous
wireless intelligence. 6G Research Visions 1, University of Oulu, Finland (2019).
15. Lemstra, W. Leadership with 5G in Europe: Two contrasting images of the future, with pol
icy and regulatory implications. Telecommunications Policy, 42(8), 587-611 (2018).
16. Letaief, K.B., Chen, W., Shi, Y., Zhang, J., Zhang, Y.A. The Roadmap to 6G: AI Empow
ered Wireless Networks. IEEE Communications Magazine 57(8), 84-90 (2019).
17. Marano, V., Tallman, S., & Teegen, H. J.. The liability of disruption. Global Strategy Jour
nal, 10(1), 174-209 (2020).
18. Matinmikko, M. Latva-aho, M., Ahokangas, P., Yrjölä, S., Koivumäki. T. Micro operators
to boost local service delivery in 5G. Wireless Personal Communications, vol. 95, no. 1, pp.
69–82, May 2017.
19. Matinmikko, M., Latva-aho, M., Ahokangas, P., & Seppänen, V. On regulations for 5G:
Micro licensing for locally operated networks. Telecommunications Policy, 42(8), 622-635
(2018).
20. Matinmikko-Blue, M., Yrjölä, S., Seppänen, V., Ahokangas, P., Hämmäinen, H., & Latva
Aho, M. Analysis of spectrum valuation elements for local 5G networks: Case study of 3.5GHz band. IEEE Transactions on Cognitive Communications and Networking, 5(3), 741753 (2019).
21. Matinmikko-Blue, M., Aalto, S., Asghar, M.I., Berndt, H., Chen, Y., Dixit, S., Jurva, R.,
Karppinen, P., Kekkonen, M., Kinnula, M., Kostakos, P., Lindberg, J., Mutafungwa, E.,
Ojutkangas, K., Rossi, E., Yrjölä, S., Öörni, A., (Eds). White Paper on 6G Drivers and the
UN SDGs. 6G Research Visions 2, University of Oulu, Finland (2020a).
22. Matinmikko-Blue, M., Yrjölä, S., Ahokangas, P. Spectrum Management in the 6G Era: Role
of Regulations and Spectrum Sharing. In: Proceedings of 2nd 6G Wireless Summit, Levi,
Finland, (2020b).
23. Morgado, A., Huq, K. M. S., Mumtaz, S., & Rodriguez, J. A survey of 5G technologies:
regulatory, standardization and industrial perspectives. Digital Communications and Networks 4(2), 87-97 (2018).
24. Nielsen, C., & Lund, M. Building scalable business models. MIT Sloan Management Re
view, 59(2), 65-69 (2018).
25. Pouttu A. (Ed.). 6G White Paper on Validation and Trials for Verticals towards 2030’s. 6G
Research Visions 4, University of Oulu, Finland (2020).
26. Pujol, F., Elayoubi, S. E., Markendahl, J., Salahaldin, L. Mobile telecommunications eco
system evolutions with 5G. Communications & Strategies, (102), 109 (2016).
27. Saad, W.; Bennis, M.; Chen, M. A Vision of 6G Wireless Systems: Applications, Trends,
Technologies, and Open Research Problems. IEEE Network (2019).
28. Schaltegger, S., Hansen, E.G., Lüdeke-Freund, F. Business models for sustainability: ori
gins, present research, and future avenues. Organization & Environment, 29(1), 3-10 (2016).
29. Stevenson, T. Anticipatory action learning: conversations about the future. Futures 34, 417
425 (2012).
30. Schoemaker, P. Scenario Planning: A tool for Strategic Thinking. (1995).
-----
17
31. Schwartz, P. The Art of the Long View (1991).
32. Stewart, C. Integral scenarios: reframing theory, building from practice. Futures 40, 160
172 (2007).
33. Taleb. T, Aguiar, R. L., Yahia, I. G. B., Chatras, B., Christensen, G., Chunduri, U., Clemm,
A., Costa, X., Dong, L., Elmirghani, J., Yosuf, B., Foukas, X., Galis, A., Giordani, M., Gurtov, A., Hecker, A., Huang, C.-W., Jacquenet, C., Kellerer, W., .... Zorzi, M. White Paper
on 6G Networking. 6G Research Visions 6, University of Oulu, Finland (2020).
34. United Nations. Global indicator framework for the Sustainable Development Goals and
targets of the 2030 Agenda for Sustainable Development (2018).
35. Viswanathan, H., Mogensen, P.E. Communications in the 6G Era. IEEE Access 8, 57063
57074 (2020).
36. Vuojala, H., Mustonen, M., Chen, X., Kujanpää, K., Ruuska, P., Höyhtyä, M., Matinmikko
Blue, M., Kalliovaara, J., Talmola, P., Nyström, A.-G. Spectrum access options for vertical
network service providers in 5G. Telecommunications Policy 44(4), 101903 (2019).
37. Yrjölä, S., Ahokangas, P. & Matinmikko-Blue, M. (2019). Novel platform ecosystem busi
ness models for future wireless communications services and networks. In: Proxceedings of
NFF 2019, Vaasa, Finland (2019).
38. Yrjölä, S.; Ahokangas, P.; Matinmikko-Blue, M., Eds. White Paper on Business of 6G. 6G
Research Visions, No. 3, University of Oulu, Finland, (2020a).
http://urn.fi/urn:isbn:9789526226767
39. Yrjölä, S. How could Blockchain transform 6G towards open ecosystemic business models?
In: Proceedings of IEEE ICC 2020 Workshop on Blockchain for IoT and CPS, Dublin, Ireland, (2020).
40. Yrjölä, S.; Ahokangas, P.; Matinmikko-Blue. Sustainability as a Challenge and Driver for
Novel Ecosystemic 6G Business Scenarios. Sustainability (12), 21 (2020b).
41. Ziegler, V., Yrjölä, S. 6G Indicators of Value and Performance. In: Proceedings of 2nd 6G
Wireless Summit, Levi, Finland, (2020).
-----
|
{
"disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references', 'abstract'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1007/978-3-030-73423-7_13?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1007/978-3-030-73423-7_13, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "CLOSED",
"url": ""
}
| 2,020
|
[
"JournalArticle"
] | false
| null |
[] | 11,494
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Mathematics",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/0247616396cf94cb16350019864f3847dad97d32
|
[
"Computer Science"
] | 0.825968
|
Combining Public Key Encryption with Schnorr Digital Signature
|
0247616396cf94cb16350019864f3847dad97d32
|
[
{
"authorId": "20843836",
"name": "Laura Savu"
}
] |
{
"alternate_issns": null,
"alternate_names": null,
"alternate_urls": null,
"id": null,
"issn": null,
"name": null,
"type": null,
"url": null
}
|
This article presents a new signcryption scheme which is based on the Schnorr digital signature algorithm. The new scheme represents my personal contribution to signcryption area. I have implemented the algorithm in a program and here are provided the steps of the algorithm, the results and some examples. The paper also contains the presentation of the original Signcryption scheme, based on ElGamal digital signature and discusses the practical applications of Signcryption in real life. The purpose of the study is to combine the public key encryption with Schnorr digital signature in order to obtain less computational and communicational costs. Signcryption primitive is a better approach then Encrypt-then-Sign or Sign-then-Encrypt methods regarding the costs. All these algorithms offer the possibility to transmit a message over an insecure channel providing both authenticity and confidentiality.
|
**_Journal of Software Engineering and Applications, 2012, 5, 102-108_**
http://dx.doi.org/10.4236/jsea.2012.52016 Published Online February 2012 (http://www.SciRP.org/journal/jsea)
# Combining Public Key Encryption with Schnorr Digital Signature
#### Laura Savu
Department of Information Security, Faculty of Mathematics and Computer Science, University of Bucharest, Bucharest, Romania.
Email: laura.savu@microsoft.com
Received December 10[th], 2011; revised January 14[th], 2012; accepted February 7[th], 2012
### ABSTRACT
This article presents a new signcryption scheme which is based on the Schnorr digital signature algorithm. The new scheme represents my personal contribution to signcryption area. I have implemented the algorithm in a program and here
are provided the steps of the algorithm, the results and some examples. The paper also contains the presentation of the
original Signcryption scheme, based on ElGamal digital signature and discusses the practical applications of Signcryption in real life. The purpose of the study is to combine the public key encryption with Schnorr digital signature in order
to obtain less computational and communicational costs. Signcryption primitive is a better approach then Encrypt-thenSign or Sign-then-Encrypt methods regarding the costs. All these algorithms offer the possibility to transmit a message
over an insecure channel providing both authenticity and confidentiality.
**Keywords: Signcryption; Schnorr; Encryption; Digital Signature; Security; Confidentiality; ElGama; RSA; ECC**
### 1. Introduction
Signcryption is the primitive that has been proposed by
Youliang Zheng in 1997 and combines public key encryption with digital signature in a single logical step,
obtaining a less cost for both communication and computation [1].
Data confidentiality and data integrity are two of the
most important functions of modern cryptography. Confidentiality can be achieved using encryption algorithms
or ciphers, whereas integrity can be provided by the use of
authentication techniques. Encryption algorithms fall into
one of two broad groups: private key encryption and public key encryption. Likewise, authentication techniques
can be categorized by private key authentication algorithms and public key digital signatures.
While both private key encryption and private key authentication admit very fast computation with minimal
message expansion, public key encryption and digital
signatures generally require heavy computation, such as
exponentiations involving very large integers, together
with message expansion proportional to security parameters (such as the size of a large composite integer or
the size of a large finite field).
Signcryption has the intention that the primitive should
satisfy “Cost (Signature & Encryption) Cost (Signature) + Cost (Encryption)” This inequality can be interpreted in a number of ways:
A signcryption scheme should be more computation
nally efficient than a native combination of public-key
encryption and digital signatures.
A signcryption scheme should produce a signcryption
“ciphertext” which is shorter than a naive combination of a
public-key encryption ciphertext and a digital signature.
A signcryption scheme should provide greater security
guarantees and/or greater functionality than a native combination of public-key encryption and digital signatures
[1].
More recently, the significance of signcryption in realworld applications has gained recognition by experts in
data security. Since 2007, a technical committee within the
International Organization for Standardization (ISO/IEC
JTC 1/SC 27) has been developing an international standard for signcryption techniques [2].
The shared secret key between the parties makes possible an unlimited number of applications. Among these
applications, one can first think of the following three:
Secure and authenticated key establishment,
Secure multicasting, and
Authenticated key recovery.
A number of signcryption-based security protocols
have been proposed for aforementioned networks and
similar environments. These include:
Secure ATM networks,
Secure routing in mobile ad hoc networks,
Secure voice over IP (VoIP) solutions,
Encrypted email authentication by firewalls,
C i h © 2012 S iR **_JSEA_**
-----
Combining Public Key Encryption with Schnorr Digital Signature 103
Secure message transmission by proxy, and
Secure message transmission by proxy, and
Mobile grid web services.
There are also various applications of signcryption in
electronic commerce, where its security properties are very
useful. Analyzing this security scheme from an application-oriented point of view, can be observed that a great
amount of electronic commerce can take advantage of
signcryption to provide efficient security solutions in the
following areas:
Electronic payment,
Electronic toll collection system,
Authenticated and secured transactions with smart cards,
etc.
My personal contribution to the article is represented
by the Schnorr Signcryption scheme which has been introduced here. Schnorr Signcryption scheme is made up
of a combination between a public key encryption scheme and a digital signature scheme. On the base of the
scheme that I present here stands the Schnorr digital sig[nature. A Schnorr signature is a digital signature produ-](http://en.wikipedia.org/wiki/Digital_signature)
ced by the Schnorr signature algorithm. Its security is
[based on the intractability of certain discrete logarithm](http://en.wikipedia.org/wiki/Discrete_logarithm)
problems. It is considered the simplest digital signature
[scheme to be provably secure in a random oracle model.](http://en.wikipedia.org/wiki/Random_oracle)
It is efficient and generates short signatures.
A signcryption scheme typically consists of five algorithms, Setup, KeyGenS, KeyGenR, Signcrypt, Unsigncrypt:
Setup-takes as input a security parameter 1^ k and outputs any common parameters _param_ required by the
signcryption schemes. This may include the security parameter 1^ k, the description of a group G and a generator g for that group, choices for hash functions or
symmetric encryption schemes, etc.
Key Generation S(Gen) generates a pair of keys for the
sender.
Key Generation R(Gen) generates a pair of keys for the
receiver.
Signcryption (SC) is a probabilistic algorithm.
Unsigncryption (USC) is a deterministic algorithm.
A signcryption scheme is a combination between a
public key encryption algorithm and a digital signature
scheme.
A public key encryption scheme consists of three polynomial-time algorithms (EncKeyGen, Encrypt, Decrypt).
**EncKeyGen—Key generation is a probabilistic algori-**
thm that takes as input a security parameter 1^ k and outputs a key pair (skenc, pkenc), written (skenc, pkenc)R←
EncKeyGen (1^ k ). The public encryption key pkenc is
widely distributed, while the private decryption key skebnc should be kept secret. The public key defines a message m ∈ M and a ciphertext ∈ C.
**Encrypt—Encryption is a probabilistic algorithm that**
takes a message m ∈ M and the public key pkenc as input and outputs a ciphertext C ∈ C, written C ← Encrypt (pkenc, m).
**Decrypt—Decryption is a deterministic algorithm that**
takes a ciphertext C ∈ C and the private key skenc as
input and outputs either a message m ∈ M or the failure symbol ⊥, written m ← Decrypt (skenc, C).
The article is structured in seven parts, as follows. Signcryption and its properties definitions are contained in
the first part. Also here, in introduction, are presented the
practical applications of Signcryption in real life. In the
second part is exposed the original signcryption primitive
introduced by Youliang Zheng, which combines public key
encryption and a derivation of ElGamal digital signature
algorithm. Part three contains the presentation of the new
sygncryption scheme, Schnorr Signcryption, as a result
of the combination of public key encryption and Schnorr
digital signature algorithm. The step-by-step implementtation of the Schnorr Signcryption scheme in a source
code program is reflected in the fourth part. Strating with
the fifth part begins the analyze of the security models on
Schnorr Signcryption. The two-users security model is presented in the sixth part and multi-user security model is
presented in the seventh part. In each of this models there
is exposed another classification for security, the insider
security and the outsider security.
### 2. Related Work
#### 2.1. Elgamal Signcryption
The original signcryption scheme that has been introduced by Youliang Zheng in 1997 is created on a derivation of ElGamal digital signature standard, combined
with a public key encryption scheme.
Based on discrete algorithm problem, ElGamal Signcryption cost is:
58% less in average computation time;
70% less in message expansion.
Here is the detailed presentation of the fifth algorithms
that make up the ElGamal signcryption scheme.
1) Setup
Signcryption parameters:
p = a large prime number, public to all;
q = a large prime factor of p − 1, public to all;
g = an integer with order q modulo p, in [1, , p − 1],
public to all;
hash = a one-way hash function;
KH = a keyed one-way hash function = KHk(m) =
hash (k, m);
(E, D) = the algorithms which are used for encryption
and decryption of a private key cipher.
Alice sends a message to Bob.
2) KeyGen sender
Alice has the pair of keys (Xa, Ya):
C i h © 2012 S iR **_JSEA_**
-----
104 Combining Public Key Encryption with Schnorr Digital Signature
Xa = Alice’s private key, chosen randomly from [1,
, q − 1]
Ya = Alice’s public key = g xa mod p.
3) KeyGen receiver
Bob has the pair of keys (Xb, Yb):
Xb = Bob’s private key, chosen randomly from [1, ,
q − 1]
Yb = Bob’s public key = g xb mod p.
4) Signcryption
In order to signcrypt a message m to Bob, Alice has to
accomplish the following operations:
Calculate
k = hash Yb x mod p
Split k in k1 and k2 of appropriate length.
Calculate r = KHk2(m) = hash(k2, m)
Calculate s = x/(r + Xa) mod q, if SDSS1 is used
Calculate s = x/(1 + Xa · r) mod q, if SDSS2 is used
Calculate c = Ek1(m) = the encryption of the message
m with the key k1.
Alice sends to Bob the values (r, s, c).
5) Unsigncryption
In order to unsigncrypt a message from Alice, Bob has
to accomplish the following operations:
Calculate k using r, s, g, p, Ya and Xb
s xb
hash Ya g r mod p, if is used SDSS1;
s xb
hash g Ya r mod p, if is used SDSS2;
Split k in k1 and k2 of appropriate length.
Calculate m using the decryption algorithm m = Dk1(c).
Accept m as a valid message only if KHk2(m) = r.
Using the two schemes SDSS1 and SDSS2, two signcryption schemes have been created, SCS1 and SCS2, respectively. The two signcryption schemes share the same
communication overhead, (|hash(*)| + |q|). SCS1 involves
one less modular multiplication in signcryption then
SCS2, both have a similar computational cost for unsigncryption [1].
#### 2.2. Rsa Signcryption
Rivest introduced for the first time in 1978 the publickey encryption scheme and digital signature scheme [3].
The RSA transform has been the basis of dozens of
public-key encryption schemes and digital signature
schemes, which have proven to be very successful and
have been very widely deployed in industry. They are widely
used in the design of public-key encryption and digital
signature schemes.
The RSA transform was introduced by Rivest, Shamir,
and Adleman in 1978 [3]. The exact definition of the problem depends upon the distribution from which the two
prime numbers p and q are drawn. For our purposes, this
is defined by a probabilistic, polynomial-time RSA parameter generation algorithm RSAGen, which takes as in
put a security parameter 1^ k and outputs two primes (p,
q) with the property that N = pq is a k-bit integer [4].
**Signcrypt (** _fS_ 1, _fR m,_ )
Bind pkS||pkR
r 0,1 _d_ | |m
c H (bind, m||r)
d m||r
w c
s G (bind, c) ○ d
C fR ( _fS_ []1 **(w||s))**
Return C
**Unsigncrypt (** _fS fR,_ 1,C )
Bind pkS||pkR
(w||s) fS ( _fR1,C_ )
m||r G (bind, w) © s
If H (bind, m||r) = w, return m
Else return ⊥
#### 2.3. Elliptic Curve Cryptography Signcryption
[The first signcryption scheme was introduced by Yuliang](http://en.wikipedia.org/wiki/Yuliang_Zheng)
[Zheng in 1997 [1]. Zheng also proposed an elliptic curve-](http://en.wikipedia.org/wiki/Yuliang_Zheng)
based signcryption scheme that saves 58% of computational and 40% of communication costs when it is compared with the traditional elliptic curve-based signature[then-encryption schemes [5].](http://en.wikipedia.org/wiki/Signcryption#cite_note-1)
Here is presented the scheme for an elliptic curve based signcryption algorithm introduced by Mohsen Toorani and Ali Asghar Beheshti Shirazi in [6].
**Signcryption (Alice)**
Choosing r in [1, n − 1]
R = rG = (xR, yR)
K = rU = (xK, yK)
s = r[]1 (H (M) + xRdA) (mod n)
e = H (M||s)
C = (M||e) © xK
**Unsigncryption (Bob)**
K = dB R = (xK, yK)
(M||e’) = C © xK
e’ = H(M||s)
If e <> e’ then rejects M’
Else
u = s[]1 H(M)
v = s[]1 xR
uG + vU = (x’R, y’R)
Signature verification: Is xR = x’R ?
The elliptic curve-based schemes are usually based on
difficulty of Elliptic Curve Discrete Logarithm Problem
(ECDLP) that is computationally infeasible under certain
circumstances [7]. The elliptic curve-based systems can
attain to a desired security level with significantly smaller
keys than those of required by their exponential-based
C i h © 2012 S iR **_JSEA_**
-----
Combining Public Key Encryption with Schnorr Digital Signature 105
counterparts. This can enhance the speed and leads to
efficient use of power, bandwidth, and storage that are
the basic limitations of resource-constrained devices [8].
Throughout the years, there have been proposed many
other signcryption schemes, each with its own problems
and limitations, while offering different level of security
services and computational costs.
### 3. Implementation of the New Signcryption Scheme
[A Schnorr signature is a digital signature produced by the](http://en.wikipedia.org/wiki/Digital_signature)
Schnorr signature algorithm. Its security is based on the
[intractability of certain discrete logarithm problems. It is](http://en.wikipedia.org/wiki/Discrete_logarithm)
considered the simplest digital signature scheme to be
[provably secure in a random oracle model [9].](http://en.wikipedia.org/wiki/Random_oracle)
**_Choosing parameters_**
[All users of the signature scheme agree on a group G](http://en.wikipedia.org/wiki/Group_(mathematics))
[with generator g of prime order q in which the discrete](http://en.wikipedia.org/wiki/Discrete_log)
[log problem is hard.](http://en.wikipedia.org/wiki/Discrete_log)
**_Key generation_**
Choose a private signing key x.
The public verification key is y = g[x].
**_Signing_**
To sign a message M:
Choose a random k.
Let r = g[k]
Let e = H (M | | r), where || denotes concatenation and r
is represented as a bit string. H is a cryptographic hash
function H : 0,1 * _q_ .
Let s = (k – xe).
The signature is the pair (s, e).
**_Verifying_**
Let rv = g[s]y[e]
Let ev = H (M | | rv)
If ev = e then the signature is verified.
**_Demonstration of correctness_**
It can be observed that ev = e if the signed message
equals the verified message:
[r]v g ys e gk xe g xe gk r, and hence ev = H (M | | rv)
= H(M | | r) = e.
It has been considered that k < q and the assumption
that the hash function is collision-resistant.
Public elements: G, g, q, y, s, e, r.
Private elements: k, x. [10]
A Schnorr Signcryption scheme is based on Schnorr
digital signature algorithm.
Here is the detailed presentation of the fifth algorithms
that make up the Schnorr signcryption scheme.
1) Setup
Schnorr Signcryption parameters:
p = a large prime number, public to all;
q = a large prime factor of p-1, public to all;
g = an integer with order q modulo p, in [1,, p − 1],
public to all;
hash = a one-way hash function;
KH = a keyed one-way hash function = KHk (m) =
hash (k, m);
(E, D) = the algorithms which are used for encryption
and decryption of a private key cipher.
Alice sends a message to Bob.
2) KeyGen sender
Alice has the pair of keys (Xa, Ya):
Xa = Alice’s private key, chosen randomly from [1,,
q − 1]
Ya = Alice’s public key = g[-]xa mod p.
3) KeyGen receiver
Bob has the pair of keys (Xb, Yb):
Xb = Bob’s private key, chosen randomly from [1,,
q − 1];
Yb = Bob’s public key = g[-]xb mod p.
4) Signcryption
In order to signcrypt a message m to Bob, Alice has to
accomplish the following operations:
Calculate
k hash Yb x mod p ;
Split k in k1 and k2 of appropriate length.
Calculate r = KHk2(m) = hash (h2, m);
Calculate s = x + (r* Xa) mod q;
Calculate c = Ek1(m) = the encryption of the message
m with the key k1.
Alice sends to Bob the values (r, s, c).
5) Unsigncryption
In order to unsigncrypt a message from Alice, Bob has
to accomplish the following operations:
Calculate k using r, s, g, p, Ya and Xb
Xb
k hash g s Ya r mod p
Split k in k1 and k2 of appropriate length.
Calculate m using the decryption algorithm m = Dk1 (c).
Accept m as a valid message only if KHk2 (m) = r.
Analyzing the two presented signcryption schemes, it
can be observed that in case of Shnorr signcryption the
computation of s, which is s = x + (r* Xa) mod q, is less
consuming comparing with the formula used in ElGamal
algorithm, where s is s = x/(r+Xa) mod q.
Another difference is on the level of unsigncryption step
as k is computing differently, using this formula for Sch
Xb
rr k hash g s Ya r mod p and this formula for El
mal
s Xb
k hash g r Ya mod p .
### 4. Security Models for Schnorr Signcryption Scheme
The first attempt to produce security models for signcrtion was given by Steinfeld and Zheng [11].
A family of security models for signcryption in both
two-user and multi-user settings was presented by An [12]
C i h © 2012 S iR **_JSEA_**
-----
106 Combining Public Key Encryption with Schnorr Digital Signature
in their work on signcryption schemes built from blackbox signature and encryption schemes.
Defining the security of signcryption in the public-key
setting is more involved than the corresponding task in
the symmetric setting [13] due to the asymmetric nature
of the former. The asymmetry of keys makes a difference
in the notions of both authenticity and privacy on two
major fronts which are addressed in this chapter.
The first difference for Schnorr signcryption is that the
security of the signcryption needs to be defined in the
multi-user setting, where issues with users’ identities need
to be addressed. On the other hand, authenticated encryption in the symmetric setting can be fully defined in a
much simpler two-user setting.
The case of Schnorr settings not only makes a difference in the multiuser and two-user settings but also
makes a difference in the adversary’s position depending
on its knowledge of the keys. There are two definitions
for security of signcryption depending on whether the
adversary is an “outsider” (a third party who only knows
the public information) or “insider” (a legal user of the
network, either the sender or the receiver, or someone
that knows the secret key of either the sender or the receiver). In the first case the security model is named
“outsider security” and in the latter “insider security”.
#### 4.1. Two-Users Security Model
In the symmetric setting, there is only one specific pair of
users who
1) Share a single key;
2) Trust each other;
3) “Know who they are”;
4) Only care about being protected from “the rest of
the world.”
In contrast, in Schnorr signcryption setting, each user
independently publishes its public keys, after which it
can send/receive messages to/from any other user. In particular, 1) each user should have an explicit identity (associated with its public key); 2) each signcryption has to
explicitly contain the (presumed) identities of the sender
S and the receiver R; 3) each user should be protected
from every other user.
The security goal is to provide both authenticity and
privacy of communicated data. In the symmetric setting,
since the sender and the receiver share the same secret
key, the only security model that makes sense is one in
which the adversary is modeled as a third party or an outsider who does not know the shared secret key. For Schnorr signcryption setting, the sender and the receiver do
not share the same secret key but each has his/her own secret key. Due to this asymmetry of the secret keys, the data
needs to be protected not only from an outsider but also
from an insider who is a legal user of the system (the sender
or the receiver themselves or someone who knows either
the sender’s secret key or the receiver’s secret key) [4].
#### 4.2. Multi-User Security Model
A central difference between the multi-user model and the
two-user models is the extra power of the adversary. In the
multi-user model, the attacker may choose receiver (resp.
sender) public keys when accessing the attacked users’
signcryption (resp. unsigncryption) oracles. For signcryption schemes that share some functionality between the
signature and the encryption components, such as are the
case for Zheng’s Signcryption scheme and Schnorr Signcryption scheme, the extra power of the adversary in the
multi-user model may be much more significant, and a
careful case-by-case analysis is required to establish security of such schemes in the multi-user model.
As in the two-user setting, the multi-user setting also
has two types of models depending on the identity of the
attacker: an insider model and an outsider model.
### 5. Experimental Results
Here is provided an example from the execution of the
program on small numbers.
Example:
p = 23, q = 11, g = 2, X=3
XA = 4 => YA=13
XB=5 => YB=18
k = 13 => hash(k) = vTB6PsMp4Qos/4+4dICCPaEU+
PQ=
k1 = vTB6PsMp4Qos/w==
k2 = j7h0gII9oRT49A==
hash(k2, m) =
E2726583242AB5CCE58AE1151DB126208F17932F
hash(k2, m) in base 10 =
1292783042124763369608714420962730428414981280559
(hash(k2,m) in base 10) mod p = 3
s mod q = x + (r*Xa) mod q = 4
Unsigncrypt k = 13
In Table 1 is presented the cost evaluation for the signature and verification in ElGamal and Schnorr signcryption schemes.
It is important the improvement for the cost consumption that has been made in the case of the proposed scheme, as at this step it is not necessary to be calculated the
modular inverse.
Texp: the time for a modular exponential computation.
Tm: the time for a modular multiplication computation.
Tinv: the time for a modular inverse computation.
Th: the time for a one way hash function f(_) computation.
### 6. Conclusions and Future Work
This paper presents a new Signcryption scheme which is
C i h © 2012 S iR **_JSEA_**
-----
Combining Public Key Encryption with Schnorr Digital Signature
107
**Table 1. The comparison between the proposed Schnorr Signcryption scheme and the initial Youliang Zheng Signcryption**
**scheme.**
The Proposed Schnorr Signcryption Scheme The Initial Youliang Zheng Signcryption Scheme
Computation cost for signature generation Th + Tm Th + Tm + Tinv
Computation cost for verifying converted signature Th + Tm + Texp Th + Tm + Tinv + Texp
based on Schnorr digital signature algorithm. This scheme is named Schnorr Signcryption and it implements in a
single logical step both public key encryption and digital
signature, offering less costs as using these two cryptographic functions individually.
In signcryption area, the following problems seem interesting in future research: 1) presenting a formal model for group signcryption, and proposing provably secure schemes; 2) Designing schemes to support dynamic
group member management in the sense that group member can join or leave the group efficiently and dynamically; 3) Optimizing the open procedure so that it does
not linearly depend on the number of group members, so
that such schemes are suitable for large groups.
### REFERENCES
[1] Y. Zheng, “Digital Signcryption or How to Achieve Cost
(Signature & Encryption) << Cost(Signature) + Cost (Encryption),” Full Version, 2011.
http://www.sis.uncc.edu/yzheng/papers/
[2] International Organization for Standardization, “IT Security Techniques—Signcryption,” ISO/IEC WD 29150,
2008.
[3] R. L. Rivest, A. Shamir and L. Adleman, “A Method for
Obtaining Digital Signatures and Public-Key Cryptosystems,” Communications of the ACM, Vol. 21, No. 2, 1978,
[pp. 120-126. doi:10.1145/359340.359342](http://dx.doi.org/10.1145/359340.359342)
[4] A. Dent and Y. L. Zheng, “Practical Signcryption, a Volume in Information Security and Cryptography,” SpringerVerlag, Berlin, 2010.
[5] Y. Zheng and H. Imai, “How to Construct Efficient Signcryption Schemes on Elliptic Curves,” Information Proc_essing Letters, Vol. 68, No. 5, 1998, pp. 227-233._
### Appendix
[doi:10.1016/S0020-0190(98)00167-7](http://dx.doi.org/10.1016/S0020-0190(98)00167-7)
[6] M. Toorani and A. A. B. Shirazi, “Cryptanalysis of an
Elliptic Curve-Based Signcryption Scheme,” International
_Journal of Network Security, Vol. 10, No. 1, 2010, pp._
51-56.
[7] D. Hankerson, A. Menezes and S. Vanstone, “Guide to
Elliptic Curve Cryptography,” Springer-Verlag, New York,
2004.
[8] M. Toorani and A. A. B. Shirazi, “LPKI—A Lightweight
Public Key Infrastructure for the Mobile Environments,”
_Proceedings of the 11th IEEE International Conference_
_on Communication Systems, Guangzhou, 19-21 Novem-_
ber 2008, pp. 162-166.
[9] C. P. Schnorr, “Efficient Identification and Signatures for
Smart Cards,” In: G. Brassard, Ed., Advances in Cryptol_ogy—Crypto’89,_ _Lecture Notes in Computer Science_ _No_
435, Springer-Verlag, 1990. pp. 239-252.
[10] C.-P. Schnorr, “Efficient Signature Generation by Smart
Cards,” _Journal of Cryptology, Vol. 4, No. 3, 1991, pp._
[161-174. doi:10.1007/BF00196725](http://dx.doi.org/10.1007/BF00196725)
[11] R. Steinfeld and Y. Zheng, “A Signcryption Scheme
Based on Integer Factorization,” In: J. Pieprzyk, E. Okamoto and J. Seberry, Eds., _Information Security Work-_
_shop, Lecture Notes in Computer Science, Vol. 1975,_
Springer, Berlin, 2000, pp. 308-322.
[12] J. H. An, Y. Dodis and T. Rabin, “On the Security of Joint
Signatures and Encryption,” In: L. Knudsen, Ed., _Ad-_
_vances in Cryptology—Eurocrypt 2002, Lecture Notes in_
_Computer Science, Vol. 2332, Springer, Berlin, 2002, pp._
83-107.
[13] M. Bellare and C. Namprempre, “Authenticated Encryption: Relations among Notions and Analysis of the Generic Composition Paradigm,” In: T. Okamoto, Ed., _Ad-_
_vances in Cryptology—Asiacrypt 2000,_ _Lecture Notes in_
_Computer Science, Vol. 1976, Springer, Berlin, 2000, pp._
531-545.
I created a source code program that verifies my algorithm. Executing this program I could generate examples. The
step-by-step implementation of the algorithm is as follows:
1) Calculate Ya and Yb
double powA = Math.Pow(g, xA);
int pow_intA = Convert.ToInt32(powA);
C i h © 2012 S iR **_JSEA_**
-----
108 Combining Public Key Encryption with Schnorr Digital Signature
int invA = modInverse(pow_intA, p);
2) Calculate k
int yB = Convert.ToInt32(textBox11.Text);
int x = Convert.ToInt32(textBox18.Text);
int p = Convert.ToInt32(textBox4.Text);
string cheie = (BigInteger.ModPow(yB, x, p)). ToString();
3) Calculate hash(k)
string HashDeCheie = _calculateHash(cheie);
textBox13.Text = HashDeCheie;
4) Split k in two keys k1 and k2 with the same lenght
byte[] k = Convert.FromBase64String(textBox13.Text);
byte[] k1 = new byte[k.Length/2];
byte[] k2 = new byte[k.Length - k.Length/2];
Buffer.BlockCopy(k, 0, k1, 0, k.Length/2);
Buffer.BlockCopy(k, k.Length/2, k2, 0, k.Length - k.Length/2);
byte[] test = new byte[k.Length];
k1.CopyTo(test, 0);
k2.CopyTo(test, k1.Length);
5) Calculate r using k2; r = hash (k2, m)
BigInteger p = BigInteger.Parse(textBox4.Text);
System.Text.ASCIIEncoding encoding = new System.Text.ASCIIEncoding();
byte[] keyByte = encoding.GetBytes(key);
HMACSHA1 hmacsha1 = new HMACSHA1(keyByte);
byte[] messageBytes =encoding.GetBytes(message);
byte[] hashmessage = hmacsha1.ComputeHash(messageBytes);
6) Calculate r using k2; transform the value obtained from hash in base 10
textBox19.Text = fn16to10(textBox15.Text).ToIntString();
7) Calculate the modulo p of the number obtained in base 10
BigInteger nr = BigInteger.Parse(textBox19.Text);
BigInteger p = BigInteger.Parse(textBox4.Text);
BigInteger rest = 0;
BigInteger.DivRem(nr, p, out rest);
8) Calculate s
BigInteger q = Convert.ToInt32(textBox5.Text);
BigInteger r = Convert.ToInt32(textBox20.Text);
BigInteger XA = Convert.ToInt32(textBox9.Text);
BigInteger X = Convert.ToInt32(textBox18.Text);
BigInteger prod = BigInteger.Multiply(r, XA);
BigInteger sum = X + prod;
BigInteger rest;
BigInteger.DivRem(sum, q, out rest);
9) Encrypt m using the k1
10) Calculate k
BigInteger rez2 = BigInteger.Pow(rez1, XB);
B igInteger invK = modInverseBI(rez2, p)
C i h © 2012 S iR **_JSEA_**
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.4236/JSEA.2012.52016?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.4236/JSEA.2012.52016, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "HYBRID",
"url": "http://www.scirp.org/journal/PaperDownload.aspx?paperID=17484"
}
| 2,012
|
[] | true
| 2012-02-27T00:00:00
|
[
{
"paperId": "0bb7d3807f7a7066a85c297bd258da37d2defdfb",
"title": "Practical Signcryption"
},
{
"paperId": "377182c155e985538ad5caddf758e936c2de78c8",
"title": "Cryptanalysis of an Elliptic Curve-based Signcryption Scheme"
},
{
"paperId": "091b7275abdb297ff7aedfa46e710352fcd21a0d",
"title": "LPKI - A lightweight public key Infrastructure for the mobile environments"
},
{
"paperId": "80b388f07313e609a0fbd7dadfbadc69ae3b653e",
"title": "Protection"
},
{
"paperId": "3b4db693363eba0e4a7bc6903129841d6e1b3c93",
"title": "The International Organization for Standardization."
},
{
"paperId": "e328960b8e36e96a163f852abcda0e48949824a6",
"title": "A Signcryption Scheme Based on Integer Factorization"
},
{
"paperId": "512e08451eb0d805c77b86e5821560f3b7dec565",
"title": "Authenticated Encryption: Relations among Notions and Analysis of the Generic Composition Paradigm"
},
{
"paperId": "bc533d2f27381d81d8e0cd3f445c54556e938816",
"title": "The State of Elliptic Curve Cryptography"
},
{
"paperId": "d21cde8bc0b11caef550a3e0e5b4cb2077347570",
"title": "How to Construct Efficient Signcryption Schemes on Elliptic Curves"
},
{
"paperId": "072e8123e534331625f52111cb5b7c0441bee8aa",
"title": "Digital Signcryption or How to Achieve Cost(Signature & Encryption) << Cost(Signature) + Cost(Encryption)"
},
{
"paperId": "8d69c06d48b618a090dd19185aea7a13def894a5",
"title": "Efficient Identification and Signatures for Smart Cards (Abstract)"
},
{
"paperId": "2642738e9977c08d4085ce1c6530d63545383d30",
"title": "Efficient signature generation by smart cards"
},
{
"paperId": null,
"title": "BigInteger prod = BigInteger.Multiply(r, XA)"
},
{
"paperId": null,
"title": "Calculate k BigInteger rez2 = BigInteger.Pow(rez1"
},
{
"paperId": null,
"title": "BigInteger sum = X + prod"
},
{
"paperId": null,
"title": "Calculate k int yB = Convert.ToInt32(textBox11"
},
{
"paperId": null,
"title": "BlockCopy(k, 0, k1, 0, k.Length/2)"
},
{
"paperId": null,
"title": "string cheie = (BigInteger.ModPow(yB, x, p))"
},
{
"paperId": null,
"title": "Encrypt m using the k1"
},
{
"paperId": null,
"title": "DivRem(nr, p, out rest)"
},
{
"paperId": null,
"title": "B igInteger invK = modInverseBI(rez2, p)"
},
{
"paperId": null,
"title": "On the Security of Joint Signatures and Encryption"
},
{
"paperId": null,
"title": "ASCIIEncoding encoding = new System"
},
{
"paperId": null,
"title": "Calculate r using k2"
}
] | 8,447
|
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/024aa4597ede5823b301593385cb892df14180da
|
[
"Computer Science"
] | 0.895508
|
Evaluating Countermeasures for Verifying the Integrity of Ethereum Smart Contract Applications
|
024aa4597ede5823b301593385cb892df14180da
|
IEEE Access
|
[
{
"authorId": "2117668992",
"name": "Suhwan Ji"
},
{
"authorId": "2111404615",
"name": "Dohyung Kim"
},
{
"authorId": "2050475",
"name": "Hyeonseung Im"
}
] |
{
"alternate_issns": null,
"alternate_names": null,
"alternate_urls": [
"http://ieeexplore.ieee.org/servlet/opac?punumber=6287639"
],
"id": "2633f5b2-c15c-49fe-80f5-07523e770c26",
"issn": "2169-3536",
"name": "IEEE Access",
"type": "journal",
"url": "http://www.ieee.org/publications_standards/publications/ieee_access.html"
}
|
Blockchain technology, which provides digital security in a distributed manner, has evolved into a key technology that can build efficient and reliable decentralized applications (called DApps) beyond the function of cryptocurrency. The characteristics of blockchain such as immutability and openness, however, have made DApps more vulnerable to various security risks, and thus it has become of great significance to validate the integrity of DApps before they actually operate upon blockchain. Recently, research on vulnerability in smart contracts (a building block of DApps) has been actively conducted, and various vulnerabilities and their countermeasures were reported. However, the effectiveness of such countermeasures has not been studied well, and no appropriate methods have been proposed to evaluate them. In this paper, we propose a software tool that can easily perform comparative studies by adding existing/new countermeasures and labeled smart contract codes. The proposed tool demonstrates verification performance using various statistical indicators, which helps to identify the most effective countermeasures for each type of vulnerability. Using the proposed tool, we evaluated state-of-the-art countermeasures with 237 labeled benchmark codes. The results indicate that for certain types of vulnerabilities, some countermeasures show evenly good performance scores on various metrics. However, it is also observed that countermeasures that detect the largest number of vulnerable codes typically generate much more false positives, resulting in very low precision and accuracy. Consequently, under given constraints, different countermeasures may be recommended for detecting vulnerabilities of interest. We believe that the proposed tool could effectively be utilized for a future verification study of smart contract applications and contribute to the development of practical and secure smart contract applications.
|
Received June 2, 2021, accepted June 16, 2021, date of publication June 21, 2021, date of current version June 30, 2021.
_Digital Object Identifier 10.1109/ACCESS.2021.3091317_
# Evaluating Countermeasures for Verifying the Integrity of Ethereum Smart Contract Applications
SUHWAN JI 1, DOHYUNG KIM 1,2, AND HYEONSEUNG IM 1,2
1Interdisciplinary Graduate Program in Medical Bigdata Convergence, Kangwon National University, Chuncheon 24341, South Korea
2Department of Computer Science and Engineering, Kangwon National University, Chuncheon 24341, South Korea
Corresponding authors: Dohyung Kim (d.kim@kangwon.ac.kr) and Hyeonseung Im (hsim@kangwon.ac.kr)
This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT)
(No. 2019R1F1A1063272, 2020R1F1A1048395, and 2020R1A4A3079947).
**ABSTRACT Blockchain technology, which provides digital security in a distributed manner, has evolved into**
a key technology that can build efficient and reliable decentralized applications (called DApps) beyond the
function of cryptocurrency. The characteristics of blockchain such as immutability and openness, however,
have made DApps more vulnerable to various security risks, and thus it has become of great significance
to validate the integrity of DApps before they actually operate upon blockchain. Recently, research on
vulnerability in smart contracts (a building block of DApps) has been actively conducted, and various
vulnerabilities and their countermeasures were reported. However, the effectiveness of such countermeasures
has not been studied well, and no appropriate methods have been proposed to evaluate them. In this
paper, we propose a software tool that can easily perform comparative studies by adding existing/new
countermeasures and labeled smart contract codes. The proposed tool demonstrates verification performance
using various statistical indicators, which helps to identify the most effective countermeasures for each type
of vulnerability. Using the proposed tool, we evaluated state-of-the-art countermeasures with 237 labeled
benchmark codes. The results indicate that for certain types of vulnerabilities, some countermeasures show
evenly good performance scores on various metrics. However, it is also observed that countermeasures that
detect the largest number of vulnerable codes typically generate much more false positives, resulting in
very low precision and accuracy. Consequently, under given constraints, different countermeasures may be
recommended for detecting vulnerabilities of interest. We believe that the proposed tool could effectively be
utilized for a future verification study of smart contract applications and contribute to the development of
practical and secure smart contract applications.
**INDEX TERMS Blockchain, countermeasure, Ethereum, smart contract, vulnerability.**
**I. INTRODUCTION**
Since Bitcoin [1], which was designed using blockchain, was
introduced, blockchain technology has evolved and interests
in its applications have greatly been increasing. With the
ability to provide digital security in a distributed manner,
blockchain has been used to develop a variety of decentralized
applications across the industry. In particular, such decentralized applications, called DApps, often operate upon the
Ethereum Virtual Machine (EVM), and they are built using
smart contracts which are a piece of code that enables DApps
to interact with the underlying Ethereum blockchain [2].
The associate editor coordinating the review of this manuscript and
approving it for publication was Hailong Sun .
Despite the great advantages of using the blockchain technology, however, it has ironically been revealed that smart
contracts are vulnerable to various security risks due to the
blockchain’s essential features, such as transparency and
immutability [3]–[6]. For example, if a smart contract incurs
a wrong transaction (by mistake or malicious attacks) and the
result is once written to the blockchain, then the transaction
can hardly be corrected. Rather, the blockchain should be
destroyed (or hardforked). Because of that reason, it is of
considerable importance to test the integrity and safety of
smart contract applications before they are actually used in
conjunction with the blockchain.
As a result of recent research on vulnerabilities
in smart contracts, representative vulnerabilities were
introduced [7], [8], and various countermeasures were
-----
proposed [5], [6], [8]–[10]. However, the effectiveness of
the countermeasures has not been properly studied. Most
performance evaluations have been performed using unlabeled data. Hence, their performance comparisons are often
not conclusive since when using unlabeled data, even if a
countermeasure detected some vulnerabilities, it is not clear
whether they were actual bugs or false positives, and also
how many vulnerabilities each countermeasure missed. For
example, recently, the authors of [10] used 47,518 smart
contracts for comparative studies, but reported the test results
without confirming that vulnerabilities actually exist in the
smart contracts. Only 69 contracts having 115 vulnerabilities
in total were used as labeled data, resulting in that it is not
clear what the most effective countermeasures are for each
type of vulnerability. Besides, a comparative study itself may
be cumbersome and time-consuming tasks since different
countermeasures may require different environments to run,
and the analysis results in different formats should be additionally arranged.
In this paper, a new software tool is designed to facilitate
validation of the existing/new countermeasures, into which
the user can easily add new countermeasures and labeled
benchmark datasets. In particular, it automatically analyzes
the results of executing the countermeasures on the available benchmark datasets, and shows their performance using
tables and graphs under various performance measures to
facilitate easy comparison. To this end, the proposed tool
is implemented using OS-level virtualization and operates
within a Docker container, allowing it to operate independently of the underlying system and eliminating the need
for the user to perform separate installation/execution for
each countermeasure. As a result, new countermeasures (as
well as labeled benchmark smart contract codes) can easily
be included and evaluated in the proposed tool, and their
performance can effectively be cross-checked using various
metrics. Using the proposed tool and 237 labeled smart contract codes, we evaluate the representative existing countermeasures in the literature.
The evaluation results show that, in general, the countermeasures identifying more vulnerable code produce a much
larger number of false positives, resulting in very low precision and accuracy. The effectiveness of countermeasures
against ‘Access Control’, ‘Denial of Service’, and ‘FrontRunning’ are questionable. The F1-scores of all countermeasures are less than 25%. Vulnerable codes with ‘Integer
Overflow/Underflow’ and ‘Timestamp Dependence’ can be
completely detected. However, the performance of the countermeasures needs to be further improved in order to reduce
false positives. As for the vulnerabilities of ‘Reentrancy’
and ‘Unchecked Low Level Call’, we confirm that there
are effective countermeasures that show both high precision
and high recall values. We believe that the proposed tool
will contribute to a future verification study of smart contracts and development of practical and secure smart contract
applications.
The main contributions are summarized as follows
:
- The nine representative vulnerabilities discussed in the
Decentralized Application Security Project (or DASP)
Top 10 of 2018 [7] and their state-of-the-art countermeasures are revisited.
- The limitations of the current countermeasures are discussed from the perspective of practicality, and an effective software tool that can evaluate the performance of
the countermeasures with great convenience is designed.
- The proposed tool is implemented using an OS-level
virtualization technique, and is open to the public via
https://github.com/93suhwan/uscv.
- The proposed tool eliminates the need to manage a separate installation/execution environment for each countermeasure and provides easy comparative analysis,
helping to identify the most effective countermeasures
for each type of vulnerability.
- Using the proposed tool and 237 labeled data, we conduct a comparative study for the representative existing
countermeasures in the literature, and their performance
is represented using various performance measures.
The rest of the paper is organized as follows. Section II
introduces Ethereum smart contracts and their vulnerabilities.
Section III summarizes the state-of-the-art countermeasures
for the vulnerabilities of smart contracts. In Section IV,
we introduce the design of the proposed tool and discuss
the results for evaluating the performance of the existing
countermeasures using the proposed tool. Finally, Section VI
concludes the paper.
**II. PRELIMINARIES**
_A. BLOCKCHAIN AND ETHEREUM SMART CONTRACTS_
The first blockchain was introduced in 2008 by a pseudonymous person or group known as Nakamoto [1]. Essentially,
a blockchain is a list of blocks that record information. Since
blocks in the chain are connected using a cryptographic hash
(more specifically, in the way that each block contains the
cryptographic hash of the previous block in the chain), any
information of a block in the chain can be changed only if all
of its subsequent blocks can also be modified. However, since
such modifications require the consent of the majority of the
network, consequently, malicious change in the blockchain is
almost impossible.
Since a blockchain can work as a distributed, verifiable
public ledger that records transactions, its first application
was a cryptocurrency, named Bitcoin [1]. In order to add
a new block to the blockchain, Bitcoin uses a consensus
mechanism called Proof-of-Work (PoW), where nodes in the
network compete for generating a right block by solving
a cryptographic puzzle. Extended from Bitcoin, Ethereum
allows to store computer code that can be used to implement unforgeable decentralized applications [2], which is
now being a building platform for running various kinds of
DApps.
-----
A smart contract is a piece of code that enables DApps
to interact with the blockchain, and it actually runs on a
quasi-Turing-complete virtual machine, called EVM. EVM
is considered as a sort of distributed machine that executes
smart contracts that embed the DApp logic by consuming
Ether (Gas in EVM). Since the blockchain has a property
of immutability, once a smart contract is deployed on the
blockchain, it cannot be modified like other transactions.
Therefore, it is significant to test the integrity and safety
of smart contracts before they are actually used upon the
blockchain. Otherwise, the blockchain must be destroyed or
hardforked if serious errors in the deployed smart contracts
are found afterwords.
_B. VULNERABILITIES OF SMART CONTRACTS_
This section briefly reviews the nine representative vulnerabilities discussed in the DASP Top 10 of 2018 [7].
- Reentrancy (Vre): Before a contract is completed (i.e.,
resolving any effects), the contract is executed recursively or other contracts are invoked to make the state
in a mess. Below is an example scenario that exploits a
reentrancy vulnerability.
function withdraw(){
- Transfer tokens to someone.
- Update balance.
}
1. The attacker invokes the function
**withdraw in succession.**
2. The second function call is done
before balance has not been updated
for the first function call.
3. balance is updated only for
the second function call.
- Access Control (Vac): Contract’s private values or functions are accessed abnormally due to an insecure visibility setting. Below is an example that describes an
access control vulnerability. This function does not
check whether the function was already called and the
state has already been initialized.
function initState(){
**owner = msg.sender**
}
1. The function can be called
abnormally via a delegatecall.
2. Then, the value of owner could
be manipulated.
- Integer Overflow/Underflow (Vio): Solidity uses variables of unsigned int type. If programmers process variables of unsigned int type as if they were the variables
of signed int type, an overflow and underflow can occur.
If such errors happen, for example, a wrong amount
of tokens can be withdrawn. Below is an example that
shows an integer underflow.
function withdraw(uint amount){
if(balance - amount > 0){
- Withdraw tokens.
}
}
1. Suppose balance = $0$ and amount = 1.
2. The value of (balance - amount) can
be interpreted positive since
**balance is a value of unsigned int.**
- Unchecked Low Level Call (Vuc): When errors happen
in low level functions in Solidity, a boolean value set to
false is returned, but the code keeps running. Therefore,
the result of such low level functions should be checked
to confirm successful execution. Below is an example
that shows an unchecked low level call vulnerability.
function withdraw(uint amount){
- balance is updated
(i.e., balance -= amount).
- Transfer tokens (as many as amount)
by calling a send function.
}
1. If send function call fails, balance
is managed incorrectly.
- Denial of Service (DoS, Vdos): When DoS attacks are
launched, smart contracts can be unavailable. Various
types of DoS implementation have been reported including increasing gas necessary, abusing access control, and
maliciously behaving. Below is an example that shows
a sort of DoS. Computation at each block is limited
by the upper bound of the amount of gas in Ethereum.
If the function (doSomething), called by the attacker,
has a heavy code that consumes too much gas, other
transactions cannot be included in the block.
function doSomething(){
for(uint i = 0; i < N; i++){
- Heavy code.
}
}
1. Attackers call doSomething.
2. Too much gas is consumed using
heavy code in doSomething.
3. Other transactions cannot be
included in the block since the gas
limit for the block is reached.
- Bad Randomness (Vbr ): Generation of a random number
is required in several applications such as games and
lotteries. However, it is tricky to implement a random
-----
number generation on the Ethereum public blockchain
since 1) Ethereum is a deterministic Turing machine
without embedding true randomness, and 2) all the data
(block variables) used for generating a random number
is open to public, even to attackers. Hence, an attacker
can predict the sources of randomness to some extent
and replicate it to attack the function relying on the random value. Obviously, instead of using block variables
open to the public, a random value can be created using
timestamps. However, as discussed below for timestamp
dependence, the timestamps can be manipulated by miners, resulting in another type of attacks. Below is an
example that exploits a bad randomness vulnerability.
function coinFlip(bool guess){
value = a hash value generated using
a block number
side = value / given denominator
if (side == guess){
- Win the game.
}
}
1. Attackers can always win using the
function exploit below.
2. Copy the function coinFlip and get
the result in advance (A).
3. Call the function coinFlip based on
the result (B).
function exploit(bool guess){
(A)
value = a hash value generated using
the same block number
side = value / given denominator
(B)
if(side == guess){
conFlip(guess);
} else{
conFlip(!guess);
}
}
- Front-Running (Vfr ): Miners perform calculation while
being compensated for the gas. The more gas (higher
fees), the more quickly the transactions can be computed. Since the public Ethereum is transparent, pending
transactions are visible to anyone. Hence, attackers can
preempt the results of an already calculated transaction by copying the transaction at a higher fee. Below
is an example scenario that exploits a front-running
vulnerability.
1. Information sent in a transaction Ta
(Ether, recipient address) is
public.
2. Time elapses until Ta is confirmed.
3. Ta is read by an attacker before it
is confirmed.
4. The attacker’s transaction Tb, which
is generated by copying Ta, is placed
before Ta.
5. The attacker can steal the result of
computing Ta.
- Timestamp Dependence (Vtd ): The timestamp of a block
is determined by the miner (they reports the time at
which the mining occurs). However, it can be manipulated by the miner (the timestamp can be changed
within 15 seconds). Hence, fake time can be advertised
by malicious miners, which allows the output of the
contract to be changed.
- Short Address (Vsa): When a contract receives data
of smaller-than-expected size, the missing portion is
padded to zeros in EVM. For example, if the user
address, signature, and the amount of token to be withdrawn are 0 12345600, 0xabcdef12, and 32(0
× ×
00000020), respectively, EVM concatenates all the
values in the order of signature, address, and token
amount, resulting in 0xabcdef121234560000000020.
If an attacker specifies a short address such as 0
×
123456 instead of 0 12345600, in the previous case,
×
EVM generates a value of 0xabcdef1212345600000020
and two zeros are padded at the end, resulting
in 0xabcdef121234560000002000. The resulting value
can be misinterpreted as having to withdraw as many
tokens as 0 00002000.
×
**III. COUNTERMEASURES USING STATIC AND**
**DYNAMIC ANALYSIS**
In this section, we briefly examine 11 publicly available,
open-sourced, representative countermeasures for Ethereum
smart contracts with a command-line interface (CLI). Table 1
summarizes the characteristics of the considered countermeasures such as the main methods, input, and DASP Top
10 vulnerabilities supported. Among the main methods used
by the countermeasures, static analysis refers to any kind of
methods for examining and analyzing the code without actually executing it, whereas dynamic analysis refers to those
for testing and evaluating the code by running it with test
cases. Typical static analysis includes abstract interpretation,
control-flow analysis, data-flow analysis, symbolic execution, etc., whereas dynamic analysis includes code coverage,
memory error detection, fault localization, security analysis,
etc. Static analysis is faster but less precise than dynamic
analysis. In addition, static analysis finds properties that hold
for all execution paths, whereas dynamic analysis finds those
for one or more execution paths, but can detect subtle or
complex vulnerabilities that static analysis may not detect.
Below we review each countermeasure in alphabetical order
of their names.
-----
**TABLE 1. Overview of the countermeasures considered in our proposed tool. We considered only publicly available, open-sourced countermeasures with**
a CLI. Year denotes the publication year of the first relevant conference, workshop, or journal paper, if any. Vulnerabilities denote either those that can
be detected by the given countermeasure (that is, the countermeasure implements a detector for the specified vulnerability) or its functionalities if the
countermeasure is a testing tool, linter, or profiler. Vac : Access control; Vdos: Denial of service; Vfr : Front-running; Vio: Integer overflow/underflow; Vre:
Reentrancy; Vtd : Timestamp dependence; Vuc : Unchecked low level call.
_A. ECHIDNA_
Echidna [11], [12] is an open-source, easy-to-use, propertybased fuzz testing tool for Ethereum smart contracts, developed and used by Trail of Bits. Instead of using a predefined
set of rules to detect vulnerabilities, it supports user-defined
properties for property-based testing [30], arbitrary assertion checking, and estimation of maximum gas usage. That
is, it automatically generates tests to detect violations in
user-defined properties and assertions, and allows us to prevent vulnerabilities caused by out-of-gas conditions. Echidna
uses the Slither static analysis tool [22], which we discuss
below, in the preprocessing step to compile and analyze
smart contracts and use information from Slither to improve
fuzz testing. Currently, Echidna can also test contracts compiled with Vyper (https://vyper.readthedocs.io/en/stable/)
and supports smart contract development frameworks
such as Truffle (https://www.trufflesuite.com/) and Embark
(https://framework.embarklabs.io/).
_B. ETHLINT_
Ethlint [13], formerly known as Solium, is a customizable, stand-alone linter for Solidity smart contracts. It provides a predefined set of various style and security rules,
which the user can configure, for example, by choosing which rules to apply to the code or by passing
options to the rules to modify their behavior. Ethlint
was originally designed to strictly adhere to the Solidity
style guide (https://solidity.readthedocs.io/en/develop/styleguide.html), but now it allows the user to not only customize
the predefined rules but also write and distribute via NPM
new plugins for their own rules. It can also automatically
fix the detected style and security issues, but there is no
benchmark result.
_C. MANTICORE_
Manticore [14], [15] is an open-source dynamic symbolic
execution framework not only for Ethereum smart contracts but also for native binaries. It consists of the Core
Engine implementing a generic platform-independent symbolic execution engine, the Native and Ethereum Execution Modules for symbolic execution of binaries and
smart contracts, respectively, and the Satisfiability Modulo
Theories (SMT) module and a Python API for supporting a customized analysis and interacting with external solvers such as Z3 (https://github.com/Z3Prover/z3),
Yices (https://yices.csl.sri.com/), and CVC4 (https://cvc4.
github.io/). Currently, Manticore supports various built-in
vulnerability detectors such as for problematic uses of
delegatecall, integer overflows, reentrancy bugs, uses of
potentially insecure instructions, reachable external calls,
reachable selfdestruct instructions, uninitialized memory and
storage usage, invalid instructions, and unused internal transaction return values. The main downside of using Manticore
is its long execution time; it is very much slower than other
static analysis tools (while it took about 24 minutes on
average, other tools just took from a few seconds to a few
minutes under experiments using 47,518 contracts) [10].
_D. MYTHRIL_
Mythril [16], [17] is an open-source, interactive, security analysis tool for Ethereum smart contracts, which
also supports other EVM-compatible blockchains
such as Quorum (https://consensys.net/quorum/), VeChain
-----
(https://www.vechain.org/), and Tron (https://tron.network/).
It is one of the earliest developed automated smart contract
analysis tools and can be used to detect various security
vulnerabilities such as use of delegatecall to untrusted contracts, integer overflows/underflows, and multiple sends in
a single transaction. It uses various program analysis techniques such as symbolic execution, SMT constraint solving,
taint analysis and control flow checking to detect such vulnerabilities. Mythril has been shown to be highly accurate
in detecting the DASP Top 10 vulnerabilities when compared with other tools [9], [10]. It can also be used in a
commercial SaaS smart contract security analysis platform
called MythX (https://mythx.io/) which is more optimized
and provides a wider range of functionalities.
_E. OYENTE_
Oyente [18], [19] is one of the first Ethereum smart contract
analysis tools, which has served as a basis for the design
and development of other tools such as HoneyBadger [31],
Maian [32], and Osiris [33]. It performs symbolic execution
and SMT constraint solving using the Z3 theorem prover
to analyze EVM bytecode and detect various vulnerabilities. The authors of [18] conducted an experiment using
existing 19,366 Ethereum smart contracts and reported that
Oyente identified 8,833 contracts as vulnerable. However,
several recent studies [9], [10] revealed that Oyente produces
a considerable number of false positives, in particular, due
to the integer overflow/underflow vulnerability, as is also
discussed in Section IV-B. That is, Oyente is not appropriate
for detecting arithmetic vulnerabilities. We also remark that
while Oyente currently reports a call stack depth attack vulnerability, it is no longer possible as of the EIP 150 hardfork.
_F. SECURIFY_
Securify [20], [21] is a security analysis tool for Ethereum
smart contracts, which currently supports more than 37 vulnerabilities including reentrancy, locked Ether, transaction
order dependence, and unrestricted write. Together with an
input contract, it takes as input a set of security patterns written in a specialized domain-specific language. More specifically, a security property is encoded into a set of compliance
and violation patterns, each of which ensures that a contract
satisfies and violates the given property, respectively. Such
patterns are checked using the Soufflé Datalog solver [34]
against the semantic facts obtained from the contract by
applying static analysis such as data- and control-flow analysis. In contrast to symbolic execution-based tools such as
Mythril [16] and Oyente [18], which do not guarantee to
explore every program path, Securify analyzes every contract
behavior, thus avoiding false negatives. Securify aims to
guarantee that if a contract matches a compliance (resp. violation) pattern, then it definitely complies with (resp. violates)
the corresponding security property. However, as discussed
in [35], most of the security patterns proposed in [20] are
not sound and can produce both false positives and false
negatives.
_G. SLITHER_
Slither [22], [23] is an open-source Solidity static analysis
framework written in Python 3, which supports automated
detection of about 45 vulnerabilities and code optimizations
that the compiler misses, and visualization of the information
about contract details, enhancing developers’ code comprehension. Given a Solidity contract source code, Slither takes
as input its abstract syntax tree generated by the Solidity compiler, and recovers its inheritance graph, control flow graph,
and list of expressions. Then, Slither transforms the contract code into an intermediate representation called SlithIR,
which uses static single assignment form [36] to facilitate
the analysis, and applies the usual program analysis techniques such as data-flow analysis and taint tracking. The
authors of [22] compares Slither with other static analysis tools such as Securify [20], SmartCheck [24], and Solhint [26] with respect to their capability to detect reentrancy
vulnerabilities using 1,000 contracts obtained from Etherscan (https://etherscan.io/), and show that Slither outperforms
the other tools for detecting reentrancy vulnerabilities with
respect to performance, robustness, and accuracy.
_H. SMARTCHECK_
SmartCheck [24], [25] is an efficient static analysis tool
for Ethereum smart contracts to detect security vulnerabilities and other code issues. It uses an XML-based intermediate representation (IR) to which Solidity source code
is translated. Potential vulnerabilities are then detected by
applying XPath [37] patterns on the generated IR. Although
SmartCheck is very fast when compared with other analysis
tools [10], since it only performs relatively simple lexical and
syntactic analysis, it cannot detect some severe bugs requiring
more advanced techniques such as taint analysis. It has also
shown that SmartCheck produces a large number of false
positives in the experiment on the reentrancy vulnerability
detection using 1,000 contracts [22]. An online version of
SmartCheck with more security patterns than the GitHub
version is available at https://tool.smartdec.net/.
_I. SOLHINT_
Solhint [26] is an open-source linter for Solidity smart contracts, similar to Ethlint [13]. It can be used not only to
validate if the Solidity code complies with the style guide
and best coding practices but also to detect syntax-related
security vulnerabilities. In addition, the user can customize
the predefined rule sets and add new rules if necessary. Solhint has shown to be fast and robust, but produce a large
number of false positives in the experiment on the reentrancy
vulnerability detection [22].
_J. SOL-PROFILER_
Sol-profiler [27] is a CLI tool to help the user to visualize
and review Solidity smart contracts by listing down various
properties of every contract method. More specifically, for
each method, it specifies the contract, interface, or library
-----
to which it belongs, its name and parameter types, its visibility (external, public, internal, or private), if it is a view
or pure function, its return type, and its modifiers. Therefore, by using Sol-profiler, the user can easily identify the
properties of the contract methods and check if there are
errors. However, Sol-profiler does not guarantee any security
properties of smart contracts.
_K. VERISMART_
VeriSmart [28], [29] is a highly precise verification tool for
detecting arithmetic bugs such as an integer overflow and
underflow in Ethereum smart contracts. It automatically discovers the transactions invariants of smart contracts, which
enable to analyze them effectively and exhaustively. More
precisely, it iteratively generates candidate transaction invariants and validates them using an off-the-shelf SMT solver
as in the usual counter example-guided inductive synthesis (CEGIS) framework [38]. By experimentally comparing
VeriSmart with other analysis tools that can detect arithmetic
bugs such as Manticore [14], Mythril [16], Osiris [33], and
Oyente [18], using 60 contracts that contains arithmetic vulnerabilities [39], the authors of [28] show that VeriSmart
far outperforms the abovementioned analyzers and detects
all arithmetic bugs with a negligible false positive rate.
Since VeriSmart outperforms Osiris, which can detect only
integer-related bugs, we do not include the latter in our proposed evaluation tool.
**IV. THE PROPOSED EVALUATION TOOL**
_A. DESIGN OF THE PROPOSED TOOL_
A number countermeasures have been introduced to detect
vulnerabilities in smart contract applications, as mentioned
in the previous section, but their effectiveness has not been
studied well. It is even not clear which countermeasures
are most effective for each type of vulnerability. When new
countermeasures are proposed, it is definitely necessary to
conduct comparative performance evaluation with existing
ones. However, since different countermeasures could require
different environments for installation and execution (e.g.,
different versions of the Solidity compiler and Z3 theorem
prover, etc.) and their verification outputs are produced in
different formats, it is not a simple task to perform a comparative study of countermeasures and compare the analysis
results. In particular, when new datasets are available, one
needs to re-execute all available countermeasures, preprocess
their verification outputs, and analyze the result under various
performance measures. To avoid such time-consuming tasks,
we provide a software tool that can
- easily be extended with existing/new countermeasures
and labeled smart contract codes,
- facilitate comparison of the countermeasures by automatically analyzing their verification outputs in terms of
various performance measures and arranging the results
in tables and graphs, and thus
- help the user to identify the most effective countermeasures for each vulnerability.
**FIGURE 1. Overall structure of the proposed evaluation tool.**
Fig. 1 shows the overall structure of our proposed evaluation tool. In the proposed tool, each countermeasure is offered
in the form of a Docker image using OS-level virtualization
and operates within the Docker container, making it easy
to meet all operational requirements. This approach helps
to effectively manage the use of computational resources
(CPU, memory) in the system, since each countermeasure is
containerized only while actually analyzing the target code.
A set of different versions of the Solidity compiler is provided
as a single Docker image and can be used in different containers where countermeasures operate. This design is effective
because it eliminates the need to update all Docker images
of the existing countermeasures when a new version of the
compiler is required to convert a new target code into binary
code.
The analyzer module analyzes the verification outputs
generated by each countermeasure and demonstrates the
verification performance using various statistical indicators.
More precisely, it preprocesses the verification outputs of
each countermeasure to check if some vulnerabilities were
found in each code in the benchmark dataset. Then, for
each countermeasure and its verification outputs, the analyzer module automatically computes various performance
measures such as the numbers of true positives, false positives, true negatives, and false negatives, precision, recall,
accuracy, F1-score, and the area under the curve (AUC),
as shown in Table 2. Finally, it organizes the analysis results
and presents them using tables and graphs as shown in
Fig. 2, 3, and 4, making it easy to conduct comparative studies
for each type of vulnerability. In particular, since the verification performance is represented using various performance
measures, users can identify the most effective countermeasure according to their own interests. (Here we note that,
different users can place higher importance on different measures. For example, some users may give higher priority to
countermeasures that maximize the number of true positives
than those that have the minimum number of false positives,
while others may prefer the opposite.) This feature can be
useful if the proposed tool is used to verify smart contract
codes in practice. In addition to selective application of each
countermeasure, an effective subset of countermeasures can
automatically be selected/recommended depending on the
target vulnerabilities, user interests, and constraints.
-----
**TABLE 2. Performance measures used in the proposed tool. The value of**
precision, recall, accuracy, and F1-score ranges from 0 to 100, and that of
AUC ranges from 0 to 1.
**FIGURE 2. Example results of using the proposed tool.**
_B. COMPARATIVE STUDY USING THE PROPOSED TOOL_
Using the proposed tool, we evaluated the performance
of the state-of-the-art countermeasures with 237 pieces of
labeled code collected from the SWC registry (Smart Contract Weakness Classification and Test Cases) [40], SmartBugs SB curated dataset [41], VeriSmart-benchmarks [39],
Zeus dataset [42], and eThor dataset [43]. Each code either
has a single type of vulnerability or is known to be secure,
i.e., without any vulnerability. The number of pieces of code
**TABLE 3. The number of smart contracts for testing each countermeasure**
for each type of vulnerability. Secure denotes smart contracts having no
vulnerability.
**TABLE 4. The number of true positives (TP).**
**TABLE 5. The number of false positives (FP).**
for each vulnerability is arranged in Table 3. The proposed
tool arranges the evaluation results in a unified manner as
shown in Fig. 2 and produces graphs for each measure as
shown in Fig. 3 and 4, which allows an easy comparative
study and cross-validation among the countermeasures.
Tables 4–9 respectively shows the TP, FP, precision, recall,
accuracy, and F1-score of each countermeasure for each type
of vulnerability for the dataset described in Table 3. We omit
the TN and FN as they are easily obtained from the FP and
TP, respectively. In the tables, ‘-’ means that the countermeasure does not support the detection of the corresponding vulnerability. In Table 4, TOTAL represents the number
of smart contracts having the corresponding vulnerability,
whereas in Table 5, it represents the number of those not
having the corresponding vulnerability. We additionally show
the F1-score of each countermeasure for each type of vulnerability in Fig. 3, which takes both precision and recall
into consideration and thus is a more appropriate metric for
imbalanced datasets. As our dataset is highly imbalanced,
the F1-score is much lower than the accuracy for every case,
but it is much more useful than the accuracy for comparing
the performance of various countermeasures.
Overall, for AC, every countermeasure shows a low detection rate of vulnerable code. That is, for all countermeasures,
-----
**FIGURE 3. F1-score of each countermeasure for each vulnerability.**
**TABLE 6. Precision (%).**
**TABLE 7. Recall (%).**
more than 80% of vulnerable codes are not detected. Mythril
detects three of the 18 vulnerable codes, showing the largest
TP value. However, considering the FP value together, Slither,
which represents a slightly smaller TP value, can be more
effective since it shows 100% precision and better accuracy.
**TABLE 8. Accuracy (%).**
**TABLE 9. F1-score (%).**
Only Mythril and Oyente work for DoS and FR, respectively. One out of six vulnerable codes with DoS is detected
by Mythril, and two out of four vulnerable codes with FR
are detected by Oyente. Both countermeasures produce more
false positives (compared to the true positives), resulting in
-----
small values in both precision and recall. These results can be
interpreted that neither countermeasure makes a sufficiently
meaningful contribution to DoS and FR detection.
VeriSmart successfully detects all vulnerable codes with
IO and reports a 100% recall value. However, it also generates 55 false positives, resulting in the precision of 50%.
In general, a large number of false positives require additional
manual examinations, which can be a large overhead. In that
sense, Manticore that detects six out of 55 vulnerable codes
without generating any false positives (100 % precision) may
be preferred.
Oyente may be the most effective tool for detecting RE,
as shown by the highest value in F1-score (Fig. 3) and good
performance for all measures (refer to Tables 6, 7 and 8).
More specifically, 29 out of 31 vulnerable codes are detected
while only six false positives are produced (among 206 secure
codes). Slither may be considered competitive in that it completely detects every vulnerable code (even though it produces 32 false positives, showing 49.2% precision and 100%
recall).
It is confirmed that both Slither and Solhint completely
detect five vulnerable codes having TD. However, Slither
can be recommended more preferentially since it produces
about half of the false positives in Solhint. As for UC, it is
reported that all vulnerable codes are detected by SmartCheck
which produces 14 false positives, resulting in the precision
of 78.8% and F1-score of 88.1%. Slither is also effective
in detecting UC. It detects 45 out of 52 vulnerable codes
while generating only one false positive, showing good performance for all measures.
In Fig. 4, we also show the performance of the countermeasures using their ROC curve and AUC value. Since
our dataset is imbalanced, the AUC is also an important
measure to be considered. Note that the AUC values in Fig. 4
do not necessarily coincide with the F1-scores in Fig. 3.
Following the general guidelines in [44], we consider the
countermeasure to be acceptable, excellent, and outstanding
if its AUC value is greater than or equal to 0.7, 0.8, and
0.9, respectively. In this regard, there is no effective countermeasure for AC, DoS, and FR; Oyente is excellent and
VeriSmart is outstanding at identifying IO; Oyente and Slither
are outstanding for RE; Slither and Solhint are outstanding
for TD; and finally Slither and SmartCheck are outstanding
for UC.
The evaluation results can be summarized as follows
:
- In general, countermeasures that identify many vulnerable codes also tend to be less precise and accurate since
they also produce much more false positives.
- There is no effective countermeasure for detecting
the ‘Access Control’, ’Denial of Service’, and ‘FrontRunning’ vulnerabilities yet.
- Vulnerable codes with ‘Integer Overflow’ and ‘Timestamp Dependence’ are completely detected. However,
the performance of the countermeasures need to be further improved in order to reduce false positives.
- As for the ‘Reentrancy’ and ‘Unchecked Low Level
Call’ vulnerabilities, effective countermeasures with
both high precision and high recall values are identified.
**V. DISCUSSION**
_A. RESULTS FROM OUR COMPARATIVE STUDY_
Similar to ours, previous studies such as [9], [10], [22],
[28] have also empirically compared various countermeasures using real-world smart contracts and discussed their
performance. In [9], four countermeasures such as Mythril,
Securify, SmartCheck, and Oyente were evaluated using
10 representative smart contracts, and the results suggested
that SmartCheck was statistically most effective in terms of
accuracy and ROC, while Mythril had the least number of
false positives. This result is consistent with our evaluation
results in the case of UC. However, due to a limited number
of test codes, the effectiveness of Oyente against IO did
not seem to be well understood. The authors of [22] proposed Slither and conducted performance comparisons with
Securify, SmartCheck, and Solhint in RE detection using
two famous contracts (DAO and SpankChain), which are
vulnerable to RE, and 1,000 unlabeled contract data. They
reported that Slither overwhelmed the other three countermeasures in terms of accuracy, execution time, and robustness, which is the same as in our evaluation results. The
performance of Oyente, which was not covered in their work,
is newly verified in our work, and it is discussed that Oyente
could be more effective than Slither in that it generates a
much smaller number of false positives for RE detection.
In [28], the authors introduced VeriSmart, a new method
for IO detection, and compared its performance with that
of Manticore, Mythril, Osiris, and Oyente using 60 labeled
vulnerable contracts. Their evaluation showed that VeriSmart
successfully detected all vulnerable codes with a negligible
false positive rate (0.41%). In our study using an increased
number of codes, VeriSmart’s effectiveness could also be
confirmed. However, in our study, it incurred a much higher
false positive rate. In [10], comprehensive evaluation for the
nine representative countermeasures were performed using a
dataset of 69 labeled vulnerable contracts and 47,518 unlabeled contracts. The authors reported that Mythril was the
most accurate countermeasure, showing 27% accuracy, when
considered the vulnerabilities altogether. However, this report
is slightly different from our findings. In our work, which
used more labeled codes and measured various evaluation
metrics for each vulnerability, Mythril is not recommended
due to its low value of precision and recall.
Obviously, our experimental results are partially consistent
and complementary with those from previous studies mentioned above. Here, we note that in our work, more countermeasures are evaluated with much more labeled codes, and
their performance is shown with various measures. Hence,
we believe that our comparative study could provide more
reliable insights into the state-of-the-art countermeasures and
-----
**FIGURE 4. ROC curve and AUC of each countermeasure for each vulnerability.**
help developers choose countermeasures that better suit their
purpose under given conditions.
_B. THREATS TO VALIDITY_
The limitations of our evaluation are summarized as follows.
In this work, we collected more labeled data than previous
studies to derive more reliable evaluation results. However,
as in other related work, a few smart contracts may have
incorrect labels, as it is very challenging to manually examine the code. Moreover, our dataset is imbalanced in that
the number of safe smart contracts is much larger than the
number of vulnerable smart contracts. In particular, among
the 237 smart contracts, only six, four, and five vulnerable
contracts for DoS, FR and TD are included, respectively.
Hence, in most cases, the detection accuracy of countermeasures against each vulnerability is highly reported. Since
new datasets and countermeasures can be easily added to
our tool, however, we believe that our tool can contribute to
achieving more accurate evaluation results with more data in
the future.
**VI. CONCLUSION**
In this paper, we revisited smart contracts using the Ethereum
blockchain technology and summarized various vulnerability
issues of smart contract applications. A number of countermeasures were briefly introduced and discussed. To assess
the effectiveness of the countermeasures, we designed and
implemented a software tool that facilitates comparative
evaluations of the countermeasures. Using the tool and
237 labeled benchmark codes, we evaluated state-of-the-art
vulnerability detection schemes. The evaluation results indicate that countermeasures that exhibit a larger TP value often
generate a much larger number of false positives, resulting
in very low precision and accuracy. In addition, among the
state-of-the-art countermeasures, Oyente and Slither are most
effective for RE detection; Slither could be recommended
for detection of TD and UC; and VeriSmart could be recommended for IO detection. Using our tool, researchers can
easily conduct performance comparisons between their own
countermeasure and other state-of-the-art schemes with a
variety of performance metrics. As for practitioners, they
can exploit our tool to find various vulnerabilities within
their smart contract applications. Since in our tool, smart
contracts can be examined by a number of countermeasures simultaneously, vulnerabilities can be easily identified.
We believe that our proposed tool will be effective in a
future verification study of smart contracts and will contribute
to the development of practical and secure smart contract
applications.
**APPENDIX. USAGE OF THE PROPOSED TOOL**
The proposed software tool is open to public via the website
https://github.com/93suhwan/uscv. This section details how
to use it.
-----
_A. INSTALLATION_
As mentioned in Section IV-A, each countermeasure is
included in the proposed software tool in the form of a Docker
image. To create a Docker image, a dockerfile can be run
under the following command:
$ docker build [dockerfile]
The content of a dockerfile is as follows:
From ubuntu:18.04
RUN [installCommand]
ENTRYPOINT [exeCommand]
/*
- 1st line indicates a layer is created
from the ubuntu:18.04 Docker image
- 2nd line is for building a
countermeasure by executing
[installCommand]
- 3rd line specifies the execution
command ([exeCommand]) for each
countermeasure (which would run by
default)
*[/]
Using the Docker image, a Docker container is generated
and executed under the following command:
$ docker build -t [dockerImage]
[dockerfile]
Since multiple versions of the Solidity compiler may be
required during the process of compiling the source files,
the proposed tool has a Docker image (.solc) that includes
different versions of the Solidity compiler. The installed
compilers can be used at each container under the following
command:
$ docker run -v [curDir]:[containerDir]\
[dockerImage] [options]
All of these processes for the 11 countermeasures discussed in Section III are executed automatically by running
‘‘createContainers.sh’’.
_B. EXECUTION_
The file named ‘‘testing.sh’’ is used to execute each countermeasure. It preprocesses the source file and determines the
version of the compiler that can be used for compiling. Then,
it can be run with the generalized options as follows:
$ testing.sh -t [schemeName] \
-f [srcFile] -l [timeout]
(e.g.,)
>> testing.sh -t oyente -f test.sol \
-o ‘‘-dl 10 -r’’ -l 100
The file named ‘‘execution.sh’’ is provided to run multiple countermeasures. ‘‘execution.sh’’ supports the following
options:
$ execution.sh -f/d [srcFile/dirName] \
-t [toolName]
(e.g.,)
>> execution.sh -d./curDir -v AC -l 100
The analysis results are recorded in the file named
tool_name.txt under the directory of ‘‘./result’’.
_C. ADDING NEW COUNTERMEASURES AND DATA_
When new countermeasures are proposed, they can easily be integrated into our proposed tool and evaluated with the embedded benchmark data. To this end,
the file named ‘‘addScheme.sh’’ is provided. By specifying
meta-information on a new countermeasure as arguments,
the new scheme can simply be included into the system.
$ addScheme.sh -l/n [dirName/imageName]\
-e [cmd] -o [option]\
-M [word]
(e.g.,)
>> addScheme.sh -l dockerfiles/
-----
smartcheck \
-e smartcheck -o p \
-a SOLIDITY_TX_ORIGIN \
-d SOLIDITY_OVERPOWERED_ROLE \
-i SOLIDITY_VAR|SOLIDITY_UINT_CANT\
-t SOLIDITY_EXACT_TIME \
-u SOLIDITY_UNCHECKED_CALL
The installed countermeasures can also be removed from
the tool as follows.
$ removeScheme.sh [countermeasureName]
(e.g.,)
>> removeScheme.sh mythril
When labeled codes are newly collected, they can be added
to our tool and used for the analysis.
$ addData.sh -d/f [dirName/fileName] \
-c [vulType]
(e.g.,)
>> addData.sh -f example.sol -c AC
**REFERENCES**
[1] S. Nakamoto, ‘‘Bitcoin: A peer-to-peer electronic cash system,’’ Tech.
Rep., 2008.
[2] G. Wood, ‘‘Ethereum: A secure decentralised generalised transaction
ledger,’’ Ethereum Project Yellow Paper, vol. 151, pp. 1–32, 2014.
[3] N. Atzei, M. Bartoletti, and T. Cimoli, ‘‘A survey of attacks on
Ethereum smart contracts (SoK),’’ in Principles of Security and Trust,
M. Maffei and M. Ryan, Eds. Berlin, Germany: Springer, 2017,
pp. 164–186.
[4] X. Li, P. Jiang, T. Chen, X. Luo, and Q. Wen, ‘‘A survey on the security of
blockchain systems,’’ Future Gener. Comput. Syst., vol. 107, pp. 841–853,
Jun. 2020.
[5] S. Sayeed, H. Marco-Gisbert, and T. Caira, ‘‘Smart contract: Attacks and
protections,’’ IEEE Access, vol. 8, pp. 24416–24427, 2020.
[6] H. Chen, M. Pendleton, L. Njilla, and S. Xu, ‘‘A survey on Ethereum
systems security: Vulnerabilities, attacks, and defenses,’’ ACM Comput.
_Surv., vol. 53, no. 3, pp. 1–3, Jun. 2020._
[7] NCC Group. Decentralized Application Security Project (or DASP) Top
_10. Accessed: Mar. 25, 2021. [Online]. Available: https://dasp.co/_
[8] M. di Angelo and G. Salzer, ‘‘A survey of tools for analyzing Ethereum
smart contracts,’’ in Proc. IEEE Int. Conf. Decentralized Appl. Infrastruc_tures (DAPPCON), Apr. 2019, pp. 69–78._
[9] R. M. Parizi, A. Dehghantanha, K.-K. R. Choo, and A. Singh, ‘‘Empirical
vulnerability analysis of automated smart contracts security testing on
blockchains,’’ in Proc. 28th Annu. Int. Conf. Comput. Sci. Softw. Eng.
_(CASCON), 2018, pp. 103–113._
[10] T. Durieux, J. A. F. Ferreira, R. Abreu, and P. Cruz, ‘‘Empirical review of
automated analysis tools on 47,587 Ethereum smart contracts,’’ in Proc.
_ACM/IEEE 42nd Int. Conf. Softw. Eng. (ICSE), New York, NY, USA,_
Jun. 2020, pp. 530–541.
[11] G. Grieco, W. Song, A. Cygan, J. Feist, and A. Groce, ‘‘Echidna: Effective,
usable, and fast fuzzing for smart contracts,’’ in Proc. 29th ACM SIGSOFT
_Int. Symp. Softw. Test. Anal., New York, NY, USA, Jul. 2020, pp. 557–560._
[12] Trail of Bits. Echidna: A Fast Smart Contract Fuzzer. Accessed:
Mar. 25, 2021. [Online]. Available: https://github.com/crytic/echidna
[13] Ethlint. Accessed: Mar. 25, 2021. [Online]. Available:
https://github.com/duaraghav8/Ethlint
[14] M. Mossberg, F. Manzano, E. Hennenfent, A. Groce, G. Grieco,
J. Feist, T. Brunson, and A. Dinaburg, ‘‘Manticore: A user-friendly symbolic execution framework for binaries and smart contracts,’’ in Proc.
_34th IEEE/ACM Int. Conf. Automated Softw. Eng. (ASE), Nov. 2019,_
pp. 1186–1189.
[15] Trail of Bits. Manticore. Accessed: Mar. 25, 2021. [Online]. Available:
https://github.com/trailofbits/manticore/
[16] B. Mueller, ‘‘Smashing Ethereum smart contracts for fun and real profit,’’
in Proc. 9th Annu. HITB Secur. Conf., 2018. Accessed: Mar. 25, 2021.
[Online]. Available: https://github.com/b-mueller/smashing-smartcontracts
[17] ConsenSys. Mythril. Accessed: Mar. 25, 2021. [Online]. Available:
https://github.com/ConsenSys/mythril
[18] L. Luu, D.-H. Chu, H. Olickel, P. Saxena, and A. Hobor, ‘‘Making smart
contracts smarter,’’ in Proc. ACM SIGSAC Conf. Comput. Commun. Secur.,
Oct. 2016, pp. 254–269.
[19] Oyente. Accessed: Mar. 25, 2021. [Online]. Available:
https://github.com/melonproject/oyente
[20] P. Tsankov, A. Dan, D. Drachsler-Cohen, A. Gervais, F. Bünzli, and
M. Vechev, ‘‘Securify: Practical security analysis of smart contracts,’’ in
_Proc. ACM SIGSAC Conf. Comput. Commun. Secur. (CCS), New York,_
NY, USA, 2018, pp. 67–82.
[21] Securify _2.0._ Accessed: Mar. 25, 2021. [Online]. Available:
https://github.com/eth-sri/securify2
[22] J. Feist, G. Greico, and A. Groce, ‘‘Slither: A static analysis framework for
smart contracts,’’ in Proc. 2nd Int. Workshop Emerg. Trends Softw. Eng.
_Blockchain (WETSEB), 2019, pp. 8–15._
[23] Trail of Bits. _Slither,_ _The_ _Solidity_ _Source_ _Analyzer._ Accessed:
Mar. 25, 2021. [Online]. Available: https://github.com/crytic/slither
[24] S. Tikhomirov, E. Voskresenskaya, I. Ivanitskiy, R. Takhaviev,
E. Marchenko, and Y. Alexandrov, ‘‘SmartCheck: Static analysis of
Ethereum smart contracts,’’ in Proc. 1st Int. Workshop Emerg. Trends
_Softw. Eng. Blockchain, May 2018, pp. 9–16._
[25] SmartCheck. Accessed: Mar. 25, 2021. [Online]. Available:
https://github.com/smartdec/smartcheck
[26] Solhint. Accessed: Mar. 25, 2021. [Online]. Available:
https://protofire.github.io/solhint/
[27] Sol-Profiler. Accessed: Mar. 25, 2021. [Online]. Available:
https://github.com/Aniket-Engg/sol-profiler
[28] S. So, M. Lee, J. Park, H. Lee, and H. Oh, ‘‘VeriSmart: A highly precise
safety verifier for Ethereum smart contracts,’’ in Proc. IEEE Symp. Secur.
_Privacy (SP), May 2020, pp. 1678–1694._
[29] VeriSmart. Accessed: Mar. 25, 2021. [Online]. Available: https://github.
com/kupl/VeriSmart-public
[30] K. Claessen and J. Hughes, ‘‘Quickcheck: A lightweight tool for random
testing of Haskell programs,’’ in Proc. 5th ACM SIGPLAN Int. Conf. Funct.
_Program. (ICFP), New York, NY, USA, 2000, pp. 268–279._
[31] C. F. Torres, M. Steichen, and R. State, ‘‘The art of the scam: Demystifying
honeypots in Ethereum smart contracts,’’ in Proc. 28th USENIX Conf.
_Secur. Symp. (SEC), 2019, pp. 1591–1607._
[32] I. Nikolić, A. Kolluri, I. Sergey, P. Saxena, and A. Hobor, ‘‘Finding the
greedy, prodigal, and suicidal contracts at scale,’’ in Proc. 34th Annu.
_Comput. Secur. Appl. Conf. (ACSAC), New York, NY, USA, Dec. 2018,_
pp. 653–663.
[33] C. F. Torres, J. Schütte, and R. State, ‘‘Osiris: Hunting for integer bugs
in Ethereum smart contracts,’’ in Proc. 34th Annu. Comput. Secur. Appl.
_Conf. (ACSAC), New York, NY, USA, 2018, pp. 664–676._
[34] H. Jordan, B. Scholz, and P. Subotić, ‘‘Soufflé: On synthesis of program
analyzers,’’ in Computer Aided Verification, S. Chaudhuri and A. Farzan,
Eds. Cham, Switzerland: Springer, 2016, pp. 422–430.
-----
[35] C. Schneidewind, I. Grishchenko, M. Scherer, and M. Maffei, ‘‘EThor:
Practical and provably sound static analysis of ethereum smart contracts,’’
in Proc. ACM SIGSAC Conf. Comput. Commun. Secur. (CCS), New York,
NY, USA, Oct. 2020, pp. 621–640.
[36] B. K. Rosen, M. N. Wegman, and F. K. Zadeck, ‘‘Global value numbers and
redundant computations,’’ in Proc. 15th ACM SIGPLAN-SIGACT Symp.
_Princ. Program. Lang. (POPL), New York, NY, USA, 1988, pp. 12–27._
[37] XML Path Language (XPath) 2.0. Accessed: Mar. 25, 2021. [Online].
Available: https://www.w3.org/TR/xpath20/
[38] A. Solar-Lezama, ‘‘Program sketching,’’ Int. J. Softw. Tools Technol.
_Transf., vol. 15, nos. 5–6, pp. 475–495, Oct. 2013._
[39] VeriSmart-Bsenchmarks. Accessed: Mar. 25, 2021. [Online]. Available:
https://github.com/kupl/VeriSmart-benchmarks
[40] SWC Registry (Smart Contract Weakness Classification and Test Cases).
Accessed: Mar. 25, 2021. [Online]. Available: https://swcregistry.io/
[41] SmartBugs: _A_ _Framework_ _to_ _Analyze_ _Solidity_ _Smart_ _Contracts._
Accessed: Mar. 25, 2021. [Online]. Available: https://github.com/
smartbugs/smartbugs
[42] S. Kalra, S. Goel, M. Dhawan, and S. Sharma, ‘‘ZEUS: Analyzing safety
of smart contracts,’’ in Proc. Netw. Distrib. Syst. Secur. Symp. (NDSS),
San Diego, CA, USA, 2018, pp. 1–12.
[43] eThor: Sound Static Analysis for Ethereum Smart Contracts. Accessed:
Mar. 25, 2021. [Online]. Available: https://secpriv.wien/ethor/
[44] D. W. Hosmer, S. Lemeshow, and R. X. Sturdivant, Applied Logistic
_Regression, 3rd ed. Hoboken, NJ, USA: Wiley, 2013._
SUHWAN JI received the B.S. and M.S. degrees
in computer science from Kangwon National University, South Korea, in 2017 and 2019, respectively, where he is currently pursuing the Ph.D.
degree majoring in AI and software with the Interdisciplinary Graduate Program in Medical Bigdata Convergence. His research interests include
programming languages, machine learning, and
blockchain.
DOHYUNG KIM received the B.S. degree
in information and computer engineering from
Ajou University, Suwon, South Korea, in February 2004, and the Ph.D. degree in computer science from the Korea Advanced Institute of Science
and Technology (KAIST), Daejeon, South Korea,
in August 2014. From 2014 to 2017, he was a
Postdoctoral Researcher and a Research Professor
with the Department of Computer Engineering,
Sungkyunkwan University. In 2018, he was an
Assistant Professor with the Department of Software and Computer Engineering, Ajou University. He is currently an Assistant Professor with the
Department of Computer Science and Engineering, Kangwon National University. His research interests include the design and analysis of computer
networking and wireless communication systems, especially for future Internet architectures.
HYEONSEUNG IM received the B.S. degree in
computer science from Yonsei University, South
Korea, in 2006, and the Ph.D. degree in computer
science and engineering from the Pohang University of Science and Technology (POSTECH),
South Korea, in 2012. From 2012 to 2015, he was
a Postdoctoral Researcher with the Laboratory for
Computer Science, Université Paris-Sud, and the
Tyrex Team, Inria, France. He is currently an Associate Professor with the Department of Computer
Science and Engineering, Kangwon National University, South Korea. His
research interests include programming languages, logic in computer science, big data analysis and management, machine learning, smart healthcare,
blockchain, and information security.
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.1109/ACCESS.2021.3091317?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/ACCESS.2021.3091317, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBYNCND",
"status": "GOLD",
"url": "https://ieeexplore.ieee.org/ielx7/6287639/9312710/09461797.pdf"
}
| 2,021
|
[
"JournalArticle"
] | true
| null |
[] | 14,468
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Mathematics",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/024b9198a432d51f0a5245bd60072e147fa2f9e1
|
[
"Computer Science"
] | 0.856956
|
Inverting Cryptographic Hash Functions via Cube-and-Conquer
|
024b9198a432d51f0a5245bd60072e147fa2f9e1
|
Journal of Artificial Intelligence Research
|
[
{
"authorId": "3326894",
"name": "O. Zaikin"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"JAIR",
"J Artif Intell Res",
"The Journal of Artificial Intelligence Research"
],
"alternate_urls": null,
"id": "aef12dca-60a0-4ca3-819b-cad26d309d4e",
"issn": "1076-9757",
"name": "Journal of Artificial Intelligence Research",
"type": "journal",
"url": "http://www.jair.org/"
}
|
MD4 and MD5 are fundamental cryptographic hash functions proposed in the early 1990s. MD4 consists of 48 steps and produces a 128-bit hash given a message of arbitrary finite size. MD5 is a more secure 64-step extension of MD4. Both MD4 and MD5 are vulnerable to practical collision attacks, yet it is still not realistic to invert them, i.e., to find a message given a hash. In 2007, the 39-step version of MD4 was inverted by reducing to SAT and applying a CDCL solver along with the so-called Dobbertin’s constraints. As for MD5, in 2012 its 28-step version was inverted via a CDCL solver for one specified hash without adding any extra constraints. In this study, Cube-and-Conquer (a combination of CDCL and lookahead) is applied to invert step-reduced versions of MD4 and MD5. For this purpose, two algorithms are proposed. The first one generates inverse problems for MD4 by gradually modifying the Dobbertin’s constraints. The second algorithm tries the cubing phase of Cube-and-Conquer with different cutoff thresholds to find the one with the minimum runtime estimate of the conquer phase. This algorithm operates in two modes: (i) estimating the hardness of a given propositional Boolean formula; (ii) incomplete SAT solving of a given satisfiable propositional Boolean formula. While the first algorithm is focused on inverting step-reduced MD4, the second one is not area-specific and is therefore applicable to a variety of classes of hard SAT instances. In this study, 40-, 41-, 42-, and 43-step MD4 are inverted for the first time via the first algorithm and the estimating mode of the second algorithm. Also, 28-step MD5 is inverted for four hashes via the incomplete SAT solving mode of the second algorithm. For three hashes out of them, it is done for the first time.
|
## Inverting Cryptographic Hash Functions via Cube-and-Conquer
**Oleg Zaikin** zaikin.icc@gmail.com
_ISDCT, Irkutsk, Russia_
### Abstract
MD4 and MD5 are seminal cryptographic hash functions proposed in early 1990s. MD4
consists of 48 steps and produces a 128-bit hash given a message of arbitrary finite size.
MD5 is a more secure 64-step extension of MD4. Both MD4 and MD5 are vulnerable to
practical collision attacks, yet it is still not realistic to invert them, i.e. to find a message
given a hash. In 2007, the 39-step version of MD4 was inverted via reducing to SAT and
applying a CDCL solver along with the so-called Dobbertin’s constraints. As for MD5,
in 2012 its 28-step version was inverted via a CDCL solver for one specified hash without
adding any additional constraints. In this study, Cube-and-Conquer (a combination of
CDCL and lookahead) is applied to invert step-reduced versions of MD4 and MD5. For
this purpose, two algorithms are proposed. The first one generates inversion problems for
MD4 by gradually modifying the Dobbertin’s constraints. The second algorithm tries the
cubing phase of Cube-and-Conquer with different cutoff thresholds to find the one with
minimal runtime estimation of the conquer phase. This algorithm operates in two modes:
(i) estimating the hardness of a given propositional Boolean formula; (ii) incomplete SATsolving of a given satisfiable propositional Boolean formula. While the first algorithm
is focused on inverting step-reduced MD4, the second one is not area-specific and so is
applicable to a variety of classes of hard SAT instances. In this study, 40-, 41-, 42-, and
43-step MD4 are inverted for the first time via the first algorithm and the estimating mode
of the second algorithm. 28-step MD5 is inverted for four hashes via the incomplete SATsolving mode of the second algorithm. For three hashes out of them this is done for the
first time.
### 1. Introduction
A cryptographic hash function maps a message of arbitrary finite size to a hash of a fixed
size. Such a function should have the following additional properties: (i) preimage resistance; (ii) second preimage resistance; (iii) collision resistance (Menezes, van Oorschot, &
Vanstone, 1996). The first property means that it is computationally infeasible to invert the
cryptographic hash function, i.e. to find any message that matches a given hash. According
to the second one, given a message and its hash, it is computationally infeasible to find
another message with the same hash. The third property means that it is computationally
infeasible to find two different messages with the same hash. A proper cryptographic hash
function must have all three types of resistance. Cryptographic hash functions are really
pervasive in the modern digital world. Examples of their applications include verification
of data integrity, passwords, and signatures.
It is well known that the resistance of a cryptographic hash function can be analyzed
by algorithms for solving the Boolean satisfiability problem (SAT) (Bard, 2009). SAT in its
decision form is to determine whether a given propositional Boolean formula is satisfiable
©2023 AI Access Foundation. All rights reserved.
-----
or not (Biere, Heule, van Maaren, & Walsh, 2021b). This is one of the most well-studied
NP-complete problems (Cook, 1971; Garey & Johnson, 1979). Over the last 25 years,
numerous crucial scientific and industrial problems have been successfully solved by SAT.
In almost all these cases, CDCL solvers, i.e. ones which are based on the Conflict-Driven
Clause Learning algorithm (Marques-Silva & Sakallah, 1999), were used.
Cube-and-Conquer is an approach for solving extremely hard SAT instances (Heule,
Kullmann, Wieringa, & Biere, 2011), for which CDCL solvers alone are not enough. According to this approach a given, problem is split into subproblems on the cubing phase
via a lookahead solver (Heule & van Maaren, 2021). Then on the conquer phase the subproblems are solved via a CDCL solver. Several hard mathematical problems from number
theory and combinatorial geometry have been solved by Cube-and-Conquer recently, e.g.,
the Boolean Pythagorean Triples problem (Heule, Kullmann, & Marek, 2016). However,
the authors of this study are not aware of any successful application of this approach to
cryptanalysis problems. This study is aimed at filling this gap by analyzing the preimage
resistance of the cryptographic hash functions MD4 and MD5 via Cube-and-Conquer.
MD4 was proposed in 1990 (Rivest, 1990). It consists of 48 steps and produces a 128-bit
hash given a message of arbitrary finite size. Since 1995 it has been known to be not collision resistant (Dobbertin, 1996). Despite this vulnerability, MD4 is still used to compute
password-derived hashes in some operating systems of the Windows family due to backwards compatibility requirements. Since MD4 still remains preimage resistant and second
preimage resistant, its step-reduced versions have been studied in this context recently. In
1998, the Dobbertin’s constraints on intermediate states of MD4 registers were proposed,
which reduce the number of preimages, but at the same time significantly simplify the inversion (Dobbertin, 1998). This breakthrough approach made it possible to easily invert
32-step MD4. In 2007, SAT encodings of slightly modified Dobbertin’s constraints were
constructed, and as a result 39-step MD4 was inverted via a CDCL solver (De, Kumarasubramanian, & Venkatesan, 2007) for one very regular hash (128 1s). Note, that it is a
common practice to invert very regular hashes such as all 1s or all 0s. Since 2007, several
unsuccessful attempts have been made to invert 40-step MD4.
The second studied cryptographic hash function, MD5, is a more secure 64-step version
of MD4 proposed in 1992 (Rivest, 1992). Thanks to elegant yet efficient designs, MD4
and MD5 have become one of the most influential cryptographic functions with several
notable successors, such as RIPEMD and SHA-1. MD5 is still used in practice to verify
the integrity of files and messages. Since 2005, MD5 has been known to be not collision
resistant (Wang & Yu, 2005). Because of a more secure design, the Dobbertin’s constraints
are not applicable to MD5 (Aoki & Sasaki, 2008). 26-step MD5 was inverted in 2007 (De
et al., 2007), while for 27- and 28-step MD5 it was done for the first time in 2012 (Legendre,
Dequen, & Krajecki, 2012). In both papers CDCL solvers were applied, at the same time
no additional constraints were added. In (Legendre et al., 2012), 28-step MD5 was inverted
for only one hash 0x01234567 0x89abcdef 0xfedcba98 0x76543210. This hash is a regular
binary sequence (it is symmetric, and the numbers of 0s and 1s are equal), but at the same
time it is less regular than 128 1s mentioned above. The same result was presented later in
two papers of the same authors. Unfortunately, none of these three papers explained the
non-existence of results for 128 1s and 128 0s. Since 2012 no further progress in inverting
step-reduced MD5 has been made.
2
-----
This paper proposes Dobbertin-like constraints, a generalization of Dobbertin’s constraints. Two algorithms are proposed. The first one generates Dobbertin-like constraints
until a preimage of a step-reduced MD4 is found. The second algorithm does sampling to
find a cutoff threshold for the cubing phase of Cube-and-Conquer with minimal runtime
estimation of the conquer phase. This algorithm operates in two modes: (i) estimating
the hardness of a given formula; (ii) incomplete SAT-solving of a given formula. The first
algorithm is MD4-specific, while the second algorithm in its estimating mode is general,
so it can be applied to any SAT instance (including unsatisfiable ones). Yet the incomplete SAT-solving mode is oriented only on satisfiable SAT instances, preferably with many
solutions.
With the help of the first algorithm and the estimating mode of the second algorithm,
40-, 41-, 42-, and 43-step MD4 are inverted for four hashes: 128 1s, 128 0s, the one from (Legendre et al., 2012), and a random hash. The first algorithm is not applicable to MD5 because
of its more secure design. The estimating mode of the second algorithm is not applicable
either because for an unconstrained inversion problem the cubing phase produces too hard
subproblems. Therefore only the incomplete solving mode of the second algorithm is applied to invert step-reduced MD5. In particular, 28-step MD5 is inverted for the same four
hashes. All the experiments were ran on a personal computer.
In summary, the contributions of this paper are:
- Dobbertin-like constraints, a generalization of Dobbertin’s constraints.
- An algorithm that generates Dobbertin-like constraints and the corresponding inversion problems to find preimages of a step-reduced MD4.
- A general algorithm for finding a cutoff threshold with the minimal runtime estimation
of the conquer phase of Cube-and-Conquer.
- For the first time, 40-, 41-, 42-, and 43-step MD4 are inverted.
- For the first time, 28-step MD5 is inverted for two most regular hashes (128 1s and
128 0s), and a random non-regular hash.
The paper is organized as follows. Preliminaries on SAT and cryptographic hash functions are given in Section 2. Section 3 proposes Dobbertin-like constraints and the algorithm
aimed at inverting step-reduced MD4. The Cube-and-Conquer-based algorithm is proposed
in Section 4. The considered inversion problems for step-reduced MD4 and MD5, as well
as their SAT encodings, are described in Section 5. Experimental results on inverting stepreduced MD4 and MD5 are presented in sections 6, 7, and 8. Section 9 outlines related
work. Finally, conclusions are drawn.
This paper builds on an earlier work (Zaikin, 2022), but extends it significantly in
several directions. First, the algorithm for generating Dobbertin-like constraints for MD4
is improved by cutting off impossible values of the last bits in the modified constraint. As
a result, in most cases the considered step-reduced versions of MD4 are inverted about 2
times faster than in (Zaikin, 2022). Second, the incomplete SAT-solving mode of the Cubeand-Conquer-based algorithm is proposed. Third, all considered step-reduced versions of
MD4 are inverted for four hashes compared to two hashes in (Zaikin, 2022). Finally, 28-step
MD5 is inverted, while in (Zaikin, 2022) only step-reduced MD4 was studied.
3
-----
### 2. Preliminaries
This section gives preliminaries on SAT, Cube-and-Conquer, cryptographic hash functions,
MD4, Dobbertin’s constraints, and MD5.
**2.1 Boolean Satisfiability**
_Boolean satisfiability problem (SAT) (Biere et al., 2021b) is to determine whether a given_
propositional Boolean formula is satisfiable or not. A formula is satisfiable if there exists a
truth assignment that satisfies it; otherwise it is unsatisfiable. SAT is historically the first
NP-complete problem (Cook, 1971). A propositional Boolean formula is in Conjunctive
_Normal Form (CNF), if it is a conjunction of clauses. A clause is a disjunction of literals,_
where a literal is a Boolean variable or its negation.
The Davis–Putnam–Logemann–Loveland (DPLL) algorithm is a complete backtracking
SAT solving algorithm (Davis, Logemann, & Loveland, 1962). A decision tree is formed,
where each internal node corresponds to a decision variable, while edges correspond to
variables’ values. Unit Propagation (UP (Dowling & Gallier, 1984)) is used to reduce the
tree after assigning a decision variable. UP iteratively applies the unit clause rule: if there
is only one remaining unassigned literal in a clause, and all other literals are assigned to
False, then this literal is assigned to True. If an unsatisfied clause is encountered, a conflict
is declared, and chronological backtracking is performed.
Lookahead is another complete SAT-solving algorithm (Heule & van Maaren, 2021). It
improves DPLL by the following heuristic. When a decision variable should be chosen, each
unassigned variable is assigned to True followed by UP, the reduction is measured, then
the same is done for False assignment. Failed literal denotes a literal for which a conflict is
found during UP. If both literals of a variable are failed, then unsatisfiability of the CNF is
proven. If there is exactly one failed literal l for some variable, then l is forced to be assigned
to False. This rule is known as failed literal elimination. If for a variable both literals are
not failed, the reduction measure for this variable is calculated as a combination of literalmeasures. A variable with the largest reduction measure is picked as a decision variable.
Thus lookahead allows one to choose good decision variables and in addition simplifies the
CNF by the described reasoning. Lookahead SAT solvers are strong on random k-SAT
formulae.
In contrast to DPLL, in Conflict-Driven Clause Learning (CDCL) when a conflict happens, the reason is found and non-chronological backtracking is performed (Marques-Silva
& Sakallah, 1999). To forbid the conflict, a conflict clause is formed based on the reason
and added to the CNF. Conflict clauses are used not only to limit the exploration of the
decision tree, but also to choose decision variables. This complete algorithm is much more
efficient than DPLL. Also, it is stronger than lookahead on non-random instances. That is
why most modern complete SAT solvers are based on CDCL.
Problems from various areas (verification, cryptanalysis, combinatorics, bioinformatics,
etc.) can be efficiently reduced to SAT (Biere et al., 2021b). When a cryptanalysis problem
is reduced to SAT and solved by SAT solvers, this is called SAT-based cryptanalysis or logical
_cryptanalysis (Cook & Mitchell, 1996; Massacci & Marraro, 2000). It is in fact a special type_
of algebraic cryptanalysis (Bard, 2009). In the last two decades, SAT-based cryptanalysis
4
-----
have been successfully applied to stream ciphers, block ciphers, and cryptographic hash
functions.
**2.2 Cube-and-Conquer**
If a given SAT instance is too hard for a sequential SAT solver, it makes sense to solve it in
parallel (Balyo & Sinz, 2018). If only complete algorithms are considered, then there are two
main approaches to parallel SAT solving: portfolio (Hamadi, Jabbour, & Sais, 2009) and
_divide-and-conquer (B¨ohm & Speckenmeyer, 1996). According to the portfolio approach,_
many different sequential SAT solvers (or maybe different configurations of the same solver)
solve the same problem simultaneously. In the divide-and-conquer approach, the problem
is decomposed into a family of simpler subproblems that are solved by sequential solvers.
_Cube-and-conquer (Heule et al., 2011; Heule, Kullmann, & Biere, 2018) is a divide-and-_
conquer SAT solving approach that combines lookahead with CDCL. On the cubing phase,
a modified lookahead solver splits a given formula into cubes. On the conquer phase, by
joining each cube with the original formula a subformula is formed. Finally a CDCL solver
is run on the subformulas. If the original formula is unsatisfiable, then all the subformulas
are unsatisfiable. Otherwise, at least one subformula is satisfiable. Since cubes can be
processed independently, the conquer phase can be easily parallelized.
As it was mentioned in the previous subsection, lookahead is a complete algorithm.
When used in the cubing phase of Cube-and-Conquer, a lookahead solver is forced to cut
off some branches thus producing cubes. Therefore, such a solver produces a decision tree,
where leaves are either refuted ones (with no possible solutions), or cubes. There are two
main cutoff heuristics that decide when a branch becomes a cube. In the first one, a
branch is cut off after a given number of decisions (Hyv¨arinen, Junttila, & Niemel¨a, 2010).
According to the second one, it happens when the number of variables in the corresponding
subproblem drops below a given threshold (Heule et al., 2011). In the present study the
second cutoff heuristic is used since it usually shows better results on hard instances.
**2.3 Cryptographic Hash Functions**
A hash function h is a function with the following properties (Menezes et al., 1996).
1. Compression: h maps an input x of arbitrary finite size to an output h(x) of fixed
size.
2. Ease of computation: for any given input x, h(x) is easy to compute.
An unkeyed cryptographic hash function h is a hash function that has the following
potential properties (Menezes et al., 1996).
1. Collision resistance: it is computationally infeasible to find any two inputs x and x[′]
such that x = x[′], h(x) = h(x[′]).
_̸_
2. Preimage resistance: for any given output y, it is computationally infeasible to find
any of its preimages, i.e. any such input x[′] that h(x[′]) = y.
3. Second-preimage resistance: for any given input x, it is computationally infeasible to
find x[′] such that x[′] = x, h(x) = h(x[′]).
_̸_
5
-----
Inputs of cryptographic hash functions are usually called messages, while outputs are
called hash values or just hashes. Hereinafter only unkeyed cryptographic hash functions
are considered.
Methods for disproving the mentioned three properties are called collision attacks, preim_age attacks, and second preimage attacks, respectively. If an attack is computationally fea-_
sible, then it is called practical. Usually it is much easier to propose a practical collision
attack than attacks of two other types. It is clear that collision resistance implies secondpreimage resistance, while second-preimage attack implies a collision attack. This study is
focused on practical preimage attacks on step-reduced cryptographic hash functions MD4
and MD5. In the rest of the paper, a practical inversion of a cryptographic hash function
implies a practical preimage attack and vise versa.
**2.4 MD4**
The Message Digest 4 (MD4) cryptographic hash function was proposed by Ronald Rivest
in 1990 (Rivest, 1990). Given a message of arbitrary finite size, padding is applied to
obtain a message that can be divided into 512-bit blocks. Then a 128-bit hash is produced
by iteratively applying the MD4 compression function to the blocks in accordance to the
Merkle-Damgard construction.
Consider the compression function in more detail. Given a 512-bit input, it produces
a 128-bit output. The function consists of three rounds, sixteen steps each, and operates
by transforming data in four 32-bit registers A, B, C, D. If a message block is the first one,
then the registers are initialized with the following constants, respectively: 0x67452301;
```
0xefcdab89; 0x98badcfe; 0x10325476. Otherwise, registers are initialized with an output
```
produced by the compression function on the previous message block. The message block
_M is divided into sixteen 32-bit words. In each step, one register’s value is updated by_
mixing one message word with values of all four registers and an additive constant. This
transformation is partially made via a nonlinear function, which is specific for each round.
Additive constants are also round-specific. As a result, in each round all sixteen words take
part in such updates. Finally, registers are incremented by the values they had after the
current block initialization, and the output is produced as a concatenation of A, B, C, D.
The nonlinear functions and additive constants are presented in Table 1.
Table 1: Characteristics of MD4 rounds.
Round Nonlinear function Additive constant
1 _F_ (x, y, z) = (x _y)_ ( _x_ _z)_ `0x0`
_∧_ _∨_ _¬_ _∧_
2 _G(x, y, z) = (x_ _y)_ (x _z)_ (y _z)_ `0x5a827999`
_∧_ _∨_ _∧_ _∨_ _∧_
3 _H(x, y, z) = x_ _y_ _z_ `0x6ed9eba1`
_⊕_ _⊕_
The full description of MD4 compression function can be found in (Rivest, 1990). Algorithm 1 presents the function when it processes the first message block. In the pseudocode
some steps are omitted, but the remaining ones are enough to understand MD4-related
results of the present paper. Here ≪ _r stands for the circular shifting to the left by r bits_
position.
6
-----
**Algorithm 1 MD4 compression function on the first 512-bit message block.**
**Input: 512-bit message block M** .
**Output: Updated values of registers A, B, C, D.**
1: AA _A_ `0x67452301`
_←_ _←_
2: BB _B_ `0xefcdab89`
_←_ _←_
3: CC _C_ `0x98badcfe`
_←_ _←_
4: DD _D_ `0x10325476`
_←_ _←_
5: A ← (A + F (B, C, D) + M [0]) ≪ 3 _▷_ ROUND 1 starts, Step 1
6: . . . _▷_ Steps 2-12
7: A ← (A + F (B, C, D) + M [12]) ≪ 3 _▷_ Step 13
8: D ← (D + F (A, B, C) + M [13]) ≪ 7 _▷_ Step 14
9: C ← (C + F (D, A, B) + M [14]) ≪ 11 _▷_ Step 15
10: B ← (B + F (C, D, A) + M [15]) ≪ 19 _▷_ Step 16
11: A ← (A + G(B, C, D) + M [0] + 0x5a827999) ≪ 3 _▷_ ROUND 2 starts, Step 17
12: D ← (D + G(A, B, C) + M [4] + 0x5a827999) ≪ 5 _▷_ Step 18
13: C ← (C + G(D, A, B) + M [8] + 0x5a827999) ≪ 9 _▷_ Step 19
14: B ← (B + G(C, D, A) + M [12] + 0x5a827999) ≪ 13 _▷_ Step 20
15: . . . _▷_ Steps 21-32
16: A ← (A + H(B, C, D) + M [0] + 0x6ed9eba1) ≪ 3 _▷_ ROUND 3 starts, Step 33
17: . . . _▷_ Steps 34-47
18: B ← (B + H(C, D, A) + M [15] + 0x6ed9eba1) ≪ 15 _▷_ Step 48
19: A _A + AA_ _▷_ Increment A by the initial value
_←_
20: A _B + BB_ _▷_ Increment B by the initial value
_←_
21: A _C + CC_ _▷_ Increment C by the initial value
_←_
22: A _D + DD_ _▷_ Increment D by the initial value
_←_
In 1995, a practical collision attack on MD4 was proposed (Dobbertin, 1996). In 2005,
it was theoretically shown that on a very small fraction of messages MD4 is not second
preimage resistant (Wang, Lai, Feng, Chen, & Yu, 2005). In 2008, a theoretical preimage
attack on MD4 was proposed (Leurent, 2008). Despite the found vulnerabilities, MD4 is
still used to compute password-derived hashes in some operating systems of the Windows
family, including Windows 10, due to backwards compatibility issues.
**2.5 Dobbertin’s constraints for MD4**
Since MD4 is still preimage resistant and second preimage resistant in practice, its stepreduced versions have been studied recently. In 1998, Hans Dobbertin introduced additional
constraints for MD4 (Dobbertin, 1998). Further they are called Dobbertin’s constraints.
Consider a constant 32-bit word K. The Dobbertin’s constraints for 32-step MD4 are as
follows: A = K in steps 13, 17, 21, 25; D = K in steps 14, 18, 22, 26; C = K in steps 15,
19, 23, 27 (numbering from 1). It means that on the first round, 3 out of 12 constraints are
applied, while the remaining 9 ones are applied on the second round. Algorithm 2 shows
how the steps from Algorithm 1 are changed when Dobbertin’s constraints are applied.
Comments for the constrained steps are marked with bold.
7
-----
**Algorithm 2 MD4 compression function on the first 512-bit message block with applied**
Dobbertin’s constraints.
**Input: 512-bit message block M** .
**Output: Updated values of registers A, B, C, D.**
1: AA _A_ `0x67452301`
_←_ _←_
2: BB _B_ `0xefcdab89`
_←_ _←_
3: CC _C_ `0x98badcfe`
_←_ _←_
4: DD _D_ `0x10325476`
_←_ _←_
5: A ← (A + F (B, C, D) + M [0]) ≪ 3 _▷_ ROUND 1 starts, Step 1
6: . . . _▷_ Steps 2-12
7: A ← (A + F (B, C, D) + M [12]) ≪ 3 ← _K_ _▷_ **Step 13, A=K**
8: D ← (D + F (A, B, C) + M [13]) ≪ 7 ← _K_ _▷_ **Step 14, D=K**
9: C ← (C + F (D, A, B) + M [14]) ≪ 11 ← _K_ _▷_ **Step 15, C=K**
10: B ← (B + F (C, D, A) + M [15]) ≪ 19 _▷_ Step 16
11: A ← (A + G(B, C, D) + M [0] + 0x5a827999) ≪ 3 ← _K_ _▷_ **ROUND 2 starts, Step**
**17, A=K**
12: D ← (D + G(A, B, C) + M [4] + 0x5a827999) ≪ 5 ← _K_ _▷_ **Step 18, D=K**
13: C ← (C + G(D, A, B) + M [8] + 0x5a827999) ≪ 9 ← _K_ _▷_ **Step 19, C=K**
14: B ← (B + G(C, D, A) + M [12] + 0x5a827999) ≪ 13 _▷_ Step 20
15: . . . _▷_ Steps 21-32
16: A ← (A + H(B, C, D) + M [0] + 0x6ed9eba1) ≪ 3 _▷_ ROUND 3 starts, Step 33
17: . . . _▷_ Steps 34-47
18: B ← (B + H(C, D, A) + M [15] + 0x6ed9eba1) ≪ 15 _▷_ Step 48
19: A _A + AA_ _▷_ Increment A by the initial value
_←_
20: A _B + BB_ _▷_ Increment B by the initial value
_←_
21: A _C + CC_ _▷_ Increment C by the initial value
_←_
22: A _D + DD_ _▷_ Increment D by the initial value
_←_
Consider step 17. A = C = D = K due to the constrained steps 13, 14, and 15, while
_B is unknown. Since G is the majority function, then G(x, y, y) = y for any x and y. So_
we have
_K = (A + G(B, C, D) + M_ [0] + 0x5a827999) ≪ 3 =
(K + G(B, K, K) + M [0] + 0x5a827999) ≪ 3 =
(K + K + M [0] + 0x5a827999) ≪ 3
Then it follows
_M_ [0] = (K ≪ 29) − 2K − `0x5a827999`
For example, if K = 0xffffffff, then
_M_ [0] = 0xffffffff 2 `0xffffffff` `0x5a827999 =`
_−_ _·_ _−_
```
0xffffffff 0x5a827999 = 0xa57d8668
```
_−_ _−_
8
-----
Thus if A is equal to a constant word in step 17, M [0] becomes a constant as well. The
same holds for M [4], M [8], M [1], M [5], M [9], M [2], M [6], and M [10] due to constrained
steps 18, 19, 21, 22, 23, 25, 26, 27, respectively. In the pseudocode, only 6 constrained steps
are shown, but for the remaining ones the picture is the same. Finally, the Dobbertin’s
constraints turn 9 message words out of 16 into constants. Therefore the constrained compression function maps 0, 1 onto 0, 1 while the original one maps 0, 1 onto
_{_ _}[224]_ _{_ _}[128]_ _{_ _}[512]_
0, 1 . As a result, for any given hash and a randomly chosen K the number of preim_{_ _}[128]_
ages (messages) is significantly reduced, maybe even to 0. The Dobbertin’s constraints is
an example of streamlined constraints (Gomes & Sellmann, 2004). Such constraints are not
implied by the formula, so they can remove some (or even all) solutions, but have a good
chance of leaving at least one solution.
The Dobbertin’s constraints were originally proposed for 32-step MD4 and they do not
guarantee that for a certain pair (hash,K), any preimage will remain when the constraints
are applied with constant K. What they do guarantee is that the corresponding system
of equations will become much smaller and easier to solve. So the idea is to try different
_K until a preimage is found. The point is that even if a few such simplified problems are_
to be solved, it may anyway be much easier than solving the original problem. The same
holds for more than 32 steps because all the constraints are applied before the 32nd step.
In other words, adding more unconstrained steps does not reduce the number of solutions.
In (Dobbertin, 1998), the Dobbertin’s constraints were used to invert 32-step MD4 by
randomly choosing values of K and B (on step 28) until a consistent system is formed and a
preimage is found. In case of 32 steps, a constant value B in addition to K on step 28 implies
values of the remaining 7 message words. This is not the case for more than 32 steps. In
2000, modified Dobbertin’s constraints were applied to invert MD4 when the second round
is omitted (Kuwakado & Tanaka, 2000). In 2007, a SAT-based implementation of slightly
modified Dobbertin’s constraints (where the constraint on step 13 was omitted) made it
possible to invert 39-step MD4 (De et al., 2007). Since 2007, several unsuccessful attempts
have been made to invert 40-step MD4, see, e.g., (Legendre et al., 2012). The present study
is aimed at inverting 40-, 41-, 42-, and 43-step MD4.
**2.6 MD5**
MD5 was proposed in 1992 by Ronald Rivest as a slightly slower, but at the same time
much more secure extension of MD4 (Rivest, 1992).
The main changes in MD5 compared to MD4 are as follows.
1. Addition of the fourth round of 16 steps with its own round function, so MD5 consists
of 64 steps;
2. Replacement of the second round’s nonlinear function by a new function;
3. Usage of a unique additive constant in each of the 64 steps;
4. Addition of output from the previous step.
The nonlinear functions are as follows:
- Round 1. F (x, y, z) = (x _y)_ ( _x_ _z)_
_∧_ _∨_ _¬_ _∧_
9
-----
- Round 2. G(x, y, z) = (x _z)_ (y _z)_
_∧_ _∨_ _∧¬_
- Round 3. H(x, y, z) = x _y_ _z_
_⊕_ _⊕_
- Round 4. I(x, y, z) = y (x _z)_
_⊕_ _∨¬_
For the first time a practical collision attack on MD5 was presented in 2005 (Wang &
Yu, 2005). In 2009, a theoretical preimage attack was proposed (Sasaki & Aoki, 2009). It is
known that the Dobbertin’s constraints are not applicable to MD5 (Aoki & Sasaki, 2008).
This is because of changes 2-4 mentioned above. In particular, when applying to MD5,
these constraints remove all solutions (preimages), so simplicity of the obtained reduced
problem does not help.
In 2007, 26-step MD5 was inverted in (De et al., 2007). In 2012, it was done for 27-, and
28-step MD5 (Legendre et al., 2012). In both papers SAT-based cryptanalysis via CDCL
solvers was applied, yet no additional constraints were added to the corresponding CNFs.
MD5 is still used in practice as a checksum to verify data integrity. Another application is
the storage of passwords’ hashes in operating systems.
### 3. Dobbertin-like Constraints for Inverting Step-reduced MD4
As it was mentioned in the previous section, the progress in inverting step-reduced MD4 was
mainly due to the Dobbertin’s constraints. This section first proposes their generalization
— Dobbertin-like constraints. Then an algorithm for inverting MD4 via Dobbertin-like
constraints is proposed.
**3.1 Dobbertin-like Constraints**
Suppose that given a constant word K, only 11 of 12 Dobbertin’s constraints hold, while
in the remaining corresponding step only b, 0 _b_ 32 bits of the register are equal to the
_≤_ _≤_
corresponding b bits of K, at the same time the remaining 32 _b bits in the register take_
_−_
values, opposite to those in K. Denote these constraints as Dobbertin-like constraints. It is
clear that the Dobbertin’s constraints is a special case of Dobbertin-like constraints when
_b = 32._
Denote an inversion problem for step-reduced MD4 with applied Dobbertin-like constraints as MD4inversion(y, s, K, p, L), where:
- y is a given 128-bit hash;
- s is the number of MD4 steps (starting from the first one);
- K is a 32-bit constant word used in the Dobbertin’s constraints;
- p 13, 14, 15, 17, 18, 19, 21, 22, 23, 25, 26, 27 is the specially constrained step;
_∈{_ _}_
- L is a 32-bit word such that if Li = 0, 0 ≤ _i ≤_ 31, then i-th bit of the register value
modified in step p is equal to Ki. Otherwise, if Li = 1, it is equal to ∼Ki.
In other words, the 32-bit word L serves as a bit mask and controls how similar the
specially constrained register is to the constant word K. To make this definition more
10
-----
clear, three examples are given below. Hereinafter 0hash and 1hash mean 128 0s and 128
1s (i.e. 4 words 0x00000000 and 0xffffffff, respectively).
**Example 3.1 (MD4inversion(0hash, 32, 0x62c7Ec0c, 21, 0x00000000)). The problem is to**
invert 0hash produced by 32-step MD4. Since L = 0x00000000, then the specially constrained register (in step 21) has value K, so all 12 Dobbertin’s constraints are applied as
usual with K = 0x62c7Ec0c. A similar inversion problem (up to choice of K) in fact was
solved in (Dobbertin, 1998).
**Example 3.2 (MD4inversion(1hash, 39, 0xfff00000, 12, 0xffffffff)). The problem is to**
invert 1hash produced by 39-step MD4. Since L = 0xffffffff, the specially constrained
register (in step 12) has value _K = 0x000fffff. In the remaining 11 Dobbertin’s steps_
_∼_
the registers have value K = 0xfff00000.
**Example 3.3 (MD4inversion(1hash, 40, 0xffffffff, 12, 0x00000003)). The problem is to**
invert 1hash produced by 40-step MD4. Since L = 0x00000003, the first 30 bits of the
specially constrained register (in step 12) are equal to those in K, while the last two bits
have values ∼K30 and ∼K31, respectively. It means that this register has value 0xfffffffc,
while in the remaining 11 Dobbertin’s steps the registers have value K = 0xffffffff.
**3.2 Inversion Algorithm**
Dobbertin-like constraints can be used for finding preimages of a step-reduced MD4 according to the following idea. For a given hash y, step s, random K, an inversion problem is
formed with L = 0x00000000. Thus all 12 Dobbertin’s constraints are applied. The inversion problem is solved and, if a preimage is found, nothing else should be done. Otherwise,
if it is proven that no preimages exist in the current inversion problem, then a new one is
formed with L = 0x00000001. In this case the specially constrained register’s value is just
1 bit shy of being K. This inversion problem is solved as well. If still no preimage exists,
then L is further modified: 0x00000002, 0x00000003, and so on. The intuition here is that
the Dobbertin’s constraints lead to a system of equations that is either consistent (with
very few solutions) or quite “close” to a consistent one. In the latter case trying different
values of L helps to form a consistent system and find its solution.
Algorithm 3 follows the described idea. In the pseudocode a complete algorithm is
_A_
used, which for a formed inversion problem returns all the preimages if they exist. Note
that it is not guaranteed that Algorithm 3 finds any preimage for a given hash. However,
as it will be shown in sections 6 and 7, in practice Algorithm 3 is able to find preimages for
step-reduced MD4. Moreover, it usually does it in just few iterations (from 1 to 3) of the
while loop.
Complete algorithms of various types can be used to solve inversion problems formed in
Algorithm 3. In particular, wide spectrum of constraint programming (Rossi, van Beek, &
Walsh, 2006) solvers are potential candidates. In preliminary experiments, state-of-the-art
sequential and parallel CDCL SAT solvers were tried to invert 40-step MD4, but even on
the first iterations (where all 12 Dobbertin’s constraints are added) CNFs turned out to
be too hard for them. That is why it was decided to use Cube-and-Conquer SAT solvers,
which are more suitable for extremely hard SAT instances. The next section describes how
11
-----
**Algorithm 3 Algorithm for inverting step-reduced MD4 via Dobbertin-like constraints.**
**Input: Hash y; the number of MD4 steps s; constant K; step p with specially constrained**
register; a complete algorithm _._
_A_
**Output: Preimages for hash y.**
1: preimages
_←{}_
2: i 0
_←_
3: while i < 2[32] **do**
4: _L_ DecimalToBinary(i)
_←_
5: _preimages_ (MD4inversion(y, s, K, p, L))
_←A_
6: **if preimages is not empty then**
7: **break**
8: _i_ _i + 1_
_←_
9: return preimages
a given problem can be properly split into simpler subproblems on the cubing phase of
Cube-and-Conquer.
### 4. Finding Cutoff Thresholds for Cube-and-Conquer
Recall (see Subsection 2.2) that in Cube-and-Conquer the following cutoff threshold n is
meant in the cubing phase: the number of variables in a subformula, formed by adding a
cube to the original CNF and applying UP. It is crucial to properly choose this threshold.
If it is too high, then the cubing phase is performed in no time, but very few extremely
hard (for a CDCL solver) cubes might be produced; if it is too low, then the cubing phase
will be extremely time consuming, and also there will be too huge number of cubes.
Earlier two algorithms aimed at finding a cutoff threshold with minimal estimated runtime of Cube-and-Conquer were proposed. Subsection 7.2 of the tutorial (Heule, 2018a)
proposed Algorithm A as follows:
Optimizing the heuristics requires selecting useful subproblems of the hard formula. This can be done as follows: First determine the depth for which the
number of refuted nodes is at least 1000. Second, randomly pick about 100
subproblems (cubes) of the partition on that depth. Second, solve these 100
subproblems and select the 10 hardest ones for the optimization.
Later Algorithm B was proposed in (Bright, Cheung, Stevens, Kotsireas, & Ganesh,
2021):
The cut-off bound was experimentally chosen by randomly selecting up to several
hundred instances from each case and determining a bound that minimizes the
sum of the cubing and conquering times.
This section proposes a new algorithm that is inspired by algorithms A and B. First the
algorithm is presented and then its novelty is described.
To find a cutoff threshold with minimal estimated runtime, first it is needed to preselect
promising values of n. On the one hand, the number of refuted leaves should be quite
12
-----
significant since it may indicate that at least some subformulas have become really simpler
compared to the original problem. On the other hand, the total number of cubes should
not be too large. An auxiliary Algorithm 4 follows this idea. Given a lookahead solver, a
CNF, and a cutoff threshold, the function LookaheadWithCut runs the solver on the
CNF with the cutoff threshold (see Subsection 2.2) and outputs cubes and the number of
refuted leaves.
**Algorithm 4 Preselect promising thresholds for the cubing phase of Cube-and-Conquer.**
**Input: CNF F; lookahead solver ls; starting threshold nstart; threshold decreasing step**
_nstep; maximal number of cubes maxc; minimal number of refuted leaves minr._
**Output: Stack of promising thresholds and corresponding cubes.**
1: function PreselectThresholds(F, ls, nstart, nstep, maxc, minr)
2: _stack_
_←{}_
3: _n ←_ _nstart_
4: **while n > 0 do**
5: _c, r_ LookaheadWithCut(ls, _, n)_ _▷_ Get cubes and number of refuted
_⟨_ _⟩←_ _F_
6: **if Size(c) > maxc then**
7: **break** _▷_ Break if too many cubes
8: **if r** _minr then_
_≥_
9: _stack.push(_ _n, c_ ) _▷_ Add threshold and cubes
_⟨_ _⟩_
10: _n_ _n_ _nstep_ _▷_ Decrease threshold
_←_ _−_
11: **return stack**
When promising values of the threshold are preselected, it is needed to estimate the
hardness of the corresponding conquer phases. It can be done by choosing a fixed number
of cubes by simple random sampling (Starnes, Yates, & Moore, 2010) among those produced
in the cubing phase. If all corresponding subproblems from the sample are solved by a CDCL
solver in a reasonable time, then an estimated total solving time for all subproblems can be
easily calculated. This idea is implemented as Algorithm 5. On the first stage, promising
thresholds are preselected by Algorithm 4, while on the second stage the one with minimal
runtime estimation of the conquer phase is chosen among them. Given a CDCL solver, a
CNF, a cube, and a time limit in seconds, the function SolveCube runs the CDCL solver
with the time limit on the CNF, and returns the runtime in seconds and an answer whether
the CNF is satisfiable or not.
The algorithm operates in two possible modes. In the estimating mode, the algorithm
terminates upon reaching a time limit by the CDCL solver on any subproblem in random
samples. In the incomplete SAT-solving mode, the algorithm terminates upon finding a
satisfying assignment. The first mode is aimed at estimating the hardness of a given CNF,
while the second one is aimed at finding a satisfying assignment of a satisfiable CNF.
The proposed algorithm has the following features.
1. A stack is used to preselect promising thresholds on the first stage in order to start
the second stage with solving the simplest subproblems (with lowest n). In practice
it allows obtaining some estimation quickly and then improve it.
13
-----
**Algorithm 5 Finding a cutoff threshold with minimal estimated runtime of the conquer**
phase.
**Input: CNF** ; lookahead solver ls; threshold decreasing step nstep; maximal number of
_F_
cubes maxc; minimal number of refuted leaves minr; sample size N ; CDCL solver cs;
CDCL solver time limit maxcst; the number of CPU cores cores; operating mode mode.
**Output: A threshold nbest with runtime estimation ebest and cubes cbest; whether a satis-**
fying assignment is found isSAT .
1: isSAT `Unknown`
_←_
2: nstart ← Varnum(F) − _nstep_
3: ⟨nbest, ebest, cbest⟩←⟨nstart, +∞, {}⟩
4: stack ← PreselectThresholds(F, ls, nstart, nstep, maxc, minr) _▷_ First stage
5: while stack is not empty do _▷_ Second stage: estimate thresholds
6: _n, c_ _stack.pop()_ _▷_ Get a threshold and cubes
_⟨_ _⟩←_
7: _sample_ SimpleRandomSample(c, N ) _▷_ Select N random cubes
_←_
8: _runtimes_
_←{}_
9: **for each cube from sample do**
10: _t, st_ SolveCube(cs, _, cube, maxcst)_ _▷_ Add cube and solve
_⟨_ _⟩←_ _F_
11: **if t > maxcst and mode = estimating then** _▷_ If CDCL was interrupted,
12: **break** _▷_ stop processing sample.
13: **else**
14: _runtimes.add(t)_ _▷_ Add proper runtime
15: **if st = True then** _▷_ If SAT,
16: _isSAT_ `True`
_←_
17: **if mode = solving then** _▷_ and if SAT solving mode,
18: **return ⟨nbest, ebest, cbest, isSAT** _⟩_ _▷_ return SAT immediately.
19: **if Size(runtimes) < N then** _▷_ If at least one interrupted in sample,
20: **break** _▷_ stop main loop.
21: _e_ Mean(runtimes) Size(c)/cores _▷_ Calculate runtime estimation
_←_ _·_
22: **if e < ebest then**
23: _⟨nbest, ebest, cbest⟩←⟨n, e, c⟩_ _▷_ Update best threshold
24: return ⟨nbest, ebest, cbest, isSAT _⟩_
2. The runtime of the cubing phase is not estimated because it is assumed that this is
negligible compared to the conquer phase.
3. In the estimating mode, if on the second stage a CDCL solver fails solving some
subproblem within time limit, the algorithm terminates. This is done because in this
case it is impossible to calculate a meaningful estimation for the threshold. Another
reason is that subproblems from the next thresholds (with higher n) will likely be
even harder.
4. It is possible that satisfying assignments are found when solving subproblems from
random samples. Indeed, if a given CNF is satisfiable, then cubes which imply satisfying assignments might be chosen to samples.
14
-----
5. In the estimating mode, even if a satisfying assignment is found when solving some
subproblem from samples, the algorithm does not terminate because in this case the
main goal is to calculate a runtime estimation.
6. In the estimating mode it is a general algorithm that is able to estimate the hardness
of an arbitrary CNF.
7. In the incomplete SAT-solving mode, a solution can be found only for a satisfiable
CNF, and even in this case this is not guaranteed because of the time limit for the
CDCL solver (see (Kautz, Sabharwal, & Selman, 2021)).
8. In fact, the runtime estimation is a stochastic costly black-box objective function
(see, e.g., (Audet & Hare, 2017; Semenov, Zaikin, & Kochemazov, 2021)) that takes
an integer value n as input. The algorithm minimizes this objective function.
Since all details of Algorithm 5 are given, it now can be compared to Algorithms A and B
(see the beginning of this section). It is clear that the idea is the same in all three algorithms
— for a certain value of the cutoff threshold, a sample of cubes is formed, the corresponding
subproblems are solved, and finally a runtime estimation is calculated. However, there are
several major differences which are listed below.
1. Algorithms A and B were described informally and briefly, while Algorithm 5 is presented formally and in detail.
2. In opposite to Algorithms A and B, Algorithm 5 takes into account the situation when
some subproblems from a sample are so hard that they can not be solved in reasonable
time by a CDCL solver.
3. Algorithms A and B assume that on the conquer phase subproblems are solved incrementally, while Algorithm 5 assumes that each subproblem is solved by an independent
call of a CDCL solver.
The main difference is the second one. This feature of Algorithm 5 is extremely important in application to cryptanalysis problems, which are considered in the rest of the
present paper. The reason is that in this case subproblems in a sample usually differ much
(by thousands and even millions of times) in CDCL solver’s runtime. A possible explanation why this feature was not taken into account in both Algorithms A and B is that they
were applied to combinatorial and geometric problems, where subproblems’ hardness in a
sample is usually uniform. Importance of the third feature follows from the second one —
incremental solving pays off in case of the uniform hardness, otherwise it can significantly
slow down the solving process.
When a cutoff threshold is found by Algorithm 5, the conquer phase operates as follows
(see Subsection 2.2). First, subproblems are created by adding cubes to the original CNF
in the form of unit clauses. Second, the subproblems are solved by the same CDCL solver
that was used to find the threshold for the cubing phase. In the present study the goal is
to find all solutions of a considered inversion problem. That is why, given a subproblem,
the CDCL solver finds all its satisfying assignments. In opposite to the cubing phase, here
the runtime of the CDCL solver is not limited.
15
-----
### 5. Considered Inversion Problems and Their SAT Encodings
This section describes the considered inversion problems for step-reduced MD4 and MD5,
as well as their SAT encodings.
Following all earlier attempts to invert step-reduced MD4 via SAT (see, e.g., (De et al.,
2007; Legendre et al., 2012; Lafitte, Jr., & Heule, 2014; Gribanova & Semenov, 2018)), the
padding is omitted (see Subsection 2.4) and only one 512-bit message block is considered. It
means that in fact a step-reduced MD4 compression function is considered when it operates
on the first block, like it was shown in Algorithm 1. The final incrementing is also omitted
since it should be done only after 48-th step. Note that these restrictions does not make
inversion problems easier since the compression function is the main component of MD4
function from the resistance point view. Inversion of step-reduced MD5 is considered in
similar way.
**5.1 Considered hashes**
The following four hashes are chosen for inversion:
1. 0x00000000 0x00000000 0x00000000 0x00000000;
2. 0xffffffff 0xffffffff 0xffffffff 0xffffffff;
3. 0x01234567 0x89abcdef 0xfedcba98 0x76543210;
4. 0x62c7Ec0c 0x751e497c 0xd49a54c1 0x2b76cff8.
Recall that 0hash and 1hash mean the first and the second hash from the list, respectively. These two hashes are chosen for inversion because it is a common practice in the
cryptographic community. The reason is that inverting some hash, that looks just like a
random word, is suspicious. Indeed, one can take a random message, produce its hash
and declare that this very hash is inverted. On the other hand, if a hash has a regular
structure, this approach does not work. All 0s and all 1s are two hashes with the most
regular structure, that is why they are usually chosen. For the first time 32-step MD4 was
inverted for 0hash (Dobbertin, 1998), while in 39-step case it was done for 1hash (De et al.,
2007), and later for 0hash (Legendre et al., 2012). As for the cryptographic hash functions
SHA-0 and SHA-1, their 23-step (out of 80) versions for the first time were inverted for
_0hash (Legendre et al., 2012)._
The third hash from the list was used to invert 28-step MD5 in (Legendre et al., 2012).
Hereinafter this hash is called symmhash. It is symmetrical — the last 64 bits are the
first 64 bits in reverse order, but at the same time it is less regular than 1hash or 0hash.
The same result for 28-step MD5 was described in later two papers of the same authors.
Unfortunately, none of these three papers explained the non-existence of the results for
_1hash and 0hash._
The fourth hash from the list is chosen randomly. The goal is to show that the proposed
approach is applicable not only to hashes with regular structure. This hash is further called
_randhash._
16
-----
**5.2 Step-reduced MD4**
While in (De et al., 2007) only K = 0x00000000 was used as a constant in the Dobbertin’s
constraints, in the present study both K = 0x00000000 and K = 0xffffffff are tried
in Dobbertin-like constraints. The constraint in step 12 is chosen for the modification (so
_p = 12, see Subsection 3.1) since in (De et al., 2007) the constraint for this very step was_
entirely omitted. Eight step-reduced versions of MD4, from 40 to 47 steps, as well as the
full MD4 are studied. Hence there are 9 4 2 = 72 MD4-related inversion problems in
_×_ _×_
total. None of these 72 inversion problems have been solved so far.
Consider 40-step MD4. Since K has two values and y has four values, Algorithm 3
should be run on eight inputs. As a result, according to the notation from Subsection 3.1,
the following eight inversion problems are formed in the corresponding first iterations of
Algorithm 3:
1. MD4inversion(0hash, 40, 0x00000000, 12, 0x00000000);
2. MD4inversion(0hash, 40, 0xffffffff, 12, 0x00000000);
3. MD4inversion(1hash, 40, 0x00000000, 12, 0x00000000);
4. MD4inversion(1hash, 40, 0xffffffff, 12, 0x00000000).
5. MD4inversion(symmhash, 40, 0x00000000, 12, 0x00000000);
6. MD4inversion(symmhash, 40, 0xffffffff, 12, 0x00000000);
7. MD4inversion(randhash, 40, 0x00000000, 12, 0x00000000);
8. MD4inversion(randhash, 40, 0xffffffff, 12, 0x00000000).
For illustrative purpose, consider the first case: invert 0hash produced by 40-step MD4
with Dobbertin’s constraints and K = 0x00000000. If no preimage exists for this inversion
problem, then on the second iteration of Algorithm 3 L is increased by 1, so the inversion
problem MD4inversion(0hash, 40, 0x00000000, 12, 0x00000001) is formed and so on.
**5.3 Step-reduced MD5**
Inversion of only 28-step MD5 compression function is considered in this study for the four
hashes presented above. Recall that in opposite to MD4, no additional constraints that
reduce the number of preimages are added. Note that for all hashes but symmhash the
inversion problems have not been solved earlier.
**5.4 SAT Encodings**
It is possible to construct CNFs that encode MD4 and MD5 via the following automatic
tools: CBMC (Clarke, Kroening, & Lerda, 2004); SAW (Carter, Foltzer, Hendrix, Huffman,
& Tomb, 2013); Transalg (Semenov, Otpuschennikov, Gribanova, Zaikin, & Kochemazov, 2020); CryptoSAT (Lafitte, 2018). In the present paper, the CNFs are constructed
via Transalg of version 1.1.5[1]. This tool takes a description of an algorithm as an input
1. https://gitlab.com/transalg/transalg
17
-----
and outputs a CNF that implements the algorithm. The description must be formulated in
a domain specific C-like language called TA language. The TA language supports the following basic constructions used in procedural languages: variable declarations; assignment
operators; conditional operators; loops, function calls. Additionally it supports various integer operations and bit operations including bit shifting and comparison that is quite handy
when describing a cryptographic algorithm. A TA program is a list of functions in TA language. All the constructed CNFs and the corresponding TA programs are available online[2].
All these CNFs can be easily reconstructed by giving the TA programs to Transalg as
inputs.
In a CNF that encodes step-reduced MD4, the first 512 variables correspond to a message, the last 128 variables correspond to a hash, while the remaining auxiliary variables
are needed to encode how the hash is produced given the message. The first 512 variables
are further called message variables, while the last 128 ones — hash variables. The Tseitin
transformations are used in Transalg to introduce auxiliary variables (Tseitin, 1970).
Characteristics of the constructed CNFs are given in Table 2.
Table 2: Characteristics of CNFs that encode the considered step-reduced MD4 and MD5.
Function Variables Clauses Literals
MD4-40 7025 70 809 317 307
MD4-41 7211 73 158 329 330
MD4-42 7397 75 507 341 353
MD4-43 7583 77 856 353 376
MD4-44 7769 80 205 365 399
MD4-45 7955 82 554 377 422
MD4-46 8141 84 903 389 445
MD4-47 8327 87 252 401 468
MD4-48 8513 89 601 413 491
MD5-28 7471 54 672 216 362
The CNF that encodes 40-step MD4 has 7 025 variables and 70 809 clauses. Then
every step adds 186 variables and 2 349 clauses, so as a result a CNF that encodes the full
(48-step) MD4 has 8 513 variables and 89 601 clauses. Note that these CNFs encode the
functions themselves, so all message and hash variables are unassigned. To obtain a CNF
that encodes an inversion problem for a given 128-bit hash, 128 corresponding one-literal
clauses are to be added, so all hash variables become assigned. The problem is to find values
of the message variables. The Dobbertin’s constraints are added as another 384 one-literal
clauses (32 clauses for each constraint). As a result, a CNF that encodes the inversion of
40-step MD4 with all 12 Dobbertin’s constraints has 7 025 variables and 71 321 clauses,
while that for the 48-step version consists of 8 513 variables and 90 113 clauses. Note that
Dobbertin-like constraints (see Subsection 3.1) are also added as 384 one-literal clauses —
the only difference is in values of the corresponding 32 variables that encode the specially
constrained register.
2. https://github.com/olegzaikin/EnCnC
18
-----
The CNF that encodes 28-step MD5 has 7 471 variables and 54 672 clauses. A CNF
that encodes an inversion problem has 7 471 variables and 54 800 clauses since only 128
one-literal clauses for hash variables are added.
### 6. Inverting 40-step MD4 via Dobbertin-like Constraints
This section describes experimental setup, simplification, and results for 40-step MD4.
Assume that several given hashes for a step-reduced MD4 are to be inverted. Then
Algorithm 3 and the estimating mode of Algorithm 5 can be used in the following combinations:
1. The estimating mode of Algorithm 5 is run on a CNF that encodes the inversion
problem for an arbitrary hash among given ones, yet the Dobbertin’s constraints are
fully applied, i.e. L = 0x00000000. When the best cutoff threshold is found, Algorithm 3 is iteratively run using a Cube-and-Conquer solver with the found threshold
as algorithm on all given hashes. It means that the threshold found for one hash
_A_
and L = 0x00000000 is used for all other hashes and values of L.
2. For each hash, its own best threshold is found for L = 0x00000000 and is used for
all other values of L. In Algorithm 3, is again a Cube-and-Conquer solver with the
_A_
found threshold.
3. For each hash and each value of L its own threshold is found. Therefore is the
_A_
estimating mode of Algorithm 5 followed by a Cube-and-Conquer solver.
In this study, the second combination is used since for any value of L the same amount
of one-literal clauses is added to a CNF.
**6.1 Experimental Setup**
Algorithm 3 was implemented in Python, while Algorithm 5 and the conquer phase of
Cube-and-Conquer were implemented in C++ as a parallel SAT solver Estimate-andCube-and-Conquer (EnCnC). The implementation is available online[3].
All experiments in this paper were held on a personal computer equipped with the 12core CPU AMD 3900X and 48 Gb of RAM. The implementations are multithreaded, so all
12 CPU cores were employed in all runs. In case of Algorithm 5, values of a cutoff threshold
_n and then subproblems from samples are processed in parallel. In case of the conquer_
phase, subproblems are processed in parallel.
The input parameters’ values of Algorithm 3 in case of 40-step MD4 were discussed in
Section 5. As for Algorithm 5, the following input parameters’ values were used:
- March cu lookahead solver (Heule et al., 2011) since it has been recently successfully
applied to several hard problems (Heule et al., 2016; Heule, 2018b).
- nstep = 5. It was chosen in preliminary experiments. If this parameter is equal to
1, then a better threshold usually can be found, but at the same time Algorithm 5
3. https://github.com/olegzaikin/EnCnC
19
-----
becomes quite time-consuming. On the other hand, if nstep is quite large, e.g., 50,
then as a rule almost all most promising thresholds are just skipped.
- maxc = 2 000 000. On the considered CNFs, March cu reaches 2 000 000 cubes in
about 30 minutes, so that value of maxc looks reasonable. Higher values were also
tried, but it did not give any improvement.
- minr = 1 000. If it is less then 1 000, then subproblems are too hard because they are
not simplified enough by lookahead. At the same time, higher value of this parameter
did not allow collecting enough amount of promising thresholds.
- N = 1 000. First N = 100 was tried, but it led to too optimistic estimations which
were several times lower than real solving time. On the other hand, N = 10 000
is too time-consuming and gives just modest improvement in accuracy compared to
_N = 1 000. The accuracy of obtained estimations is discussed later in Subsection 7.1._
- Kissat CDCL solver of version sc2021 (Biere, Fleury, & Heisinger, 2021a). The reason
is that Kissat and its modifications won SAT Competitions 2020 and 2021.
- maxst = 5 000 seconds. It is a standard time limit in SAT Competitions (see, e.g., (Balyo, Froleyks, Heule, Iser, J¨arvisalo, & Suda, 2021)), so modern CDCL solvers are
designed to show all their power within this time.
- cores = 12.
- mode = estimating. Here the goal is not just to find one preimage, but rather find
all preimages for a given inversion problem (up to added Dobbertin-like constraints).
It should be noted that in both Algorithm 5 and the conquer phase of Cube-and-Conquer
subproblems were solved by Kissat in the non-incremental mode, i.e. it solved them independently from each other.
**6.2 Simplification**
In case of 40-step MD4, two parameters were varied for each of four considered hashes
(see Section 5.1). The first one is the value of the Dobbertin’s constant K (see Section 3):
```
0x00000000 and 0xffffffff. The second one is simplification type applied to a CNF. A
```
motivation behind varying the second parameter is as follows. First, it is crucial to simplify
a CNF before giving it to a lookahead solver. Second, in preliminary experiments it was
found out that the simplification type can significantly alter the effectiveness of Cube-andConquer on the considered problems.
The CDCL solver CaDiCaL of version 1.5.0 (Froleyks & Biere, 2021) was used to
simplify the CNFs. This solver uses inprocessing, i.e. a given CNF is simplified during the
CDCL search (Biere, 2011). The more conflicts have been generated by a CDCL solver
so far, the more simplified (in terms of the number of variables) the CNF has been made.
A natural simplification measure in this case is the number of generated conflicts. In
the experiments related to 40-step MD4, the following limits on the number of generated
conflicts were tried: 1, 10 thousand, 100 thousand, 1 million, 10 million. Note that 1 conflict
20
-----
as the limit in some cases gives the same result as UP (see Subsection 2.1), while in the
remaining cases the corresponding CNF is slightly smaller.
For example, consider problem MD4inversion(1hash, 40, 0xffffffff, 12, 0x00000000).
Table 3 presents characteristics of six CNFs which encode this problem. The original
(unsimplified) CNF is described by the number of variables, clauses, and literals. For
those simplified by CaDiCaL also the runtime on 1 CPU core is given.
Table 3: CNFs that encode MD4inversion(1hash, 40, 0xffffffff, 12, 0x00000000). The
best values are marked with bold.
Simplif. type Variables Clauses Literals Simplif. runtime
no (original CNF) 7025 71 321 317 819 1 conflict 3824 33 371 138 820 0.02 sec
10 thousand conflicts 2969 27 355 116 618 0.31 sec
100 thousand conflicts 2803 23 121 94 250 4.29 sec
1 million conflicts 2756 **22 391** **90 412** 1 min 19 sec
10 million conflicts **2054** 24 729 110 267 33 min
It is clear, that first the number of variables, clauses, and literals decrease, but then 10
million conflicts provides lower number of variables yet the number of clauses and literals
is higher than that on 1 million conflicts. For other hashes and values of L the picture is
similar.
In preliminary experiments also the limits of 1 thousand and 100 million conflicts were
tried. However, it turned out that the first variant is usually similar to 1 conflict in number
of variables and clauses. The second variant in all cases was similar to 10 million conflicts
in number of variables, though the number of clauses was a bit lower. Yet generating 100
million conflicts is quite time consuming — it takes about 1 day on average. On the other
hand, 10 million conflicts are generated in about a half an hour. That is why these two
simplification types, 1 thousand and 100 million conflicts, were omitted.
**6.3 Experiments**
Of course, more parameters can be varied for each hash in addition to K and simplification
type mentioned in the previous subsection. One of the most natural is a CDCL solver used
in Cube-and-Conquer. For example, a cryptanalysis-oriented solver can be chosen (Soos,
Nohl, & Castelluccia, 2009; Nejati & Ganesh, 2019; Kochemazov, 2021). Moreover, internal
parameters of the chosen CDCL solver can be varied as well.
Recall that there are 4 hashes, 5 simplification types, while K has 2 values. Therefore
in total 4 5 2 = 40 CNFs were constructed with fully applied Dobbertin’s constraints
_×_ _×_
(L = 0x00000000) for MD4-40. On each of them the first iteration of Algorithm 3 was
run. It turned out, that Algorithm 5 could not find any estimations for all 20 CNFs with
_K = 0x00000000. The reason is because in all these cases Kissat was interrupted due to_
the time limit even for the simplest (lowest) values of the cutoff threshold n. On the other
hand, for K = 0xffffffff much more positive results were achieved. For 0hash, symmhash,
and randhash, estimations for all simplification types were successfully calculated, and the
21
-----
best one was 1 conflict in all these cases. On the other hand, for 1hash no estimations were
found for 1 conflict and 10 thousand conflicts, while the best estimation was gained for 1
million conflicts.
The results are presented in Table 4. For each pair (simplification type, hash) the best
estimation for 12 CPU cores, the corresponding cutoff threshold, and the number of cubes
are given. Here “-” means that no estimation was obtained because Kissat was interrupted
on the simplest threshold. Runtimes of Algorithm 5 are not presented there, but on average
it took about 2 hours for K = 0x00000000 and about 3 hours for K = 0xffffffff.
Table 4: Runtime estimations for 40-step MD4. The best estimations are marked with bold.
Hash Simplif. conflicts _ebest_ _nbest_ _|cbest|_
_0_
_1_
_symm_
_rand_
**1** **15 h 33 min** **3290** **303 494**
10 thousand 21 h 43 min 2530 210 008
100 thousand 52 h 32 min 2485 107 657
1 million 22 h 19 min 2400 148 518
10 million 34 h 27 min 1895 69 605
1 - - 10 thousand - - 100 thousand 81 h 31 min 2535 362 429
**1 million** **42 h 43 min** **2510** **182 724**
10 million 991 h 12 min 1890 1 671 849
**1** **19 h 16 min** **3395** 80 491
10 thousand 29 h 47 min 2725 181 267
100 thousand 22 h 44 min 2615 60 403
1 million 21 h 11 min 2530 151 567
10 million 59 h 28 min 1945 189 744
**1** **14 h 27 min** **3400** **75 823**
10 thousand 227 h 54 min 2660 1 098 970
100 thousand 20 h 22 min 2540 159 942
1 million 17 h 33 min 2455 225 854
10 million 81 h 3 min 1915 242 700
Figure 1 depicts how the objective function was minimized on the inversion problem for
_0hash. Here 10k stands for 10 thousand conflicts, 1m for 1 million conflicts and so on. The_
figures for the remaining three hashes can be found in Appendix A.
In Section 4 it was mentioned that in the estimating mode of Algorithm 5 it is possible
to find satisfying assignments of a given satisfiable CNF. That is exactly what happened for
_symmhash — a satisfying assignments was found for the CNF simplified by 100 thousand_
conflicts. It means that a preimage for symmhash generated by 40-step MD4 was found
just in few hours during the search for good thresholds for the cubing phase. However, the
goal was to find all preimages of the considered inversion problems (up to chosen value of
_L). That is why using the cubes produced with the help of the best cutoff thresholds, the_
conquer phase was run on all four inversion problems: 1-conflict-based for 0hash, symmhash,
22
-----
15
10
5
1
1m
10k
100k
10m
0 500000 1000000 1500000 2000000
1
0.5
|Col1|1 1m 10k 100 10m|
|---|---|
|||
Cubes
Figure 1: Minimization of the objective function on 40-step MD4, 0hash. The intersection
of two dotted lines shows the best estimation.
and randhash; 1-million-conflicts-based for 1hash. As a result, all subproblems were solved
successfully. The subproblems’ solving times in case of 0hash are shown in Figure 2.
mean
median
Figure 2: Kissat runtimes on subproblems from the conquer phase applied to
```
MD4inversion(0hash, 40, 0xffffffff, 12, 0x00000000)
```
.
23
-----
For 0hash and 1hash, no satisfying assignments were found, therefore the corresponding
inversion problems have no solutions. On the other hand, satisfying assignments were
found for hashes symmhash and randhash. The found thresholds, estimations, and the real
runtimes are presented in Table 5. In the header, sol stands for the number of solutions.
Note that the best estimation ebest was calculated only for L = 0x00000000, so for other
values of L it is equal to “-”. The right three columns present subproblems’ statistics: mean
solving time; maximum solving time; and standard deviation of times (when they are in
seconds). The minimum solving time is not reported since it was equal to 0.007 seconds in
all cases.
Table 5: Estimated and real runtimes (on 12 CPU cores) of the conquer phase for inversion
problems related to 40-step MD4. The best estimations from Table 4 are presented.
Hash _L_ _ebest_ real time sol mean max _sd_
`0x00000000` 15 h 33 min 20 h 9 min 0 2.84 sec 1 h 13.68
_0_ `0x00000001` - 19 h 25 min 0 2.61 sec 29 min 10.22
`0x00000002` - 34 h 27 min 1 5.7 sec 38 min 20.62
`0x00000000` 42 h 43 min 48 h 29 min 0 11.5 sec 26 min 30.23
_1_ `0x00000001` - 59 h 7 min 0 4.08 sec 17 min 11.4
`0x00000002` - 28 h 1 min 1 7.7 sec 17 min 18.68
_symm_ `0x00000000` 19 h 16 min 20 h 45 min 2 11.24 sec 18 min 21.1
_rand_ `0x00000000` 14 h 27 min 15 h 48 min 1 9.08 sec 38 min 21.59
The next iteration of Algorithm 3 (with L = 0x00000001) was executed for 0hash and
_1hash. Note that the same simplification and cutoff threshold as for L = 0x00000000 were_
applied to the corresponding CNFs. The conquer phase again did not find any satisfying
assignment. Finally, preimages for both hashes were found on the third iteration (L =
```
0x00000002), see Table 5. All found preimages are presented in Table 6. The obtained
```
results will be discussed in the next section.
### 7. Inverting 41-, 42-, and 43-step MD4 via Dobbertin-like Constraints
This section presents results on inverting 41-, 42-, and 43-step MD4. Finally, all MD4related results are discussed.
Recall that in the previous section on inverting 40-step MD4, Algorithm 3 was run on 40
CNFs: for each of 4 hashes, 2 values of K and 5 simplification types were tried. Note that
_K = 0x00000000 did not allow solving any 40-step-related problem. As for simplification_
types, for 3 hashes out of 4 the best estimations were obtained on 1-conflict-based CNFs,
while for the remaining one 1 million conflicts was the best. Following these results, in this
section only K = 0xffffffff is used, as well as only two mentioned simplification types.
Therefore only 8 CNFs were constructed for 41-step MD4, and the same for 42-, 43-, and
44-step MD4. Also it turned out that the best 40-step-related estimations were achieved
when at most 303 494 cubes were produced, see Table 4. That is why in this section the
24
-----
Table 6: Found preimages for 40-step MD4.
Hash Preimages
0xe57d8668 0xa57d8668 0xa57d8668 0xbc8c857b 0xa57d8668 0xa57d8668 0xa57d8668 0xcb0a1178
_0_
_1_
_symm_
0xa57d8668 0xa57d8668 0xa57d8668 0x307bc4e7 0xad02e703 0xe1516b23 0x981c2a75 0xc08ea9f7
0xe57d8668 0xa57d8668 0xa57d8668 0x1d236482 0xa57d8668 0xa57d8668 0xa57d8668 0x97a13204
0xa57d8668 0xa57d8668 0xa57d8668 0x991ede3 0x301e2ac3 0x5bed2a3d 0xe167a833 0x890d22f0
0xa57d8668 0xa57d8668 0xa57d8668 0xc8cf2f7c 0xa57d8668 0xa57d8668 0xa57d8668 0x61915bc1
0xa57d8668 0xa57d8668 0xa57d8668 0x2c017cc4 0xda6acfa2 0x55e9f993 0x50d83f7b 0x2d7d47a6
0xa57d8668 0xa57d8668 0xa57d8668 0x154f3b86 0xa57d8668 0xa57d8668 0xa57d8668 0x95b7616d
0xa57d8668 0xa57d8668 0xa57d8668 0xf3ca15df 0x7eb66f5e 0x446dc43f 0x7d8e2888 0xafe37a76
0xa57d8668 0xa57d8668 0xa57d8668 0xbb809ab0 0xa57d8668 0xa57d8668 0xa57d8668 0xab67285f
_rand_
0xa57d8668 0xa57d8668 0xa57d8668 0x85517639 0xc3eab3d 0x6edfba39 0xa1512693 0xaa686ac9
value of maxc is reduced from 2 000 000 to 500 000. The remaining input parameters of
Algorithm 5 are the same.
The same approach was applied as in the previous section: for each pair (steps, hash)
first the best cutoff threshold was found via Algorithm 5 for a CNF with added Dobbertin’s
constraints (L = 0x00000000), and then Algorithm 3 used the found threshold to run Cubeand-Conquer as a complete algorithm on each iteration. For 44 steps, no estimations were
obtained. On the other hand, for 41, 42, and 43 steps estimations were successfully calculated and they turned out to be comparable with that for 40 steps. Moreover, Algorithm 5
found preimages for two problems: 41 step and 1hash; 42 steps and 0hash. In Section 4
it was discussed that such a situation is possible if a given CNF is satisfiable. The found
estimations for 43-step MD4 are presented in Table 7. For all hashes, 1 conflict was the
best. For 41 steps, 1 conflict was better on 0hash and 1hash, while on remaining two hashes
1-million-conflicts based simplification was the winner. On 42-step MD4, 1 conflict was the
best for all hashes except 1hash.
Table 7: Runtime estimations for 43-step MD4. The best estimations are marked with bold.
Hash Simplif. conflicts _ebest_ _nbest_ _|cbest|_
**1** **15 h 26 min** **3 390** **103 420**
_0_
1 million - -
**1** **39 h 10 min** **3 395** **98 763**
_1_
1 million 52 h 5 min 2 575 121 969
**1** **37 h 51 min** **3 395** **81 053**
_symm_
1 million 50 h 7 min 2 555 253 489
**1** **49 h 13 min** **3 385** **120 619**
_rand_
1 million 86 h 23 min 2 565 246 972
25
-----
Figure 3 depicts how the objective function was minimized on the inversion problem
for 1hash in case of 43 steps. Figures for the remaining three 43-steps-related inversion
problems can be found in Appendix A.
4
3
2
1
|Col1|1 1m|
|---|---|
|||
0 100000 200000 300000 400000 500000
Cubes
Figure 3: Minimization of the objective function on the inversion problem
```
MD4inversion(1hash, 43, 0xffffffff, 12, 0x00000000). The intersection of two dotted lines
```
shows the best estimation among all simplification types.
Using the found cutoff thresholds, Algorithm 3 was run on all inversion problems with
_L = 0x00000000, and, as a result, for 43 steps preimages were found for all four hashes. For_
41 and 42 steps, preimages were found on the first or the second iteration of Algorithm 3.
The results are presented in Table 8. Here values 0 and 1 of L stand for 0x00000000 and
```
0x00000001, respectively, while sd stands for standard deviation in seconds. It can be seen
```
that at least some inversion problems turned out to be easier compared to that for 40-step
MD4. This phenomenon is discussed in the next subsection.
The subproblems’ solving times in case of 43 steps and 1hash are shown in Figure 4. In
Table 9, the found preimages for 43-step MD4 are presented. The corresponding tables for
41 and 42 steps can be found in Appendix B.
**7.1 Discussion**
**Correctness** The correctness of the found preimages was verified by the reference implementations from (Rivest, 1990). This verification can be easily reproduced since MD4 is
hard to invert, but the direct computation is extremely fast. First, the additional actions
(padding, incrementing, see Section 5), as well as the corresponding amount of the last steps
should be deleted. Then the found preimages should be given as inputs to a compression
function.
26
-----
Table 8: Estimated and real runtimes (on 12 CPU cores) of the conquer phase for inversion
problems related to 41-, 42, and 43-step MD4.
Steps Hash _L_ _ebest_ real time sol mean max _sd_
41
42
43
0 8 h 40 min 10 h 11 min 0 6.4 sec 17 min 16.77
_0_
1 - 21 h 23 min 1 12.41 sec 14 h 23 min 421.41
_1_ 0 37 h 45 h 10 min 3 9.78 sec 52 min 44.73
0 19 h 54 min 20 h 10 min 0 12.08 sec 17 min 24.28
_symm_
1 - 20 h 15 min 4 11.57 sec 17 min 23.66
_rand_ 0 16 h 6 min 17 h 25 min 1 10.05 sec 43 min 31.07
_0_ 0 19 h 36 min 22 h 32 min 3 11.68 sec 19 min 25.51
0 25 h 15 min 29 h 19 min 0 10.91 sec 1 h 14 min 45.61
_1_
1 - 39 h 1 16.38 sec 2 h 18 min 86.32
_symm_ 0 28 h 20 min 29 h 35 min 1 12.25 sec 32 min 19.98
0 21 h 16 min 21 h 30 min 0 10.22 sec 15 min 18.51
_rand_
1 - 20 h 35 min 3 9.34 sec 13 min 16.71
_0_ 0 15 h 26 min 17 h 14 min 2 7.23 sec 16 min 16.6
_1_ 0 39 h 10 min 42 h 16 min 1 18.64 sec 39 min 29.88
_symm_ 0 37 h 51 min 41 h 55 min 1 22.59 sec 34 min 46.44
_rand_ 0 49 h 13 min 51 h 21 min 1 18.51 sec 46 min 30.41
mean
median
Figure 4: Kissat runtimes on subproblems from the conquer phase applied to
```
MD4inversion(1hash, 43, 0xffffffff, 12, 0x00000000).
```
**Simplification** According to the estimations, in most cases the 1-conflict-based simplification is better than more advanced simplifications. On the other hand if only this simpli
27
-----
Table 9: Found preimages for 43-step MD4.
Hash Preimages
0xa57d8668 0xa57d8668 0xa57d8668 0xf48a97a3 0xa57d8668 0xa57d8668 0xa57d8668 0xd330e8ed
_0_
0xa57d8668 0xa57d8668 0xa57d8668 0x37c9ca21 0xe1df551f 0x7f49d66a 0x135a1c93 0x9e744bdb
0xa57d8668 0xa57d8668 0xa57d8668 0xb289afa0 0xa57d8668 0xa57d8668 0xa57d8668 0xaf2c850e
0xa57d8668 0xa57d8668 0xa57d8668 0x19c5ce09 0xcae6b29e 0xb2595b20 0xab3a433d 0xf6cdee42
0xa57d8668 0xa57d8668 0xa57d8668 0x82ef987a 0xa57d8668 0xa57d8668 0xa57d8668 0xe18fbc3b
_1_
0xa57d8668 0xa57d8668 0xa57d8668 0x558f3513 0xbf09004d 0x8fb490dd 0x502eca9 0xbd0e1a80
0xa57d8668 0xa57d8668 0xa57d8668 0xd1c33d35 0xa57d8668 0xa57d8668 0xa57d8668 0xc8519181
_symm_
0xa57d8668 0xa57d8668 0xa57d8668 0x8157aaf2 0xd7bdc37b 0xe52f3348 0xf17901d9 0x7e2de5a4
0xa57d8668 0xa57d8668 0xa57d8668 0x24f0e099 0xa57d8668 0xa57d8668 0xa57d8668 0xe57e4c54
_rand_
0xa57d8668 0xa57d8668 0xa57d8668 0x8fbbadcd 0xc0326ae6 0xe0e6a048 0x6217a3b9 0x15ee5a3b
fication type had been chosen, then the inversion problem for 1hash produced by 40-step
MD4 would have remained unsolved. The non-effectiveness of the advanced simplifications
is an interesting phenomenon which is worth investigating in the future.
**Classes of subproblems** Figures 2 and 4 show that in the conquer phase about 25% of
subproblems are extremely easy (runtime is less than 0.1 second) and there is a clear gap
between these subproblems and the remaining ones. Since this gap is much lower than mean
and median runtime, is seems promising to solve all extremely easy subproblems beforehand
and apply the corresponding reasoning to the remaining subproblems.
**Accuracy of estimations** The obtained estimations can be treated as accurate ones
since they are close to the real solving times (see tables 5 and 8). On average the real time
on inversion problems with L = 0x00000000 is 11 % higher than the estimated time, while
in the worst case for 40-step MD4 and 0hash the real time is 30 % higher. As for real time on
inversion problems with L = 0x00000001 and L = 0x00000002, the picture is different. In
some cases, the real time is still close to the estimated time for L = 0x00000000. However,
for MD4inversion(0hash, 41, 0xffffffff, 12, 0x00000001) the real time is 2.5 times higher,
while the standard deviation is also very high. It can be concluded that the heavy-tail
behavior occurs in this case (Gomes & Sabharwal, 2021). These results might indicate that
it is better to find its own cutoff threshold for each value of L, that corresponds to the 3rd
combination of Algorithm 3 and Algorithm 5 described at the beginning of Section 6. Note
that for those problems where their own thresholds were used, i.e. when L = 0x00000000,
the heavy-tail behavior does not occur.
**Hardness of inversion problems** It might seem counterintuitive that for 40-43 steps
the hardness of the inversion problems in fact is more or less similar. Recall that when
Dobbertin’s constraints are applied, values of 9 32-bit message words (our of 16) with
indices 0, 1, 2, 4, 5, 6, 8, 9, 10 are derived automatically (in a CNF this is done by UP), so
only 7 words remain unknown (see Subsection 2.5). It means that in the CNF 224 message
bits are unknown compared to 512 message bits when Dobbertin’s constraints are not added.
28
-----
It holds true for Dobbertin-like constraints as well. On the 40th step, the register value
is updated via a nonlinear function that takes as input an unknown word M [14] along
with registers’ values. That is why the 40th step gives a leap in hardness compared to 39
steps. On the next 8 steps, message words with the following indices are used for updating:
1, 9, 5, 13, 3, 11, 7, 15. It means that on steps 41, 42, and 43 the nonlinear function operates
with known (constant) M [1], M [9], and M [5], respectively. Therefore steps 41-43 do not
add any hardness. Rather, additional connections between registers’ values are added. As
for the remaining steps 44-48, unknown message words are used for updating, so each of
these steps gives a new leap in hardness. That is why no estimations were calculated for 44
steps earlier in this section — these inversion problems are much harder.
**Partially constant preimages** In all found preimages for steps 40 and 43, 9 out of
16 32-bit message words are equal to 0xa57d8668. These are the automatically derived
message words which depend on K. Recall that K = 0xffffffff was used in all cases.
However, in some preimages for 41 and 42 steps M [0] = 0x257d8668 while all remaining 8
message words are equal to 0xa57d8668. The reason is that in these cases the preimages
were found not in the first iteration of Algorithm 3, so on the 13th step the constant was
not K, but rather its slightly modified value.
### 8. Inverting Unconstrained 28-step MD5
As it was mentioned in Subsection 2.6, Dobbertin’s constraints are not applicable to MD5.
That is why in this study 28-step MD5 is inverted without adding any additional constraints,
like it was done in (Legendre et al., 2012). Recall that in this case for an arbitrary hash
there are about 2[384] preimages, but it is not easy to find any of them. Algorithm 5 in
its estimating mode is not applicable to MD5 either because the cubing phase gives too
hard subproblems for an unconstrained inversion problem, so no runtime estimation can be
calculated in reasonable time. On the other hand, since the considered inversion problem
has huge number of solutions, the incomplete solving mode of Algorithm 5 suits well for it.
First a CNF that encodes 28-step MD5 was constructed based on the encoding from
Subsection 5.3. The CNF has 7 471 variables and 54 672 clauses. The same four hashes
were considered for inversion as for MD4: 0hash, 1hash, symmhash, randhash. Therefore, 4
CNFs were constructed by adding corresponding 128 one-literal clauses to the original CNF.
Then these CNFs were simplified by CaDiCaL such that at most 1 conflict was generated.
Characteristics of the simplified CNFs are presented in Table 10.
Table 10: Characteristics of simplified CNFs that encode inversion problems for 28-step
MD5.
Hash Variables Clauses Literals
_0_ 6 814 50 572 199 596
_1_ 6 844 50 749 200 153
_symm_ 6 842 50 737 200 114
_rand_ 6 842 50 741 200 110
29
-----
The SAT solver EnCnC (see the beginning of Subsection 6.1) was run on these CNFs
in the incomplete solving mode. The following input parameters’ values were used:
- March cu.
- nstep = 10.
- minr = 0.
- N = 1 000.
- Kissat sc2021.
- maxst = 5 000 seconds.
- cores = 12.
- mode = solving.
The key parameter here is maxc (maximal number of generated cubes), for which the
following values were tried: 2 000 000; 1 000 000; 500 000; 250 000; 125 000; 60 000. Note
that the default value of maxc in EnCnC is 1 000 000. Recall that in the incomplete solving
mode, EnCnC stops if a satisfying assignment is found; if a CDCL solver is interrupted
due to a time limit on some subproblem, EnCnC continues working. The corresponding 6
versions of EnCnC with different values of maxc were run on the CNFs with the wall-clock
time limit of 1 day. In Table 11, the wall-clock solving times are presented. Also, the same
data is presented in Figure 5.
Table 11: Wall clock time for 28-step MD5 on a 12-core CPU. Here “-” means that the
solver was interrupted due to the time limit of 1 day. The best results are marked with
bold.
Solver _0hash_ _1hash_ _symmhash_ _randhash_
EnCnC-maxc=2m 1 h 47 min 1 h 41 min 39 min 1 h 36 min
EnCnC-maxc=1m 42 min 53 min 13 min 59 min
EnCnC-maxc=500k 48 min 32 min 22 min **15 min**
EnCnC-maxc=250k 38 min 4 min 41 min 37 min
EnCnC-maxc=125k 16 min 35 min **6 min** 20 min
EnCnC-maxc=60k **4 min** **3 min** 14 min 1 h 32 min
P-MCOMSPS - - -
Treengeling - - -
Additionally, two complete parallel SAT solvers were tried. The first one, P-MCOMSPS,
is the winner of the Parallel track in SAT Competition 2021 (Vallade, Frioux, Oanea, Baarir,
Sopena, Kordon, Nejati, & Ganesh, 2021). It is a portfolio solver built upon the widely-used
Painless framework (Frioux, Baarir, Sopena, & Kordon, 2017). The second one, treengeling (Biere, 2016), is a Cube-and-Conquer solver. It was chosen to compare EnCnC
30
-----
Figure 5: Runtimes of EnCnC in the incomplete solving mode on four MD5-28-related
inversion problems.
with a competitor built upon a similar strategy. Besides this, treengeling won several
prizes in SAT Competitions and SAT Races.
Let us discuss the results. Based on average runtime, the best version of EnCnC is
EnCnC-maxc=125k, while the worst is EnCnC-maxc=2m. Nevertheless, all versions
managed to find satisfying assignments for all 4 CNFs within the time limit. On 23 runs
out of 24, versions of EnCnC did it during solving the first 12 subproblems from the first
random sample (for the lowest values of the cutoff threshold). It means that Kissat did
not reach the time limit of 5 000 seconds in these cases. The only exception is EnCnCmaxc=60k on randhash, where on all 12 first subproblems Kissat was interrupted due to
the time limit, and then a satisfying assignment was found in one of the next 12 subproblems
from the same sample. As for competitors, they could not solve anything within the time
limit. In Table 12, the preimages found by EnCnC-maxc=2m are presented. It should be
noted that preimages for 0hash, 1hash, and randhash have not been published so far.
The found preimages were verified by the reference implementation from (Rivest, 1992).
It can be easily reproduced in the same way that was discussed in Subsection 7.1.
### 9. Related Work
In fact, SAT-based cryptanalysis was proposed in 1996 (Cook & Mitchell, 1996), but for the
first time it was applied to solve a real cryptanalysis problem in 2000 (Massacci & Marraro,
2000). In particular, a reduced version of the block cipher DES was analyzed via a SAT
solver. Since that publication, SAT-based cryptanalysis has been successfully applied to
analyze various block ciphers, stream ciphers, and cryptographic hash functions.
31
-----
Table 12: Preimages found by EnCnC-maxc=2m for 28-step MD5.
Hash Preimages
0xd825e4fb 0xa73fcaa9 0x660cd53d 0xb9308515 0x4677d4e0 0xcadcee62 0x40722cb3 0xf41a4b12
_0_
0xac2fdec3 0x9cbcb4a3 0xffcca30f 0x9a0e2026 0x475763e5 0x30ce233b 0xbef0cd57 0x1a6b39d
0xdfe6feeb 0xc4437a85 0x11af5182 0xe3b13f03 0x5103e1fc 0xea231da2 0xc3b513d1 0xb95fa9d7
_1_
0x7a2a331c 0x2ddf2607 0x699a2dae 0xc1827561 0xfe80aeed 0xcf45b09a 0x5b596c8f 0xd0265347
0x54032182 0x2a1693f1 0x1053aef3 0x9f4d7c87 0x9f0d5ba1 0xb43a63f8 0x4310aa89 0x9df4e0d8
_symm_
0xada73cbf 0x63fd55c2 0x49f1f4a0 0x5e05beff 0x6c149122 0x54a25f8e 0x12ef4bb0 0x78482fb4
0x120686db 0xad5834c6 0x7d660963 0x71c408fe 0x17cf4511 0x75df78de 0x544ae232 0x13745ecc
_rand_
0x9190f8a2 0x4878ab8d 0x43229cc7 0x5013f2de 0xd49b395a 0xa151b704 0x5f1dd4ec 0xc860dfb5
SAT-based cryptanalysis via CDCL solvers has been earlier applied to cryptographic
hash functions as follows. For the first time it was done in (Jovanovic & Janicic, 2005) to
construct benchmarks with adjustable hardness. In (Mironov & Zhang, 2006), a practical
collision attack on MD4 was performed. 39-step MD4 was inverted in (De et al., 2007;
Legendre et al., 2012; Lafitte et al., 2014; Gribanova, Zaikin, Kochemazov, Otpuschennikov,
& Semenov, 2017; Gribanova & Semenov, 2018). In (Gladush, Gribanova, Kondratiev,
Pavlenko, & Semenov, 2022), the hardness of practical preimage attacks on 43-, 45-, and
47-step MD4 was estimated. In (Gribanova & Semenov, 2020), an MD4-based function
was constructed and the full (48-step) version of this function was inverted. As for MD5,
in (Mironov & Zhang, 2006) and later in (Gribanova et al., 2017), practical collision attacks
on MD5 were performed. In (De et al., 2007), 26-step MD5 was inverted, while in (Legendre
et al., 2012) it was done for 27- and 28-step MD5. For the first time a collision for SHA-1
was found in (Stevens, Bursztein, Karpman, Albertini, & Markov, 2017) (in this very case it
was done partially by a CDCL solver). Step-reduced versions of SHA-0, SHA-1, SHA-256,
SHA-3, BLAKE-256, and JH-256 were inverted in (Nossum, 2012; Legendre et al., 2012;
Homsirikamol, Morawiecki, Rogawski, & Srebrny, 2012; Nejati, Liang, Gebotys, Czarnecki,
& Ganesh, 2017). An algebraic fault attack on SHA-1 and SHA-2 was performed in (Nejati,
Hor´acek, Gebotys, & Ganesh, 2018), while that on SHA-256 was done in (Nakamura, Hori,
& Hirose, 2021).
The following hard mathematical problems have been solved via Cube-and-Conquer:
the Erd˝os discrepancy problem (Konev & Lisitsa, 2015); the Boolean Pythagorean Triples
problem (Heule et al., 2016); Schur number five (Heule, 2018b); Lam’s problem (Bright
et al., 2021). In (Weaver & Heule, 2020), new minimal perfect hash functions were found.
Note that these hash functions are not cryptographic ones and find their application in
lookup tables. In the present paper, for the first time significant cryptanalysis problems
were solved via Cube-and-Conquer.
The present paper presents a general Cube-and-Conquer-based algorithm for estimating
hardness of SAT instances. Usually this is done by other approaches: the tree-like space
complexity (Ans´otegui, Bonet, Levy, & Many`a, 2008); supervised machine learning (Hutter,
Xu, Hoos, & Leyton-Brown, 2014); the popularity–similarity model (Almagro-Blanco &
Gir´aldez-Cru, 2022); backdoors (Williams, Gomes, & Selman, 2003).
32
-----
Backdoors are closely connected with Cube-and-Conquer. Informally, backdoor is a
subset of variables of a given formula, such that by varying all possible values of the backdoor’s variables simpler subproblems are obtained which can be solved independently via
a CDCL solver (Williams et al., 2003; Kilby, Slaney, Thi´ebaux, & Walsh, 2005). In fact,
such a set of values can be considered a cube, while choosing a proper backdoor and varying all corresponding values is a special way to generate cubes on the cubing phase of
Cube-and-Conquer. For given SAT instance and backdoor, hardness of the instance can
be estimated by processing a (relatively small) sample of subproblems (Semenov, Zaikin,
Bespalov, & Posypkin, 2011). The search for a backdoor with a minimal hardness was
reduced to minimization of a costly stochastic black-box functions in application to SATbased cryptanalysis in (Semenov & Zaikin, 2015; Kochemazov & Zaikin, 2018; Zaikin &
Kochemazov, 2021; Semenov, Pavlenko, Chivilikhin, & Kochemazov, 2022). In the present
paper, a similar function is minimized to find a cutoff threshold of the cubing phase of
Cube-and-Conquer rather than a backdoor.
### 10. Conclusions and Future Work
This paper proposed two algorithms. Given a hash, the first algorithm gradually modifies
one of twelve Dobbertin’s constraints for MD4 until a preimage for a given hash is found.
Potentially, this algorithm is applicable to some other cryptographic hash functions. The
second algorithm can operate with a given CNF in two modes. In the estimating mode,
values of the cutoff threshold of the cubing phase of Cube-and-Conquer are varied, and
the CNF’s hardness for each value is estimated via sampling. The threshold with the best
estimation can be naturally used to choose a proper computational platform and solve the
instance if the estimation is reasonable. This mode is general, so it can be applied to
estimate the hardness and solve hard SAT instances from various classes. In the incomplete
SAT solving mode, the second algorithm is a SAT solver, oriented on satisfiable CNFs with
many satisfying assignments.
The preimage resistance of two seminal cryptographic hash functions, MD4 and MD5,
was analyzed. In case of MD4, a combination of the first algorithm and the estimating mode
of the second algorithm was used. As a result, 40-, 41-, 42-, and 43-step MD4 were inverted
for the first time. In opposite to MD4, MD5 served as an example of a cryptographic hash
function for which no problem-specific constraints are added. 28-step MD5 was inverted
for two most regular hashes (128 1s and 128 0s) for the first time via the incomplete
SAT solving mode of the second algorithm. In other words, the first practical SAT-based
preimage attacks on the mentioned step-reduced MD4 and MD5 were proposed.
In the future it is planned to apply the proposed algorithms to analyze other cryptographic hash functions. Also we are going to investigate two MD4-related phenomena which
were figured out during the experiments. The first one is non-effectiveness (in most cases)
of an advanced simplification in application to the constructed CNFs. The second one is
an evident division of subproblems in the conquer phase to extremely simple ones and the
remaining ones.
33
-----
### Appendix A. Estimations for Step-reduced MD4
The following figures depict how the objective function was minimized on 40- and 43-step
MD4.
15
10
5
5
4
3
2
1
0.5
1
0.5
|Col1|1m 100k|
|---|---|
|||
|Col1|1|
|---|---|
0 500000 1000000 1500000 2000000
Cubes
0 100000 200000 300000 400000 500000
Cubes
(a) MD4-40, 1hash
500000 1000000 1500000
Cubes
(c) MD4-40, symmhash
|Col1|1 1m|1 1m|
|---|---|---|
||||
(d) MD4-43, symmhash
(b) MD4-43, 0hash
5
4
3
2
1
0.5
0 100000 200000 300000 400000 500000
Cubes
5
4
3
2
1
0.5
|Col1|1 1m|1 1m|
|---|---|---|
||||
0 100000 200000 300000 400000 500000
Cubes
(e) MD4-40, randhash
(f) MD4-43, randhash
Figure 6: Minimization of the objective function on 40- and 43-step MD4.
34
-----
### Appendix B. Found Preimages for Step-reduced MD4
Table 13: Found preimages for 41-step MD4.
0x257d8668 0xa57d8668 0xa57d8668 0xdafb914d 0xa57d8668 0xa57d8668 0xa57d8668 0x1edf9f78
_0_
0xa57d8668 0xa57d8668 0xa57d8668 0x12984195 0x97f0b6c 0xd9e5df17 0xabe482c7 0x23d98522
0xa57d8668 0xa57d8668 0xa57d8668 0x5c31dc3 0xa57d8668 0xa57d8668 0xa57d8668 0x52f59fb2
_1_
0xa57d8668 0xa57d8668 0xa57d8668 0x1e8a7cbb 0x3982e99f 0x812d980d 0x27b8d0b5 0xb81a00d1
0x257d8668 0xa57d8668 0xa57d8668 0xeaaf86e 0xa57d8668 0xa57d8668 0xa57d8668 0xc3b97274
0xa57d8668 0xa57d8668 0xa57d8668 0x21b8d189 0x15fc5540 0xd283c2c4 0x7d27396b 0x7bb74632
0x257d8668 0xa57d8668 0xa57d8668 0x5e8d818a 0xa57d8668 0xa57d8668 0xa57d8668 0x8fc29cce
_symm_
_rand_
_0_
0xa57d8668 0xa57d8668 0xa57d8668 0x8c6b49cc 0xe31a2c8d 0x9a5e1c5d 0x2dd896f5 0x1ed72fab
0x257d8668 0xa57d8668 0xa57d8668 0x9278c8f 0xa57d8668 0xa57d8668 0xa57d8668 0x4e3194eb
0xa57d8668 0xa57d8668 0xa57d8668 0x22efb603 0xe2b4a054 0xd74ec43 0xf09b0821 0xe4ca9fca
0x257d8668 0xa57d8668 0xa57d8668 0x6172bd01 0xa57d8668 0xa57d8668 0xa57d8668 0x8e35540f
0xa57d8668 0xa57d8668 0xa57d8668 0x4b8210a9 0xd5c0fedb 0x45c28d93 0x1b542bb8 0x74c28676
0xa57d8668 0xa57d8668 0xa57d8668 0x4b11d0ca 0xa57d8668 0xa57d8668 0xa57d8668 0x4c195670
0xa57d8668 0xa57d8668 0xa57d8668 0x76529071 0x68d3862d 0xdd3779df 0x768ce847 0x77e1b04e
0xa57d8668 0xa57d8668 0xa57d8668 0xcfbf3444 0xa57d8668 0xa57d8668 0xa57d8668 0xaac69f2f
0xa57d8668 0xa57d8668 0xa57d8668 0xbdaf1de9 0xfb9496dc 0x537e7a8c 0xd083975f 0xf3a5fc76
0xa57d8668 0xa57d8668 0xa57d8668 0xbfbf37eb 0xa57d8668 0xa57d8668 0xa57d8668 0xf3252a5c
0xa57d8668 0xa57d8668 0xa57d8668 0x3f829fe3 0x28c0fe6 0x27eadfa1 0xc87af86e 0x48fcd23d
Table 14: Found preimages for 42-step MD4.
0xa57d8668 0xa57d8668 0xa57d8668 0xecdab667 0xa57d8668 0xa57d8668 0xa57d8668 0xe3844a01
0xa57d8668 0xa57d8668 0xa57d8668 0xa3205929 0xfad1ea59 0xd2cae4d2 0x52149d55 0xc82cffbf
0xa57d8668 0xa57d8668 0xa57d8668 0xae60af85 0xa57d8668 0xa57d8668 0xa57d8668 0x8bcd69e3
0xa57d8668 0xa57d8668 0xa57d8668 0x59b8bf6 0x7755a76 0xfbe0b515 0xf9a31765 0x14d516a6
0xa57d8668 0xa57d8668 0xa57d8668 0xa9210d09 0xa57d8668 0xa57d8668 0xa57d8668 0xba9694ea
0xa57d8668 0xa57d8668 0xa57d8668 0x6a8157fe 0xd6566aae 0xbacb3d6c 0x1ec4854d 0x22357d65
0x257d8668 0xa57d8668 0xa57d8668 0xd8f77148 0xa57d8668 0xa57d8668 0xa57d8668 0x88275d15
_1_
0xa57d8668 0xa57d8668 0xa57d8668 0xcf6b92d0 0x4a8e498d 0x3beb0878 0xb55e027 0x87b4d62c
0xa57d8668 0xa57d8668 0xa57d8668 0xd1dce7ea 0xa57d8668 0xa57d8668 0xa57d8668 0xcbc2a90
_symm_
0xa57d8668 0xa57d8668 0xa57d8668 0xd9834f6d 0x5267d5d6 0x41a9cf18 0x71469663 0xbd507731
_rand_
0x257d8668 0xa57d8668 0xa57d8668 0xbd7389e6 0xa57d8668 0xa57d8668 0xa57d8668 0x3eb8ae3a
0xa57d8668 0xa57d8668 0xa57d8668 0x162c323e 0xa4056a04 0x9da74aac 0xfee2c77 0x8b25de8e
0x257d8668 0xa57d8668 0xa57d8668 0xc1748842 0xa57d8668 0xa57d8668 0xa57d8668 0xd7e32a57
0xa57d8668 0xa57d8668 0xa57d8668 0x21c5baab 0x552a7372 0xa21b2963 0x2fe88ffb 0xadfddb3
0x257d8668 0xa57d8668 0xa57d8668 0xc455558f 0xa57d8668 0xa57d8668 0xa57d8668 0xff87976a
0xa57d8668 0xa57d8668 0xa57d8668 0x3e82e858 0x46ad9cde 0x76f3b1d0 0x31aadb79 0x45cc1c91
35
-----
### References
Almagro-Blanco, P., & Gir´aldez-Cru, J. (2022). Characterizing the temperature of SAT
formulas. Int. J. Comput. Intell. Syst., 15 (1), 69.
Ans´otegui, C., Bonet, M. L., Levy, J., & Many`a, F. (2008). Measuring the hardness of SAT
instances. In AAAI, pp. 222–228.
Aoki, K., & Sasaki, Y. (2008). Preimage attacks on one-block MD4, 63-step MD5 and more.
In SAC, pp. 103–119.
Audet, C., & Hare, W. (2017). Derivative-Free and Blackbox Optimization. Springer Series
in Operations Research and Financial Engineering. Springer International Publishing.
Balyo, T., Froleyks, N., Heule, M., Iser, M., J¨arvisalo, M., & Suda, M. (Eds.). (2021). Pro_ceedings of SAT Competition 2021: Solver and Benchmark Descriptions. Department_
of Computer Science, University of Helsinki.
Balyo, T., & Sinz, C. (2018). Parallel Satisfiability. In Handbook of Parallel Constraint
_Reasoning, pp. 3–29. Springer._
Bard, G. V. (2009). Algebraic Cryptanalysis (1st edition). Springer Publishing Company,
Incorporated.
Biere, A. (2011). Preprocessing and inprocessing techniques in SAT. In HVC 2011.
Biere, A. (2016). Splatz, Lingeling, Plingeling, Treengeling, YalSAT Entering the SAT
Competition 2016. In Proc. of SAT Competition 2016 – Solver and Benchmark De_scriptions, pp. 44–45._
Biere, A., Fleury, M., & Heisinger, M. (2021a). CaDiCaL, Kissat, Paracooba entering the
SAT Competition 2021. In SAT Competition 2021 – Solver and Benchmark Descrip_tions, pp. 10–13._
Biere, A., Heule, M., van Maaren, H., & Walsh, T. (Eds.). (2021b). Handbook of Satisfiability
_- Second Edition, Vol. 336 of Frontiers in Artificial Intelligence and Applications. IOS_
Press.
B¨ohm, M., & Speckenmeyer, E. (1996). A fast parallel SAT-solver - efficient workload
balancing. Ann. Math. Artif. Intell., 17 (3-4), 381–400.
Bright, C., Cheung, K. K. H., Stevens, B., Kotsireas, I. S., & Ganesh, V. (2021). A SATbased resolution of Lam’s problem. In AAAI, pp. 3669–3676.
Carter, K., Foltzer, A., Hendrix, J., Huffman, B., & Tomb, A. (2013). SAW: the software
analysis workbench. In HILT, pp. 15–18.
Clarke, E., Kroening, D., & Lerda, F. (2004). A tool for checking ANSI-C programs. In
_Tools and Algorithms for the Construction and Analysis of Systems (TACAS 2004)._
Cook, S. A. (1971). The complexity of theorem-proving procedures. In STOC, pp. 151–158.
ACM.
Cook, S. A., & Mitchell, D. G. (1996). Finding hard instances of the satisfiability problem:
A survey. In Satisfiability Problem: Theory and Applications.
36
-----
Davis, M., Logemann, G., & Loveland, D. W. (1962). A machine program for theoremproving. Commun. ACM, 5 (7), 394–397.
De, D., Kumarasubramanian, A., & Venkatesan, R. (2007). Inversion attacks on secure
hash functions using SAT solvers. In SAT, pp. 377–382.
Dobbertin, H. (1996). Cryptanalysis of MD4. In FSE, pp. 53–69.
Dobbertin, H. (1998). The first two rounds of MD4 are not one-way. In FSE, pp. 284–292.
Dowling, W. F., & Gallier, J. H. (1984). Linear-time algorithms for testing the satisfiability
of propositional horn formulae. J. Log. Program., 1 (3), 267–284.
Frioux, L. L., Baarir, S., Sopena, J., & Kordon, F. (2017). PaInleSS: A framework for
parallel SAT solving. In SAT 2017, pp. 233–250.
Froleyks, N., & Biere, A. (2021). Single clause assumption without activation literals to
speed-up IC3. In FMCAD.
Garey, M. R., & Johnson, D. S. (1979). Computers and Intractability: A Guide to the Theory
_of NP-Completeness. W. H. Freeman._
Gladush, A., Gribanova, I., Kondratiev, V., Pavlenko, A., & Semenov, A. (2022). Measuring
the effectiveness of SAT-based guess-and-determine attacks in algebraic cryptanalysis.
In Parallel Computational Technologies, pp. 143–157.
Gomes, C. P., & Sabharwal, A. (2021). Exploiting runtime variation in complete solvers.
In Handbook of Satisfiability - Second Edition, Vol. 336 of Frontiers in Artificial In_telligence and Applications, pp. 463–480. IOS Press._
Gomes, C. P., & Sellmann, M. (2004). Streamlined constraint reasoning. In CP, pp. 274–289.
Gribanova, I., & Semenov, A. A. (2018). Using automatic generation of relaxation constraints to improve the preimage attack on 39-step MD4. In MIPRO, pp. 1174–1179.
Gribanova, I., & Semenov, A. A. (2020). Constructing a set of weak values for full-round
MD4 hash function. In MIPRO, pp. 1212–1217.
Gribanova, I., Zaikin, O., Kochemazov, S., Otpuschennikov, I., & Semenov, A. (2017). The
study of inversion problems of cryptographic hash functions from MD family using algorithms for solving Boolean satisfiability problem. In Mathematical and Information
_Technologies, pp. 98–113._
Hamadi, Y., Jabbour, S., & Sais, L. (2009). ManySAT: a parallel SAT solver. J. Satisf.
_Boolean Model. Comput., 6_ (4), 245–262.
Heule, M. (2018a). Cube-and-Conquer Tutorial. https://github.com/marijnheule/CnC/.
Heule, M. (2018b). Schur number five. In AAAI, pp. 6598–6606.
Heule, M., Kullmann, O., & Biere, A. (2018). Cube-and-conquer for satisfiability. In Hand_book of Parallel Constraint Reasoning, pp. 31–59. Springer._
Heule, M., Kullmann, O., & Marek, V. W. (2016). Solving and verifying the Boolean
Pythagorean triples problem via Cube-and-Conquer. In SAT, pp. 228–245.
Heule, M., Kullmann, O., Wieringa, S., & Biere, A. (2011). Cube and conquer: Guiding
CDCL SAT solvers by lookaheads. In HVC, pp. 50–65.
37
-----
Heule, M. J. H., & van Maaren, H. (2021). Look-ahead based SAT solvers. In Handbook
_of Satisfiability - Second Edition, Vol. 336 of Frontiers in Artificial Intelligence and_
_Applications, pp. 183–212. IOS Press._
Homsirikamol, E., Morawiecki, P., Rogawski, M., & Srebrny, M. (2012). Security margin
evaluation of SHA-3 contest finalists through SAT-based attacks. In CISIM, pp. 56–67.
Hutter, F., Xu, L., Hoos, H. H., & Leyton-Brown, K. (2014). Algorithm runtime prediction:
Methods & evaluation. Artificial Intelligence, 206, 79–111.
Hyv¨arinen, A. E. J., Junttila, T. A., & Niemel¨a, I. (2010). Partitioning SAT instances for
distributed solving. In LPAR, pp. 372–386.
Jovanovic, D., & Janicic, P. (2005). Logical analysis of hash functions. In FroCoS, pp.
200–215.
Kautz, H. A., Sabharwal, A., & Selman, B. (2021). Incomplete algorithms. In Handbook
_of Satisfiability - Second Edition, Vol. 336 of Frontiers in Artificial Intelligence and_
_Applications, pp. 213–232. IOS Press._
Kilby, P., Slaney, J. K., Thi´ebaux, S., & Walsh, T. (2005). Backbones and backdoors in
satisfiability. In AAAI, pp. 1368–1373.
Kochemazov, S. (2021). Exploring the limits of problem-specific adaptations of SAT solvers
in SAT-based cryptanalysis. In Parallel Computational Technologies, pp. 149–163.
Kochemazov, S., & Zaikin, O. (2018). ALIAS: A modular tool for finding backdoors for
SAT. In SAT, pp. 419–427.
Konev, B., & Lisitsa, A. (2015). Computer-aided proof of Erd˝os discrepancy properties.
_Artif. Intell., 224, 103–118._
Kuwakado, H., & Tanaka, H. (2000). New algorithm for finding preimages in a reduced
version of the md4 compression function. IEICE TRANS, Fundamentals, A, 83 (1),
97–100.
Lafitte, F. (2018). CryptoSAT: a tool for SAT-based cryptanalysis. IET Information Secu_rity, 12_ (6), 463–474.
Lafitte, F., Jr., J. N., & Heule, D. V. (2014). Applications of SAT solvers in cryptanalysis:
Finding weak keys and preimages. J. Satisf. Boolean Model. Comput., 9 (1), 1–25.
Legendre, F., Dequen, G., & Krajecki, M. (2012). Encoding hash functions as a SAT
problem. In ICTAI, pp. 916–921.
Leurent, G. (2008). MD4 is not one-way. In FSE, pp. 412–428.
Marques-Silva, J. P., & Sakallah, K. A. (1999). GRASP: A search algorithm for propositional
satisfiability. IEEE Trans. Computers, 48 (5), 506–521.
Massacci, F., & Marraro, L. (2000). Logical cryptanalysis as a SAT problem. J. Autom.
_Reasoning, 24_ (1/2), 165–203.
Menezes, A., van Oorschot, P. C., & Vanstone, S. A. (1996). Handbook of Applied Cryptog_raphy. CRC Press._
Mironov, I., & Zhang, L. (2006). Applications of SAT solvers to cryptanalysis of hash
functions. In SAT, pp. 102–115.
38
-----
Nakamura, K., Hori, K., & Hirose, S. (2021). Algebraic fault analysis of SHA-256 compression function and its application. Inf., 12 (10), 433.
Nejati, S., & Ganesh, V. (2019). Cdcl(crypto) SAT solvers for cryptanalysis. In CASCON.
Nejati, S., Hor´acek, J., Gebotys, C. H., & Ganesh, V. (2018). Algebraic fault attack on
SHA hash functions using programmatic SAT solvers. In CP, pp. 737–754.
Nejati, S., Liang, J. H., Gebotys, C. H., Czarnecki, K., & Ganesh, V. (2017). Adaptive restart
and CEGAR-based solver for inverting cryptographic hash functions. In VSTTE.
Nossum, V. (2012). SAT-based preimage attacks on SHA-1. Master’s thesis, University of
Oslo, Department of Informatics.
Rivest, R. L. (1990). The MD4 message digest algorithm. In CRYPTO, pp. 303–311.
Rivest, R. L. (1992). The MD5 message-digest algorithm. RFC, 1321, 1–21.
Rossi, F., van Beek, P., & Walsh, T. (Eds.). (2006). Handbook of Constraint Programming,
Vol. 2 of Foundations of Artificial Intelligence. Elsevier.
Sasaki, Y., & Aoki, K. (2009). Finding preimages in full MD5 faster than exhaustive search.
In Joux, A. (Ed.), EUROCRYPT, pp. 134–152.
Semenov, A., Zaikin, O., & Kochemazov, S. (2021). Finding effective SAT partitionings via
black-box optimization. In Black Box Optimization, Machine Learning, and No-Free
_Lunch Theorems, pp. 319–355. Springer International Publishing._
Semenov, A. A., Otpuschennikov, I. V., Gribanova, I., Zaikin, O., & Kochemazov, S. (2020).
Translation of algorithmic descriptions of discrete functions to SAT with applications
to cryptanalysis problems. Log. Methods Comput. Sci., 16 (1).
Semenov, A. A., Pavlenko, A., Chivilikhin, D., & Kochemazov, S. (2022). On probabilistic
generalization of backdoors in Boolean satisfiability. In AAAI.
Semenov, A. A., & Zaikin, O. (2015). Using Monte Carlo method for searching partitionings
of hard variants of Boolean satisfiability problem. In PaCT.
Semenov, A. A., Zaikin, O., Bespalov, D., & Posypkin, M. (2011). Parallel logical cryptanalysis of the generator A5/1 in BNB-grid system. In Parallel Computing Technologies,
pp. 473–483.
Soos, M., Nohl, K., & Castelluccia, C. (2009). Extending SAT solvers to cryptographic
problems. In SAT, pp. 244–257.
Starnes, D., Yates, D., & Moore, D. (2010). The Practice of Statistics. W. H. Freeman.
Stevens, M., Bursztein, E., Karpman, P., Albertini, A., & Markov, Y. (2017). The first
collision for full SHA-1. In CRYPTO, pp. 570–596.
Tseitin, G. S. (1970). On the complexity of derivation in propositional calculus. In Studies in
_constructive mathematics and mathematical logic, part II, Seminars in mathematics,_
Vol. 8, pp. 115–125. V.A. Steklov Mathematical Institute, Leningrad.
Vallade, V., Frioux, L. L., Oanea, R., Baarir, S., Sopena, J., Kordon, F., Nejati, S., &
Ganesh, V. (2021). New concurrent and distributed Painless solvers: P-MCOMSPS, PMCOMSPS-COM, P-MCOMSPS-MPI, and P-MCOMSPS-COM-MPI. In SAT Com_petition 2021 – Solver and Benchmark Descriptions, pp. 40–41._
39
-----
Wang, X., Lai, X., Feng, D., Chen, H., & Yu, X. (2005). Cryptanalysis of the hash functions
MD4 and RIPEMD. In EUROCRYPT, pp. 1–18.
Wang, X., & Yu, H. (2005). How to break MD5 and other hash functions. In Cramer, R.
(Ed.), EUROCRYPT, pp. 19–35.
Weaver, S. A., & Heule, M. (2020). Constructing minimal perfect hash functions using SAT
technology. In AAAI.
Williams, R., Gomes, C. P., & Selman, B. (2003). Backdoors to typical case complexity. In
_IJCAI, pp. 1173–1178._
Zaikin, O. S., & Kochemazov, S. E. (2021). On black-box optimization in divide-and-conquer
SAT solving. Optimization Methods and Software, 36 (4), 672–696.
Zaikin, O. (2022). Inverting 43-step MD4 via Cube-and-Conquer. In IJCAI-ECAI, p. in
press.
40
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/2212.02405, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "GOLD",
"url": "https://doi.org/10.1613/jair.1.15244"
}
| 2,022
|
[
"JournalArticle"
] | true
| 2022-12-05T00:00:00
|
[
{
"paperId": "2bfa20069212d04c1662b960f52d8a404678b461",
"title": "A SAT Solver and Computer Algebra Attack on the Minimum Kochen-Specker Problem (Student Abstract)"
},
{
"paperId": "6094988c9ebe2621e6cccc7bc6f52248117e19d0",
"title": "Frontiers in Artificial Intelligence and Applications"
},
{
"paperId": "9ccd3a54cdf81c9afd5bdaa94dd2d166d619ea37",
"title": "Characterizing the Temperature of SAT Formulas"
},
{
"paperId": "9b39397b07bef4594e2c16ef0785dfff362d546b",
"title": "Better Decision Heuristics in CDCL through Local Search and Target Phases"
},
{
"paperId": "daca018a1564f76df66b40d455bb5d2e11573fa7",
"title": "Inverting 43-step MD4 via Cube-and-Conquer"
},
{
"paperId": "fd2df3e9941353581aaaf0151fec27332370022e",
"title": "On Probabilistic Generalization of Backdoors in Boolean Satisfiability"
},
{
"paperId": "a37a79d09d12de03d01435bd29abf06416f48e21",
"title": "Single Clause Assumption without Activation Literals to Speed-up IC3"
},
{
"paperId": "d643ab573da72863f23fd41439b7a05e69f30aad",
"title": "Algebraic Fault Analysis of SHA-256 Compression Function and Its Application"
},
{
"paperId": "b7bd88de7ad9057c407818261e4bc0759ddef9b9",
"title": "Exploring the Limits of Problem-Specific Adaptations of SAT Solvers in SAT-Based Cryptanalysis"
},
{
"paperId": "429b90bbaa51d77a85218b4ab1511b3e504aee27",
"title": "Incomplete Algorithms"
},
{
"paperId": "b0cc5cb047b4172a7584796b5844eaa109eeade1",
"title": "Exploiting Runtime Variation in Complete Solvers"
},
{
"paperId": "74d57bf83d227253d224944fba2991c1fd4a3887",
"title": "Look-Ahead Based SAT Solvers"
},
{
"paperId": "cb392d0364323c2b7b6ade3ec5878829f3db7364",
"title": "A SAT-based Resolution of Lam's Problem"
},
{
"paperId": "b8612d7866be434e8b6b7c6f4c1d891af8e683c5",
"title": "Constructing a Set of Weak Values for Full-round MD4 Hash Function"
},
{
"paperId": "fa7fa280009fef4e37a92e6fb8fe5616f439d991",
"title": "CDCL(Crypto) SAT solvers for cryptanalysis"
},
{
"paperId": "ce1e092f8c30c1e51b279335e5e34af82dfe9d7d",
"title": "Constructing Minimal Perfect Hash Functions Using SAT Technology"
},
{
"paperId": "055326dd234e552d99d35fb8050048befc760e6b",
"title": "On black-box optimization in divide-and-conquer SAT solving"
},
{
"paperId": "5f27124b4767e1c10ca89ee7e66c37c529152d1d",
"title": "The Resolution of Keller’s Conjecture"
},
{
"paperId": "61b87dd4da58a5297c166281564dc3a475db4a50",
"title": "C. Audet and W. Hare: Derivative-free and blackbox optimization. Springer series in operations research and financial engineering"
},
{
"paperId": "83721103a6fd5535e943b1b575cf70862c2322a8",
"title": "Handbook of Applied Cryptography"
},
{
"paperId": "5cb07d365d3d8748e8859242616aae2f1a1ba918",
"title": "CryptoSAT: a tool for SAT-based cryptanalysis"
},
{
"paperId": "f21a2281efaebfe84896dcec5cc7cf50afb0ac70",
"title": "Algebraic Fault Attack on SHA Hash Functions Using Programmatic SAT Solvers"
},
{
"paperId": "ce0a8071295381cf0fd5afc7a50b9c93895fc8f6",
"title": "ALIAS: A Modular Tool for Finding Backdoors for SAT"
},
{
"paperId": "ceb4e06ad017a2a8135fe0b633a53f7c1dccab21",
"title": "Translation of Algorithmic Descriptions of Discrete Functions to SAT with Applications to Cryptanalysis Problems"
},
{
"paperId": "aeaec30319e3fe46b1ca0d07128474b9858e0a2d",
"title": "On Cryptographic Attacks Using Backdoors for SAT"
},
{
"paperId": "fbce880aa2d637fc1d1258b8e49bdbe4bcb90f90",
"title": "Using automatic generation of relaxation constraints to improve the preimage attack on 39-step MD4"
},
{
"paperId": "db9c1c061dfbf30f325e69565d6f5b0ca99172b1",
"title": "Schur Number Five"
},
{
"paperId": "1f8fb24ce9fdff1c4b3aae52c54a7c0d22e5b8ce",
"title": "Proceedings of SAT Competition 2017: Solver and Benchmark Descriptions"
},
{
"paperId": "8d6870ab6c3abbb7b952fde8ffb2e57da8bf2747",
"title": "PaInleSS: A Framework for Parallel SAT Solving"
},
{
"paperId": "f648783368d2010858070efe778a6a253e656295",
"title": "The First Collision for Full SHA-1"
},
{
"paperId": "fe5093abefd1ded617159182381c99415d9c910b",
"title": "Handbook of Petroleum Refining"
},
{
"paperId": "61d4b2a48bf4bd73853abccf9c30e1dcf8f73a47",
"title": "Adaptive Restart and CEGAR-Based Solver for Inverting Cryptographic Hash Functions"
},
{
"paperId": "10d8f3ef07b15ee728061b9830d848896cdcb9ca",
"title": "Encoding Cryptographic Functions to SAT Using TRANSALG System"
},
{
"paperId": "1be4e96f3ae691fe7f4fbc2a480c820884a0a7ca",
"title": "Solving and Verifying the Boolean Pythagorean Triples Problem via Cube-and-Conquer"
},
{
"paperId": "c602c0296a7c0f06d34ce72370307175a9908e87",
"title": "Using Monte Carlo Method for Searching Partitionings of Hard Variants of Boolean Satisfiability Problem"
},
{
"paperId": "90dccd01df4e1d727fa60b2138e153506e6832a4",
"title": "Applications of SAT Solvers in Cryptanalysis: Finding Weak Keys and Preimages"
},
{
"paperId": "223a9ae6eefd65f7992eb4fa667196840f90d62e",
"title": "Computer-aided proof of Erdős discrepancy properties"
},
{
"paperId": "84ceeba430725f3c674d31740a24a465da8e2ebc",
"title": "SAW: the software analysis workbench"
},
{
"paperId": "9ab2c60ed17e936223677abbb067f226ddee0709",
"title": "From a logical approach to internal states of Hash functions how SAT problem can help to understand SHA-⋆ and MD⋆"
},
{
"paperId": "16367cdea1bc84a3ccafed980e06bca397faf536",
"title": "Encoding Hash Functions as a SAT Problem"
},
{
"paperId": "14e1f20764d6733499e0a07062de42433522bddf",
"title": "Algorithm runtime prediction: Methods & evaluation"
},
{
"paperId": "5cf85100d09c743cbb4b87d0d02c41e8eb3729f0",
"title": "Security margin evaluation of SHA-3 contest finalists through SAT-based attacks"
},
{
"paperId": "18ed960ea0146e674e91275a02abc67ff443f099",
"title": "Cube and Conquer: Guiding CDCL SAT Solvers by Lookaheads"
},
{
"paperId": "6aa30550fe4fb7ef608fc71da766de0c2754b9cc",
"title": "Preprocessing and Inprocessing Techniques in SAT"
},
{
"paperId": "a829753d623f6adc0de2ca73d5474f6d429170e3",
"title": "Parallel Logical Cryptanalysis of the Generator A5/1 in BNB-Grid System"
},
{
"paperId": "feca806dc20be4a72a57b45b9ce7ebfcded08038",
"title": "A machine program for theorem-proving"
},
{
"paperId": "1da2db87c1f4627d2213fe596070e9a60a613e95",
"title": "Advanced Meet-in-the-Middle Preimage Attacks: First Results on Full Tiger, and Improved Results on MD4 and SHA-2"
},
{
"paperId": "508f96b85440dad366be3067fb5e537ffd948a6d",
"title": "Partitioning SAT Instances for Distributed Solving"
},
{
"paperId": "e36075def919440a084f59516403a4ce88702ab2",
"title": "Preimage Attacks on One-Block MD4, 63-Step MD5 and More"
},
{
"paperId": "fbae47bf980dd14adc0dea81ec3dbf90a8aa67c9",
"title": "Extending SAT Solvers to Cryptographic Problems"
},
{
"paperId": "e8da2dc0b6481f30b2b4ffd5bad6b835616e3f37",
"title": "ManySAT: a Parallel SAT Solver"
},
{
"paperId": "1dc4a98755b542437e6e0df316282a47134aa702",
"title": "Finding Preimages in Full MD5 Faster Than Exhaustive Search"
},
{
"paperId": "577cd1ef41f6f2f12f391516cb87029b86ad5aa7",
"title": "Measuring the Hardness of SAT Instances"
},
{
"paperId": "31e1077691e174fdcb3de4d2b326bfd013459b55",
"title": "Backdoor Trees"
},
{
"paperId": "7c5f4b9e9a486d60cc04eac9d395ff640b0433eb",
"title": "MD4 is Not One-Way"
},
{
"paperId": "04347767ad76a974bd019f1174c1ef1c03f4fed2",
"title": "Tradeoffs in the Complexity of Backdoor Detection"
},
{
"paperId": "f072e529551ca1307957fd84300c56384ab271a6",
"title": "Inversion Attacks on Secure Hash Functions Using satSolvers"
},
{
"paperId": "f4f4eeabc6863a3d3571ffa67e5df19fcd9f02f2",
"title": "Applications of SAT Solvers to Cryptanalysis of Hash Functions"
},
{
"paperId": "4aa5a98543ffdf74f475c971b34ef3747d552c26",
"title": "Logical Analysis of Hash Functions"
},
{
"paperId": "f13245779a5d5f7eb419992de69261059249b335",
"title": "Backbones and Backdoors in Satisfiability"
},
{
"paperId": "9ed6d2833ae315921f754e40802916340b148701",
"title": "How to Break MD5 and Other Hash Functions"
},
{
"paperId": "45a2ce8717638071073477dbd0b535b7c87b220b",
"title": "Cryptanalysis of the Hash Functions MD4 and RIPEMD"
},
{
"paperId": "66966f6cc6fc7e5b018ff6e157c991eaa1a3f1b9",
"title": "Streamlined Constraint Reasoning"
},
{
"paperId": "5f5335a25d2b4ef0fb099ae917ff98d8afea97e3",
"title": "A Tool for Checking ANSI-C Programs"
},
{
"paperId": "b84109df224e4f19963d6ffd99d11c9d0ec89db2",
"title": "Backdoors To Typical Case Complexity"
},
{
"paperId": "745f8201364a5c1e22c0ff54b28e1b2eeed34822",
"title": "The Practice of Statistics"
},
{
"paperId": "71efcfb787d75e3718aee49270943d7c514972f7",
"title": "Logical Cryptanalysis as a SAT Problem"
},
{
"paperId": "cf6c60f8875e69216056d652f734da5d0f253a5c",
"title": "New Algorithm for Finding Preimages in a Reduced Version of the MD4 Compression Function(Special Section on Cryptography and Information Security)"
},
{
"paperId": "38be6e613f2c30d21bffe8b468bc0cd46edba0d0",
"title": "GRASP: A Search Algorithm for Propositional Satisfiability"
},
{
"paperId": "b7dde73e8867e9c0f2c0825fb614f3dbba23b9ea",
"title": "The First Two Rounds of MD4 are Not One-Way"
},
{
"paperId": "f3ee74ac186d75938635b1673e5b9972a01c4f17",
"title": "Tools and Algorithms for the Construction and Analysis of Systems"
},
{
"paperId": "e8d8f6eb224942cf294bd499efe21f752f5907a5",
"title": "A fast parallel SAT-solver — efficient workload balancing"
},
{
"paperId": "e8783ef34f73773fb679aef58d987839313e1b3c",
"title": "Cryptanalysis of MD4"
},
{
"paperId": "8861608e6c6b42f8883aec62b98997477229eeb8",
"title": "Foundations of Artificial Intelligence"
},
{
"paperId": "e9ce5ad132f753624a017dc036f45eff45839265",
"title": "The MD4 Message-Digest Algorithm"
},
{
"paperId": "3a500741f6d989f8e672d8aadfd979678b87d09f",
"title": "A Certified Digital Signature"
},
{
"paperId": "d2712ce067a604c61a28778babebeced19b6bf8e",
"title": "A Design Principle for Hash Functions"
},
{
"paperId": "7e33296dfff963d595d2121f14a7a0bd5c187188",
"title": "Linear-Time Algorithms for Testing the Satisfiability of Propositional Horn Formulae"
},
{
"paperId": "048ce334f8b749edcf6c327a8172552cbcced865",
"title": "The complexity of theorem-proving procedures"
},
{
"paperId": null,
"title": "Inverting Cryptographic Hash Functions via Cube-and-Conquer — Re-sults and Source code"
},
{
"paperId": "020f9c70c1b19d95ebcff9e3e00a2eca4f74bf4e",
"title": "contrast of"
},
{
"paperId": null,
"title": "Measuring the effectiveness of SAT-based guess-and-determine attacks in algebraic cryptanalysis"
},
{
"paperId": "88f708e1fae27707df1f786b5f53ab13a4c736e0",
"title": "Black Box Optimization, Machine Learning, and No-Free Lunch Theorems"
},
{
"paperId": "3fe726775b69e763a0e0f5cef0adb1c9e0aa8b55",
"title": "Evaluating the Hardness of SAT Instances Using Evolutionary Optimization Algorithms"
},
{
"paperId": "e954c49711559956144a148a427a534464230c36",
"title": "Finding Effective SAT Partitionings Via Black-Box Optimization"
},
{
"paperId": null,
"title": "CaDiCaL, Kissat, Paracooba entering the SAT Competition 2021"
},
{
"paperId": null,
"title": "New concurrent and distributed Painless solvers: P-MCOMSPS, PMCOMSPS-COM, P-MCOMSPS-MPI, and P-MCOMSPS-COM-MPI"
},
{
"paperId": null,
"title": "2018a). Cube-and-Conquer Tutorial. https://github.com/marijnheule/CnC"
},
{
"paperId": "93adadb9a64b39f19fcd1e7551da2b0eb9be1a0c",
"title": "Cube-and-Conquer for Satisfiability"
},
{
"paperId": "533afa1a39800a966040a124ed930e8d96b53efe",
"title": "Parallel Satisfiability"
},
{
"paperId": "1fd6a36125d944522c4adbfadb6b9abe6613ac27",
"title": "Derivative-Free and Blackbox Optimization"
},
{
"paperId": "e6d3fbf35cdffe39b203eab15ade8c055fc79e0a",
"title": "Splatz , Lingeling , Plingeling , Treengeling , YalSAT Entering the SAT Competition 2016"
},
{
"paperId": "684b5498d99dfef1fe1beac1a6b3864f0088d13b",
"title": "Computers And Intractability A Guide To The Theory Of Np Completeness"
},
{
"paperId": "480dbd4ca40a1e94ca6a724b7e61f39ae296a075",
"title": "Inverting Thanks to SAT Solving - An Application on Reduced-step MD*"
},
{
"paperId": "880abfd4c5421973fd0b25325bedf3248f5065a9",
"title": "SAT-based preimage attacks on SHA-1"
},
{
"paperId": null,
"title": "Algebraic Cryptanalysis (1st edition)"
},
{
"paperId": "027282e993a97cad07797a980c32b0bc83f61989",
"title": "Fundaments of Branching Heuristics"
},
{
"paperId": "9e3ea009e70de20efcd4614e79b9a0aabbb621e0",
"title": "The First Two Rounds of MD 4 are Not One-Way Extended Abstract"
},
{
"paperId": "cd8c2397f984313d5450f8d3c8aca7a8c3f5009e",
"title": "Finding Hard Instances of the Satissability Problem: a Survey"
},
{
"paperId": "e860216dba46960dec044ab477e8cd823cd608c3",
"title": "Finding hard instances of the satisfiability problem: A survey"
},
{
"paperId": "674c0d017a03c9716d490517ad1a58c72f8d9c5d",
"title": "On the Complexity of Derivation in Propositional Calculus"
},
{
"paperId": "ccfe55b17272e2ebf855844e96d5ac2aefa4d841",
"title": "The Study of Inversion Problems of Cryptographic Hash Functions From MD Family Using Algorithms for Solving Boolean Satisfiability Problem"
},
{
"paperId": null,
"title": "Second-preimage resistance : for any given input x , it is computationally infeasible to find x (cid:48) such that x (cid:48) (cid:54) = x, h ( x ) = h ( x (cid:48) )"
},
{
"paperId": null,
"title": "Algorithm 5 can be easily modified to be oriented on finding only one solution in the conquer phase"
},
{
"paperId": null,
"title": "The proposed runtime estimation is a stochastic costly black-box objective function (Audet & Hare, 2017; Semenov, Zaikin, & Kochemazov, 2021). The algorithm minimizes this objective function"
},
{
"paperId": null,
"title": "For each hash and each value of L its own threshold is found. Therefore A is the estimating mode of Algorithm 5 followed by a Cube-and-Conquer solver"
},
{
"paperId": null,
"title": "the estimating mode"
},
{
"paperId": null,
"title": "the incomplete SAT solving mode, a solution can be found only for a satisfiable CNF, and even in this case this is not guaranteed because of the time limit for the CDCL solver"
}
] | 32,295
|
en
|
[
{
"category": "Political Science",
"source": "external"
},
{
"category": "Political Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/024c65380b37cfbcc1fc5a595dd16fa20a382507
|
[
"Political Science"
] | 0.964944
|
The performance of power and citizenship: David Cameron meets the people
|
024c65380b37cfbcc1fc5a595dd16fa20a382507
|
International journal of cultural studies
|
[
{
"authorId": "145562991",
"name": "P. Lunt"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"Int j cult stud",
"Int J Cult Stud",
"International Journal of Cultural Studies"
],
"alternate_urls": [
"http://ics.sagepub.com/",
"http://www.sagepub.com/journals/Journal200946/title",
"https://journals.sagepub.com/home/ics"
],
"id": "19c52db0-c559-4319-93e6-289771c58034",
"issn": "1367-8779",
"name": "International journal of cultural studies",
"type": "journal",
"url": "http://www.sagepub.com/journal.aspx?pid=196"
}
|
How do citizens respond to and engage with the performance of political power in the context of mainstream media? Through an analysis of two television programmes aired during the UK Brexit referendum campaign of 2016, a picture emerges of citizenship as the performative disruption of the performance of power. In the programmes the then UK prime minister, David Cameron, met members of the public for a mediated discussion of key issues in the Brexit referendum. Their interactions are analysed here as a confrontation between the performance of citizenship and power reflecting activist modalities of disruptive citizenship played out in the television studio. The article ends with reflections on questions about political agency as individualistic forms of disruptive political autonomy.
|
**The performance of power and citizenship: David Cameron meets the people**
Peter Lunt
School of Media, Communication and Sociology
University of Leicester, UK
[Pl108@le.ac.uk](mailto:Pl108@le.ac.uk)
**Abstract**
How do citizens respond to and engage the performance of political power in the context of
mainstream media? Through an analysis of two television programmes aired during the UK
Brexit referendum campaign of 2016 a picture emerges of citizenship as the performative
disruption of the performance of power. In the programmes the then UK Prime Minister, David
Cameron, met members of the public for a mediated discussion of key issues in the Brexit
referendum. Their interactions are analysed here as a confrontation between the performance
of citizenship and power reflecting activist modalities of disruptive citizenship played out in
the television studio. The article ends with reflections on questions about political agency as
individualistic forms of disruptive political autonomy.
**Keywords: political discourse; debate; civility; autonomy; performance**
**Introduction**
In this article I examine the mediated juxtaposition and interrelation of the performance of
power and citizenship in the context of two television programmes aired during the UK Brexit
referendum campaign of 2016. The Prime Minister of the day, David Cameron, head of the
campaign to remain in the European Union (EU), appeared on the shows, one at the launch of
the campaign and one just days before the actual referendum. The shows were adaptions of the
popular BBC current affairs panel discussion programme Question Time on which Cameron
fielded questions from members of the public moderated by a programme host. In the first
programme he was also interviewed by a political journalist in front of the television audience
before taking questions.
There are several reasons why these shows are significant in relation to the intersection
of political communication and the theme of this special issue, ‘Citizenship and performance’.
First, as popular culture, the programmes are part of the diversity of forms of political
communication ranging from set piece party political broadcasts, political interviews, televised
-----
debates, talk shows and appearances by politicians on current affairs and popular daytime
television programmes (Craig, 2016). Second, the mediated engagement between the Prime
Minister and members of the public raises questions about the role of the media in public
engagement that crosses boundaries between public discourse and politics. Third, as relatively
unscripted public exchanges these engagements are performative as ‘individuals,
organizations, and parties moved “instinctively” to hook their actions into the background
culture in a lively and compelling manner, working to create an impression of sincerity and
authenticity rather than one of calculation and artificiality, to achieve verisimilitude’
(Alexander et al., 2006: 1). The analysis illustrates that the performance of power by the Prime
Minister was a construction of personal authenticity and political authority, and that the
performance of citizenship by lay participants was as a disruption of the performance of power
in the form of individualized dissent (Ruiz, 2014; 2016). This article provides an analysis of
the two television programmes drawing on dramaturgy (Goffman, 1959) and performance
studies followed by an analysis of the genealogy of both the performance of power and
citizenship. The article ends with a discussion of the meaning and modality of the performance
of citizenship as a subjectivity constructed as autonomy.
The Brexit referendum was a major political event that stood in a complex relation to
traditional party political affiliation and engaged the public in a relatively open debate between
sides representing the answer to a single question: whether to remain in the EU or to leave.
Cameron, it has been widely acknowledged, as an ex-public relations man, was a consummate
political performer across a range of media contexts (Craig, 2016). In a similar way to former
US President Barack Obama, he had developed a style of political leadership that sought to
overcome the excesses of spin and media management characteristic of the Clinton and Blair
years (Craig, 2016). Until the EU referendum campaign, Cameron appeared as a highly skilled
media performer accomplished at managing a variety of communication contexts such as press
conferences, interrogative interviews with political journalists and set piece speeches such as
the annual party conference. He was equally at home meeting the people in mediated town hall
meetings or sitting on the sofa of current affairs television shows as he was when debating in
the Chamber of the House of Commons. Craig (2016) argues that such multiply skilled
performances across varied communication genres and contexts aims to manage, if not resolve,
tensions between authenticity and performance, between the public politician and the private
individual, between factual broadcasting and entertainment, and between legitimacy based on
expertise and public popularity. Such a leadership style also aims to avoid or overcome the
public cynicism that potentially results from the visibility of techniques of media management
-----
and spin that draws attention to the strategies of political communication rather than
substantive claims and policy commitments (Capella and Jamieson, 1997).
Two television programmes were aired on free-to-air channels during the Brexit
campaign in which Cameron came face to face with members of the public. The programmes
were an extension of a series of similar encounters with members of the public that he made
during his time as Prime Minister of a coalition government between his Conservative Party
and the Liberal Party from 2010 to 2015. These ‘meetings’, called PM Direct, were made
available on YouTube supported by transcripts available on the website of the Prime Minister’s
office, and were held in workplaces (e.g. Caterpillar, EasyJet and Rolls Royce). In these
contexts, Cameron stood, often shirt-sleeved, amongst the employees or members who were
seated at floor level between him and a single fixed camera and on a bank of chairs behind him.
This production format created a space in which Cameron was framed by the audience as he
delivered intense, short statements of the key points of his campaign agenda in response to
questions from members of the audience. Unlike similar examples of political discussion
programmes, the shows were unmoderated and Cameron managed the questions from the floor
as well as being the only ‘guest’ on the show. An important feature of such encounters was that
members of the audience were restricted to asking questions and there were no follow-up or
supplementary questions. This lack of interactivity allowed Cameron the opportunity to treat
questions as cues, to which he responded by delivering well-rehearsed statements of his
policies or campaign agenda.
The Conservative Party adapted the PM Direct format for the 2015 General Election
campaign. The context moved from workplace meetings to spaces in which greater control
could be exercised over access and the production format of the events, and in which the
audience acted as cheerleaders creating an excited emotional climate as Cameron pronounced.
However, these occasions were constructed to create the impression of being public meetings.
There is a long tradition in UK parliamentary election campaigns in which candidates hold
public meetings in their constituencies in which they meet members of the public and address
their questions and concerns. Such occasions are often robust and boisterous exchanges in
which political discourse meets vernacular, committed expressions of politics. In contrast, in
the versions of PM Direct aired during the 2015 General Election campaigns, another feature
of the shows was the generation of the emotional climate of an election rally in which the
‘audience’ reacted positively and emotionally; Cameron was fired up and the audience was
fired up. This was a simulation of the traditional campaign stump, but as a highly controlled
and disciplined occasion in which enthusiastic party members created a sense of spontaneity
-----
as a background to Cameron’s mini speeches. The Conservative Party won an unexpected
majority in the 2015 General Election and one of the election promises had been to hold a
referendum on Britain’s membership of the EU that led to the referendum in 2016 when the
two programmes analysed here were broadcast.
**Background to the EU referendum**
In a political campaign, notwithstanding the increasing importance of digital communication
technologies, television remains a key site for performative embedding of campaign messages
and engagement with national audiences. The communication styles of political leaders have
adapted to make use of the diverse forms and contexts of communication balanced by
disciplined campaigning and media management strategies. In the three weeks of Brexit
campaigning between 2 and 22 June 2016, 15 mainstream television programmes were aired
across several genres. These included political interviews conducted by well-known
journalists, debates between leading representatives of the Remain and Leave campaigns,
audience discussion programmes with members of the public and the programmes examined
here, variants of Question Time in which key campaigners faced questions from members of
the public. The BBC played a central role in staging 10 of the 15 television programmes during
the campaign; ITV held three events, Sky News two and Channel 4 one. Recent commentators
(Chadwick, 2013; Craig, 2016) have suggested that after a period of hyperbole about digital
campaigning there is growing recognition that television is finding its place in contemporary
campaigns, partly through innovations in programme forms and partly by complementing and
intersecting with digital and social media campaigns.
In the UK a referendum once triggered by an Act of Parliament is managed not by
government or political parties but by campaign groups that are chosen by the Electoral
Commission to act as the official voice of the two campaigns representing the two sides of the
referendum decision – in this case, Remain or Leave the EU. Bids are invited by groups that
wish to represent each side of the referendum and the two groups chosen attract public
campaign funds. The campaign for Remain was modelled on the Conservative Party campaign
of the 2015 General Election. Having governed as part of a coalition with the Liberal
Democrats since 2010, the Conservatives won an unexpected parliamentary majority in 2015.
The results recorded increased support for both Conservatives and Labour, the Liberal vote
collapsed and notably, there was a dramatic increase in nationalist votes for both the Scottish
National Party (SNP) and the UK Independence Party (UKIP). The Daily Telegraph offered
-----
an insightful analysis of the successful Conservative election campaign organized by Lynton
Crosby that demanded discipline from members of the Conservative Party, a campaign agenda
that focused on economic policy, negative campaigning against their main rivals – Labour and
the Liberal Democrats – focused on the party leaders (Ed Miliband and Nick Clegg), and David
Cameron fronting the campaign in presidential style (Swinford, 2015).
The deployment of Cameron ‘front and centre’ formed a key part of the Remain
campaign strategy as it had in the 2015 Conservative General Election campaign. This was
partly justified by his high opinion poll ratings with 41 per cent approval during this period
(Boffey, 2015), although these were moderated by perceptions of Cameron as uncommitted
and unemotional, and by negative reactions to his upper-class social background. Craig (2016)
discusses the strategy adopted to overcome these public perceptions of deploying Cameron’s
high-level media skills to make a direct appeal to the broader electorate. For example, Cameron
handled interviews such as that by the BBC’s political journalist Andrew Marr by skilfully
challenging the host’s framing of Conservative policy, answering only the questions he wanted
to answer and refusing to be drawn into areas that might be problematic (Craig, 2016). The
challenge facing Cameron and his advisors was to find ways of bringing his undoubted
rhetorical and presentational skills into contact with a broader public to popularize his
leadership. Consequently, skilful performance in political interviews was supplemented by a
mixed communication strategy that kept Cameron in the public eye and aimed to soften his
public image and spread his popular appeal.
**The television programmes**
**_Sky News_**
Cameron kickstarted the Remain campaign with an appearance on a _Sky News special_
programme on 2 June 2016. The show began with an interview conducted by Faisal Islam, a
political journalist, in front of a live television audience followed by a moderated Q&A session
with members of the studio audience hosted by Sky newsreader and presenter, Kay Burley.
Islam opened his questioning by asking Cameron to stick to the facts about migration and to
outline the figures for net migration during his leadership. This was challenging ground for
Cameron because he had made a feature of his critique of the Brexit campaign by stating that
it was based on false claims, anticipating subsequent debates about fake news and post-truth
political discourse, and here he was being asked about the failure of his government to meet
promises made in the General Election campaign to reduce migration to tens of thousands a
-----
year. Cameron gave a straight answer by admitting that 600,000 more people had entered the
UK than had moved to other countries since he had come to power. When pressed as to whether
he had broken a manifesto promise, he provided an intriguing justification by shifting the
ground from a manifesto ‘promise’ to an ‘ambition’, suggesting that the relatively better
performance of the UK economy during his period of office compared to Continental Europe
had led to the creation of many new jobs that had attracted workers from abroad. A strategic
advantage of this answer is that it shifted the focus on to Cameron’s central campaigning
agenda, the economic benefits of EU membership and the risks of leaving. He argued that the
target “remains the right ambition for Britain” and that trying to cut immigration by leaving
the EU and pulling out of the single market would be “madness” because of the economic
damage it would cause. A number of further questions followed from Islam, most significantly
challenging Cameron’s references to the First World War to illustrate the potential security
dangers of Brexit, which Islam suggested was an example of “fearmongering”.
After further questions that were less challenging, the show changed gear (and genre),
morphing into a mediated popular press conference or political talk show (Craig, 2016). Burley
moderated the Q&A session between members of the audience and Cameron. This combination
of production formats was a challenge as the robust exchange with a professional journalist
was followed immediately by a context that required softer skills to engage members of the
public. What became immediately apparent was that Cameron continued with his strategy from
PM Direct of treating questions as triggers or cues for campaign sound bites or as an expression
of concern or lack of understanding of social or political issues. He saw his role as combining
the provision of public information and a therapeutic alleviation of public fears and concerns.
Rather than seeing indignation about the EU as an expression of substantial political critique
addressing substantive political questions, his stance was that it reflected ignorance and
anxiety. For example, one participant, identified as a businessman, asked Cameron to reflect
on the “personal damage the scaremongering has done to your legacy.” Cameron appealed to
personal authenticity in the shape of his core political commitments: “I don’t accept it is
scaremongering. I am genuinely worried about Britain leaving the single market.” He then
linked his campaign focus on the economic risks of leaving the EU to his political authority:
“Frankly, I think the job of the prime minister is to warn about potential dangers as well as to
talk about the upsides and the opportunities there are by being a member of this organization.”
In addition, Cameron emphasized his reliance on and trust in a variety of experts who supported
the claim that leaving the EU would be to the economic detriment of the UK and linking this
to his political authority: “But if I didn’t listen to the IMF, to the OECD, to the TUC, to the
-----
CBI, to the governor of the Bank of England – if I didn’t listen to any of these people, I would
not be doing my job and I would not be serving this country.” As the campaign unfolded, Brexit
campaigners were able to characterize such claims to authority as representing the interests of
the great and the good: in other words, the establishment. Aversion to and criticism of the
establishment is a key plank of populist political discourse (Jagers and Walgrave, 2007), which
Cameron opened himself up to by invoking a consensus of expertise in favour of Remain.
Another questioner expressed concerns that during the Brexit campaign the Prime
Minister shared a platform with the Mayor of London who he had strongly opposed as the
Labour candidate in the mayoral election of the previous year. Criticizing the Mayor’s apparent
support or legitimation of terrorist groups, Cameron responded:
We had a lively election campaign in London, I didn’t think it was the right choice
some of the people he shared a platform with. The right thing for the PM to do is to
work together. Sadiq and I disagree about many things; we’ll try and work together and
on this issue of Europe we agree. We buried our disagreements and appeared on a
platform.
From Cameron’s perspective, the contingencies of a referendum necessarily realigned politics
across party lines and, as leader of the Conservative Party and Prime Minister, he would now
find himself opposing the arguments and positions of some of his own party colleagues and
working for the Remain campaign, which included many liberal or left-leaning organizations,
public figures, and politicians. From the perspective of members of the public, however, the
dissociation of Cameron from his role as leader of the Conservative Party and Prime Minister
was not taken lightly. For example, one participant suggested the referendum was a vote of
confidence in the government and in Cameron himself. Intriguingly, some political
commentators ridiculed participants for such questions, but the difficulty of separating political
commitments and allegiances from the question of membership of the EU was and remains
non-trivial.
The limits of Cameron’s communication strategy on this programme were well
illustrated by his exchange with a literature student on the Sky News programme. The student,
identified as Soraya Bouazzaoui, stated her concerns: “The entire campaign was nothing but
scaremongering; no valid facts; no pros and cons and that everything I’ve seen makes voting
into the EU look worse.” The campaign, in other words, was high on persuasion and low on
fact and argument, and significantly, this intelligent, informed, articulate member of the public
was thinking of the referendum as a choice between ‘voting in’ and ‘voting out’. Referenda
are, however, usually deployed following parliamentary agreement on legislation that has
-----
significant constitutional implications, which is then put to the public for their assent.
Cameron’s original strategy was to negotiate significant changes to the UK’s position in
Europe, to get parliamentary approval for the changes, and then to put these new conditions to
a referendum. In this scenario, the question put to the public would have been to ask them
whether they agreed or disagreed with the new conditions for UK membership of the EU.
However, Cameron was only able to negotiate adjustments to the UK’s conditions of
membership and, in this context, the referendum was drafted as an in/out vote giving equal
weight to both sides and triggering more existential questions about membership of the EU.
However, as is evident in Cameron’s performance on this show, the Remain campaign avoided
substantive political questions of migration and sovereignty to focus on economic policy.
Having expressed her concerns about the conduct of the Brexit campaigns, Bouazzaoui
put her substantive question about the reassurances that Cameron had repeatedly made that
remaining in the EU would make the UK safer in response to terrorist threats. Referring to
concerns raised by Middle East states about Turkey’s relations to and perceived support for
terrorist groups, she questioned whether being in the EU meant that there were no risks in
foreign policy. Cameron’s response was characteristic, saying that he would address the two
issues that Bouazzaoui had raised:
First, the positive case for staying. I think there is a positive case. I think we’ll be better
off as a country, with more jobs. I think we’ll keep our country moving forward, we’ll
get things done in the world, whether it’s tackling climate change or indeed standing
up to Islamic terrorism … and also, we’ll be safer; strength in numbers.
This is a graphic illustration of Cameron’s approach to questions from members of the public
as a cue to deliver his campaign messages. However, the question was a legitimate and serious
one, and Cameron’s response clearly frustrated Bouazzaoui, who interrupted him:
That’s not answering my question. Let me finish now, because I’ve seen you interrupt
many people before. Let me finish. I’m an English Literature student, I know waffling
when I see it, OK. I’m sorry, but you’re not answering my question – how can you
reassure people who want to vote out that we are safe from extremism when we are
willing to work with a government like Turkey who want to be part of the EU when
they are under heavy accusation?
Cameron’s discomfort was evident and he tried to get back on track by saying that he had “got
it” and addressed the question of Turkey’s potential accession to the EU:
There is no prospect of Turkey joining the EU in decades. They applied in 1987, they
have to complete 35 chapters. One has been completed so far. At this rate they will join
-----
in the year 3000. There are lots of reasons to vote one way or vote the other way. Turkey
is not going to join the EU any time soon, every country, every parliament, has a veto.
There are lots of things to worry about in this referendum campaign. I absolutely think
that is not a prospect, it’s not going to happen.
This exchange illustrates a number of aspects of the interaction between the performance of
power and of citizenship in this programme. It demonstrates Cameron’s strategy of taking
questions as cues to which he responds with rehearsed campaign speeches. The passage also
demonstrates that an important aspect of the performance of citizenship in this context is
refusing the subject position of the audience to Cameron’s pronouncements, to bring power to
account by insisting on the relevance of answers, and disrupting the performance of power.
Cameron’s strategy of treating questions as expressions of concern that he takes as needing
reassurance, information or contradiction leaves this participant, members of the audience, and
by extension, the public, frustrated.
The programme demonstrated that the public were not in agreement with the Remain
campaign’s focus on the economic consequences of leaving the EU, and that a combination of
substantive political questions related to migration, the legal framework of the EU, sovereignty,
the impact of migration on public services and the efficacy of the government’s austerity
policies were all implicated in deciding how to vote in the referendum. Furthermore,
Cameron’s deflection of questions and his skilled practice of turning to his own agenda raised
serious questions about both his claims to authenticity and political authority, on which his
enviable popularity ratings had been based up to this point.
Press reaction to the programme was ambivalent. There was recognition that Cameron
had got his agenda across despite the distraction of a hostile interview and having to manage
the relationship with members of the public. In contrast, there was an acknowledgement that
the interactions with members of the public were less convincing and seemed to illustrate a gap
between the campaign agenda and public concerns. Interestingly this did not split neatly along
the political affiliation of the papers – for example, Michael Deacon, writing in The Telegraph:
‘The studio audience didn’t think much of him, and he knew it. It was no disaster. But if you
wondered why Mr Cameron didn’t fancy a proper debate: now you know’ (Deacon, 2016).
**The BBC**
Shortly before the EU referendum Cameron appeared on a BBC programme to meet the people
in an adaptation of the Question Time format, moderated by resident host David Dimbleby.
-----
This version of the programme differed in significant ways from the standard version of the
show on which members of the studio audience are selected by the host, guided by the
production team, to ask questions to a panel representing the main political parties plus
celebrity guests. In the programme commissioned for the referendum, there was no panel and
instead, David Cameron fielded all the questions
The producers and the host had learned from the Sky News programme and dealt with
Cameron’s tendency to not answer questions and shift topic onto his campaign agenda by
clustering questions thematically. Consequently, although Cameron shifted topic in response
to the questions, he found himself back on the same ground in the next question. The effect of
this was exacerbated by the programme format which was unlike in the panel version of the
programme in which different members of the panel voice alternative responses to audience
questions, and to contest these among each other before the host turns back to the audience for
supplementary questions and comments. In contrast, in this version of the programme one
question to the Prime Minister was rapidly followed by another.
The first cluster of questions addressed the impact on the political culture of the Brexit
campaign asking, for example, if it had “soured the political climate in the UK” by amplifying
antagonisms. In response, Cameron attempted to draw a distinction between political
commitment, passion and aggression, arguing that the committed use of reason, argument and
rhetoric is essential to politics. He then invited the audience to contemplate what distinguishes
reasonable/appropriate from unreasonable/inappropriate arguments and sentiments in political
discourse and public debate. In this he positioned himself as on the ‘right’ side of these
oppositions, claiming that his politics combined authentic personal commitment with political
authority backed by arguments and claims backed by evidence. His opponents, by implication,
were represented as political opportunists prepared to say anything to win, and consequently
lacking both personal authenticity and political authority (Craig, 2016).
These reflections on civility in public and political discourse are all very interesting,
but Cameron sidestepped the point that the questions were addressed to the conduct of his
campaign as much as to the Brexit campaign and to the use of negative campaigning to discredit
opponents. Nevertheless, Cameron pressed ahead, aiming to justify the comparison between
himself and his political opponents. He focused on Nigel Farage, leader of the populist UKIP
and a key figure in the campaign to leave the EU, although not part of the official ‘Leave’
group. Cameron referred to a Brexit campaign poster by UKIP that used a photograph of
refugees crossing the border into Bosnia-Herzegovina with the headline ‘Breaking Point’. He
argued that Farage was “wrong in fact and wrong in motivation”, and that the aim of Brexit
-----
campaigners was an “attempt to frighten and divide people.” In the campaign, Brexit
campaigners, particularly Boris Johnson, were able to turn this argument against Cameron by
pointing to inconsistencies in his position on Europe, thereby challenging the authenticity of
his position and characterizing his focus on the potential economic ills of leaving the EU as
‘project fear’. At this point, the host intervened to ask, “Has your side been guilty of that?”
articulating a commitment to impartiality as a moderator of the broader public debate. The
theme continued including a question that challenged Cameron on the ‘Brexit budget’ prepared
by the Chancellor to demonstrate the effects of leaving Europe on taxation and public spending.
Cameron’s reply suggested that his concerns were authentic, expressing his “genuine concern
for the economic impact of leaving the EU” and citing, once more, the support of independent
experts.
Following the exchange on the conduct of the campaigns was a series of questions and
answers on Cameron’s own future: would he resign if the country voted to leave? Would he
call a general election if the vote was to leave the EU? These questions reflected the central
role that Cameron played in the campaign, and although he tried to argue that the campaign
was not about him, he shifted to his main agenda that we should remain for the sake of the
economy, jobs, safety, security, and because being part of Europe strengthened the UK: “It
comes down to a question of the economy and we need to work together – to grow the economy
and beat terrorists.”
How did Cameron find himself in such a difficult, compromised performative context?
In the language of the history of the present (Foucault, 1977, 1984; Garland, 2014), a line can
be traced back to his previous ‘meetings’ with members of the public in PM Direct allied to a
leadership style that aimed to combine personal authenticity with political authority, and a
disciplined approach to campaigning that included a presidential style with Cameron at the
centre, negative campaigning against rivals and a focus on economic policy. The field of
emergence for this configuration of leadership and campaigning styles was partly due the
demands of the UK coalition government of Conservative and Liberal parties between 2010
and 2015 that demanded efforts at public communication as policies did not always clearly
follow party lines. During this time there was also increasing public support for nationalist and
populist parties that required renewed forms of popular communication from established
political parties. However, the PM Direct events did not create a stage that afforded the
opportunity for authenticated engagement with members of the public but instead, were
‘managed shows’ (Thompson, 1995) in which Cameron and his team selected the places and
audiences and set the rules of interaction. In contrast, as we have seen, the two television shows
-----
in which he met the people in the Brexit campaign were managed by the broadcasters and gave
opportunities for the performance of disruptive citizenship. Instead of a controlled context that
afforded the illusion of public engagement while allowing Cameron to deliver his campaign
message, he found himself involved in a contested performative space. So how did members
of the public find themselves occupying space in the television studio and challenging the
performance of power?
The difficulties experienced by Cameron and the opportunity afforded to citizens was
a function of the production format of the programme as a mise en scène for the performances
of power and citizenship. The two programmes included significant variations on the Question
_Time format, a popular political panel talk show in which guest members of the public put_
topical social and political questions to a panel. Question Time is a microcosm of the role of
factual programming in public service broadcasting assuming a politics of pluralism,
represented by the different panel members who stand for the main political parties as
competitive interest groups and, therefore, representing a particular idea of public
accountability understood as a fair and balanced representation of the views of different
competing interests in the political sphere (Karppinen, 2007). The transformation of the
programme format in which the panel was replaced by Cameron represents a shift from the
idea of communication of politics in which different positions are put in front of the public
(democratic pluralism) with commentary from expertise (elite democracy), to the appropriation
of public broadcasting as a vehicle for a political campaign. In the traditional formulation, the
responsibility of public service broadcasting is to create contexts in which competing interests
have equal opportunities to state their arguments and to provide an expert commentary on those
views. In contrast, placing Cameron in the place of the panel made the Prime Minister the
single recipient of questions, transforming the programme into a popular version of the press
conference in comparison to the panel format adopted in Question Time, which constituted a
debate between different political positions.
Episodes of Question Time usually proceed in a sequence in which the host invites a
question from the studio audience and then asks panel members to answer. In this sense,
_Question Time_ is characterized by contestation, argument and often conflict between panel
members as they debate alternative answers to the questions under the direction and scrutiny
of the programme host. The host then goes back to the audience for supplementary questions
and reactions, and finally the person who asked the question gives their reactions and
reflections. In contrast, Cameron’s previous mediated town hall PM Direct ‘meetings’ were a
pale reflection of the dynamics of _Question Time. There was no variety of responses to_
-----
questions, just Cameron’s, and no display of divergent views or contestation in front of the
audience. The rhythm of exchanges and arguments and the emotional flow of the programmes
were altered considerably by these changes, becoming a series of Q&A rather than a question
followed by a robust exchange and opportunity for further audience engagement. In terms of
the flow of emotions, instead of a dispersed exchange of feelings, views and political
commitments, the direction of sentiment was focused on Cameron.
**Reflections and conclusions**
How are we to think about the performance of citizenship in these programmes and the
implications for understanding political subjectivity and agency? These disruptive encounters
appear to be the work of individuals asserting their rights to visibility in public and to challenge
the performance of power. In terms of asserting communicative rights participants
communicate in a performative practice similar to Isin and Ruppert’s (2015) account of digital
citizenship as rights claiming practice. However, these appear to be political acts undertaken
by individuals expressing their autonomy by occupying a space (mainstream television studios)
in which the performance of power is made visible and realized through the interaction between
the performance of power and citizenship. The lay participants on these programmes are not
there to press individual claims and nor are they there to represent an emerging collective or
connective (Bennett and Segerberg, 2012); they are there as citizens, as individuals aiming to
have their say, bring power to account and disrupt attempts to persuade. This is a form of what
Dayan (2009) calls ‘monstration’ and takes up his argument that as television broadcasting
converges with digital culture it can be understood by analogy to the bulletin board on which
individual citizens post their messages to the public.
Such performances of citizenship are agentic, skilful deployments of material and
symbolic resources in staged interactions articulated as individualized forms of dissent (Ruiz,
2014, 2016). The television discussion programmes in which Cameron met the people reflect
‘the ways in which citizens – from protestors in Occupy movements … to participants [in]
street performances reclaim public space as a place to play out, both expressively and
deliberatively, struggles for recognition and new political subjectivities’ (Rovisco and Ong,
2016: 3). In this sense although appearing on television their actions reflected recent activism
and protest in what Gerbaudo (2016) terms ‘the digital popular’. The invasion of the television
studio and programme space reflects the transgressions of space by the Occupy movement.
This is an incursion into mainstream media culture deploying some of the tactics of new protest
-----
movements in the name of individual citizens. Along with the invasion of public spaces, new
social movements experiment with radical forms of democracy by making use of the resources
of digital media in reclaiming the square (Rovisco and Ong, 2016). In so doing, they engage
with activities within the square that reflect conceptions of direct democracy and civic virtue
(Dagger, 1997). Just as digital and social media provide social movements with new resources
that ameliorate their lack of access to mainstream media resources so, here, performative
disruption seeks to influence through visibility and public impact and by providing models of
alternative political practices.
Perhaps what is at stake here is that the disruption of political communication and the
occupation of the places of media production appear to express the position of individual
citizens intervening in public communication and debate. Such activism appears to suggest a
trajectory from personal concerns that are projected into the public sphere through performance
as a social practice that inserts personal issues and commitments into mediated public life.
There are other examples of a trajectory from personal commitment and action to political
debate, such as when individuals protest about the environmental implications of global
systems of production and distribution through individual and localized practices of consumer
activism (Lekakis, 2013). One way to make sense of this is as a combination of political
autonomy and democracy as a social practice (Gray, 2000). Or, following Dagger (1997), to
argue that participants perform individualized dissent related to questions of public interest in
civically virtuous expressions of individual political autonomy as:
… social actors, embedded in collective representations and working through symbolic
and material means, implicitly orient towards others as if they were actors on a stage
seeking identification with their experiences and understandings from their audiences.
(Alexander et al., 2006: 2)
These programmes stage an encounter between performances of power and citizenship
(Goffman, 1961), that instantiates the blurring of boundaries in contemporary political culture
reflecting tropes of digital citizenship in their disruption of power (Isin and Rupert, 2015) and
the digital popular as sites of occupation that engage new political subjectivities (Gerbaudo,
2016). In this article I have explored the way that these trends spill over into mainstream, linear
media to disrupt the performance of power in a staged encounter by autonomous, individual
political subjects (Gray, 2000). Traditional differentiations between the state and the body
politic (Habermas, 1987), between questions of politics and values (Rawls, 1993), and between
civil and uncivil discourses and actions (Mouffe, 2005), are all potentially blurred in the current
conjuncture typified by the example of when Cameron met the people.
-----
**References**
Alexander JC, Giesen B and Mast JL (eds) (2006) _Social Performance: Symbolic Action,_
_Cultural Pragmatics, and Ritual. Cambridge: Cambridge University Press._
Bennett WL and Segerberg A (2013) The Logic of Connective Action. Cambridge: Cambridge
University Press.
Boffey D (2015) David Cameron maintains high approval rating, despite Labour’s poll lead.
_The_ _Guardian,_ 15 February. Available at:
www.theguardian.com/politics/2015/feb/14/opinium-poll-david-cameron-maintains
approval-rating (accessed on 31 July 2018).
Capella JN and Jamieson KH (1997) Spiral of Cynicism: The Press and the Public Good. New
York: Oxford University Press.
Chadwick A (2013) _The Hybrid Media Systems: Politics and Power. Oxford: Oxford_
University Press.
Craig G (2016) Performing Politics. Cambridge: Polity Press.
Dagger R (1997) _Civic Virtues: Rights, Citizenship, and Republican Liberalism. Oxford:_
Oxford University Press.
Dayan D (2009) Sharing and showing: Television as monstration. The Annals of the American
_Academy of Political and Social Science 625, The end of television? Its impact on the world_
(so far), September: 19–31.
Deacon M (2016) ‘I know waffling when I see it!’ David Cameron takes a Brexit roasting. The
_Telegraph, 2 June. Available at: www.telegraph.co.uk/news/2016/06/02/i-know-waffling-_
when-i-see-it-david-cameron-takes-a-brexit-roast/ (accessed on 31 July 2018).
Foucault M (1977) Discipline and Punish: The Birth of the Prison. New York: Pantheon.
Foucault M (1984) Nietzsche, genealogy, history. In: Rabinow P (ed) The Foucault Reader.
New York: Pantheon. 76 - 100
Garland D (2014) What is a ‘history of the present’? On Foucault’s genealogies and their
critical preconditions. Punishment & Society 16(4): 365–384.
Gerbaudo P (2016) Occupying the digital-popular. In: Rovisco M and Ong J (eds) Taking the
_Square: Mediated Dissent and Occupations of Public Space. London: Rowman & Littlefield,_
37 – 54.
Goffman E (1959) The Presentation of Self in Everyday Life. Hamondsworth: Penguin.
Goffman E (1961) Encounters: Two Studies in the Sociology of Interaction. New York: The
Bob Merrell Company.
-----
Gray J (2000) Two Faces of Liberalism. Cambridge: Polity Press.
Habermas J (1987) The Theory of Communicative Action, Volume 2: Lifeworld and System: A
_Critique of Functionalist Reason. Cambridge: Polity Press._
Isin E and Ruppert E (2015) Being Digital Citizens. London: Rowman & Littlefield.
Jagers J and Walgrave S (2007) Populism as political communication style: An empirical study
of political parties’ discourse in Belgium. European Journal of Political Research 46: 319–
345.
Karppinen K (2007) Against naïve pluralism in media politics: On the implications of the
radical-pluralist approach to the public sphere. Media, Culture & Society 29(3): 405–508.
Lekakis EJ (2013) Coffee Activism and the Politics of Fair Trade and Ethical Consumption in
_the Global North._ London: Palgrave Macmillan.
Mouffe C (2005) On the Political. London: Sage.
Rovisco M and Ong J (eds) (2016) Taking the Square: Mediated Dissent and Occupations of
_Public Space. London: Rowman & Littlefield._
Rawls J (1993) Political Liberalism. New York: Colombia University Press.
Ruiz P (2014) Articulating Dissent: Protest and the Public Sphere. London: Pluto.
Ruiz P (2016) Identity, place and politics: From picket lines to occupation. In: Rovisco M and
Ong J (eds) Taking the Square: Mediated Dissent and Occupations of Public Space. London:
Rowman & Littlefield, 15 – 36.
Swinford S (2015) Election 2015: How David Cameron’s Conservatives won. The Telegraph,
8 May. Available at: www.telegraph.co.uk/news/general-election-2015/11592230/Election
2015-How-David-Camerons-Conservatives-won.html (accessed on 31 July 2018).
Thompson J (1995) _The Media and Modernity: A Social Theory of the Media. Cambridge:_
Polity Press.
-----
|
{
"disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1177/1367877919849960?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1177/1367877919849960, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "other-oa",
"status": "GREEN",
"url": "https://figshare.le.ac.uk/articles/journal_contribution/The_performance_of_power_and_citizenship_David_Cameron_meets_the_people/10243721/1/files/18491303.pdf"
}
| 2,019
|
[
"JournalArticle"
] | true
| 2019-07-23T00:00:00
|
[] | 9,769
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Medicine",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Environmental Science",
"source": "s2-fos-model"
},
{
"category": "Engineering",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/024c9e9327edd1f6ed69eb9850cfaf5c2b266be9
|
[
"Computer Science",
"Medicine"
] | 0.846175
|
Cloud Servers: Resource Optimization Using Different Energy Saving Techniques
|
024c9e9327edd1f6ed69eb9850cfaf5c2b266be9
|
Italian National Conference on Sensors
|
[
{
"authorId": "5956296",
"name": "Mohammad Hijji"
},
{
"authorId": "107848585",
"name": "B. Ahmad"
},
{
"authorId": "23148907",
"name": "Gulzar Alam"
},
{
"authorId": "2275954260",
"name": "Ahmed M. Alwakeel"
},
{
"authorId": "31240499",
"name": "Mohammed M. Alwakeel"
},
{
"authorId": "2191399741",
"name": "L. A. Alharbi"
},
{
"authorId": "31111836",
"name": "Ahd Aljarf"
},
{
"authorId": "2115774945",
"name": "M. Khan"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"SENSORS",
"IEEE Sens",
"Ital National Conf Sens",
"IEEE Sensors",
"Sensors"
],
"alternate_urls": [
"http://nbn-resolving.de/urn/resolver.pl?urn=urn:nbn:ch:bel-142001",
"http://www.mdpi.com/journal/sensors",
"https://www.mdpi.com/journal/sensors"
],
"id": "3dbf084c-ef47-4b74-9919-047b40704538",
"issn": "1424-8220",
"name": "Italian National Conference on Sensors",
"type": "conference",
"url": "http://www.e-helvetica.nb.admin.ch/directAccess?callnumber=bel-142001"
}
|
Currently, researchers are working to contribute to the emerging fields of cloud computing, edge computing, and distributed systems. The major area of interest is to examine and understand their performance. The major globally leading companies, such as Google, Amazon, ONLIVE, Giaki, and eBay, are truly concerned about the impact of energy consumption. These cloud computing companies use huge data centers, consisting of virtual computers that are positioned worldwide and necessitate exceptionally high-power costs to preserve. The increased requirement for energy consumption in IT firms has posed many challenges for cloud computing companies pertinent to power expenses. Energy utilization is reliant upon numerous aspects, for example, the service level agreement, techniques for choosing the virtual machine, the applied optimization strategies and policies, and kinds of workload. The present paper tries to provide an answer to challenges related to energy-saving through the assistance of both dynamic voltage and frequency scaling techniques for gaming data centers. Also, to evaluate both the dynamic voltage and frequency scaling techniques compared to non-power-aware and static threshold detection techniques. The findings will facilitate service suppliers in how to encounter the quality of service and experience limitations by fulfilling the service level agreements. For this purpose, the CloudSim platform is applied for the application of a situation in which game traces are employed as a workload for analyzing the procedure. The findings evidenced that an assortment of good quality techniques can benefit gaming servers to conserve energy expenditures and sustain the best quality of service for consumers located universally. The originality of this research presents a prospect to examine which procedure performs good (for example, dynamic, static, or non-power aware). The findings validate that less energy is utilized by applying a dynamic voltage and frequency method along with fewer service level agreement violations, and better quality of service and experience, in contrast with static threshold consolidation or non-power aware technique.
|
# sensors
### Article
## Cloud Servers: Resource Optimization Using Different Energy Saving Techniques
### Mohammad Hijji [1,]*, Bilal Ahmad [2], Gulzar Alam [3], Ahmed Alwakeel [1], Mohammed Alwakeel [1], Lubna Abdulaziz Alharbi [1], Ahd Aljarf [4] and Muhammad Umair Khan [5,]*
1 Faculty of Computers & Information Technology, University of Tabuk, Tabuk 71491, Saudi Arabia
2 Warwick Manufacturing Group, University of Warwick, Coventry CV4 7AL, UK
3 School of Computing, Ulster University, Belfast BT15 1AP, UK
4 College of Computers & Information Systems, Umm Al Qura University, Mecca 21955, Saudi Arabia
5 School of Computing, Gachon University, Seongnam-si 13120, Korea
***** Correspondence: m.hijji@ut.edu.sa (M.H.); mumairkhan@gachon.ac.kr (M.U.K.)
**Citation: Hijji, M.; Ahmad, B.; Alam,**
G.; Alwakeel, A.; Alwakeel, M.;
Abdulaziz Alharbi, L.; Aljarf, A.;
Khan, M.U. Cloud Servers: Resource
Optimization Using Different Energy
Saving Techniques. Sensors 2022, 22,
[8384. https://doi.org/10.3390/](https://doi.org/10.3390/s22218384)
[s22218384](https://doi.org/10.3390/s22218384)
Academic Editor: Sung-Bae Cho
Received: 21 September 2022
Accepted: 26 October 2022
Published: 1 November 2022
**Publisher’s Note: MDPI stays neutral**
with regard to jurisdictional claims in
published maps and institutional affil
iations.
**Copyright:** © 2022 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
[Attribution (CC BY) license (https://](https://creativecommons.org/licenses/by/4.0/)
[creativecommons.org/licenses/by/](https://creativecommons.org/licenses/by/4.0/)
4.0/).
**Abstract: Currently, researchers are working to contribute to the emerging fields of cloud computing,**
edge computing, and distributed systems. The major area of interest is to examine and understand
their performance. The major globally leading companies, such as Google, Amazon, ONLIVE, Giaki,
and eBay, are truly concerned about the impact of energy consumption. These cloud computing
companies use huge data centers, consisting of virtual computers that are positioned worldwide
and necessitate exceptionally high-power costs to preserve. The increased requirement for energy
consumption in IT firms has posed many challenges for cloud computing companies pertinent to
power expenses. Energy utilization is reliant upon numerous aspects, for example, the service
level agreement, techniques for choosing the virtual machine, the applied optimization strategies
and policies, and kinds of workload. The present paper tries to provide an answer to challenges
related to energy-saving through the assistance of both dynamic voltage and frequency scaling
techniques for gaming data centers. Also, to evaluate both the dynamic voltage and frequency scaling
techniques compared to non-power-aware and static threshold detection techniques. The findings
will facilitate service suppliers in how to encounter the quality of service and experience limitations
by fulfilling the service level agreements. For this purpose, the CloudSim platform is applied for
the application of a situation in which game traces are employed as a workload for analyzing the
procedure. The findings evidenced that an assortment of good quality techniques can benefit gaming
servers to conserve energy expenditures and sustain the best quality of service for consumers located
universally. The originality of this research presents a prospect to examine which procedure performs
good (for example, dynamic, static, or non-power aware). The findings validate that less energy
is utilized by applying a dynamic voltage and frequency method along with fewer service level
agreement violations, and better quality of service and experience, in contrast with static threshold
consolidation or non-power aware technique.
**Keywords: cloud computing; distributed systems; data centers; virtual machines; energy saving**
### 1. Introduction
This Virtualization techniques distribute the physical server into many remote and single-performance computer system environments by implementing a layer like a hyper- visor or virtual machine manager on hardware or operating systems. In the implemented performance environment, every single-performance computer, such as a virtual machine, runs freely, combined with an operating system and other relevant applications devoid of mutual interference. The virtualization method was not trendy before due to some challenges, such as separate hardware resources, memory, and inadequate network [1–3]. Virtualization has emerged with advancements in technology, such as enhancements in hardware, cloud computing, IT networks, etc. [4,5]. The research community and
-----
_Sensors 2022, 22, 8384_ 2 of 13
### practitioners started to work on the effective operation of virtualization when more users’ demands and use of cloud data centers for performing their tasks with other applications increased [6,7]. Issues were raised, such as overloaded and idle servers; if one server fails to operate, then all virtual machines will be affected, protection of virtual machines and hardware failure, etc. These issues were resolved with the beginning of virtual machine migration initiated from process migration [8]. The greater part of cloud computing operations is encouraged by virtual machine migration, such as server consolidation, hardware maintenance, energy, and flow management [9–11]. Numerous cloud computing models have been developed in which control and man- agement of computing resources are provided. This helps businesses and clients use resources according to their design needs [12–14]. As an alternative to acquiring increased amounts in obtaining information technology infrastructure and dealing with hardware and software maintenance and updates, companies can outsource their computational requirements to the cloud. Large-size data centers have developed that consist of thousands of processing nodes and expend massive volumes of electric power. According to the latest survey, information technology impacts 25% of the total cost of managing and using data centers [15,16]. Energy consumption is overwhelming not only due to idle computing resources but also because of the ineffective management of these computational hardware and software resources. Servers commonly operate up to 50% complete capacity ahead of additional costs on over-provision and total cost of acquisition [17]. Energy management can be used to leverage resources through virtualization techniques and technology [18,19]. It permits cloud providers to generate many virtual machine occurrences on a separate physical server to enhance the efficient management and utilization of computational resources. This will increase the return on investment. Amiri et al. [20] recommended an SDN (Software Defined Network) model for choos- ing DC (Data Centers) for novel gaming sessions. They used a hierarchy-based model for transport/response delay with bandwidth status by using the Lagrange algorithm and logarithmic techniques. Similarly, they used a new approach to reduce end-to-end latency in a cloud-based gaming data center environment [21]. Cai et al. [22] conducted a comprehensive survey on cloud gaming by involving various facets such as the platform used for cloud gaming, various optimization techniques, and commercial services for cloud gaming. Further, they explored the experience factor for gamers and energy utilization with network metrics. Chen et al. [23] proposed an approach for describing energy usage for virtual machines using measurement attributes such as performance, execution time, power (utilization and effectiveness), and energy usage. Therefore, to reduce the cost related to the cloud and to improve energy saving needed appropriate optimization techniques to enhance the user gamer experience. GreenCloud architecture aims to reduce data center power consumption while guaran- teeing performance from the users’ perspective. GreenCloud architecture enables compre- hensive online monitoring, live virtual machine migration, and VM placement optimization. For experimentation, the CloudSim framework is used. CloudSim is a free and open-source framework based on Java language used for cloud computing infrastructure and services simulations. Similarly, this framework is utilized to model and simulate a cloud computing setting to perform tests and produce results. Further, it maintains various functionalities such as the generation of cloud-based entities, relations among entities, processing events, jobs and tasks queue, and implementation of broker policies [24,25]. The major contribution of the proposed research will be as follows:
To investigate how resource optimization can be performed in gaming data centers
•
Utilizing real-time gaming workload
•
• To measure service quality during online gaming data by utilizing its two features, i.e., energy consumption and SLA (Service Level Agreement)
-----
_Sensors 2022, 22, 8384_ 3 of 13
### To test and implement DVFS (Dynamic Voltage and Frequency Scaling), non-power
• aware, and static threshold virtual machine consolidation techniques for improving service quality.
The remainder of the paper is organized as follows: Section 2 explains the literature review, followed by Section 3.1, which presents challenges related to the migration of a single virtual machine; Section 3.2 addresses the challenges related to the migration of the dynamic virtual machine; Section 4 discusses system methodology; Section 5 describes performance analysis and discussion while Section 6 represents conclusions and future work close the article.
2. Literature Review
Nathuji and Schwan [26] did initial work on the application of power management in virtualized data centers by proposing an architecture called a data center resource management system by splitting it into two categories: local policies and global policies. Then, [27] worked on virtualized heterogeneous environment power management and proposed the problem of sequential optimization by addressing it through the concept of limited lookahead control. This research work aims to increase resource providers’ profits by reducing power consumption. Similarly, [28] researched the issue of scheduling for multi-tier web applications related to virtualizing heterogeneous systems to decrease power consumption by maintaining performance. Further, [29] recommended a method on the issue of efficient allocation of power in virtual machines over the complete environment of a virtualized heterogeneous computing system. [30] worked on and used continuous optimization to solve the difficulty of power-aware dynamic placement of applications in interaction with a virtualize heterogeneous environment. [31] have worked on the allocation of available power budgets among servers related to virtualized server farms in heterogeneous environments to decrease the mean response time. Furthermore, they used the proposed model to detect optimal power allocation. Jung et al. [32] analyzed the issue of dynamic consolidation of virtual machines running on multi-tier web applications while using live migration. However, the proposed method is only implemented on individual web application setups and cannot be used as a service system for multitenant infrastructure. Similarly, [33] worked on the same issue of capacity planning and resource allocation by proposing three controllers: the longest, shorter, and shortest time scales. Every controller operates at various time scales. Kumar et al. [34] developed a method for dynamic virtual machine consolidation based on estimation stability. Further, they mentioned that the resource demands of application estimation are performed by utilizing the time-varying probability density function. They stated that the values can be achieved by utilizing offline profiling of applications and calibration; however, offline profiling is impractical for infrastructure as a service system. Likewise, [35] researched a similar issue of dynamic consolidation of virtual machine- running applications using machine learning algorithms to optimize the combined energy consumption. However, this method was applied for high-performance computing and is not appropriate for various workloads. Arshad et al. [36] proposed an algorithm based on energy proficiency heuristics by uti- lizing virtual machine consolidation to reduce greater usage of energy consumption in the cloud data server environment. They build up a model for virtual machines relocation from one physical host to the other with an aim to lower energy consumption. Moura et al. [37] used the internal value fuzzy logic approach to overcome the problems of resources us- ing vagueness and inaccuracies to save energy with the lowest performance deprivation. They increased energy effectiveness by 2.3% in cloud computing simulation environments. Similarly, Shaw et al. look at the application of reinforcement machine learning to address the virtual machine consolidation issue related to the dissemination of virtual machines throughout the cloud data centers to enhance the management of resources. They enhance energy efficiency by 25% and lower service violations by 63%. Liu et al. [38] proposed a method to overcome the problem of virtual machine consolidation to optimize energy
-----
_Sensors 2022, 22, 8384_ 4 of 13
### utilization. They presented a new algorithm to choose the optimal solution for energy usage optimization by accomplishing an average conservation of 42% energy. Further, Gharehpasha et al. [39] presented an approach to combine both Sine and Cosine algorithms with the salp swarm algorithm for the best possible virtual machine placement. Also, their research work aims to decrease energy utilization in cloud data centers environment with SLA reduction. Hussain et al. [40] developed a schedule-based algorithm to decrease energy usage in the heterogenous virtual machine cloud environment. After all, Katal et al. [41] conducted a thorough survey on energy efficiency in a cloud computing data center environment. They discussed various methods to lower the power usage in data centers with hardware component level for decreasing the usage by components. As a variation to the above literature findings, we propose that the central research field consists of single servers and exclusive tasks. Though, presently, huge cloud comput- ing platforms such as Gaikai and Amazon EC2 come up with servers that are spending versatile applications which are further disseminated universally. Conversely, there is an examination disparity in gaming, particularly for multi-player scale games with consumers located remotely. In contrast to this, less evidence has been found regarding energy saving in the context of large data in single-objective applications. The notion of virtualization is employed by researchers using a local regression and robust migration algorithm. The findings propose that latency and service quality can be attained in huge data servers with this virtualization technique. Still, adjustment is a prerequisite between the quality of service and experience [42]. Table 1 shows the comparison among different optimization techniques with an applied method, category, and problem resolution.
**Table 1. Different Optimization Techniques.**
**Method** **Categories** **Technique** **Resolves**
Data Centre Resource
Local and Global Policies Virtualization
Management [26,27]
Scheduling for multi-tier web Virtualizing heterogeneous
Virtualization
applications [28] systems
Power-aware dynamic
Dynamic Virtualization Continuous Optimization
placement of applications [30]
Sequential optimization by
addressing it through the
concept of limited lookahead
control
Decreases power
consumption by maintaining
performance for multi-web
applications
Power-aware dynamic
placement of applications in
interaction with a virtualized
heterogeneous environment
Resolves resource
optimization for small
applications
Saves power and resolves
resource optimization issues
based on workload for servers
placed locally and globally
Dynamic virtual machine Dynamic VM consolidation
consolidation [34] based on estimation stability
Resource demands by
utilizing the time-varying
probability density function
Dynamic Voltage and
Single and Multi-server DVFS, based on workload
Frequency (DVFS)—Proposed
### 3. Challenges
### The main challenges are explored in two domains such as (1) migration to a single virtual machine and (2) migration to a dynamic virtual machine.
3.1. Migration of a Single Virtual Machine
Virtual machines offer benefits to the system consumption, workload, and flexibility of the data center. However, challenges remain, such as waste of resources, network conges- tion, and consolidation, which will cause server hardware failures. Single virtual machine migration is used by researchers to define a data center with particular properties [43,44]. Similarly, [45] worked to increase the server average utilization and experiments on the
-----
Virtual machines offer benefits to the system consumption, workload, and flexibility
_Sensors 2022, 22, 8384_ 5 of 13
of the data center. However, challenges remain, such as waste of resources, network congestion, and consolidation, which will cause server hardware failures. Single virtual machine migration is used by researchers to define a data center with particular properties
[43,44]. Similarly, [45] worked to increase the server average utilization and experiments historical data to predict the future servers’ demands, as well as migrating the virtual
on the historical data to predict the future servers’ demands, as well as migrating the vir-machine in conditions of future needs.
tual machine in conditions of future needs. Unstable length and long latency are the key challenges of migrating virtual machines
### in wide-area networks. Therefore, [Unstable length and long latency are the key challenges of migrating virtual ma-46] get significantly responsive in wide area network mi
chines in wide-area networks. Therefore, [46] get significantly responsive in wide area gration by proposing a three-phase solution. Most importantly, virtual machine migration
network migration by proposing a three-phase solution. Most importantly, virtual ma-is widely utilized to conserve power using the consolidation of idle desktop virtual [47].
chine migration is widely utilized to conserve power using the consolidation of idle desk-Moreover, researchers have developed algorithms with the objective of decreasing power
top virtual [47]. Moreover, researchers have developed algorithms with the objective of mode transition latency [48].
decreasing power mode transition latency [48].
### 3.2. Migration of a Dynamic Virtual Machine
_3.2. Migration of a Dynamic Virtual Machine_
### Virtual machine migration (VMM) is the movement of some or all parts of virtual
Virtual machine migration (VMM) is the movement of some or all parts of virtual
### machine data from one place to a different place, with live migration having no interruption
machine data from one place to a different place, with live migration having no interrup
### of the provided services. VMM is organized in two ways: live migration and non-live
tion of the provided services. VMM is organized in two ways: live migration and non-live
### migration. In non-live migration, the virtual machine is suspended earlier migration and
migration. In non-live migration, the virtual machine is suspended earlier migration and
### conditional on whether the virtual machine needs to remain the running services later
conditional on whether the virtual machine needs to remain the running services later
### migration or not. If it is suspended, then the states will be moved into the target site.
migration or not. If it is suspended, then the states will be moved into the target site.
### In the case of migration, all the connections are restored after virtual machine continu
In the case of migration, all the connections are restored after virtual machine contin
### ation because no open network connection is preserved, as shown in Figure 1.
uation because no open network connection is preserved, as shown in Figure 1.
**Figure 1. Figure 1.Non-Live Migration. Non-Live Migration.**
Live migration is the movement of a virtual machine operating on one physical host Live migration is the movement of a virtual machine operating on one physical host
_Sensors 2022, 22, x FOR PEER REVIEW to a different host devoid of interrupting the usual operations or triggering any stoppage to a different host devoid of interrupting the usual operations or triggering any stoppage6 of 14_
or other undesirable causes for the end user, as shown in Figure 2. or other undesirable causes for the end user, as shown in Figure 2.
In live migration, data migration memory and network connection continuity are two
problems. However, a few challenges are associated with the migration of dynamic virtual
machines, such as the consideration of multiple hosts and multiple virtual machines [49].
Other challenges include memory data migration, storage data migration, and network
connection connectivity [42].
problems. However, a few challenges are associated with the migration of dynamic virtual
machines, such as the consideration of multiple hosts and multiple virtual machines [49].
Other challenges include memory data migration, storage data migration, and network
connection connectivity [42].
**Figure 2. Figure 2.Live VM Migration. Live VM Migration.**
**4. System Methodology In live migration, data migration memory and network connection continuity are two**
### problems. However, a few challenges are associated with the migration of dynamic virtual
The overall system methodology is shown in Figure 3, which consists of the software
layer of the system, which is tied up with local as well as global management modules. machines, such as the consideration of multiple hosts and multiple virtual machines [49].
Local managers represent individual nodes as a component of the VMM. The main pur
-----
**y** **gy**
_Sensors 2022, 22, 8384_ The overall system methodology is shown in Figure 3, which consists of the software 6 of 13
layer of the system, which is tied up with local as well as global management modules.
Local managers represent individual nodes as a component of the VMM. The main purpose of this is to continuously monitor all the nodes contributing to the CPU utilization
### Other challenges include memory data migration, storage data migration, and network
and then adjust all resources that are needed for a virtual machine, and finally to decide
### connection connectivity [42].
about the node’s migration timing and place related to a virtual machine, as shown in
point 4 of Figure 3.
### 4. System Methodology
- The global manager represents a master node to gather information from all local
### The overall system methodology is shown in Figure 3, which consists of the software
managers to preserve the total layout of the consumption of related resources, as
### layer of the system, which is tied up with local as well as global management modules. Local
shown in point 2 of Figure 3.
### managers represent individual nodes as a component of the VMM. The main purpose of this
- The global manager provided instructions for the optimization of virtual machine
### is to continuously monitor all the nodes contributing to the CPU utilization and then adjust
positioning, as shown in point 3 of Figure 3.
### all resources that are needed for a virtual machine, and finally to decide about the node’s
- The main function of VMMs is to resize and migrate the virtual machines and shift
### migration timing and place related to a virtual machine, as shown in point 4 of Figure 3.
the power modes of the nodes, as presented in point 5 of Figure 3.
**Figure 3.Figure 3. Overall System Methodology. 1 defines the user type as a global user and each nodeOverall System Methodology. 1 defines the user type as a global user and each node**
communicates to the global manager through its local manager represented by 2. Each node is
communicates to the global manager through its local manager represented by 2. Each node is
divided into the number of VMs represented as 5 that are managed by their local manager for mi
divided into the number of VMs represented as 5 that are managed by their local manager for
gration presented by 4. The global manager issues commands for the optimisation of the VM as
migration presented by 4. The global manager issues commands for the optimisation of the VMsignments shown in 3.
assignments shown in 3.
### The global manager represents a master node to gather information from all local
• managers to preserve the total layout of the consumption of related resources, as shown in point 2 of Figure 3.
The global manager provided instructions for the optimization of virtual machine
• positioning, as shown in point 3 of Figure 3.
• The main function of VMMs is to resize and migrate the virtual machines and shift the power modes of the nodes, as presented in point 5 of Figure 3.
5. Performance Analysis and Discussion
Some tests have been conducted on CloudSim simulation settings to determine differ- ent characterizations of resource optimization. All these tests were executed on the same datasets by applying “Eclipse Luna and Java IDE Developers. 283 MB; 144,793 DOWN- LOADS”. Different optimization techniques have been used, namely dynamic voltage and frequency techniques, non-power awareness, and static virtualization techniques. These tests have been designed and carried out on a data set from World of WarCraft that is a mas- sively multiplayer online games (MMOs) game that is multi-location multi-environment. Test environments consist of multiple avatars over 3.5 years collected from an online cloud environment. This helps to test the limits of resource optimization for cloud environments for different features, such as energy optimization. service level agreement, service level
-----
_Sensors 2022, 22, 8384_ 7 of 13
### agreement violations, virtualization, host timing, etc. Virtualization techniques will be used for the management of load for virtual machines (VMs) that are over or underloaded in the system, and relocation of these will be performed based on techniques such as minimum migration time (MMT), maximum correlation (MC), and minimum utilization (MU). DVFS, non-power aware (NPA), and static threshold virtualization technique (STVM) techniques will be compared in the same environment. For STVM techniques, defined resources are used in terms of random-access memory (RAM), bandwidth, storage, and input-output file size, whereas in dynamic technique, resources are allocated based on central processing unit voltage and frequency fluctuations. Different evaluation metrics will be used to gauge the performance of the proposed system. Initially, the tests are divided into different techniques for example DVFS, NPA, and STVM. The reason for dividing them into sub-techniques is to see how the proposed system will behave under different conditions. Test environment and workload are standard for all methods. All these proposed methods will be measured against certain defined parameters such as energy consumption, VM selection time, VM relocation time, host selection meantime, and service level agreement violations. These matrices will help to determine which technique will perform better under static and dynamic workloads in the proposed test environment. The comparison method will also help to determine which technique performs better for energy saving and resource optimization for small and large servers placed globally. A test has been carried out to distinguish how dynamic frequency scaling will behave with non-power-aware techniques for the same workload. The results in Figure 4 are plotted using the reality check method. The results show that the non-power-aware method consumes more power compared to the dynamic voltage and frequency methods. DVFS shows a linear trend for energy consumption and less consumption of power. The DVFS method results in increased profits and minimum SLAs per host compared to the NPA technique. However, using NPA with the same host numbers and fixed millions
_Sensors 2022, 22, x FOR PEER REVIEW of instructions per second (MIPS) consumes more energy in the setup, emitting higher8 of 14_
### CO2 emissions.
**_Energy Comparion_**
**100%**
**80%**
**60%**
**40%**
**20%**
**0%**
DVFS NPA
**Hosts**
**-20%**
**Figure 4.Figure 4. Illustrations of Energy Utilization in a Data Center.Illustrations of Energy Utilization in a Data Center.**
### A similar test is further extended, and the static threshold virtualization technique (STVM) has been added to determine the energy consumption. In these experimental results, as shown in Figure 5, three virtualization techniques were used to relocate the
**_Energy Comparion_**
**100%**
**80%**
**60%**
**40%**
**20%**
**0%**
DVFS NPA
**Hosts**
**-20%**
-----
_Sensors 2022, 22, 8384_ 8 of 13
**0%**
DVFS NPA
### virtual machines for overloaded and underloaded hosts. This relocation of virtual machinesHosts is done using minimum migration time (MMT), minimum correlation (MC), and maximum-20% utilization (MU) in a static threshold environment.
**Figure 4. Illustrations of Energy Utilization in a Data Center.**
**Figure 5.Figure 5. Evaluation of Energy Utilization in the Recommended System.Evaluation of Energy Utilization in the Recommended System.**
### In STVM, higher, and lesser threshold boundaries are specified for any test envi-In STVM, higher, and lesser threshold boundaries are specified for any test environ- ronment. In the static threshold technique, MC has less energy consumption comparedment. In the static threshold technique, MC has less energy consumption compared to the to the MU or MMT method. When compared with the dynamic voltage and frequencyMU or MMT method. When compared with the dynamic voltage and frequency tech
niques, the results are different. Static threshold behaves better for small workloads as
### techniques, the results are different. Static threshold behaves better for small workloads
upper and lower limits are definable for required parameters. In comparison to the dy
### as upper and lower limits are definable for required parameters. In comparison to the
namic workload environment, DVFS again proves to have less service level agreement
### dynamic workload environment, DVFS again proves to have less service level agreement
violation (SLAV) and maintains higher SLAs, resulting in a better quality of service and
### violation (SLAV) and maintains higher SLAs, resulting in a better quality of service and better user experience compared to the NPA method. It can also be concluded that STVM virtual machine relocation methods are supported with smaller workloads, which verifies the theoretical concept. All three techniques are used to compare the execution times for three techniques for different levels of hosts with the same configuration setup in Figure 6. Virtual machine selection, relocation, and host selection time remained similar for DVFS and NPA.
**0%**
DVFS NPA
### virtual machines for overloaded and underloaded hosts. This relocation of virtual machinesHosts is done using minimum migration time (MMT), minimum correlation (MC), and maximum-20%
-----
p
_Sensors 2022, 22, 8384_ All three techniques are used to compare the execution times for three techniques for 9 of 13
different levels of hosts with the same configuration setup in Figure 6. Virtual machine
selection, relocation, and host selection time remained similar for DVFS and NPA.
**_DVFS_**
**_0.03_**
**_NPA_**
**_0.025_**
**_MMT_**
**_0.02_** **_MC_**
**_MU_**
**_0.015_**
**_0.01_**
**_0.005_**
**_0_**
**VM Selection** **VM Relocation** **Host Selection**
**Mean Time** **Mean Time** **Mean Time**
**Figure 6. Virtual Machine Performance Time for Every Host.**
**Figure 6. Virtual Machine Performance Time for Every Host.**
### MC has the highest VM selection time in a static environment, and MC takes more
MC has the highest VM selection time in a static environment, and MC takes more
### time for VM relocation when compared with other techniques. In a static environment, all
time for VM relocation when compared with other techniques. In a static environment, all
### three techniques have similar host selection meantime because of defined threshold limits
three techniques have similar host selection meantime because of defined threshold limits
### as compared to a dynamic environment. The results also support the theoretical concept
as compared to a dynamic environment. The results also support the theoretical concept
### that no relocation of VMs is done for DVFS, and resource optimization is done using centralthat no relocation of VMs is done for DVFS, and resource optimization is done using cen- processing unit (CPU) voltage and frequency methods.tral processing unit (CPU) voltage and frequency methods.
If a proper virtualization technique is selected, downtime in the network can be re-If a proper virtualization technique is selected, downtime in the network can be
### reduced for overloaded and underloaded environments. The results in Figureduced for overloaded and underloaded environments. The results in Figure 7 show that 7 show that in the STVM method, MC has the lowest number of virtual machines that are migrated,in the STVM method, MC has the lowest number of virtual machines that are migrated, whereas maximum utilization has the highest number of migrations. NPA and DVFS dowhereas maximum utilization has the highest number of migrations. NPA and DVFS do
_Sensors 2022, 22, x FOR PEER REVIEW not carry any VM migrations, which second the theoretical concept of dynamic voltage andnot carry any VM migrations, which second the theoretical concept of dynamic voltage 10 of 14_
### frequency scaling and non-power aware techniques.and frequency scaling and non-power aware techniques
Service level agreement and service level agreement degradation were administered
for all three techniques. DVFS has a minimum service level degradation when compared
to the rest of the techniques. NPA has the highest number of SLAV. If better service quality
is required, fewer SLAV methods need to be selected. The MMT technique needs to be
selected for a better user experience, as this has a minimum number of SLAVs and SLAs
for the static threshold environment, as shown in Figure 8.
**Figure 7.Figure 7. Sum of Virtual Machine Migrations.Sum of Virtual Machine Migrations.**
### Service level agreement and service level agreement degradation were administered for all three techniques. DVFS has a minimum service level degradation when compared to the rest of the techniques. NPA has the highest number of SLAV. If better service quality is required, fewer SLAV methods need to be selected. The MMT technique needs to be
for all three techniques. DVFS has a minimum service level degradation when compared
to the rest of the techniques. NPA has the highest number of SLAV. If better service quality
is required, fewer SLAV methods need to be selected. The MMT technique needs to be
selected for a better user experience, as this has a minimum number of SLAVs and SLAs
for the static threshold environment, as shown in Figure 8.
-----
_Sensors 2022, 22, 8384_ 10 of 13
### selected for a better user experience, as this has a minimum number of SLAVs and SLAs for the static threshold environment, as shown in Figure 8.
**Figure 7. Sum of Virtual Machine Migrations.**
### selected for a better user experience, as this has a minimum number of SLAVs and SLAs
**Figure 8.Figure 8. Analysis of the Service Level Agreement Violation.Analysis of the Service Level Agreement Violation.**
### In dynamic environments, DVFS has less energy consumption associated with NPAIn dynamic environments, DVFS has less energy consumption associated with NPA methods. In a static environment, MMT has the highest number of host shutdowns, as VMsmethods. In a static environment, MMT has the highest number of host shutdowns, as are selected and relocated for loaded hosts to save resources and energy. MMT, therefore,VMs are selected and relocated for loaded hosts to save resources and energy. MMT, also has less mean and standard deviation time in a static environment compared to othertherefore, also has less mean and standard deviation time in a static environment com
pared to other virtual machine relocation techniques.
### virtual machine relocation techniques.
Therefore, the overall detailed analysis of the proposed system is shown in Figure 9.
### Therefore, the overall detailed analysis of the proposed system is shown in Figure 9.
So, depending on whether the test environment is dynamic or static, resource optimiza
### So, depending on whether the test environment is dynamic or static, resource optimization,
tion, service quality, and better user experience can be achieved if proper methods are
### service quality, and better user experience can be achieved if proper methods are selected
selected for loaded hosts in a cloud environment. Proper selection of optimization tech
### for loaded hosts in a cloud environment. Proper selection of optimization techniques
niques will help in energy and resource optimization for large-scale servers that are placed
_Sensors 2022, 22, x FOR PEER REVIEW will help in energy and resource optimization for large-scale servers that are placed and11 of 14_
and operating globally.
### operating globally.
**_Detailed Analysis of the Proposed System_**
**DVFS**
**NPA**
**900**
**800** **MMT**
**700**
**MC**
**600**
**MU**
**500**
**400**
**300**
**200**
**100**
**0**
**Energy**
**Concumption** **Host Shutdown**
**Mean Time**
**Before Host** **St Dev Time**
**Shutdown** **Before Host**
**Shutdown**
**Figure 9.Figure 9. Overall detailed evaluation of the developed system.Overall detailed evaluation of the developed system.**
**DVFS**
**NPA**
**MMT**
**MC**
**MU**
**100**
**0**
**Energy**
**Concumption** **Host Shutdown**
**Mean Time**
**6 Conclusions**
-----
_Sensors 2022, 22, 8384_ 11 of 13
### 6. Conclusions
Different simulation experiments are designed using the CloudSim simulation envi- ronment to test resource optimization in cloud gaming servers. These experiments suggest different resource optimization techniques for large and small servers. Gaming datasets are versatile in nature and consist of different audio, video, avatars, locations, etc. The data versatility helps to challenge resource optimization in terms of energy consumption, execu- tion time, virtual machine relocation, and service level agreement violations for different user levels. From the results, it is evident that different resource optimization techniques are required to be selected for under-and overloaded hosts depending on servers and user data type. If the data that is being processed has defined limits, then the static threshold technique will be used with another virtualization discussed above. In terms of a dynamic environment with multiple users and a large pool of resources, dynamic resource optimiza- tion behaves better. Therefore, for large servers, DVFS saves more energy, has fewer service level agreement violations, and results in a better quality of service and experience. In the future, this work will be enhanced to explore new energy-saving techniques and compared them with the current methods. This work will also be extended to other domains of computing for example Internet of Things (IoT), Big Data, and Artificial Intelligence (AI).
**Author Contributions: Conceptualization, M.H., B.A. and M.U.K.; methodology M.H., B.A., G.A.**
and M.U.K.; software, B.A.; validation, M.H., B.A. and G.A.; formal analysis, M.H., B.A., A.A. (Ahmed
Alwakeel), A.A. (Ahd Aljarf) and M.A.; investigation, M.A., L.A.A., A.A. (Ahmed Alwakeel) and A.A.
(Ahd Aljarf); data curation, M.H., B.A. and A.A. (Ahd Aljarf); writing—original draft preparation,
M.H., B.A. and G.A.; writing—review and editing, M.H., B.A., G.A., M.A., L.A.A., A.A. (Ahmed
Alwakeel), A.A. (Ahd Aljarf) and M.U.K.; visualization, B.A. and G.A.; supervision, M.H.; project
administration, M.H. and B.A.; funding acquisition, M.H. All authors have read and agreed to the
published version of the manuscript.
**Funding: This research received no external funding.**
**Institutional Review Board Statement: Not Applicable.**
**Informed Consent Statement: Not Applicable.**
**Data Availability Statement: Data will be available upon request through correspondence email.**
**Acknowledgments: We acknowledge the support of the University of Tabuk, Saudi Arabia.**
**Conflicts of Interest: The authors declare no conflict of interest.**
### References
1. Huber, N.; von Quast, M.; Hauck, M.; Kounev, S. Evaluating and Modeling Virtualization Performance Overhead for Cloud
Environments. CLOSER 2011, 11, 563–573.
2. Khan, F.; Ahmad, S.; Gürüler, H.; Cetin, G.; Whangbo, T.; Kim, C.-G. An Efficient and Reliable Algorithm for Wireless Sensor
[Network. Sensors 2021, 21, 24. [CrossRef] [PubMed]](http://doi.org/10.3390/s21248355)
3. Khan, F.; Khan, A.W.; Shah, K.; Qasim, I.; Habib, A. An Algorithmic Approach for Core Election in Mobile Ad-hoc Network.
_J. Internet Technol. 2019, 20, 4._
4. Wang, L.; Tao, J.; Kunze, M.; Castellanos, A.C.; Kramer, D.; Karl, W. Scientific Cloud Computing: Early Definition and Experience.
In Proceedings of the 2008 10th IEEE International Conference on High Performance Computing and Communications, Dalian,
[China, 25–27 September 2008; pp. 825–830. [CrossRef]](http://doi.org/10.1109/HPCC.2008.38)
5. Cheng, L.; Tachmazidis, I.; Kotoulas, S.; Antoniou, G. Design and evaluation of small–large outer joins in cloud computing
[environments. J. Parallel Distrib. Comput. 2017, 110, 2–15. [CrossRef]](http://doi.org/10.1016/j.jpdc.2017.02.007)
6. [Wu, J.; Guo, S.; Li, J.; Zeng, D. Big Data Meet Green Challenges: Greening Big Data. IEEE Syst. J. 2016, 10, 873–887. [CrossRef]](http://doi.org/10.1109/JSYST.2016.2550538)
7. Khan, F.; Gul, T.; Ali, S.; Rashid, A.; Shah, D.; Khan, S. Energy aware cluster-head selection for improving network life time in
wireless sensor network. Sci. Inf. Conf. 2018, 857, 581–593.
8. Osman, S.; Subhraveti, D.; Su, G.; Nieh, J. The design and implementation of Zap: A system for migrating computing environ[ments. ACM SIGOPS Oper. Syst. Rev. 2002, 36, 361–376. [CrossRef]](http://doi.org/10.1145/844128.844162)
9. Dillon, T.; Wu, C.; Chang, E. Cloud Computing: Issues and Challenges. In Proceedings of the 2010 24th IEEE International
[Conference on Advanced Information Networking and Applications, Perth, Australia, 20–23 April 2010; pp. 27–33. [CrossRef]](http://doi.org/10.1109/AINA.2010.187)
10. [Bianchini, R.; Rajamony, R. Power and energy management for server systems. Computer 2004, 37, 68–76. [CrossRef]](http://doi.org/10.1109/MC.2004.217)
-----
_Sensors 2022, 22, 8384_ 12 of 13
11. Jiang, J.W.; Lan, T.; Ha, S.; Chen, M.; Chiang, M. Joint VM placement and routing for data center traffic engineering. In Proceedings
[of the 2012 Proceedings IEEE INFOCOM, Orlando, FL, USA, 25–30 March 2012; pp. 2876–2880. [CrossRef]](http://doi.org/10.1109/INFCOM.2012.6195719)
12. Buyya, R.; Yeo, C.S.; Venugopal, S.; Broberg, J.; Brandic, I. Cloud computing and emerging IT platforms: Vision, hype, and reality
[for delivering computing as the 5th utility. Future Gener. Comput. Syst. 2009, 25, 599–616. [CrossRef]](http://doi.org/10.1016/j.future.2008.12.001)
13. Khan, F.; Tarimer, I.; Taekeun, W. Factor Model for Online Education during the COVID-19 Pandemic Using the IoT. Processes
**[2022, 10, 7. [CrossRef]](http://doi.org/10.3390/pr10071419)**
14. Khan, F.; Zahid, M.; Gürüler, H.; Tarımer, I.; Whangbo, T. An Efficient and Reliable Multicasting for Smart Cities.[˙] _Mathematics_
**[2022, 10, 3686. [CrossRef]](http://doi.org/10.3390/math10193686)**
15. In the data center, power and cooling costs more than the it equipment it supports. Electron. Cool. 2007, 13, 24. Avail[able online: https://www.electronics-cooling.com/2007/02/in-the-data-center-power-and-cooling-costs-more-than-the-it-](https://www.electronics-cooling.com/2007/02/in-the-data-center-power-and-cooling-costs-more-than-the-it-equipment-it-supports/)
[equipment-it-supports/ (accessed on 31 May 2022).](https://www.electronics-cooling.com/2007/02/in-the-data-center-power-and-cooling-costs-more-than-the-it-equipment-it-supports/)
16. Khan, F.; Khan, A.W.; Khan, S.; Qasim, I.; Habib, A. A secure core-assisted multicast routing protocol in mobile ad-hoc network.
_J. Internet Technol. 2020, 21, 375–383._
17. Fan, X.; Weber, W.-D.; Barroso, L.A. Power provisioning for a warehouse-sized computer. ACM SIGARCH Comput. Archit. News
**[2007, 35, 13–23. [CrossRef]](http://doi.org/10.1145/1273440.1250665)**
18. Barham, P. Xen and the Art of Virtualization. In Proceedings of the 19th ACM Symposi-um on Operating Systems Principles,
Bolton Landing, NY, USA, 19–22 October 2003; ACM Press: New York, NY, USA, 2003.
19. Ahmad, S.; Mehmood, F.; Khan, F.; Whangbo, T.K. Architecting Intelligent Smart Serious Games for Healthcare Applications: A
[Technical Perspective. Sensors 2022, 22, 810. [CrossRef]](http://doi.org/10.3390/s22030810)
20. Amiri, M.; Osman, H.A.; Shirmohammadi, S. Resource optimization through hierarchical SDN-enabled inter data center network
for cloud gaming. In Proceedings of the 11th ACM Multimedia Systems Conference, New York, NY, USA, 27 May 2020;
[pp. 166–177. [CrossRef]](http://doi.org/10.1145/3339825.3391868)
21. Amiri, M.; Osman, H.A.; Shirmohammadi, S.; Abdallah, M. Toward Delay-Efficient Game-Aware Data Centers for Cloud Gaming.
_[ACM Trans. Multimed. Comput. Commun. Appl. 2016, 12, 71. [CrossRef]](http://doi.org/10.1145/2983639)_
22. Cai, W.; Shea, R.; Huang, C.Y.; Chen, K.T.; Liu, J.; Leung, V.C.M.; Hsu, C.H. A Survey on Cloud Gaming: Future of Computer
[Games. IEEE Access 2016, 4, 7605–7620. [CrossRef]](http://doi.org/10.1109/ACCESS.2016.2590500)
23. Chen, Q.; Grosso, P.; van der Veldt, K.; de Laat, C.; Hofman, R.; Bal, H. Profiling Energy Consumption of VMs for Green Cloud
Computing. In Proceedings of the 2011 IEEE Ninth International Conference on Dependable, Autonomic and Secure Computing,
[Sydney, Australia, 12–14 December 2011; pp. 768–775. [CrossRef]](http://doi.org/10.1109/DASC.2011.131)
24. Calheiros, R.N.; Ranjan, R.; Beloglazov, A.; de Rose, C.A.; Buyya, R. CloudSim: A toolkit for modeling and simulation of cloud
[computing environments and evaluation of resource provisioning algorithms. Softw. Pract. Exp. 2011, 41, 23–50. [CrossRef]](http://doi.org/10.1002/spe.995)
25. Khan, F.; Abbas, S.; Khan, S. An efficient and reliable core-assisted multicast routing protocol in mobile Ad-Hoc network. Int. J.
_[Adv. Comput. Sci. Appl. 2016, 7, 231–242. [CrossRef]](http://doi.org/10.14569/IJACSA.2016.070533)_
26. Nathuji, R.; Schwan, K. Virtualpower: Coordinated power management in virtualized enterprise systems. ACM SIGOPS Oper.
_[Syst. Rev. 2007, 41, 265–278. [CrossRef]](http://doi.org/10.1145/1323293.1294287)_
27. Kusic, D.; Kephart, J.O.; Hanson, J.E.; Kandasamy, N.; Jiang, G. Power and performance management of virtualized computing
[environments via lookahead control. Clust. Comput. 2009, 12, 1–15. [CrossRef]](http://doi.org/10.1007/s10586-008-0070-y)
28. Srikantaiah, S.; Kansal, A.; Zhao, F. Energy aware consolidation for cloud computing. In Proceedings of the 2008 conference on
[Power Aware Computing and Systems, San Diego, CA, USA, 7 December 2008; p. 10. Available online: https://www.usenix.org/](https://www.usenix.org/legacy/event/hotpower08/tech/full_papers/srikantaiah/srikantaiah_html/)
[legacy/event/hotpower08/tech/full_papers/srikantaiah/srikantaiah_html/ (accessed on 20 October 2022).](https://www.usenix.org/legacy/event/hotpower08/tech/full_papers/srikantaiah/srikantaiah_html/)
29. Cardosa, M.; Korupolu, M.R.; Singh, A. Shares and utilities based power consolidation in virtualized server environments. In
Proceedings of the 2009 IFIP/IEEE International Symposium on Integrated Network Management, New York, NY, USA, 1–5 June
[2009; pp. 327–334. [CrossRef]](http://doi.org/10.1109/INM.2009.5188832)
30. Verma, A.; Ahuja, P.; Neogi, A. pMapper: Power and Migration Cost Aware Application Placement in Virtualized Systems. In
_[Middleware 2008; Springer: Berlin/Heidelberg, Germany, 2008; pp. 243–264. [CrossRef]](http://doi.org/10.1007/978-3-540-89856-6_13)_
31. Gandhi, A.; Harchol-Balter, M.; Das, R.; Lefurgy, C. Optimal power allocation in server farms. ACM Sigmetrics Perform. Eval. Rev.
**[2009, 37, 157–168. [CrossRef]](http://doi.org/10.1145/2492101.1555368)**
32. Jung, G.; Joshi, K.R.; Hiltunen, M.A.; Schlichting, R.D.; Pu, C. A Cost-Sensitive Adaptation Engine for Server Consolidation of
[Multitier Applications. In Middleware 2009; Springer: Berlin/Heidelberg, Germany, 2009; pp. 163–183. [CrossRef]](http://doi.org/10.1007/978-3-642-10445-9_9)
33. Zhu, X.; Young, D.; Watson, B.J.; Wangm, Z.; Rolia, J.; Singhal, S.; McKee, B.; Hyser, C.; Gmach, D.; Gardner, R.; et al. 1000 Islands:
Integrated Capacity and Workload Management for the Next Generation Data Center. In Proceedings of the 2008 International
[Conference on Autonomic Computing, Chicago, IL, USA, 2–6 June 2008; pp. 172–181. [CrossRef]](http://doi.org/10.1109/ICAC.2008.32)
34. Kumar, S.; Talwar, V.; Kumar, V.; Ranganathan, P.; Schwan, K. vManage: Loosely coupled platform and virtualization management
in data centers. In Proceedings of the 6th International Conference on Autonomic Computing, New York, NY, USA, 15–19 June
[2009; pp. 127–136. [CrossRef]](http://doi.org/10.1145/1555228.1555262)
35. Berral, J.L.; Goiri, I.; Nou, R.; Ferran, J.; Guitart, J.; Gavalda, R.; Torres, J. Towards energy-aware scheduling in data centers using
machine learning. In Proceedings of the 1st International Conference on Energy-Efficient Computing and Networking, New York,
[NY, USA, 14–15 October 2010; pp. 215–224. [CrossRef]](http://doi.org/10.1145/1791314.1791349)
-----
_Sensors 2022, 22, 8384_ 13 of 13
36. Arshad, U.; Aleem, M.; Srivastava, G.; Lin, J.C.-W. Utilizing power consumption and SLA violations using dynamic VM
[consolidation in cloud data centers. Renew. Sustain. Energy Rev. 2022, 167, 112782. [CrossRef]](http://doi.org/10.1016/j.rser.2022.112782)
37. Moura, B.M.P.; Schneider, G.B.; Yamin, A.C.; Santos, H.; Reiser, R.H.S.; Bedregal, B. Interval-valued Fuzzy Logic approach for
[overloaded hosts in consolidation of virtual machines in cloud computing. Fuzzy Sets Syst. 2022, 446, 144–166. [CrossRef]](http://doi.org/10.1016/j.fss.2021.03.001)
38. Liu, X.; Wu, J.; Chen, L.; Zhang, L. Energy-aware virtual machine consolidation based on evolutionary game theory. Concurr.
_[Comput. Pract. Exp. 2022, 34, e6830. [CrossRef]](http://doi.org/10.1002/cpe.6830)_
39. Gharehpasha, S.; Masdari, M.; Jafarian, A. Power efficient virtual machine placement in cloud data centers with a discrete and
[chaotic hybrid optimization algorithm. Clust. Comput. 2021, 24, 1293–1315. [CrossRef]](http://doi.org/10.1007/s10586-020-03187-y)
40. Hussain, M.; Wei, L.-F.; Lakhan, A.; Wali, S.; Ali, S.; Hussain, A. Energy and performance-efficient task scheduling in heteroge[neous virtualized cloud computing. Sustain. Comput. Inform. Syst. 2021, 30, 100517. [CrossRef]](http://doi.org/10.1016/j.suscom.2021.100517)
41. Katal, A.; Dahiya, S.; Choudhury, T. Energy efficiency in cloud computing data center: A survey on hardware technologies. Clust.
_[Comput. 2022, 25, 675–705. [CrossRef]](http://doi.org/10.1007/s10586-021-03431-z)_
42. Zhang, F.; Liu, G.; Fu, X.; Yahyapour, R. A Survey on Virtual Machine Migration: Challenges, Techniques, and Open Issues. IEEE
_[Commun. Surv. Tutor. 2018, 20, 1206–1243. [CrossRef]](http://doi.org/10.1109/COMST.2018.2794881)_
43. Duggan, M.; Duggan, J.; Howley, E.; Barrett, E. A network aware approach for the scheduling of virtual machine migration
[during peak loads. Clust. Comput. 2017, 20, 2083–2094. [CrossRef]](http://doi.org/10.1007/s10586-017-0948-7)
44. Nathan, S.; Bellur, U.; Kulkarni, P. Towards a comprehensive performance model of virtual machine live migration. In Proceedings
of the Sixth ACM Symposium on Cloud Computing, Kohala Coast, HI, USA, 27–29 August 2015; pp. 288–301.
45. Bobroff, N.; Kochut, A.; Beaty, K. Dynamic placement of virtual machines for managing sla violations. In Proceedings of the 2007
10th IFIP/IEEE International Symposium on Integrated Network Management, Munich, Germany, 21–25 May 2007; pp. 119–128.
46. Zhang, W.; Lam, K.T.; Wang, C.L. Adaptive live vm migration over a wan: Modeling and implementation. In Proceedings of the
2014 IEEE 7th International Conference on Cloud Computing, Anchorage, AK, USA, 27 June–2 July 2014; pp. 368–375.
47. Bila, N.; de Lara, E.; Joshi, K.; Lagar-Cavilla, H.A.; Hiltunen, M.; Satyanarayanan, M. Jettison: Efficient idle desktop consolidation
with partial VM migration. In Proceedings of the 7th ACM European Conference on Computer Systems, Bern, Switzerland, 10–13
April 2012; pp. 211–224.
48. Liu, H.; He, B. Vmbuddies: Coordinating live migration of multi-tier applications in cloud environments. IEEE Trans. Parallel
_[Distrib. Syst. 2014, 26, 1192–1205. [CrossRef]](http://doi.org/10.1109/TPDS.2014.2316152)_
49. Beloglazov, A.; Buyya, R. Optimal online deterministic algorithms and adaptive heuristics for energy and performance efficient
[dynamic consolidation of virtual machines in cloud data centers. Concurr. Comput. Pract. Exp. 2012, 24, 1397–1420. [CrossRef]](http://doi.org/10.1002/cpe.1867)
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://pmc.ncbi.nlm.nih.gov/articles/PMC9659174, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "GOLD",
"url": "https://www.mdpi.com/1424-8220/22/21/8384/pdf?version=1667292344"
}
| 2,022
|
[
"JournalArticle"
] | true
| 2022-11-01T00:00:00
|
[
{
"paperId": "736f9cc4bd0adad85682481ab74c411518e72610",
"title": "Development of a Model for Spoofing Attacks in Internet of Things"
},
{
"paperId": "edc2118d69cd42e51751a38337b2650f9ae0a0d8",
"title": "Utilizing power consumption and SLA violations using dynamic VM consolidation in cloud data centers"
},
{
"paperId": "b5b9f3424d73a80ef010c3e5928c5c355cc14a8e",
"title": "Factor Model for Online Education during the COVID-19 Pandemic Using the IoT"
},
{
"paperId": "6dc63f26c474942c8a504aa033a5f8aba0586812",
"title": "Architecting Intelligent Smart Serious Games for Healthcare Applications: A Technical Perspective"
},
{
"paperId": "c2f87230368f3f66274e67d80a1afebe51aa19d6",
"title": "Energy‐aware virtual machine consolidation based on evolutionary game theory"
},
{
"paperId": "77a81471325bbd4c3e0528a53585eef455c3f0d5",
"title": "An Efficient and Reliable Algorithm for Wireless Sensor Network"
},
{
"paperId": "95f36a84877001dabe83e3cf701e0db838b527c0",
"title": "Energy efficiency in cloud computing data center: a survey on hardware technologies"
},
{
"paperId": "2b65c5dff56f9e0fc45f9cbf15a2b2ebcb8698a9",
"title": "Energy and performance-efficient task scheduling in heterogeneous virtualized cloud computing"
},
{
"paperId": "2efb9f335df498215368a22e32d931a2f0344a04",
"title": "Interval-valued Fuzzy Logic approach for overloaded hosts in consolidation of virtual machines in cloud computing"
},
{
"paperId": "5df07bb2b87a8e5b568c2c0973c21b6903122c97",
"title": "Power efficient virtual machine placement in cloud data centers with a discrete and chaotic hybrid optimization algorithm"
},
{
"paperId": "b8db853865fafa782900fc6896b97c5ee7d3217d",
"title": "Resource optimization through hierarchical SDN-enabled inter data center network for cloud gaming"
},
{
"paperId": "c4a42142419e5d841e354162856d136d0114bfa9",
"title": "A Secure Core-Assisted Multicast Routing Protocol in Mobile Ad-Hoc Network"
},
{
"paperId": "b6e4b4fbec34c832dc1941e59ac4d39611e86e0c",
"title": "An Algorithmic Approach for Core Election in Mobile Ad-hoc Network"
},
{
"paperId": "2c1247d7dd6c319c134708c05374788316041dcd",
"title": "Energy Aware Cluster-Head Selection for Improving Network Life Time in Wireless Sensor Network"
},
{
"paperId": "0036975a5e51251e3ccde743ecc3aa18a9ae7623",
"title": "A Survey on Virtual Machine Migration: Challenges, Techniques, and Open Issues"
},
{
"paperId": "ebccfd28ff291a790ef4f313a2640435529ae0da",
"title": "Design and evaluation of small-large outer joins in cloud computing environments"
},
{
"paperId": "54937a64210e0946b678618e160a6646fbeac59f",
"title": "A network aware approach for the scheduling of virtual machine migration during peak loads"
},
{
"paperId": "9e0eb64b8085d9b73166775317c7d0e02d80b9b0",
"title": "A Survey on Cloud Gaming: Future of Computer Games"
},
{
"paperId": "f4900599936427fd1c7714cae15ff00590396a7c",
"title": "Big Data Meet Green Challenges: Greening Big Data"
},
{
"paperId": "5e4fedddbfb7fc57ece9fbd77cf3061285291ebb",
"title": "Towards a comprehensive performance model of virtual machine live migration"
},
{
"paperId": "acec0a12a1946279a35b79828cab4f4cb13761fd",
"title": "VMbuddies: Coordinating Live Migration of Multi-Tier Applications in Cloud Environments"
},
{
"paperId": "5723fc37b2a41250faaefdda5ec5b46d314de666",
"title": "Adaptive Live VM Migration over a WAN: Modeling and Implementation"
},
{
"paperId": "85011e8d503e28c098249e935a1ad52a673dc9f0",
"title": "Optimal online deterministic algorithms and adaptive heuristics for energy and performance efficient dynamic consolidation of virtual machines in Cloud data centers"
},
{
"paperId": "5ba59323988d329f824175f5a99e587979e08422",
"title": "Jettison: efficient idle desktop consolidation with partial VM migration"
},
{
"paperId": "532b2033ed22b81333cdc5a60c4a545849388ca9",
"title": "Joint VM placement and routing for data center traffic engineering"
},
{
"paperId": "4a82222f4234f980f1ce3a476824ed379cd9c530",
"title": "Profiling Energy Consumption of VMs for Green Cloud Computing"
},
{
"paperId": "30a82a63a339c1e69aac36b23900544fe9ec97bb",
"title": "CloudSim: a toolkit for modeling and simulation of cloud computing environments and evaluation of resource provisioning algorithms"
},
{
"paperId": "04bc9c61c81c9125e39a6c52bee56f794e463a6f",
"title": "Cloud Computing: Issues and Challenges"
},
{
"paperId": "8f54b85fcd5074482a74f32a96e808b5be3a9da7",
"title": "Towards energy-aware scheduling in data centers using machine learning"
},
{
"paperId": "022bc8718b12a4667efbf24babcfaa7032e74397",
"title": "A Cost-Sensitive Adaptation Engine for Server Consolidation of Multitier Applications"
},
{
"paperId": "fe6901f657ea8457cf30a5360916f2f14de58870",
"title": "vManage: loosely coupled platform and virtualization management in data centers"
},
{
"paperId": "ad3705596de246d50745db402cbd046cc482b8d7",
"title": "Optimal power allocation in server farms"
},
{
"paperId": "2a4a6255e8218e29bad093cc58a39fec19d92dd0",
"title": "Shares and utilities based power consolidation in virtualized server environments"
},
{
"paperId": "3e19046c665867bbe557685da60738a40738010a",
"title": "Energy aware consolidation for cloud computing"
},
{
"paperId": "2e7647a07fe21c18ab5b7037de3038157338f1db",
"title": "pMapper: Power and Migration Cost Aware Application Placement in Virtualized Systems"
},
{
"paperId": "b254ad4b9a2735a72e13761ed9076b15147ccbde",
"title": "Scientific Cloud Computing: Early Definition and Experience"
},
{
"paperId": "563f7d19eb0e7e1a7197fb498bf47263b91f8f9c",
"title": "Power and performance management of virtualized computing environments via lookahead control"
},
{
"paperId": "f707ae4755139d83da0ca0482defa0413ce897b5",
"title": "1000 Islands: Integrated Capacity and Workload Management for the Next Generation Data Center"
},
{
"paperId": "69e7ae7d28285e1e85056970b0aeded89d366480",
"title": "VirtualPower: coordinated power management in virtualized enterprise systems"
},
{
"paperId": "b70283748f5edb07efed762f1645481e7b0f88ca",
"title": "Dynamic Placement of Virtual Machines for Managing SLA Violations"
},
{
"paperId": "b44e3a9b0afb925b8cd6477495981d7b73d1d0af",
"title": "Power provisioning for a warehouse-sized computer"
},
{
"paperId": "642029f803ee3bd68303df50d663cea3f4473122",
"title": "Power and energy management for server systems"
},
{
"paperId": "3af27466b0b648aaedd25ea7d087e3954329e0dd",
"title": "Xen and the art of virtualization"
},
{
"paperId": "ddcf503f1ea0d0e0a43cef7bce7819556a11e84f",
"title": "An Efficient and Reliable Multicasting for Smart Cities"
},
{
"paperId": "f13a16d437c9c83984e6f06ab0b157f424f7d7eb",
"title": "Towards Delay-Efficient Game-Aware Data Centers For Cloud Gaming"
},
{
"paperId": "2940e7bba0cf534b5caf6a8f7545b5012d1cdf6d",
"title": "An Efficient and Reliable Core-Assisted Multicast Routing Protocol in Mobile Ad-Hoc Network"
},
{
"paperId": "76aab02968d76197d7373e2f724c76f6463ebdec",
"title": "Evaluating and Modeling Virtualization Performance Overhead for Cloud Environments"
},
{
"paperId": null,
"title": "In the data center , power and cooling costs more than the it equipment it supports"
},
{
"paperId": "8fc928bb430d3f72ac876ca156042ad1860acacd",
"title": "Article in Press Future Generation Computer Systems ( ) – Future Generation Computer Systems Cloud Computing and Emerging It Platforms: Vision, Hype, and Reality for Delivering Computing as the 5th Utility"
},
{
"paperId": null,
"title": "In the data center, power and cooling costs more than the it equipment it supports"
},
{
"paperId": "6437561175be151f2a2ce5747443564b54f7140c",
"title": "Proceedings of the 5th Symposium on Operating Systems Design and Implementation"
}
] | 13,322
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Engineering",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/024e06bc350d5bb5599bd58a89ee18c3fc3ed122
|
[
"Computer Science"
] | 0.897986
|
A Data Security Enhanced Access Control Mechanism in Mobile Edge Computing
|
024e06bc350d5bb5599bd58a89ee18c3fc3ed122
|
IEEE Access
|
[
{
"authorId": "2106945079",
"name": "Yichen Hou"
},
{
"authorId": "144529280",
"name": "S. Garg"
},
{
"authorId": "2117271124",
"name": "Lin Hui"
},
{
"authorId": "1581170567",
"name": "DushanthaNalin K. Jayakody"
},
{
"authorId": "2068300460",
"name": "Rui Jin"
},
{
"authorId": "2237935030",
"name": "M. S. Hossain"
}
] |
{
"alternate_issns": null,
"alternate_names": null,
"alternate_urls": [
"http://ieeexplore.ieee.org/servlet/opac?punumber=6287639"
],
"id": "2633f5b2-c15c-49fe-80f5-07523e770c26",
"issn": "2169-3536",
"name": "IEEE Access",
"type": "journal",
"url": "http://www.ieee.org/publications_standards/publications/ieee_access.html"
}
|
Mobile edge computing, with characteristics of position awareness, mobile support, low latency, decentralization, and distribution, has received widespread attention from industry and academia, and has been applied to application areas such as intelligent transportation, smart city, and real-time big data analysis. However, it also brings the new security threats, especially data security threats during data access that leads to unauthorized/unauthorized access, alteration and disclosure of data, affecting the confidentiality and integrity of the data. Therefore, access control, as an important method to ensure the security of user data during data access, began to be applied to mobile edge computing. However, the existing access control has the disadvantages of coarse-grain, poor flexibility and accuracy, lack of internal attack considerations, etc., which cannot meet the needs of data security in practical applications of mobile edge computing. In this paper, a data security enhanced Fine-Grained Access Control mechanism (FGAC) is proposed to ensure data security during data access in mobile edge computing. In FGAC, a dynamic fine-grained trusted user grouping scheme based on attributes and metagraphs theory was first designed. Secondly, the scheme was combined with the traditional role-based access control mechanism to assign roles to users based on user group credibility. And then, based on attribute matching the user authentication further verifies whether the user is allowed to perform the access operations to achieve fine-grained data protection. Experimental results show that FGAC can effectively identify malicious users and make group adjustments, while achieving fine-grained access control and assure the data security during the data access process in mobile edge computing.
|
Received June 30, 2020, accepted July 18, 2020, date of publication July 23, 2020, date of current version August 5, 2020.
_Digital Object Identifier 10.1109/ACCESS.2020.3011477_
# A Data Security Enhanced Access Control Mechanism in Mobile Edge Computing
YICHEN HOU 1, SAHIL GARG 2,3, (Member, IEEE), LIN HUI 1,
DUSHANTHA NALIN K. JAYAKODY[3], (Senior Member, IEEE),
RUI JIN[4], AND M. SHAMIM HOSSAIN 5, (Senior Member, IEEE)
1College of Mathematics and Informatics, Fujian Normal University, Fuzhou 350117, China
2École de Technologie Supérieure (ETS), Montreal, QC H3C 1K3, Canada
3School of Computer Science and Robotics, Tomsk Polytechnic University, 634050 Tomsk, Russia
4College of Engineering, Mathematics, and Physical Sciences, University of Exeter, Exeter EX4 4QF, U.K.
5Department of Software Engineering, College of Computer and Information Sciences, King Saud University, Riyadh 11543, Saudi Arabia
Corresponding author: Lin Hui (linhui@fjnu.edu.cn)
This work was supported in part by the Competitive Enhancement Program of the Tomsk Polytechnic University, Russia No
VIU-ISHITR-180/2020; and in part by the Researchers Supporting Project number (RSP-2020/32), King Saud University, Riyadh,
Saudi Arabia.
**ABSTRACT** Mobile edge computing, with characteristics of position awareness, mobile support, low
latency, decentralization, and distribution, has received widespread attention from industry and academia,
and has been applied to application areas such as intelligent transportation, smart city, and real-time big data
analysis. However, it also brings the new security threats, especially data security threats during data access
that leads to unauthorized/unauthorized access, alteration and disclosure of data, affecting the confidentiality
and integrity of the data. Therefore, access control, as an important method to ensure the security of user
data during data access, began to be applied to mobile edge computing. However, the existing access control
has the disadvantages of coarse-grain, poor flexibility and accuracy, lack of internal attack considerations,
etc., which cannot meet the needs of data security in practical applications of mobile edge computing. In this
paper, a data security enhanced Fine-Grained Access Control mechanism (FGAC) is proposed to ensure
data security during data access in mobile edge computing. In FGAC, a dynamic fine-grained trusted user
grouping scheme based on attributes and metagraphs theory was first designed. Secondly, the scheme was
combined with the traditional role-based access control mechanism to assign roles to users based on user
group credibility. And then, based on attribute matching the user authentication further verifies whether the
user is allowed to perform the access operations to achieve fine-grained data protection. Experimental results
show that FGAC can effectively identify malicious users and make group adjustments, while achieving
fine-grained access control and assure the data security during the data access process in mobile edge
computing.
**INDEX TERMS Mobile edge computing, access control, data security, data confidentiality, data integrity,**
metagraph theory.
**I. INTRODUCTION**
In recent years, the development of intelligent mobile terminal technology such as smartphones, tablets, various Internet
of Things devices, and mobile communication technologies
s uch as 5G, the types of mobile applications such as face
recognition, augmented reality, virtual reality, live webcasting, etc. are also constantly enriched. Due to constraints
such as size, many mobile devices still have relatively scarce
resources such as computing, storage, network, and electrical
The associate editor coordinating the review of this manuscript and
approving it for publication was Md Zakirul Alam Bhuiyan .
energy, and cannot meet application requirements. To this
end, scholars have proposed the Mobile Cloud Computing
(MCC) [1] that expanding physical resources of device by
migrating tasks to cloud data center to meet all kinds of application of resource requirements. However, since the rapid
growth of the mobile devices and applications, the mobile
cloud computing mode is overly centralized, and the number
of server connections is extremely large, which will cause
huge pressure on the server and the network, resulting in
server downtime and excessive network delays, which seriously affects the user experience [2]. In view of the above
problems, the traditional centralized computing model needs
-----
**FIGURE 1. Architecture of mobile edge computing.**
to be further optimized and improved, and is developing
towards flattening and marginalization. In this context, as an
emerging technology, Mobile Edge Computing (MEC) [3],
[4] integrates the mobile access network with various network
services and has become an inevitable product that conforms
to this trend of development. By migrating the server from a
cloud data center to the mobile network edge, MEC reduces
physical distance between the mobile terminal and the
server. On the one hand, it can reduce the transmission delay
and ease the pressure on the backbone network. On the other
hand, it can also share the concentration heavy server load.
As shown in Fig. 1, a typical MEC is divided into 4 layers,
mobile terminal layer, edge network layer, edge data center
layer, and core infrastructure layer [5], [6]. In the MEC,
the edge terminal equipment is responsible for data perception and reception, and performs some preliminary data
processing. The wireless network is connected to the edge
network, and the edge network integrates a variety of communication networks to interconnect the mobile terminal and
the sensor network to upload the data to the edge data center.
The edge data center is deployed at the edge of the network
and is connected to the cloud center. And, the edge data center
performs data fusion processing according to the processing
results to feedback information or provide related services,
or transfer the processed data to the core infrastructure. The
of data storage, processing, and access operations are performed at the core infrastructure layer. The MEC architecture
built on this can provide a platform for data analysis of
Intelligent transportation, smart cities, and the Vehicle Area
network, etc.
With the vigorous development of technologies such as
5G, Internet of Things, and artificial intelligence [7], new
service models and services based on mobile edge computing
[9] will show an explosive growth trend, and generate ‘‘massive’’ data [10]. And, it also brings new security threats to
mobile edge computing [11], [12], especially data security
threats during data access. These security threats will lead to
unauthorized/unauthorized access, alteration and disclosure
of data [13], affecting the confidentiality and integrity [8] of
the data. Therefore, access control, as an important method to
ensure the security of user data during data access, began to
be applied to mobile edge computing. At present, the access
control mechanisms used in mobile edge computing are
mainly divided into two categories: Role-Based Access Control (RBAC) and Attribute-Based Access Control (ABAC)
[14]. However, existing mechanisms have the disadvantages
of coarse-grain, poor flexibility and accuracy, lack of internal
attack considerations, etc., which cannot meet the needs of
data security in practical applications of MEC.
To enhance the data security such as data confidentiality
and data integrity during data access process, a data security
enhanced Fine-Grained Access Control mechanism (FGAC)
is proposed, and the contributions of this work include:
(1) Combining the traditional RBAC with metagraph
theory based user grouping strategy and user attributes,
in FGAC, a novel role and attribute based access control
mechanism is proposed to achieve fine granularity of data
confidentiality and integrity assurance through fine-grained
grouping and access rights settings for users.
(2) In order to realize the fine-grained grouping of users,
a dynamic fine-grained trusted user grouping scheme based
on user attributes and metagraph theory is proposed. The
scheme divides user groups according to the attribute relevance between users and uses the metagraph theory to establish trust relationships based on the access behavior between
users. At the same time, a user group update module is also
designed to achieve the dynamic adjustment of user groups
within the user group.
(3) In order to reduce the probability of internal attacks and
achieve fine-grained data protection, a user re-authentication
based on attribute matching is proposed. The new authentication mechanism further verifies the matching of user
attributes and access data attributes after the user passes preliminary identity verification, restricts the malicious unauthorized access of authorized users, and realizes the fine-grained
protection of data.
**II. RELATED WORK**
In order to achieve more secure, efficient, and dynamic access
control to meet various application requirements, recently,
researchers combine RBAC and ABAC [14], and propose
some improved solutions.
Kuhn et al. [17] combined attribute-based and role-based
access control schemes for the first time to achieve effective
distributed access control and support dynamic role assignment and permission management. Wang et al. [18] proposed
an attribute encryption based novel RBAC scheme to provide more flexible access control by introducing the user
attributes into RBAC to implement attribute-based user roles
-----
and permission assignment. Mon and Naing [20] provide
an attributes and roles based access control method, and
formulate corresponding access policies to ensure personal
data and clouds private. Barkha and Sahani [21] designed
a context-based role activation and permission revocation
method. The proposed method effectively overcome the
shortcomings of traditional ABAC and ABAC, and achieve
the advantages of context-aware, fine-grained, etc. For the
SaaS model of cloud computing, Geetha and Anbarasi [22]
proposed a role-based and attribute-based Web service access
control mechanism to ensure the security of the service composition by ranking the possible chains of services based on
user’s role and sensitivity of related data. Yu et al. [23] combined attribute encryption algorithm with FAHP-based user
trust evaluation methods, and proposed an attribute and user
trust based RBAC to implement the fine-grained dynamic
authorization of access control.
Although the existing research results can provide certain
data access security, the implementation of the program will
generate a lot of additional overhead and cannot be directly
applied to mobile terminals with limited resources. At the
same time, these solutions lack the flexibility to meet the
fine-grained data security requirements associated with different scenarios and multiple services in mobile edge computing and the need to ensure that multiple categories of users
access different data. Besides, the lack of consideration of
internal attacks also makes these methods impossible to apply
directly to practice. Therefore, introducing an internal attack
defense mechanism and designing a fine-grained, flexible,
and accurate security access control mechanism against internal attacks will be a powerful guarantee for improving the
security of mobile edge computing data.
**III. ATTACK MODEL**
In FGAC, all users are divided into different groups, and each
user accesses data resources according to the role assigned
by the user group’s credibility. We consider collusion attacks
and self-improvement attacks initiated by internal attackers.
Attackers can increase their access to important resources
through collaboration, thereby threatening data security. The
specific attack is defined as follows:
- Collusion attack: Multiple attackers can cooperate and
provide false information to increase the reputation
value of malicious users and reduce the reputation value
of normal users, thereby affecting the security level of
users.
- Self-promotion attacks: Attackers try to increase their
reputation by mistake by providing false information
or exploiting calculation loopholes, thereby improving
their security level.
**IV. A DATA SECURITY ENHANCED FINE-GRAINED**
**ACCESS CONTROL MECHANISM (FGAC)**
Because of the existing access control problems such as
coarse-grained access control strategy, poor flexibility, and
accuracy, lack of internal attack considerations, etc., which
cannot meet the data security access requirements in practical
**TABLE 1. Main symbols.**
applications of mobile edge computing, this section proposes
a data security-oriented fine-grained access control mechanism FGAC. Table 1 shows the main symbols used in this
paper and their meanings. The overall architecture of FGAC
is shown in Fig. 2, which mainly contains two modules:
user role assignment and authority assignment. Among them,
the user role assignment module divides all users into different groups according to the evaluation result of the user
attribute relevance, and then assigns roles to each user group
according to the user group’s credibility. The rights assignment module re-authenticates the module based on the user
based on the attribute matching degree assign appropriate permissions to users. FGAC converts the user-role-permission
relationship into a user-user group-role-permission relationship, divides users into different groups according to the
user’s attribute values and access requirements, assigns corresponding roles and permissions to the user group, and also
validates the user role Perform user authentication with the
attribute matching degree, and then screen more qualified
users for access operations, and meet the different access
needs of users under the premise of ensuring user data
security.
The constituent elements in FGAC are defined as follows:
1) Users: a collection of data access requesters, denoted
as U, defined as
:
_U = {u1, u2, . . ., un},_
(n ∈ _N_ )(i, j, if i ̸= j then ui ̸= uj). (1)
2) Attribute relevance (AR): The similarity of the user’s
own attribute set. The higher the attribute correlation
between users, the closer the functions, access data
preferences, and security levels of different users are,
and the easier they are to be classified into a user
group.
3) User group (G): a group divided according to the evaluation results of the user attribute relevance, and the user
group is used as a transition between connecting users
and roles to form a user-user group-role authorization
method. Users in the same user group have similar
functions, similar security levels, access requirements,
and so on.
4) User group credibility: Measure the value of user group
credibility. Each user has a different security level,
and users in the same user group have similar security levels. User group credibility is determined by the
minimum security level of users in the group.
-----
**FIGURE 2. Overall architecture of FGAC.**
5) User group role (Roles): A role is a collection of responsibilities and access rights. In FGAC, role assignment
is performed for user groups, and different roles are
assigned to user groups with different credibility. At the
same time, the user roles in the group are divided
into A1 level roles and A2 level roles according to
the security level. The highest level A1 role is responsible for updating users in the group, etc.; the other
level roles are responsible for access operations without
change User group permissions. The roles and role
sets are collectively denoted as r and R, respectively,
defined as
:
�
_ri = {ui1, ui2, . . ., uik_ }, (k ∈ _N_ ) (2)
_R = {r1, r2, . . ., rm},_ (m ∈ _N_ ).
6) Permissions: It represents the specific access permission for different information content. Data owners
will add attributes to resources and data according to
their requirements, thereby restricting access by unauthorized users; operations are specific access modes
that users can perform, such as readable, modifiable,
or denied access, etc.
7) Attribute matching degree (AM ): The data owner
further restricts access users after verifying the user
role and can screen more suitable users for access
operations to ensure the security of their own data.
The data owner not only requires the user to have the
relevant role to obtain access qualification but also
further authenticates the access user. It requires that
the matching degree between the user attribute and the
where UAS[′] is the attribute set used in this interaction.UASi[i][nt]
and UASj[i][nt] are the attribute set used in each interaction
between users i and j, respectively. n is the total number
of interactions between users i and j. w is the threshold of
the proportion of attribute intersections.R(i,j) is the reputation
of j versus i stored in i’s local reputation database.τ is a
access data attribute is greater than the set threshold
before the user is allowed to access related data.
_A. USER GROUPING SCHEME BASED ON ATTRIBUTES_
_AND METAGRAPHS_
In this scheme, firstly, the data needs to be divided into
different levels according to the data sensitivity hierarchy
(sh). The data sensitivity hierarchy is determined by the
data owner. The higher the hierarchy, the greater the need
for confidentiality and data security. Secondly, according to
the evaluation results of Attribute Relevance (AR) between
users, all users are divided into different groups by using the
metagraph theory [16], [19].
Assume that each user has a set of attributes that including
specialty, access data preference, security level, etc., and
denoted as UAS = {uas1, uas2, . . ., uask }. The attribute
relevance AR(i,j)evaluated by user j for user i can be calculated
as follows
:
[1]
_n_
[×]
UASinti ∩ UAS[int]j
��� ���
��UAS′��
_n_ UASinti ∩ UAS[int]j
� ��� ���
_AR(i,j) = R(i:j) × τ ×_ _n[1]_ [×] int=1 ��UAS′��
_s.t._ UASinti ∩ UAS[int]j _> w._ (3)
��� ���
-----
_e1_ _< 0.4; 0.7 > indicates that the attribute correlation_
between user x1 and user x4 is 0.4 and there is a trust relationship. The trust relationship between the two is 0.7.
1) TRUST RELATIONSHIP BETWEEN USERS
According to the evaluation result of attribute relevance,
all users are divided into different groups using metagraph
theory. Assuming that user u and user u[′] belong to different
groups, the trust relationship between user u and user u[′] is
expressed as �TR �u, u[′][��], which is calculated as follows:
(1) When user u and user u[′] have direct interaction, the trust
relationship TR[direct]
(u,u[′]) [between][ u][ and][ u][′][ is calculated as follows][:]
**FIGURE 3. User grouping based on attributes and metagraphs.**
time factor that determines how much interaction time affects
_R(i,j). Then, τ is defined as follows:_
_τ = τi:j,Tn × θTn._ (4)
where θTnindicates the frequency of historical interactions
between users i and j up to time Tn. τi:j,Tnis a weighting factor,
which determines the degree of influence of the distribution
of the historical interactions of users i and j on R(i,j) up to Tn.
The calculation of τi:j,Tn and θTn is as follows:
|SH|
� _Nsh_
_θTn = 1 −_ _e_ [(][−] _sh=m1×n_ ) . (5)
_n_
�
_τi:j,Tn =_ _l=1_ ( _[T]m[l]_ [×][ l]n [)][.] (6)
where Nsh is the number of historical interactions performed
by users i and j based on the data sensitivity hierarchy(sh),
and m and n are the number of time slots and period T,
respectively.
The user grouping method based on metagraph theory is
defined as follows:
1) Construct the metagraph S =< X _, E > into a graph_
construction specified by its generation set X (user set)
and a set of edges E defined on the generation set.
2) Among them, the generation set X represents the user;
the edge between the meta nodes users) represents the
trust relationship between them. For example, edge
_e =< Ve, We >∈_ _E indicates that there is a trust_
relationship between user Ve and user We.
3) The weight of the edge e =< Ve, We _>∈_ _E is_
represented by a binary <ar _wr>, where ar represents_
;
the attribute correlation between the user Ve and the
user We; wr represents the trust relationship between
the user Ve and the user We, and the value range is [0,1].
As an example, consider the metagraph S =< X _, E >_
in Fig. 3. Generate set X = {x1, x2, x3, x4, x5, x6, x7} with
edge set E = {e1, e2, e3, e4}, where e1 =< x1, x4 _>_
_, e2 =< x4, x6 >, e3 =< x3, x5 >, e4 =< x5, x7 >._
First, divide X into 4 groups (G1, G2, G3, G4) according
to the attribute correlation between users, where G1
=
{x1}, G2 = �x2,, x3� _, G3 =_ �x4,, x5� _, G4 =_ �x6,x7�. Then,
the trust relationship between users is established according to the historical interaction between users. For example,
where slmax is the maximum security level of the person
directly recommended in DirR.
Then, each user updates the reputation value of the interacted user according to the calculated trust relationship value
between users. Assuming that user i sends an access request
to user j, hoping that j provides corresponding services, then
the credibility value from j to i can be calculated as follows
:
_R(i,j) = UQi × TR(i,j)._ (10)
Among them, TR(i,j) is the trust relationship between the
current users i and j. UQi is the user qualification of user i in
the user group. Because each user may have different status
and influence in a group, the higher the user’s UQ in the
group, the more likely their behavior will meet the group’s
1
_TR[direct](u,u[′])_ [=] SH
| | [×]
|SH|
�
_sh=i_
� _SI sh_
_TI_ _[sh][ ×][ ξ][sh]_
�
_._ (7)
_ξ = E(γt_ )
|SH|
�
γt = _j=i_ _IAj_
� |SH|
� (t = 1 . . . Nslot ). (8)
_IAj,_
_j=1_
where i is the lowest data sensitivity level. SI _[sh]_ and TI _[sh]_
represent the number of successful data interactions with the
sensitivity hierarchy(sh) and the total number of interactions,
respectively. ξ is a weighting factor, which determines the
degree to which the sensitivity hierarchy (sh) affects TR[direct]
(u,u[′])
when the two interact. γt is the ratio between the number
of interactions with a sensitivity hierarchy higher than the
currently required sensitivity hierarchy i and the total number
of interactions at all sensitivity hierarchies. IAj represents
the number of times the sensitivity hierarchy in the historical
interaction is confirmed as j, and Nslot represents the number
of time slots.
(2) When users u and u[′] do not directly interact, assume
_DirR = {dir −_ _reci|i = 1 . . . m} is a set of direct recom-_
menders. The direct recommender uj has direct interaction
with the user u[′] and has the result of direct trust relationship evaluation about u[′]. Then the indirect trust relationship
_TR[indirect]_ between u and u[′] is calculated as follows
(u,u[′]) :
�
_._ (9)
_TR[indirect](u,u[′])_ = [1]
_m_
[×]
_m_
�
_j=1,uj∈DirR_
� _slj_
_slmax_ × TR[direct](u,uj)
-----
standards. Let _g be the group, and the UQ of the user in_ _g is_
¯ ¯
defined as follows
:
�
_UQ = κ1 ×_ [1]g _AR(u¯, u) + κ2 ×_ _g[1]_
|¯| [×] _u∈¯g,u̸=¯u_ |¯|
�
× _TR(u¯, u)_
_u∈¯g,u̸=¯u_ (11)
_TR(u¯, u) = ρ1 × TR[direct](u¯,u)_ [+][ ρ][2][ ×][ TR]([indirect]u¯,u)
_κ1 + κ2 = 1_
ρ1 + ρ2 = 1.
Because user i will interact with multiple users, according
to the change of the trust relationship between the data owner
and user i and the update of the reputation value after each
interaction, the comprehensive reputation value R[sum]i of user
_i can be calculated as follows_
:
_kn_
�
_R[sum]i_ = _k[1]n_ _SLj[k][n]_ × λsl × R(i,j). (12)
_n=1_
where kn is the total number of interactions between user i
and other users. SLj[k][n] is the security level of the data owner
_j during the kn interaction of user i. λsl is the proportion of_
the reputation value of user i provided by data owners with
different security levels.
Assuming that the security level is divided into n levels,
the security level of user i is divided according to the comprehensive reputation value of user i. When R[sum]i ∈ [TSj, TSj+1]
is satisfied, the security level of user i is j + 1, j ∈ [j, n],
where TSj+1 and TSj is the upper limit of the credibility value
corresponding to different security levels.
2) USER GROUP UPDATE
After the initial grouping of users, it is assumed that user
_x belongs to user group g. After some access operations,_
the change of user attributes may no longer meet the requirements of user group g. At this time, user x needs to be
comprehensively evaluated to determine whether the user still
meets the Group g requirements.
(1) If the following constraints are met, the original
grouping remains unchanged, and user x still belongs to user
group g;
1
�
_AR(u, x) > θ_
_G_ 1
| | − [×] _u∈G,x̸=u_
1
�
_TR(u, x) > θ_ [′]
_G_ 1
| | − [×] _u∈G,x̸=u_ (13)
_R[sum]x_ _> CG_
_TR(u, x) = ρ1 × TR[direct](u,x)_ [+][ ρ][2][ ×][ TR]([indirect]u,x)
ρ1 + ρ2 = 1.
where CG is the reputation threshold set by the current user
group G. θ and θ [′] are the thresholds of attribute relevance and
trust relationship set by group G, respectively.
(2) If the user x does not meet the constraints set by the
user group g, the user group update module (GUM) is used
to update the user x grouping.
where γj[y] is a weighting factor, which determines the
importance of the jth attribute of the attributes required by
the data y, and γj[y] [is set by the data owner.]
Finally, the data owner z compares the attribute matching degree AM(x,y) of the user x and the data y with the
attribute matching degree threshold Tsy, where is the threshold of the attribute matching degree set by the access data y.
If AM(x,y) ≥ _Tsy, it is determined that user x is granted_
**FIGURE 4. User group update module.**
The user group update module (GUM) mainly provides
two functions, as shown in Fig. 4.
One is the redistribution of user groups. This function first
integrates the constraints set by all user groups into a list,
then calculates the relevant value of user x according to the
constraints set by the user group, and finally compares the
calculation results with the constraints in the list to assign user
_x to In the corresponding group._
The second is the change of constraints. The constraints
here refer to the constraints set for each group in the user
group redistribution function list. This function mainly provides the update of user group constraints. If the user group
has not changed much within a certain period of time, this
function will regularly update the constraints set by the user
group; if the user changes within the user group are too
large, the originally set constraints will no longer meet the
group status, the user The group can immediately submit the
constraint condition update to the user group update module,
and replace the constraint condition of the group in the user
group redistribution function list.
_B. USER AUTHENTICATION BASED ON ATTRIBUTE_
_MATCHING DEGREE_
The user requests access to certain data. After verifying that the user role is qualified to access the data,
the data owner needs to further authenticate the access user
by calculating the attribute matching degree. Assume that
_UcA = {ucai|i = 1 . . . n} is a set of user attributes corre-_
sponding to the data attribute requirements. When user x
sends an access request to data owner z, indicating that he
wants to access data y, the attribute matching degree of user
_x and data y is calculated as follows_
:
_AM(x,y) =_
_n_
� _γj[y]_ [×][ uca][j][.] (14)
_j=1,aj∈UcA_
-----
relevant permissions and user x is allowed to perform the
access operation.
_C. FINE-GRAINED ACCESS CONTROL MECHANISM_
_BASED ON USER GROUPING_
To ensure the security of user data, the FGAC access control
strategy is mainly divided into two parts: role assignment
strategy and user authorization strategy.
- Role assignment strategy
FGAC first divides all users into different groups based
on user attribute relevance. Users in the same user
group have similar functions, similar security levels,
and access requirements, etc. Therefore, role assignment
is performed for the entire user group, only the user
group When the credibility is greater than the threshold
set by the role, users in the user group can obtain the
corresponding role.
- User authorization strategy
When a user wants to access a certain item of data,
the data owner will often further set the access rights
for the item of data according to his requirements, not
just the role constraints. After verifying that the user
role is qualified to access the data, the data owner
will re-authenticate the user based on the attribute
matching degree, and calculate the matching degree
between the user attribute and the data attribute. Only
when the matching degree of the two attributes is greater
than the threshold set by the data owner can the user
obtain the corresponding authority, and then access the
data for related operations. This can ensure the security
of the data owner’s data, and prevent users with relevant
roles and attributes who do not meet the requirements
from accessing relevant data.
The specific implementation process of FGAC is shown
in Fig. 5, and the access control process is described as
follows:
(1) User u sends an access request to a certain data;
(2) The data owner performs an authorization check on
the access request of user u, first verifying whether the role
owned by user u is in the set of roles defined in the data and
determining whether user u is qualified to access the data.
If the role of user u is in the set of accessible roles of this item
of data, step (3) is performed; otherwise, the access request
of user u is denied;
(3) After the user role is verified, the user re-authentication
based on the attribute matching degree is then performed to
calculate the matching degree between the user u attribute and
the data attribute. If the attribute matching degree of the two
meets the threshold defined by the data, the user is granted
the corresponding permission to allow user u to perform the
access operation; otherwise, the access request of user u is
denied;
(4) After the user, u’s visit is over, first update the trust
relationship between users according to the user’s access
behavior, and then update the user’s reputation value to adjust
the user’s security level.
**FIGURE 5. FGAC implementation process.**
**V. SIMULATION VERIFICATION AND ANALYSIS**
The experiments in this section mainly verify and analyze the
user security and authorization fine-grained aspects. In the
Windows 7 environment, the configuration is i7-5500U CPU,
8.0GB memory, 1TB hard disk, and simulation verification
using MATLAB2017b. In the experiment, we assume that
there are 100 mobile terminal users, among which a certain
number of malicious users. Malicious users are not always
performing malicious visits, while normal users’ visits are
always benign.
Among the parameters used in this paper, κ1 and κ2 are the
weighting factors of equation (11). We set κ1 and κ2 to 0.4 and
0.6 respectively, which determine the degree of influence of
attribute relevance and the trust relationship between users on
user qualifications(UQ); ρ1 and ρ2 are the weighting factors
in equation (11) and equation (13). We set ρ1 and ρ2 to 0.6 and
0.4 respectively, which determine the degree of influence of
the direct and indirect trust relationship between users on the
trust relationship(TR).
-----
**FIGURE 6. User’s reputation changes.**
_A. USER SAFETY ANALYSIS_
The user security is determined by the user’s security level,
and the user security level is adjusted by updating the trust
relationship between users and the user’s reputation value
after each interaction. The trust relationship between users
reflects the historical interaction between users based on
different data sensitivity hierarchies.
In Fig. 6(a), it is assumed that two users are in the same user
group and the reputation values are equal. To prevent malicious users from excluding the user group and thereby update
the user group, we set the user group reputation threshold
_CG = 0. From the results in the figure, it can be found_
that with the increase of time, the reputation values of the
two users change significantly. On the one hand, when normal users interact with other users, their normal and benign
behavior causes their reputation value to continue to increase;
on the other hand, when malicious users interact, their malicious behavior makes their reputation value continue to
decrease, This is the same as what we estimated. Fig. 6(b)
shows the changes in the reputation value and security level
of users with high reputation values when their proportion
of malicious behavior continues to increase. As can be seen
from the results in the figure, even users who performed
well in the previous historical interactions will have their
**FIGURE 7. Average UIA with different proportions of malicious users.**
reputation value lower as the malicious behavior continues to
increase in the later period, and the user’s security level will
gradually adjust from the high level ‘‘1’’ To the lower level
‘‘4’’, the user’s safety is re-evaluated.
Besides, based on the historical interaction between users,
we consider comparing and evaluating FGAC, TARAS [15],
and RBE in terms of user recognition accuracy and successful acceptance rate, because they are all role-based access
control mechanisms, in which TARAS provides users with
permissions based on the estimation of the dynamic trust
relationship between users, similar to the FGAC mechanism.
- User identification accuracy(UIA): the accuracy of
identifying normal users and malicious users;
- Successful acceptance rate(SAR): The ratio of the number of access requests that do not meet the security
requirements to the total number of access requests.
1) USER IDENTIFICATION ACCURACY
First, we compared the accuracy of user identification
between the two schemes of FGAC and TARAS under the
proportion of 20% and 30% malicious users with an attack
probability of 1, where the attack probability determines the
possibility of malicious users attacking. The greater the probability, the higher the frequency of malicious user attacks.
Fig. 7(a) and Fig. 7(b) show the comparison between the
-----
accuracy of identifying normal users and malicious users
when the proportion of malicious users is 20% and 30%,
respectively. It can be seen from the figure that as the proportion of malicious users increases, the accuracy of user
identification in both schemes decreases. But at the same
time, it can also be found that in the case of a fixed proportion
of malicious users (20% or 30%), after a long period of
observation and comprehensive evaluation of users, the accuracy of user identification in both schemes has increased,
and the accuracy of the FGAC scheme is higher. Although
both schemes restrict the access of malicious users by setting
thresholds, FGAC combines the division of user groups based
on attribute correlation and the establishment of trust relationships, and FGAC sets trust thresholds for user groups. The
range of users in the group is small and similar, so users in
the group can provide more accurate evaluation references,
which improves the accuracy of evaluating users’ security
level, and it is easier to detect malicious users and adjust the
user group. Therefore, the accuracy of user identification is
slightly Higher than TARAS.
At the same time, we also compared the accuracy of user
identification of the two schemes under different malicious
user attack probabilities when the proportion of malicious
users was 20%. Fig. 8(a) and Fig. 8(b) show the comparison
between the accuracy of identifying normal users and malicious users when the attack probability of malicious users is
30% and 70%, respectively. As can be seen from the figure,
as the probability of malicious user attacks increases, the possibility of malicious user exposure increases accordingly,
so the accuracy of user identification in both schemes has
increased. But at the same time, it can also be found that,
regardless of the increase in time or the probability of malicious user attacks, the accuracy of FGAC user identification
is still higher than that of TARAS. The reason is that the
user group division scheme based on attribute relevance in
FGAC divides users with similar security levels into a group.
If there is a malicious user in the group and the proportion
of the user’s malicious behavior increases, GRM can identify
the malicious user in time by establishing a trust relationship
between users and setting a user group trust threshold.
2) SUCCESSFUL ACCEPTANCE RATE
Fig. 9 is a comparison of the successful acceptance rate of the
three schemes of FGAC, TARAS, and RBE. As can be seen
from the figure, as the number of interactions, and the proportion of malicious users increase, the successful acceptance
rate of the three schemes has increased. In general, the successful acceptance rate of FGAC and TARAS is better than
RBE. As shown in Fig. 9(b), when the proportion of malicious
users is 0-20%, the overall successful acceptance rate of
the two schemes is not much different. As the proportion
of malicious users continues to increase, TARAS’s successful acceptance rate has increased, while FGAC’s successful
acceptance rate has changed less and is relatively stable. This
is because the establishment of the trust relationship between
users makes the adjustment of the user’s security level more
**FIGURE 8. Average UIA with different attack probabilities.**
accurate so that more credible users can be selected during
data access. Besides, the user re-authentication based on
attribute matching proposed in the FGAC can screen out users
who are more in line with the access requirements based
on the user’s true attributes and reduce the probability of
collusion attacks, which also improves the security of the data
access process, and decreases the successful acceptance rate
of FGAC.
_B. AUTHORIZED FINE-GRAINED VERIFICATION_
Authorized fine-grained verification is mainly to determine
whether more fine-grained access control is achieved than
the traditional RBAC model. In the simulation experiment,
7 users are specifically set, and each user’s attribute set
includes ID, name, department, job title, work experience,
the annual number of operating tables, and security level.
The security level is determined by the user’s comprehensive
reputation value. Table 2 lists the detailed information of each
user.
After preliminary experiment setting, the threshold of
user group credibility corresponding to the role is shown
in Table 3. Table 4 is the attribute requirements set by the
data Data_1 and the data Data_2.
The user access results are shown in Table 5. If user Staff _0
and Staff _3 request access to data Data_1 at the same time,
-----
**FIGURE 9. Average SAR.**
**TABLE 2. User information.**
first verify whether the roles of the two users meet the requirements of data Data_1. At this time, the roles owned by both
users are Role_2, which is consistent with the data Data_1
request. Then further verify other attributes. User Staff _0
and Staff _3 are the director physicians of the Department of
Neurology. The work experience and the number of operating
tables are different. At this time, the matching degree of
the user attribute and the data attribute can be calculated
**TABLE 3. The credibility of the user group corresponding to the role.**
**TABLE 4. Data attribute requirements.**
**TABLE 5. Access results.**
according to equation (14). Assuming that the weight of
work experience in the data Data_1 is 0.4 and the weight
of the annual number of operating tables is 0.6. According
to the calculation, the user Staff _0 is more in line with the
requirements of the data Data_1, then the user Staff _0 is
allowed to perform the access operation, and the user Staff _3
is denied the access request.
In addition, if user Staff _6 and user Staff _5 request access
to data Data_2 at the same time, the roles owned by the
two users meet the requirements of Data_2. Although user
_Staff _6 and user Staff _5 belong to internal medicine, user_
_Staff _5 belongs to respiratory medicine, which is more in line_
with the requirements of data Data_2. After attribute matching calculation, user Staff _5 is allowed to perform access
operations. In the traditional RBAC model, for example,
the user Staff _0 and the user Staff _3 are all assigned the role
of Role_2, so in the subsequent data access process, the two
have the same permissions. The FGAC scheme proposed in
this article adds the user re-authentication module based on
the attribute matching degree. According to the matching
degree of different attribute values and data attributes of the
user, even if the user Staff _0 and the user Staff _3 have
the same role, the permissions they have will be different,
thus enabling more fine-grained authorization to ensure the
security of user data.
**VI. CONCLUSION**
Aiming at the problems that the existing access control policies have coarse granularity, poor flexibility and accuracy,
and lack of internal attack considerations, which cannot meet
the data security access requirements in practical applications of MEC, this paper proposes a data security enhanced
-----
Fine-Grained Access Control mechanism(FGAC) based on
user grouping. First, the attribute relevance evaluation for
users is carried out, and a dynamic fine-grained trusted user
grouping scheme is designed based on the above evaluation results and metagraph theory. Then, combined with
role-based access control, the scheme assigns roles based on
the credibility of user groups and further verifies users based
on attribute matching, to achieve fine-grained protection of
data and reduce the risk of internal attacks. Experimental
results show that FGAC can effectively limit the access of
malicious users and update user groups in time, and ensure
the security of user’s data by implementing more fine-grained
access control. For future work, we intend to introduce
blockchain technology into the access control mechanism in
mobile edge computing to solve data security issues in the
process of data access further.
**REFERENCES**
[1] J. X. Zhai, ‘‘Research on authentication protocol in mobile cloud computing,’’ Ph.D. dissertation, Dept. Elect. Eng., Harvard Univ., JiangSu, China,
2019.
[2] Y. Chen, D. J. Xu, and L. Xiao, ‘‘Survey on network security based on
blockchain,’’ Telecommun. Sci., vol. 34, no. 3, pp. 10–16, Mar. 2018.
[3] K. Kaur, S. Garg, G. Kaddoum, M. Guizani, and D. N. K. Jayakody,
‘‘A lightweight and privacy-preserving authentication protocol for mobile
edge computing,’’ in Proc. IEEE Global Commun. Conf. (GLOBECOM),
Waikoloa, HI, USA, Dec. 2019, pp. 1–6.
[4] N. Abbas, Y. Zhang, A. Taherkordi, and T. Skeie, ‘‘Mobile edge computing:
A survey,’’ IEEE Internet Things J., vol. 5, no. 1, pp. 450–465, Feb. 2018.
[5] X. Y. Ma, ‘‘Research on trusted cooperative mechanism based on edge
computing,’’ M.S. thesis, Dept. Abbrev., Beijing Univ. Posts Telecommun.,
Beijing, China, 2019.
[6] X. Li, S. Liu, F. Wu, S. Kumari, and J. J. P. C. Rodrigues, ‘‘Privacy
preserving data aggregation scheme for mobile edge computing assisted
IoT applications,’’ IEEE Internet Things J., vol. 6, no. 3, pp. 4755–4763,
Jun. 2019.
[7] X. Guo, H. Lin, Z. Li, and M. Peng, ‘‘Deep-reinforcement-learning-based
QoS-aware secure routing for SDN-IoT,’’ IEEE Internet Things J., vol. 7,
no. 7, pp. 6242–6251, Jul. 2020.
[8] T. Wang, M. Z. A. Bhuiyan, G. Wang, L. Qi, J. Wu, and T. Hayajneh,
‘‘Preserving balance between privacy and data integrity in edge-assisted
Internet of Things,’’ IEEE Internet Things J., vol. 7, no. 4, pp. 2679–2689,
Apr. 2020.
[9] Y. Mao, C. You, J. Zhang, K. Huang, and K. B. Letaief, ‘‘A survey on
mobile edge computing: The communication perspective,’’ IEEE Commun.
_Surveys Tuts., vol. 19, no. 4, pp. 2322–2358, 4th Quart., 2017._
[10] W. S. Shi, H. Sun, J. Cao, Q. Zhang, and W. Liu, ‘‘Edge computing-an
emerging computing model for the Internet of everything era,’’ J. Comput.
_Res. Develop., vol. 54, no. 5, pp. 907–924, Feb. 2017._
[11] H. Li, G. Shou, Y. Hu, and Z. Guo, ‘‘Mobile edge computing: Progress and
challenges,’’ in Proc. 4th IEEE Int. Conf. Mobile Cloud Comput., Services,
_Eng. (MobileCloud), Oxford, U.K., Mar. 2016, pp. 83–84._
[12] S. Wang, X. Zhang, Y. Zhang, L. Wang, J. Yang, and W. Wang, ‘‘A survey
on mobile edge networks: Convergence of computing, caching and communications,’’ IEEE Access, vol. 5, pp. 6757–6779, Mar. 2017.
[13] H. Tao, M. Z. A. Bhuiyan, M. A. Rahman, G. Wang, T. Wang,
M. M. Ahmed, and J. Li, ‘‘Economic perspective analysis of protecting
big data security and privacy,’’ Future Gener. Comput. Syst., vol. 98,
pp. 660–671, Mar. 2019.
[14] J. L. Zhang, Y. C. Zhao, B. Chen, F. Hu, and K. Zhu, ‘‘Survey on data
security and privacy-preserving for the research of edge computing,’’
_J. Commun., vol. 39, no. 3, pp. 1–21, Mar. 2018._
[15] B. Gwak, J.-H. Cho, D. Lee, and H. Son, ‘‘TARAS: Trust-aware role-based
access control system in public Internet-of-Things,’’ in Proc. 17th IEEE
_Int. Conf. Trust, Secur. Privacy Comput. Commun., New York, NY, USA,_
2018, Aug. 2018, pp. 74–85.
[16] Y. Zhu, B. Li, H. Fu, and Z. Li, ‘‘Core-selecting secondary spectrum
auctions,’’ IEEE J. Sel. Areas Commun., vol. 32, no. 11, pp. 2268–2279,
Nov. 2014.
[17] D. R. Kuhn, E. J. Coyne, and T. R. Weil, ‘‘Adding attributes to role-based
access control,’’ Computer, vol. 43, no. 6, pp. 79–81, Jun. 2010.
[18] Y. Wang, Y. Ma, K. Xiang, Z. Liu, and M. Li, ‘‘A role-based access control
system using attribute-based encryption,’’ in Proc. Int. Conf. Big Data
_Artif. Intell. (BDAI), Beijing, China, Jun. 2018, pp. 128–133._
[19] M. Ezhei and B. T. Ladani, ‘‘GTrust: A group based trust model,’’ Int. J. Inf.
_Secur., vol. 5, no. 2, pp. 155–169, Mar. 2014._
[20] E. E. Mon and T. T. Naing, ‘‘The privacy-aware access control system
using attribute-and role-based access control in private cloud,’’ in Proc. 4th
_IEEE Int. Conf. Broadband Netw. Multimedia Technol., Shenzhen, China,_
Oct. 2011, pp. 447–451.
[21] P. Barkha and G. Sahani, ‘‘Flexible attribute enriched role based access
control model,’’ in Proc. Int. Conf. Inf., Commun., Instrum. Control (ICI_CIC), Indore, India, Aug. 2017, pp. 1–6._
[22] N. Geetha and M. S. Anbarasi, ‘‘Role and attribute based access control
model for Web service composition in cloud environment,’’ in Proc. Int.
_Conf. Comput. Intell. Data Science(ICCIDS), Chennai, India, Jun. 2017,_
pp. 1–4.
[23] B. Yu, X. Q. Tai, and Z. J. Ma, ‘‘Research on RBAC model based on
attribute and trust in cloud computing environment,’’ Comput. Eng. Appl.,
vol. 56, no. 9, pp. 84–92, May 2020.
YICHEN HOU received the bachelor’s degree in
software engineering from Xinyang Normal University, China, in 2018. She is currently pursuing
the master’s degree with the School of Mathematics and Information, Fujian Normal University.
Her research interests include blockchain, access
control, and network security.
SAHIL GARG (Member, IEEE) received the Ph.D.
degree from the Thapar Institute of Engineering
and Technology, Patiala, India, in 2018. He is
currently a Postdoctoral Research Fellow at École
de technologie supérieure, Université du Québec,
Montréal, Canada. He has many research contributions in the area of machine learning, big data analytics, security & privacy, internet of things, and
cloud computing. He has over 60 publications in
high ranked Journals and Conferences, including
40+ top-tier journal papers and 20+ reputed conference articles. He was
awarded the IEEE ICC best paper award in 2018 at Kansas City, Missouri.
He is currently a Managing Editor of Springer’s Human-centric Computing
and Information Sciences (HCIS) journal. He is also an Associate Editor of
the IEEE NETWORK MAGAZINE, IEEE SYSTEM JOURNAL, Elsevier’s Applied Soft
Computing, Elsevier’s Future Generation Computer Systems (FGCS), and
_Wiley’s International Journal of Communication Systems (IJCS). In addition,_
he also serves as the Workshops and Symposia Officer for the IEEE ComSoc
Emerging Technology Initiative on Aerial Communications. He guest-edited
a number of special issues in top-cited journals, including IEEE T-ITS, IEEE
TII, IEEE IoT Journal, IEEE NETWORK, and Future Generation Computer
Systems (Elsevier). He serves/served as the workshop chair/publicity cochair for several IEEE/ACM conferences, including the IEEE INFOCOM, IEEE
GLOBECOM, and IEEE ICC and ACM MobiCom. He is a member of ACM.
LIN HUI received the Ph.D. degree in computing
system architecture from the College of Computer
Science, Xidian University, China, in 2013. He is
currently a Professor with the College of Mathematics and Informatics, Fujian Normal University,
Fuzhou, China, where he is also a M.E. Supervisor.
He has published more than 50 papers in international journals and conferences. His research
interests include mobile cloud computing systems,
blockchain, and network security.
-----
DUSHANTHA NALIN K. JAYAKODY (Senior
Member, IEEE) received the M.Sc. degree (Hons.)
in electronics and communications engineering
from Eastern Mediterranean University, Turkey
(under the University Graduate Scholarship), and
the Ph.D. degree in electronics and communications engineering from University College Dublin,
Ireland, under the supervision of Prof. M. Flanagan
(Science Foundation Ireland Grant). From 2014 to
2016, he has held a Postdoctoral position at the
Coding and Information Transmission Group, University of Tartu, Estonia,
and the University of Bergen, Norway. Since 2016, he has been a Professor
with the School of Computer Science and Robotics, National Research
Tomsk Polytechnic University, Russia. He has held various visiting positions
at the Texas A&M University, Qatar, the University of Jyväskylä, Finland,
and the National Institute of Technology, Trichy, India. He has served as
the Session Chair or a Technical Program Committee Member for various
international conferences, such as IEEE PIMRC 2014–2020, IEEE WCNC
2014–2020, and IEEE VTC 2015–2019.
RUI JIN received the bachelor’s degree in
computer science from the University of Science
and Technology Beijing. She is currently pursuing
the Ph.D. degree in computer science with the
University of Exeter, U.K. Her research interests
include network security, machine learning, and
mobile edge computing.
M. SHAMIM HOSSAIN (Senior Member, IEEE) is currently a Professor
with the Department of Software Engineering, College of Computer and
Information Sciences, King Saud University, Riyadh, Saudi Arabia. He is
also an Adjunct Professor with the School of Electrical Engineering and
Computer Science, University of Ottawa, Canada. He has authored and
coauthored more than 260 publications including refereed journals (200+
SCI/ISI-Indexed papers, 100+ IEEE/ACM Transactions/Journal papers, 10+
ESI highly cited papers, 1 hot paper), conference papers, books, and book
chapters. Recently, he co-edited a book on ‘‘Connected Health in Smart
Cities’’, published by Springer. He has served as cochair, general chair,
workshop chair, publication chair, and TPC for over 12 IEEE and ACM conferences and workshops. He is currently the cochair of the 3rd IEEE ICME
workshop on Multimedia Services and Tools for smart-health (MUST-SH
2020). He is a recipient of a number of awards, including the Best Conference
Paper Award and the 2016 ACM Transactions on Multimedia Computing, Communications and Applications (TOMM) Nicolas D. Georganas Best
Paper Award, and the 2019 King Saud University Scientific Excellence
Award (Research Quality). He is on the editorial board of the IEEE
TRANCTIONS ON MULTIMEDIA, the IEEE NETWORK, the IEEE MULTIMEDIA, the
IEEE WIRELESS COMMUNICATIONS, IEEE ACCESS, the Journal of Network and
Computer Applications (Elsevier), and the International Journal of Multimedia Tools and Applications (Springer). He also presently serves as a lead
guest editor of IEEE Network, ACM Transactions on Internet Technology,
ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM) and Multimedia systems Journal. Previously, he served as
a guest editor of IEEE Communications Magazine, IEEE Network, the IEEE
TRANCTION INFORMATION TECHNOLOGY IN BIOMEDICINE (currently JBHI), the
IEEE TRANCTIONS ON CLOUD COMPUTING, International Journal of Multimedia
_Tools and Applications (Springer), Cluster Computing (Springer), Future_
_Generation Computer Systems (Elsevier), Computers and Electrical Engi-_
_neering (Elsevier), Sensors (MDPI), and International Journal of Distributed_
_Sensor Networks. He is a senior member of the ACM._
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.1109/access.2020.3011477?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/access.2020.3011477, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "GOLD",
"url": "https://ieeexplore.ieee.org/ielx7/6287639/8948470/09146646.pdf"
}
| 2,020
|
[
"JournalArticle"
] | true
| null |
[
{
"paperId": "ad42f67fda2b83b46ab0d07d137c49c148c1ebc9",
"title": "Deep-Reinforcement-Learning-Based QoS-Aware Secure Routing for SDN-IoT"
},
{
"paperId": "e240ed44143e68d8a61954a5df230b11657e0f5f",
"title": "Preserving Balance Between Privacy and Data Integrity in Edge-Assisted Internet of Things"
},
{
"paperId": "9de4028b8e63a9ab7b7dd13ec4114c04a282a7f5",
"title": "Economic perspective analysis of protecting big data security and privacy"
},
{
"paperId": "0cc5e894cdba0f33c56877e54e1ca493b000f432",
"title": "A Lightweight and Privacy-Preserving Authentication Protocol for Mobile Edge Computing"
},
{
"paperId": "5ec0565869e9c024c3ac98da9f59f1a2f4d744bb",
"title": "Privacy Preserving Data Aggregation Scheme for Mobile Edge Computing Assisted IoT Applications"
},
{
"paperId": "a8029d5a0d45c9f1ce1ea4ca46564b00eaf18f42",
"title": "TARAS: Trust-Aware Role-Based Access Control System in Public Internet-of-Things"
},
{
"paperId": "c117ee3eef61ad6dc2313b38a6caf0efcfbef745",
"title": "A Role-Based Access Control System Using Attribute-Based Encryption"
},
{
"paperId": "e7f84b1d7f8378ffaadbf85c33bacc8bcd9e28dd",
"title": "Mobile Edge Computing: A Survey"
},
{
"paperId": "dead20583beaea14ec2d9a5451c2d28c13d0be5a",
"title": "Flexible attribute enriched role based access control model"
},
{
"paperId": "4095249c5a04dc38e91a2b7c7af5cdcb08b5acbe",
"title": "Role and attribute based access control model for web service composition in cloud environment"
},
{
"paperId": "1be70c1cc40865c281b3f59e973f9bd8a8cf06c8",
"title": "A Survey on Mobile Edge Networks: Convergence of Computing, Caching and Communications"
},
{
"paperId": "e9b28ed88dd50cc80bde3c7cb835b1d7b0feea68",
"title": "Edge Computing—An Emerging Computing Model for the Internet of Everything Era"
},
{
"paperId": "8c5293da3ad1a463cb9694edfbf1bf19b8cbd698",
"title": "A Survey on Mobile Edge Computing: The Communication Perspective"
},
{
"paperId": "638c4339628b8b4cf34136afda4072bbb3e4296d",
"title": "Mobile Edge Computing: Progress and Challenges"
},
{
"paperId": "52135387c4cbc881b2220d7d10f53874129ebc44",
"title": "Core-Selecting Secondary Spectrum Auctions"
},
{
"paperId": "1e22da07e6694af27e3adb517dc938b219660185",
"title": "GTrust: a group based trust model"
},
{
"paperId": "14ffb5fb85c8a7fd11254a5601821891f084d31e",
"title": "The privacy-aware access control system using attribute-and role-based access control in private cloud"
},
{
"paperId": "3c647cbcbebe293712ca9678b1377d50b581a922",
"title": "Adding Attributes to Role-Based Access Control"
}
] | 13,592
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/025150bc246547d7eabeae51326b6d24c154d4b3
|
[
"Computer Science"
] | 0.849105
|
Verifier-based Password Authenticated 3P-EKE Protocol using PCLA Keys
|
025150bc246547d7eabeae51326b6d24c154d4b3
|
[
{
"authorId": "9380768",
"name": "A. Raghuvamshi"
},
{
"authorId": "52538632",
"name": "Premchand Parvataneni"
}
] |
{
"alternate_issns": null,
"alternate_names": null,
"alternate_urls": null,
"id": null,
"issn": null,
"name": null,
"type": null,
"url": null
}
|
This paper endeavors to present a novel framework for the generic structure of a verifier-based password authenticated Three-Party Encrypted Key Exchange (3P-EKE) protocol which yields more efficient protocol than the ones knew before. A previous framework presented by Archana and Premchand is more secured against all types of attacks like password guessing, replay, pre-play, man-in-the-middle attack etc. But unfortunately, this protocol does not solve the problem of a server compromise. These proofs help as inspiration to search for another framework. The framework we offer produces more efficient 3P-EKE protocol, and, in addition, delivers perceptive clarification about the existing attacks that do not solve in the previous framework. Moreover, it allows direct change from a class of verge private-key encryption to a hybrid (symmetric & Asymmetric) one without significant overhead.
|
Published Online June 2016 in MECS (http://www.mecs-press.org/)
DOI: 10.5815/ijcnis.2016.06.07
# Verifier-based Password Authenticated 3P-EKE
Protocol using PCLA Keys
## Archana Raghuvamshi
Adikavi Nannaya University /CSE Department, Rajahmundry, 533296, India
E-mail: archana_anur@yahoo.in
## Premchand Parvataneni
Osmania University/CSE Department, Hyderabad, 500007, India
E-mail: profpremchand.p@gmail.com
**_Abstract—This paper endeavors to present a novel_**
framework for the generic structure of a verifier-based
password authenticated Three-Party Encrypted Key
Exchange (3P-EKE) protocol which yields more efficient
protocol than the ones knew before. A previous
framework presented by Archana and Premchand is more
secured against all types of attacks like password
guessing, replay, pre-play, man-in-the-middle attack etc.
But unfortunately, this protocol does not solve the
problem of a server compromise. These proofs help as
inspiration to search for another framework. The
framework we offer produces more efficient 3P-EKE
protocol, and, in addition, delivers perceptive clarification
about the existing attacks that do not solve in the previous
framework. Moreover, it allows direct change from a
class of verge private-key encryption to a hybrid
(symmetric & Asymmetric) one without significant
overhead.
**_Index Terms—Verifier–based protocols, Password –_**
based Authentication, Three Party Encrypted Key
Exchange Protocol (3P-EKE), Public-Key Cryptosystem
Based on Logarithmic Approach (PCLA).
I. INTRODUCTION
A vital job of cryptography is to guard the
confidentiality of messages transferred over the
unsecured network. To provide security, messages can be
encrypted by using a key (secret information) so that an
intruder cannot decode the message. However, encoding
the messages may not be only the solution, because an
intruder may take the most active role as the network is
open and reachable. We may need to change the key
according to the session to establish a secure
communication through the unsecured open network.
Consecutively, a password-authenticated two-party
encrypted key exchange (2P-EKE) protocols are used to
exchange a session key based on a low-entropy password.
In this network, each party who wants to communicate
needs to memorize a low-entropy password, which
implies high maintenance of passwords. Due to this
drawback, a password-authenticated three-party
encrypted key exchange (3P-EKE) protocols has its
demand as on date. According to Ding & Hoster [1],
many of such 3P-EKE protocols suffer from any one of
the three types of password guessing attacks. An intruder
can guess the correct password by continuously trying
until he succeeds, which is known as password guessing
attack.
An Ideal password authenticated key exchange
protocol should satisfy the security requirements like,
Mutual Authentication, Resistant to password guessing
attacks, Session Key (SK) security, Resistant to Trivial
Attack, Resistant to Pre-play Attack, Resistant to Replay
Attack, Resistant to Man-in-the-middle Attack, Server
spoofing security, Perfect forward secrecy, backward
secrecy, Known-Key Security, etc.
Based on the low entropy passwords shared between a
user and a server, the password-authenticated key
exchange (PAKE) protocols are classified into two types.
They are:
_A._ _Symmetric model_
As the name implies, symmetric (identical) low
entropy password shared between a Trusted Party (server)
and a user in establishing a secure session key. If a
Trusted Party is compromised, an intruder will get
succeed in performing an attack on the legitimate user.
_B._ _Asymmetric model_
As the name implies, the password distribution is
asymmetric in nature; i.e., a user will share the different
knowledge (a verifier) about the low entropy password
with the Trusted Party in establishing a secure session
key. If a Trusted Party is compromised, the password
table does not reveal the direct information about the
password. In this way, a server spoofing is avoided.
Henceforth, the design of a novel framework which is
smart in establishing a secure session key with less
computational overhead; which prove to be secure against
the attacks like password guessing, a server spoofing and
also provide mutual authentication, Backward Secrecy,
and Forward Secrecy is the need of the hour. This paper
endeavors to propose a novel framework for establishing
a secure session key based on an asymmetric model by
using PCLA keys. PCLA is a new public key
-----
cryptosystem based on the logarithmic approach
proposed by Archana et al. in 2012[2].
Further, the rest of the paper is organized as follows:
Related works is discussed in section II. In section III, we
listed notations used in the proposed protocol. A
framework for the proposed protocol is described in
section IV. Security Analysis of the proposed protocol is
done in section V. Finally, we made concluding remarks
in section VI.
II. RELATED WORK
Diffie-Hellman (1976) [3] key exchange protocol is
suffered from a man-in-the-middle attack due to the lack
of an authentication. To assure a good access control,
many applications require a robust client authentication.
In such scenario, password-authenticated key exchange
(PAKE) protocols have their own identity.
Bellovin and Merritt (1992)[4] have first proposed
password-based authenticated encrypted key exchange
protocol for the two-party network. But, due to the server
compromise (server hacking: e.g., In 2012, more than
million LinkedIn passwords are stolen) this protocol no
longer proved to be secure. Hence, to eliminate such a
problem he proposed an improvement over it known as
Augmented EKE protocol (1993) [5], where a server
instead of storing the actual passwords, it stores the
verifiers of the passwords which prevents from a server
compromise but it does not solve the problem of off-line
dictionary attacks.
Subsequently, Gong et al.(1993)[6] proposed a three
party password-based authenticated key exchange
protocol using a server‘s public key, where the clients are
given a risk to verify and keep the public key safely.
Many improvements proposed by various researchers in
terms of a security and computational efficiency [7, 8, 9,
10].
Abdalla et al. (2005) [11] proposed a ‗provable secure‘
one-time password-based authentication and key
exchange (OPKeyX) technology for grid computing;
where a user changes the password from one session to
another session to eliminate the problem of password
sniffing. Lin et al. (2008)[12] proposed an efficient
verifier-based password-authentication key exchange
protocol by using elliptic curve cryptography.
Unfortunately, Yang et al. (2011) [13] showed the flaws
of Lin et al. protocol and proposed an improvement over
the Efficient verifier-based password-authentication key
exchange protocol via elliptic curves.
A Novel ECC-3PEKE protocol is proposed by Chang
et al. (2004) [14], which proved to be practical, efficient
and secure. However, Yoon et al. (2008) [15] notified an
undetectable online password guessing attack and
proposed an improvement over ECC-3PEKE protocol.
Subsequently, PSRJ protocol has been proposed by
Padmavathy et al. (2009) [16], which is also an
improvement over ECC-3PEKE protocol. They claimed
that the proposed protocol achieves better computational
complexity and also secure against dictionary attacks.
Later Chang et al. (2009) [17] discussed why Yoon
Yoo‘s Protocol is still insecure. R. Padmavathy (2010)
[18] cryptanalyzed the PSRJ protocol and to overcome an
attack she proposed an improvement over the existing
one by using reduced modular exponentiation operations.
Successively, an impersonation attack has been shown on
the ECC-3PEKE protocol by Shirisha Tallapally (2010)
[19]. Next, Archana et al. (2012) [20] showed detectable
online password guessing the attack on PSRJ protocol.
Also, Kulkarni et al.(2007) [21] proposed a novel key
exchange protocol based on verifier-based password
authentication for three parties; where each client instead
of storing the direct password itself it computes a oneway hash function on each password and stores the
corresponding result in a server's password table.
Subsequently, Shaban et al.(2008)[22] proposed an
improvement over the Kulkarni et al.‘s protocol in terms
of computational complexity, by showing the reduced
rounds from 7 to 4 without using symmetric
encryption/decryption. But unfortunately, Archana et al.
(2015) [23] cryptanalyzed the Shaban et al.‘s protocol by
showing the detectable online password guessing attack.
Kulkarni et al.'s protocol are proved as secure against the
dictionary attacks but it is computationally more
expensive than our proposed protocol.
A previous framework presented by Archana et al. ―in
press‖ [24] is more secured against all types of attacks
like password guessing, replay, pre-play, man-in-themiddle attack etc. But unfortunately, this protocol does
not solve the problem of a server compromise. These
proofs help as inspiration to search for another
framework which eliminates the problems, may occur in
the previous framework.
III. NOTATIONS
The list of notations along with their descriptions used
in this paper is given in Table 1. In fact, Ida, Idb, Idtp are
the identities of client-A, client-B, Trusted Party-TP
respectively, which are known publicly.
-----
Table 1. List of Notations
At this 1[st] stage, the clients who want to communicate
IV. FRAMEWORK FOR PROPOSED PROTOCOL each other have to register with Trusted Party in advance.
The procedure for registration is as follows:
This section endeavors to propose a novel verifier
based password-authenticated 3P-EKE protocol using the
**Step 0:** Client-A and Client-B compute‘s the verifiers
PCLA keys. PCLA is a new public key cryptosystem
based on the logarithmic approach proposed by Archana VA=H(Ida, Idtp, Pwda) and VB=H(Idb, Idtp, Pwdb) by
et al. More details about this algorithm (PCLA) are given choosing low random passwords Pwda and Pwdb
respectively. Now client-A and client-B sends the
in the reference paper [2].The proposed protocol has been
divided into three stages. They are: verifiers VA & VB to Trusted Party respectively through a
secure channel.
- Initialization Stage **i.e., Client-A Trusted Party: {VA}, and Client-B**
**Trusted Party: {VB}.**
- Key Agreement Stage
Then Trusted Party stores the verifiers in its password
- Key Computation Stage verifier‘s table. The detail of the initialization stage is
depicted in Fig.1.
_A._ _Initialization Stage_
Fig.1. Initialization Stage
_B._ _Key Agreement Stage_ **i.e., Client-ATrusted Party: {Ida, Idb, Idtp, {EPKeyp**
(Pwda), VA}, EPwda(Ma ra), htp(Pwda Ma), FKatp(Ma)}.
In this 2[nd] stage of the protocol, the actual procedure Similarly, client-B generates two random numbers viz.,
for session key agreement begins. The protocol executes REb, rb ∈RZP, and computes Kbtp=Mbrb mod p where Mb=
as per the following steps of the protocol. g[REb] mod p and sends the credentials
**Step KA1:** Client-A generates two random numbers {Ida,Idb,Idtp,{EPKeypu(Pwdb),VB}, EPwdb(Mbrb),htp(Pwdb
viz., REa, ra ∈ZR, and computes Katp=Mara mod p where Mb), FKbtp(Mb)}to Trusted Party.
Ma=g[REa] mod p and sends the credentials {Ida, Idb, Idtp, **i.e., Client-BTrusted Party: {Ida,Idb,Idtp,{EPKeypu**
{EPKeypu(Pwda),VA},EPwda(Mara),htp(PwdaMa),FKatp (Pwdb),VB},EPwdb(Mbrb),htp(PwdbMb), FKbtp(Mb)}.
(Ma)}to Trusted Party. **Step KA2: Upon receiving the credentials from client-**
-----
A and client-B, Trusted Party decrypts EPKeypu(Pwda) &
EPKeypu(Pwdb) by using its PCLA private key Keypr i.e.,
DPKeypr(EPKeypu(Pwda)) & DPKeypr(EPKeypu(Pwdb)) and
gets the low entropy passwords Pwda, Pwdb respectively.
Now Trusted Party computes H(Ida, Idtp, Pwda) & H(Idb,
Idtp, Pwdb) and retrieves the verifier VA & VB from its
table and checks whether both the numbers are equal. If
not, then it terminates the protocol at the current session.
If yes, then it implies that client-A & client-B is
verified at first level and Trusted Party continues with the
residual procedure of the protocol. Trusted party retrieves
PwdaMa & PwdbMb from htp(PwdaMa) &
htp(PwdbMb) by using trapdoor [25] ‗tp‘ and compute‘s
Ma=(PwdaMa) Pwda [& M]b[=(Pwd]b[][M]b[)][][ Pwd]b
respectively. Now, it gets ra=(Mara)Ma &
rb=(Mbrb)Mb from the credential EPwda(Ma ra) &
EPwdb(Mb rb) by decrypting with low entropy password
Pwda & Pwdb i.e., DPwda(EPwda(Ma ra)) &
DPwdb(EPwdb(Mb rb)) . Next, Trusted Party performs the
second level of verification by calculating FKatp(Ma) &
FKbtp(Mb) after the computation of Katp=Mara mod p &
Kbtp=Mbrb mod p respectively. That is, it compares the
computed value of FKatp(Ma) (or FKbtp(Mb)) with the
received value of FKatp(Ma)(or FKbtp(Mb)). If not identical,
it terminates the protocol at the current session.
If both are identical then verification of client-A &
client-B is passed and it continues with the residual
procedure of the protocol. Now, a Trusted Party chooses
a random exponent REtp∈RZP to compute MbREtp mod p
and MaREtp mod p and encrypts these values with its
PCLA private key. Then Trusted Party sends these
credentials to client-A and client-B simultaneously.
**i.e., Trusted PartyClient-A:{EPKeypr(MbREtp** mod
p)}, and Trusted Party Client-B:{EPKeypr(MaREtp mod
p)}.
The detail of key agreement stage is depicted in Fig.2.
Fig.2. Key Agreement Stage
**Step KC2:** Upon receiving the incoming credentials
_C._ _Key Computation Stage_
FSK(Idb, SK) and FSK(Ida, SK) from client-A and client-B
In order to compute a secure session key, this stage respectively, they verify each other and can confirm that
accomplishes to begin with the verification of the Trusted the mutual session key is SK=( MbREtp) REa (mod p) =
Party in a smart way. ( MaREtp)REb (mod p).
**Step KC1: Upon receiving the credentials from a** The detail of key computation stage is illustrated in
Trusted Party, client-A decrypts EPKeypr(MbREtp mod p) by Fig.3.
using the PCLA public key of Trusted Party i.e.,
DPKeypu(EPKeypr(MbREtp mod p)) to get MbREtp mod p. In _D._ _Password Change Mechanism_
this way, client-A authenticates the Trusted Party. If any one of the client (say Client-A) suspects a ‗leak
Similarly, client-B also authenticates the Trusted Party of information‘, then it invokes a password change
in the same way. mechanism of our proposed protocol, which is helpful in
Now, client-A computes a mutual session key SK= providing backward secrecy. The steps in this mechanism
(MbREtp) REa mod p= ((gREb) REtp) REa mod p & FSK(Ida,SK) are as follows:
and sends it to client-B. Similarly, client-B computes a
mutual session key SK= (MaREtp) REb mod p= ((gREa) REtp) **Step PC1: Client-A preserves the session key SK from**
REb mod p & FSK(Idb,SK) and sends it to client-A. the previous session.
**i.e., Client-A Client-B: {FSK(Ida,SK)}, and Client-** **Step PC2: Client-A, first encrypts the session key SK**
**B Client-A: {FSK(Idb,SK)}.** by using PCLA public key Keypu of Trusted Party. Next,
-----
it computes H(Ida, Idtp, Pwda) H(Ida, Idtp, NewPwda)
SK & htp(H(Ida, Idtp, NewPwda) and finally sends the
following credentials to Trusted Party to reset the new
password verifier.
**i.e., Client-A Trusted Party: {Ida, Idtp, EPKeypu(SK),**
H(Ida, Idtp, Pwda) H(Ida, Idtp, NewPwda) SK, htp(H(Ida,
Idtp, NewPwda)}
**Step PC3: Upon receiving the credentials from client-**
A, Trusted Party compute‘s SK by decrypting EPKeypu(SK)
with the PCLA private key i.e., DPKeypr (EPKeypu(SK)) and
also retrieves the old verifier from the password table and
computes the new password verifier as follows: H(Ida, Idtp,
NewPwda)= (H(Ida, Idtp, Pwda) H(Ida, Idtp, NewPwda)
SK) H(Ida, Idtp, Pwda) SK.
Now, Trusted party check for validation by extracting
the _new verifier from the received credential htp(H(Ida,_
Idtp, NewPwda) by using a trapdoor ‗tp‘; if computed
H(Ida, Idtp, NewPwda) is equal to the received one,
Trusted Party updates the password table accordingly and
sends ‗Accepted‘ message to the client-A. Otherwise,
Trusted Party rejects the request by sending the ‗Denied‘
message to the client.
**i.e., Trusted Party: Accepted/Denied.**
Fig.3. Key Computation Stage
Fig.4. Password Change Mechanism
-----
V. SECURITY ANALYSIS
The following security requirements are satisfied by
the proposed protocol; which proves that the proposed
protocol is not only efficient but also secure.
_A._ _Resistant to an off-line dictionary attack_
An attacker Eve-E may try to mount off-line password
guessing attack to guess the password. She intercepts {Ida,
Idb, Idtp, {EPKeypu(Pwda),VA}, EPwda(Ma ra), htp(Pwda
Ma), FKatp(Ma)} and may guess a password to extract
(Mara), but it is impossible for her to get Ma until
trapdoor ‗tp‘ is known, which is known only to Trusted
Party. This implies that she cannot verify the hash value
FKatp(Ma) which ascertains an offline password guessing
attack on the proposed protocol is impossible.
Hence, the proposed protocol is resistant to off-line
dictionary attack.
_B._ _Resistant to server spoofing attack_
Assume an intruder _Eve-E succeeds in getting the_
password table of Trusted Party. Since only the verifier V
of clients is stored in password table, Eve-E cannot mimic
the client and compute SK.
Hence, the proposed protocol is resistant to a server
spoofing attack.
_C._ _Provides the mutual authentication_
The proposed protocol promotes the mutual
authentication and realizes the session key security to a
great extent. The following are the scenarios where the
mutual authentication can be proved.
- **_First Scenario: Client-A and Client-B use the_**
public key Keypu of Trusted Party to hide the
corresponding passwords. Only Trusted Party
knows the private key Keypr to decrypt it. Hence, for
an intruder Eve-E, it is not possible to get the
passwords of client-A & client-B.
- **_Second Scenario: Client-A and Client-B use the_**
trapdoor ‗tp‘ to hide the random exponents REa in
Ma & Pwda and REb in Mb & Pwdb. Since only
Trusted Party knows the trapdoor ‗tp‘ and
passwords Pwda & Pwdb he can very well
authenticate Client-A and Client-B after receiving
the messages sent in step KA1 of the protocol.
- **_Third_** **_Scenario:_** Trusted Party sends
{EPKeypr(MbREtp mod p)} to client-A &
{EPKeypr(MaREtp mod p)} to client-B in _step KA2 of_
the protocol. This message can be used to
authenticate Trusted Party.
- **_Fourth Scenario: Client-A and Client-B derive a_**
key from MbREtp and MaREtp respectively, as
mentioned in _step KC1 of the protocol. With the_
help of FSK(Idb, SK) & FSK(Ida, SK) both client-A
and client-B can authenticate each other respectively
as mentioned in step KC2 of the protocol.
Hence, the mutual authentication is provided by the
proposed protocol.
_D._ _Provides backward secrecy_
The proposed protocol can provide backward secrecy,
where compromise of Pwda will not lead to the
compromise of NewPwda. Assume, a client-A suspects a
‗leak of information‘ to Eve-E, then immediately client-A
request to Trusted Party to change its password from
Pwda to NewPwda. Let us assume, subsequently _Eve-E_
intercepted the password change request, i.e., {Ida, Idtp,
EPKeypu(SK), H(Ida, Idtp, Pwda) H(Ida, Idtp, NewPwda)
SK, htp(H(Ida, Idtp, NewPwda)} sent to Trusted Party by
client-A. However, in this process Eve-E cannot compute
SK by using Pwda, hence, he cannot compute NewPwda
from the intercepted message {Ida, Idtp, EPKeypu(SK),
H(Ida, Idtp, Pwda) H(Ida, Idtp, NewPwda) SK, htp(H(Ida,
Idtp, NewPwda)}.
Hence, the backward secrecy is provided by the
proposed protocol.
_E._ _Provides the forward secrecy_
The session key is computed as follows: SK=
(MbREtp)REa (mod p)=(MaREtp)REb (mod p). If the Eve-E gets
{EPKeypr(MbREtp mod p)} or {EPKeypr(MaREtp mod p)}, then
in order to obtain the session key, she should know the
public key of Trusted Party and REb or REa. The session
keys generated in different sessions are independent since
REa and REb are randomly chosen by client-A and clientB respectively. This indicates that _Eve-E cannot obtain_
previous session keys even if she obtains the session key
used in this run.
Hence, the forward secrecy is provided by the
proposed protocol.
VI. CONCLUSION
In this paper, we proposed a novel verifier-based
password authenticated 3P-EKE protocol using PCLA
keys, which provides perceptive justification about the
existing attacks that do not solve in the previous
framework. That is, our proposed protocol is proved to be
secure against offline dictionary attacks and server
spoofing attack. Further, we have also proved that our
protocol provides mutual authentication, backward
secrecy and also forward secrecy.
REFERENCES
[1] Y. Ding and P. Horster. ―Undetectable online password
guessing attacks,‖ ACM Operating Systems Review
vol.29, pp.77-86, 1995.
[2] Archana Raghuvamshi, P.Premchand and P.Venkateswara
Rao. ―PCLA: A New Public-key Cryptosystem Based on
Logarithmic Approach‖, International Journal of
Computer Science Issues(IJCSI), vol.9,no.1, pp.355-359,
2012.
[3] W. Diffie and M. E. Hellman. ―New directions in
cryptography‖, IEEE Transactions on Information Theory,
vol.22, no.6, pp.644–654, 1976.
[4] S. M. Bellovin and M. Merritt, ―Encrypted key exchange:
Password-based protocols secure against dictionary
-----
attacks‖, IEEE Symposium on Security and Privacy, IEEE
Computer Society Press, pp.72–84 May 1992.
[5] S. M. Bellovin and M. Merritt, ―Augmented encrypted
key exchange: A password-based protocol secure against
dictionary attacks and password file compromise‖, ACM
CCS, ACM Press vol.93, pp.244–250, November 1993.
[6] L. Gong, M. Lomas, R. Needham, and J. Saltzer,
―Protecting poorly chosen secrets from guessing attacks‖,
_IEEE Journal on Selected Areas in Communications,_
vol.11,no.5,pp. 648-656, 1993.
[7] W.M. Li, and Q.Y. Wen, ―Efficient verifier-based
password-authentication key exchange protocol via
elliptic curves‖, _Proceedings of 2008 International_
_Conference_ _on_ _Computer_ _Science_ _and_ _Software_
_Engineering, pp. 1003-1006, 2008._
[8] E.J. Yoon, and K.Y. Yoo, ―Robust User Password Change
Scheme based on the Elliptic Curve Cryptosystem‖,
_Fundamenta Informaticae, pp 483-492, 2008._
[9] Zeng, Yong and Ma, Jianfeng, ―An improvement on a
password authentication scheme over insecure networks‖
_Journal of Computational Information Systems, vol.5,_
no.4, pp.1331-1336, 2009.
[10] Chunling Liu, Yufeng Wanga and Qinxi Bai, ―A New
Three-party Key Exchange Protocol Based on DiffieHellman,‖ I.J. Wireless and Microwave Technologies, vol.
1, no.4, pp. 65-69, 2011.
[11] M. Abdalla, O. Chevassut, and D. Pointcheval. ―One-time
verifier-based encrypted key exchange‖, PKC LNCS,
Springer, vol. 3386, pp.47–64, January 2005.
[12] W.M. Lin, and Q.Y. Wen, ―Efficient verifier-based
password-authentication key exchange protocol via
elliptic curves‖, _Proceedings of 2008 International_
_Conference_ _on_ _Computer_ _Science_ _and_ _Software_
_Engineering, pp.1003-1006, 2008._
[13] Junhan YANG and Tianjie CAO, “A Verifier-based
Password-Authenticated Key Exchange Protocol via
Elliptic Curves‖, Journal of Computational Information
Systems, Binary Information Press, pp.548-553, 2011.
[14] Chin-Chen Chang and Ya-fen Chang, ―A novel three
party encrypted key exchange protocol‖, Elsevier,
Computer Standards & Interfaces, vol.26 pp.471 – 476,
2004.
[15] [Eun-Jun Yoon, and Kee-Young Yoo, ―Improving the](https://www.researchgate.net/researcher/33971269_Eun-Jun_Yoon)
novel three-party encrypted key exchange protocol‖,
Elsevier, _Computer Standards and Interfaces, vol. 30,_
pp.309-314, 2008.
[16] R.Padmavathy, Tallapally Shirisha, M.Rajkumar, and
Jayadev Gyani, ―Improved analysis on Chang and Chang
Password Key Exchange Protocol‖, IEEE International
Conference on Advances in Computing, Control, and
Telecommunication Technologies, pp.781-783, 2009.
[17] Ya-Fen Chang, Wei-Cheng Shiao, and Chung-Yi Lin,
―Comments on Yoon and Yoo‘s Three-party Encrypted
Key Exchange Protocol‖, International Conference on
Advanced Information Technologies (AIT), 2009.
[18] R. Padmavathy, ―Improved Three Party Eke Protocol‖,
Information Technology and Control, Vol.39, No.3,
pp.220-226, 2010.
[19] Shirisha Tallapally, ―Impersonation Attack on EKE
Protocol‖, International Journal of Network Security & Its
Applications (IJNSA), vol.2, no. 2, pp. 114-121, 2010.
[20] Archana Raghuvamshi, P.Venkateshwara Rao, and
Prof.P.Premchand, ―Cryptanalysis of Authenticated Key
Exchange 3P-EKE Protocol and its Enhancement‖, IEEEInternational Conference on Advances in Engineering,
Science and Management (ICAESM -2012), pp.659-666,
March 30, 31, 2012.
[21] S. Kulkarni, D. Jena, and S.K. Jena, "A Novel Secure Key
Agreement Protocol using Trusted Third Party",
Computer Science and Security Journals (IJCSS), vol.1,
no.1, pp. 11 – 18, 2007.
[22] Dina Nabil Shaban, Maged H. Ibrahim, and Zaki
B.Nossair, **“Enhanced** Verifier-Based Password
Authenticated Key Agreement Protocol For ThreeParties‖, Journal of Engineering Sciences, vol. 36, no. 6,
pp.1513- 1522, 2008.
[23] Archana Raghuvamshi and Premchand Parvataneni.
―Cryptanalysis of Verifier-Based Password-Authenticated
Key Agreement Protocol for Three Parties‖, Research
Journal of Recent Sciences. Vol. 4, pp. 5-8, Feb 2015.
[24] Archana Raghuvamshi and Premchand Pavataneni,
―Design of a Robust, Computation-Efficient and Secure
3P-EKE Protocol using Analogous Message
Transmission‖, International Journal of Computer
Network and Information Security (IJCNIS), In Press.
[25] Y. Gertner, T. Malkin, and O. Reingold, ―On the
impossibility of basing trapdoor functions on trapdoor
predicates‖, Proceedings of the 42nd IEEE Symposium on
foundations of Computer Science, Las Vegas, Nevada,,
pp. 126 – 135, October 2001.
**Authors’ Profiles**
**Archana Raghuvamshi is presently**
working as an Assistant Professor in Dept.
of CSE, UCOE, Adikavi Nannaya
University, Rajahmundry. She is having
13+ year of teaching experience.
She received her Bachelor‘s Degree
BSc (M.S.Cs), Master‘s Degrees M.C.A
and M.Tech(CSE) from Osmania
University, Hyderabad. She did course
work in ADS and WMN in IITM (Indian Institute of
Technology, Madras). She is perusing Ph.D. (CSE) in JNTUK,
Kakinada. She published four research papers in IEEE Digital
library and another six research papers in various peer reviewed
International Journals. Her research interest includes
Cryptography and Information Security, Security in Cloud
Computing etc.
Ms. Archana Raghuvamshi is a,
1. Professional Member of ACM
2. Member of Professional Body IAENG
3. Member of IACSIT
4. Associate Member of theIRED
**Prof.** **Premchand** **Parvataneni** is
presently working as a professor in
Department of Computer Science and
Engineering at University College of
Engineering, Osmania University,
Hyderabad (Telangana). He received his
Bachelor‘s Degree B.Sc (Engg.) from
RIT, Jamshedpur. He received his
Master‘s M.E (CE) from AU (Andhra
University), Visakhapatnam. He received his Ph.D.(CSSE) from
AU. He has published more than 50 publications in various
International Journals and Conference proceedings. His research
Interest includes Cryptography and Network Security, Image
Processing, Software Engineering etc.
Prof.Premchand is having 40+ years of teaching experience
-----
in various Universities. He was as a Director in AICTE, New
Delhi. And also, he has been held for the various positions like
Head, Chairman of BOS, Additional Controller of Examinations
in the Professional wing, Osmania University, Hyderabad.
**How to cite this paper:** Archana Raghuvamshi, Premchand Parvataneni,"Verifier-based Password Authenticated 3PEKE Protocol using PCLA keys", International Journal of Computer Network and Information Security(IJCNIS), Vol.8,
No.6, pp.59-66, 2016.DOI: 10.5815/ijcnis.2016.06.07
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.5815/IJCNIS.2016.06.07?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.5815/IJCNIS.2016.06.07, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "GOLD",
"url": "http://www.mecs-press.org/ijcnis/ijcnis-v8-n6/IJCNIS-V8-N6-7.pdf"
}
| 2,016
|
[] | true
| 2016-06-08T00:00:00
|
[
{
"paperId": "04e9a7d9625b38452aad458dfd9bf6e6e8358dfe",
"title": "Design of a Robust, Computation-Efficient and Secure 3P-EKE Protocol using Analogous Message Transmission"
},
{
"paperId": "175e547be9705b621b06c3a231ba254aad84d599",
"title": "PCLA: A new public-key cryptosystem based on logarithmic approach"
},
{
"paperId": "9b565623e6662ec68983d16e22c87a3399cd90ad",
"title": "Cryptanalysis of authenticated key exchange 3P-EKE protocol and its enhancement"
},
{
"paperId": "4ad98e966fe095e49597599669a8c093a16dc656",
"title": "A New Three-party Key Exchange Protocol Based on Diffie-Hellman"
},
{
"paperId": "ed98c3b046b5895ecc7e8d58c608e08ab01caf30",
"title": "Verifier-based password authenticated key exchange protocol via elliptic curve"
},
{
"paperId": "5d4d6ff7a52c3c3f97c5ac26e045b26533cd684b",
"title": "IMPERSONATION ATTACK ON EKE PROTOCOL"
},
{
"paperId": "53a44301e06bdfd9e17a510b183be4ffabc361cf",
"title": "Improved Analysis on Chang and Chang Password Key Exchange Protocol"
},
{
"paperId": "92ea1592743d40c9819c9880d9b8c6c16cd967ed",
"title": "Efficient Verifier-Based Password-Authentication Key Exchange Protocol via Elliptic Curves"
},
{
"paperId": "a568d98af58850a26ef1def71b53e981e23cf5fb",
"title": "ENHANCED VERIFIER-BASED PASSWORD AUTHENTICATED KEY AGREEMENT PROTOCOL FOR THREE-PARTIES"
},
{
"paperId": "df1e6550400c3d5f38a340618309a253569a5d0a",
"title": "Improving the novel three-party encrypted key exchange protocol"
},
{
"paperId": "2b3bf2ce47d4baf857fea8104205aaeec8d8d57b",
"title": "Robust User Password Change Scheme based on the Elliptic Curve Cryptosystem"
},
{
"paperId": "7a8985c16a0551c464bbfcb46fbc6d2d48377137",
"title": "A password authentication scheme over insecure networks"
},
{
"paperId": "16dde8415501f1a4f5d63b292343088836ebe995",
"title": "One-Time Verifier-Based Encrypted Key Exchange"
},
{
"paperId": "686a5383620fd2be566ad06893d6e5f12a286fbc",
"title": "A novel three-party encrypted key exchange protocol"
},
{
"paperId": "97fe9e35e87a45811e7e0e9b05e7822ae575148c",
"title": "Undetectable on-line password guessing attacks"
},
{
"paperId": "8e39cad74131ad5c1f09d44ef3b2b65a3ae29e35",
"title": "Augmented encrypted key exchange: a password-based protocol secure against dictionary attacks and password file compromise"
},
{
"paperId": "6b479047219fc565a478efbe95572806cd03a7a1",
"title": "Protecting Poorly Chosen Secrets from Guessing Attacks"
},
{
"paperId": "0c0fbe79e49c4859f4d63052d27074049733e092",
"title": "Encrypted key exchange: password-based protocols secure against dictionary attacks"
},
{
"paperId": "ba624ccbb66c93f57a811695ef377419484243e0",
"title": "New Directions in Cryptography"
},
{
"paperId": "8fa5b000d25aed21a3f2288c6571b827a644b626",
"title": "IMPROVED THREE PARTY EKE PROTOCOL"
},
{
"paperId": null,
"title": "A Novel Secure Key Agreement Protocol using Trusted Third Party"
},
{
"paperId": "f346674cb9fc26d243a0b63eb43385e9efaaf860",
"title": "Authors' Profiles"
},
{
"paperId": null,
"title": "She received her Bachelor‘s Degree BSc (M.S.Cs), Master‘s Degrees M.C.A and M.Tech(CSE) from Osmania University, Hyderabad"
}
] | 8,421
|
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/0254fed1ad7983ac9d834822d66a553845e0f7de
|
[
"Computer Science"
] | 0.919658
|
Privacy-Aware and Highly-Available OSN Profiles
|
0254fed1ad7983ac9d834822d66a553845e0f7de
|
2010 19th IEEE International Workshops on Enabling Technologies: Infrastructures for Collaborative Enterprises
|
[
{
"authorId": "1796529",
"name": "Rammohan Narendula"
},
{
"authorId": "1766169",
"name": "Thanasis G. Papaioannou"
},
{
"authorId": "1751802",
"name": "K. Aberer"
}
] |
{
"alternate_issns": null,
"alternate_names": null,
"alternate_urls": null,
"id": null,
"issn": null,
"name": null,
"type": null,
"url": null
}
| null |
# Privacy-aware and highly-available OSN profiles
Rammohan Narendula, Thanasis G. Papaioannou, and Karl Aberer
School of Computer and Communication Sciences, EPFL, Switzerland
Email: firstname.lastname@epfl.ch
**_Abstract—The explosive growth of online social networks_**
**(OSNs) and their wide popularity suggest the impact of OSNs on**
**today’s Internet. At the same time, concentration of vast amount**
**of personal information within a single administrative domain**
**causes critical privacy concerns. As a result, privacy-conscious**
**users feel dis-empowered with today’s OSNs. In this paper, we**
**report on an on-going research work and introduce a privacy-**
**aware decentralized OSN called porkut. Our system exploits trust**
**relationships in the social network for decentralized storage of**
**OSN profiles and their content. By taking users’ geographical**
**locations and online time statistics into account, it also addresses**
**availability and storage performance issues. We finally advocate**
**indexing of social network content and present an approach for**
**indexing in a privacy-preserving manner.**
**_Keywords-online social network; privacy-preserving index; con-_**
**nected dominating set; trust**
I. INTRODUCTION
Online social networks (e.g. Facebook.com, Orkut.com)
have recently seen an explosive growth. Facebook received 130
million visitors in a single month in 2008 [1] and currently
has more than 200 million users. As a result, these OSNs
have become store houses of unprecedented amount of data in
the form of messages, photos, links, and personal information.
Facebook has grown to be the the world’s largest photo sharing
service surpassing even dedicated photo-sharing online applications (e.g. Flickr). It is also the largest instant messaging
service on the web. Researchers argue that future Internet will
be very much influenced by social networks regarding the
location of content and knowledge, and the user interactions
[2]. However, most of the current social networks operate on
infrastructure administered by a single authority (big-brother),
such as Google, Facebook, etc. These organizations perform
mining of personal data hosted inside users profiles and exploit
it for targeted advertisements, in order to be compensated for
their huge investments in infrastructure. During sign-up time,
users consciously or unconsciously permit the organizations to
share their personal information with third-parties in whatever
form the organizations choose to [3]. In addition, the leakage
of personal information from OSNs can be associated with the
user activity on non-OSN sites as well [4].
However, the exponential growth of the OSNs suggests that
users are ready to trade privacy over utility of the services
offered. As a result, there is almost negligible motivation for
the OSN operators to address privacy concerns of the users.
In order to address privacy concerns of OSN users, research
community has resorted to the P2P paradigm for OSN content
management. Replacing the big-brother with a community of
users, enables OSN users to have complete control on their
profile content.
In this paper, we present an initial design of such a system,
referred to as porkut, where users organize a social network
over a P2P overlay with privacy-preserving data access. We
briefly outline the system architecture and mainly focus on
the distributed storage layer. Specifically, we propose a decentralized mechanism for users to manage their own online
social network on top of resources collectively contributed by
themselves. Such a design is motivated with several goals
in mind: a) It eliminates the requirement for a single bigbrother who can exploit the users’ profile data for his own
interest without users’ consent. b) It preserves the privacy
of individuals social profile content, as they have complete
control on who can access which parts of the content. c)
It exploits the trust relationships among users in the social
network to improve the content availability and the storage
performance. Both issues are non-trivial in a P2P setting.
Three approaches with different goals for improving storage
performance are introduced, while maintaining high content
availability. A user’s profile content is hosted only on a set
of self-defined trusted nodes that enforce access control on
the content. This set of trusted nodes is selected intuitively
keeping the availability and performance goals in mind. Other
issues such as the structure of the profile content, the format of
the access control policies, trusted identity management, and
other data integrity issues are beyond our scope.
In addition, the system constructs a privacy-preserving
index of the social network content that enables privacyaware searching. We argue that such an index enables content
discovery among friends in OSNs and helps system users
discover new friends (based on content, such as common
interests etc.) within the OSN application and establish new
social connections. This is an add-on feature over existing
OSNs like Facebook that do not allow content-based search.
Such an index is hosted over the P2P overlay in a distributed
hash table (DHT). Users can specify their privacy objectives
during content publishing and thus content existence and ownership are only revealed according to their preferences. This
index could be used to serve advertisements on searches and
distribute the revenues to the users according to their published
content. This way, users can benefit from their content without
compromising their privacy. In contrast, the current OSN
applications exploit users’ content for own monetary gains.
The rest of the paper is organized as follows. In Section II,
a brief description of the system is provided. The storage layer
is discussed in Section III. The privacy preserving indexing is
-----
described in Section IV. In Section V, we discuss the related
work and, finally in Section VI, we conclude this paper and
outline our future work.
II. SYSTEM OVERVIEW
As mentioned earlier, the porkut system exploits the trust
relationships among friends and social network connections to
improve the availability and search performance of the system.
We assume that a user of porkut runs the client on his office
or personal laptop/computer. Hence, for the rest of the paper,
we use the terms user and node interchangeably.
A user u’s profile content is hosted only on a set of selfdefined trusted nodes, which enforce access control on the
content on behalf of the user. This set of trusted nodes for a
user is referred to as his trusted proxy set (TPS). The TPS
members for a user are properly selected with respect to the
availability and performance goals. We observe that every
user in an OSN has friends scattered over a limited set of
geographical locations (e.g. his home town, working location,
home country, location of previous institute etc.). Moreover,
we observe that each user’s online timings are predictable
to a large extent (e.g. his office hours, completely offline
on weekends). Exploiting these facts, we populate this set
of trusted nodes in such a way that, at any given time, one
node in this set is online to satisfy the profile access requests,
while at the same time, the content is located at a node falling
within a geographical neighborhood away from the user that
frequently asks for it. The computation of the set TPS based
on a user’s social graph is explained in next section. Each user
u is identified by a unique identifier denoted by UIdu. Note
that a TPS is a set of UIds.
The porkut system employs a distributed hash table (DHT)
hosted at the resources contributed by the users. This DHT
is used for storing the privacy-preserving index of the profile
content and other meta information, e.g. the current IP address
of a user. A user u and his TPS mapping is stored in the DHT
in the form of (key,value) pair with key being the UIdu and
_value being the members of the TPS. Using cryptographic_
signatures, it should be trivial to test the authenticity of such
an entry in the DHT. This user-to-TPS mapping in the DHT is
useful for contacting the nodes where the profile of a particular
user is stored.
We assume that, with a reasonable replication factor, one
can ensure that the data items stored inside DHT are highly
available in spite of node churn. As a trusted storage is not
required by the system design, such a DHT could be hosted
at a highly available cloud storage or in publicly available
OpenDHT-like services [5]. The porkut storage architecture is
illustrated in Figure 1. Therein, the user u1 has 5 friends in
the OSN, namely u2 to u6. The set TPS = fu1; u2; u4g is
shown in the figure and a mapping between u1 and TPS[u1]
is inserted into the DHT. The user social graph is represented
as online time graph, which is explained in the next section.
u3
u5
u6
u1
(u1,TPS[u1]) TPS[u1]={u1,u2,u4}
DHT
Fig. 1. The porkut storage layer
III. STORAGE LAYER
In this section, we discuss the storage mechanism of the
_porkut system, and mainly address the construction of the set_
TPS(u) for a user u from his social graph.
The social network graph is denoted as G(U; R), where U
is the set of users represented by the vertices in the graph and
R is the set of friendship relations represented by edges. For
example, an edge between two vertices u1 and u2 models the
fact that users u1; u2 are friends. We assume that friendship
relationships are symmetric. This is the default assumption in
current OSN applications, e.g. Orkut, Facebook. We use the
notation NG(u) to represent the set of neighbors (i.e. friends
on the OSN) of user u in the social graph G, and NG[u] to
represent NG(u) [ fug.
We assume that each user u in the social network is
characterized with two parameters: his geographical location
and online time period. For instance, the location can be set to
the country/city where the user is currently located. We exploit
location information of friends of a user, in order to place data
as close as possible to the nodes that most frequently access
the data for getting profile updates etc. Therefore, data is stored
on nodes falling within a certain geographical proximity from
its most-frequent access points.
This is quantified by the metric access cost Cu[u]1[2] [between]
two geographical locations/users/nodes u1 and u2, which is
defined as the cost of the communication link between them
(i.e. unit cost for transferring data in between these two nodes).
This could be measured, for example, in terms of RTT between
these two nodes.
Online time period represents the usual time that the user is
online in the social network. This is the time window in which
the user contributes his resources (i.e. bandwidth, storage, and
processing power) for the social network operation. The node
can only reply to the data access requests (for the data it hosts)
that are generated during this time window. Beyond this time
window frame, the user is offline. We denote the location and
online time period parameters for a user u as Lu and OTu
respectively. Given two users u1 and u2’s locations and online
time settings, we argue that they can contact each other and
thus exchange data if and only if their online time intervals
overlap, which we represent by the condition that OTu1 \
OTu2 6= ;.
_A. Trusted Proxy Set_
Each user u selects some of the neighbors in his social
network as trusted nodes. The user trusts these nodes both for
-----
storing his profile content and for enforcing access control
on the access requests. We believe that storing content in
plain text and leveraging mutual trust relationships for access
control enforcement simplifies the system to a great extent.
This way of exploiting trust relationships for access control
was first introduced by the authors in [6] and employed for
the social network case in [1]. We assume that users mutually
cooperate for hosting content and delegating access control
with some social contracts. The intuition is that users do not
breach the delegation responsibilities because of social pressure and monitoring. This is left for future study. Alternative
solutions, which employ encryption mechanisms for access
control and content storage [7], not only involve complicated
key management issues, but also, they are highly inefficient in
terms of storage overhead, as the same data item may need to
be encrypted multiple times for different users with different
access rights.
Let T (u) NG(u) be the set of trusted users/nodes for user
u based on his social relationships. T [u] also includes the user
u himself in the set of trusted nodes. The user selects a subset
of these trusted users for hosting his content. We call this set
as trusted proxy set (TPS) (TPS(u) T (u)). The content of
user u is stored on the members of the set TPS(u) and itself,
which is denoted as
TPS[u] = TPS(u) [ fUIdug
We propose the following criteria to select the proper set
of members into TPS from the set of all the trusted users
of a user: i) low access and consistency costs and ii) high
_data availability. To this end, the number of replicas should_
consider access and update costs and replica placement should
consider users online time settings.
Next, we describe three approaches for the computation of
the set TPS[u] that satisfy high availability but have different
cost minimization objectives. In every approach, once TPS[u]
is computed, for each friend/user in the social neighborhood
of user u (i.e., 8v 2 NG(u)), a mount point is configured
(represented by Mv) for accessing u’s profile. In other words,
for a certain friend of u, u’s profile is said to be mounted at
a certain node. Note that, by definition, the mount point is
available at some point in time during the friend’s online time
frame so that he can access u’s profile. However, a singlemount-point-per-user technique allows to access the profile
replica only when that mount point is online. To increase the
availability, we can use all the nodes in TPS[u] as mount
points. In this case, Mv would be the primary mount point
and the remaining would be the secondary ones. In the rest
of the discussion, we assume that content accesses are being
done from the primary mount point.
Given the above, the purpose of the following algorithms is
to compute a storage configuration for user u, which is given
by:
the set TPS[u], and
8v 2 NG(u), the mount point Mv, where Mv 2 TPS[u].
_B. Computing the storage configuration_
Computing the storage configuration for a user u involves
two steps:
i) Constructing the online time graph.
ii) Storage configuration computation from this graph based
on some criterion.
For simplicity, we assume that geographical locations are
considered at the granularity of country, assuming an OSN user
has friends scattered over several countries. First, we construct
the online time graph (denoted by OGu) for user u. This graph
will be used to compute TPS(u).
_Definition 1: Online time graph: for a user u (denoted by_
OGu) is defined as (NG[u]; E) where NG[u] is the set of
vertices and E is the set of edges, such that
8v1; v2 2 NG[u], 9 an edge(v1; v2) 2 E iff
(v1 2 T [u] _ v2 2 T [u]) ^ (OTv1 \ OTv2 6= ;)
Next, we specify the following two conditions on the graph
OGu, which are necessary and sufficient in order to compute
a valid storage configuration.
1) OGu must be connected. Only then, every user in the
set NG[u] can access u’s content.
2) The sub-graph induced by the set T [u] i.e., the graph
OGu[T [u]] must also be connected, in order to allow content synchronization across TPS members pass
through only trusted nodes[1].
We suppose that each user constructs OGu offline locally from
the set of friendship relations that he has in the social network
and their online time (OT ) specifications. The construction of
OGu is explained with the following example. Assume a user
u1 with neighbors in the OSN u2 to u7 and their locations
set as follows: Lu1 is Switzerland, Lu2 and Lu3 are India,
and finally the rest are US-West. Assume OT set to 8am to
5pm local time for all users. Let T [u1] = fu1; u2; u4; u6g.
The resulting OGu1 is shown in Figure 2.
Note that OGu[T [u]] is expected to be connected for a
reasonable number of trusted friends with overlapping online
times (given 120 friends per user in Facebook and 100 in
Orkut on average [2]). Otherwise, another node v 2 OGu, yet
v =2 T [u], has to be employed in the TPS construction as
well. However, profile data stored at v has to be encrypted by
a key shared by the T [u] members. This approach would be
particularly useful in the bootstrap phase of the social network.
In the next subsections, we describe three algorithms with
different cost minimization objectives for TPS generation and
1However, as long as the first condition is met, nodes from the set T [u] can
be removed one by one until the resulting induced graph becomes connected.
u3
u5
u7
u4
u2
u6
u1
T[u1]={u1,u2,u4,u6}
Fig. 2. The graph OGu1
-----
_3) Minimize storage cost: This approach quantifies the stor-_
_age cost of a given storage configuration (x = (M; TPS[u]))_
and, by exploring the entire solution space, picks the storage
configuration with the minimum effective cost. The storage
cost is measured in terms of the total cost incurred for
accessing and updating the profile content by a user’s friends
in addition to that of replica synchronization among all TPS
members. We do not consider the access cost incurred by
non-friend users, even though the system allows such users
to access the profile content on case-by-case basis based on
the access control settings.
Let n[v]a [be the number of times a user][ v][ accesses a user]
u’s profile content with each access involving s[v]a [units of]
data access on average. n[v]u [and][ s]u[v] [represent number of]
updates and update sizes respectively. Note that this update
is performed on Mv, which must be then pushed to the other
members of the TPS as well. We assume that these parameters
are approximated from the statistics collected over a certain
period. To this end, the user u selects the configuration x that
minimizes its storage cost, i.e.
h
argx [min] v2N�G(u) n[v]a [�] [s]a[v] [�] [C]M[v] v [+][ n]u[v] [�] [s]u[v] [�] [C]M[v] v
+�v02T P S[u]�fMvg n[v]u [�] [s]u[v] [�] [C]M[v][0]v i
We refrain from further discussion of this approach for brevity
reasons.
_C. Handling updates in social graph and TPS_
As social relations evolve, there will be updates in a user’s
social graph. Moreover, breach of trust or of the social contract
to host and enforce access control on behalf of others, may
result to updates in the set TPS. Once a node v is removed
from a user u’s TPS, it is no longer contacted for u’s content.
All users in NG(u) for which v is the mount point are
informed of this change. Such nodes are mapped to a new
temporary mount point (say the node u itself), until one of
the three aforementioned algorithms are run to assign them
new mount points. We assume the user periodically invokes
TPS computation process to accommodate the updates made
on OG graph because of updates in the set T (u) or updates
in friendship relationships.
Since revocations can happen from the set TPS, users must
choose TPS members carefully. Such revocation can happen
either because one of the three aforementioned algorithms
excludes an existing member from the set TPS, or a breach in
the social contract is noticed. However, we believe that mutual
social contracts (i.e. reciprocative hosting of data between
users) restrict users from maliciously exploiting their hosted
data after their removal from the TPS. Handling additions to
the set TPS is simple: user u copies the replica of the profile
to this new member, which there on, serves access requests.
When a new social relationship is made by user u, we assign
as default mount point for the new member the node u itself,
or another TPS node that has an overlapping online time
interval. Later, the new friend could be assigned a different
u3
u5
u7
u4
u2
u6
u1
TPS[u1]={u1,u2,u6}
Fig. 3. _MAC approach_
u3
u5
u7
u4
u2
u6
u1
TPS[u1]={u1}
Fig. 4. _MNR approach_
user-mount point mappings. If two TPS members are not
directly connected in OGu, synchronization has to happen
through another node v 2 T [u]. In this case, a profile replica is
stored at node v as well; however, still v is not considered as a
member of TPS, as it is not a mount point for any neighbor.
_1) Minimize the access cost (MAC): The MAC approach_
prioritizes only the access cost for each friend in a user’s social
network. Hence, for every user v in OGu, it assigns the nearest
(i.e., with minimum access cost) trusted node connected to v
as the mount point, i.e.
8v 2 OGu; M (v) = v[0] : Cv[v][0] [][ C]v[i] [;] 8i 2 T [u]
Then,
TPS(u) = fv : v 2 T (u) ^ 9 v[0] 2 NG(u) : M (v[0]) = vg
The set TPS(u) contains all members of T (u), which are
assigned as mount points for friends of u.
In OGu1 (Figure 2), assume that CIndia[Switzerland] = 1 and
CUS[Switzerland]�W est = 2. The resulting storage configuration for the
MAC approach is shown in Figure 3.
_2) Minimize the number of replicas (MNR): The MNR ap-_
proach determines the number of replicas to be maintained for
a user, so as to minimize the storage and replica management
overhead. In addition, it applies an optimization step in order
to minimize the access costs as well.
Our approach exploits the fact that the set TPS can be
modeled as the minimum connected dominating set (MCDS)
on the graph OGu, with the additional constraint that the
members of the MCDS must belong to T [u]. Hereby, we
modify a greedy algorithm from [8] to solve this variant of
the MCDS problem.
**Algorithm 1 The MNR algorithm**
1: Mark all v 2 OGu as white
2: Mark u as black
3: Mark all neighbors of u in OGu as grey
4: while 9 a white node in OGu do
5: Select a grey v[0] 2 T (u) such that v[0] has the highest
number of white neighbors in OGu
6: Mark v[0] as black and its neighbors as grey
7: end while
8: TPS[u] is the set of all black nodes in OGu
9: for all grey nodes v in OGu do
10: Mv = v[0] : Cv[v][0] [][ C]v[i] [,][ 8][i][ 2][ TPS][[][u][]]
11: end for
-----
mount point based on the result of the execution of above
algorithms.
When there is a change in the location of some trusted
nodes, the graph OGu may get disconnected. Noticing this,
node u should set itself as mount point of the disconnected
nodes. We suggest u to adjust its online time frame OTu in
order to make the TPS graph connected in this case.
_D. Replica synchronization_
We propose that after every update, the concerned mount
point pushes the update to other TPS members during their
online time frame. Note that OGu[T [u]] is connected. Assume
that each TPS member is informed of other members by
the user u during TPS creation. Until recent updates reach
a mount point, it continues to serve access requests with
out-dated content, which is acceptable, as porkut aims to
eventual consistency among replicas with tolerable temporary
inconsistencies.
_E. Accessing a user’s profile_
A user u’s profile content is available to his friends in
the social network directly through their mount points. New
nodes which are not assigned any mount point, can reach the
TPS members via the DHT index and access the content after
appropriate authorization. However, as already mentioned, the
exact organization of the profile content, the request format,
and the access control policies are beyond our scope.
IV. PRIVACY PRESERVING INDEXING
We advocate privacy-aware indexing of social networking
content of users in the system. Such index facilitates content
discovery on OSN among friends and allows users with
specialized interesting content to reach new potential friends.
Furthermore, this index allows for short-lived friendship relations for the exchange of a particular content.
_A. Privacy objectives_
_porkut’s indexing service addresses various levels of pri-_
vacy, which are described below:
No privacy: Content with no privacy requirements is
freely accessible by any social network participant.
Owner privacy: The owner of a particular content (i.e.
the user in whose profile the content exists) should not
be able to be determined with certainty by the index entry
for the content.
Content and Owner privacy: In addition to owner privacy,
the index entry should not allow someone to determine
with certainty whether a particular content item exists in
the system or not.
_B. Index creation_
A conventional DHT-based index has entries in the form
_(key,value) pairs, where a content identifier (i.e., search term_
on the index) maps to the key and the user profile identifier
(UId) maps to the value field. In order to achieve content
and owner privacies, porkut indexing mechanism uses kanonymization techniques [9] and (key,value) pairs are replaced by (key[],value[]) pairs i.e., a list of keys are now
mapped to a list of values. We call such an index entry as
(c; o)- entry, where c is the size of the key list and o is the size
of value list. A user inspecting a (c; o)-entry cannot identify
which of the content items exist in the system. By analogy,
the conventional index entries are referred to as (1; 1)- entries.
When a user creates an index entry for a content item, he
mixes the item identifier with c � 1 randomly chosen yet
meaningful item identifiers and the owner identifier with o � 1
randomly chosen user identifiers, thus creating a (c; o)-entry
from a (1; 1)-entry. Each user uses a dictionary of content
items which, for example, can be constructed from all of his
accessible content items in the social network. This dictionary
is used as input to the content anonymization technique.
Content entries that require no privacy use c = 1; o = 1.
When only owner privacy is needed, c = 1; o > 1 are
employed. Using c > 1; o > 1 results in index entries that
support both content and owner privacies.
Once a user constructs (c; o)-entry, he publishes this entry
into the DHT anonymously by employing a Crowds-like
source anonymization technique [10], where a crowd is the set
of these o users in the index entry. At the end of anonymous
routing, a (c; o)-entry is inserted into the DHT as c separate
(1; o) entries with each of them having one of the c keys as a
pivot. The detailed privacy preserving index construction and
its evaluation for a P2P system are described in [11].
A user retrieves from the DHT, the list of UIds associated
with his searched key. Then, for each of the UIds, he contacts
one (again k-anonymized) of its corresponding TPS members
for the content item that he looks for. Our index allows
strangers (i.e. non-friend users) to contact each other based
on interesting content. Authentication and authorization follow
this step.
V. RELATED WORK
There is significant related work on privacy issues in social
networks. The possibility for involuntary personal information
leakage in current social networks is highlighted in [12], e.g.
by means of certain OSN features like annotating or tagging
user photos, and its effects are demonstrated in [4].
Lockr system [13] improves the privacy of centralized
and decentralized content sharing systems. It allows users
to control their own social information by decoupling the
social networking information from other OSN functionality
using social attestations, which act like capabilities. However,
these social attestations are used only for authentication and
authorization is enforced using separate authorization policies.
Persona [14] uses attribute-based encryption to realize privacypreserving OSNs. The attributes a user has (e.g., friend, family
member, colleague) determine what data he can access. The
NOYB approach [3] adopts a novel approach for preserving
content privacy. They observe that if users address their privacy
issues themselves by hosting encrypted content on OSNs, they
could be expelled from the OSN by the OSN operator. Hence,
-----
they propose to replace users profile content items with “fake”
items randomly picked from a dictionary. NOYB encrypts the
index of the user’s item in this dictionary and uses the ciphered
index to pick the substitute. On the other hand, flyByNight [15]
encrypts the users’ content that hosts on the OSN.
Recently, the issue of using decentralized infrastructures
for organizing OSNs in a privacy-preserving manner, was addressed by the research community [1], [7], [16]. PeerSon [16]
adopts encryption mechanisms for content storage and access
control enforcement. It uses a two-tier architecture in which
the first tier is a DHT, which is used as a common storage by
all participants. The second tier consists of peers and contains
the user data. The DHT stores the meta-data required to find
users. Peers connect each other directly, exchange the content,
and then disconnect. [7] addresses privacy in OSNs by storing
profile content in a P2P storage infrastructure. Each user in
the OSN defines his own view (“matryoshka”) of the system.
In this view, nodes are organized in concentric rings, having
nodes at each ring trusted by the nodes in its immediate inner
ring, with the user node being the center of all rings. The user’s
profile data is stored encrypted at the innermost ring, which
is accessed by other users through multi-hop anonymous
communication across this set of concentric rings. In the DHT,
an entry for a user with the list of nodes in the outermost
ring is added. Thus, [7] achieves both content privacy (using
encryption) and anonymity of searcher and hosting nodes, yet
limited content discovery and profile availability, as opposed
to our approach.
In [1], a decentralized OSN, Vis-`a-Vis is proposed, where
a user’s profile content is stored at his own machine called as
virtual individual server (VIS). VISs self-organize into P2P
overlays, one overlay per social group what has access to
content stored on a VIS. Three different storage environments
are considered: cloud alone, P2P storage on top of desktops,
a hybrid storage, and their availability, cost, and privacy
trade-offs were studied. In desktop-only storage model, a
_socially-informed replication scheme was proposed, where a_
user replicates his content to his friend nodes and delegates
access control to them. However, normally, a uses trusts only
a fraction of his friends to the extent of delegating access
control enforcement, as considered in our porkut approach
along with online time information. Our earlier work [6]
considered access control delegation in P2P systems in terms
of trust transitivity.
Tribler [17] is a P2P file sharing application which exploits
friendship relationships, tastes and preferences of users to
increase the performance of file sharing. However, in Tribler,
users host their own profile and therefore profile placement
for high availability and low access or consistency cost are
not considered. Finally, LifeSocial [18] is a P2P-hosted OSN
where users employ public-private key pairs to encrypt profile
data that is stored in a distributed way and is indexed in a
DHT. Friends can read a user’s profile based on a symmetric
key that is encrypted with their public keys. However, data
privacy and profile availability are not considered in [18].
VI. CONCLUSION AND FUTURE WORK
In this paper, we presented the initial design of porkut,
a privacy-preserving decentralized OSN. We emphasized on
satisfying high availability and lookup efficiency of scattered
OSN profiles. The users geographical locations and online time
statistics were exploited in deciding the user’s profile storage
points. Three algorithms with different cost minimization
objectives were presented for selecting the set of nodes that
host OSN profiles, while preserving high availability. As a
future work, we plan to deploy the porkut system, and study its
performance, availability and privacy characteristics in detail.
ACKNOWLEDGEMENT
This work was funded by the Swiss Nano-Tera OpenSense
project (Nano-Tera ref. 839 401).
REFERENCES
[1] A. Shakimov, A. Varshavsky, L. P. Cox, and R. C´aceres, “Privacy, cost,
and availability tradeoffs in decentralized osns,” in Proc. of the WOSN,
2009.
[2] A. Mislove, M. Marcon, K. P. Gummadi, P. Druschel, and B. Bhattacharjee, “Measurement and analysis of online social networks,” in Proc. of
_the 7th Internet measurements conference, 2007._
[3] S. Guha, K. Tang, and P. Francis, “Noyb: privacy in online social
networks,” in Proc. of the WOSP, Seattle, WA, USA, 2008.
[4] B. Krishnamurthy and C. E. Wills, “On the leakage of personally
identifiable information via online social networks,” in Proc. of the
_WOSN, 2009._
[5] S. Rhea, B. Godfrey, B. Karp, J. Kubiatowicz, S. Ratnasamy, S. Shenker,
I. Stoica, and H. Yu, “Opendht: a public dht service and its uses,”
_SIGCOMM Comput. Commun. Rev., vol. 35, no. 4, pp. 73–84, 2005._
[6] N. Rammohan, Z. Miklos, and K. Aberer, “Towards access control
aware p2p data management systems,” in Proc. of the 2nd International
_workshop on data management in peer-to-peer systems, 2009._
[7] L. A. Cutillo, R. Molva, and T. Strufe, “Privacy preserving social
networking through decentralization,” in Proc. of the WONS, 2009.
[8] L. Ruan, H. Du, X. Jia, W. Wu, Y. Li, and K.-I. Ko, “A greedy
approximation for minimum connected dominating sets,” Theoretical
_Computer Science, vol. 329, no. 1-3, pp. 325 – 330, 2004._
[9] L. Sweeney, “k-anonymity: a model for protecting privacy,” Int. J.
_Uncertain. Fuzziness Knowl.-Based Syst., vol. 10, no. 5, pp. 557–570,_
2002.
[10] M. K. Reiter and A. D. Rubin, “Crowds: anonymity for web transactions,” ACM Trans. Inf. Syst. Secur., vol. 1, no. 1, 1998.
[11] R. Narendula, T. G. Papaioannou, and K. Aberer, “Panacea: Tunable
privacy for access controlled data in peer-to-peer systems,” 2010, EPFL
Technical Report 148337. http://infoscience.epfl.ch/record/148337.
[12] I.-F. Lam, K.-T. Chen, and L.-J. Chen, “Involuntary information leakage
in social network services,” in Proc. of the 3rd International Workshop
_on Security, 2008._
[13] A. Tootoonchian, S. Saroiu, Y. Ganjali, and A. Wolman, “Lockr: better
privacy for social networks,” in Proc. of the CoNEXT, 2009.
[14] R. Baden, A. Bender, N. Spring, B. Bhattacharjee, and D. Starin,
“Persona: an online social network with user-defined privacy,” in Proc.
_of the ACM SIGCOMM, 2009._
[15] M. M. Lucas and N. Borisov, “Flybynight: mitigating the privacy risks
of social networking,” in Proc. of the WPES, 2008.
[16] S. Buchegger, D. Schi¨oberg, L.-H. Vu, and A. Datta, “Peerson: P2p
social networking: early experiences and insights,” in Proc. of the ACM
_EuroSys Workshop on Social Network Systems, 2009._
[17] J. A. Pouwelse, P. Garbacki, J. Wang, A. Bakker, J. Yang, A. Iosup,
D. H. J. Epema, M. Reinders, M. R. van Steen, and H. J. Sips, “Tribler:
a social-based peer-to-peer system: Research articles,” Concurr. Comput.
_: Pract. Exper., vol. 20, no. 2, pp. 127–138, 2008._
[18] K. Graffi, P. Mukherjee, B. Menges, D. Hartung, A. Kovacevic, and
R. Steinmetz, “Practical security in p2p-based social networks,” in Proc.
_of the IEEE LCN, October 2009._
-----
|
{
"disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references', 'abstract'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1109/WETICE.2010.40?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/WETICE.2010.40, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "GREEN",
"url": "http://infoscience.epfl.ch/record/148642/files/cops.pdf"
}
| 2,010
|
[
"JournalArticle"
] | true
| 2010-06-28T00:00:00
|
[] | 9,130
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Engineering",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/02562dc2b524dceb6a535c2baddc1e8d88463ad9
|
[
"Computer Science"
] | 0.881622
|
Practical Fully-Decentralized Secure Aggregation for Personal Data Management Systems
|
02562dc2b524dceb6a535c2baddc1e8d88463ad9
|
International Conference on Statistical and Scientific Database Management
|
[
{
"authorId": "2123019524",
"name": "Julien Mirval"
},
{
"authorId": "1714085",
"name": "Luc Bouganim"
},
{
"authorId": "40576371",
"name": "I. S. Popa"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"Stat Sci Database Manag",
"SSDBM",
"Statistical and Scientific Database Management",
"Int Conf Stat Sci Database Manag"
],
"alternate_urls": null,
"id": "070a8b94-9234-4a39-bff0-23e0c6b74464",
"issn": null,
"name": "International Conference on Statistical and Scientific Database Management",
"type": "conference",
"url": "http://www.ssdbm.org/"
}
|
Personal Data Management Systems (PDMS) are flourishing, boosted by legal and technical means like smart disclosure, data portability and data altruism. A PDMS allows its owner to easily collect, store and manage data, directly generated by her devices, or resulting from her interactions with companies or administrations. PDMSs unlock innovative usages by crossing multiple data sources from one or many users, thus requiring aggregation primitives. Indeed, aggregation primitives are essential to compute statistics on user data, but are also a fundamental building block for machine learning algorithms. This paper proposes a protocol allowing for secure aggregation in a massively distributed PDMS environment, which adapts to selective participation and PDMSs characteristics, and is reliable with respect to failures, with no compromise on accuracy. Preliminary experiments show the effectiveness of our protocol which can adapt to several contexts with varying PDMSs characteristics in terms of communication speed or CPU resources and can adjust the aggregation strategy to the estimated selective participation.
|
# Practical Fully-Decentralized Secure Aggregation for Personal Data Management Systems
## Julien Mirval, Luc Bouganim, Iulian Sandu Popa
To cite this version:
#### Julien Mirval, Luc Bouganim, Iulian Sandu Popa. Practical Fully-Decentralized Secure Aggrega- tion for Personal Data Management Systems. 33rd International Conference on Scientific and Sta- tistical Database Management, SSDBM 2021, Jul 2021, Tampla, FL, United States. pp.259-264, 10.1145/3468791.3468821. hal-03329878
## HAL Id: hal-03329878
https://hal.science/hal-03329878
#### Submitted on 8 Oct 2021
HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non, lished or not. The documents may come from émanant des établissements d’enseignement et de teaching and research institutions in France or recherche français ou étrangers, des laboratoires abroad, or from public or private research centers. publics ou privés.
```
Copyright
```
-----
# Practical Fully-Decentralized Secure Aggregation for Personal Data Management Systems
### Julien Mirval
##### julien.mirval@cozycloud.cc Cozy Cloud Inria-Saclay UVSQ, Université Paris-Saclay France
#### ABSTRACT
### Luc Bouganim
##### luc.bouganim@inria.fr Inria-Saclay UVSQ, Université Paris-Saclay France
### Iulian Sandu-Popa
##### iulian.sandu-popa@uvsq.fr UVSQ, Université Paris-Saclay Inria-Saclay France
Personal Data Management Systems (PDMS) are flourishing, boosted
by legal and technical means like smart disclosure, data portability and data altruism. A PDMS allows its owner to easily collect,
store and manage data, directly generated by her devices, or resulting from her interactions with companies or administrations.
PDMSs unlock innovative usages by crossing multiple data sources
from one or many users, thus requiring aggregation primitives.
Indeed, aggregation primitives are essential to compute statistics
on user data, but are also a fundamental building block for machine
learning algorithms. This paper proposes a protocol allowing for
secure aggregation in a massively distributed PDMS environment,
which adapts to selective participation and PDMSs characteristics,
and is reliable with respect to failures, with no compromise on
accuracy. Preliminary experiments show the effectiveness of our
protocol which can adapt to several contexts with varying PDMSs
characteristics in terms of communication speed or CPU resources
and can adjust the aggregation strategy to the estimated selective
participation.
#### CCS CONCEPTS
- Computer systems organization → **Architectures; • Infor-**
**mation systems →** _Data management systems._
#### KEYWORDS
Privacy, secure aggregation, decentralized, machine learning.
**ACM Reference Format:**
Julien Mirval, Luc Bouganim, and Iulian Sandu-Popa. 2021. Practical FullyDecentralized Secure Aggregation for Personal Data Management Systems.
In 33rd International Conference on Scientific and Statistical Database Man_agement (SSDBM 2021), July 6–7, 2021, Tampa, FL, USA. ACM, New York,_
[NY, USA, 6 pages. https://doi.org/10.1145/3468791.3468821](https://doi.org/10.1145/3468791.3468821)
#### 1 INTRODUCTION
The new privacy-protection regulations (e.g., GDPR) and smart
disclosure initiatives in the last decade have boosted the development and adoption of Personal Data Management Systems (PDMSs)
[2]. A PDMS (e.g., Cozy Cloud, Nextcloud, Solid) is a data platform
Publication rights licensed to ACM. ACM acknowledges that this contribution was
authored or co-authored by an employee, contractor or affiliate of a national government. As such, the Government retains a nonexclusive, royalty-free right to publish or
reproduce this article, or to allow others to do so, for Government purposes only.
_SSDBM 2021, July 6–7, 2021, Tampa, FL, USA_
© 2021 Copyright held by the owner/author(s). Publication rights licensed to ACM.
ACM ISBN 978-1-4503-8413-1/21/07...$15.00
[https://doi.org/10.1145/3468791.3468821](https://doi.org/10.1145/3468791.3468821)
allowing users to easily collect, store and manage into a single place
data directly generated by user devices (e.g., quantified-self data,
smart home data, photos) and data resulting from user interactions
(e.g., social interaction data, health, bank, telecom). Users can then
leverage the power of their PDMS to benefit from their personal
data for their own good and in the interest of the community [7].
Consequently, the PDMS paradigm leads to an important shift
in the personal data ecosystem since data becomes massively distributed, at the user-side. It also holds the promise of unlocking
innovative usages. An individual can now cross her data from different data silos, e.g., health records and physical activity data.
Moreover, individuals can cross data within large communities of
users, e.g., to compute statistics for epidemiological studies or to
train a machine learning model (ML) for recommender systems
or automatic classification of user data. However, these exciting
perspectives should not eclipse the security issues –user data must
be kept private– and the right for any PDMS user to consent, or
not, in participating in each computation.
Aggregation primitives (e.g., sum or average) are obviously essential to compute basic statistics on user data but are also a fundamental building block for machine learning algorithms. Thus,
to enable such new usages, we need scalable, privacy-preserving
protocols implementing data aggregation primitives with selective
(i.e., consenting) participants. Ideally, the proposed protocol should
provide an accurate result that fully takes advantage of high-quality
data available in PDMSs. Efficiency (i.e., protocol latency and total
load of the system) is of prime importance and the protocol should
adapt to several contexts: the PDMSs could be limited by their communication speed or by their computation power. Finally, given the
scale of such decentralized aggregation, such protocols must also
be robust to node failures. To summarize, our goal is to propose an
aggregation protocol for basic aggregate functions that fulfills the
following properties:
- fully decentralized and highly scalable, with the number of
participants.
- privacy-preserving, i.e., it protects the confidentiality of user
data.
- accurate, i.e., it does not require a trade-off between accuracy
and privacy.
- adaptable, i.e, it can adapt to a large spectrum of computation
selectivity values (reflecting the subset of contributor nodes)
and system configurations (network and cryptographic latency).
- reliable, i.e., it handles node failures or voluntary disconnections.
-----
SSDBM 2021, July 6–7, 2021, Tampa, FL, USA Mirval, Bouganim and Sandu-Popa
The rest of this paper is organized as following. After discussing
the related works w.r.t. the required properties in Section 2, we
introduce the considered architecture and threat model in Section
3. Sections 4 and 5 focus on the proposed protocol and preliminary
results. Section 6 concludes with future issues.
#### 2 RELATED WORKS
Secure aggregation is an intense research area since many years
and many approaches were proposed: secure multi-party computation (SMC) and (fully) homomorphic threshold encryption (HTE),
(local) differential privacy and gossip-based protocols. However,
the existing solutions are not adapted to the PDMS context and fail
to cover all the required properties listed above.
HTE and SMC-based solutions [4, 6, 9, 10] generally target applications in which central servers orchestrate and coordinate the
participating nodes (e.g. federated learning). Such solutions are not
scalable with a large number of participants and a fully decentralized setting such as in the PDMS context (e.g., the server(s) load is
linear [9] or quadratic [6] with the number of participants).
Local differential privacy (LDP) has gained significant momentum in the recent years addressing problems such as machine learning [14] or basic statistics based on range queries [8]. However, LDP
requires more noise than classical DP [1], either affecting accuracy
or requiring a large number of participants to reduce the impact of
noise, contradicting adaptability to selective participation.
Gossip-based protocols are scalable, fully decentralized, reliable
and have an adjustable accuracy. Unfortunately, classical gossipbased protocols do not protect the user privacy. In [5], participants
collectively learn a machine learning model in a privacy preserving
way by gossiping differentially private models, impacting accuracy.
In [12], participants introduce noise in the first iterations and gradually remove it in subsequent iterations. This approach makes such
solutions unreliable w.r.t. node failures. Finally, we are not aware
of gossip protocols tolerating selective participation and trivial
adaptations produce inaccurate results.
#### 3 SYSTEM OVERVIEW AND THREAT MODEL
In this section, we introduce first the system architecture and the
related concepts. We then present the considered threat model for
the proposed secure protocol.
#### 3.1 System Architecture
**P2P network. We envision a fully distributed Peer-to-Peer (P2P)**
system relying only on PDMSs, thus requiring an efficient communication overlay. Distributed Hash Tables (DHT) are structured
overlays which enable a logarithmic scalability with the number of
nodes. Our protocol is currently built on top of the Chord DHT [13].
Each node has an Id obtained by hashing a static property of the
node and stores a fingertable (FT) to route Chord messages. FT is a
table with a number of entries equal to the size of the Id space in
bits. If X is a node Id, the i[th] entry of the FT contains the IP address
of the node whose Id is closest but lower than X + 2[i] . Routing is
done by searching in the FT the closest entry to the target address
and transmitting recursively the message until it reaches its target,
with a worse case of O(loд(N )) message complexity, where N is
the number of DHT nodes.
**Computation model. An aggregate computation can be triggered**
by any node, called querier. The querier broadcasts the computation
and each node consents or not to contribute, and in the positive case
is called contributor. The ratio between the number of contributors
and total number of nodes defines the selectivity σ, 0 < σ ≤ 1.
Finally, each node (contributor or not) is a potential data processor
and is then called aggregator.
#### 3.2 Threat Model
We consider the honest-but-curious threat model, in which, an attacker can access, without altering, the data manipulated by the
attacked nodes, then called leaking nodes. The rationale behind the
honest-but-curious model is that a PDMS can hold the entire digital
life of her owner and therefore needs to be highly protected against
privacy threats. Recent works [3] indicate that Trusted Execution
Environments (TEEs) are prime candidates to offer this protection
since they guarantee that executed code and manipulated data cannot be observed. In our context, this property allows sharing data
between PDMS nodes without breaking data confidentiality.
We thus consider that each PDMS is protected by a secure hardware solution, such as Intel SGX or ARM TrustZone, providing a
TEE. Such a hardware protection makes attacks difficult to produce, but since no security measure is unbreakable, we consider
that some PDMS owners have succeeded in tampering their PDMS.
Since attackers may collude and thus, de facto, control more than
one PDMS, the worst-case attack is represented by the maximum
number of colluding nodes controlled by a single “attacker”, i.e., C
leaking nodes.
Additionally, the TEE of each PDMS is equipped with a trustworthy certificate. Thus, any node can verify the authenticity of
other participants by checking their certificate. This prevents Sybil
attacks (i.e., forging nodes to master a large portion of the system).
Finally, attackers can also observe the communications between
the nodes, thus requiring secure communication channels (e.g., TLS)
to protect sensitive data exchanges.
Our objective is to provide a protocol that fully protects the
confidentiality of the contributors’ data and all the intermediary
results, with high and tunable probability, the final result not being
confidential. Also, we consider that being a contributor for a given
computation is not a sensitive information.
#### 4 PROPOSED PROTOCOL
In the protocol overview, we analyze the properties listed in Section 1 and present the main ideas and techniques behind each
property and its impact on the protocol. Due to space constraints,
we cannot describe in detail the proposed protocol and thus discuss
some identified key elements of the protocol under the form of
questions/answers, considering first the privacy aspects and then
the efficiency perspective.
#### 4.1 Protocol Overview
**Scalability: The DHT achieves de facto a fully decentralized and**
efficient architecture for inter-nodes communication. Achieving
a scalable aggregation process requires multiple aggregators, arranged in a tree structure. Building and broadcasting this aggregation tree can be very costly since the tree itself can be large. We
-----
Practical Fully-Decentralized Secure Aggregation for Personal Data Management Systems SSDBM 2021, July 6–7, 2021, Tampa, FL, USA
thus employ a divide-and-conquer approach to parallelize the tree
construction and diffusion and use the finger table structure to
minimize communications. Finally, we reduce the knowledge (and
thus the diffusion) of the tree to the minimal part strictly necessary
to perform the aggregation: basically, each node of the tree only
knows its parent(s) and its children.
**Privacy and accuracy: We use a secret sharing scheme (without**
threshold) in which each contributor splits its data into s shares,
making them unreadable unless someone collects all s shares. s
is computed such that the probability to obtain s shares for an
attacker, controlling C nodes, is inferior to a security threshold
_α (e.g., α = 10[−][6]). Each i[th]_ share has the value xi = x + ϵi such
that [�]i[n]=1 _[ϵ][i][ =][ 0, where][ x][ is the private value. This way, shares]_
from different contributors can be aggregated separately and if
no share is missing (the reliability is discussed below), the final
result will be equal to the exact sum of all private data. Hence, our
protocol provides also, by construction, accurate results. Note that
the protocol works for complex values of x, such as an array or a
matrix, which is useful for advanced aggregations, e.g., training a
naive Bayes ML model.
**Adaptability: The number of aggregators (i.e., the tree fan-out and**
its height) is tuned as a function of the number of contributors,
the communication costs (i.e., the latency to send a message between two nodes) and the processing costs (i.e., the asymmetric
cryptographic costs to secure a communication or to sign or verify
a signature, which is, by far, the most important processing costs).
This allows the protocol to always offer near-optimal performance
(i.e., aggregation latency) and achieve adaptability w.r.t. the computation selectivity and PDMSs characteristics. Furthermore, our
protocol can also be conveniently configured to offer the desired
trade-off between the latency and the total cost of the aggregation, which are conflicting optimization objectives as discussed in
Section 5.
**Reliability: Handling failures and disconnections is mainly im-**
plemented at two levels. First, the aggregators in the last level of
the tree (just above the leaves) execute a synchronization protocol
to make sure that contributors have sent all the s shares before
disconnection and remove the shares for the contributors that have
sent less than s shares. This ensures that the aggregation result
stays accurate despite contributors failure. Second, a list of backup
aggregators is created before the tree creation. Its size depends on
the observed node failure/disconnection ratio. In case an aggregator fails, it is automatically replaced with a backup node during
the aggregation process (the parents monitor their children). This
allows the protocol to be robust to node failures and avoids losing
aggregation subtree results.
#### 4.2 Privacy Issues
_What is the impact of the secret sharing on the aggregation tree?_
Considering _s shares for each contributor and partial results leads to_
build s separate aggregation trees, with exactly the same structure,
to avoid inferences from an attacker on any of the intermediate
results. The final sum of the shares is done by the querier (tree root).
A simple means to construct such trees is to consider that each
node of the tree is a group of s nodes (see Figure 1 with s = 2). The
protocol to build the tree is described in Section 4.3 considering
that, at each step, s nodes are selected instead of one. To make this
selection efficient, each node in the DHT maintains a cache with the
addresses of the s − 1 successor nodes that will form the aggregator
group.
_How is the number of shares computed? An attacker could cleverly_
locate her controlled nodes in the DHT to obtain the s shares of
a group (typically controlling a node and its s − 1 successors). We
avoid this attack by reusing the concept of imposed location that we
proposed in [11]: the node Id in the DHT is computed by hashing the
public key from the PDMS certificate (see Section 3.2). The nodes are
then uniformly distributed in the DHT space and the PDMS owner
(here the attacker) cannot influence this placement: the uniform
distribution also applies to leaking nodes. As a consequence, s can
be easily computed and is minimal when s = ⌈log(α)/log(C/N )⌉.
_Do contributors/aggregators have to check the correctness of the_
_received query? Basically, the answer is yes. Indeed, a trivial attack_
would be to impersonate s aggregators (at the bottom of the tree)
and ask a set of contributors for their shares, with the same protocol.
If no control is done, the contributor cannot distinguish a real query
from a fake query. To avoid such an attack, every aggregator must
check the signature of the incoming query using the public key of
the sender, having previously checked the validity of the sender’s
certificate. Since all the nodes are honest but curious, they must
follow the protocol and thus cannot create a specific query that
would lead to the disclosure of certain data.
#### 4.3 Efficiency Issues
_What is the divide-and-conquer approach to build the aggregation_
_tree? Assuming the querier knows the height h and the fan-out_
_f of the aggregation tree, it starts creating a tree assigning the_
whole DHT to its successor(s). Recursively, each aggregator in the
tree (i.e., a parent node) is assigned to a DHT region that it will
subdivide and delegate to other aggregators in that region. When
an aggregator oversees a DHT region, it looks for f nodes that
are (almost) evenly spaced across the region. The node responsible
for finding peers is a parent aggregator, while the selected nodes
are child aggregators. Each child then becomes the parent of the
region between itself and the next sibling. This process goes on
until the height h is reached. At the last tree level, the tree leaves
(i.e., the contributors) are found by using a localized DHT broadcast
in the respective region. Contributors willing to participate reply
with their private data, after establishing a secure channel with
their aggregator parent. The aggregators at level d aggregate the
data they receive before sending them to the previous level of the
tree down to the root (i.e., the querier) which performs the final
aggregation to obtain the result. Figure 1 illustrates this process
with two nodes per group (blue and red) by using letters to represent
a group. The fan-out is 4 and the height is 3 (excluding the querier
Q). Q selects his successor, A, who is responsible of the whole DHT
and is the root of the tree. A uses its finger table to contact C, E and
B. C recursively contacts D. The second level of the tree is built.
Then B, E, C and D contact recursively the nodes for the second
level (the figure only shows what happens with E for readability).
When leaves are contacted, they send one share to each aggregators
of the group (i.e., blue and red) which are summed-up separately
in each aggregation tree and finally summed-up by Q.
-----
SSDBM 2021, July 6–7, 2021, Tampa, FL, USA Mirval, Bouganim and Sandu-Popa
### QAB Q
1 4 Aggregator share 1
##### Aggregator share 2 A Contributor
4 5 Querier B E C D
3 E
### D 2 3
6 I
4 5 I H F G
4
### H
5
### F
## Chord
### G
## DHT Aggregation tree C
**Figure 1: Building the aggregation tree based on DHT**
_How does an aggregator contribute? If a node selected as aggrega-_
tor in the tree wishes to contribute, it can simply add its data to the
partial aggregate it computes before sending it to its parent. Note
that it will add it without splitting the data into shares since its parent cannot guess this addition. To compute an average, we need to
count the number of contributors and thus, the aggregator will add
_s to the count of share contributions: each aggregator accounts the_
number of shares it received, and the total will be finally divided by
_s to obtain the number of contributors. Consequently, aggregators_
do not appear as leaves in the aggregation tree. Note that this is
not the case for backup nodes which must have the possibility to
appear as leaves of the tree in case they wish to contribute.
_How are the tree fan-out f and height_ _h computed? At one extreme,_
a binary tree (f = 2) distributes the query load on a maximum
number of aggregator nodes but increases the communications
costs, including the creation of many secure channels to transfer
the intermediate results. At the other extreme, a tree limited to a
unique aggregator (f = σ × N ) minimizes the communications and
thus the number of secure channels (1 per contributor). It minimizes
the total system load induced by the query but concentrates most
of that load on this unique aggregator (that becomes overloaded by
asymmetric crypto operations for the communication decryption).
An "ideal" aggregation tree would be completely balanced, with
the same fan-out all along the tree. Moreover, this fan-out (and
thus the height of the tree) would be cleverly chosen to optimize
the query latency without impacting too much the total load. Note
that this depends on the PDMS characteristics, i.e., communication
speed or computation power. Finally, the tree height is simply
computed based on the number of contributors (σ × _N_ ) and the tree
fan-out. σ can be estimated, for instance by contacting all nodes
within a region of the DHT, and checking the ratio of nodes willing
to participate. Since nodes are uniformly distributed in the DHT
thanks to the hash of their public key, choosing a sample of the
population should give a good estimation of σ .
#### 5 PRELIMINARY RESULTS
As in most evaluations of distributed systems [13], we implemented
a simulator allowing varying any parameter: number of nodes N,
of colluding nodes C, security threshold α, selectivity σ, and β, a
ratio defined below. Our simulator captures two metrics: (i) for the
network utilization, we consider the number of exchanged messages
as the most important metric (compared to, e.g., the message size);
(ii) for the PDMS resource utilization, the simulator counts the
_asymmetric cryptographic operations which are, by far, the most_
expensive operations. The output of the simulator is the protocol
latency and total work. They depend on β, the relative cost of
one asymmetric cryptographic operation denoted crypt and the
latency when sending a message between two PDMSs denoted com.
Specifically, β = crypt/(crypt + com) with 0 ≤ _β ≤_ 1. However,
note that the two extremes values of β are not realistic, i.e., β = 0
when crypt = 0 or com = +∞, β = 1 when com = 0 or crypt = +∞.
Our protocol is adaptive to σ and β, thus called Adaptive in
this section. To measure the impact of these two parameters on the
aggregation costs, we compare the Adaptive protocol to two other
simplified versions. First, Full tree is a classical aggregation tree
that does not adapt to the query selectivity, i.e., it considers σ = 1:
a tree is created recursively until all nodes are included, but only
those willing to contribute will send back shares. Second, Single
**_level considers that β = 0, i.e., the communication cost is so high_**
that we must minimize it, thus concentrating all the computation
on a single group, collecting the shares from all participants, and
sending the results to the querier.
We consider a network with N = 1, 000, 000 nodes, a quite large
attack level (C = 10, 000) and a high security threshold (α = 10[−][6])
and compare the above protocols in relative terms, i.e., dividing the
latency/total work of Full tree/Single Level by the one of Adaptive.
We first confirmed that the adaptive protocol is scalable. With
increasing values of N, we obtained a logarithmic increase of the
latency, thanks to the DHT and the divide-and-conquer approach.
We also verified that the number of colluding nodes C has a small
impact on the protocol latency, with reasonable values of C w.r.t. N
-----
Practical Fully-Decentralized Secure Aggregation for Personal Data Management Systems SSDBM 2021, July 6–7, 2021, Tampa, FL, USA
#### 10[4]
10[3]
10[2]
10[1]
10[0] 10[−][4] 10[−][3] 10[−][2] 10[−][1] 10[0]
#### 10[4]
10[3]
#### 10[0] 0.0 0.2 0.4 0.60 0.80 1.00
𝛽 ratio
#### 10[2]
10[1]
#### selectivity (𝜎)
#### 10[3]
10[2]
10[1]
10[0]
10[−][1] 10[−][4] 10[−][3] 10[−][2] 10[−][1] 10[0]
#### 10[3]
10[2]
#### 10[1]
#### 10[0]
10[−][1] 0.0 0.2 0.4 0.6 0.8 1.0
|103|Col2|
|---|---|
|102 work total 101 relative 100 1||
|||
|||
|||
#### selectivity (𝜎)
**Figure 2: Latency and total work relatively to the Adaptive**
**_strategy varying σ_**
#### 𝛽 ratio
**Figure 3: Latency and total work relatively to the Adaptive**
**_strategy varying β_**
in accordance with the considered threat model. Thus, in the rest
of this section, we focus on the adaptability feature of our protocol
and leave the evaluation of its reliability for future works. We vary
the selectivity σ (keeping β = 0.5) and the PDMSs characteristics
_β (keeping σ = 0.01). The results are presented in Figures 2 and 3_
(log scale on Y axis for all graphs and on X axis for selectivity only).
Let’s first focus on the Single Level protocol studied to show the
impact of an extreme strategy, i.e., concentrating all the load on a
single (group of) node(s). As expected, Single Level always provides
a better total work than Adaptive and Full tree. However, the latency
increases linearly with the number of participants leading rapidly
to prohibitive costs. Practically, Single Level is competitive only
if the selectivity is extremely high (i.e., tens to a few hundreds of
contributors) or β = 0 (i.e., unrealistic setting).
Execution based on aggregation trees (Full tree or Adaptive) are
much scalable for handling many contributors by distributing the
workload. Note that for a maximal selectivity, both approaches have
exactly the same latency, as their structure is identical. However,
_Full tree becomes more costly for both latency and total work as_
soon as the selectivity is below 1. Indeed, the adaptive fan-out and
tree depth of Adaptive can reduce the latency up to a factor of 3
and especially the total work up to two orders of magnitude, which
indicates the importance of adapting the aggregation structure to
the computation and system settings.
In the last part of our experimental evaluation, we study in more
details the Adaptive protocol. In particular, we evaluate the impact
of the tree fan-out on the latency and the total work of the protocol
with different values of β while keeping σ = 0.01. The results are
presented in Figure 4 (log scale on the X axis for both graphs). As
above and to increase the readability, we represent relative values
for both the latency and the total work, i.e., the ratio between the
latency value (or the total work value) and its minimum observed
value.
As expected, increasing the fan-out, decreases the total work,
as the aggregation tree includes less nodes. This reduces the total
amount of communications (and hence reduces the number of
secure channels), but concentrates the cryptographic load on a
-----
SSDBM 2021, July 6–7, 2021, Tampa, FL, USA Mirval, Bouganim and Sandu-Popa
_𝛽_ = 0.2 _𝛽_ = 0.4 _𝛽_ = 0.6 _𝛽_ = 0.8 _𝛽_ = 1.0
#### 3 2
#### 2.75
2.5
#### 2.25
2
#### 1.75
1.5
#### 1.25
1
#### 2 4 8 16 32
fan-out
#### 2 4 8 16 32
fan-out
**Figure 4: Relative latency and relative total work w.r.t. the minimum value varying the fan-out**
few nodes, leading generally to a higher latency. However, we
observed an exception to this behavior for small fan-out values.
In this case, the communication overhead required to construct
the tree leads to sub-optimal latency. With small values of β (i.e.,
the communication cost is larger than the cryptographic cost),
this overhead is more prominent. On the contrary, once the fanout increases, smaller values of β result in a decreased latency, as
the cryptographic operations, which are dominant, are relatively
cheaper.
Our results confirm that there is a sweet spot for the fan-out
depending on the PDMSs and network characteristics. The results
also indicate that, depending on the application requirements, the
fan-out can be adjusted to obtain a better trade-off between latency
and total work. For example, training a machine learning model
on user’s data may be less restricted in terms of latency than a
real-time traffic analysis. For instance, when β = 0.6, choosing a
fan-out of 8 leads to a total work only 3% higher than the optimal
value, while the latency is 32% larger than the optimal value.
#### 6 CONCLUSION AND FUTURE WORKS
In this short paper, we made the first steps towards the design of
an aggregation protocol providing interesting properties: highly
scalable, privacy preserving, adaptable to selective participation, to
several system settings, with a tree-like structure enabling robustness to failure; all this without compromise on the result quality.
This protocol could be a building block to compute statistics on
large communities of PDMS users or even to train ML algorithms.
There is still a long way to go before providing all the required
properties with efficient and secure protocols. Our next steps are
to focus on the reliability aspect, selectivity estimation, and performance enhancements in the case of ML algorithms manipulating
large datasets and requiring many iterations on users’ data. This is
an exciting research agenda with innovative usages in perspective.
#### REFERENCES
[1] Mário S. Alvim, Konstantinos Chatzikokolakis, Catuscia Palamidessi, and Anna
Pazii. 2018. Local Differential Privacy on Metric Spaces: Optimizing the Trade-Off
[with Utility. In IEEE CSF. 262–267. https://doi.org/10.1109/CSF.2018.00026](https://doi.org/10.1109/CSF.2018.00026)
[2] Nicolas Anciaux, Philippe Bonnet, Luc Bouganim, Benjamin Nguyen, Philippe
Pucheral, Iulian Sandu Popa, and Guillaume Scerri. 2019. Personal data management systems: The security and functionality standpoint. Information Systems
80 (2019), 13–35.
[3] Nicolas Anciaux, Luc Bouganim, Philippe Pucheral, Iulian Sandu Popa, and
Guillaume Scerri. 2019. Personal Database Security and Trusted Execution
Environments: A Tutorial at the Crossroads. Proc. VLDB Endow. 12, 12 (2019),
1994–1997.
[4] Yoshinori Aono, Takuya Hayashi, Lihua Wang, Shiho Moriai, et al. 2017. Privacypreserving deep learning via additively homomorphic encryption. IEEE Transac_tions on Information Forensics and Security 13, 5 (2017), 1333–1345._
[5] Aurélien Bellet, Rachid Guerraoui, Mahsa Taziki, and Marc Tommasi. 2018. Personalized and private peer-to-peer machine learning. In International Conference
_on Artificial Intelligence and Statistics. PMLR, 473–481._
[6] Keith Bonawitz, Vladimir Ivanov, Ben Kreuter, Antonio Marcedone, H Brendan
McMahan, Sarvar Patel, Daniel Ramage, Aaron Segal, and Karn Seth. 2017. Practical secure aggregation for privacy-preserving machine learning. In ACM CCS.
1175–1191.
[7] EU Commission. 25 October 2020. Proposal for a Regulation on European data
[governance (Data Governance Act), COM/2020/767. [eur-lex].](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:52020PC0767)
[8] Graham Cormode, Tejas Kulkarni, and Divesh Srivastava. 2019. Answering
[Range Queries Under Local Differential Privacy. PVLDB 12, 10 (2019). https:](https://doi.org/10.14778/3339490.3339496)
[//doi.org/10.14778/3339490.3339496](https://doi.org/10.14778/3339490.3339496)
[9] Henry Corrigan-Gibbs and Dan Boneh. 2017. Prio: Private, robust, and scalable
computation of aggregate statistics. In NSDI. 259–282.
[10] David Froelicher, Juan Ramón Troncoso-Pastoriza, Joao Sa Sousa, and Jean-Pierre
Hubaux. 2020. Drynx: Decentralized, secure, verifiable system for statistical
queries and machine learning on distributed datasets. IEEE Transactions on
_Information Forensics and Security 15 (2020), 3035–3050._
[11] Julien Loudet, Iulian Sandu Popa, and Luc Bouganim. 2019. SEP2P: Secure and
Efficient P2P Personal Data Processing. In EDBT.
[12] Yilin Mo and Richard M Murray. 2016. Privacy preserving average consensus.
_IEEE Trans. Automat. Control 62, 2 (2016), 753–765._
[13] Ion Stoica, Robert Morris, David Karger, M Frans Kaashoek, and Hari Balakrishnan. 2001. Chord: A scalable peer-to-peer lookup service for internet applications.
_ACM SIGCOMM 31, 4 (2001), 149–160._
[14] Kai Zheng, Wenlong Mou, and Liwei Wang. 2017. Collect at Once, Use Effectively:
Making Non-interactive Locally Private Learning Possible. In ICML, Vol. 70.
-----
|
{
"disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1145/3468791.3468821?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1145/3468791.3468821, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "other-oa",
"status": "GREEN",
"url": "https://hal.archives-ouvertes.fr/hal-03329878/file/3468791.3468821.pdf"
}
| 2,021
|
[
"JournalArticle",
"Book",
"Conference"
] | true
| 2021-07-06T00:00:00
|
[] | 8,564
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Engineering",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/025659bfc224e9420db7c3c50027478355a32fb2
|
[
"Computer Science"
] | 0.874027
|
An Improved Deep Learning Model for DDoS Detection Based on Hybrid Stacked Autoencoder and Checkpoint Network
|
025659bfc224e9420db7c3c50027478355a32fb2
|
Future Internet
|
[
{
"authorId": "144871713",
"name": "Amthal K. Mousa"
},
{
"authorId": "2112605919",
"name": "M. N. Abdullah"
}
] |
{
"alternate_issns": null,
"alternate_names": null,
"alternate_urls": [
"http://nbn-resolving.de/urn/resolver.pl?urn=urn:nbn:ch:bel-156830",
"https://www.mdpi.com/journal/futureinternet"
],
"id": "c3e5f1c8-9ba7-47e5-acde-53063a69d483",
"issn": "1999-5903",
"name": "Future Internet",
"type": "journal",
"url": "http://www.e-helvetica.nb.admin.ch/directAccess?callnumber=bel-156830"
}
|
The software defined network (SDN) collects network traffic data and proactively manages networks. SDN’s programmability makes it excellent for developing distributed applications, cybersecurity, and decentralized network control in multitenant data centers. This exceptional architecture is vulnerable to security concerns, such as distributed denial of service (DDoS) attacks. DDoS attacks can be very serious due to the fact that they prevent authentic users from accessing, temporarily or indefinitely, resources they would normally expect to have. Moreover, there are continuous efforts from attackers to produce new techniques to avoid detection. Furthermore, many existing DDoS detection methods now in use have a high potential for producing false positives. This motivates us to provide an overview of the research studies that have already been conducted in this area and point out the strengths and weaknesses of each of those approaches. Hence, adopting an optimal detection method is necessary to overcome these issues. Thus, it is crucial to accurately detect abnormal flows to maintain the availability and security of the network. In this work, we propose hybrid deep learning algorithms, which are the long short-term memory network (LSTM) and convolutional neural network (CNN) with a stack autoencoder for DDoS attack detection and checkpoint network, which is a fault tolerance strategy for long-running processes. The proposed approach is trained and tested with the aid of two DDoS attack datasets in the SDN environment: the DDoS attack SDN dataset and Botnet dataset. The results show that the proposed model achieves a very high accuracy, reaching 99.99% in training, 99.92% in validation, and 100% in precision, recall, and F1 score with the DDoS attack SDN dataset. Also, it achieves 100% in all metrics with the Botnet dataset. Experimental results reveal that our proposed model has a high feature extraction ability and high performance in detecting attacks. All performance metrics indicate that the proposed approach is appropriate for a real-world flow detection environment.
|
## future internet
_Article_
# An Improved Deep Learning Model for DDoS Detection Based on Hybrid Stacked Autoencoder and Checkpoint Network
**Amthal K. Mousa * and Mohammed Najm Abdullah**
Computer Engineering Department, University of Technology-Iraq, Baghdad P.O. Box 10071, Iraq;
mohammed.n.abdullah@uotechnology.edu.iq
*** Correspondence: amthal.k.mousa@uotechnology.edu.iq**
**Citation: Mousa, A.K.; Abdullah,**
M.N. An Improved Deep Learning
Model for DDoS Detection Based on
Hybrid Stacked Autoencoder and
Checkpoint Network. Future Internet
**[2023, 15, 278. https://doi.org/](https://doi.org/10.3390/fi15080278)**
[10.3390/fi15080278](https://doi.org/10.3390/fi15080278)
Academic Editor: Izzat Alsmadi
Received: 20 July 2023
Revised: 11 August 2023
Accepted: 17 August 2023
Published: 19 August 2023
**Copyright:** © 2023 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
[Attribution (CC BY) license (https://](https://creativecommons.org/licenses/by/4.0/)
[creativecommons.org/licenses/by/](https://creativecommons.org/licenses/by/4.0/)
4.0/).
**Abstract: The software defined network (SDN) collects network traffic data and proactively manages**
networks. SDN’s programmability makes it excellent for developing distributed applications, cybersecurity, and decentralized network control in multitenant data centers. This exceptional architecture
is vulnerable to security concerns, such as distributed denial of service (DDoS) attacks. DDoS attacks
can be very serious due to the fact that they prevent authentic users from accessing, temporarily or
indefinitely, resources they would normally expect to have. Moreover, there are continuous efforts
from attackers to produce new techniques to avoid detection. Furthermore, many existing DDoS
detection methods now in use have a high potential for producing false positives. This motivates us
to provide an overview of the research studies that have already been conducted in this area and point
out the strengths and weaknesses of each of those approaches. Hence, adopting an optimal detection
method is necessary to overcome these issues. Thus, it is crucial to accurately detect abnormal flows
to maintain the availability and security of the network. In this work, we propose hybrid deep
learning algorithms, which are the long short-term memory network (LSTM) and convolutional
neural network (CNN) with a stack autoencoder for DDoS attack detection and checkpoint network,
which is a fault tolerance strategy for long-running processes. The proposed approach is trained
and tested with the aid of two DDoS attack datasets in the SDN environment: the DDoS attack SDN
dataset and Botnet dataset. The results show that the proposed model achieves a very high accuracy,
reaching 99.99% in training, 99.92% in validation, and 100% in precision, recall, and F1 score with the
DDoS attack SDN dataset. Also, it achieves 100% in all metrics with the Botnet dataset. Experimental
results reveal that our proposed model has a high feature extraction ability and high performance in
detecting attacks. All performance metrics indicate that the proposed approach is appropriate for a
real-world flow detection environment.
**Keywords: DDoS detection; distributed denial of service; software defined networking; SDN;**
network security
**1. Introduction**
Software defined networking, also known as SDN, is a novel approach to the networking paradigm which separates control decisions from the forwarding hardware. The
primary objective is to make it as simple as possible for software developers to rely on the
resources provided by the network for storage and computation [1]. The SDN comprises
switches that support open-flow, a controller, and a secure channel for the controller and
the switches [2]. SDN focuses on four main features [3]:
Separation of the data plane from the control plane.
_•_
A centralized management system and network perspective.
_•_
Open connections between the devices in the control plane and the data plane.
_•_
The network can be programmed by an outside administration.
_•_
SDN has two main assets, which are the centralization of control and the ability to
control the whole network through software. Those two assets are attractive features for
-----
_Future Internet 2023, 15, 278_ 2 of 16
attackers. Thus, several security challenges affect the SDN, including the distributed denial
of service attack (DDoS), man-in-the-middle attack, side channel attack, application manipulation, diversion of traffic, application exploitation, traffic sniffing, password guessing
or brute force, and network manipulation [4]. Recently, the DDoS attack has become one
of the most serious attacks due to the inability to access the controller. The process and
communication capacity of the controller are overloaded when DDoS attacks occur against
the SDN controller because of the unnecessary flow produced by the controller for the
attack packets. The capacity of the switch flow table becomes full, leading the network
performance to decline to a critical threshold [5]. Machine learning (ML) and powerful
deep learning (DL) are two of the most common techniques to protect any network from
DoS/DDoS attacks.
This work proposes a novel model of DL-based DDoS attack detection algorithms
in SDN, evaluates those efforts, and then compares those findings to the recent related
papers. The motivation of using a proposed model to find and stop DDoS attacks on SDN
is to give an overview of the research studies that have already been conducted in this
area and point out the strengths and weaknesses of each of those approaches. Also, DDoS
attacks are a big problem for SDN networks. Traditional methods of defense may not be
able to find and stop these attacks, because attackers now use new methods to flood SDN
using different types of traffic (high and low rates), that slow down the SDN controller
and make it inaccessible to legitimate users. Additionally, many recent DDoS detection
methods have a high potential for producing false alarms, which can be time-consuming to
analyze and cause alert fatigue. Consequently, techniques that can lessen false positives as
well as increase the accuracy of DDoS detection are required.
This study proposes a model-based CNN-LSTM as a stacked autoencoder with a
checkpoint network for DDoS detection to achieve high accuracy DDoS detection. We have
demonstrated how this particular structure can enhance performance, accurately estimate
attacks, and remarkably suppress false alarms. Additionally, we provide details of the
dataset and hyperparameter values. Furthermore, we produce a comparative analysis of
the proposed approach against some recently published work. The main contributions of
this work are as follows:
Propose a deep-stacked autoencoder-based CNN-LSTM for detecting DDoS attacks on a
_•_
network. This model can extract features effectively in an unsupervised learning approach.
Utilize a checkpoint network model: a fault tolerance strategy for long-running pro
_•_
cesses that permits the definition of checkpoints for the model weights at certain
locations and improves inference accuracy in real time.
After this introductory section of the paper, Section 2 provides a background of DDoS
attacks and the detection mechanism in the SDN. Section 3 presents an overview of the
most recent related studies for DDoS detection. The proposed system structure appears in
Section 4, and the experimental results of the proposed model classifiers for DDoS attack
detection in SDN appear in Section 5, along with a comparison to some relevant research.
The last topic of discussion in the paper is the conclusion in Section 6.
**2. Concept of SDN and the Detection Mechanism of DoS/DDoS Attacks**
The development of technology to detect and mitigate distributed denial of service
attacks in SDN environments [6] provides a significant obstacle to these attacks. A distributed denial of service attack sends many packets to the target network. Unmatched
flows are considered new if the target and source IP addresses of the forwarded packets are
fake, and switches cannot locate these packets in their flow table entries. Next, the switch
will forward the packet directly to the SDN controller or send the mismatched packet to the
SDN controller [7]. Finding the appropriate routes for these packets lies within the purview
of the SDN controller. Many disguised DDoS flows are in legitimate traffic. These flows
continually consume the controller’s resources, and as a result, those resources eventually
become unavailable for use by incoming packets. As a direct consequence of this attack, the
SDN controller goes offline, which causes the entire network to enter a downstate. Even if
-----
_Future Internet 2023, 15, 278_ 3 of 16
a backup controller is available, this security flaw still exists [8]. The characteristics of a
DDoS attack in a software-defined networking system are subtly distinct from those of an
attack on a traditional network. The following is a conclusion reached after researching the
DDoS attack techniques utilized against the SDN controller [9]:
In traditional networks, there are one or more network links, and DDoS attackers go
_•_
after servers that are the endpoints. In SDN, the controller is hit with a DDoS attack. In
SDN, the main goal of a DDoS attack is to make the controller’s resources unavailable
by failing at a single point.
_•_ The IP addresses of packets in traditional networks are real. As a result, DDoS attackers
typically target the terminal server. To conduct a DDoS assault in SDN, the attacker
attempts to counterfeit the IP addresses of the destination, involving the controller in
constant processing with fresh flows. The controller’s resources are made unavailable.
_•_ In traditional networks, when a DDoS attack occurs, the server stops providing services
to actual users. But in SDN, if the DDoS attack occurs, the controller in the SDN loses
contact with the data plane and cannot provide services for moving data packets.
Traditional ways to find DDoS attacks use a stochastic analysis and the randomness
of network traffic to find unusual intrusions. When the detection software finds an attack
event, traffic rate-limiting and filtering are used to lessen the damage. But if it uses
mitigation strategies carelessly, they will affect legitimate traffic. Even though the victim is
not receiving a lot of traffic, a poor response like this can make it difficult for regular users
to get online. So, the detection technique must be capable of determining when a DDoS
attack occurs and distinguishing between attack traffic and normal traffic.
The current trend in DDoS detection is to use machine learning to classify and detect
malicious traffic. These techniques can learn the attributes of the underlying data smartly
without needing to be told what is normal and what is dangerous. Even though machine
learning-based techniques show promise, most focus on offline traffic analysis and have
trouble staying current with how DDoS attacks change over time [10]. Lastly, the detection
method should try to reduce false alarms, which can hurt sources that are not doing
anything wrong. So, the defense system stops attack traffic and ensures that legitimate
traffic gets to the end users reliably [11].
**3. Related Works**
Recent DDoS detection research utilizing machine learning approaches has achieved
promising results. These systems can intelligently understand the underlying data properties without explicitly specifying normal and harmful behaviors, bypassing the limits
of conventional detection schemes. The DDoS detection problem is a binary classification
problem in which the observed traffic is either normal or attack traffic. Moreover, detection
techniques have used deep learning more often in recent years to find DDoS attacks and
presented several approaches. Of various recent notable works in this field, some utilized
convolutional networks, some utilized recurrent neural networks, especially LSTM and
bidirectional LSTM, and some used autoencoder, an unsupervised learning approach, to discover non-linear characterizations from input data, and would then perform a classification
algorithm to differentiate malicious traffic from genuine traffic.
In 2017, Yuan, Li, and Li [12] developed a deep learning algorithm named “DeepDefense”, a model that uses a deep learning model to detect DDoS attacks. To carry out their
research, they used CNN, as well as several distinct variants of RNN (such as LSTM and
the gated recurrent unit neural network (GRUNN)), and the random forest (RF) method.
The study included a comparison analysis between several deep learning methodologies
and between deep learning and machine learning algorithms (it selected RF). DeepDefense
put into action four deep learning models: LSTM, CNN-LSTM, GRU, and 3-LSTIM. We
compared the results of these models with one another. With an accuracy of 98.410% and
an area under the curve (AUC) score of 99.450%, the top deep learning model, 3LSTM,
was able to identify DDoS attacks. Shone et al. [13] found that stacking two autoencoders
allowed for the learning of more complex feature-based correlations. For intrusion detec
-----
_Future Internet 2023, 15, 278_ 4 of 16
tion, they combined the stacked autoencoder with a random forest classifier. They asserted
that the soft-max layer was less effective than traditional classifiers. In 2019, Pekta¸s and
Acarman [14] presented a model-based deep learning method that utilized CNN and LSTM
to train the spatial-temporal characteristics of network flows. It used two datasets for
training and testing: the ISCX2012 dataset [15] and CICIDS2017 [16]. The results show
that the model achieved 0.9669 in precision, 0.9649 in recall, 0.9657 in F1-score, and 0.9666
in accuracy. The model also returned good results when using CI-CIDS2017, where it
achieved 0.9797 in precision, 0.9765 in recall, 0.9780 in F1-score, and 0.9772 in accuracy.
Some studies developed hybrid deep learning models. Gadze et al., 2021 [17] proposed
a model that combined two types of deep learning, LSTM and CNN, to detect an attack.
Mininet generated the dataset dynamically and utilized OpenFlow switches and Floodlight
as an external controller. Based on the findings, RNN LSTM outperformed linear-based
models like SVM (86.85%) and Naive Bayes (82.61%), achieving an accuracy of 89.63%
compared to their respective scores. Their model had an accuracy of 99.4%, while the
KNN technique, based on linear models, had an even higher accuracy. In addition, the
model functioned most effectively when it split the data in a 70/30 train/test split ratio.
Singh and Jang-Jaccard (2022) [18] created a hybrid autoencoder model dubbed MSCNNLSTM-AE. This model found anomalies in network traffic by utilizing a combination of a
multi-scale convolutional neural network (MSCNN) and LSTM. The MSCNN autoencoder
was employed initially to evaluate the spatial characteristics of the dataset. Next, it used
an LSTM-based autoencoder network to identify the temporal features of the latent space
features learned from the MSCNN-AE. The authors analyzed their work with the UNSWNB15 [19], NSL-KDD [20], and CICDDoS2019 tests. The accuracy score for their model
(MSCNN-LSTM-AE) came in at 93.76%, while the recall score was 92.26%. Elubeyd and
Yiltas-Kaplan [21] presented a hybrid deep learning approach for detecting and countering
DoS/DDoS attacks in SDNs. The selection of a hybrid model that included a 1D CNN, a
dense neural network (DNN), and a gated recurrent unit (GRU) took advantage of their
individual strengths that synergistically addressed the intricacies of the problem. The
model achieved good results when using CICDDoS 2019, where it achieved 0.9981 in
accuracy, 0.9996 in precision, 0.999 in recall, and 0.9993 in F1-score.
Some recent studies used a stacked autoencoder to improve DDoS detection accuracy.
Yaser et al., 2022 [22] proposed a novel approach for detecting DDoS attacks, which involved
integrating deep learning with feedforward neural networks in the form of autoencoders.
The training and evaluation of the model were analyzed using two datasets, initially
through a static approach and subsequently through an iterative technique. They developed
the autoencoding model through a layer-by-layer stacking of the input layer and the hidden
layer of self-encoding models, wherein each self-encoding model employed a hidden layer.
They assessed the performance of their model by employing a three-fold data partitioning
strategy comprising training, testing, and validating subsets. The test result showed that
the model yielded superior accuracy for the static dataset. Specifically, for the ISCXIDS-2012
dataset, the model attained a maximum accuracy of 99.35% during training, 99.3% during
validation, 99.78% for precision, 99.99% for recall, and 99.87 for F1-score. The UNSW-2018
dataset exhibited high levels of accuracy during training, with values of 99.95% for training
and 99.94 for validation, and 99.99 for recall, precision, and F1-score. Jiang et al., 2018 [23],
presented a new method (DLGraph) for detecting malware based on deep learning along
with graph embedding. Their architecture for deep learning was comprised of two stacked
denoising autoencoders (SDA). One SDA was able to learn the latent structure of functioncall graphs in programs. The other SDA was capable of learning a latent representation
of Windows API calls made by programs. They utilized the node2vec technique when
incorporating a function-call graph in a feature space. The experimental results on three
distinct datasets demonstrated that the proposed DLGraph method obtained high levels
of accuracy and exceeded the closely related DL4MD method, where it gained 99.14% in
accuracy for dataset 1, 99.36% for dataset 2, and 99.31 for dataset 3. Table 1 shows the
comparison between these related works.
-----
_Future Internet 2023, 15, 278_ 5 of 16
**Table 1. Comparison of related works in terms of methods, performance measures, and achievement.**
**Ref** **Model** **Achievement**
3-LSTIM outperformed the other models which gained
Yuan, Li [12] LSTM, GRU, CNN-LSTM, and 3-LSTIM
99.450% in accuracy
combined the stacked autoencoder with a They asserted that the soft-max layer was less effective
Shone et al. [13]
random forest classifier than traditional classifiers
For the ISCX2012 dataset, the model achieved 0.9669 in
precision, 0.9649 in recall, 0.9657 in F1-score, and 0.9666
Pekta¸s and Acarman [14] LSTM and CNN in accuracy. For CI-CIDS2017, the model achieved 0.9797
in precision, 0.9765 in recall, 0.9780 in F1-score, and
0.9772 in accuracy
The model outperformed the other ML models, in which
it gained 99.4% in accuracy compared with RNN LSTM
Gadze et al., 2021 [17] LSTM and CNN
that archived 89.63%, SVM achieved 86.85%, and Naive
Bayes achieved 82.61%
Hybrid autoencoder model dubbed The accuracy score was 93.76% and the recall
Singh and Jang-Jaccard, 2022 [18]
MSCNN-LSTM-AE score was 92.26%
Hybrid deep learning approaches (1D CNN, a The model achieved good results when using CICDDoS
Elubeyd and Yiltas-Kaplan [21] dense neural network (DNN), and a gated 2019, where it achieved 0.9981 in accuracy, 0.9996 in
_uture Internet 2023, 15, x FOR PEER REVIEW_ recurrent unit (GRU)) precision, 0.999 in recall, and 0.9993 in F1-score
For the ISCXIDS-2012 dataset, the model attained a
maximum accuracy of 99.35% during training, 99.3%
during validation, 99.78% for precision, 99.99% for recall,
Yaser et al., 2022 [22] LSTM-Autoencoder
and 99.87 for F1-score. For UNSW-2018, it gained 99.95%
for training accuracy and 99.94 for validation accuracy,
#### 4. Proposed Model Structure and 99.99 for recall, precision, and F1-score
DLGraph based on two stacked denoising It gained 99.14% in accuracy for dataset 1, 99.36% for
Jiang et al., 2018 [23]
#### Our approach uses autoencoders, a method that is now popular in deepautoencoders (SDA) dataset 2, and 99.31 for dataset 3
autoencoder is an unsupervised neural network-based feature extraction learns the best feasible factors to reproduce faithfully its output given some 4. Proposed Model Structure
Our approach uses autoencoders, a method that is now popular in deep learning.
#### its many appealing features is its potential to provide a non-linear and more
An autoencoder is an unsupervised neural network-based feature extraction method that
#### eralization than the principal component analysis (PCA). It achieves this relearns the best feasible factors to reproduce faithfully its output given some input. One propagation with input-equivalent target values. To rephrase, it tries to figuof its many appealing features is its potential to provide a non-linear and more efficient
generalization than the principal component analysis (PCA). It achieves this result by
#### predict the occurrence of itself as closely as possible. The typical architectur
backpropagation with input-equivalent target values. To rephrase, it tries to figure out
#### encoder consists of three layers: an input layer, an output layer, and a hiddhow to predict the occurrence of itself as closely as possible. The typical architecture of an hidden layer’s dimensions are lower than that of the input [24]. Figure 1 shoautoencoder consists of three layers: an input layer, an output layer, and a hidden layer.
The hidden layer’s dimensions are lower than that of the input [24]. Figure 1 shows the
#### tional (single) autoencoders.
traditional (single) autoencoders.
##### Figure 1. Figure 1. Single autoencoder [Single autoencoder [24]. 24].
-----
_Future Internet 2023, 15, 278_ 6 of 16
_Future Internet 2023, 15, x FOR PEER REVIEW In the proposed method, we utilize a deep autoencoder. Unlike traditional autoen-7 of 17_
coders, deep autoencoders consist of two typical deep-belief networks, one for encoding
and one for decoding, with four or five shallow layers each. Deep learning can be applied to
autoencoders by a stacked autoencoder, in which many hidden layers build depth, and the
hidden layers reflect fundamental concepts. As a result of this increased depth, computing𝑦𝑖= 𝑠(𝑊𝑥𝑖 + 𝑏) (1)
costs will reduce, the amount of instruction data required will decrease, and accuracy willThrough
improve. The output of one buried layer serves as the input to a later, more advanced
𝑧𝑖 = 𝑠(𝑊[′]𝑦𝑖+ 𝑏[′]) (2)
step. First-order features are often learned from unprocessed data by the first layer of a
stacked autoencoder. Second-order features based on trends in the presence of first-orderThis concealed representation is mapped back into a reconstruction of the same shape
as input x. Here, s represents a non-linear function, such as the sigmoid function. The first traits are typically learned by the second layer. Subsequent layers build our understandcomponent is the encoder, while the second is the decoder. This model’s parameters min-ing of higher-order characteristics. Figure 2 shows the structure of the proposed deep
imize the average reconstruction error. autoencoder model.
**Figure 2. Figure 2. The general structure of the proposed deep autoencoder.The general structure of the proposed deep autoencoder.**
The model consists of one input layer, one convolutional layer (Conv1D), two LSTM The first layer is the input layer, which receives input Xi and uses numerous hidden
layers, one max pooling layer, and one dense layer in output. Figure 3 shows the training layers to encode and decode it (encoder and decoder blocks). The encoding process
model with the proposed deep autoencoder scheme.compresses the attributes to make them smaller than the input data, and the decoding
process restores these attributes in reverse order to begin the final output at the deepest
layer. When processed, the output feature vector Xi is virtually identical to the input. The
convolutional layer and LSTM are combined with an autoencoder to generate a robust
DDoS attack classifier. LSTM is excellent at understanding the context of Internet packets,
identifying long- and short-term dependencies, and identifying trends in DDoS attack
sequences. LSTM is particularly proficient at categorizing processes such as time series and
learning from experience. After the encoding is complete, based on the output result of the
hidden layer, the output layer is decoded and reconstructed according to Equation (2) to
produce an output of the same size as the input layer neuron.
The purpose of the autoencoder section is to map input x [0, 1][d] to a latent represen_∈_
tation y [0, 1]d[′], where the mapping is performed by the function
_∈_
_yi = s(Wxi + b)_ (1)
Through
_zi = s�W[′]yi + b[′][�]_ (2)
This concealed representation is mapped back into a reconstruction of the same shape
as input x. Here, s represents a non-linear function, such as the sigmoid function. The
**Figure 3. The training model of the proposed deep autoencoder.**
first component is the encoder, while the second is the decoder. This model’s parameters
minimize the average reconstruction error.
In addition, the checkpoint network improves the weights. Checkpointing is a crucial
The model consists of one input layer, one convolutional layer (Conv1D), two LSTM
functionality that expedites failure recovery, reducing the total training time and ensuring
continuous progress. Checkpoints are periodic captures of the current state of a running layers, one max pooling layer, and one dense layer in output. Figure 3 shows the training
process, which are then stored in a durable storage medium. The individual loads the model with the proposed deep autoencoder scheme.
most recent checkpoint to recover from a setback and recommence training. In addition
tation y [0, 1]d[′]
_∈_
Through
-----
### y, p g y, y p g
_Future Internet 2023, 15, 278_ model with the proposed deep autoencoder scheme. 7 of 16
**Figure 3. The training model of the proposed deep autoencoder.**
##### Figure 3. The training model of the proposed deep autoencoder.
In addition, the checkpoint network improves the weights. Checkpointing is a crucial
functionality that expedites failure recovery, reducing the total training time and ensuring
_Future Internet 2023, 15, x FOR PEER REVIEW In addition, the checkpoint network improves the weights. Checkpointi8 of 17_
continuous progress. Checkpoints are periodic captures of the current state of a running
### functionality that expedites failure recovery, reducing the total training timeprocess, which are then stored in a durable storage medium. The individual loads the most continuous progress. Checkpoints are periodic captures of the current staterecent checkpoint to recover from a setback and recommence training. In addition to the
to the imperative of failure recovery, the utilization of checkpoints is necessary for trans
imperative of failure recovery, the utilization of checkpoints is necessary for transferring
### process, which are then stored in a durable storage medium. The individtraining processes across various nodes or clusters. This transition may be necessary forferring training processes across various nodes or clusters. This transition may be neces
sary for server maintenance (for instance, urgent security updates that cannot be delayed),
### most recent checkpoint to recover from a setback and recommence traininserver maintenance (for instance, urgent security updates that cannot be delayed), hardware
malfunctions, network complications, and the optimization or reallocation of resources.hardware malfunctions, network complications, and the optimization or reallocation of
Another significant application of checkpoints involves the real-time publication of snapsresources. Another significant application of checkpoints involves the real-time publicaof trained models to enhance the accuracy of inference, commonly referred to as onlinetion of snaps of trained models to enhance the accuracy of inference, commonly referred
training. For example, we can employ an interim model obtained by checkpointing forto as online training. For example, we can employ an interim model obtained by checkprediction serving. This method allows the model to continue training on more recentpointing for prediction serving. This method allows the model to continue training on
datasets, ensuring the freshness of the inference model. We can also utilize checkpointsmore recent datasets, ensuring the freshness of the inference model. We can also utilize
for transfer learning, a technique where an intermediate structure state is an initial pointcheckpoints for transfer learning, a technique where an intermediate structure state is an
for training toward a distinct objective [initial point for training toward a distinct objective [10]. Figure 4 shows the training loop 10]. Figure 4 shows the training loop with a
checkpoint network.with a checkpoint network.
**Figure 4. The training looping with checkpoint network [10].**
**Figure 4. The training looping with checkpoint network [10].**
The evaluation of the model occurs at the conclusion of each epoch and the weights
-----
_Future Internet 2023, 15, 278_ 8 of 16
The evaluation of the model occurs at the conclusion of each epoch, and the weights
corresponding to the highest accuracy and lowest loss during that specific epoch are
retained and saved. In the event that the weights in the model during a specific epoch
fail to yield the optimal accuracy or loss, as determined by the user-defined criteria, the
weights will not be preserved. However, the training process will persist, commencing
from the aforementioned condition.
**5. Results and Discussion**
This section discusses the experimental results of our proposed method. We use
several performance metrics for evaluation. We use two datasets to test the performance of the proposed model in detecting DDoS attacks. Then, we compare these results
with some recent related works using the same datasets and with some other machine
learning algorithms.
_5.1. Datasets_
Most datasets are imperfect, and the row samples employed to cover the application manners are insufficient in these cases. The most common DDoS datasets involve
CI-CIDS2019, CICIDS2017, KDDCUP99, ISCX2012, Kyoto 2006+, and NSL-KDD, which
researchers have significantly utilized for intrusion detection. In this work, we chose two
datasets to validate our proposed DDoS classifier, which are:
1- DDoS attack SDN dataset (Mendeley Data): This set is an SDN-specific dataset created by the Mininet emulator and utilized by machine learning and deep learning
algorithms for traffic classification. The project begins by constructing 10 Mininet
topologies with switches connected to a single Ryu controller. It simulates a network
for benign TCP, UDP, and ICMP traffic and malicious traffic, which consists of a
TCP Syn attack, UDP Flood attack, and ICMP assault. The data collection contains
23 features, of which some are from the switches, and others are calculated, such as
the Packet count, Switch-id, duration sec, byte count, Destination IP, Source IP, Port
number, etc.
2- Botnet dataset (UNSW_2018_IoT_Botnet): Even though several datasets have been
proposed for detecting intrusions, most datasets are not updated and do not reflect
actual data. The Canadian Institute for Cybersecurity addressed these issues by
developing the Intrusion Detection Evaluation Dataset, ISCX-IDS 2012, and [25]
generated by monitoring network activity for seven days. The labeled dataset consists
of approximately 1,512,000,000 packets with 20 features. The primary characteristics
of this dataset are discussed in [25] and include real, normal, and malicious streams
comprising FTP, HTTP, IMAP, POP3, SMTP, and SSH protocols collected using real
devices. All data are categorized and marked. The collected datasets contain a variety
of intrusion kinds (Infiltrating, DoS, DDoS, and Brute Force SSH).
_5.2. Performance Evaluation Metrics_
A performance evaluation is the process of measuring the of a classification model
after assigning cases to their various predetermined labels. The performance evaluation
considers measures including accuracy, recall, precision, F1-score, and confusion matrix [26].
These metrics are described as follows:
1. Accuracy: The ratio of all correct predictions over the total number of packets in the
dataset [24]:
Tp + Tn
Accuracy(Ac) = (3)
Tp + Tn + Fp + Fn
where Tp is the “True Positive,” which describes the rate where the actual instance of a
certain label is categorized as that label; Tn is the “True Negative,” which is the attack
instance’s value that is classified as an attack; Fp is the “False Positive,” which is the
number incorrectly classified for a certain class label, i.e., the instance categorized
-----
_Future Internet 2023, 15, 278_ 9 of 16
value as an additional class label for a given dataset; Fn is the “False Negative,” which
is the value of normal traffic which is classified as an attack.
2. Recall: Recall is the number of correctly predicted positive records over all the positive
records: a metric that can detect DDoS attack traffic compared to normal traffic [24].
Tp
Recall = (4)
Tp + Fn
3. Precision: Precision is the proportion of actual positive instances that were correctly
predicted, i.e., it is a metric that can detect DDoS attack traffic among normal traffic [24].
Tp
Precision = (5)
Tp + Fp
4. F1-score: The F1-score is the balance between the recall and the precision [24].
F1 score = 2 [Recall][ ×][ precision] (6)
_−_ _×_
Recall + precision
5. Confusion Matrix: The data classification results appear in a table format. The
accuracy of a classification model is evaluated by applying it to test data for which
the results have already been determined. The study uses it to show the distribution
of the expected outcomes, despite its poor suitability for anything beyond binary
_Future Internet 2023, 15, x FOR PEER REVIEW classification [24]._ 10 of 17
_5.3. Results and Discussion of Base Classifier Models_
5.3.1. Mendeley DDoS Attack SDN Dataset
The dataset utilized in this study is a software defined networking (SDN) dataset
The dataset utilized in this study is a software defined networking (SDN) dataset
generated by implementing ten distinct topologies within the Mininet framework,
generated by implementing ten distinct topologies within the Mininet framework, wherein
wherein switches are interconnected to a singular Ryu controller. The network simulation
switches are interconnected to a singular Ryu controller. The network simulation encom
encompasses benign traffic, including TCP, UDP, and ICMP, as well as malicious traffic,
passes benign traffic, including TCP, UDP, and ICMP, as well as malicious traffic, which
which comprises TCP Syn, UDP Flood, and ICMP attacks. The dataset contains 23 fea
comprises TCP Syn, UDP Flood, and ICMP attacks. The dataset contains 23 features, in
tures, including extracted data from switches and calculated variables. At first, we extract
cluding extracted data from switches and calculated variables. At first, we extract packet
packet fields from the DDoS attack SDN dataset. The number of samples used is 100000.
fields from the DDoS attack SDN dataset. The number of samples used is 100,000. The
The extracted feature appears in Figure 5.
extracted feature appears in Figure 5.
**Figure 5. Extracted feature from the original packets.**
**Figure 5. Extracted feature from the original packets.**
The extracted features are Packet_count, Switch-id, byte_count, duration_sec (repre
The extracted features are Packet_count, Switch-id, byte_count, duration_sec (repre
senting the duration in seconds), duration_nsec (representing the duration in nanoseconds),
senting the duration in seconds), duration_nsec (representing the duration in nanosec
and the overall duration obtained by summing duration_sec and duration_nsec. Addi
onds), and the overall duration obtained by summing duration_sec and duration_nsec.
tionally, the characteristics contain Source IP and Destination IP. The numerical identifier
Additionally, the characteristics contain Source IP and Destination IP. The numerical iden
assigned to a specific communication endpoint within a computer network is commonly
tifier assigned to a specific communication endpoint within a computer network is com
referred to as a port number. The variable “tx_bytes” represents the quantity of bytes
monly referred to as a port number. The variable “tx_bytes” represents the quantity of
that have been moved via the switch port, whereas “rx_bytes” denotes the quantity of
bytes that have been moved via the switch port, whereas “rx_bytes” denotes the quantity
bytes that have been received on the switch port. The “dt” field represents the numerical
of bytes that have been received on the switch port. The “dt” field represents the numeri
-----
_Future Internet 2023, 15, 278_ 10 of 16
representation of both the date and time. This field is utilized to monitor the flow of a
particular process at regular intervals of 30 s. The calculated features encompass the term
“packet per flow”, which refers to the count of packets transmitted during a single flow.
Similarly, “byte per flow” represents the count of bytes transmitted during a single flow.
“Packet Rate” denotes the number of packets sent per second and may be computed by
splitting the packet for each flow by the monitoring interval. Additionally, the number of
“Packet_ins” messages and the total flow inputs in the switch are relevant factors in this
context. The variables tx_kbps and rx_kbps represent the rates at which data are transferred and received, respectively. Port bandwidth refers to the cumulative value of both
the transmitted kilobits per second (tx_kbps) and received kilobits per second (rx_kbps).
The final column denotes the class label, which serves as an indicator to determine if
the traffic class is normal or malicious. In the classification scheme utilized, benign traffic is
assigned a label of 0, whereas malicious traffic is assigned a label of 1. A network simulation
is conducted over a duration of 250 min, resulting in the collection of 104,345 rows of data.
The simulation is executed repeatedly within a specified time frame, allowing for the
accumulation of further data.
We split the dataset into 80% for training and 20% for testing. Attack traffic is labeled
with 1, whereas the normal traffic is labeled with 0, and we train the model for 500 epochs
to study the effect of the checkpoint strategy on improving classification accuracy. The
classification result appears in Table 2 and Figures 6 and 7.
**Table 2. Results of tests using the Mendeley dataset of the proposed model.**
**Precision (%)** **Recall (%)** **F1-Score (%)**
**Accuracy (%)** **Validation Accuracy (%)**
**Normal** **Attack** **Normal** **Attack** **Normal** **Attack**
_Future Internet 99.99_ **2023, 15, x FOR PEER REVIEW 99.923** 100 100 100 100 100 10011 of 17
_Future Internet 2023, 15, x FOR PEER REVIEW_ 11 of 17
where Normal is the normal traffic, and Attack is DDoS attack traffic.
(a) (b)
(a) (b)
**Figure 6. A screenshot of the output terminal (a) training and validation accuracy results; (b) metrics**
**Figure 6. Figure 6.A screenshot of the output terminal (A screenshot of the output terminal (a) training and validation accuracy results; (a) training and validation accuracy results;b) metrics**
results.
results. (b) metrics results.
(a) (b) (c)
(a) (b) (c)
**Figure 7. Classification results of the CNN-LSTM-autoencoder model. (a) Accuracy results of train-**
**Figure 7. Figure 7. Classification results of the CNN-LSTM-autoencoder model. (Classification results of the CNN-LSTM-autoencoder model. (a) Accuracy results of traininga) Accuracy results of train-**
ing and validation per epoch, (b) loss per epoch, and (c) confusion matrix.
ing and validation per epoch, (b) loss per epoch, and (c) confusion matrix.
and validation per epoch, (b) loss per epoch, and (c) confusion matrix.
**Table 2. Results of tests using the Mendeley dataset of the proposed model.**
**Table 2. Results of tests using the Mendeley dataset of the proposed model.**
-----
_Future Internet 2023, 15, 278_ 11 of 16
Table 2 and Figure 6 show the final results of training and Figure 7a, b represent
the training/test loss and training/test accuracy after 500 epochs, respectively. Figure 7c
shows the confusion matrix of the proposed model. These results prove that the model
achieves very high classification results and stability with close results between training
and validation, where it gains 99.99% in training accuracy, 99.923% in validation accuracy,
and gain 100% in precision, recall, and F1-score. From Figure 7a, the model training and
validation loss are very low, about 3.98 10[−][4] for training and 3.3 10[−][3] for validation.
_×_ _×_
From Figure 7b, the model reaches the best train/validation accuracy at epoch (19). From
Figure 7c, the model achieves a significant degree of prediction accuracy. It reaches an
accuracy of about 100% in correctly detecting attacks and normal traffic flows. Moreover,
the proposed model has fewer false alarms since it shows a False Positive Rate (FPR) of
0.00086 and a False Negative Rate (FNR) of 0.00071. These results show the importance of
using a checkpoint network and many epochs, which can highly affect accuracy, as shown
in Figure 6.
_nternet 2023, 15, x FOR PEER REVIEW_
5.3.2. UNSW_2018_IoT_Botnet Dataset
The Bot-IoT dataset was developed in 2018 and published in 2019 by the New South
Wales University (UNSW). It is a contemporary and authentic dataset for training machine
learning models to effectively identify and mitigate Botnet attacks within Internet of Things
#### attacks, such as OS and Service Scan, DoS, DDoS, Data exfiltration, and Keyloggi(IoT) networks. The dataset comprises 72 million instances, consisting of three dependent ditionally, the DoS and DDoS attacks are further categorized based on the specifiand forty-three independent features. The dataset encompasses various cyber-attacks, such
as OS and Service Scan, DoS, DDoS, Data exfiltration, and Keylogging. Additionally, the
#### col. At first, we extract packet fields from the DDoS attack SDN dataset. The nu
DoS and DDoS attacks are further categorized based on the specific protocol. At first, we
#### samples used is 100000. The extracted feature appears in Figure 8. extract packet fields from the DDoS attack SDN dataset. The number of samples used is
100000. The extracted feature appears in Figure 8.
**Figure 8. Extracted feature from the original UNSW 2018 dataset packets.**
##### Figure 8. Extracted feature from the original UNSW 2018 dataset packets.
The UNSW 2018 dataset involves pkSeqID to represent the row identifier, Proto is
the representation of textual for transaction protocols that are resent in the network flow,
#### The UNSW 2018 dataset involves pkSeqID to represent the row identifier,
saddr is the IP address of the source, sport is the port number of the source, daddr is
#### the representation of textual for transaction protocols that are resent in the netwothe IP address of the destination, dport is the port number of the destination, seq is the saddr is the IP address of the source, sport is the port number of the source, daddargus sequence number, stddev is the aggregated records of the standard deviation, min
is the minimum duration of the standard deviation, state number represents the feature
#### IP address of the destination, dport is the port number of the destination, seq is th
state numerical representation, mean is the aggregated records of the average deviation,
#### sequence number, stddev is the aggregated records of the standard deviation, mdrate is the packets per second of destination-to-source, srate is the packets per second of minimum duration of the standard deviation, state number represents the featusource-to-destination, max is the aggregated records’ maximum duration, attack is the class
label where 0 represents normal traffic and 1 represents the attack traffic, category is the
#### numerical representation, mean is the aggregated records of the average deviatio
category of traffic, subcategory is the subcategory of traffic, and dbytes is the byte count of
#### is the packets per second of destination-to-source, srate is the packets per sedestination-to-source. source-to-destination, max is the aggregated records’ maximum duration, attacHence, like the previous dataset, we split the data into 80% for training and 20% for
testing. Attack traffic is labeled 1, the normal traffic is labeled 0, and we train the model for
#### class label where 0 represents normal traffic and 1 represents the attack traffic, cat
500 epochs. The classification results appear in Table 3 and Figures 9 and 10.
#### the category of traffic, subcategory is the subcategory of traffic, and dbytes is t count of destination to source
-----
count of destination to source.
_Future Internet 2023, 15, 278_ Hence, like the previous dataset, we split the data into 80% for training and 20% for 12 of 16
testing. Attack traffic is labeled 1, the normal traffic is labeled 0, and we train the model
for 500 epochs. The classification results appear in Table 3 and Figures 9 and 10.
**Table 3. Results of tests using the UNSW 2018 dataset of the proposed model.**
**Table 3. Results of tests using the UNSW 2018 dataset of the proposed model.**
**Precision (%)** **Recall (%)** **F1-Score (%)**
**Accuracy (%)** **Validation Accuracy (%)** **Precision (%)** **Recall (%)** **F1-Score (%)**
**Accuracy (%)** **Normal[Validation ]** **Attack** **Normal** **Attack** **Normal** **Attack**
**Accuracy (%)** **Normal** **Attack** **Normal** **Attack** **Normal** **Attack**
100 100 100 100100 100100 100 100 100 100 100 100100 100100
where Normal is the normal traffic, and Attack is DDoS attack traffic.where Normal is the normal traffic, and Attack is DDoS attack traffic.
_Future Internet 2023, 15, x FOR PEER REVIEW_ 13 of 17
(a) (b)
_Future Internet 2023, 15, x FOR PEER REVIEW Figure 9. Screenshot of the output terminal showing (a) training and validation accuracy results; (13 of 17 b)_
**Figure 9. Screenshot of the output terminal showing (a) training and validation accuracy results;**
metrics results.
(b) metrics results.
###### (a) (b) (c)
metrics results.
(b) metrics results.
###### (b)
(a) (b) (c)
**Figure 10. Classification results of the CNN-LSTM-autoencoder model. (a) Accuracy results of train-**
ing and validation per epoch, (Figure 10. Classification results of the CNN-LSTM-autoencoder model. (b) loss per epoch, and (c) confusion matrix. a) Accuracy results of train
**Figure 10. Classification results of the CNN-LSTM-autoencoder model. (a) Accuracy results of**
ing and validation per epoch, (b) loss per epoch, and (c) confusion matrix.
training and validation per epoch, (b) loss per epoch, and (c) confusion matrix.
###### Table 3 and Figure 9 show the final results of training and Figure 10a,b represent the
training/test loss and training/test accuracy after 500 epochs, respectively. Figure 10c Table 3 and Figure 9 show the final results of training and FigureTable 3 and Figure 9 show the final results of training and Figure 1010a,b represent the a,b represent the shows the confusion matrix of the proposed model. These results prove that the model training/test loss and training/test accuracy after 500 epochs, respectively. Figure 10c training/test loss and training/test accuracy after 500 epochs, respectively. Figure 10c achieves very high classification results and stability with close results between training shows the confusion matrix of the proposed model. These results prove that the model shows the confusion matrix of the proposed model. These results prove that the model
achieves very high classification results and stability with close results between training achieves very high classification results and stability with close results between training
###### and validation, where it gains 100% in all metrics. From Figure 10a, the model training
and validation, where it gains 100% in all metrics. From Figure 10a, the model training and validation, where it gains 100% in all metrics. From Figure 10a, the model training
###### and validation loss are very low, about 3.98 × 10[−][4] for training and 3.3 × 10[−][3 ]for validation.
and validation loss are very low, about 3.98 × 10and validation loss are very low, about 3.98 10[−][4][−] for training and 3.3 × 10[4] for training and 3.3 _[−] 10[3 ]for validation. [−][3]_ for valida
_×_ _×_
###### From Figure 10b, the model reaches the best train/validation accuracy at epoch 19, which
From Figure 10b, the model reaches the best train/validation accuracy at epoch 19, which tion. From Figure 10b, the model reaches the best train/validation accuracy at epoch 19,
###### shows the importance of using a checkpoint network and many epochs, which can signif
shows the importance of using a checkpoint network and many epochs, which can signif-which shows the importance of using a checkpoint network and many epochs, which can
###### icantly affect accuracy, as shown in Figure 11. icantly affect accuracy, as shown in Figure 11. significantly affect accuracy, as shown in Figure 11.
**Figure 11. Best accuracy results of training and validation at epoch 19.**
**Figure 11. Best accuracy results of training and validation at epoch 19.**
**Figure 11. Best accuracy results of training and validation at epoch 19.**
###### (a)
For further checking, we utilize the standard deviation metrics to quantify the extent
###### For further checking, we utilize the standard deviation metrics to quantify the extent
to hi h the att ibute alue of a featu e de iate f o it ea alue The ta da d de i
-----
_Future Internet 2023, 15, 278_ 13 of 16
**Figure 11. Best accuracy results of training and validation at epoch 19.**
For further checking, we utilize the standard deviation metrics to quantify the extent toFor further checking, we utilize the standard deviation metrics to quantify the e
which the attribute value of a feature deviates from its mean value. The standard deviationto which the attribute value of a feature deviates from its mean value. The standard
categorization aids in the identification of features that deviate from the average value byation categorization aids in the identification of features that deviate from the ave
highlighting values that are both above and below the mean. Figurevalue by highlighting values that are both above and below the mean. Figure 12 sh 12 shows the standard
deviation result for the proposed model for two datasets.the standard deviation result for the proposed model for two datasets.
13 of 16
(a) (b)
**Figure 12. The standard deviation metrics for proposed model using (Figure 12. The standard deviation metrics for proposed model using (a) Mendeley dataset anda) Mendeley dataset an**
(b) UNSW 2018 dataset.UNSW 2018 dataset.
As shown in Figure 12 and Table 4, the proposed model has a lower variance. As a
result, the proposed approach outperforms in terms of accuracy and reliability, and the
learning curves are smoother, indicating that the proposed model is consistent. As a result,
it is not only more accurate, but it is also more robust and consistent.
**Table 4. Results of tests using the UNSW 2018 dataset of the proposed model.**
**Dataset** **Accuracy (%)** **Standard Deviation (%)** **Validation Accuracy (%)** **Standard Deviation (%)**
Mendeley 99.99 0.0118198 99.923 0.000954
UNSW 2018 100 0.9250918 100 0.023263
5.3.3. Comparison of Results with Some Machine Learning and Deep Learning Algorithms
This section compares the proposed CNN-LSTM-autoencoder model with the LSTM
model and three other ML algorithms includes K-nearest neighbors algorithm (KNN), SVM,
and XGBoost. We use the same datasets and number of epochs to determine the difference
in performance between the proposed model and these two ML models. The results appear
in Table 5.
**Table 5. Results of tests using the Mendeley dataset for ML models.**
**Precision (%)** **Recall (%)** **F1-Score (%)**
**Model** **Accuracy (%)** **Val. Accuracy (%)**
**Normal** **Attack** **Normal** **Attack** **Normal** **Attack**
LSTM 92.6 79.433 80 80 6 62 11 70
KNN - - 97 95 97 95 97 95
SVM - 95 98 91 94 96 96 94
XGBoost 99.55 99.54 - - - - -
where Normal is the normal traffic, and Attack is DDoS attack traffic.
From Table 5, the results show the lower performance of the LSTM model with
the DDoS attack SDN dataset when it gains 92.6% in training accuracy and 79.433% in
validation accuracy. It also achieves very low results in recall, 6% in normal and 62% in
attack, and the precision is the same in both normal and attack (about 80%). So, the model
-----
_Future Internet 2023, 15, 278_ 14 of 16
gains poor results in the F1-score, 11% for normal, and 70% in attack. Table 5 shows good
results for the KNN model with the DDoS attack SDN dataset when it gains 97% for normal
and 95% for attack for precision, and for recall, it gains 97% in normal and 95% in attack.
The F1-score is then good at 97% for normal and 95% in attack. However, these results are
less than the proposed CNN-LSTM-autoencoder model. The SVM also gains good results
which are 98% for normal and 91% for attack for precision, and for recall, it gains 94% in
normal and 96% in attack. The F1-score is then good at 96% for normal and 94% in attack.
The XGBoost achieves higher accuracy and reaches up to 99.54%.
By comparing the deep learning model (LSTM) with the proposed CNN-LSTMautoencoder model, the proposed model is more accurate and stable than the LSTM
model and achieves higher accuracy in lower epochs. Compared with the machine learning
model (KNN), the proposed CNN-LSTM-autoencoder model is also more accurate than all
ML for the DDoS attack SDN dataset.
5.3.4. Comparison of Results with Published (Base)
We compared the result of the proposed system with some recent related works using
the DDoS attack SDN dataset and UNSW2018 BoTIoT (Table 6). The obtained results
showed that the model achieves very high classification results. For the UNSW2018 dataset,
the proposed model achieves 100% in all metrics and a very low loss, about 3.98 10[−][4]
_×_
for training and 3.3 10[−][3] for validation. Table 6 proves that our model outperforms
_×_
Yaser et al. [22] in all metrics using the same dataset (UNSW 2018). Ivanova et al. [27] and
Prasad et al. [28] had models that achieved an accuracy of 99.99%, whereas our model
achieves 100% accuracy. Our model outperforms their models in all metrics. Although it
had high accuracy, they showed lower precision, recall, and F1-score for normal flows. So,
our model can accurately detect and recognize normal and abnormal flows since it shows
100% in all metrics.
**Table 6. Comparison results between the proposed CNN-LSTM-autoencoder model and some**
recent works.
**Precision (%)** **Recall (%)** **F1-Score (%)**
**Ref.** **Dataset** **Algorithm** **Accuracy (%)** **Val. Accuracy (%)**
**Normal** **Attack** **Normal** **Attack** **Normal** **Attack**
Proposed CNN-LSTM
100 100 100 100 100 100 100 100
model autoencoder
Yaser et al. [22] UNSW2018 LSTM-autoencoder 99.95 99.94 95 99 94 99 95 99
optimized
Ivanova et al.
feed-forward neural 99.99 99.99 82.55 99.99 66.35 99.99 73.57 99.87
[27]
network
Prasad et al.
VMFCVD 99.99 99.99 87.72 99.99 82.55 99.99 81.97 99.99
[28]
Proposed CNN-LSTM
99.99 99.923 100 100 100 100 100 100
model autoencoder
DDOS attack
CNN 98.74 - 98.75 98.73 98.9 98.55 98.83 98.64
SDN Dataset
LSTM 95.60 - 96.20 94.90 95.64 95.56 95.92 95.23
(Mendeley
Ahuja et al. [29] CNN-LSTM 99.48 - 99.43 99.55 99.66 99.26 99.54 99.40
dataset)
SVC-SOM 95.45 - 96.71 93.75 95.40 95.51 96.05 94.62
SAE-MLP 99.75 - 99.96 99.69 99.77 99.94 99.87 99.82
Generated
Yaser et al. [22] LSTM-autoencoder 97.62 97.68 98 88 92 97 95 93
SDN dataset
where Normal is the normal traffic, and Attack is DDoS attack traffic.
Meanwhile, for the DDoS attack SDN dataset, our system gains an accuracy of up to
99.99% in training and 99.923% in validation and achieves 100% in precision, recall, and
F1-score. Also, the model training and validation losses are very low, about 3.98 × 10[−][4] for
training and 3.3 10[−][3] for validation. Our model outperforms all proposed models by
_×_
Ahuja et al. [29] in all factors. Experimental results reveal that our proposed model has a
high feature extraction ability and high performance in detecting attacks. All performance
metrics indicate that the proposed approach is the most appropriate choice to apply to a
real-world flow detection environment.
-----
_Future Internet 2023, 15, 278_ 15 of 16
**6. Conclusions**
Network virtualization imposes new risks and exploitable attacks in addition to those
currently on traditional networks. The DDoS attack group is one of the most aggressive
attack types in recent years, devastating the entire network infrastructure. To defend against
the DDoS attack, within the scope of this project, we developed and deployed a DDoS
detection system based on deep learning to detect multi-vector attacks within an SDN
environment. The proposed approach has a success rate of 99.99% in train and 99.923%
in validation and 100% for all metrics (precision, recall, and F1-score) for identifying
individual DDoS attacks in all DDoS datasets. It does so with an accuracy of 100% and an
extremely low False-Positive Rate compared to other efforts, and it categorizes the traffic
into normal and attack groups. One of our future goals is to test the proposed model as a
real-time classifier in an SDN environment under real-time DDoS traffic and normal traffic
to address its accuracy and time of detection using an emulator such as Mininet or in a
real SDN environment. In addition, our goal is to lessen the strain placed on the controller
by putting in place a network intrusion detection system that can identify not only DDoS
attacks but also others.
**Author Contributions: Conceptualization, A.K.M. and M.N.A.; methodology, A.K.M.; formal analy-**
sis, A.K.M. and M.N.A.; investigation, A.K.M.; writing—original draft preparation, A.K.M.; supervision, M.N.A. All authors have read and agreed to the published version of the manuscript.All authors
have read and agreed to the published version of the manuscript.
**Funding: This research received no external funding.**
**Data Availability Statement: Data derived from public domain resources.**
**Conflicts of Interest: The authors declare no conflict of interest.**
**References**
1. Urrea, C.; Benítez, D. Software-Defined Networking Solutions, Architecture and Controllers for the Industrial Internet of Things:
[A Review. Sensors 2021, 21, 6585. [CrossRef] [PubMed]](https://doi.org/10.3390/s21196585)
2. Nadeau, T.D.; Gray, K. SDN: Software Defined Networks; O’Reilly Media: Newton, MA, USA, 2013.
3. Feamster, N.; Rexford, J.; Zegura, E.; Tech, G. The Road to SDN: An Intellectual History of Programmable Networks. ACM
_[SIGCOMM Comput. Commun. Rev. 2014, 44, 87–98. [CrossRef]](https://doi.org/10.1145/2602204.2602219)_
4. Pradhan, A.; Mathew, R. Solutions to Vulnerabilities and Threats in Software Defined Networking (SDN). Procedia Comput. Sci.
**[2020, 171, 2581–2589. [CrossRef]](https://doi.org/10.1016/j.procs.2020.04.280)**
5. Silva, F.S.D.; Silva, E.; Neto, E.P.; Lemos, M.; Neto, A.J.V.; Esposito, F. A Taxonomy of DDoS Attack Mitigation Approaches
[Featured by SDN Technologies in IoT Scenarios. Sensors 2020, 20, 3078. [CrossRef] [PubMed]](https://doi.org/10.3390/s20113078)
6. Abdulkarem, H.S.; Alethawy, A.D. DDoS attack detection and mitigation at SDN enviroment. Iraqi J. Inf. Commun. Technol. 2021,
_[4, 1–9. [CrossRef]](https://doi.org/10.31987/ijict.4.1.115)_
7. Tan, L.; Pan, Y.; Wu, J.; Zhou, J.; Jiang, H.; Deng, Y. A New Framework for DDoS Attack Detection and Defense in SDN
[Environment. IEEE Access 2020, 8, 161908–161919. [CrossRef]](https://doi.org/10.1109/ACCESS.2020.3021435)
8. Choudhary, A.R.; Associates, C. OpenFlow switch controller as a policy-based system. Issues Inf. Syst. 2021, 22, 320–334.
[[CrossRef]](https://doi.org/10.48009/1_iis_2021_320-334)
9. Wang, T.; Chen, H.; Cheng, G.; Lu, Y. SDNManager: A Safeguard Architecture for SDN DoS Attacks Based on Bandwidth
[Prediction. Secur. Commun. Netw. 2018, 2018, 7545079. [CrossRef]](https://doi.org/10.1155/2018/7545079)
10. Lakshmanan, V.; Robinson, S.; Munn, M. Machine Learning Design Patterns; O’Reilly Media, Inc.: Newton, MA, USA, 2020;
Chapter 4; ISBN 9781098115784.
11. Doshi, K.; Yilmaz, Y.; Uludag, S. Timely Detection and Mitigation of Stealthy DDoS Attacks via IoT Networks. arXiv 2020,
[arXiv:abs/2006.08064. Available online: http://arxiv.org/abs/2006.08064 (accessed on 1 May 2023). [CrossRef]](http://arxiv.org/abs/2006.08064)
12. Yuan, X.; Li, C.; Li, X. DeepDefense: Identifying DDoS Attack via Deep Learning. In Proceedings of the 2017 IEEE International
[Conference on Smart Computing, SMARTCOMP, Hong Kong, China, 29–31 May 2017; pp. 1–8. [CrossRef]](https://doi.org/10.1109/SMARTCOMP.2017.7946998)
13. Shone, N.; Ngoc, T.N.; Phai, V.D.; Shi, Q. A Deep Learning Approach to Network Intrusion Detection. IEEE Trans. Emerg. Top.
_[Comput. Intell. 2018, 2, 41–50. [CrossRef]](https://doi.org/10.1109/TETCI.2017.2772792)_
14. Pekta¸s, A.; Acarman, T. A deep learning method to detect network intrusion through flow-based features. Int. J. Netw. Manag.
**[2018, 29, e2050. [CrossRef]](https://doi.org/10.1002/nem.2050)**
15. [IDS 2012|Datasets|Research|Canadian Institute for Cybersecurity|UNB. Available online: https://www.unb.ca/cic/datasets/](https://www.unb.ca/cic/datasets/ids.html)
[ids.html (accessed on 11 December 2022).](https://www.unb.ca/cic/datasets/ids.html)
-----
_Future Internet 2023, 15, 278_ 16 of 16
16. [IDS 2017|Datasets|Research|Canadian Institute for Cybersecurity|UNB. Available online: https://www.unb.ca/cic/datasets/](https://www.unb.ca/cic/datasets/ids-2017.html)
[ids-2017.html (accessed on 11 December 2022).](https://www.unb.ca/cic/datasets/ids-2017.html)
17. Gadze, J.D.; Bamfo-Asante, A.A.; Agyemang, J.O.; Nunoo-Mensah, H.; Opare, K.A.-B. An Investigation into the Application of
[Deep Learning in the Detection and Mitigation of DDOS Attack on SDN Controllers. Technologies 2021, 9, 14. [CrossRef]](https://doi.org/10.3390/technologies9010014)
18. Singh, A.; Jang-Jaccard, J. Autoencoder-based Unsupervised Intrusion Detection using Multi-Scale Convolutional Recur-rent
Networks. arXiv 2022, arXiv:2204.03779.
19. [The UNSW-NB15 Dataset|UNSW Research. Available online: https://research.unsw.edu.au/projects/unsw-nb15-dataset](https://research.unsw.edu.au/projects/unsw-nb15-dataset)
(accessed on 12 December 2022).
20. [NSL-KDD|Datasets|Research|Canadian Institute for Cybersecurity|UNB. Available online: https://www.unb.ca/cic/datasets/](https://www.unb.ca/cic/datasets/nsl.html)
[nsl.html (accessed on 6 December 2019).](https://www.unb.ca/cic/datasets/nsl.html)
21. Elubeyd, H.; Yiltas-Kaplan, D. Hybrid Deep Learning Approach for Automatic DoS/DDoS Attacks Detection in Software-Defined
[Networks. Appl. Sci. 2023, 13, 3828. [CrossRef]](https://doi.org/10.3390/app13063828)
22. Yaser, A.L.; Mousa, H.M.; Hussein, M. Improved DDoS Detection Utilizing Deep Neural Networks and Feedforward Neural
[Networks as Autoencoder. Futur. Internet 2022, 14, 240. [CrossRef]](https://doi.org/10.3390/fi14080240)
23. Jiang, H.; Turki, T.; Wang, J.T. DLGraph: Malware detection using deep learning and graph embedding. In Proceedings of the 2018
17th IEEE international conference on machine learning and applications (ICMLA), Orlando, FL, USA, 17–20 December 2018;
IEEE: Piscataway, NJ, USA, 2018.
24. Elsayed, M.S.; Le-Khac, N.-A.; Dev, S.; Jurcut, A.D. Network Anomaly Detection Using LSTM Based Autoencoder. In Proceedings
[of the Q2SWinet’20, Alicante, Spain, 16–20 November 2020. [CrossRef]](https://doi.org/10.1145/3416013.3426457)
25. Shiravi, A.; Shiravi, H.; Tavallaee, M.; Ghorbani, A.A. Toward developing a systematic approach to generate benchmark datasets
[for intrusion detection. Comput. Secur. 2012, 31, 357–374. [CrossRef]](https://doi.org/10.1016/j.cose.2011.12.012)
26. Tonkal, Ö.; Polat, H.; Ba¸saran, E.; Cömert, Z.; Kocao˘glu, R. Machine Learning Approach Equipped with Neighbourhood
[Component Analysis for DDoS Attack Detection in Software-Defined Networking. Electronics 2021, 10, 1227. [CrossRef]](https://doi.org/10.3390/electronics10111227)
27. Ivanova, V.; Tashev, T.; Draganov, I. Detection of IoT based DDoS Attacks by Network Traffic Analysis using Feedforward Neural
[Networks. Int. J. Circuits Syst. Signal Process. 2022, 16, 653–662. [CrossRef]](https://doi.org/10.46300/9106.2022.16.81)
28. Prasad, A.; Chandra, S. VMFCVD: An Optimized Framework to Combat Volumetric DDoS Attacks using Machine Learning.
_[Arab. J. Sci. Eng. 2022, 47, 9965–9983. [CrossRef] [PubMed]](https://doi.org/10.1007/s13369-021-06484-9)_
29. Ahuja, N.; Singal, G.; Mukhopadhyay, D.; Kumar, N. Automated DDOS attack detection in software defined networking. J. Netw.
_[Comput. Appl. 2021, 187, 103108. [CrossRef]](https://doi.org/10.1016/j.jnca.2021.103108)_
**Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual**
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3390/fi15080278?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/fi15080278, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "GOLD",
"url": "https://www.mdpi.com/1999-5903/15/8/278/pdf?version=1692439584"
}
| 2,023
|
[
"JournalArticle",
"Review"
] | true
| 2023-08-19T00:00:00
|
[
{
"paperId": "4cb78da23ad10b8c967b4f2b88d661eef251e538",
"title": "Hybrid Deep Learning Approach for Automatic Dos/Ddos Attacks Detection in Software Defined Networks"
},
{
"paperId": "8f689a70f8803a9bbb24fa6287976035c0b7883c",
"title": "Improved DDoS Detection Utilizing Deep Neural Networks and Feedforward Neural Networks as Autoencoder"
},
{
"paperId": "3906dc23018d4af33f5286b6828b1f5760e7765b",
"title": "VMFCVD: An Optimized Framework to Combat Volumetric DDoS Attacks using Machine Learning"
},
{
"paperId": "e61531a73ee56bba9d713eb3c6945944ad82aaa1",
"title": "Detection of IoT based DDoS Attacks by Network Traffic Analysis using Feedforward Neural Networks"
},
{
"paperId": "171008558ed39a467879635c9472b7a024b8d70a",
"title": "Software-Defined Networking Solutions, Architecture and Controllers for the Industrial Internet of Things: A Review"
},
{
"paperId": "d337ce71613e6d6c74fe6b88822852b5782681b4",
"title": "Automated DDOS attack detection in software defined networking"
},
{
"paperId": "aa1d3d17c48673ee968d729dd443811d93aac128",
"title": "Machine Learning Approach Equipped with Neighbourhood Component Analysis for DDoS Attack Detection in Software-Defined Networking"
},
{
"paperId": "80a5a2b2e37f7dd8a7aa1b92d98da216bfc53f58",
"title": "DDOS ATTACK DETECTION AND MITIGATION AT SDN ENVIROMENT"
},
{
"paperId": "4befc58cb4bbde0c86db40d44e8344bec77f2053",
"title": "An Investigation into the Application of Deep Learning in the Detection and Mitigation of DDOS Attack on SDN Controllers"
},
{
"paperId": "53f74493cb86cb733d0210b6bfa3accfd23e9506",
"title": "A Taxonomy of DDoS Attack Mitigation Approaches Featured by SDN Technologies in IoT Scenarios"
},
{
"paperId": "fe8f6bc91042ba9f18f6f284e6f3b984bf8265e7",
"title": "A deep learning method to detect network intrusion through flow‐based features"
},
{
"paperId": "78f7c6818c97383b5ae4b61664ffc6cde7974466",
"title": "A Deep Learning Approach to Network Intrusion Detection"
},
{
"paperId": "7968cabb6a0e45ddc56ae8af5bc3e7881baa966d",
"title": "SDNManager: A Safeguard Architecture for SDN DoS Attacks Based on Bandwidth Prediction"
},
{
"paperId": "4e62c6c212e51e2033654bd1f2cc26f394d5dd3a",
"title": "The road to SDN: an intellectual history of programmable networks"
},
{
"paperId": "846fcf30dc75f04886092891e754791e9704f69f",
"title": "Toward developing a systematic approach to generate benchmark datasets for intrusion detection"
},
{
"paperId": null,
"title": "OpenFlow switch controller as a policy-based system"
},
{
"paperId": "6fdc744218dba580e8104c3be0111eef9369c8b2",
"title": "A New Framework for DDoS Attack Detection and Defense in SDN Environment"
},
{
"paperId": "a13a0f244eb633a8e870b69b56be3c41ec97078f",
"title": "Solutions to Vulnerabilities and Threats in Software Defined Networking (SDN)"
}
] | 16,812
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/025bde8270278a539aa25dfc08ee9e73b569f0d0
|
[
"Computer Science"
] | 0.908816
|
DocCert: Nostrification, Document Verification and Authenticity Blockchain Solution
|
025bde8270278a539aa25dfc08ee9e73b569f0d0
|
International Conference on Blockchain Computing and Applications
|
[
{
"authorId": "1776184",
"name": "Monther Aldwairi"
},
{
"authorId": "2258548115",
"name": "Mohamad Badra"
},
{
"authorId": "2074769883",
"name": "Rouba Borghol"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"BCCA",
"Int Conf Blockchain Comput Appl"
],
"alternate_urls": null,
"id": "6cbed6ad-fb46-4c9a-8477-75227ff3a47e",
"issn": null,
"name": "International Conference on Blockchain Computing and Applications",
"type": "conference",
"url": "https://ieeexplore.ieee.org/xpl/conhome/1839124/all-proceedings"
}
|
Many institutions and organizations require nostrification and verification of qualification as a prerequisite for hiring. The idea is to recognize the authenticity of a copy or digital document issued by an institution in a foreign country and detect forgeries. Certificates, financial records, health records, official papers and others are often required to be attested from multiple entities in distinct locations. However, in this digital era where most applications happen online, and document copies are uploaded, the traditional signature and seal methods are obsolete. In a matter of minutes and with a simple photo editor, a certificate or document copy may be plagiarized or forged. Blockchain technology offers a decentralized approach to record and verify transactions without the need for huge infrastructure investment. In this paper, we propose a blockchain based nostrification system, where awarding institutions generate a digital certificate, store in a public but permissioned blockchain, where students and other stakeholders may verify. We present a thorough discussion and formal evaluation of the proposed system.
|
#, Authenticity Blockchain Solution
Monther Aldwairi
_College of Computer and Information_
_Technology_
_Jordan University of Science and_
_Technology_
Irbid, Jordan
munzer@just.edu.jo
Mohamad Badra
_College of Technological Innovation_
_Zayed University_
Dubai, UAE
mohamd.badra@zu.ac.ae
Rouba Borghol
_Science and Liberal Arts Dept._
_Rochester Institute of Technology_
Dubai, UAE
rbbcad@rit.edu
**_Abstract— Many institutions and organizations require_**
**nostrification and verification of qualification as a prerequisite**
**for hiring. The idea is to recognize the authenticity of a copy or**
**digital document issued by an institution in a foreign country**
**and detect forgeries. Certificates, financial records, health**
**records, official papers and others are often required to be**
**attested from multiple entities in distinct locations. However, in**
**this digital era where most applications happen online, and**
**document copies are uploaded, the traditional signature and seal**
**methods are obsolete. In a matter of minutes and with a simple**
**photo editor, a certificate or document copy may be plagiarized**
**or forged. Blockchain technology offers a decentralized**
**approach to record and verify transactions without the need for**
**huge infrastructure investment. In this paper, we propose a**
**blockchain based nostrification system, where awarding**
**institutions generate a digital certificate, store in a public but**
**permissioned** **blockchain,** **where** **students** **and** **other**
**stakeholders may verify. We present a thorough discussion and**
**formal evaluation of the proposed system.**
**_Keywords— nostrification, antiforgery, plagiarism, document_**
**_authentication, blockchain._**
I. INTRODUCTION
Often times, job applicants are required to certify their
documents, attest certificates, or equalize a degree or course.
The attestation process is lengthy, time consuming, costly and
cumbersome, especially when the candidate did graduate long
time ago or he graduated from a foreign country, he no longer
has access to. Equivalency process requires those attested
certificates and transcripts and awards an equivalent and
recognized degree. The same process applies for international
trade agreements, customs’ forms, birth certificates, etc. The
process may involve universities, schools, notaries,
embassies, departments ministries, education boards, etc. The
process takes several months and may be costly if you live
abroad. However, after all of that trouble, the final attested or
sealed documents may be easily forged digitally. Therefore,
not much verification is achieved after this lengthy and costly
process [1]. Below is a sample of foreign degree equivalency
process requirements.
1. Certified copy of degree (and or all previous degrees) in
English or a translated version from an official service.
Copies must be attested by
a. Ministry of education in issuing country,
b. ministry of exterior in issuing country,
c. embassy of country seeking equivalency,
d. ministry of exterior in country of equivalency, and
e. ministry of education in country of equivalency.
2. Copy of transcript or diploma indicating dates of
admission and completion.
3. Copy of passport with visa, entry and exit stamps.
4. Equivalency fees.
5. All original documents.
In most of the above documents, notaries’ services may be
required. Public and private notaries are authorized by the
judicial system to attest documents and certify their
originality. Online notaries have been using cameras to verify
identity and attest documents. However, the admissibility in
court of digital signatures continues to be challenged. Courts
often accepts digitally signed documents only when a copy of
the originally signed document is presented! [2].
We believe blockchain is a game changer and would
present a perfect solution for the online nostrification issue [3].
Blockchain has been made popular with the wide spread of
Bitcoin. Bitcoin is one of the earliest and most popular
cryptocurrencies. It is a digital currency that can be exchanged
between people without the need for a central bank or
authority. It benefits from peer-to-peer networks and
blockchain technologies to keep an anonymous record of all
Bitcoins [4].
Blockchain is a shared immutable ledger for securely
recording transaction history. A blockchain could be public or
private. As indicated by the name, the blocks are chained, with
each block storing one or more transactions. Transactions are
kept indefinitely and the blockchain may be queried to verify
any transaction, which makes it ideal for nostrification. There
have been few attempts to use blockchain for nostrification.
EduCTX was one of the first attempts to use blockchain as a
higher education credit platform. It is supposed to serve as a
centralized repository of all students’ records and completed
courses [5].
In this paper we propose to use blockchain in a novel
manner to implement a secure, shared authenticated and
public repository of student records. Students may access their
records, so can universities and any other participant who
wishes to verify a record. Security and privacy are of the at
most importance and therefore access to records is
authenticated. The rest of the paper is organized as follows.
Section II explains blockchain in more details. Section III
surveys the literature, covers the related work and points out
advantages and disadvantages of exiting approaches. Section
IV discusses the proposed approach and Section V presents
the formal evaluation.
-----
II. BLOCKCHAIN
Blockchain maintains a shared record including full details
of every single transaction over a network. Blockchain is
based on peer to peer networks, making it distributed and not
controlled by any third party [4]. A transaction is any
exchange of assets between participants and is represented by
a block. Each block tracks and stores data and those blocks are
chained together chronologically. Blocks are not editable,
which means once a transaction is committed to the
blockchain it can no longer be modified. A transaction is
reversed by creating a new block, which maintains a timeline
of events and changes to the data. Each block contains the
transaction data, timestamp, unique hash and the hash of the
previous block. The latter maintains the chain and the
timestamp ensures timeliness. Unlike databases that are files
stored on a single system, blockchain is decentralized and
identical copies of the shared ledger are distributed across all
participating nodes. This distributed nature of the ledger
reduces the chances of data tempering. If a party chooses to
add a block to his copy of ledger, it will be inconsistent with
all other blockchain participants [6].
Before any block is added to the chain, a consensus of the
majority of the endorsing nodes must be reached. Consensus
may be through solving a cryptographic puzzle called “proofof-work”, which is the case in many cryptocurrencies. Proofof-state on the other hand requires validators to hold a
cryptocurrency in escrow trusted service. While proof-oftime-elapsed randomizes blocks waiting for trusted
environment. Solo-NoOps requires no consensus and
validator applies transactions, which may lead to divergent
chains or ledgers. Finally, Byzantine-Fault-Tolerance (BFT)
achieves consensus in a peer to peer network while some
nodes are malicious or faulty [7].
In blockchain for business, we have a shared ledger, where
every participant has his own copy. Those ledgers are
permissioned and proper credentials are required to access the
ledger. The ledger is immutable, in that no participant may
tamper with a transaction after it was agreed upon.
Transactions cannot be altered deleted or inserted back in
time. Smart contracts are a set of business rules in chain code
format that when executed a block/transaction is created. The
shared ledger has the final say of an asset ownership and
provenance. Contrary to cryptocurrencies that emphasize
anonymity, blockchain for business is private permissioned
network that values identity and permissions over anonymity
[8].
Turkanovic et al. explained that higher education
institutions (HEI) keep their students’ completed courses’
records in databases that are structured and only available to
institution’s staff [4]. Thus, leaving students with limited
access where they can only view or print their document.
Moreover, these student documents are stored in different
standards, which contributes to the problem of transferring
student documents to another HEI [9]. Correspondingly, if a
student wants to apply for a job in a foreign country, he/she
has to translate and nostrificate their academic certificate,
which is complex and time consuming. In addition, if a student
loses his/her academic certificate, he/she has to visit their HEI
and ask for a new copy. Andrejs Rauhvargers discussed
qualifications frameworks for recognizing qualification in the
European higher education. The paper details the recognition
of foreign higher education qualifications [9].
Blockchain is idea for keeping record of student’s diploma
certificates, transcript, courses, grades, achievements, skills
and research experience. All of these may be securely
registered in a shared ledger that can be accessed by many
institutions or stakeholders. This will help reduce fraud,
forgery and false claims [10]. All of the above data can be
logged in the form of timely transactions into the shared
ledger. The data in the blockchain is permissioned and
associated with a student ID, organization ID and stored
security in the blockchain. Using blockchain means
performance may be sacrificed for secure recordkeeping of
transactions. Nonetheless, the blockchain would be much
more efficient as opposed to the manual attestation process
described earlier [11].
Using blockchain in education might be a new concept, but
it surely is very beneficial. It will make it much easier for
students to have all of their completed courses certificated,
verified and in one place. Not only this will facilitate
attestation and verification of qualification but also will help
in the cases of credit transfer between institutions. It will be
very easy for any workplace anywhere in the world to
subscribe to the blockchain and verify graduate credentials, of
course with the applicant’s consent [12].
III. RELATED WORK
There are a few research papers concerned with
blockchain for nostrification. In this section we summarize
each paper and present a critical analysis.
Wibke et al. used blockchain technology to store and
handle educational data [13]. They offer the possibility to
store different types of immutable educational information on
blockchain technology. A total of 58.1 % of the education
technologies were based on Ethereum, 3.2 % on Bitcoin, 9.7
% on EOS, and 1.6 % on NEM; 1.6 % used a private
blockchain, 4.8 % could be used more than one blockchain
and 6.5 % used other blockchain technologies. Their results
provide a deeper understanding of blockchain technology in
education and serve as a signal to educational stakeholders by
underlining the importance of blockchain technology in
education.
EduCTX is a blockchain-based higher education credit
platform from University of Maribor [5]. This EduCTX
platform is anticipated to use ECTX tokens as academic
credits. It rests on peer-to-peer networks where the peers of
the network are HEI and users of the platform are students and
other various organizations. These ECTX tokens represent
students credit amount for completed courses. Every student
will have an EduCTX wallet for collection of ECTX tokens
that will be transferred by his/her HEI. The transferred
information is stored in blockchain alongside with the
sender’s identity with HEI official name, the recipients (which
is student and is anonymously presented), the token (course
credit value) and the course identification. Therefore, students
can access and provide his/her completed courses by directly
presenting their blockchain address.
EduCTX is still a prototype based on Ark blockchain
platform, and the real-world perception cannot be evaluated.
EduCTX enables organizations and students to check
academic records of a student’s (potential employee) in
transparent way. Moreover, since the system is based on
blockchain platform, it maintains the possibility of fraud
detection and prevention. On the other hand, in the case of a
student’s losing his/her private key, they have to visit to their
-----
home HEI and request a new blockchain address, which is
time consuming and almost similar to the current approach for
certification. Moreover, it is expected that user and
organizations have to protect and backup their private keys,
signatures and stamps to be secure, because this platform is
yet to have additional level of protection against
impersonation.
Gresch et al. from the University of Zurich proposed a
blockchain-based Architecture for Transparent Certificate
Handing [14]. The work used a questionnaire to shed the light
on the wide spread of people with fake diplomas, and how
ineffective is the current accreditation system. The system
identifies three stake holders: the certificate issuer, companies
and institutes wanting to verify diplomas, and the graduates or
applications who submitted the diploma. The system has two
stages. First, the issuing organization has to create the digital
diploma, with one-way hash function and the hash will be
stored in a smart contract. Second, the verifier company
verifies the authenticity of the document without contacting
the university. A prototype was built using an Ethereum
blockchain and deployed on University of ZuricH BlockChain
(UZHBC). They concluded that granting an organization the
ability to issue certificates is one of the most critical aspects
of the blockchain. In addition, they only stored the diploma
has on the blockchain for privacy concerns. As opposed to
storing an encrypted diploma and risking losing the data
forever if the key is lost.
Azael Capetllo proposed a blockchain education longstanding model for academic institutions [15]. The paper
described the technology of storing student records, which can
be shared openly with third parties, offering a safe and lasting
record. The technology is strong against data damage or loss,
and those third parties can verify student record directly by
accessing the University blockchain. Two applications of
Blockchain in education have been mentioned in the research
paper, the first one is Smart Contracts, and it is to form an
autonomous learning experience by consuming an analogy
from the financial application of blockchain. The second
application is the use of Blockchain to offset the cost learning
using peer-to-peer networks, offering financial prize for
students offering services to university.
Mike Sharples and John Domingue from University of
Nicosia proposed blockchain and Kudos, a distributed system
for educational record, reputation and reward [16]. It was the
first higher education institution to issue academic certificates
whose authenticity can be verified through the Bitcoin
blockchain. They proposed to use Bitcoin payments as a
reward for academic achievements as tasks such as peer
review or assessments. Then they proposed an “educational
reputation currency’, called Kudos. Each recognized
educational institution, innovative organization, and
intellectual worker is given an initial award of educational
reputation currency, the initial award might be based on some
existing metric: Times Higher Education World Reputation
Rankings for Universities, H-index for academics, Amazon
author rank for published authors etc. An institution could
allocate some of its initial fund of Kudos to staff whose
reputation it wishes to promote. Each person and institution
store its fund of reputation in a virtual wallet on a universal
educational blockchain. They used Ethereum smart contracts
to implement OpenLearn badges on a private blockchain,
where student enroll on courses and institution award them
badges.
Wolfgang et al. proposed blockchain in the context of
education and certification [17]. The blockchain technology
supports counterfeit protection of certificates, easy
verification of certificates even if the certification authority no
longer exists and automation of monitoring processes for
certificates with a time-limited validity. It ensures higher
efficiency and improved security for certification authorities
through digitization of current processes, issuing and
registering of certificates in a blockchain as well as automatic
monitoring of certificates. It comprises a blockchain including
smart contracts, a public storage holding profile information
of certification authorities, a document management system
managing the actual payload of certificates tracked by the
blockchain and the parties involved in the system, namely
accreditation and certification authorities, certifiers, learners
and employers.
John Rooksby and Kristiyan Dimitrov from University of
Glasgow implemented Ethereum based blockchain
technology for permanent and tamper proof grading system
[18]. The system was able to store student information on
courses enrolled, grades and their final degree. It supported
the university specific cryptocurrency called Kelvin Coin.
Payment of the cryptocurrency can be made by smart contract
to the top performing student in a course. However, there were
some drawbacks involved by implementing the system.
Scenario-based and focus group evaluation methods were
implemented to address the advantages and disadvantage
derived from the system. Because universities rely on trust and
confidentiality the blockchain system was found to be not
trustworthy. Blockchain system was global scope idea,
however universities tend to set their own boundaries, at least
at institutional level. Moreover, using smart contracts to store
grades in blockchain was problematic due the fact that there is
no formal algorithm for calculating grades. Unfortunately, the
Ethereum Blockchain system needed to change the way of
administrative system of the university work. Finally, the
prototype of the blockchain system was found to prioritize
transparency over efficiency.
Cheng et al. [19] proposed a system that uses Ethereum to
generate digital certificates and confirm the eligibility of
graduation certificates [19]. The system functions as follows.
The HEI enters student’s certificate and academic records into
the system. The system verifies all the data. The student
receives a quick response (QR) code, inquiry number and
electronic file of their certificate. Whenever students want to
apply for a job or apply for higher education, he/she has only
to send the e-certificate alongside with the QR code to the
respective organization. The organization can retrieve the
student’s certificate and academic records once the credentials
are verified. Moreover, the QR is used to asses if the certificate
is tampered or forged.
There have been several industry projects that were
concerned with student records and online digital badges.
Many projects capitalized on the opportunity of digital
diplomas as countermeasure to fake diplomas. Those projects
offered technologies to both mange the complete educational
past of students by gathering all digital badges awarded by
different academic organizations. Sony Global Education for
example, has announced development of a new blockchain for
storing academic records [20]. Their platform allows secure
sharing of exams results and academic proficiency levels with
third-party evaluating organizations. Mozilla Foundation
Open Badges are a digital record of the different
-----
accomplishments encoded into an image with associated
infrastructure for verification [21]. MS Global Learning
Consortium was managing this central open source repository
of badges with over 1500 participating organizations until
2017. More recently, Mozilla migrated all users to Badgr, as a
replacement for Open Badges as the standard verifying
credentials [22]. Finally, Acclaim and IBM offer digital
badges as a form of organizations recognizing individuals’
skills and competencies [23]. Contrary to all of the above
industry efforts based in central repositories of badges,
BCDiploma is using blockchain to provide security,
immutability, ease of use for certifying diplomas and other
achievements [24].
All of the above research agreed that counterfeit
certificates, credentials and documents is a major problem that
can be solved with blockchain. Record all student’s academic
history from completed courses, skills and qualification in one
trusted and secure blockchain is a perfect solution [25]. Yet,
all of them tried to tweak current cryptocurrencies blockchain
to be used to store certificates and award badges (reputation)
and that resulted in low usability. Crypto currencies
blockchains and smart contracts are not customized to student
records. We propose a permissioned and custom blockchain
for business, designed specifically for storing student records
or any other document for that matter.
IV. PROPOSED APPROACH
In this section, we propose an efficient solution that is
based on a Merkle tree to provide nostrification and
verification of qualifications to guarantee data integrity on the
certificates through non-repudiation.
A Merkle hash tree [26] is a data structure used to
efficiently verify data integrity and authenticity. As illustrated
in Fig. 1, each non-leaf node in the tree, from the bottom up
until reaching the root node in the tree, holds the hash of the
concatenated hashes of its sub-nodes; example, ℎ�� �
ℎ�ℎ� | ℎ��. The hash held by the root is represented as the root
hash, which can be shared in a trusted way for verification
purposes; example ℎ���� �ℎ�ℎ�� | ℎ��� . In [27], a hash
calendar is proposed to include the generated root hashes to
verify the integrity of the contents of large data structures.
In our proposed solution, we propose forming a Merkle
tree where the leaves are documents. The first objective is to
provide a periodic publication of the root hash in the
blockchain to enhance transparency and protection against
any modification in the hashed content, and to provide proof
of existence of contents. Each document is certified by its
issuer, so we include, in a blockchain, either the hash of that
document, or the hash root of a set of documents issued by the
same issuer when multiple documents are to be included. The
included hash is authenticated by digitally signing it by the
same issuer.
Fig. 1. Example of Merkle tree with 4 leaves (depth = 2).
Interested parties can authenticate the existence of any
document and the verification is legally acceptable. The
verification process relies on the authentication path of a given
node in the tree to validate the hashed content held by the
node, without Merkle tree traversal [28]. The authentication
path of a node in the tree consists of a set of siblings on the
path from that node to the root. The content of a document can
be authenticated using the hashed contents held by the root
node and by the corresponding authentication path as well. For
example, and with reference to Fig.1, to verify ℎ�, the verifier
needs the values of ℎ����, and the authentication path ℎ�� and
ℎ� . Hence, the verifier computes ℎ��� �ℎ�ℎ�� | ℎ�� and
ℎ����� �ℎ�ℎ��� | ℎ���, and then compares ℎ����� with ℎ���� for
equality. Choosing the hash function or algorithm [29]
depends on many factors such as speed, digest length, number
of rounds, collisions and ease of implementation both in
hardware and software [30].
_A._ _Transaction Structure_
When a transaction is generated by our entities, it should
include the hash root and a set of hash values, where each hash
is the digest of a document belonging to the user.
## hroot Set of hash values (i.e., h1, h2, …)
_B._ _Nostrification’s Generation of Document’s Qualification_
The proposed system is very versatile and can be applied
to any document and authentication process. In this section,
we describe our solution using three different scenarios. In the
first one, we describe the case where the user has several
documents issued by the same entities, whereas in the second
case, the user has several documents issued by different
entities. The third scenario is concerned with the case of one
document that will be certified by a hierarchy of different
institutions.
The proposed nostrification solutions supports 3 cases or
operating scenarios, because of space limitation we present
cases 1 and 3.
**_Case 1_**
In this case, the issuer (entity) of the documents will form
a Merkle tree where each leaf is the hash of a document (Fig.
2). Next, the entity will generate the authentication path for
each intermediate node in the tree, the digest of each
document, and the hash root of the tree. Then the entity will
sign the hash root and publish it along with the hash value of
each document in the blockchain. The entity will next issue
the documents to the user after stamping each document. The
stamp consists of adding to each document, the identifier of
the transaction that is already stored in the blockchain.
**_Case 3_**
This case is similar to the two previous cases; however, each
tree is dedicated to one document only and each layer of the
tree is associated to an organization that will certify the
document. Each organization has a private and a public key.
When we want to nostrify a document, we start by selecting
all organizations that will certify the document. Then, with
the hash of the document, we sign that hash with the private
key of the first organization. Finally, we create the right node
of the layer with the pair constituted by the signature and the
location of the public key of the organization, required to
verify the signature. The parent node of the layer is created
-----
by hashing the result of the concatenation of the hash from
the left node and the signature from the right node. The
process is then repeated for each organization. When a layer
has been created for each organization, we calculate the last
hash that will become the hash root and we have the Merkle
tree (Fig. 3).
Fig. 2. Nostrification of several documents issued by the same entities.
As everyone is able to get public keys of organization with
the location included in the tree, it is very simple to verify the
authenticity of a document certify by any number of
organizations.
Fig. 3. Nostrification of a single document by different entities.
_C._ _Nostrification’s Verification of Document’s_
_Qualification_
Any third party who is willing to verify any document that
is shown by the user, the third party shall first query the
blockchain to extract the root hash and the set of hash values
as well, using the transaction identifier stored on the document
presented by the user. Then, the third party generates the
digest of the presented document and compares it for equality
with the one of the hash values stored in the extracted set of
the hash values. Then, the third party regenerates the hash root
and compares the result for equality with the hash root that is
downloaded from blockchain. Then, it verifies the signature
on the hash root that is generated by the document issuer, and
if the signature is validate, the third party approve the
document.
In case a third party is willing to verify more than one
document that were nostrified by more than one entity, then
the user should send the transaction identifier of the most
recent nostrified document, which includes a hash root, and a
set of hash values. This latter set includes the hash of the most
recent nostrified document and the hash of any other
document belonging to the user and nostrified prior to
nostrifying the most recent nostrified document.
V. IMPLEMENTATION AND SECURITY ANALYSIS
In this section, we present a detailed analysis of the
proposed approach’s security and we demonstrate its
effectiveness in providing in providing long-term integrity
protection, proof of existence, authenticity, non-repudiation
and privacy. Additionally, we evaluate the efficiency in terms
of the processor and time usage.
We start with one of the most popular attacks on data
integrity, False Data Injection (FDI). In FDI the attackers
intentionally change the data in such a way that the receiver
will be unable to detect forged data. Blockchain by its design
is secured against tampering and revision, which makes it very
difficult to the adversary to inject or add forged or malformed
document into the blockchain. In addition, the signature of the
issuer over the document, make it much more difficult, even
impossible, to inject malformed data into the blockchain.
In addition to the long-term data integrity and the proof of
existence, our solution ensures the authenticity and the nonrepudiation of origin because the hash root will be signed by
the last document’s issuer when we have several documents
from different entities, and by the documents’ issuer when
those documents were issued by the same issuer. It is worth
noting that the issued documents will be always valid if the
issuer’s certificate will expire or revoked. In fact, our solution
leverages blockchain approach properties to provide longterm integrity of documents.
Privacy concern usually arises in many applications,
particularly in is a public distributed database like the
blockchain. The privacy concerns are mostly related to the
publication of the user’s documents. In our solution, the
privacy is preserved since the digest of the documents are
stored in the blockchain, but not the document itself. Hence,
the adversary needs to crack the digest in order to find out the
original document.
Our approach maintains the above security services while
reducing the computation overhead. In fact, instead of
generating a signature for each document issued by the same
entity to the user, the entity will only need to sign the hash
root. However, our solution will introduce very limited
computation overhead related to the hash function that will be
applied to generate the hash of each document that is required
to compute the hash root. But consider the asymmetric
encryption computational overhead when compared to the
hash function computational overhead, this latter is usually
negligible.
The system proof of concept was implemented using
Python v3, Flask webserver, and Ganache is used to create the
blockchain test server. To evaluate the efficiency of the
system we measure the CPU usage, memory consumption,
and time for adding a document and the nostrification process.
The PyCharm IDE was used for measurements and average of
five runs. In Table I, we can observe that adding a user will
relatively take more time because of the deployment of the
contract on the blockchain.
Our system proof of concept allows everyone to verify the
authenticity of a document after accessing both the
authentication path and the hash root, which are stored into a
smart contract that is publicly available on a blockchain. Our
used smart contracts are based on the same cryptography
technology being used by cryptocurrency; therefore, they have
the same level of security. All details of deployments or
-----
updates of contracts are written into transactions to help
finding the data at any time. Particularly, it allows storing data
like the username and the user’s Merkle Tree. When the
issuing institution adds a user along with its documents to the
blockchain, a contract is deployed, and a transaction initializes
it with the data. At any time, the issuing institution can add
several documents to the user profile, in which the user’s
Merkle Tree is then updated, and a new transaction is also
needed to update the data stored on the blockchain.
TABLE I. SIMULATED PERFORMANCE
**_Add Time (s)_** **_Time (s)_**
**_Memory_**
**_CPU_** **_1 user & 4_** **_Nostrification_**
**_(MB)_**
**_documents_**
2% 37 7.5 0.001
Case1
5% 49 0.22 0.028
Case3
VI. CONCLUSIONS
In this paper, we proposed a blockchain-based
nostrification system, where awarding institutions generate a
digital certificate, store in a public but permissioned
blockchain, where students and other stakeholders may verify.
We present a thorough discussion and formal evaluation of the
proposed system. In addition, we implemented a prototype of
the solutions supporting 3 use-cases. The formal analysis
shows resistance to all sorts of common attacks with excellent
performance in terms of CPU and memory usage as well as
negligible blockchain programing and query times.
ACKNOWLEDGMENT
This project was supported in part by Zayed University
Research incentive grant #R22018.
REFERENCES
[1] Guang Chen, Bing Xu, Manli Lu and Nian-Shing Chen, Exploring
blockchain technology and its potential applications for education,
Smart Learning Environments 5:1, 2018,
[2] Ernie Brickell, Jan Camenisch, and Liqun Chen. 2004. Direct
anonymous attestation. In Proceedings of the 11th ACM conference on
Computer and communications security (CCS '04). ACM, New York,
NY, USA, 132-145.
[3] Nelson Bore, Samuel Karumba, Juliet Mutahi, Shelby Solomon
Darnell, Charity Wayua, and Komminist Weldemariam. 2017.
Towards Blockchain-enabled School Information Hub. In Proceedings
of the Ninth International Conference on Information and
Communication Technologies and Development (ICTD '17). ACM,
New York, NY, USA, Article 19, 4 pages.
[4] [Satoshi Nakamoto, Bitcoin: A Peer-to-Peer Electronic Cash System.
White paper, October 31, 2008. Available
https://bitcoin.org/en/bitcoin-paper [Last Access 20/5/2019].
[5] Turkanovic, M., Holbl, M., Kosic, K., Hericko, M., & Kamisalic, A.
(2018). EduCTX: A Blockchain-Based Higher Education Credit
Platform. IEEE Access,6, 5112-5127.
[6] Swan M. Blockchain: Blueprint for a New Economy. “O’Reilly Media,
Inc.”; 2015.
[7] Vincent Gramoli, From blockchain consensus back to Byzantine
consensus, Future Generation Computer Systems, 2017.
[8] Underwood, Sarah. (2016). Blockchain beyond Bitcoin.
Communications of the ACM. 59. 15-17. 10.1145/2994581.
[9] Gavin Heron, Pam Green Lister. (2014) Influence of National
Qualifications Frameworks in Conceptualising Feedback to Students.
Social Work Education 33:4, pages 420-434.
[10] Blockcerts, The open standard for blockchain credentials. 2016.
Available https://www.blockcerts.org/ [Last Access 20/5/2019].
[11] Grech, A., & Camilleri, A. F. (2017). Blockchain in Education. JRC
Science for Policy Report. https://doi.org/10.2760/60649
[12] Guang Chen, Bing Xu, Manli Lu and Nian-Shing Chen, Exploring
blockchain technology and its potential applications for education.
Smart Learning Environments, 5:1, 3 January 2018.
[13] Lévy, W., Stumpf-Wollersheim, J., & Welpe, I. M. (2018). Disrupting
Education Through Blockchain-Based Education Technology? SSRN
Electronic Journal. doi:10.2139/ssrn.3210487
[14] Gresch J., Rodrigues B., Scheid E., Kanhere S.S., Stiller B. (2019) The
Proposal of a Blockchain-Based Architecture for Transparent
Certificate Handling. In: Abramowicz W., Paschke A. (eds) Business
Information Systems Workshops. BIS 2018. Lecture Notes in Business
Information Processing, vol 339. Springer, Cham.
[15] Azael Capetillo, Blockchained education: Challenging the longstanding model of academic institutions. Available
http://www.iacee.org/docs/P10Blockchained_education_Challenging_the_longstanding_model_of_academic_institutions._52.pdf [Last Access
20/5/2019].
[16] Sharples, M., & Domingue, J. (2016). The Blockchain and Kudos: A
Distributed System for Educational Record, Reputation and Reward.
Adaptive and Adaptable Learning Lecture Notes in Computer
Science,490-496. doi:10.1007/978-3-319-45153-4_48.
[17] Wolfgang Gräther, Sabine Kolvenbach, Rudolf Ruland, Julian Schütte,
Christof Torres, Florian Wendland, Blockchain for Education:
Lifelong Learning Passport, Proceedings of 1st ERCIM Blockchain
Workshop 2018, ERCIM-Blockchain 2018: Blockchain Engineering:
Challenges and Opportunities for Computer Science Research.
Amsterdam, Netherlands, 8 - 9 May 2018.
doi:10.18420/blockchain2018_05.
[18] John Rooksby and Kristiyan Dimitrov, “Trustless Education? A
Blockchain System for University Grades”. New value transactions:
Understanding and Designing for Distributed Autonomous
Organizations, Workshop at DIS2017, 10th of June 2017. Edinburgh
[19] J. Cheng, N. Lee, C. Chi and Y. Chen, "Blockchain and smart contract
for digital certificate," 2018 IEEE International Conference on Applied
System Invention (ICASI), Chiba, 2018, pp. 1046-1051. doi:
10.1109/ICASI.2018.8394455
[20] Sony Global Education Develops Technology Using Blockchain for
Open Sharing of Academic Proficiency and Progress Records.” Sony
Global, Sony Global Headquarters. February 22, 2016. Available:
http://www.sony.net/SonyInfo/News/Press/201602/160222E/index.html [Last Access 20/5/2019].
[21] Open Badges Standard, 2014. Available:
http://www.badgealliance.org/open-badges-standard/. [Last Access
20/5/2019].
[22] Badgr, 2019. Available: https://badgr.com/ [Last Access 20/5/2019].
[23] Acclaim Digital badges, 2018. Available:
https://www.youracclaim.com/ [Last Access 20/5/2019].
[24] BCDiploma, 2019. Available: https://www.bcdiploma.com/index.html
[Last Access 20/5/2019].
[25] DJ Skiba, The potential of Blockchain in education and health care.
Nurs. Educ. Perspect. 38(4), 220–221 (2017)
https://doi.org/10.1097/01.NEP.0000000000000190.
[26] R.C. Merkle, “A Digital Signature based on Conventional Encryption
Function,” in proc CRYPTO, 1987, Springer Verlag.
[27] A. Buldas, A. Kroonmaa, R. Laanoja, “Keyless signatures
infrastructure: How to build global distributed hashtrees,” in H. Riis
Nielson and D. Gollmann (Eds.), NordSec 2013, LNCS 8208, 2013,
pp. 313-320.
[28] M. Jakobsson, T. Leighton, S. Micali, M. Szydlo, “Fractal Merkle tree
representation and traversal,” in Cryptographer’s Track at RSA
Conference, 2003, pp.314-326.
[29] M. Aldwairi, D. Alansari, “n-Grams exclusion and inclusion filter for
intrusion detection in Internet of Energy big data system”s. Trans
Emerging Tel Tech. 2022; 33:e3711.
[30] M.r Aldwairi, Y. Flaifel, K. Mhaidat, “Efficient Wu-Manber Pattern
Matching Hardware for Intrusion and Malware Detection”,
International Conference on Electrical, Electronics, Computers,
Communication, Mechanical and Computing (EECCMC), 28-29th
January 2018, Tamil Nadu, India
|Col1|CPU|Memory (MB)|Add Time (s) 1 user & 4 documents|Time (s) Nostrification|
|---|---|---|---|---|
|Case1|2%|37|7.5|0.001|
|Case3|5%|49|0.22|0.028|
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/2310.09136, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "GREEN",
"url": "https://arxiv.org/pdf/2310.09136"
}
| 2,023
|
[
"JournalArticle",
"Conference"
] | true
| 2023-10-13T00:00:00
|
[
{
"paperId": "70be9c86d51ea0f3b8932484a084f65837c014d5",
"title": "Blockchained education: challenging the long-standing model of academic institutions"
},
{
"paperId": "3c6e884fb6adae5bf90dc56acb3a64bb9ece5ae4",
"title": "Efficient Wu-Manber Pattern Matching Hardware for Intrusion and Malware Detection"
},
{
"paperId": "546d8aad284e72be2e02802abbc52ae2bb196c4c",
"title": "Trustless education? A blockchain system for university grades1"
},
{
"paperId": "713e07e26178b03586b438ccd1035472086b4714",
"title": "n-Grams exclusion and inclusion filter for intrusion detection in Internet of Energy big data systems"
},
{
"paperId": "82410757efad7a5db47676ab4b5bfb3ccff6553b",
"title": "The Proposal of a Blockchain-Based Architecture for Transparent Certificate Handling"
},
{
"paperId": "3aae59b3d9ef1b9f4808d8095c4ba1c99b8dcaad",
"title": "Disrupting Education Through Blockchain-Based Education Technology?"
},
{
"paperId": "50f725695a38b38e43f83757c289e7f6239398db",
"title": "Blockchain and smart contract for digital certificate"
},
{
"paperId": "ce99ac88a0baefd749ca4d7db8e5ed07dbaf51bf",
"title": "Exploring blockchain technology and its potential applications for education"
},
{
"paperId": "ea54e5e33145a511c88cd04727e8f5ff39cb0212",
"title": "Towards Blockchain-enabled School Information Hub"
},
{
"paperId": "d6853c98da0e95757dfc8fe31da4cea5f47503cc",
"title": "Blockchain in education"
},
{
"paperId": "9df21c23ccd29676889aea3d1e42e7fea76b4536",
"title": "EduCTX: A Blockchain-Based Higher Education Credit Platform"
},
{
"paperId": "7d7bbd407359122decdc5fd90afd58fd11b68476",
"title": "From blockchain consensus back to Byzantine consensus"
},
{
"paperId": "9a446ba6ca75630407aa7d5deee1b0a2d0ec7ed8",
"title": "The Potential of Blockchain in Education and Health Care."
},
{
"paperId": "efe573cbfa7f4de4fd31eda183fefa8a7aa80888",
"title": "Blockchain beyond bitcoin"
},
{
"paperId": "0200d453f5c995c87761e50976ed07692e257a30",
"title": "The Blockchain and Kudos: A Distributed System for Educational Record, Reputation and Reward"
},
{
"paperId": "97fddbbfd681bce9eeb8e0a013353b4d5b2ba0db",
"title": "Blockchain: Blueprint for a New Economy"
},
{
"paperId": "c7864e71ae2e4c8b3cc8daa91a651b4c09c90835",
"title": "Influence of National Qualifications Frameworks in Conceptualising Feedback to Students"
},
{
"paperId": "25a17e483599215949cb3961fd945f6867d3bcae",
"title": "Keyless Signatures' Infrastructure: How to Build Global Distributed Hash-Trees"
},
{
"paperId": "60739cb3e92415cff6f76a34374045d3f5cb6f27",
"title": "Direct anonymous attestation"
},
{
"paperId": "a616473a13cf6f3a9deba8cf30847dd83d59979b",
"title": "Fractal Merkle Tree Representation and Traversal"
},
{
"paperId": "5bcd990b11e068234c3a13b021f3266bb45a2964",
"title": "A Digital Signature Based on a Conventional Encryption Function"
},
{
"paperId": "120de9ed3ad57dfa2f9297aaad25f63a708be89a",
"title": "Blockchain for Education: Lifelong Learning Passport"
},
{
"paperId": null,
"title": "Last Access 20"
},
{
"paperId": null,
"title": "Acclaim Digital badges"
},
{
"paperId": null,
"title": "Sony Global Education Develops Technology Using Blockchain for Open Sharing of Academic Proficiency and Progress Records"
},
{
"paperId": null,
"title": "The open standard for blockchain credentials"
},
{
"paperId": null,
"title": "Open Badges Standard,"
},
{
"paperId": "4e9ec92a90c5d571d2f1d496f8df01f0a8f38596",
"title": "Bitcoin: A Peer-to-Peer Electronic Cash System"
},
{
"paperId": "ee1b2191e6de66a8c45d3cbafda96a7262780b61",
"title": "Bitcoin: A Peer-to-Peer Electronic Cash System"
},
{
"paperId": null,
"title": "Satoshi"
},
{
"paperId": null,
"title": "Blockchain Engineering: Challenges and Opportunities for Computer Science Research"
}
] | 9,121
|
en
|
[
{
"category": "Economics",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/025db54d117d024615422b3cee16c1d9f46e15ce
|
[] | 0.929581
|
Could the Issuance of CBDC Reduce the Likelihood of Banking Panic?1
|
025db54d117d024615422b3cee16c1d9f46e15ce
|
Journal of Central Banking Theory and Practice
|
[
{
"authorId": "2218880205",
"name": "Soraya Ben Souissi"
},
{
"authorId": "103423786",
"name": "M. Nabi"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"J Central Bank Theory Pract"
],
"alternate_urls": null,
"id": "b665ae3f-05b5-443a-98de-ecbe4779954d",
"issn": "1800-9581",
"name": "Journal of Central Banking Theory and Practice",
"type": "journal",
"url": "http://www.degruyter.com/view/j/jcbtp"
}
|
Abstract This paper delves into the relationship between the issuance of Central Bank Digital Currencies (CBDC) and the likelihood of banking panic. The issuance of CBDC acts as a disturbing shock that incentivizes depositors to withdraw all/part of their deposits from the commercial banks, to swap it for CBDC which are offered by the central bank. We determine a variety of tools that central banks can use in order for the issuance of CBDC to act as a stabilizing factor of the banking system (by reducing the likelihood of banking panic).
|
_Journal of Central Banking Theory and Practice, 2023, 2, pp. 83-101_
_Received: 08 February 2022.; accepted: 13 June 2022_
## Soraya BEN SOUISSI [*], Mahmoud Sami NABI[ **]
# Could the Issuance of CBDC Reduce the Likelihood of Banking Panic?[1]
**Abstract: This paper delves into the relationship between the issu-**
ance of Central Bank Digital Currencies (CBDC) and the likelihood
of banking panic. The issuance of CBDC acts as a disturbing shock
that incentivizes depositors to withdraw all/part of their deposits
from the commercial banks, to swap it for CBDC which are offered
by the central bank. We determine a variety of tools that central
banks can use in order for the issuance of CBDC to act as a stabilizing factor of the banking system (by reducing the likelihood of
banking panic).
**Keywords: Central bank digital currency, liquidity, financial stabil-**
ity.
**JEL classifications: E31, E42, G11.**
## 1. Introduction[1]
_UDK: 336.711:004_
_DOI: 10.2478/jcbtp-2023-0015_
_* University of Carthage,_
_LEGI-Tunisia Polytechnic School_
_and FSEG Nabeul, Tunisia_
_E-mail:_
_soraya.souissi.ss@gmail.com_
_** University of Carthage,_
_LEGI-Tunisia Polytechnic School_
_and FSEG Nabeul, Tunisia_
_ERF, Economic Research Forum,_
_Egypt_
_E-mail (Corresponding author):_
_mahmoudsami.nabi@ept.rnu.tn_
The determinants and impacts of Central Bank Digital Currency (CBDC)’s issuance is the subject of an increasingly number of research papers and experimental projects by central banks. Unlike crypto-currencies which are not backed by
any sovereign authority, CBDC is considered as a new form of central banks currencies. This new form of sovereign currency is expected to contribute to faster,
easier, cheaper and more secure financial transactions. The effects of issuing this
new currency are not yet well understood. While some researchers and financiers (e.g. Davoodalhosseini, 2018; Panetta, 2018; Cooper, Esser and Allen, 2019;
1 The authors declare that the current research has not benefited from any source of funding and
that there are no conflicts of interest with third parties.
-----
Kaczmarek, 2022) emphasize its benefits, others are more sceptic. The main argument against the issuance of CBDC is related to the financial stability issue
that might be exacerbated by bank run. For example, Genberg (2020) argues that
the issuance of CBDC could threaten the intermediation function of commercial
banks. Besides, it is not clear if CBDC will replace cash or if it will be considered
a financial asset?
This paper tries to contribute to this nascent literature by investigating the impacts of CBDC’s issuance on financial stability. Our paper is in line with Kim and
Kwon (2019) and Brunnermeier and Niepelt (2019) which analyse the conditions
under which CBDC issuance does not affect financial stability. It studies the effects of CBDC issuance on financial stability through a simplified model based
on Kim and Kwon` paper with three main modifications: space, time, and investment choice. In our model, CBDC does not exist initially, and is issued at the
end of the first period. This shock incentivizes the depositors to withdraw their
deposits from the commercial banks and swap it (totally or partially) for CBDC.
We then study the effect of this event on financial stability and the possible options to preserve it. We show that avoiding the bank run is possible if the central
bank transfers the CBDC into loans for the commercial banks, in an attempt to
preserve the stability of the reserve-deposit ratio. In addition, we show that this
is not the only possible option. Indeed, the central bank could also intervene by
restricting the access to CBDC accounts, either by limiting its available amount
or by imposing a substitution fee. The next option is to suspend the convertibility
of bank deposits into CBDC (à la Diamond and Dybvig, 1983). For this option, we
show that the proportion of lenders converting their deposits into CBDC shall be
kept below an endogenously determined bank panic cut-off.
The remaining of this paper is organized as follows. We begin with a literature
review on the economics of CBDC. The second section presents the theoretical
model which is used to analyse the effects of CBDC issuance on financial stability. The third section is devoted to the analysis of the equilibrium without CBDC
issuance. In the fourth section, we study the impacts of CBDC issuance on financial stability. Finally, we determine various other options that could limit the
impacts of CBDC issuance on financial stability.
## 2. The literature review
There are emerging studies analysing the determinants and impacts of central
bank digital currency (CBDC) issuance. Auer and Böhme (2020) focus on the
economic and institutional drivers of CBDC projects. It suggests a CBDC pro
-----
ject index taking higher values in countries where mobile phone use is widespread and innovation capacity is developed. Cooper et al. (2019) and Vučinić
and Luburić (2022) show that CBDC can accelerate the financial inclusion by facilitating the interoperability of the payment systems, improving their efficiency,
and reducing the financial costs and risks. These studies show that the impacts
of CBDC are diverse and not all positive. Indeed, this digital currency can have a
destabilizing impact on banking intermediation and a negative effect on financial
and digital equality. For Panetta (2018), if the CBDC are used as a means of payment, they will have a positive effect on financial inclusion. The main idea is that
a proportion of consumers who do not have bank accounts could use it without
incurring the cost of holding bank accounts. From the perspective of the central
banks, the use of CBDC are expected to reduce the cost of using cash. If CBDC
are used as a store of value, they will be considered as assets without cost for the
economic agents (who will no longer have to bear the fees due to the management
of their deposit accounts). CBDC would be more suitable than bank deposits if
they are issued as liquidity-free assets with a rate of return. But this option is not
free of impacts on the financial stability, since it could generate bank runs and
impact the intermediation role of commercial banks. In this context, Vučinić
(2020) shows that FinTech could have an adverse systemic impact on financial
stability through microfinancial and macrofinancial risks. CBDCs can have this
same impact on financial stability since they are part of fintech.
Bindseil (2020) analyses the effect of CBDC creation in two forms: by replacing
banknotes and by replacing bank deposits. It concludes that the first form has a
neutral effect on financial stability while the second form does not. Some other
studies tried to analyse the impacts of CBDC on financial stability by using general equilibrium models or attempting to analyse the neutrality conditions of the
introduction of this new currency. Brunnermeier and Niepelt (2019) develop a
general model of money, liquidity, and financial frictions and attempts to define
the equivalence conditions between different monetary systems. The exchange
equivalence between a private and a public currency is studied. The authors analyse if CBDC’s issuance affects the allocations and equilibrium prices. They show
that the issuance shall be accompanied by measures that guarantee wealth and liquidity neutrality. Besides, they show that a substitution operation accompanied
by open-market operations and transfers has no effects on wealth and liquidity.
Kim and Kwon (2019) propose an OLG model with agents that move between
locations, in which CBDC competes with bank deposits and is accessible to all
agents in all locations. It shows that an increase in CBDC deposits leads to a
reduction in commercial bank deposits, and to an increase of the probability of
banking panic. The authors argue that the financial system could keep its stability if the central bank uses CBDC deposits to extend credit to commercial
-----
banks. Mersch (2017) confirms this idea regarding the destabilizing effect of the
introduction of CBDC, mainly through the increase in the risk of deposit flight.
Bitter (2020) tries to study how the introduction of CBDC affects the likelihood
of aggregate bank runs. It shows that CBDC does not affect the aggregate output
and prices in a steady state. However, it changes the composition of household
savings, bank funding and capital investment. Besides, central banks can accommodate CBDC in their balance sheet via some options as loans to banks and corporate asset purchases. The author concludes that these two CBDC policies have
a stabilizing effect on the economy during crises.
Brunnermeier and Niepelt (2019) and Kim and Kwon (2019) suggest different
approaches of measures to reduce the negative impacts of CBDC on financial
stability. The former considers the neutrality of the equilibrium allocations while
the latter focuses on a cut-off threshold that triggers banking panic. Nevertheless, the two studies converge to quite similar conclusions. Indeed, they show that
for the CBDC’s issuance not to affect financial stability, the central bank should
activate specific instruments. Brunnermeier and Niepelt suggest open market
operations and clearing transfers, whereas Kim and Kwon (2019) propose the
refinancing of commercial banks by the central bank. In the same vein, Kumhof
and Noone (2018) outline the following four principles that must be followed
in order to control the impacts of CBDC issuance on financial stability: i) Payment of an adjustable interest rate for CBDC, ii) Distinction between reserves
and (non-convertible) CBDC, iii) No convertibility of bank deposits into CBDC
hold in commercial banks, iv) Issuance of CBDC against eligible securities. The
payment interest rate on CBDC should be adjustable so that it can be used as a
monetary policy tool to maintain financial stability, price stability, and parity
between bank deposits and CBDC.
## 3. The model
We consider a two-periods and three-dates _t = 0,1,2 economy. There is a [0,1]_
continuum of agents with a unit mass. Agents live for two periods. In the first
period, they are young and starting from t = 1 they become old. Half of the agents
are lenders while the other half are borrowers. The preferences of agents are described by the following utility function:
(1)
Where is consumption in period j. β is a stochastic discount factor. Lenders have an initial endowment x _> 0 of consumable good when they are young_
-----
and no donations when they are old. Borrowers have no donation when they are
young and have a donation y > 0 when they are old. We assume that βx > y. At
time _t = 0, there is a continuum of old agents with a unit mass. They have an_
initial donation of money M > 0 and from this time there is no injection or withdrawal of money. At the beginning of the first period, agents receive their donations. The young lenders will use their allocation to purchase goods and services
and invest the rest as deposits in commercial banks that will be remunerated at
the end of the first period. Young borrowers contact commercial banks to get
loans in the first period which they will repay increased by interest at the beginning of the second period. Finally, we consider that cash exists in the economy
and is used by agents to make transactions.
We assume that the central bank chooses to issue the CBDC during at t = 1 as a
liquid and non-risky asset which is accessible directly to agents. The central bank
keeps the CBDC accounts and pays a remuneration (r[c]) that compete with bank
deposits. The central bank purchases government securities at a rate Rc. The model has a finite number of commercial banks that live forever. They hold reserves,
collect deposits and grant loans. Each bank announces its repayment schedule at
_t = 0 and the interest rate that will be charged on deposits for each unit deposited_
according to the type of lender.
After the issuance of the CBDC during the first period t = 1, a random fraction
_π of young lenders called "swappers" decide to invest a part or all of their com-_
mercial bank deposits in CBDC. Thus, a lender may have a diversified portfolio of
commercial bank deposits and CBDC deposits. Swappers contact their banks to
withdraw their deposits. We denote F(π) the distribution function of the random
variable π and f(π) its continuously differentiable density function.
## 4. The financial equilibrium without CBDC
### 4.1. Agents’ problem
At the beginning of the first period, lenders receive their initial donations, consume and decide on the amount to deposit. Commercial banks decide on the
interest rate that will remunerate each type of lender: r[s](π) if it is a swapper and
_r(π) if it is not. Once the repayment schedule is announced, banks accept deposits_
and set the interest rate R applied to loans. Each lender chooses the deposit level
_d[b] that maximizes its expected utility given the payment scheme announced at_
time t = 1. At the beginning of the first period (t = 0), the lender invests all his
capital in a commercial bank. His expected utility is:
-----
The volume of deposits that maximizes this utility is given by:
(2)
(3)
On the other hand, a borrower observes the competitive interest rate R given at
time t = 0 and determines the amount of the credit he will apply for in order to
maximise his expected utility, whose expression is:
(4)
The borrower chooses the optimal amount of credit whose expression is:
(5)
### 4.2. Commercial banks’ problem
Commercial banks hold reserves, collect deposits and grant credit. They hold reserves z for any positive amount of deposits in commercial banks d[b]. They grant
credit for the remaining amount:
(6)
Let be the reserve-deposit ratio decided by the central bank. The reserves are remunerated at the rate where pt is the inverse of the price level
at time t =1, 2. To simplify, we consider that we are in the case of a stationary
equilibrium where pt+1 = pt.
### 4.3. Equilibrium without CBDC
The gross interest rate is given by R > 1. At the equilibrium, the total amount of
deposits made by the young lenders reduced by the number of bank reserves,
should allow the commercial banks to cover the demand for credits. Consequently, using equation (5), we obtain the market equilibrium condition:
(7)
We can then derive the expression of the nominal interest rate in equilibrium in
a regime where only fiat money exists:
-----
(8)
## 5. Effects of CBDC issuance on financial stability
We now consider that the central bank decides to issue the CBDC and charges an
interest rate r[c]. In the following sections, we study the effects of this new introduction on the behaviour of agents and the initial equilibrium values.
### 5.1. Agents and commercial banks problem
When the central bank announces the issuance of CBDC during a specified period t = 1, the young lender is proposed three strategies:
- Keep the full deposit in the commercial bank: d[b] = d,
- Withdraw all its deposits from the commercial bank and transfer it to a
CBDC account at the central bank: d[c] = d,
- Withdraw a proportion θ of its deposits to convert it into CBDC, where
In case the lender decides to invest in a CBDC account at the central bank, he will
be qualified as a swapper. The expression of the utility is then given by:
(9)
The lender will always choose the optimal deposit level that maximizes his utility:
(10)
For borrowers, nothing is altered and the amount of credit they apply for is always the same. Once γ is chosen, and credits are granted, lenders who choose to
invest (swappers) in CBDC and whose proportion is equal to (π) they will make
withdrawals in an amount that equals:
(11)
This withdrawal amount is paid by the commercial bank in the form of banknotes since the latter do not convert deposits into CBDC. It is assumed that there
is a fraction α(π) of bank reserves intended for swappers, where α ∈ [0,1].
-----
### 5.2. Equilibrium in case of CBDC investment
Considering that commercial banks do not make profit at the equilibrium, the
values of r[s] (π), r(π), r[c], α(π) and γ should be chosen in order to maximize the utility which is expressed as follows:
(12)
Such as:
(13)
(14)
(15)
The optimal solution must satisfy equations (13) and (14) as equalities. In this
case, and assuming [2], we can determine the maximal level of bank
reserves that can be withdrawn by swappers:
(16)
where and . This fraction of reserves cannot exceed the
maximum value of 1. Therefore, we can define as the bank panic cut-off point
for which, i.e. all bank reserves will be liquidated by agents switching to CBDC, and above which commercial banks can no longer satisfy liquidity
withdrawal demand. Considering equation (16), we have:
(17)
This value is interpreted as the cut-off value for the probability of migration to
CBDC that can generate a deposit run. If the amount of CBDC deposit per individual is low (θ is small), the cut off value of will be high, signifying low
exposure to bank panic. In particular, there is no banking panic for since
. It is already clear that this restriction emanates as one of the tools that
the central bank can use to mitigate the negative effects of the CBDC issuance on
financial stability. As shown in Graph 1, below the threshold, the bank panic
threshold is too high and the stability of the financial system can be preserved.
2 Under the pressure of bank competition and under the condition of no bank panic, which will
be made explicit later, following Kim and Kwon (2019).
-----
**Graph 1: Cut-off variation as a function of CBDC conversion ratio**
Source: Authors` simulations
The bank panic threshold depends not only on but also on the reserve-deposit
ratio y. In Graph 2, we vary this ratio for a given interest rate R. We observe that
as y increases, the bank panic threshold increases for a given CBDC conversion
ratio. This means that the more reserves the commercial banks hold to satisfy
the liquidity needs of their customers (who want to convert their deposits into
CBDC), the longer it takes for the banking panic to trigger. This same threshold
of banking panic is weakly sensitive to a variation in the interest rate R applicable
to bank loans, if we keep the reserve-deposit ratio of commercial banks at the
same level (Graph 3).
**Graph 2: Impact of the variation in the reserve-deposit ratio**
Source: Authors’ simulations
-----
**Graph 3: Impact of interest rate changes**
Source: Authors` simulations
However, what is more interesting is the analysis of determinants of the highest level (see section V) for a given swap proportion . From another viewpoint, condition (17) means that if depositors wish to convert all their deposits
into CBDC ( ) corresponding to ), the cut-off probability of
switching to CBDC will be equal to the level, which itself depends on the
interest rate charged on the loans and on the reserve-deposit ratio. The higher
this ratio is, the more reserves commercial banks have to pay to swappers and
the banking panic phenomenon takes longer to appear. On the other hand, if
deposits are not converted into CBDC or if this proportion is close to zero, the
cut-off value of banking panic will incline towards infinity. Thus, bank panic is
less likely to occur.
If the probability of lenders leaving to CBDC remains below the bank panic cutoff ( ), then we will have and commercial banks have enough
reserves to honour all liquidity demands. In the presence of banking competition
and in a market characterized by stability, i.e. the absence of a banking panic, the
interest rate applied to deposits for swappers and non-swappers is the same and
is given by the expression:
(18)
However, if then, the commercial banks use all their bank reserves to pay the liquidity requests and we are then faced with a situation of bank
run. In this case, the interest rates charged to swappers and non-swappers are no
longer equivalent and we have:
-----
(19)
(20)
These two expressions allow us to conclude that . Hence, in case of bank
run, swappers will be disadvantaged compared to non-swappers, since they will
have a lower remuneration. The optimal strategy for commercial banks is then:
By analysing equation (17) the cut-off of banking panic is an increasing function of the reserve-deposit ratio. The higher this ratio, the higher the threshold,
and thus the lower the probability of a banking panic. On the other hand, if
lenders choose to decrease their bank deposits in favour of CBDC deposits at the
central bank, they will make massive liquidity withdrawals, bank reserves will
decrease and the bank panic cut-off will also decrease.
It can be concluded that if lenders choose to convert their deposits into CBDC
following its, they will need to make liquidity withdrawals. Commercial
banks can meet this withdrawal demand as long as it does not exceed the threshold of its reserves. Otherwise, reserves will fall to zero and the proportion of
lenders leaving banks will approach the limit for the banking panic. Following
Kim and Kwon (2019), we define the optimal reserve-deposit ratio. Let us define
the function as:
(21)
The optimal reserve-deposit ratio can be expressed as:
(22)
Considering equation (17), we can then rewrite equation (22) as:
(23)
The function is decreasing and concave in
and for all . If, then (23) is satisfied only by
. If, then (23) has two solutions and the interior solution solves the
optimization problem. Results of the optimal choice of can be summarized as
follows: the optimal reserve-deposit ratio is given by ; with
and .
-----
### 5.3. The general equilibrium in case of CBDC issuance
**_Proposition 1_**
In equilibrium, the CBDC issuance increases the nominal interest rate which is
given by:
(24)
**Proof. At the equilibrium, the total amount of bank deposits must cover the**
bank's reserves and the granted loans. In other words:
(25)
We then have the following equilibrium condition:
(26)
It is clear from equation (26) that if the proportion of bank deposits converted
into CBDC increases, then the nominal interest rate also increases. Consequently, there will be less lending by commercial banks, for a given deposit-to-reserve
ratio. Therefore, if (absence of CBDC), the nominal interest rate is at its
minimum threshold compared to interest rate, in the presence of CBDC. In other
words, the issuance of CBDC will be more expensive for the borrowers. The increase in the interest rate could have an impact on the volume of granted loans.
However, the decrease is not only due to the increase in the nominal interest rate,
but also to the declining volume of private deposits since there will be a run-off
of lenders` deposits to CBDC accounts at the central bank.
## 6. Limiting the negative effects of CBDC on financial stability
In this section, the various strategies for dealing with the potential effects of
CBDC issuance on financial stability are analysed. We have seen previously, that
following this issuance, which occurs at, lenders could withdraw their deposits from commercial banks and convert it into CBDC accounts at the central
bank. Commercial banks must hold enough reserves to meet this need for liquidity. Beyond a certain limit withdrawals could generate a bank panic and reduce the amount of granted loans. To overcome this panic, Kim and Kwon (2019)
propose that the central bank use CBDC deposits to extend credits to commercial
banks, which could then use the new reserves to pay lenders. Bitter (2020) shows
-----
that under two different scenarios: loans to banks and corporate asset purchases,
CBDC issuance does not destabilize the economy. On the contrary, it could improve financial stability by postponing the emergence of bank run equilibrium.
The authors opts for the principle of managing the issuance of CBDC through
the interest rate. In the same context, Gross and Schiller (2021) show that the
central bank can decrease the remuneration of CBDC in order to reduce the volume held. However, the authors show that if CBDC are not interest-bearing, the
central bank cannot govern the demand for its digital currency.
### 6.1. Intervention on the volume of CBDC
Here, we study the possibility of central bank intervention through the volume of
issued CBDC . This intervention instrument can be applied to cases where CBDC
are issued with or without interest.
**_Proposition 2_**
To prevent a bank run and provide the required liquidity to lenders, conversion
to CBDC should not exceed the highest level given by:
(27)
**Proof. We saw in the previous section that when the central bank issues its new**
form of money at, a proportion of young lenders (swappers) equal to will
choose to invest a volume of their deposits in CBDC. We showed that there is
a threshold at which there is a banking panic phenomenon for any,
leading to financial instability. Let us consider an overlapping generation model,
extending the time horizon to infinity. In this context, it is possible for the central
bank to prevent this banking panic with new generations being born starting
from, by constraining the volume of CBDC to be converted to a ceiling, based on previous periods observations. Accordingly, and in order
to avoid a banking panic, the demand for withdrawals of deposits that would
be converted into CBDC has to be at most equal to the proportion of reserves
left for swappers. Since the central bank has already observed the proportion of
swappers at, it has to implement a new strategy to avoid a banking panic
arising from the conversion of deposits into CBDC. This implies that:
(28)
Replacing by and by its expression in (17), we can easily deduce the
expression of .
-----
The highest level of conversion to CBDC is established by replacing in (28) by
its expression in equation (18). Therefore, if the proportion of swappers in the
preceding period is observed, and knowing the cut-off point for the bank panic,
the central bank can limit the issuance of CBDC so that it does not exceed the
limit defined in equation (28). The highest volume of CBDC to be issued is a
decreasing function of the swapper ratio observed, given a defined reserve-todeposit ratio, as shown in Graph 4. In case the bank reserves of a period
are fully liquidated, the proportion of swappers at time is equal to
the bank panic threshold, then . The expression (28) turns into:
(29)
**Graph 4: Relation between CBDC issuance**
**and bank run**
Source: Authors’ simulations
Then the maximum volume of CBDC
issued by the central bank during a period following a banking panic, will
depend on the volume of CBDC issued in the previous period. We agree
with the findings of Panetta (2018) that
the strategy of limiting the volume of
CBDC can reduce the risk of a bank
run, while developing an approximate
expression of this maximum volume.
If the central bank chooses to adopt
a strategy of limiting the volume of
CBDC, it will not be able to issue an additional quantity of this new money if
the demand for it increases. In this case,
the adjustment shall be made through
the interest rate applied to CBDC.
### 6.2. Intervention through commercial banks' suspension of convertibility
We saw previously that the central bank can intervene through the limitation
of CBDC issuance. Reducing the probability of deposit flight is not limited to
the central bank, but can be done by commercial banks using the convertibility
suspension tool, as analysed in Diamond and Dybvig (1983). In this model, a
deposit contract between the commercial bank and the lender fixes the deposit
remuneration. Besides, the withdrawal of liquidity by the agents is done sequentially until the bank reserves are exhausted. We have already defined as the
amount of deposits that will be withdrawn from the bank in the form of cash to
-----
be converted into CBDC. We also assume that the central bank sets a maximum
threshold for conversion to individual CBDC that is the same for all lenders.
We assume that liquidity demanders can make liquidity withdrawals of deposits
in a sequential manner and in a well-defined order. Furthermore, we assume that
banks accept each agent's withdrawal request given only his position in the queue
and without any additional information about the behaviour of agents who are
ranked after him. Finally, we assume that lenders are served in order. We denote
by V1 the remuneration of deposits that will be withdrawn. This remuneration
depends on the lender's position in the line-up at period t. We will define as the
total number of deposit withdrawal demands and as the number of withdrawal
demands served before individual j. is a fraction of . Thus, we have:
(30)
**_Proposition 3_**
The bank deposit convertibility is suspended as soon as in order to avoid
the exhaustion of the fraction of reserves intended for swappers and so the
emergence of bank run.
(31)
**Proof. In order to determine the expression of, we propose the following rea-**
soning. We know that each lender will invariably have a total deposit amount
equal to . Once and are decided and
the volume of CBDC allowed for conversion, is chosen by the central bank,
commercial banks must set the number of withdrawals made by lenders such
that:
(32)
For all withdrawal sequences that occur before reaching the point, we are in
an equilibrium situation with CBDC where agents are seeking to maximize their
utility described in equation (8). Yet, we know that at the equilibrium, each lender
(j) will choose the level of deposit that maximises its utility, i.e.:
-----
(33)
By substituting this individual deposit expression into equation (32), we can then
derive the expression for .
If the number of withdrawals by lenders for conversion into CBDC, reaches the
level of, the banks no longer pay any remuneration and the agents have an incentive to keep their bank deposits until the end of the first period to receive their
remuneration. **This critical limit for the number of withdrawals is a decreas-**
**ing function of CBDC volume. An increase in the CBDC conversion volume**
**generates a lower threshold of convertibility suspension. In a situation where**
lenders choose to convert their deposits into CBDC, banks may choose to serve
them sequentially according to their position in the line-up until their reserves
are exhausted. At that point, the commercial banks pay no further remuneration
and agents have an incentive to keep their deposits at their banks until receiving
their final remuneration.
## 7. Conclusion
The issuance of CBDC by central banks is generating a lot of interest. New research is emerging to study their various economics impacts. Several authors
have shown that the introduction of CBDC could remain without effect on financial stability if it is accompanied by some measures by central banks such as open
market operations or granting of credits to commercial banks to guarantee the
stability of their reserves.
In this paper, we tried to analyse the impacts of CBDC’s issuance on financial
stability through a simplified model inspired from Kim and Kwon (2019). CBDC
are assimilated to non-risky liquid financial assets competing with bank deposits. The issuance of this new money takes place over a period of time after lenders
have already invested in bank deposits. We enable lenders to convert a part or
their total bank deposits into CBDC, and we analyse the impacts of such behaviour on the likelihood of banking panic. We show that under certain conditions, two strategies are possible to avoid the negative effects of CBDC issuance
on financial stability. The central bank could limit the volume of issued CBDC
to a predetermined threshold in order to avoid the occurrence of a bank run.
Commercial banks could also limit the convertibility of deposits into cash as
soon as the banking panic cut-off is reached. Although our results are consistent
with existing work on the impacts of CBDC issuance on financial stability, we of
-----
fer original recommendations in relation to the options available for the central
bank to mitigate these negative impacts.
-----
## References
1. Auer, R. and Böhme, R, (2020), “CBDC architectures, the financial system,
and the central bank of the future”, VOXEU – Center for Economic Policy
Research.
2. Bindseil, U., (2020), “Tired CBDC and the financial system”, European
Central Bank, Working Paper Series, No. 2351.
3. Bitter, L., (2020), “Banking crises under a Central Bank Digital Currency
(CBDC), Beiträge zur Jahrestagung des Vereins für Socialpolitik 2020 :
Gender Economics.
4. Brunnermeier, M.K. and Niepelt, D. (2019), “On the equivalence of private
and public money”, Journal of Monetary Economics, vol.106, October 2019,
p27-41.
5. Brunnermeier, M.K. and Sannikov, Y. (2016), “The I Theory of Money”,
NBER Working Paper Series, Working Paper 22533.
6. Cooper, B., Esser, A. and Allen, M. (2019), “The use cases of central bank
digital currency for financial inclusion: A case for mobile money”, The
_Centre for Financial Regulation and Inclusion._
7. Davoodalhosseini, S.M.R. (2018), “Central Bank Digital Currency and
Monetary Policy”, Bank of Canada, Staff working paper, 2018-36.
8. Diamond, D. and Dybvig, P. (1983), “Bank runs, Deposit Insurance, and
Liquidity”, Journal of Political Economy, 1983, vol.91, No. 3.
9. Genberg, H. (2020), “Digital transformation: Some Implications for
Financial and Macroeconomic Stability”, Macroeconomic stabilization in
the digital age, Asian Development Bank Institute, 68 – 85.
10. Gross, J. and Schiller, J. (2021), “A Model for Central Bank Digital
Currencies: Implications for Bank Funding and Monetary Policy”, Social
Science Research Network.
11. Kaczmarek, P. (2022) “Central Bank Digital Currency: Scenarios of
Implementation and Potential Consequences for Monetary System,” Journal
_of Central Banking Theory and Practice, Central Bank of Montenegro, vol._
11(3), pages 137-154.
12. Kim, Y. and Kwon, O. (2019), “Central Bank Digital Currency and Financial
Stability”, Bank Of Korea Working Paper.
13. Kumhof, M. and Noone, C. (2018), “Central bank digital currencies- design
principles and balance sheet implications”, Bank of England, Staff Working
_Paper No. 725._
14. Mersch, Y. (2017), “Digital base money : an assessment from the ECB
perspective”, speech, Helsinki,16 January 2017.
-----
15. Panetta, F. (2018), “21st Century Cash : Central Banking, technological
innovation and digital currencies”, Bocconi University.
16. Vučinić, M. (2020) “Fintech and financial stability : Potential influence of
Fintech on financial stability, risks and benefits,” Journal of Central Banking
_Theory and Practice, Central bank of Montenegro, vol. 9(2), pages 43-66._
17. Vučinić, M. and Radoica, L. (2022) “Fintech, Risk-based thinking and
cyber risk,” Journal of Central Banking Theory and Practice, Central bank of
Montenegro, vol. 11 (2), pages 27-53.
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.2478/jcbtp-2023-0015?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.2478/jcbtp-2023-0015, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "GOLD",
"url": "https://sciendo.com/pdf/10.2478/jcbtp-2023-0015"
}
| 2,023
|
[
"JournalArticle"
] | true
| 2023-05-01T00:00:00
|
[
{
"paperId": "e51c5aae3d405ef4a71b853fd49690df082956e8",
"title": "Central Bank Digital Currency: Scenarios of Implementation and Potential Consequences for Monetary System"
},
{
"paperId": "c9179af62222c2badd45cd18242903fb2d3a530c",
"title": "Fintech, Risk-Based Thinking and Cyber Risk"
},
{
"paperId": "b061a57abfada65588a3f4d5a8f44fc36de1f548",
"title": "Fintech and Financial Stability Potential Influence of FinTech on Financial Stability, Risks and Benefits"
},
{
"paperId": "5aa2691b106ea71b10c218a915f23b3fabc33f66",
"title": "Tiered CBDC and the Financial System"
},
{
"paperId": "64e80e34fc664d9d7d86bae35a398a164c1c3590",
"title": "On the Equivalence of Private and Public Money"
},
{
"paperId": "0debfa9cb8631487c8946d50a4658b76ed1bb33d",
"title": "Central Bank Digital Currency and Financial Stability"
},
{
"paperId": "1ca2403d065e0715c560978012972b7e71c21f47",
"title": "Central Bank Digital Currency and Monetary Policy"
},
{
"paperId": "665efab907e67c5a7ed457c1f5ad70bc0c030717",
"title": "Central Bank Digital Currencies - Design Principles and Balance Sheet Implications"
},
{
"paperId": "97396fbef37644f319c4ba035886137df55959f8",
"title": "The I Theory of Money"
},
{
"paperId": "5d2a97be97ed70c86addf5229ae2ffa5b39227e0",
"title": "Bank Runs, Deposit Insurance, and Liquidity"
},
{
"paperId": "b365b874e5fc8f1bd860014912acd47e656db0dd",
"title": "Digital Base Money: a few considerations from a central bank’s perspective"
}
] | 8,290
|
en
|
[
{
"category": "Medicine",
"source": "external"
},
{
"category": "Medicine",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/025e07ce9e6fd253b53c18e097d0ea6df186490f
|
[
"Medicine"
] | 0.897687
|
Knowledge structure and emerging trends on osteonecrosis of the femoral head: a bibliometric and visualized study
|
025e07ce9e6fd253b53c18e097d0ea6df186490f
|
Journal of Orthopaedic Surgery and Research
|
[
{
"authorId": "2119019381",
"name": "Haiyang Wu"
},
{
"authorId": "2150057421",
"name": "Kunming Cheng"
},
{
"authorId": "2077571148",
"name": "Linjian Tong"
},
{
"authorId": "2115663919",
"name": "Yulin Wang"
},
{
"authorId": "2109351991",
"name": "Weiguang Yang"
},
{
"authorId": "104642162",
"name": "Zhiming Sun"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"J Orthop Surg Res"
],
"alternate_urls": [
"http://www.josr-online.com/",
"http://www.josr-online.com/home"
],
"id": "0cd86699-aba1-4cf4-b9d4-b95bb87da7f6",
"issn": "1749-799X",
"name": "Journal of Orthopaedic Surgery and Research",
"type": "journal",
"url": "https://josr-online.biomedcentral.com/"
}
|
Background Osteonecrosis of the femoral head (ONFH) is a common disabling disease with considerable social and economic impacts. Although extensive studies related to ONFH have been conducted in recent years, a specific bibliometric analysis on this topic has not yet been performed. Our study attempted to summarize the comprehensive knowledge map, development landscape, and future directions of ONFH research with the bibliometric approach. Methods All publications concerning ONFH published from 2001 to 2020 were identified from Web of Science Core Collection. Key bibliometric indicators were calculated and evaluated using CiteSpace, VOSviewer, and the online bibliometric analysis platform. Results A total of 2594 publications were included. Our analysis revealed a significant exponential growth trend in the annual number of publications over the past 20 years ( R 2 = 0.9663). China, the USA, and Japan were the major contributors both from the quality and quantity points of view. Correlation analysis indicated that there was a high positive correlation between the number of publications and gross domestic product ( r = 0.774), and a moderate positive correlation between publications and demographic factor ( r = 0.673). All keywords were categorized into four clusters including Cluster 1 (etiology and risk factors study); Cluster 2 (basic research and stem cell therapy); cluster 3 (hip-preserving study); and Cluster 4 (hip replacement study). Stem cell therapy-related research has been recognized as an important research hotspot in this field. Several topics including exosomes, autophagy, biomarkers, osteogenic differentiation, microRNAs, steroid-induced osteonecrosis, mesenchymal stem cells, double-blind, early-stage osteonecrosis, and asymptomatic osteonecrosis were considered as research focuses in the near future. Conclusion Over the past two decades, increasing attention has been paid to global ONFH-related research. Our bibliometric findings provide valuable information for researchers to understand the basic knowledge structure, identify the current research hotspots, potential collaborators, and future research frontiers in this field.
|
_p_ _g y_ _(_ _)_
https://doi.org/10.1186/s13018-022-03068-7
## RESEARCH ARTICLE
## Open Access
# Knowledge structure and emerging trends on osteonecrosis of the femoral head: a bibliometric and visualized study
### Haiyang Wu[1*†], Kunming Cheng[2†], Linjian Tong[1], Yulin Wang[1], Weiguang Yang[1] and Zhiming Sun[1,3*]
**Abstract**
**Background: Osteonecrosis of the femoral head (ONFH) is a common disabling disease with considerable social and**
economic impacts. Although extensive studies related to ONFH have been conducted in recent years, a specific bibliometric analysis on this topic has not yet been performed. Our study attempted to summarize the comprehensive
knowledge map, development landscape, and future directions of ONFH research with the bibliometric approach.
**Methods: All publications concerning ONFH published from 2001 to 2020 were identified from Web of Science Core**
Collection. Key bibliometric indicators were calculated and evaluated using CiteSpace, VOSviewer, and the online
bibliometric analysis platform.
**Results: A total of 2594 publications were included. Our analysis revealed a significant exponential growth trend in**
the annual number of publications over the past 20 years (R[2] 0.9663). China, the USA, and Japan were the major
=
contributors both from the quality and quantity points of view. Correlation analysis indicated that there was a high
positive correlation between the number of publications and gross domestic product (r 0.774), and a moderate
=
positive correlation between publications and demographic factor (r 0.673). All keywords were categorized into four
=
clusters including Cluster 1 (etiology and risk factors study); Cluster 2 (basic research and stem cell therapy); cluster 3
(hip-preserving study); and Cluster 4 (hip replacement study). Stem cell therapy-related research has been recognized
as an important research hotspot in this field. Several topics including exosomes, autophagy, biomarkers, osteogenic
differentiation, microRNAs, steroid-induced osteonecrosis, mesenchymal stem cells, double-blind, early-stage osteonecrosis, and asymptomatic osteonecrosis were considered as research focuses in the near future.
**Conclusion: Over the past two decades, increasing attention has been paid to global ONFH-related research. Our**
bibliometric findings provide valuable information for researchers to understand the basic knowledge structure, identify the current research hotspots, potential collaborators, and future research frontiers in this field.
**Keywords: Osteonecrosis of the femoral head, Bibliometric analysis, Hotspots, VOSviewer, CiteSpace**
*Correspondence: nfykdxwhy@126.com; szhm618@163.com
†Haiyang Wu and Kunming Cheng contributed equally to the study.
1 Graduate School of Tianjin Medical University, No. 22 Qixiangtai Road,
Tianjin 300070, China
3 Department of Orthopaedic Surgery, Tianjin Huanhu Hospital, No 6,
Jizhao Road, Jinnan District, Tianjin 300350, China
Full list of author information is available at the end of the article
**Introduction**
Osteonecrosis of the femoral head (ONFH) is a common
progressive disease typically characterized by reduction
in vascular supply, bone metabolism disorder, and necrosis of the subchondral bone and eventually resulting in
bone collapse of the femoral head [1]. It can be classified
as traumatic and non-traumatic ONFH on the basis of
diverse etiologies. Although the pathophysiology of this
process has not yet been clearly elucidated, corticosteroid
© The Author(s) 2022. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which
permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the
original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or
other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line
to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory
regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this
[licence, visit http://creativecommons.org/licenses/by/4.0/. The Creative Commons Public Domain Dedication waiver (http://creativeco](http://creativecommons.org/licenses/by/4.0/)
[mmons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data.](http://creativecommons.org/publicdomain/zero/1.0/)
-----
use, alcoholism, smoking, inherited coagulation disorders, as well as systemic lupus erythematosus are typically considered to be high-risk factors for non‐traumatic
ONFH [1, 2]. In the USA alone, more than 20,000 new
patients were diagnosed with non-traumatic ONFH each
year, contributing to approximately 10% of the total number of total hip arthroplasties (THA) performed annually
[3]. Another epidemiological study estimated that around
8.12 million Chinese population aged 15 years or over
were affected by this condition, among which 55.75% of
females and 26.35% of males have reported corticosteroid
use [4]. Therefore, ONFH represent a major challenge in
the orthopedic arena due to its high morbidity and disability rate, especially among young and middle-aged
people.
Many tactics for treatment of ONFH depend on the
severity of their condition, and the staging system formulated by the Association Research Circulation Osseous
(ARCO) is one of the commonly used staging method
in clinical practice [2]. According to the changes in the
intraosseous blood supply in different phases of disease
progression, the corresponding surgical and nonsurgical treatment strategies are recommended to prevent, or
at least delay the progression toward the stage of femoral head collapse, in which THA is unavoidable [2, 5].
Despite the availability of numerous hip‐preserving surgical methods including core decompression, osteotomy,
and bone graft, there still exists controversy regarding
whether these treatment modalities are meaningful and
valuable, as more than 80% patients with ONFH finally
require THA [6, 7]. Additionally, some other controversies derived from ONFH, such as the pathogenesis, the
optimal classification system, practicability of pharmacological treatments, optimal treatment protocols and surgical timing of THA, predictors of outcomes, and so on
[2, 3]. Motivated by these concerns, ONFH have piqued
the interest of researchers worldwide and a large number of related papers on this challenging topic have been
published. To our knowledge, although some systematic
reviews focusing on a specific subfield of ONFH research
have been published, the global knowledge structure and
research trends in this area have not been systematically
studied yet.
Notably, the appearance of bibliometric method has
compensated the shortage of literature reviews in a
complementary fashion. Bibliometrics, first defined by
Pritchard in 1969, is a visualization method to quantitatively assess the contribution of a research field by using
mathematical and statistical approaches [8]. It is also
regarded as an important approach to reveal the research
trends and predict the research hotspots in a certain field
[9, 10]. Over recent years, the application of bibliometric analysis is very extensive in the biomedical sciences,
due to the explosion in the quantity of scientific publications and the availability of several freeware bibliometric
tools [11, 12]. In the field of hip surgery, one recent study
explored the global trends and hotspots of scientific
research on femoroacetabular impingement from 2000 to
2019, based on 2471 originals articles indexed in Web of
Science (WOS) [13]. Our research team and other groups
have also investigated the publications on hip fracture
[14], developmental dysplasia of the hip [15], and THA
[16] by using bibliometric methods and assessed the coauthorship and co-citation network in these areas. However, to date, no studies have applied the bibliometric
method to analyze the global research trends on ONFH.
In view of this, the aims of this study were to (1) identify the current status of ONFH domain, including the
distribution of annual outputs, the major players such as
countries, institutions, and individuals; (2) analyze the
cooperation networks at the level of countries, institutions, and authors; (3) summarize main research directions and hotspots; and (4) propose research frontiers
and potential hotspots in the near future.
**Methods**
**Source of bibliometric data and search strategies**
Based on previous studies [14, 17, 18], the Science Citation Index Expanded (SCI-Expanded) of the Web of
Science Core Collection (WOSCC) was selected as the
main data source. The scientific literature was searched
based on the titles (TI), abstracts (AB), and author keywords (AK) with the following search strategy: “femoral
head necrosis” OR [(osteonecrosis OR necrosis) NEAR/2
(“femoral head”)] OR ONFH. The proximity operator
of “NEAR/2” was used to combine search terms, which
means two terms may have separated by a maximum
of two words in any order (e.g., osteonecrosis NEAR/2
femoral head would have identified “osteonecrosis of
the femoral head” and “osteonecrosis of femoral head”).
A timespan of 20 years was set, and thus, only literature
published from 2001 to 2020 was included. Publication
language was restricted to English, and only original
articles and reviews were eligible for this bibliometric
analysis. All data utilized in this work were downloaded
from public databases and, therefore, ethics committee
approval or informed consent was not required.
**Data export and extraction**
Considering that the database is regularly updated, all
searches were done on a separate day to avoid this potential bias. By using the function of “export” in WoSCC,
“full records and cited references” of retrieved records
were exported as “tab delimited text (.txt)” to bibliometric tools for additional processing. Then the detailed data
on the general information including annual publications,
-----
countries, institutions, authors, source journals, funding
sources, research areas, number of citations, and Hirschindex (H-index) were extracted. The above procedure
was completed by two investigators independently, and
any disagreements were solved through discussion or, if
necessary, by the senior author. Moreover, journal impact
factors (JIF) and quartile ranks were collected from the
2020 Journal Citation Reports. The detailed literature
search and selection process are shown in Fig. 1.
**Bibliometric analysis**
To obtain a more comprehensive analysis, three bibliometric tools, including an online platform and two software, were used to perform this study. First, the online
[bibliometric analysis platform (available at: https://bibli](https://bibliometric.com/)
[ometric.com/) was used to conduct academic coop-](https://bibliometric.com/)
eration networks between countries. Then, VOSviewer
1.6.16 and Citespace V 5.7 R2 software were further
used for mapping and visualizing bibliometric networks
of scientific publications. VOSviewer, a freely available
Java-based software developed by van Eck and Waltman at Erasmus University, is one of the frequently
used bibliometric tools for quantitatively analyzing
the academic literature [11]. In this study, VOSviewer
was used to visualize the following network maps of
ONFH research: network map of co-citation authors
and journals; co-occurrence analysis of keywords.
Specifically, co-citation network means that two items
appear together in the bibliography of a third citing
item, while co-occurrence network represents that the
relationship of items is built according to the quantity
of publications where they occur together [8, 11]. Generally speaking, the visualization maps mainly consist
of nodes and links with different colors. Nodes in the
visualization map represented the analyzed elements
such as author, journal, or keyword, and the size of the
nodes indicated the number of citations or occurrences
[14]. The links between nodes reflected the relationship
**Fig. 1** Flow diagram of the literature search and selection process
-----
of co-citation or co-occurrence. An important parameter, total link strength (TLS), was used to quantitatively
evaluate the strength of links [11, 14]. And the detailed
descriptions of the maps could be found in the software
[manual at https://www.vosviewer.com/documentation.](https://www.vosviewer.com/documentation)
Apart from that, we also employed another bibliometric software, called Citespace, which was developed
by Professor Chaomei Chen of Drexel University, to
perform further bibliometric analysis [12]. In the present study, CiteSpace was applied to conduct research
cooperation relationships of authors and institutions;
timeline view map of co-citation references; and references with the strongest citation bursts. CiteSpace is
capable of generating different types of visualization
map, such as the network map, the cluster view map,
and timeline view map [12]. Overall, all these visualization maps are also comprised of nodes and lines representing different meanings. Betweenness centrality
(BC) is an important indicator that could identify the
relative importance of a node within the networks, and
nodes with the highest BC value (≥ 0.1) are usually
known as hubs nodes that usually marked with purple
rings [17, 19]. More detailed software utilization skills
and information about the visualization maps can be
[found in the CiteSpace manual (available at http://clust](http://cluster.ischool.drexel.edu/~cchen/citespace/CiteSpaceManual.pdf)
[er.ischool.drexel.edu/~cchen/citespace/CiteSpaceM](http://cluster.ischool.drexel.edu/~cchen/citespace/CiteSpaceManual.pdf)
[anual.pdf](http://cluster.ischool.drexel.edu/~cchen/citespace/CiteSpaceManual.pdf) ).
**Statistical analyses**
R software (v3.6.3.), SPSS (IBM SPSS Statistics 21, Inc.,
Chicago, IL, USA), and Microsoft Excel 2019 were used
for descriptive analysis, statistical evaluation, data fitting,
and plotting graphs. We computed the growth rate of
publications over time with the following calculation formula as described previously by Guo et al. [20]. Pearson’s
correlation coefficient test was used to assess the correlation between continuous variables, and correlations were
considered significant when p value < 0.05.
**Results**
**Publication outputs and trends**
Following the aforementioned screening strategy, a total
of 2594 documents including 2394 original articles and
200 reviews related to ONFH were identified covering the
period 2001–2020. Figure 2 presents the specific numbers of annual documents about ONFH. Model fitting
curve revealed a significant exponential growth trend in
the annual number of publications over the past 20 years
(y = 44.943e[0.088][x], _R[2]_ = 0.9663). From 2001 to 2020, the
average growth rate of scientific outputs was 33.41%.
**Analysis of countries/regions and funding agencies**
All publications were distributed among 71 countries/
regions. China had published the most publications
with 1077 (41.52%) articles/reviews, followed by the
**Fig. 2** The specific numbers of annual documents regarding ONFH
-----
USA [392 (15.11%)], Japan [278 (10.72%)], South Korea
[167 (6.44%)], and Germany [123 (4.74%)] (Table 1). The
results of correlation analysis indicated that there was a
high positive correlation between the number of publications and gross domestic product (GDP) (r = 0.774,
_p < 0.001), and a moderate positive correlation was_
also found between publications and demographic factor (r = 0.673, _p_ = 0.001). International collaboration of
countries in this domain was also analyzed. As demonstrated in Fig. 3A, extensive collaboration was observed
between productive countries. For instance, China collaborated closely with the USA, Australia, Germany, and
South Korea. The USA, South Korea, Greece, Japan, and
Italy have demonstrated active cooperation as well. In
addition, the annual number of publications in the top
10 prolific countries from 2001 to 2020 is illustrated in
Fig. 3B. Figure 3C lists the top 10 most active funding
agencies in this field. Of these, four of them were from
Japan, three from China, and the remaining three agencies were from the USA.
**Analysis of the most prolific institutions**
The top 10 most prolific institutions were laid out in
Fig. 3D. All these institutions were from Asian institutions including 7 Chinese institutions, 2 Japanese institutions, and 1 Korean institution. Among them, Shanghai
**Table 1 Top 20 most productive countries related to ONFH research**
Jiao Tong University had the largest number of publications (77), followed by Kyushu University (58), Guangzhou University of Chinese Medicine (51), and Xi’an
Jiaotong University (51). As for the other parameters,
H-Index for Osaka University exhibited the highest value
(18), followed closely by Shanghai Jiao Tong University
(17). And Seoul National University had the highest value
of average number of citations (23.05). A network visualization map of institution cooperation was generated by
CiteSpace and illustrated in Fig. 4A.
**Analysis of the most influential authors**
As shown in Fig. 4B, Zhang CQ from Shanghai Jiao Tong
University contributed the highest number of papers, followed by Zhao DW from Dalian University, and Motomura G from Kyushu University. Figure 4C illustrates
the cooperation network map of authors, none of the
included authors had a BC value of more than 0.1. In
addition, the co-citation network among authors was
conducted using VOSviewer. As displayed in Fig. 4D,
only authors with a minimum of 100 citations were
included. There were 55 nodes, 5 clusters, and 1439 links
in the network map. Among them, Mont MA from Sinai
Hospital of Baltimore has occupied maximum node with
the largest citations and TLS.
**Ranking** **Countries** **Publications, n** **% of 2594** **H-index** **Average**
**citations per**
**document**
1 China 1077 41.52 45 11.61
2 USA 392 15.11 55 29.35
3 Japan 278 10.72 33 16.28
4 South Korea 167 6.44 32 19.89
5 Germany 123 4.74 27 18.29
6 UK 102 3.93 28 24.54
7 France 73 2.81 23 25.14
8 India 59 2.27 15 12.63
9 Turkey 56 2.16 13 11.84
10 Italy 49 1.89 20 19.69
11 Greece 47 1.81 23 26.72
12 Canada 41 1.58 19 28.2
13 Switzerland 36 1.39 15 34.31
14 Belgium 31 1.20 18 36.13
15 Australia 27 1.04 12 27.52
16 Spain 27 1.04 10 11.85
17 Israel 24 0.93 12 38.13
18 Brazil 22 0.85 8 10.45
19 Austria 20 0.77 11 22.75
20 Iran 20 0.77 7 10.45
Ranking: according to the number of total publications
-----
**Fig. 3** **A International collaboration of countries in this domain. The area occupied per country is proportional to the number of documents. Line**
thickness reflects the closeness between countries, and a thicker line represents a stronger collaboration. B The annual number of publications in
the top 10 prolific countries from 2001 to 2020. From 2010 onward, China has surpassed the USA for the first time and has remained that way since
then. C The top 10 most active funding agencies in ONFH-related research. D The total number of publications, H-index, and average citations per
item of top 10 most prolific countries in this field
**Analysis of the higher‑impact journals**
The top 10 most prolific journals were listed in Table 2.
_International Orthopaedics (JIF 3.075) has published_
the greatest number of 123 papers, followed by _Clinical_
_Orthopaedics and Related Research (JIF 4.291), and Jour-_
_nal of Arthroplasty (JIF 4.757), with 103 and 81 publica-_
tions, respectively. According to JIF, JCR splits journals
belonging to the same discipline into four equal parts, of
the top 25% classified as Q1 and the top 25–50% being
Q2, and so on. More than half of the top 10 journals were
categorized in Q1 or Q2. Figure 5 shows the network visualization map of journal co-citation analysis. Only journals with more than 200 citations were depicted. Of the
60 journals satisfying the criteria, the top 5 co-cited journals were _Clinical Orthopaedics and Related Research,_
_Journal of Bone and Joint Surgery American Volume,_
_Bone & Joint Journal, Journal of Arthroplasty, and Inter-_
_national Orthopaedics._
**Analysis of highly cited references**
Reference co-citation analysis is one of the most attractive functions of CiteSpace, which is often applied to
determine the research focuses on a given field. As
shown in Fig. 6 and Table 3, all the nodes in the reference
co-citation network map could be grouped into 13 major
clusters. In CiteSpace, weighted mean silhouette value (S
value) and the modularity value (Q value) are two indicators to evaluate the significance of clustering, and it is
generally believed that _S > 0.7 and_ _Q > 0.3 represent the_
clusters are convincing. In this study, the mean value of
-----
**Fig. 4** **A The cooperation network map of institutions generated by CiteSpace. Each node represents an institution and the node size is**
proportional to the number of publications by that institution. The node with the highest BC value (≥ 0.1) are usually known as hubs nodes that
usually marked with purple rings. The connecting line between nodes indicates the cooperation relationship, and the value of link strength is also
displayed between lines. B The total number of publications, H-index, and average citations per item of top 10 productive authors in this field. C The
cooperation network map of productive authors generated by CiteSpace. The graphical explanations are the same as in A. D Authors co-citation
analysis by VOSviewer. Each node represents a different author, and the node size is proportional to the quantity of citations. The thickness of the
connecting line between nodes indicates link strength of a co-citation relationship, which could be weighted by a quantitative indicator that is TLS
**Table 2 Top 10 journals with most publications in the field of ONFH research**
**Ranking** **Journal title** **Output** **% of 2594** **JIF (2020)** **Quartile in**
**category**
**(2020)**
1 International Orthopaedics 123 4.74 3.075 Q2
2 Clinical Orthopaedics and Related Research 103 3.97 4.291 Q1
3 Journal of Arthroplasty 81 3.12 4.757 Q1
4 Journal of Bone and Joint Surgery American Volume 79 3.05 5.284 Q2
5 Archives of Orthopaedic and Trauma Surgery 69 2.66 3.067 Q3
6 Bone & Joint Journal 67 2.58 5.082 Q1
7 Hip International 54 2.08 2.135 Q4
8 Medicine 53 2.04 1.889 Q3
9 BMC Musculoskeletal Disorders 50 1.93 2.355 Q3/Q4
10 Journal of Orthopaedic Surgery and Research 48 1.85 2.359 Q2
-----
**Fig. 5** The network visualization map of journal co-citation analysis by VOSviewer. Each node represents a different journal, and the node size is
proportional to the quantity of citations. Other graphical explanations are the same as in Fig. 4D
**Fig. 6** The timeline view map of reference co-citation analysis. For each cluster, the position of each node shows the time of publication of the
document, and the node size represents the number of citations
S equals 0.7481 and that Q equals 0.7794, indicating the
rationality of this clustering strategy. Moreover, as shown
in Fig. 7, the top 25 references with the strongest citation
bursts were identified in terms of their burst values.
**Analysis of the most concerned keywords**
In this study, keywords that occurred at least 10 times
were extracted from 2594 publications and analyzed
by VOSviewer. After deleting meaningless keywords
-----
**Table 3 Main clusters of co-cited references**
**Cluster ID** **Size** **Silhouette** **Label** **Mean (year)**
#0 32 0.965 Risk factor 2015
#1 31 0.9 Glucocorticoid-induced osteonecrosis 2010
#2 25 0.881 Autologous transplantation 2006
#3 21 0.932 Non-weight bearing 1998
#4 21 0.972 Intertrochanteric osteotomy 1998
#5 21 0.934 Fibular grafting 2001
#6 20 0.915 Prognostic value 1999
#7 19 0.893 Stem cell therapy 2015
#8 19 0.788 t-786c enos polymorphism 2002
#9 17 0.846 Gene expression 2007
#10 17 0.967 Concentrated autologous bone marrow 2010
#11 14 1 Propylene fumarate 2004
#12 14 0.951 Early collapse 2002
**Fig. 7** Top 25 references with the strongest citation burst in the ONFH field. The red segment represents the begin and end year of the burst
duration
and merging keywords with the same meaning, a total
of 319 keywords were identified. Based on the research
categories of these keywords, VOSviewer software was
able to divide all keywords into several major clusters with different colors. As shown in the network
visualization map of Fig. 8A, all the included keywords
were classified into the following four clusters: Cluster
1(etiology and risk factors study, green nodes); Cluster 2 (basic research and stem cell therapy, red nodes);
cluster 3 (hip-preserving study, blue nodes); and Cluster 4 (hip replacement study, purple nodes). In addition
-----
**Fig. 8** **A Network visualization map of the co-occurrence network of keywords using VOSviewer. Each node represents a certain keyword. Nodes**
and font size represent the number of keyword occurrences. Keywords with close correlation will be assigned to one cluster with the same color. B
Overlay visualization map of keywords analysis in the ONFH field. The color of each node shows the average appearing year (AAY) of the keyword,
according to the color gradient shown at the bottom right. The blue-purple color reflected the keywords appeared relatively earlier, while the dark
red nodes represented the recent occurrence
to this, we also provided an overlay visualization map
of keywords co-occurrence analysis in Fig. 8B.
**Discussion**
**Worldwide research tendency of ONFH from 2001 to 2020**
The number of publications in a certain field is able to
reflect the productivity and development of the topic over
the years. ONFH-related research has drawn increasing
attention among scholars from 2001 to 2020. The global
number of publications in this field has been gradually
increased from 58 in 2001 to 297 in 2020, 71.05% of which
was published in the last 10 years. One important reason
for this growth was that the incidence of ONFH has been
increasing worldwide and this devastating condition has
become an increasingly prominent issue globally [1, 2, 21,
22]. Based on current trends, Mont et al. [23] reported
that the total number of individuals affected by ONFH is
estimated to reach 20 million by next 10 years worldwide.
Besides, the increase in annual publications is inseparable from the advances in basic research and clinical trials
in recent years [24, 25].
**Knowledge structure of ONFH‑related publications**
**_Countries_**
It is not difficult to see that the research centers of this
field were mainly concentrated in East Asia, North
America and Western Europe. The results of correlation analysis indicated that part of the discrepancy in
the quantity of publications across different countries
can be explained by factors of economy or population.
The H-index is a bibliometric indicator that simultaneously measures the quality (mainly depends on citations)
and quantity (number of documents) of publications in
a journal, author, or country [19, 26]. In this study, the
USA, China, and Japan were the top three countries with
the highest H-index. Therefore, this result further proved
-----
that China, the USA, and Japan were the major contributors both from the quality and quantity points of view.
**_Institutions_**
In terms of research institutions, the most prolific institution was Shanghai Jiao Tong University, Kyushu University, Guangzhou University of Chinese Medicine, and
Xi’an Jiaotong University. Nevertheless, inter-institutional cooperation levels were relatively low and primarily
conducted in Asian research institutions. Furthermore,
none of the institutions had BC value greater than 0.1,
indicating that there was no one institution occupied the
absolutely central position in the collaboration network
[17, 19]. In view of this, strengthening cooperative networks in different research institutions or teams will be
important for future studies whether for basic scientific
researches or clinical trials.
**_Authors_**
An analysis of the most influential authors is helpful
for scholars to learn existing partnerships and identify
potential cooperative subjects at home and abroad. As
shown in Fig. 4B, Zhang CQ, Zhao DW, and Motomura
G were the top three contributors in this field. Zhang
CQ and colleagues mainly focused on the application
of free vascularized fibula grafting for the treatment of
ONFH [27]. Apart from that, a study conducted by their
research team, which reported the potential preventative effect of exosomes secreted by induced pluripotent
stem cell-derived mesenchymal stem cells (iPS-MSCExos) on ONFH via promoting local angiogenesis, also
received a great attention [28]. The co-citation analysis
is usually considered to be a better method to evaluate
the academic influence of a journal or a scholar [17]. As
a result, the co-citation network among authors was conducted using VOSviewer. Mont MA from Sinai Hospital
of Baltimore has occupied maximum node with the largest citations and TLS. Further analysis found that several
high-quality reviews regarding diagnosis, classification
systems, and treatment for ONFH, published by Mont
et al. have achieved a high number of citations [29, 30].
**_Journals_**
As for journal analysis, International Orthopaedics, Clin_ical Orthopaedics and Related Research, and_ _Journal of_
_Arthroplasty were the top three journals with the most_
publications. Of the top 10 journals, although China is
the largest publishing country, there is no one Chinese
journal, indicating that China should strengthen several
international journals in this field so as to attract more
scientific publications and spread academic perspective.
Notably, to address this issue, Chinese government has
continuously increased its investment in the construction
of first-class academic journals in recent years [31].
**An overview of research focuses and frontiers**
**_Reference analysis_**
Reference co-citation analysis is often applied to determine the research focuses on a given field. All the publications and their references data were used to create
homogeneous clusters, thus references that were connected tightly were divided into the same clusters, and
conversely in different clusters. Our findings demonstrated that there were 13 major clusters in the co-citation network map. The largest cluster was “risk factor”
(#0) [1, 2, 32]. Figure 6 shows the timeline view for the
major clusters, which could illustrate the temporal and
evolution characteristics of each cluster. The development of cluster 3, cluster 4, and cluster 6 occurred earliest, whereas cluster 0 (risk factor) and cluster 7 (stem
cell therapy) were the recent research topics in the field
of ONFH, which reflects the shift in research focus.
Apart from that, burst detection of reference was another
approach to track and capture the research hotspots. References with the strongest citation burst, indicating that
they have received special attention during a period, are
generally acknowledged as the research basics of frontiers in a certain field. As shown in Fig. 7, the strongest
burst starting from 2015 was from the paper published by
Mont MA and colleagues [30], followed by Moya-Angeler et al. [3] in 2015 and Mont MA et al. [29] in 2006. It
also can be observed that reference with citation bursts
first emerged in 2004, due to an article in 2002, and continued through 4 years. Of note, the burst of several references after 2015 is still ongoing, suggesting that these
topics have gained considerable attention in recent years
and deserve further attention for future periods of time.
It is worth noting that most of these references involved
stem cells therapy [33–35].
**_Keywords analysis_**
Generally, the author keywords of an article are usually
the most representative terms used to give a brief overview of research theme, and the co-occurrence analysis
of keywords is a common bibliometric method to present
the knowledge content and structure visually and also
uncover the evolution process and hot topics of a field
[14]. Based on the research categories of these keywords,
VOSviewer software was able to divide all keywords into
several major clusters with different colors. As shown in
the network visualization map of Fig. 8A, all the included
keywords were classified into four main research clusters including Cluster 1(etiology and risk factors study);
Cluster 2 (basic research and stem cell therapy); cluster 3 (hip-preserving study); Cluster 4 (hip replacement
-----
study). In addition to this, the VOSviewer software could
impart all the included keywords with different colors
based on their AAY. It can be seen that early research
prior to 2015, the ONFH research was mainly focused on
“hip replacement study” in cluster 4 and “hip-preserving
study” in cluster 3, whereas keywords belong to in cluster 2 (“basic research and stem cell therapy”) had the
relatively latest AAY than other clusters. Stem cells possess the ability to self-renew and differentiate into various cell types such as osteoblasts and endothelial cells to
promote angiogenesis as well as bone regeneration [33].
In the meantime, they could also secrete a broad range
of biological factors including multiple cytokines, growth
factors, and exosomes to promote new blood vessel formation and rebuild blood supply in the necrotic regions
[35]. As far as we know, multiple rigorous random controlled trials (RCTs) on the efficacy of stem cell therapy
for early-stage ONFH have been initiated or currently in
progress [23]. Yet at the same time, their clinical values
remain to be further elucidated. Thus, stem cell therapy
has become one of the most promising areas for ONFH.
Additionally, in cluster 2, these keywords with relatively
latest AAY, such as exosomes, autophagy, biomarkers,
osteogenic differentiation, microRNAs, steroid-induced
osteonecrosis, and mesenchymal stem cells, may have
great potential to be hot topics in the near future. For
example, miRNAs are small non-coding RNAs that
broadly regulate gene expression by specifically binding to complementary sequences in the 3’‐untranslated
regions of their target RNAs. In recent years, the field of
miRNAs has been emerged as a focus of ONFH research,
and has received extensive attention from researchers
in China and other countries [33]. Some scholars have
used the microarray method to compare miRNA expression in patients with ONFH and in patients with femoral
neck fracture. Of the 17 miRNAs identified with differential expression, 12 were up-regulated and 5 were downregulated, suggesting that aberrant miRNAs expression
might be involved in the pathogenesis of ONFH, and
thus become diagnostic markers for ONFH [36]. Additionally, accumulating evidence demonstrated that multiple miRNAs could act as novel therapeutic targets for
the prevention and treatment of ONFH by regulating
osteogenic and adipogenic differentiation in MSCs [33,
35]. In terms of steroid-induced osteonecrosis, it is worth
emphasizing that as the ongoing spread of coronavirus
disease 2019 (COVID-19) globally, despite great strides
in management, corticosteroids remain the mainstay for
the treatment of moderate to severe acute respiratory
syndrome (SARS), and with it arises challenges such as
steroid-induced ONFH, especially in patients with the
long-term or high doses use [37]. Some scholars have
noted the potential risk and called for judicious use of
corticosteroids in COVID-19 patients, particularly not
recommended for routine use [38]. Aside from cluster
2, several topics with relatively latest AAY in other clusters including double-blind [39], early-stage osteonecrosis [40, 41], and asymptomatic osteonecrosis [42] also
deserve further attention.
**Limitation**
Despite the rigorous bibliometric analysis of this study,
there were still several inevitable shortcomings. For
example, we only analyzed bibliometric data from
WOSCC database, which potentially missed several relevant publications reported in other databases, such as
Scopus and PubMed [43]. Moreover, in consideration of
only English publications in the study, it is unavoidable
that several important publications in non-English language were omitted. As for the keyword clustering analysis, it might not be appropriate to combine the keywords
with the same meaning into one node as different keywords might belong to clinical or basic research, respectively. And finally, the latest publications in 2021 were not
incorporated since they lack sufficient time to accumulate considerable citations, which might in part affect our
conclusions due to the rapid updating of research hotspots and frontiers.
**Conclusion**
Overall, the ascending trend in the annual number of
publications indicates that ONFH has attracted a great
deal of interest from researchers worldwide, especially
in the last 10 years. China, the USA, and Japan were the
major contributors. And part of the discrepancy in the
quantity of publications across different countries can be
explained by the factors of economy or population. The
most prolific institution was Shanghai Jiao Tong University. Professor Zhang CQ and Mont MA were the most
influential authors with the highest number of publications and citations, respectively. According to keywords
analysis, all the selected keywords could be categorized into four major clusters. Stem cell therapy-related
research has been recognized as an important research
hotspot in this field. It is recommended to pay more
attention to these topics including exosomes, autophagy,
biomarkers, osteogenic differentiation, microRNAs,
steroid-induced osteonecrosis, mesenchymal stem cells,
double-blind, early-stage osteonecrosis, and asymptomatic osteonecrosis, which have great potential to continue to be the research focuses in the near future.
**Abbreviations**
ONFH: Osteonecrosis of the femoral head; THA: Total hip arthroplasty; WOSCC:
Web of Science Core Collection; GDP: Gross Domestic Product; JIF: Journal
-----
impact factors; TLS: Total link strength; BC: Betweenness centrality; AAY: Average appearing year; MSCs: Mesenchymal stem cells.
**Acknowledgements**
The authors thank Dr. Zhou Yan of Tianjin Medical University and “home-for[researchers” Company (https://www.home-for-researchers.com/) for their help](https://www.home-for-researchers.com/)
in in polishing our English writing.
**Author’s contributions**
All authors contributed to the study conception and design. LT, YW, and WY
collected the data and material preparation. Data collection and analysis
were performed by HW and KC. The first draft of the manuscript was written
by HW and KC. ZS revised the work. All authors read and approved the final
manuscript.
**Funding**
This work was supported by the Tianjin Municipal Health Bureau (Grant Number 14KG115) and Key Program of the Natural Science Foundation of Tianjin
(Grant Number 20JCZDJC00730).
**Availability of data and materials**
All the data can be downloaded from Web of Science Core Collection.
**Declarations**
**Ethics approval and consent to participate**
Ethical approval was not required for this study, as all data were downloaded
from public databases and did not involve any human or animal participants.
**Consent for publication**
All authors agreed with the content and gave explicit consent to submit
and they obtained consent from the responsible authorities at the institute/
organization where the work has been carried out.
**Competing interests**
Haiyang Wu, Kunming Cheng, Linjian Tong, Yulin Wang, Weiguang Yang, Zhiming Sun declare that they have no competing interests.
**Author details**
1 Graduate School of Tianjin Medical University, No. 22 Qixiangtai Road, Tianjin 300070, China. [2] Department of Intensive Care Unit, The Second Affiliated
Hospital of Zhengzhou University, Zhengzhou 450014, Henan, China. [3] Department of Orthopaedic Surgery, Tianjin Huanhu Hospital, No 6, Jizhao Road,
Jinnan District, Tianjin 300350, China.
Received: 19 August 2021 Accepted: 15 March 2022
**References**
1. Tan B, Li W, Zeng P, et al. Epidemiological study based on China osteonecrosis of the femoral head database. Orthop Surg. 2021;13:153–60.
2. Zhao D, Zhang F, Wang B, et al. Guidelines for clinical diagnosis and
treatment of osteonecrosis of the femoral head in adults (2019 version). J
Orthop Transl. 2020;21:100–10.
3. Moya-Angeler J, Gianakos AL, Villa JC, et al. Current concepts on osteonecrosis of the femoral head. World J Orthop. 2015;6:590–601.
4. Zhao DW, Yu M, Hu K, et al. Prevalence of nontraumatic osteonecrosis of
the femoral head and its associated risk factors in the Chinese population: results from a nationally representative survey. Chin Med J (Engl).
2015;128:2843–50.
5. Wu CT, Yen SH, Lin PC, et al. Long-term outcomes of Phemister bone
grafting for patients with non-traumatic osteonecrosis of the femoral
head. Int Orthop. 2019;43:579–87.
6. Johnson AJ, Mont MA, Tsao AK, et al. Treatment of femoral head osteonecrosis in the United States: 16-year analysis of the Nationwide Inpatient
Sample. Clin Orthop Relat Res. 2014;472:617–23.
7. Liu L, Gao F, Sun W, et al. Investigating clinical failure of core decompression with autologous bone marrow mononuclear cells grafting for
the treatment of non-traumatic osteonecrosis of the femoral head. Int
Orthop. 2018;42:1575–83.
8. Wu H, Tong L, Wang Y, et al. Bibliometric analysis of global research trends
on ultrasound microbubble: a quickly developing field. Front Pharmacol.
2021;12:646626.
9. Li C, Ojeda-Thies C, Xu C, et al. Meta-analysis in periprosthetic joint infection: a global bibliometric analysis. J Orthop Surg Res. 2020;15:251.
10. Li C, Wang L, Perka C, et al. Clinical application of robotic orthopedic
surgery: a bibliometric study. BMC Musculoskelet Disord. 2021;22:968.
11. van Eck NJ, Waltman L. Software survey: VOSviewer, a computer program
for bibliometric mapping. Scientometrics. 2010;84:523–38.
12. Synnestvedt MB, Chen C, Holmes JH. CiteSpace II: visualization and
knowledge discovery in bibliographic databases. AMIA Annu Symp Proc.
2005;2005:724–8.
13. Tang F, Dai WB, Li XL, et al. Publication trends and hot spots in femoroacetabular impingement research: a 20-year bibliometric analysis. J
Arthroplasty. 2021;36:2698–707.
14. Wu H, Li Y, Tong L, et al. Worldwide research tendency and hotspots on
hip fracture: a 20-year bibliometric analysis. Arch Osteoporos. 2021;16:73.
15. Wu H, Wang Y, Tong L, et al. The global research trends and hotspots on
developmental dysplasia of the hip: a bibliometric and visualized study.
Front Surg. 2021;8:671403.
16. Zhang W, Tang N, Li X, et al. The top 100 most cited articles on total hip
arthroplasty: a bibliometric analysis. J Orthop Surg Res. 2019;14:412.
17. Gao J, Xing D, Dong S, et al. The primary total knee arthroplasty: a global
analysis. J Orthop Surg Res. 2020;15:190.
18. Giannoudis PV, Chloros GD, Ho YS. A historical review and bibliometric
analysis of research on fracture nonunion in the last three decades. Int
Orthop. 2021;45:1663–76.
19. Wu H, Zhou Y, Xu L, et al. Mapping knowledge structure and research
frontiers of ultrasound-induced blood-brain barrier opening: a scientometric study. Front Neurosci. 2021;15:706105.
20. Guo Y, Hao Z, Zhao S, et al. Artificial intelligence in health care: bibliometric analysis. J Med Internet Res. 2020;22:e18228.
21. Sodhi N, Acuna A, Etcheson J, et al. Management of osteonecrosis of the
femoral head. Bone Joint J. 2020;102:122–8.
22. Cooper C, Steinbuch M, Stevenson R, et al. The epidemiology of osteonecrosis: findings from the GPRD and THIN databases in the UK. Osteoporos
Int. 2010;21:569–77.
23. Mont MA, Zywiel MG, Marker DR, et al. The natural history of untreated
asymptomatic osteonecrosis of the femoral head: a systematic literature
review. J Bone Joint Surg Am. 2010;92:2165–70.
24. Mao L, Jiang P, Lei X, et al. Efficacy and safety of stem cell therapy for the
early-stage osteonecrosis of femoral head: a systematic review and metaanalysis of randomized controlled trials. Stem Cell Res Ther. 2020;11:445.
25. Huang C, Wen Z, Niu J, Lin S, Wang W. Steroid-induced osteonecrosis of
the femoral head: novel insight into the roles of bone endothelial cells in
pathogenesis and treatment. Front Cell Dev Biol. 2021;9:777697.
26. Hirsch JE. An index to quantify an individual’s scientific research output.
Proc Natl Acad Sci U S A. 2005;102:16569–72.
27. Zhang CQ, Gao YS, Zhu ZH, et al. Why we choose free vascularized
fibular grafting for osteonecrosis of the femoral head? Microsurgery.
2011;31:417–8.
28. Liu X, Li Q, Niu X, et al. Exosomes secreted from human-induced pluripotent stem cell-derived mesenchymal stem cells prevent osteonecrosis of the femoral head by promoting angiogenesis. Int J Biol Sci.
2017;13:232–44.
29. Mont MA, Jones LC, Hungerford DS. Nontraumatic osteonecrosis of the
femoral head: ten years later. J Bone Joint Surg Am. 2006;88:1117–32.
30. Mont MA, Cherian JJ, Sierra RJ, et al. Nontraumatic osteonecrosis of the
femoral head: where do we stand today? A ten-year update. J Bone Joint
Surg Am. 2015;97:1604–27.
31. Zu GA. WJG sets an example of internationalization for other Chinese
academic journals. World J Gastroenterol. 2010;16:2707–9.
32. Wu H, Shang R, Cai X, et al. Single ilioinguinal approach to treat complex
acetabular fractures with quadrilateral plate involvement: outcomes
using a novel dynamic anterior plate-screw system. Orthop Surg.
2020;12:488–97.
33. Hao C, Yang S, Xu W, et al. MiR-708 promotes steroid-induced osteonecrosis of femoral head, suppresses osteogenic differentiation by targeting
SMAD3. Sci Rep. 2016;6:22599.
-----
34. Houdek MT, Wyles CC, Packard BD, et al. Decreased osteogenic activity
of mesenchymal stem cells in patients with corticosteroid-induced
osteonecrosis of the femoral head. J Arthroplasty. 2016;31:893–8.
35. Hernigou P, Flouzat-Lachaniette CH, Delambre J, et al. Osteonecrosis
repair with bone marrow cell therapies: state of the clinical art. Bone.
2015;70:102–9.
36. Yuan HF, Von Roemeling C, Gao HD, et al. Analysis of altered microRNA
expression profile in the reparative interface of the femoral head with
osteonecrosis. Exp Mol Pathol. 2015;98:158–63.
37. Gálvez-Romero JL, Palmeros-Rojas O, Real-Ramírez FA, et al. Cyclosporine
A plus low-dose steroid treatment in COVID-19 improves clinical outcomes in patients with moderate to severe disease: a pilot study. J Intern
Med. 2021;289:906–20.
38. Zhang S, Wang C, Shi L, et al. Beware of steroid-induced avascular necrosis of the femoral head in the treatment of COVID-19-experience and
lessons from the SARS epidemic. Drug Des Dev Ther. 2021;15:983–95.
39. Hauzeur JP, De Maertelaer V, Baudoux E, et al. Inefficacy of autologous
bone marrow concentrate in stage three osteonecrosis: a randomized
controlled double-blind trial. Int Orthop. 2018;42:1429–35.
40. Yue J, Gao H, Guo X, et al. Fibula allograft propping as an effective treatment for early-stage osteonecrosis of the femoral head: a systematic
review. J Orthop Surg Res. 2020;15:206.
41. Hua KC, Yang XG, Feng JT, et al. The efficacy and safety of core decompression for the treatment of femoral head necrosis: a systematic review
and meta-analysis. J Orthop Surg Res. 2019;14:306.
42. Wei QS, Hong GJ, Yuan YJ, et al. Huo Xue Tong Luo capsule, a vasoactive
herbal formula prevents progression of asymptomatic osteonecrosis of
femoral head: a prospective study. J Orthop Transl. 2018;18:65–73.
43. Wu H, Sun Z, Tong L, et al. Bibliometric analysis of global research trends
on male osteoporosis: a neglected field deserves more attention. Arch
Osteoporos. 2021;16:154.
**Publisher’s Note**
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://pmc.ncbi.nlm.nih.gov/articles/PMC8960091, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "GOLD",
"url": "https://josr-online.biomedcentral.com/track/pdf/10.1186/s13018-022-03068-7"
}
| 2,022
|
[
"JournalArticle"
] | true
| 2022-03-28T00:00:00
|
[
{
"paperId": "505776b938f60eecd327a3d8e3563c5bfaed0073",
"title": "Steroid-Induced Osteonecrosis of the Femoral Head: Novel Insight Into the Roles of Bone Endothelial Cells in Pathogenesis and Treatment"
},
{
"paperId": "d845899e33ccd1a56c53a130f08a7bd29fa6246c",
"title": "The Global Research Trends and Hotspots on Developmental Dysplasia of the Hip: A Bibliometric and Visualized Study"
},
{
"paperId": "66f503781cd203bd9e7f1dc744a8acd955f0535e",
"title": "Bibliometric analysis of global research trends on male osteoporosis: a neglected field deserves more attention"
},
{
"paperId": "4e71222a7c12c7df5310e538e054b7e7dc657c89",
"title": "Mapping Knowledge Structure and Research Frontiers of Ultrasound-Induced Blood-Brain Barrier Opening: A Scientometric Study"
},
{
"paperId": "7b25720059f573000f59eecf78b5809c418d19a6",
"title": "Clinical application of robotic orthopedic surgery: a bibliometric study"
},
{
"paperId": "5f45eee62a0778f34dbbaa10fbf42ba403c32b8d",
"title": "Bibliometric Analysis of Global Research Trends on Ultrasound Microbubble: A Quickly Developing Field"
},
{
"paperId": "fc24e5ff6525ca1d8cdf7ea6ee47712ef7993174",
"title": "A historical review and bibliometric analysis of research on fracture nonunion in the last three decades"
},
{
"paperId": "71920364958c3d9a8b32c3084c2e94e1fa733f7b",
"title": "Worldwide research tendency and hotspots on hip fracture: a 20-year bibliometric analysis"
},
{
"paperId": "3a064dcea528939e0ca4055afefb8e106ec9d2b1",
"title": "Publication Trends and Hot Spots in Femoroacetabular Impingement Research: A 20-Year Bibliometric Analysis."
},
{
"paperId": "fb514be4d63ae41fcb05e2cdbcf222a03f5f1f00",
"title": "Beware of Steroid-Induced Avascular Necrosis of the Femoral Head in the Treatment of COVID-19—Experience and Lessons from the SARS Epidemic"
},
{
"paperId": "31cb39d93748309295b0b4db9f0a2147bf2b5464",
"title": "Epidemiological Study Based on China Osteonecrosis of the Femoral Head Database"
},
{
"paperId": "0c866c3f7bfcef1839bf41dd4cdde80f99608fce",
"title": "Cyclosporine A plus low‐dose steroid treatment in COVID‐19 improves clinical outcomes in patients with moderate to severe disease: A pilot study"
},
{
"paperId": "eed028c1f96ff0d45cb3524b8dbd391e501404dc",
"title": "Efficacy and safety of stem cell therapy for the early-stage osteonecrosis of femoral head: a systematic review and meta-analysis of randomized controlled trials"
},
{
"paperId": "f17aff43609f02c542abbe8989e02881e0b5ed75",
"title": "Fibula allograft propping as an effective treatment for early-stage osteonecrosis of the femoral head: a systematic review"
},
{
"paperId": "17d180dae73468d8c6ca37176f6d965ad24ff0d2",
"title": "The primary total knee arthroplasty: a global analysis"
},
{
"paperId": "08c1943d4d11486157432b0b2cd97c8e538883b8",
"title": "Meta-analysis in periprosthetic joint infection: a global bibliometric analysis"
},
{
"paperId": "9f031630e63769594b8e5e35a9476a9614c23142",
"title": "Single Ilioinguinal Approach to Treat Complex Acetabular Fractures with Quadrilateral Plate Involvement: Outcomes Using a Novel Dynamic Anterior Plate–Screw System"
},
{
"paperId": "2bbc7e46425d8dabb2c2ebbf28dbbb0462d3b5e3",
"title": "Artificial Intelligence in Health Care: Bibliometric Analysis"
},
{
"paperId": "d07a582c5207c70972627aa5e1ffb94d26d18717",
"title": "Guidelines for clinical diagnosis and treatment of osteonecrosis of the femoral head in adults (2019 version)"
},
{
"paperId": "083f4fbbfe2ada334e6aae348594645394d5ef74",
"title": "The top 100 most cited articles on total hip arthroplasty: a bibliometric analysis"
},
{
"paperId": "1b25cff2ff1be73ead0d4527fd9dd5005788d049",
"title": "The efficacy and safety of core decompression for the treatment of femoral head necrosis: a systematic review and meta-analysis"
},
{
"paperId": "30461c5ec1fcde2a29d59517dbfa4cc5ed8946dd",
"title": "Huo Xue Tong Luo capsule, a vasoactive herbal formula prevents progression of asymptomatic osteonecrosis of femoral head: A prospective study"
},
{
"paperId": "f86d81143f243a1fe99afe85f05d835efb232387",
"title": "Inefficacy of autologous bone marrow concentrate in stage three osteonecrosis: a randomized controlled double-blind trial"
},
{
"paperId": "27afdb535a2e7516adeaf0b8950c91b89c9b17bd",
"title": "Long-term outcomes of Phemister bone grafting for patients with non-traumatic osteonecrosis of the femoral head"
},
{
"paperId": "6cfc74f4cf3eb5daa644ef277bd8524d79731285",
"title": "Investigating clinical failure of core decompression with autologous bone marrow mononuclear cells grafting for the treatment of non-traumatic osteonecrosis of the femoral head"
},
{
"paperId": "bb6189f8113a2eea9c6b2bc30b5aacf4b8328a69",
"title": "Publisher's Note"
},
{
"paperId": "99d5833dc2b1093fbb9987d521ada2930e749e02",
"title": "Exosomes Secreted from Human-Induced Pluripotent Stem Cell-Derived Mesenchymal Stem Cells Prevent Osteonecrosis of the Femoral Head by Promoting Angiogenesis"
},
{
"paperId": "a1d466daa94aa17b4a674c86622e643605e16dc1",
"title": "Decreased Osteogenic Activity of Mesenchymal Stem Cells in Patients With Corticosteroid-Induced Osteonecrosis of the Femoral Head."
},
{
"paperId": "121016d380f14ce696d909b4748c73356f3d5827",
"title": "MiR-708 promotes steroid-induced osteonecrosis of femoral head, suppresses osteogenic differentiation by targeting SMAD3"
},
{
"paperId": "489d706a40201afb180058e5d81b731f3687aacb",
"title": "Prevalence of Nontraumatic Osteonecrosis of the Femoral Head and its Associated Risk Factors in the Chinese Population: Results from a Nationally Representative Survey"
},
{
"paperId": "420bba2ca599c8d0a2d591f1cb825750162a4122",
"title": "Nontraumatic Osteonecrosis of the Femoral Head: Where Do We Stand Today? A Ten-Year Update."
},
{
"paperId": "5cfaf64b0931b978a4a1738d871b148b8b0e0ccd",
"title": "Current concepts on osteonecrosis of the femoral head."
},
{
"paperId": "b23ad5150713b61190be59462a855ceff0b66552",
"title": "Analysis of altered microRNA expression profile in the reparative interface of the femoral head with osteonecrosis."
},
{
"paperId": "db49fefe6389e14e6773db37c7f7ed407883cf92",
"title": "Treatment of Femoral Head Osteonecrosis in the United States: 16-year Analysis of the Nationwide Inpatient Sample"
},
{
"paperId": "a60da185617baf7adbb699b0fbaeba31dcba3b0e",
"title": "Why we choose free vascularized fibular grafting for osteonecrosis of the femoral head?"
},
{
"paperId": "c5b7af49b7a40d248449eb0acf743ab72547b83d",
"title": "The natural history of untreated asymptomatic osteonecrosis of the femoral head: a systematic literature review."
},
{
"paperId": "4fe772fadd6f63df784541947d5fd2cc9a47859c",
"title": "WJG sets an example of internationalization for other Chinese academic journals."
},
{
"paperId": "edea882e8294537c76e16339ce32c2dd96753025",
"title": "Software survey: VOSviewer, a computer program for bibliometric mapping"
},
{
"paperId": "a0f16815404e86b2b4a732e5a090fca7be64c18a",
"title": "The epidemiology of osteonecrosis: findings from the GPRD and THIN databases in the UK"
},
{
"paperId": "a7f5aff22f72fc30c4f3c771c3ec92b281dfd5bd",
"title": "Nontraumatic osteonecrosis of the femoral head: ten years later."
},
{
"paperId": "c177ba654583987567377ed9d17a58f26b3580a3",
"title": "F1000Prime recommendation of An index to quantify an individual's scientific research output."
},
{
"paperId": "f5153b4371f362c6fef3aa6395db5fc0ee4b96da",
"title": "Management of osteonecrosis of the femoral head."
},
{
"paperId": "e455ec7012de2d49be108a52b7152caa02189e1a",
"title": "Osteonecrosis repair with bone marrow cell therapies: state of the clinical art."
},
{
"paperId": "1cdb5cc417c041acb85dd5d1b8f8b59fee38fa71",
"title": "CiteSpace II: Visualization and Knowledge Discovery in Bibliographic Databases"
},
{
"paperId": null,
"title": "Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations"
}
] | 11,840
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Engineering",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/0260f2fe0b329bbacd07ecaf3731bf48e7c8ebc3
|
[
"Computer Science"
] | 0.895421
|
Secured Decentralized Confidential Data Distributed in the Disruption-Tolerant Military Network
|
0260f2fe0b329bbacd07ecaf3731bf48e7c8ebc3
|
[
{
"authorId": "2056988783",
"name": "Aniruddha Singh Chauhan"
},
{
"authorId": "66266387",
"name": "Nikita Umare"
}
] |
{
"alternate_issns": null,
"alternate_names": null,
"alternate_urls": null,
"id": null,
"issn": null,
"name": null,
"type": null,
"url": null
}
| null |
**_https://dx.doi.org/10.22161/ijaers/3.10.11 ISSN: 2349-6495(P) | 2456-1908(O)_**
# Secured Decentralized Confidential Data Distributed in the Disruption-Tolerant Military Network
## Aniruddha Singh Chauhan[1], Prof. Nikita Umare[2]
1ME 3rd Sem. WCC Student, Abha Gaikwad-Patil College of Engineering, Nagpur, India
2Department of CSE/WCC, Abha Gaikwad-Patil College of Engineering, Nagpur, India
**_Abstract— Disruption tolerant network technologies are_** authorities and access the confidential information or
_becoming successful solutions that allow wireless devices_ command reliably by exploiting external storage nodes.
_carried by soldiers to communicate with each other and_ Some of the most challenging issues in this scenario are the
_access the confidential information or command reliably by_ enforcement of authorization policies and the policies
_exploiting external storage nodes. Some of the most_ update for secure data retrieval. Cipher text-policy attribute_challenging issues in this scenario are the enforcement of_ based encryption (CP-ABE) is a promising cryptographic
_authorization policies and the policies update for secure_ solution to the access control issues.
_data retrieval. Cipher text-policy attribute-based encryption_ However, the problem of applying CP ABE in decentralized
_is a promising cryptographic solution to the access control_ DTNs introduces several security and privacy challenges
_issues. However, the problem of applying CP-ABE in_ with regard to the attribute revocation, key escrow, and
_decentralized DTNs introduces several security and privacy_ coordination of attributes issued from different authorities.
_challenges with regard to the attribute revocation, key_ We propose a secure data retrieval scheme using CPABE
_escrow, and coordination of attributes issued from different_ for decentralized DTNs where multiple key authorities
_authorities. We propose a secure data retrieval scheme_ manage their attributes independently. We demonstrate how
_using idea for decentralized DTNs where multiple key_ to apply the proposed mechanism to securely and efficiently
_authorities manage their attributes independently. We_ manage the confidential data distributed in the disruption_demonstrate how to apply the proposed mechanism to_ tolerant military network.
_securely and efficiently manage the confidential data_
_distributed in the disruption-tolerant military network._ **II.** **PROBLEM** **DEFINITION**
**_Keywords—Access control, attribute-based encryption_** Military applications require increased protection of
**_(ABE), disruption-tolerant network (DTN)._** confidential data including access control method In many
cases, it is desirable to provide differentiated access services
**I.** **INTRODUCTION** such that Data access policies are defined over user attributes
Mobile nodes in military environments such as a battlefield or roles, which are managed by the key authorities.
or a hostile region are likely to suffer from intermittent
network connectivity and frequent partitions.Disruptiontolerant network (DTN) technologies are becoming
successful solutions that allow wireless devices carried by
soldiers to communicate with each other in this extreme
networking environment typically when there is no end to
end connection between source and destination pairs, the
messages from the source node may need to wait in the
intermediate nodes for a substantial amount of time until the
connection would be eventually established. Storage nodes
in DTNs where data is stored or replicated in a way such
that only authorized mobile nodes can access the necessary
information quickly and efficiently. Many military
applications require increased protection of confidential
data including access control methods that are
cryptographically enforced and provide differentiated access
service such that data access policies, which are defined as
_Fig.1: System Flow_
per user attributes or roles, which are managed by the key
-----
**_https://dx.doi.org/10.22161/ijaers/3.10.11 ISSN: 2349-6495(P) | 2456-1908(O)_**
**III.** **PROPOSED METHOD**
1) Key Authorities: They are key generation centres that
generate public/secret parameters for CP-ABE. The key
authorities consist of a central authority and multiple local
authorities. We assume that there are secure and reliable
communication channels between a central authority and
each local authority during the initial key setup and
generation phase.
Each local authority manages different attributes and issues
corresponding attribute keys to users.
They grant differential access rights to individual users
based on the users’ attributes. The key authorities are
assumed honest-but-curious. That is, they will honestly
execute the assigned tasks in the system; however they
would like to learn information of encrypted contents as
much as possible.
2) Storage node: This entity stores data from senders and
provide corresponding access to users. It may be mobile or
static, we also assume the storage node to be semi trusted,
that is honest-but-curious.
3) Sender: This entity owns confidential messages or data
(e.g., a commander) and wishes to store them into the
external data storage node for ease of sharing or for reliable
delivery to users in the extreme networking environments.
A sender is responsible for defining (attribute based) access
policy and enforcing it on its own data by encrypting the
data under the policy before storing it to the storage node.
4) User: This mobile node wants to access the data stored at
the storage node (e.g., a soldier). If a user possesses a set of
attributes satisfying the access policy of the encrypted data
defined by the sender, and is not revoked in any of the
attributes, then he will be able to decrypt and obtain the
data. Since the key authorities are semi-trusted, they should
be deterred from accessing plaintext of the data in the
storage node; meanwhile, they should be still able to issue
secret keys to users.
In order to realize this contradictory requirement, the central
authority and the local authorities engage in the arithmetic
2PC protocol with master secret keys of their own and issue
independent key components to users during the key issuing
phase.
The 2PC protocol prevents them from knowing each other’s
master secrets so that none of them can generate the whole
set of secret keys of users individually. Thus, we take an
assumption that the central authority does not collude with
the local authorities (otherwise, they can guess the secret
keys of every user by sharing their master secrets).
**Problems in Proposed Method:**
1) Collusion-resistance: If multiple users collude, they may
be able to decrypt a cipher text by combining their attributes
even if each of the users cannot decrypt the cipher text alone
[11]–[13]. For example, suppose there exist a user with
attributes {”Battalion 1”, “Region 1”} and another user with
attributes {”Battalion 2”, “Region 2”}. They may succeed in
decrypting a cipher text encrypted under the access policy
of (“Battalion 1” AND “Region 2”), even if each of them
cannot decrypt it individually. We do not want these
colluders to be able to decrypt the secret information by
combining their attributes. We also consider collusion
attack among curious local authorities to derive users’ keys
2)Backward and forward Secrecy: In the context of ABE,
backward secrecy means that any user who comes to hold
an attribute (that satisfies the access policy) should be
prevented from accessing the plaintext of the previous data
exchanged before he holds the attribute. On the other hand,
forward secrecy means that any user who drops an attribute
should be prevented from accessing the plaintext of the
subsequent data exchanged after he drops the attribute,
unless the other valid attributes that he is holding satisfy the
access policy.
3) Key Escrow: In CP-ABE, the key authority generates
private keys of users by applying the authority’s master
secret keys to users’ associated set of attributes. Thus, the
key authority can decrypt every cipher text addressed to
specific users by generating their attribute keys. If
adversaries when deployed in the hostile environments
compromise the key authority, this could be a potential
threat to the data confidentiality or privacy especially when
the data is highly sensitive. The key escrow is an inherent
problem even in the multiple-authority systems as long as
each key authority has the whole privilege to generate their
own attribute keys with their own.
**IV.** **REASERCH CONTRIBUTION**
Technologies are becoming successful solutions that allow
wireless devices carried by soldiers to communicate with
each other and access the confidential information or
command by exploiting external storage nodes. Some of the
most challenging issues in this scenario are the enforcement
of authorization policies and the policies update for secure
data retrieval. Cipher text-policy attribute-based encryption
is a promising cryptographic solution to the access control
issues. However, the problem of applying CP-ABE in
decentralized DTNs introduces several security and privacy
challenges with regard to the attribute revocation, key
escrow, and coordination of attributes issued from different
authorities.
Hence, we propose a secure data retrieval scheme using
IDEA Algorithm as 3DES with MD5 Algorithm Known as
Crypto Hybrid Algorithm for decentralized DTNs where
multiple key authorities manage their attributes
independently. We demonstrate how to apply the proposed
mechanism to securely and efficiently manage the
confidential data distributed in the disruption-tolerant
military network.
-----
**_https://dx.doi.org/10.22161/ijaers/3.10.11 ISSN: 2349-6495(P) | 2456-1908(O)_**
**3DES with MD5 ALGORTHIM**
Use of multiple length keys leads us to the Triple-DES
algorithm, in which DES is applied three times. Triple DES
is simply another mode of DES operation. It takes three 64bit keys, for an overall key length of 192 bits. In Private
Encryption, you simply type in the entire 192-bit key rather
than entering each of the three keys individually. The Triple
DES then breaks the user provided key into three sub keys,
padding the keys if necessary so they are each 64 bits long.
The procedure for encryption is the same as regular DES,
but it is repeated three times. Hence, the name Triple DES,
The data is encrypted with the first key, decrypted with the
second key, and finally encrypted again with the third key. In particular, the following requirements must be supported
Triple DES, also known as 3DES. by the key management scheme, in order to facilitate data
Consequently, Triple DES runs three times slower than aggregation and dissemination process:
standard DES, but is much more secure if used properly. 1. Data aggregation is possible only if intermediate nodes
The procedure for decrypting something is the same as the have access to encrypted data so that they can extract
procedure for encryption, except it is executed in reverse. measurement values and apply to them aggregation
Like DES, data is encrypted and decrypted in 64-bit chunks. functions. Therefore, nodes that send data packets toward
Unfortunately, there are some weak keys that one should be the base station must encrypt them with keys available to
aware of: if all three keys, the first and second keys, or the the aggregator nodes.
second and third keys are the same, then the encryption 2. Data dissemination implies broadcasting of a message
procedure is essentially the same as standard DES. This from the aggregator to its group members. If an aggregator
situation is to be avoided because it is the same as using a shares a different key (set of keys) with each of the sensor
slow version of regular DES. within its group, then it will have to make multiple
Note that although the input key for DES is 64 bits long, the transmissions, encrypted each time with a different key, in
actual key used by DES is only 56 bits in length. The least order
significant (right most) bit in each byte is a parity bit, and to broadcast a message to all of the nodes .But transmissions
should be set so that there are always an odd number of 1s must be kept as low as possible because of their high-energy
in every byte. These parity bits are ignored, so only the consumption rate.
seven most significant bits of each byte are used, resulting 3. Confidentiality: In order to protect sensed data and
in a key length of 56 bits. This means that the effective key communication-changes between sensor nodes it is
strength for Triple DES is actually 168 bits because each of important to guarantee the secrecy of messages. In the
the three keys contains 8 parity bits that are not used during sensor network, case this is usually achieved by the use of
the encryption process. symmetric cryptography as asymmetric or public key
A commonly used technique in the Internet is to provide a cryptography in general is considered too expensive.
MD5 -Hash String so the receiver can compare if the file However, while encryption protects against outside attacks,
has been transmitted without any modifications. it does not protect against inside attacks/node compromises,
3DES encrypts a 64-bit block of plaintext to 64-bit block of as an attacker can use recovered cryptographic key material
ciphertext. It uses a 128-bit key. The algorithm consists of to successfully eavesdrop, impersonate or participate in the
eight identical rounds and a “half” roundfinal secret communications of the network Furthermore, while
Transformation. There are 216 possible 16-bitblocks confidentiality guarantees the security of communications
0000000000000000, 1111111111111111. Each operation inside the network it does not prevent the misuse of
with the set of possible 16-bit blocks is an algebraic group. information reaching the base station Hence,confidentiality
Bitwise XOR is bitwise addition modulo 2, and addition must also be coupled with the right control policies so that
modulo 216 is the usual group operation. Some spin must be only authorized users can have access to confidential
put on the elements – the 16-bit blocks – to make sense of information
multiplication modulo 216 + 1, however. 0 (i.e., 4. Integrity and Authentication: Integrity and authentication
0000000000000000) is not an element of the multiplicative is necessary to enable sensor nodes to detect modified,
group. injected, or replayed packets. While it is clear that safety
critical applications require authentication, it is still wise to
use it even for the rest of applications since otherwise the
owner of the sensor network may get the wrong picture of
the sensed world thus making inappropriate decisions.
-----
**_https://dx.doi.org/10.22161/ijaers/3.10.11 ISSN: 2349-6495(P) | 2456-1908(O)_**
However, authentication alone does not solve the problem
of node takeovers as compromised nodes can still authenticate themselves to the network. Hence, authentication
mechanisms should be “collective” and aim at securing the
entire network.
First, we focused on the establishment of trust relationship
among wireless sensor nodes, and presented a key
management protocol for sensor networks. The protocol
includes support for establishing four types of keys per
Fig 7.2 Packets Vs Time Graph
sensor node:
individual keys shared with the base station, pairwise keys
shared with individual neighboring nodes, cluster keys
**V.** **CONCULSION**
shared with a set of neighbors, and a group key shared with
The corresponding attribute group keys are updated and
all the nodes in the network. We showed how the keys
delivered to the valid attribute group members securely
could be distributed so that the protocol can support in
(including the user). In addition, all of the components
network processing and efficient dissemination, while
encrypted with a secret key in the cipher text are re
restricting the security impact of a node compromise to the
encrypted by the storage node with a random, and the cipher
immediate network neighborhood of the compromised node.
text components corresponding to the attributes are re
Applying the protocol makes it hard for an adversary to
encrypted with the updated attribute group keys. Even if the
disrupt the normal operation of the network.
user has stored the previous cipher text exchanged before he
In Hybrid Cryptosystem System, security is combination of
obtains the attribute keys and the holding attributes satisfy
more algorithm than base paper but still requires less time to
the access policy, he cannot decrypt the pervious cipher
Verify and process. While they are not present in the base
text.
paper. Hybrid Cryptosystem to enhance the security we use
combination of algos
**REFERENCES**
1) Idea algo.
[1] J. Burgess, B. Gallagher, D. Jensen, and B. N. Levine,
2) MD5
“Maxprop: Routing for vehicle-based disruption
3) ECB (ELECTRONIC CODE BOOK)
tolerant networks,” in _Proc. IEEE INFOCOM, 2006,_
4) Hashing code
_pp. 1–11._
[2] M. Chuah and P. Yang, “Node density-based adaptive
**Comparative Result analysis**
routing scheme for disruption tolerant networks,” in
In my Base Paper we have used CP-ABE systems i.e.
_Proc. IEEE MILCOM, 2006, pp.1–6._
Cipher text-policy attribute-based encryption which is a
[3] M. M. B. Tariq, M. Ammar, and E. Zequra, “Mesage
promising cryptographic solution to the access control
ferry route design for sparse ad hoc networks with
issues. While its communication Cost is higher than New
mobile nodes,” in Proc. ACM _MobiHoc, 2006, pp. 37–_
Hybrid Cryptography Technique. Comparative results can
_48._
see in Graph as:
[4] S. Roy andM. Chuah, “Secure data retrieval based on
ciphertext policy attribute-based encryption (CP-ABE)
system for the DTNs,” Lehigh CSE Tech. Rep., 2009.
[5] M. Chuah and P. Yang, “Performance evaluation of
content-based information retrieval schemes for
DTNs,” in Proc. IEEE MILCOM,2007, pp. 1–7.
[6] M. Kallahalla, E. Riedel, R. Swaminathan, Q. Wang,
and K. Fu,“Plutus: Scalable secure file sharing on
untrusted storage,” in _Proc.Conf. File Storage_
_Technol., 2003, pp. 29–42._
[7] [7] L. Ibraimi, M. Petkovic, S. Nikova, P. Hartel, and
_Fig.7.1: Communication cost in CP-ABE System_ W. Jonker, “Mediated ciphertext-policy attribute-based
encryption and its application,”in _Proc. WISA, 2009,_
Number of conversion and verification time is more in base LNCS 5932, pp. 309–323.
paper CP-ABE System then Hybrid Encryption by Using [8] [8] N. Chen, M. Gerla, D. Huang, and X. Hong,
Idea Algorithm and MD5. “Secure, selective group broadcast in vehicular
networks using dynamic attribute based encryption,”in
_Proc. Ad Hoc Netw. Workshop, 2010, pp. 1–8._
-----
**_https://dx.doi.org/10.22161/ijaers/3.10.11 ISSN: 2349-6495(P) | 2456-1908(O)_**
[9] D. Huang and M. Verma, “ASPE: Attribute-based
secure policy enforcement n vehicular ad hoc
networks,” Ad Hoc Netw.,vol. 7, no. 8,pp. 1526–1535,
2009.
[10] A. Lewko and B. Waters, “Decentralizing attributebased encryption,”Cryptology ePrint Archive:
Rep. 2010/351, 2010.
[11] A. Sahai and B. Waters, “Fuzzy identity-based
encryption,” in Proc.Eurocrypt, 2005, pp. 457–473.
[12] V. Goyal, O. Pandey, A. Sahai, and B. Waters,
“Attribute-based encryption for fine-grained access
control of encrypted data,” in _Proc. ACM Conf._
_Comput. Commun. Security, 2006, pp. 89–98._
[13] J. Bethencourt, A. Sahai, and B. Waters, “Ciphertextpolicy attributebased encryption,” in _Proc. IEEE_
_Symp. Security Privacy, 2007, pp. 321–334_
-----
|
{
"disclaimer": "Notice: The following paper fields have been elided by the publisher: {'abstract', 'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.22161/IJAERS/3.10.11?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.22161/IJAERS/3.10.11, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "BRONZE",
"url": "https://doi.org/10.22161/ijaers/3.10.11"
}
| 2,016
|
[] | true
| 2016-10-01T00:00:00
|
[] | 4,669
|
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Engineering",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/02620aa4121370fb8727553bb72e8bab1b95450f
|
[
"Computer Science"
] | 0.856484
|
Adaptive Restructuring of Merkle and Verkle Trees for Enhanced Blockchain Scalability
|
02620aa4121370fb8727553bb72e8bab1b95450f
|
Internet of Things
|
[
{
"authorId": "2267503406",
"name": "Oleksandr Kuznetsov"
},
{
"authorId": "2282961530",
"name": "Dzianis Kanonik"
},
{
"authorId": "2282962463",
"name": "Alex Rusnak"
},
{
"authorId": "2282967712",
"name": "Anton Yezhov"
},
{
"authorId": "2283137914",
"name": "Oleksandr Domin"
}
] |
{
"alternate_issns": [
"2199-1073"
],
"alternate_names": [
"Internet Thing"
],
"alternate_urls": [
"https://www.sciencedirect.com/journal/internet-of-things",
"http://www.springer.com/series/11636"
],
"id": "2989732e-2668-4b47-9c29-326646a60273",
"issn": "2542-6605",
"name": "Internet of Things",
"type": null,
"url": "https://www.journals.elsevier.com/internet-of-things"
}
| null |
# Adaptive Restructuring of Merkle and Verkle Trees
for Enhanced Blockchain Scalability
Oleksandr Kuznetsov[ 1,2*], Dzianis Kanonik [1], Alex Rusnak [1], Anton Yezhov[ 1], Oleksandr Domin[1 ]
1 Proxima Labs, 1501 Larkin Street, suite 300, San Francisco, USA
2 Department of Political Sciences, Communication and International Relations, University of
Macerata, Via Crescimbeni, 30/32, 62100 Macerata, Italy
[*Corresponding author. E-mail(s): kuznetsov@karazin.ua](mailto:kuznetsov@karazin.ua)
[Contributing authors: alex@proxima.one](mailto:alex@proxima.one)
**Abstract:** The scalability of blockchain technology remains a pivotal challenge, impeding its
widespread adoption across various sectors. This study introduces an innovative approach to
address this challenge by proposing the adaptive restructuring of Merkle and Verkle trees,
fundamental components of blockchain architecture responsible for ensuring data integrity and
facilitating efficient verification processes. Unlike traditional static tree structures, our adaptive
model dynamically adjusts the configuration of these trees based on usage patterns, significantly
reducing the average path length required for verification and, consequently, the computational
overhead associated with these processes. Through a comprehensive conceptual framework, we
delineate the methodology for adaptive restructuring, encompassing both binary and non-binary
tree configurations. This framework is validated through a series of detailed examples,
demonstrating the practical feasibility and the efficiency gains achievable with our approach.
Moreover, we present a comparative analysis with existing scalability solutions, highlighting the
unique advantages of adaptive restructuring in terms of simplicity, security, and efficiency
enhancement without introducing additional complexities or dependencies. This study's
implications extend beyond theoretical advancements, offering a scalable, secure, and efficient
method for blockchain data verification that could facilitate broader adoption of blockchain
technology in finance, supply chain management, and beyond. As the blockchain ecosystem
continues to evolve, the principles and methodologies outlined herein are poised to contribute
significantly to its growth and maturity.
**Keywords:** Blockchain Scalability, Merkle and Verkle Trees, Adaptive Restructuring, Data
Verification, Efficiency Optimization, Blockchain Architecture
**1. Introduction**
The advent of blockchain technology has heralded a new era in digital transactions, offering
unparalleled security, transparency, and decentralization [1]. At its core, blockchain leverages
cryptographic principles to create a distributed ledger system, where data integrity and transaction
veracity are maintained across a network of nodes without the need for a central authority [2]. This
innovative approach has found applications far beyond its initial cryptocurrency origins, extending
into finance, supply chain management, healthcare, and more [3].
However, as blockchain technology ventures into more complex and demanding applications, it
encounters a fundamental challenge that threatens its broader adoption: scalability [4]. The
-----
scalability issue primarily revolves around the capacity of a blockchain network to handle a large
volume of transactions quickly and efficiently. Current blockchain architectures, while robust and
secure, are hampered by their inherent design, which leads to bottlenecks in transaction processing
and data verification [5,6]. These limitations not only increase transaction costs but also extend
the time required to achieve consensus across the network, thereby reducing the system's overall
throughput.
**1.1. The Blockchain Paradigm and the Challenge of Scalability**
In the burgeoning landscape of blockchain technology, Ethereum stands out as a beacon of
innovation and application diversity. The network's capacity to support a wide array of
decentralized applications (dApps) and smart contracts has positioned it at the forefront of
blockchain development. This prominence is underscored by the data depicted in Figure 1, which
reveals a significant uptick in the total number of unique Ethereum addresses over the past year,
surging from 219.26 million to 254.66 million [7].
**Figure 1: Growth in Total Unique Ethereum Addresses**
However, the rapid expansion of the Ethereum network brings to light the pressing challenges
associated with managing an ever-growing blockchain ecosystem. The core issue revolves around
the efficient verification and management of data within the blockchain's infrastructure, a task that
becomes increasingly complex as the network scales. The juxtaposition of the network's growth
against the slight decrease in daily active Ethereum addresses, as shown in Figure 2, further
complicates this challenge [8]. Despite a year-over-year decrease of 1.70% in daily active
addresses, the fact remains that a mere 0.17% of all unique addresses engage with the network on
a daily basis. This discrepancy between the total number of addresses and the proportion of active
participants underscores a critical aspect of blockchain management: the network's activity is
highly concentrated within a relatively small segment of the overall user base.
This concentration of activity presents a unique set of challenges in optimizing the blockchain's
state tree. The state tree must be updated frequently to reflect the transactions and interactions
occurring within the network, a process that is predominantly influenced by the small fraction of
actively participating addresses. The need to efficiently manage and verify these updates, without
compromising the integrity or performance of the network, is paramount.
-----
**Figure 2: Daily Active Ethereum Addresses**
Given this backdrop, our research aims to address the pressing need for an optimized approach to
managing the blockchain's state tree, particularly within the context of Ethereum's rapidly
expanding network. The goal of our study is to explore innovative methods for restructuring
Merkle and Verkle Trees adaptively, thereby enhancing the efficiency of data verification
processes. By focusing on dynamic adjustments to tree configurations in response to usage
patterns, we seek to minimize verification path lengths and reduce the computational overhead
associated with maintaining data integrity. This research endeavor not only aims to bolster the
scalability of blockchain systems but also to contribute to the ongoing discourse on optimizing
blockchain infrastructure for the next generation of decentralized applications.
**1.2. State of the art**
In the realm of blockchain scalability, Kottursamy et al. (2023) [9] introduce a novel blockchain
architecture termed Mutable Block with Immutable Transaction (MBIT), aiming to enhance
scalability through a trapdoor cryptographic hash function for quantifying unspent coins. While
their approach significantly reduces verification and confirmation times, it primarily focuses on
transaction efficiency without addressing the broader scalability challenges related to blockchain's
data structure and state management.
Li et al. (2023) [10] propose PRI, a Payment Channel Hubs (PCH) solution enhancing privacy,
reusability, and interoperability for blockchain scalability. Despite its innovative approach to
solving the deposit lock-in problem and supporting multi-party participation, PRI's reliance on
trusted hardware and its limited scope in addressing the fundamental architectural scalability issues
of blockchain remain unaddressed.
Nasir et al. (2022) [11] provide a systematic review of scalable blockchains, identifying key
strategies for enhancing blockchain capabilities and analyzing scalability solutions. Their work
highlights the multifaceted nature of blockchain scalability but leaves a gap in practical
implementation strategies for optimizing blockchain data structures, particularly in the context of
state management and verification processes.
-----
Sanka and Cheung (2021) [12] offer a comprehensive review of blockchain scalability issues and
solutions, emphasizing the need for efficient consensus mechanisms and system throughput
improvements. While their analysis sheds light on the scalability challenges, the exploration of
adaptive data structures for optimizing blockchain's underlying architecture is not thoroughly
explored.
Sharma et al. (2023) [13] introduce BLAST-IoT, a blockchain-assisted scalable trust model for the
Internet of Things (IoT), focusing on secure dissemination and storage of trust information. Their
model addresses scalability in the context of IoT devices but does not extend to the broader
blockchain scalability challenges, particularly in relation to adaptive restructuring of blockchain
data structures.
Wang and Wu (2024) [14] present Lever-FS, a validation framework for intensive blockchain
validation, achieving scalability through optimistic execution and dispute resolution. While their
work advances the scalability of validation processes, it does not directly tackle the optimization
of blockchain's state tree structure for overall network efficiency.
Wang et al. (2023) [15] propose a scalable, efficient, and secured consensus mechanism for
Vehicle-to-Vehicle (V2V) energy trading, leveraging blockchain technology. Their consensus
mechanism addresses scalability in the specific context of V2V energy trading but does not address
the broader application of scalable data structures within the blockchain.
Xiao et al. (2024) [16] develop CE-PBFT, a high availability consensus algorithm for large-scale
consortium blockchain, focusing on improving system throughput and reducing latency. While
their algorithm enhances consensus efficiency, the exploration of adaptive and scalable blockchain
data structures remains an area for further research.
Yu et al. (2023) [17] introduce OverShard, a full sharding approach for scaling blockchain, which
significantly improves throughput and reduces confirmation latency. However, the application of
sharding to optimize blockchain's state tree and the exploration of adaptive restructuring
techniques are not fully addressed.
Zhen et al. (2024) [18] propose a dynamic state sharding blockchain architecture for scalable and
secure crowdsourcing systems, addressing the scalability and security of blockchain in
crowdsourcing applications. While their architecture offers improvements in throughput and
security, the potential for adaptive restructuring of blockchain data structures to further enhance
scalability is not explored.
Transitioning from the broader challenges of blockchain scalability, we delve into the specific
realm of tree-based data structures within blockchain technology. These structures, notably Merkle
and Verkle trees, are pivotal for ensuring data integrity, enhancing verification processes, and
optimizing storage within blockchain systems. The following literature review highlights
significant advancements and identifies gaps that our research aims to fill.
Ayyalasomayajula and Ramkumar (2023) [19] explore the optimization of Merkle Tree structures,
comparing linear and subtree implementations. Their findings favor the subtree method for its
efficiency in handling large-scale databases. However, their work primarily focuses on theoretical
advantages without addressing the practical challenges of integrating these optimized structures
into existing blockchain frameworks.
-----
Jeon et al. (2023) [20] introduce a hardware-accelerated approach for generating reusable Merkle
Trees for Bitcoin blockchain headers, significantly reducing execution time and power
consumption. While their solution enhances the efficiency of block candidate generation, it is
tailored to Bitcoin and does not explore the adaptability of their approach to other blockchain
architectures or the potential for further optimization in tree structure.
Jing, Zheng, and Chen (2021) [21] provide a comprehensive review of Merkle Tree's technical
principles and applications across various fields. Their work underscores the versatility and
potential of Merkle Trees but stops short of proposing innovative methods for dynamic
restructuring or optimization of these trees in response to the evolving needs of blockchain
systems.
Knollmann and Scheideler (2022) [22] present a self-stabilizing protocol for the Hashed Patricia
Trie, a distributed data structure enabling efficient prefix search. Their protocol addresses selfstabilization in distributed systems but does not explore the scalability implications of their data
structure within the broader context of blockchain technology.
Lin and Chen (2023) [23] propose a file verification scheme based on Verkle Trees, highlighting
the efficiency and security benefits over traditional Merkle Trees. While their work demonstrates
the potential of Verkle Trees in file verification, the exploration of these trees for broader
blockchain scalability and optimization remains limited.
Liu et al. (2021) [24] offer systematic insights into Merkle Trees, emphasizing their role in
blockchain data verification and retrieval. Their discussion on the advantages and applications of
Merkle Trees provides a solid foundation but lacks a detailed exploration of innovative approaches
to enhance these trees' efficiency and scalability in blockchain systems.
Mardiansyah, Muis, and Sari (2023) [25] introduce the Multi-State Merkle Patricia Trie (MSMPT)
for high-performance data structures in multi-query processing. Their work addresses performance
and efficiency in lightweight blockchain but does not delve into the scalability challenges of more
complex blockchain systems.
Mitra, Tauz, and Dolecek (2023) [26] propose the Graph Coded Merkle Tree to mitigate Data
Availability Attacks in blockchain systems. While their approach offers a novel solution to a
specific problem, the broader application of their design for general blockchain scalability and data
structure optimization is not addressed.
Zhao et al. (2024) [27] focus on minimizing block incentive volatility through Verkle tree-based
dynamic transaction storage. Their innovative approach addresses a crucial aspect of blockchain
economics but does not explore the structural optimization of Verkle Trees for enhanced scalability
and efficiency in blockchain systems.
Our research fills these gaps by proposing adaptive restructuring techniques for Merkle and Verkle
Trees, aiming to enhance blockchain scalability, optimize data verification and storage processes,
and provide a flexible framework adaptable to various blockchain architectures and applications.
**1.3. Our contribution**
This work introduces a novel approach to optimizing tree-based data structures within blockchain
technology, focusing on adaptive restructuring techniques for Merkle and Verkle Trees. Our
contributions are twofold: First, we propose a dynamic restructuring algorithm that enhances the
scalability and efficiency of blockchain systems by optimizing the verification and storage
processes. Second, we extend the applicability of these optimized tree structures beyond traditional
-----
blockchain applications, demonstrating their versatility in various blockchain architectures and
scenarios. Through rigorous analysis and experimentation, our research addresses the critical
scalability challenges faced by blockchain technology, offering a scalable, efficient, and adaptable
solution.
**1.4. Article structure**
The structure of this article is designed to provide a comprehensive overview of our research and
findings. Section 2 conceptualizes the problem of blockchain scalability and the role of tree-based
data structures in addressing this challenge. Section 3 introduces our idea for optimizing trees in
blockchain, detailing the theoretical foundation of our approach. Section 4 evaluates the efficiency
of adaptive Merkle trees through analytical and empirical methods. Section 5 describes the
algorithm for Merkle Tree restructuring, followed by Section 6, which presents examples of the
algorithm's execution in various scenarios. Section 7 delves into the specifics of path encoding in
adaptive Merkle Trees, and Section 8 explores the enhancement of Verkle Trees through adaptive
restructuring. The discussion in Section 9 synthesizes our results, comparing them with existing
solutions and highlighting our contribution to the field. Finally, the conclusion in Section 10
summarizes our research contributions and outlines future directions for this promising area of
study.
**2. Conceptualizing the Problem**
The core issue addressed in this research is the optimization of tree structures in blockchain
systems for efficient and cost-effective verification. Currently, blockchain data is stored in
balanced trees, with Merkle paths for data verification being approximately equal in length and
complexity across all data. This uniformity results in a consistent verification cost and complexity,
regardless of the frequency of data use.
Figure 3 depicts a balanced Merkle Tree, a fundamental data structure used in blockchain for
ensuring data integrity. Each leaf node (A-P) represents a block of data with a unique hash value,
while the non-leaf nodes (AB, CD, etc.) are hashes of their respective child nodes. The root node
(ABCDEFGHIJKLMNOP) encompasses the entire tree's hash, providing a single point of
reference for the entire dataset's integrity.
The Merkle Tree's structure ensures that any alteration in a single data block can be quickly
detected by recalculating the hashes up the tree to the root. However, this balanced structure, while
efficient in evenly distributing the data, does not account for the frequency of data access or
modification (frequency is indicated in brackets). As a result, frequently used data and rarely
accessed data have the same level of complexity and cost in terms of verification, leading to
inefficiencies in resource utilization.
-----
**Figure 3: Balanced Merkle Tree Structure**
Figure 4 highlights the Merkle Path (nodes B, CD, EFGH, IJKLMNOP) for verifying the integrity
of leaf node A (with a high frequency of 0.2041). The Merkle Path is marked in red, indicating the
nodes whose hashes are required to verify A's integrity up to the root. The leaf node A and the root
are highlighted in green, while the intermediate nodes (AB, ABCD, ABCDEFGH) involved in
hash calculations are in yellow.
The verification process involves recalculating and comparing the hashes from node A up to the
root, ensuring data integrity. However, this method, while straightforward, applies the same
verification complexity to all data, regardless of usage frequency. This "one-size-fits-all" approach
is suboptimal, especially for data that is accessed and modified frequently, as it incurs unnecessary
computational overhead.
**Figure 4: Merkle Path (red nodes B, CD, EFGH, IJKLMNOP) for Leaf Node A**
In Figure 5, the Merkle Path for verifying leaf node G (with a frequency of 0.0612) is shown. The
path (nodes H, EF, ABCD, IJKLMNOP) is marked in red, with node G and the root in green, and
the intermediate nodes (GH, EFGH, ABCDEFGH) in yellow. The verification process for G
-----
follows the same principle as for A, recalculating hashes along the red path to validate the data's
integrity.
**Figure 5: Merkle Path (red nodes H, EF, ABCD, IJKLMNOP) for Leaf Node G**
The Figure 6 demonstrates the Merkle Path for leaf node P (with a frequency of 0.0102), with the
path (nodes O, MN, IJKL, ABCDEFGH) in red, P and the root in green, and intermediate nodes
(OP, MNOP, IJKLMNOP) in yellow. The process for verifying P's integrity mirrors that of A and
G, emphasizing the consistent approach across the tree.
**Figure 6: Merkle Path (red nodes O, MN, IJKL, ABCDEFGH) for Leaf Node P**
This consistency in verification, while ensuring uniform security and integrity checks, does not
account for the varying frequencies of data access and modification. It leads to a rigid and
sometimes inefficient system, especially in a dynamic environment like blockchain, where data
access patterns can vary significantly.
Thus, the current Merkle Tree verification process, as illustrated in these figures, is a rather
primitive and blunt approach. It treats all data equally, irrespective of its usage frequency, leading
-----
to potential inefficiencies in computational resources. Our proposed solution aims to revolutionize
this process by introducing adaptive Merkle Trees. These trees will optimize verification paths
based on data usage frequency, significantly reducing the complexity and cost of verifying
frequently accessed data. This innovative approach promises to enhance the efficiency and
scalability of blockchain systems, tailoring the verification process to the dynamic needs of the
network. By differentiating between frequently and infrequently accessed data, adaptive Merkle
Trees can allocate computational resources more effectively, ensuring faster and more costefficient data verification. This method not only optimizes the blockchain's performance but also
aligns with the evolving nature of blockchain usage, where certain data nodes may become
hotspots of activity.
**3. Our Idea for Optimizing Trees in Blockchain**
Figure 7 represents an innovative adaptation of the traditional Merkle Tree, incorporating
principles of Shannon-Fano and Huffman statistical coding. Unlike the balanced Merkle Tree, this
adaptive structure is intentionally unbalanced to optimize the verification process based on the
frequency of data usage. Each leaf node (A-P) still represents a block of data with a unique hash
value, but their placement in the tree now correlates with the probability of their usage.
In this adaptive Merkle Tree, the most frequently used data nodes (A, B, C, D) are positioned
closer to the root, significantly shortening the path required for their verification. This strategic
placement reduces the computational complexity and time required for verifying frequently
accessed data. Conversely, less frequently used data nodes (M, N, O, P) are placed further from
the root, reflecting their lower probability of access.
**Figure 7: Adaptive Merkle Tree**
-----
The structure of this tree is a direct application of Shannon-Fano and Huffman coding principles,
where the most common elements are given shorter codes (or paths in the case of a Merkle Tree).
This approach ensures that the average path length for verification is minimized, aligning the
computational effort with the actual usage patterns of the data within the blockchain.
In the Figure 8, the Merkle Path for leaf node A (highlighted in green) is significantly shorter than
in a balanced Merkle Tree. The path (marked in red) includes nodes DHG and CJLONFBEMPKI,
with intermediate calculations (in yellow) at node ADHG. This optimized path reflects the high
frequency of usage for node A, making the verification process faster and more cost-effective. The
integrity of node A can be verified with fewer computational steps, demonstrating the efficiency
of the adaptive Merkle Tree in handling frequently used data.
For leaf node G (Figure 9), the Merkle Path includes nodes H, D, and CJLONFBEMPKI, with
intermediate calculations at nodes HG and ADHG. This path, while longer than that for node A, is
still optimized based on the usage frequency of G. The adaptive tree structure ensures that the
verification process remains efficient, even for nodes with moderate usage. This approach balances
the need for data integrity with computational efficiency, tailoring the verification complexity to
the usage pattern of each node.
**Figure 8: Optimized Merkle Path for High-Frequency Leaf Node A**
-----
**Figure 9: Adaptive Merkle Path for Moderately Used Leaf Node G**
The Merkle Path for leaf node P (Figure 10), a less frequently used node, is longer, including nodes
M, K, I, E, B, CJLONF, and ADHG. The path reflects P's lower usage frequency, with more
intermediate calculations (nodes MP, MPK, MPKI, EMPKI, BEMPKI, and CJLONFBEMPKI)
required for verification. While this makes the verification process for P more resource-intensive,
it is justified by the node's infrequent use. This example illustrates how the adaptive Merkle Tree
allocates computational resources more efficiently, focusing on optimizing the paths for more
frequently used nodes.
-----
**Figure 10: Extended Merkle Path for Low-Frequency Leaf Node P**
Thus, the adaptive Merkle Tree approach significantly enhances the efficiency of data verification
in blockchain systems. For high-frequency nodes like A, the verification process is streamlined,
requiring fewer computational steps and resources. This optimization can lead to a verification
process that is up to twice as fast and cost-effective compared to a balanced Merkle Tree.
Conversely, for nodes with lower usage frequencies, like P, the longer verification path is a
reasonable trade-off, considering their infrequent access.
**4. Efficiency of adaptive Merkle trees**
In this work, we delve into the comparative complexity of data integrity verification between the
conventional balanced Merkle Tree and the proposed adaptive Merkle Tree model. The balanced
Merkle Tree's average path length is determined by
### k logm n,
where _m represents the maximum allowable number of child nodes per node (the arity of the tree),_
and n is the count of unique symbols within the alphabet.
Conversely, the adaptive Merkle Tree's average path length mirrors the average length of a
Huffman code, calculated as the weighted sum of all code lengths, with the probabilities of the
corresponding symbols serving as weights:
_n_
_k_ _A_ = _pi_ li,
_i=1_
where _[p]i[ is the probability of the ]_ _[i][ th symbol, and ]_ _[l]i[ is the length of the code for the ]_ _[i][ th symbol. ]_
-----
The theoretical minimum average length of a Huffman code, given a specific probability
distribution, can be derived from the entropy formula:
_n_
_k_ _A_ _H_ = − _pi_ log (m _pi_ ) .
_i=1_
Thus, the efficiency of a Huffman code increases as its average code length approaches the entropy
of the distribution.
For the binary tree example ( _m =_ 2 ) discussed, the Huffman code's average length is
approximately 3.49 bits per symbol, closely approximating the entropy of the symbol probabilities
distribution, which is about 3.46 bits per symbol. These figures suggest that the Huffman code
from our example is remarkably close to the theoretical minimum average code length defined by
entropy. Ideally, if the code were perfectly optimal, its average length would equal the entropy.
Transforming these assessments into a comparison of the complexity of data integrity verification
in both the classical balanced and the proposed adaptive Merkle Tree yields:
- For a balanced binary tree, the average Merkle path length is _k _ 4 ;
- For an adaptive binary Merkle Tree, the average path length is _k А_ 3.49, indicating an
efficiency gain of approximately 13%. This gain is reflected in the reduced average number
of hash computations required for verifying the integrity of leaf data.
The efficiency gain increases with the growing disparity between the probabilities of leaf data. In
the extreme case, where one leaf has a 100% probability and all others have 0%, the maximum
efficiency gain—up to 100%—can be observed. Although this represents a hypothetical scenario,
it is intriguing to model real adaptive Merkle Trees, including non-binary types, and assess the
effectiveness of our proposed solution. In Ethereum, Patricia trees are utilized, and our aim is to
extend our approach to this case as well. Furthermore, algorithms for the gradual restructuring of
balanced trees into an unbalanced form are of particular interest. We propose a protocol for such
gradual restructuring, which utilizes newly added nodes to replace high-frequency nodes in the
existing tree. These high-frequency nodes are relocated within the tree to positions that correspond
to their usage probability, allowing us to incrementally modify the tree's configuration and enhance
the efficiency of blockchain integrity checks without a complete overhaul.
**5. Algorithm for Merkle Tree Restructuring**
The restructuring of a Merkle Tree, aimed at optimizing its efficiency for blockchain applications,
necessitates adherence to two primary criteria:
- Minimization of Average Path Length: The restructuring process must account for the
usage frequency of each leaf, ensuring that the average path length, _k, approaches the A_
theoretical minimum, or the average entropy, _H . The deviation between_ _k and A_ _H is_
assessed through the average discrepancy:
### = kA − H (1)
with each elemental discrepancy defined as:
### =i p li ( i + log (m pi )), (2)
where
-----
_n_
= i .
_i=1_
This requirement mandates the availability of a list of probabilities, _pi, and path lengths,_
_li, for each leaf during the restructuring process, updating only as necessary._
- Minimization of Altered Paths: The algorithm should limit modifications to a minimal
subset of nodes, reflecting the reality that only a few accounts are activated in any given
transaction, including complex smart transactions. This approach ensures that inactive
accounts retain their positions and paths within the tree, preserving the integrity of user
data and access pathways. To adhere to this criterion, the algorithm must maintain a list of
leaves (nodes) eligible for restructuring, focusing solely on those affected by current
transactions.
**Restructuring Algorithm (A Single Iteration)**
Input:
- A tree (or tree fragment) with its root, intermediate nodes, and leaves (the bottom layer of
the tree nodes).
- The probability distribution (frequencies) of the tree's leaves.
- A set (list) of leaves available for restructuring.
- A new leaf and/or a new probability distribution for all tree leaves.
Output:
- A restructured tree (or tree fragment) optimized according to the criterion of minimizing
the average discrepancy ( ).
Algorithm Steps:
- Utilize the set (list) of leaves available for restructuring to formulate all possible
restructuring alternatives for the tree (or tree fragment).
- Evaluate the average discrepancy ( ) for each alternative.
- Select the alternative with the lowest average discrepancy ( ).
- Adopt the selected alternative as the algorithm's output.
The most challenging aspect of this algorithm is Step 1, which involves generating all possible
restructuring alternatives for the tree. This process is crucial for identifying the most efficient tree
configuration that minimizes the average path length while accommodating the dynamic nature of
blockchain transactions. To demonstrate the algorithm's functionality amidst the increasing
number of alternatives, several illustrative examples will be provided, showcasing its application
in various scenarios.
**6. Examples of Merkle Tree Restructuring Algorithm Execution**
To illustrate the algorithm's application, let's consider the case of a binary tree where one new leaf
is added at each iteration. Initially, we form a list of alternatives for all leaves in the previous tree
configuration. Each leaf can be transformed into an intermediate node with two child nodes: one
from the previous configuration and one new (added) leaf.
**6.1 Example 1: Restructuring a Binary Tree by Adding One Leaf**
-----
Suppose we have a small binary ( _m =_ 2 ) balanced tree consisting of two nodes A and B, with
probabilities 7/8 and 1/8, respectively (see Fig. 11, a).
a) b) c)
**Figure 11: Binary Tree Restructuring (First Iteration)**
Assume that at the next moment, a new leaf C is created with probabilities now equal to:
A (1/2), B (1/4), C (1/4).
Our goal is to add this new leaf C in such a way as to minimize the average discrepancy (1). Here
and subsequently, we assume that all branches from the previous tree configuration are available
for addition.
**6.1.1. First Iteration**
On the first iteration, we have two alternatives for adding the new leaf to the previous tree
configuration. These alternatives are presented in Fig. 11 and Table 1.
The first alternative (see Fig. 11, b) corresponds to adding node C (1/4) to the branch with node A
(1/2). As we can see from Table 1, this increases the discrepancy . The second alternative (see
Fig. 11, c) is more preferable as the discrepancy (1) here is significantly lower (equals zero), i.e.,
node C (1/4) should be added to the branch with node B (1/4).
Table 1. Discrepancy Values for Two Alternative Ways of Tree Restructuring
First alternative, Fig. 11, b Second alternative, Fig. 11, c
_pi_ _li_ _pli i_ −pi log (2 _pi_ ) i _pi_ _li_ _pli i_ −pi log (2 _pi_ ) i
A ½ 2 1 ½ ½ A ½ 1 ½ ½ 0
B ¼ 1 ¼ ½ -¼ B ¼ 2 ½ ½ 0
C ¼ 2 ½ ½ 0 C ¼ 2 ½ ½ 0
½ 0
Therefore, by the criterion of minimizing (1), we select the second alternative, i.e., the tree
presented in Fig. 11, c. From the perspective of path length, this option is optimal as its average
discrepancy (1) equals zero. Essentially, this indicates that we have achieved an ideal structure for
this probability distribution.
Continuing from the initial iteration of the Merkle Tree restructuring algorithm, let us delve into
subsequent iterations to further elucidate the process and its outcomes.
**6.1.2. Second Iteration**
|First alternative, Fig. 11, b|Col2|Col3|Col4|Col5|Col6|Col7|Second alternative, Fig. 11, c|Col9|Col10|Col11|Col12|Col13|
|---|---|---|---|---|---|---|---|---|---|---|---|---|
||p i|l i|pl i i|−p log (p) i 2 i| i|||p i|l i|pl i i|−p log (p) i 2 i| i|
|A|½|2|1|½|½||A|½|1|½|½|0|
|B|¼|1|¼|½|-¼||B|¼|2|½|½|0|
|C|¼|2|½|½|0||C|¼|2|½|½|0|
||||||½|||||||0|
-----
Let us assume that in the second iteration, a new leaf, D, is introduced, leading to a new probability
distribution among the leaves:
A (1/2), B (1/8), C (1/4), and D (1/8).
Building upon the tree's previous state (refer to Fig. 11, c), we are presented with three distinct
restructuring alternatives (illustrated in Fig. 12):
a) Integrating the new node D (1/8) into the branch containing node A (1/2);
b) Integrating the new node D (1/8) into the branch containing node B (1/8);
c) Integrating the new node D (1/8) into the branch containing node C (1/4).
For each alternative, we calculate the average discrepancy, as detailed in Table 2. The most
favorable alternative, marked by a zero discrepancy, is highlighted in the table.
a) b) c)
**Figure 12: Binary Tree Restructuring (Second Iteration)**
Table 2. Comparison of Alternatives (Second Iteration)
_k_ _A_ _H_
Fig. 12, a 2 1.75 0.25
Fig. 12, b 1.75 1.75 0
Fig. 12, c 1.875 1.75 0.125
Accordingly, the second alternative (see Fig. 12, b) emerges as the most preferable, characterized
by a zero average discrepancy, indicating an optimal restructuring choice under the given criteria.
**6.1.3. Third Iteration**
Advancing to the third iteration, let's hypothesize the addition of another new leaf, E, resulting in
the following probability distribution:
A (1/2), B (1/8), C (1/8), D (1/8), E (1/8).
The tree structure that achieves a zero discrepancy, indicative of an optimal configuration
minimizing the average path length, is depicted in Fig. 13.
|Col1|k A|H||
|---|---|---|---|
|Fig. 12, a|2|1.75|0.25|
|Fig. 12, b|1.75|1.75|0|
|Fig. 12, c|1.875|1.75|0.125|
-----
**Figure 13: Binary Tree Restructuring (Third Iteration)**
Achieving a zero discrepancy signifies that we have attained an optimal tree structure, effectively
minimizing the average path length.
**6.1.4. Fourth Iteration**
During the fourth iteration, we face the task of incorporating an additional leaf, resulting in a new
probability distribution:
A (1/2), B (1/4), C (1/16), D (1/16), E (1/16), F (1/16).
Ideally, a tree structure with a zero discrepancy ( = 0) would align with the configuration
depicted in Figure 14.a. However, this ideal structure cannot be achieved by simply adding a leaf
to the previous configuration (as shown in Figure 13).
a) b)
**Figure 14: Binary Tree Restructuring (Fourth Iteration)**
Utilizing the tree from Figure 13, we identify five potential alternatives (refer to Table 3), none of
which lead to the desired configuration seen in Figure 14.a.
In Table 3, we present a comparison of all possible alternatives based on the average discrepancy
value (1). It becomes evident that the last three alternatives are equivalent in terms of their potential
-----
outcomes, and thus, the choice of restructuring can be made arbitrarily. We decide to employ a
lexicographical ordering rule, selecting alternative "C" as illustrated in Figure 14.b. Although this
new structure is not optimal, it achieves the minimum average discrepancy of 0.125 among all
possible restructuring scenarios.
Table 3. Comparison of Alternatives (Fourth Iteration)
_k_ _A_ _H_
Restructuring Branch with Leaf A 2.4375 2 0.4375
Restructuring Branch with Leaf B 2.3125 2 0.3125
Restructuring Branch with Leaf C 2.125 2 0.125
Restructuring Branch with Leaf D 2.125 2 0.125
Restructuring Branch with Leaf E 2.125 2 0.125
**6.1.5. Iterations 5-10**
For further illustration, let's assume that each subsequent iteration introduces one additional leaf,
with the probability distribution for the tree's leaves as specified in Table 4. This table also
indicates the number of available restructuring alternatives and (in parentheses) the number of
alternatives with the minimum discrepancy value. The last column provides the minimum
discrepancy value (11) across all alternatives. Figures 15-20 showcase the restructuring outcomes
based on the chosen optimal alternative.
Table 4. List of Leaves and Their Probabilities per Iteration, Number of Restructuring
Alternatives, and Minimum Discrepancy Value
|Col1|k A|H||
|---|---|---|---|
|Restructuring Branch with Leaf A|2.4375|2|0.4375|
|Restructuring Branch with Leaf B|2.3125|2|0.3125|
|Restructuring Branch with Leaf C|2.125|2|0.125|
|Restructuring Branch with Leaf D|2.125|2|0.125|
|Restructuring Branch with Leaf E|2.125|2|0.125|
|Iteration number|List of Leaves and Their Probabilities|Number of Restructurin g Alternatives|Minimum Discrepancy Value|
|---|---|---|---|
|5|A (1/2), B (1/4), C (1/16), D (1/16), E (1/32), F (1/16), G (1/32)|6 (1)|0.125|
|6|A (1/4), B (1/4), C (1/16), D (1/8), E (1/16), F (1/8), G (1/16), H (1/16)|7 (6)|0.25|
|7|A (1/4), B (1/4), C (1/16), D (1/8), E (1/32), F (1/8), G (1/32), H (1/16), I (1/16)|8 (1)|0.1875|
|8|A (1/4), B (1/4), C (1/16), D (1/8), E (1/32), F (1/16), G (1/32), H (1/16), I (1/16), J (1/16)|9 (2)|0.125|
|9|A (1/8), B (1/4), C (1/16), D (1/8), E (1/16), F (1/8), G (1/32), H (1/16), I (1/16), J (1/16), K (1/32)|10 (2)|0.185|
|10|A (1/8), B (1/4), C (1/16), D (1/8), E (1/32), F (1/8), G (1/32), H (1/16), I (1/16), J (1/16), K (1/32), L (1/32)|11 (2)|0.185|
-----
**Figure 15: Binary Tree Restructuring**
(Iterations 5)
**Figure 17: Binary Tree Restructuring**
(Iterations 7)
**Figure 19: Binary Tree Restructuring**
(Iterations 9)
**Figure 16: Binary Tree Restructuring**
(Iterations 6)
**Figure 18: Binary Tree Restructuring**
(Iterations 8)
**Figure 20: Binary Tree Restructuring**
(Iterations 10)
-----
The table reveals that, despite an increase in the number of alternatives with each iteration, the
number of most preferable restructuring options remains limited. Moreover, despite fluctuations
in the probabilities of individual leaves, we consistently approach an optimal tree structure.
To further enhance the efficiency of tree restructuring, we propose expanding the algorithm's
parameters. This involves forming potential alternatives not just by adding leaves but by
considering various scenarios for swapping the positions of leaf pairs. This approach is particularly
relevant in blockchain transactions, which typically involve at least two accounts, thus allowing
for the repositioning of leaves by exchanging their locations.
This expansion of the algorithm's capabilities demonstrates our commitment to optimizing tree
structures for improved verification efficiency, paving the way for more dynamic and efficient
blockchain architectures.
**6.2. Example 1.1: Binary Tree Restructuring Through Leaf Node Swapping**
To demonstrate the algorithm's functionality, let's revisit the outcome of the sixth iteration from
the previous example, which represented the worst case in terms of average discrepancies, i.e., it
was the most suboptimal structure depicted in Figure 16. We now aim to improve this structure by
swapping the positions of leaf pairs to minimize the discrepancy (Δ).
Instead of considering all possible leaf pairs, we focus only on those leaves whose elementary
discrepancies ( i ) are non-zero. Essentially, the criterion i 0 indicates that the i -th leaf is "out
of place." For the graph in Figure 16, we have:
- Leaves: A, B, C, D, E, F, G, H;
- Probabilities ( _pi ): 0.25, 0.25, 0.06, 0.13, 0.06, 0.13, 0.06, 0.06;_
- Path Lengths ( il ): 2, 3, 4, 3, 4, 4, 4, 2;
- Discrepancies ( i ): 0, 0.25, 0, 0, 0, 0.125, 0, -0.125.
Thus, leaves B, F, and H are candidates for swapping positions, yielding three alternatives:
- Swapping B and F results in Δ = 0.375;
- Swapping B and H results in Δ = 0.0625;
- Swapping F and H results in Δ = 0.125.
Clearly, the best alternative for our example is to swap leaves B and H. Visually, this corresponds
to the same graph shown in Figure 16.1 – identical to Figure 16 but with B and H swapped.
After this restructuring, we observe the following distribution of elementary discrepancies:
- Leaves: A, H, C, D, E, F, G, B;
- Probabilities ( _pi ): 0.25, 0.0625, 0.0625, 0.125, 0.0625, 0.125, 0.0625, 0.25;_
- Path Lengths ( il ): 2, 3, 4, 3, 4, 4, 4, 2;
- Discrepancies ( i ): 0, -0.0625, 0, 0, 0, 0.125, 0, 0.
-----
Now, the only viable alternative is to swap leaves H and F, which results in a zero average
discrepancy, indicating an optimal structure in terms of minimizing the average path length in the
binary tree. The outcome of this optimization is depicted in Figure 16.2.
**Figure 16.1: Graph Optimization Post-Sixth**
Iteration (Swapping Leaves B and H)
**Figure 16.2: Graph Optimization Post-Sixth**
Iteration (Swapping Leaves H and F)
It's important to note that non-binary trees are often used in practical scenarios. For instance,
Ethereum's blockchain utilizes a Patricia-Merkle tree, which can have up to 16 child nodes, i.e.,
### m =16.
Let's demonstrate the algorithm's operation on a simple example of a non-binary ( _m =_ 4 ) tree.
**6.3. Example 2.1: Restructuring a Non-Binary Tree by Adding a Single Leaf**
In this example, we explore a non-binary tree where each node can have up to four children (
### m = 4 ), closely mirroring a simplified real-world scenario of Patricia-Merkle trees in the
Ethereum blockchain.
Let's assume the initial state of the tree is as shown in Figure 11.a, similar to the previous example.
Also, let the newly added leaves and their probability distributions follow the pattern established
in Example 1. These changes in probabilities are summarized in Table 5, which also lists the
number of alternatives and the unique discrepancy value (1). Figures 21-30 depict the
corresponding tree graphs.
Table 5. List of Leaves and Their Probabilities per Iteration, Number of Restructuring
Alternatives, and Minimum Discrepancy Value
|Iteration Number|List of Leaves and Their Probabilities|Number of Alternatives|Minimum Discrepancy Value (1)|
|---|---|---|---|
|1|A (1/2), B (1/4), C (1/4)|3 (1)|0. 25|
|2|A (1/2), B (1/8), C (1/4), D (1/8)|4 (1)|0.125|
|3|A (1/2), B (1/8), C (1/8), D (1/8), E (1/8)|4 (3)|0.25|
|4|A (1/2), B (1/4), C (1/16), D (1/16), E (1/16), F (1/16)|6 (1)|0.375|
-----
|5|A (1/2), B (1/4), C (1/16), D (1/16), E (1/32), F (1/16), G (1/32)|7 (1)|0.34375|
|---|---|---|---|
|6|A (1/4), B (1/4), C (1/16), D (1/8), E (1/16), F (1/8), G (1/16), H (1/16)|7 (1)|0.25|
|7|A (1/4), B (1/4), C (1/16), D (1/8), E (1/32), F (1/8), G (1/32), H (1/16), I (1/16)|9 (1)|0.21875|
|8|A (1/4), B (1/4), C (1/16), D (1/8), E (1/32), F (1/16), G (1/32), H (1/16), I (1/16), J (1/16)|10 (1)|0.15625|
|9|A (1/8), B (1/4), C (1/16), D (1/8), E (1/16), F (1/8), G (1/32), H (1/16), I (1/16), J (1/16), K (1/32)|10 (1)|0.21875|
|10|A (1/8), B (1/4), C (1/16), D (1/8), E (1/32), F (1/8), G (1/32), H (1/16), I (1/16), J (1/16), K (1/32), L (1/32)|12 (1)|0.21875|
**Figure 21: Tree Restructuring**
(Iterations 1)
**Figure 22: Tree Restructuring**
(Iterations 2)
**Figure 23: Tree Restructuring**
(Iterations 3)
**Figure 24: Tree Restructuring (Iterations 4)**
**Figure 25: Tree Restructuring (Iterations 5)**
**Figure 26: Tree Restructuring (Iterations 6)**
-----
**Figure 27: Tree Restructuring (Iterations 7)**
**Figure 28: Tree Restructuring (Iterations 8)**
**Figure 29: Tree Restructuring (Iterations 9)**
**Figure 30: Tree Restructuring (Iterations 10)**
-----
The calculations of discrepancies in Table 5 show that the configuration of the restructured trees
tends towards optimality by minimizing the average path length.
Now, let's demonstrate the algorithm's operation in the mode of swapping positions between pairs
of nodes, as in Example 1.1.
**6.4. Example 2.2: Restructuring a Non-Binary Tree Through Leaf Pair Swapping**
Let's delve into the most challenging scenario from Example 2, which resulted in the highest
discrepancy value. This scenario corresponds to the tree graph obtained after the fourth iteration,
as depicted in Figure 24. We will demonstrate how swapping the positions of leaf pairs can enhance
this structure.
To form a set of alternatives for the graph in Figure 24, we observe:
- Leaves: A, B, C, D, E, F;
- Probabilities ( _pi ): 0.50, 0.25, 0.06, 0.06, 0.06, 0.06;_
- Path Lengths ( il ): 1, 2, 1, 1, 2, 2;
- Discrepancies ( i ): 0.25, 0.25, -0.0625, -0.0625, 0, 0.
Focusing on pairs with differing path lengths, we identify:
- Swapping positions between leaves A and B, Δ=0.625;
- Swapping positions between leaves B and C, Δ=0.1875;
- Swapping positions between leaves B and D, Δ=0.1875.
The last two alternatives are equivalent and halve the discrepancy (1), thus optimizing the final
tree structure (see Figure 24.1).
**Figure 24.1: Graph Optimization After the Fourth Iteration (Swapping Positions Between**
Leaves B and C)
This approach allows for the dynamic restructuring of trees, minimizing the divergence between
the current and optimal graph structures. By combining different rules (adding new leaves and
swapping positions of existing leaves), we can achieve highly efficient structures that minimize
the average path length.
-----
Through this methodology, we underscore the algorithm's capability to swiftly adapt tree
structures, ensuring an optimal configuration that aligns closely with the theoretical minimum
discrepancy. This adaptability is crucial for maintaining efficient data verification processes in
blockchain technologies, where the dynamic nature of transactions necessitates a flexible yet
robust system for ensuring data integrity.
The proposed algorithm exemplifies a significant advancement in optimizing tree structures for
blockchain applications, particularly in scenarios where non-binary trees, such as Patricia-Merkle
trees used in Ethereum, are prevalent. By judiciously applying leaf swapping and addition
strategies, we can significantly enhance the efficiency of these cryptographic structures, paving
the way for more scalable and cost-effective blockchain operations.
In conclusion, let's explore another example of a tree with _m =16, which can be considered as_
restructuring a fragment of the Patricia-Merkle tree in the Ethereum blockchain.
**6.5. Example 2.3: Restructuring a Patricia-Merkle Tree Fragment Through Leaf Pair**
**Swapping**
Imagine we have a fragment of the Patricia-Merkle tree with leaves (and probabilities) assigned
as follows (see Figure 31):
A (0.003906), B (0.0625), C (0.16529), D (0.0625), E (0.0625), F (0.0625), G (0.0625), H
(0.0625), I (0.003906), J (0.0625), K (0.0625), L (0.0625), M (0.0625), N (0.003906), O
(0.0625), P (0.000244), Q (0.0625), R (0.01), S (0.0625), T (0.000244).
For this tree configuration, we have an average discrepancy Δ=0.1297. By swapping the positions
of leaves A (0.003906) and E (0.0625), we achieve the lowest discrepancy Δ=0.0711 among all
possible alternatives. This tree is depicted in Figure 32.
Continuing to apply the algorithm, we swap the positions of leaves P (0.000244) and R (0.01),
resulting in a discrepancy Δ=0.0614 and a graph as shown in Figure 33.
**Figure 31: Fragment of a Patricia-Merkle Tree with** _m =16_
**Figure 32: Result of the First Optimization of the Patricia-Merkle Tree with** _m =16_
-----
**Figure 33: Result of the Second Optimization of the Patricia-Merkle Tree with** _m =16_
Thus, even a small number of algorithm iterations allows for a significant reduction in discrepancy
(1) and optimization of the tree structure, reducing the average path length.
This example underscores the potential of our restructuring algorithm to enhance the efficiency of
Patricia-Merkle trees in blockchain applications. By judiciously swapping the positions of leaf
pairs, we can significantly improve the tree's structure, aligning it closer to the optimal
configuration. This process not only minimizes the average path length but also contributes to the
overall efficiency and scalability of blockchain operations, particularly in systems like Ethereum
where Patricia-Merkle trees play a crucial role in data integrity verification.
**7. Path Encoding in the Adaptive Merkle Tree**
The integration of adaptive Merkle Trees into existing blockchain systems like Ethereum presents
a paradigm shift in data integrity verification. This shift, while promising significant efficiency
gains, also necessitates substantial modifications to current protocols. In Ethereum's existing
structure, an account's address directly determines its encoding path in the Patricia-Merkle Tree.
This encoding, defined by a series of nibbles (four-bit blocks), uniquely maps each address from
the root to a specific leaf in the vast tree structure. The current system's design allows for the
seamless integration of new addresses into this expansive tree.
Adopting an adaptive approach fundamentally alters this scenario. Instead of a balanced structure,
we would deal with a highly unbalanced tree where frequently used leaves are positioned closer
to the root, and less probable leaves are relegated to lower levels. Implementing this directly in the
existing Patricia-Merkle Tree structure is not feasible. However, creating a new tree during a
protocol update in Ethereum could allow for the incorporation of this adaptive approach, radically
changing the concept of path encoding in the tree.
The challenge arises in reconciling existing addresses with new path encodings. In the new
structure, a random address would no longer be tied to a specific path encoding but would merely
determine the leaf's value, not its path from the root. This concept is illustrated in Figures 34 and
35, where Figure 34 shows the simplified path encoding in a balanced tree with corresponding
addresses, and Figure 35 depicts these addresses in an adaptive tree with Huffman code-based path
encodings.
A practical solution to address compatibility issues in the adaptive tree is the tabular storage of
two structures: "account address – path encoding." This approach allows for the adaptation of the
tree based on the usage probabilities of addresses, leading to significant savings in verification
complexity and cost. Simultaneously, it preserves the existing mechanism for generating random
addresses, including the ability to transfer funds to not-yet-created accounts.
-----
**Figure 34: Path Encoding in a Balanced Merkle Tree**
**Figure 35: Adaptive Merkle Tree with Huffman Code-Based Path Encoding**
Figure 34 illustrates the path encoding mechanism within a traditional balanced Merkle Tree, as
utilized in current blockchain systems like Ethereum. Each account address is directly linked to a
unique path encoded by a series of nibbles, efficiently mapping the journey from the tree's root to
-----
the respective leaf. This representation underscores the systematic and predictable nature of path
encoding in a balanced tree structure, highlighting the ease with which new addresses can be
integrated into the expansive tree.
Figure 35 depicts the transformative approach of an adaptive Merkle Tree, where path encodings
are based on Huffman codes. This figure contrasts sharply with Figure 34, showcasing a more
dynamic and usage-frequency-oriented structure. In this adaptive model, the path encoding is no
longer a straightforward derivative of the account address but is instead determined by the
frequency of data access, leading to a highly unbalanced but efficient tree structure. This figure
effectively demonstrates the shift from a uniform to a tailored approach in path encoding, aligning
more closely with the actual usage patterns within the blockchain network.
**Table 6: Correlation between Account Addresses and Path Encodings in Adaptive Merkle Tree**
Leaves Leaf Path Encoding in a Balanced Huffman Code-Based Path
probabilities Merkle Tree (Account Addresses) Encoding in Adaptive Merkle Tree
A 0.2041 0000 00
B 0.1531 0001 110
C 0.1224 0010 100
D 0.1020 0011 010
E 0.0816 0100 1110
F 0.0714 0101 1011
G 0.0612 0110 0111
H 0.0510 0111 0110
I 0.0408 1000 11111
J 0.0306 1001 10100
K 0.0204 1010 111101
L 0.0204 1011 101010
M 0.0102 1100 1111000
N 0.0102 1101 1010111
O 0.0102 1110 1010110
P 0.0102 1111 1111001
Average code length 4 **3.49 (13% more efficient)**
Table 6 presents a crucial solution to the challenge posed by the transition to an adaptive Merkle
Tree: a tabular representation of the correlation between account addresses and their new path
encodings. This table exemplifies the practical approach to reconciling the existing system of
address generation with the innovative path encoding mechanism of the adaptive tree. By
maintaining a record of these correlations, the table ensures that the integrity and functionality of
the blockchain are preserved, even as the system evolves to embrace more efficient data
verification methods. This tabular approach symbolizes a bridge between the legacy structures of
blockchain and the future-oriented adaptive Merkle Tree, ensuring a seamless transition and
operational continuity.
**8. Enhancing Verkle Trees Through Adaptive Restructuring**
The advent of Verkle trees represents a significant leap forward in the optimization of blockchain
storage and verification processes. By combining the succinctness of vector commitments with the
hierarchical structure of Merkle trees, Verkle trees offer a promising solution to scalability and
efficiency challenges in blockchain systems. This section delves into the potential applications of
|Leaves|Leaf probabilities|Path Encoding in a Balanced Merkle Tree (Account Addresses)|Huffman Code-Based Path Encoding in Adaptive Merkle Tree|
|---|---|---|---|
|A|0.2041|0000|00|
|B|0.1531|0001|110|
|C|0.1224|0010|100|
|D|0.1020|0011|010|
|E|0.0816|0100|1110|
|F|0.0714|0101|1011|
|G|0.0612|0110|0111|
|H|0.0510|0111|0110|
|I|0.0408|1000|11111|
|J|0.0306|1001|10100|
|K|0.0204|1010|111101|
|L|0.0204|1011|101010|
|M|0.0102|1100|1111000|
|N|0.0102|1101|1010111|
|O|0.0102|1110|1010110|
|P|0.0102|1111|1111001|
|Average code length||4|3.49 (13% more efficient)|
-----
our adaptive restructuring approach to Verkle trees, exploring how dynamic adjustments to tree
configurations can further enhance their efficiency and applicability in blockchain technologies.
**8.1. Application of Adaptive Trees in Verkle Tree Technology**
Verkle trees, a novel data structure, merge the benefits of Merkle trees with vector commitments,
providing a compact, efficient means of storing and verifying blockchain state. They stand poised
to revolutionize data storage in blockchain by significantly reducing the size of proofs required for
state verification. Our approach, centered on adaptive restructuring, introduces a method to
dynamically adjust Verkle tree configurations based on usage patterns, thereby optimizing both
storage efficiency and verification speed.
Adaptive restructuring in the context of Verkle trees involves the dynamic adjustment of tree
branches and nodes based on the frequency and patterns of data access and updates. This method
leverages statistical analysis to predict which parts of the tree are accessed more frequently,
allowing for a more efficient organization of data. By applying Huffman or Shannon-Fano coding
principles, we can ensure that the most accessed elements are closer to the root, thereby reducing
the path length for common operations.
**8.2. Technology and Advantages**
- Reduced Proof Sizes: By optimizing the structure of Verkle trees to reflect access patterns, we
can significantly reduce the size of proofs required for verifying transactions. This is because
frequently accessed data can be positioned closer to the root, making it quicker and less
resource-intensive to generate and verify proofs.
- Enhanced Verification Speed: Adaptive restructuring can lead to a more efficient verification
process. Shorter paths for frequently accessed data mean that less computational effort is
required to verify transactions, enhancing the overall throughput of the blockchain network.
- Dynamic Scalability: As blockchain systems evolve, so do their storage and access patterns.
Adaptive restructuring allows Verkle trees to dynamically adjust to these changes, ensuring
that the data structure remains optimized for current usage trends. This adaptability is crucial
for maintaining high performance as the system scales.
- Cost Efficiency: By optimizing the path lengths for data access and verification, the proposed
approach can also reduce the cost associated with these operations. In blockchain systems
where transaction costs are a significant concern, such as Ethereum, this can lead to substantial
savings for users and applications.
- Application in Sharding: Verkle trees are particularly well-suited for sharded blockchain
architectures. Adaptive restructuring can enhance the efficiency of cross-shard communication
by optimizing the storage and retrieval of shard-specific data, further improving the scalability
of sharded networks.
Thus, the integration of adaptive restructuring techniques with Verkle tree technology presents a
promising avenue for enhancing blockchain efficiency. By dynamically optimizing data storage
and access patterns, we can achieve significant improvements in proof size, verification speed, and
overall system scalability. This approach not only addresses current scalability and efficiency
challenges but also provides a flexible framework that can adapt to future developments in
blockchain technology. As we continue to explore the potential of adaptive Verkle trees, it becomes
increasingly clear that this innovative approach could play a pivotal role in the next generation of
blockchain systems.
-----
**9. Discussion**
In this work, we have embarked on a comprehensive exploration of optimizing tree structures
within the blockchain ecosystem, addressing the critical challenge of scalability that plagues
current blockchain technologies. Our investigation spans from conceptualizing the inherent
problems associated with traditional Merkle trees to proposing and validating an innovative
approach for adaptive restructuring of these trees to enhance efficiency and scalability in
blockchain systems.
The blockchain paradigm, while revolutionary, faces significant scalability challenges, primarily
due to the inherent limitations of its underlying data structures and consensus mechanisms.
Traditional Merkle trees, despite their widespread adoption for ensuring data integrity and
facilitating efficient verifications, contribute to these scalability issues due to their static nature
and the increasing cost of operations as the blockchain grows.
Existing solutions to blockchain scalability, such as sharding and layer 2 protocols, offer partial
remedies by distributing the workload or offloading transactions. However, these approaches often
introduce complexity or compromise on decentralization and security. Our review of the state of
the art highlights a gap in dynamically optimizing the data structures themselves to directly address
the root causes of inefficiency.
**9.1. Our Contribution**
Our primary contribution lies in the introduction of adaptive Merkle trees, a novel concept that
leverages dynamic restructuring based on usage patterns to optimize path lengths and reduce the
computational overhead associated with data verification and integrity checks. By applying
principles from Huffman and Shannon-Fano coding to the organization of tree nodes, we ensure
that frequently accessed data is more accessible, thereby reducing the average path length and
associated costs.
Through rigorous analysis and examples, we demonstrated the efficiency gains achievable with
adaptive Merkle trees. Our algorithm for Merkle tree restructuring, detailed in Section 5, provides
a systematic approach for dynamically adjusting tree structures, significantly improving upon the
static nature of traditional Merkle trees.
Extending our concept to Verkle trees, we showcased how adaptive restructuring could be applied
to this advanced data structure, further enhancing its efficiency and making it even more suitable
for large-scale blockchain applications. This application not only underscores the versatility of our
approach but also its potential to contribute to the next generation of blockchain technologies.
**9.2. Comparison with Existing Solutions**
In the quest to address blockchain scalability, several innovative solutions have been proposed and
implemented across various platforms. Each of these solutions presents unique advantages and
challenges. Below (Table 7), we provide a comparative analysis of these solutions, including our
adaptive restructuring approach, to highlight their relative strengths and limitations.
**Table 7: Comparison of Scalability Solutions in Blockchain Technology**
|Solution Type|Examples|Advantages|Disadvantages|
|---|---|---|---|
-----
**- Reduces average path length**
**and verification costs.**
**- Concept is newer and less**
**tested in real-world scenarios.**
|Sharding|Ethereum 2.0, Zilliqa|- Distributes workload across multiple chains. - Enhances transaction throughput.|- Increases complexity. - Potential security risks due to smaller validator sets.|
|---|---|---|---|
|Layer 2 Protocols|Lightning Network, Plasma|- Offloads transactions from the main blockchain. - Facilitates faster and cheaper transactions.|- Can introduce centralization points. - Complex to manage and integrate.|
|State Channels|Raiden Network, Celer Network|- Enables off-chain transaction channels. - Instantaneous transaction settlement.|- Requires on-chain settlement for disputes. - Limited to participants in the channel.|
|Sidechains|Liquid Network, POA Network|- Allows for customizable blockchains linked to the main chain. - Facilitates specific use cases and scalability.|- Security is often reliant on the main chain. - Interoperability challenges.|
|Adaptive Merkle Trees|Our Approach|- Dynamically optimizes data structure based on usage. - Reduces average path length and verification costs.|- Requires initial restructuring and maintenance. - Concept is newer and less tested in real-world scenarios.|
The comparative analysis underscores the diversity of approaches to tackling blockchain
scalability, each with its unique trade-offs. Sharding and Layer 2 protocols, while promising
significant throughput improvements, introduce additional layers of complexity and potential
security concerns. State channels and sidechains offer more specialized solutions but are limited
by their applicability and integration challenges.
Our approach, adaptive restructuring of Merkle and Verkle trees, stands out by directly optimizing
the underlying data structure of the blockchain. This method offers a fundamental improvement in
efficiency without introducing external dependencies or significantly altering the blockchain's
operational principles. While it necessitates initial efforts for restructuring and ongoing
maintenance, the benefits of reduced path lengths and lower verification costs present a compelling
case for its adoption. Moreover, being a relatively new concept, it opens up extensive opportunities
for further research and development to fully realize its potential and address any emerging
challenges.
Thus, our work contributes a novel perspective to the field of blockchain research, opening new
avenues for the development of more scalable and efficient blockchain systems. By addressing
scalability at the data structure level, we provide a foundational solution that can be integrated
with other scalability and efficiency-enhancing techniques, offering a comprehensive approach to
overcoming one of the most significant barriers to blockchain adoption.
**10. Conclusion**
The exploration of adaptive restructuring in Merkle and Verkle trees within this study presents a
novel approach to addressing the enduring challenge of blockchain scalability. By dynamically
adjusting the structure of these trees based on usage patterns, we propose a method that
-----
significantly reduces the average path length for verification processes, thereby enhancing the
efficiency and scalability of blockchain systems.
Our contribution to the field of blockchain technology is twofold. Firstly, we introduce a
conceptual framework for the adaptive restructuring of Merkle trees, which lays the groundwork
for practical implementations in existing blockchain infrastructures. Secondly, through a series of
detailed examples, we demonstrate the feasibility and benefits of our approach, highlighting its
potential to optimize verification processes and reduce associated costs.
Comparative analysis with existing scalability solutions reveals that while many approaches offer
improvements in transaction throughput and efficiency, they often introduce additional complexity
or security concerns. In contrast, adaptive restructuring directly targets the underlying data
structure of the blockchain, offering foundational improvements without compromising on
security or introducing external dependencies.
The implications of our research extend beyond theoretical advancements. By providing a scalable
and efficient method for data verification, adaptive restructuring has the potential to facilitate
broader adoption of blockchain technology across various sectors, including finance, supply chain
management, and beyond. It opens up new avenues for blockchain applications that require high
throughput and efficient data integrity verification.
In conclusion, the adaptive restructuring of Merkle and Verkle trees represents a significant step
forward in the quest for blockchain scalability. It offers a unique blend of efficiency, security, and
practicality, making it a promising solution for the next generation of blockchain systems. As the
blockchain ecosystem continues to evolve, the principles and methodologies outlined in this study
will undoubtedly contribute to its growth and maturity, paving the way for more scalable, efficient,
and versatile blockchain architectures.
**References**
[1] Xu K, Zhu J, Song X, Lu Z, editors. Blockchain Technology and Application: Third CCF
China Blockchain Conference, CBCC 2020, Jinan, China, December 18-20, 2020, Revised
Selected Papers. vol. 1305. Singapore: Springer; 2021. https://doi.org/10.1007/978-981-336478-3.
[2] Lee S-W, Singh I, Mohammadian M, editors. Blockchain Technology for IoT Applications.
Singapore: Springer; 2021. https://doi.org/10.1007/978-981-33-4122-7.
[3] Krichen M, Ammi M, Mihoub A, Almutiq M. Blockchain for Modern Applications: A
Survey. Sensors 2022;22:5274. https://doi.org/10.3390/s22145274.
[4] IEEE Draft Standard for Blockchain-based Digital Asset Classification. IEEE P3206/D10,
April 2023 2023:1–20.
[5] Zeba S, Suman P, Tyagi K. Chapter 4 - Types of blockchain. In: Pandey R, Goundar S,
Fatima S, editors. Distributed Computing to Blockchain, Academic Press; 2023, p. 55–68.
https://doi.org/10.1016/B978-0-323-96146-2.00003-6.
[6] Tiwari A. Chapter 14 - Cryptography in blockchain. In: Pandey R, Goundar S, Fatima S,
editors. Distributed Computing to Blockchain, Academic Press; 2023, p. 251–65.
https://doi.org/10.1016/B978-0-323-96146-2.00011-5.
[7] Ethereum Cumulative Unique Addresses n.d.
https://ycharts.com/indicators/ethereum_cumulative_unique_addresses (accessed January
11, 2024).
[8] Ethereum Daily Active Addresses n.d.
https://ycharts.com/indicators/ethereum_daily_active_addresses (accessed January 11,
2024).
-----
[9] Kottursamy K, Sadayapillai B, AlZubi AA, Bashir AK. A novel blockchain architecture
with mutable block and immutable transactions for enhanced scalability. Sustainable
Energy Technologies and Assessments 2023;58:103320.
https://doi.org/10.1016/j.seta.2023.103320.
[10] Li Y, Weng J, Wu W, Li M, Li Y, Tu H, et al. PRI: PCH-based privacy-preserving with
reusability and interoperability for enhancing blockchain scalability. Journal of Parallel and
Distributed Computing 2023;180:104721. https://doi.org/10.1016/j.jpdc.2023.104721.
[11] Nasir MH, Arshad J, Khan MM, Fatima M, Salah K, Jayaraman R. Scalable blockchains —
A systematic review. Future Generation Computer Systems 2022;126:136–62.
https://doi.org/10.1016/j.future.2021.07.035.
[12] Sanka AI, Cheung RCC. A systematic review of blockchain scalability: Issues, solutions,
analysis and future research. Journal of Network and Computer Applications
2021;195:103232. https://doi.org/10.1016/j.jnca.2021.103232.
[13] Sharma A, Pilli ES, Mazumdar AP, Jain A. BLAST-IoT: BLockchain Assisted Scalable
Trust in Internet of Things. Computers and Electrical Engineering 2023;109:108752.
https://doi.org/10.1016/j.compeleceng.2023.108752.
[14] Wang M, Wu Q. Fast intensive validation on blockchain with scale-out dispute resolution.
Computer Standards & Interfaces 2024;89:103820.
https://doi.org/10.1016/j.csi.2023.103820.
[15] Wang Y, Li Y, Suo Y, Qiang Y, Zhao J, Li K. A scalable, efficient, and secured consensus
mechanism for Vehicle-to-Vehicle energy trading blockchain. Energy Reports
2023;10:1565–74. https://doi.org/10.1016/j.egyr.2023.07.035.
[16] Xiao J, Luo T, Li C, Zhou J, Li Z. CE-PBFT: A high availability consensus algorithm for
large-scale consortium blockchain. Journal of King Saud University - Computer and
Information Sciences 2024;36:101957. https://doi.org/10.1016/j.jksuci.2024.101957.
[17] Yu B, Zhao H, Zhou T, Sheng N, Li X, Xu J. OverShard: Scaling blockchain by full
sharding with overlapping network and virtual accounts. Journal of Network and Computer
Applications 2023;220:103748. https://doi.org/10.1016/j.jnca.2023.103748.
[18] Zhen Z, Wang X, Lin H, Garg S, Kumar P, Hossain MS. A dynamic state sharding
blockchain architecture for scalable and secure crowdsourcing systems. Journal of Network
and Computer Applications 2024;222:103785. https://doi.org/10.1016/j.jnca.2023.103785.
[19] Ayyalasomayajula P, Ramkumar M. Optimization of Merkle Tree Structures: A Focus on
Subtree Implementation. 2023 International Conference on Cyber-Enabled Distributed
Computing and Knowledge Discovery (CyberC), 2023, p. 59–67.
https://doi.org/10.1109/CyberC58899.2023.00021.
[20] Jeon K, Lee J, Kim B, Kim JJ. Hardware Accelerated Reusable Merkle Tree Generation for
Bitcoin Blockchain Headers. IEEE Computer Architecture Letters 2023;22:69–72.
https://doi.org/10.1109/LCA.2023.3289515.
[21] Jing S, Zheng X, Chen Z. Review and Investigation of Merkle Tree’s Technical Principles
and Related Application Fields. 2021 International Conference on Artificial Intelligence,
Big Data and Algorithms (CAIBDA), 2021, p. 86–90.
https://doi.org/10.1109/CAIBDA53561.2021.00026.
[22] Knollmann T, Scheideler C. A self-stabilizing Hashed Patricia Trie. Information and
Computation 2022;285:104697. https://doi.org/10.1016/j.ic.2021.104697.
[23] Lin K-W, Chen Y-C. A File Verification Scheme Based on Verkle Trees. 2023 International
Conference on Consumer Electronics - Taiwan (ICCE-Taiwan), 2023, p. 295–6.
https://doi.org/10.1109/ICCE-Taiwan58799.2023.10226788.
[24] Liu H, Luo X, Liu H, Xia X. Merkle Tree: A Fundamental Component of Blockchains.
2021 International Conference on Electronic Information Engineering and Computer
Science (EIECS), 2021, p. 556–61. https://doi.org/10.1109/EIECS53707.2021.9588047.
[25] Mardiansyah V, Muis A, Sari RF. Multi-State Merkle Patricia Trie (MSMPT): High
Performance Data Structures for Multi-Query Processing Based on Lightweight
-----
Blockchain. IEEE Access 2023;11:117282–96.
https://doi.org/10.1109/ACCESS.2023.3325748.
[26] Mitra D, Tauz L, Dolecek L. Graph Coded Merkle Tree: Mitigating Data Availability
Attacks in Blockchain Systems Using Informed Design of Polar Factor Graphs. IEEE
Journal on Selected Areas in Information Theory 2023;4:434–52.
https://doi.org/10.1109/JSAIT.2023.3315148.
[27] Zhao X, Zhang G, Long H-W, Si Y-W. Minimizing block incentive volatility through Verkle
tree-based dynamic transaction storage. Decision Support Systems 2024;180:114180.
https://doi.org/10.1016/j.dss.2024.114180.
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/2403.00406, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "GREEN",
"url": "http://arxiv.org/pdf/2403.00406"
}
| 2,024
|
[
"JournalArticle"
] | true
| 2024-03-01T00:00:00
|
[
{
"paperId": "c6f8f769b41cee1cbedf2d2cf9b77f93b64e4120",
"title": "Minimizing Block Incentive Volatility Through Verkle Tree-Based Dynamic Transaction Storage"
},
{
"paperId": "5902810bbbcfa673f365de85cc492879654ebffb",
"title": "Fast intensive validation on blockchain with scale-out dispute resolution"
},
{
"paperId": "561124f31bd25e111424789993ab44911146e4c6",
"title": "Optimization of Merkle Tree Structures: A Focus on Subtree Implementation"
},
{
"paperId": "9e17bfe2de77407c7166fa160900f7be862596d9",
"title": "A dynamic state sharding blockchain architecture for scalable and secure crowdsourcing systems"
},
{
"paperId": "399066925c3cb0f8de15a90ed4b7631df16f2256",
"title": "A scalable, efficient, and secured consensus mechanism for Vehicle-to-Vehicle energy trading blockchain"
},
{
"paperId": "903d15c229e3b5787a2924cc5c47edc1568d965d",
"title": "OverShard: Scaling blockchain by full sharding with overlapping network and virtual accounts"
},
{
"paperId": "cf360b7e1c5cd0c50556c3bf85baa5561cda28ec",
"title": "A novel blockchain architecture with mutable block and immutable transactions for enhanced scalability"
},
{
"paperId": "121c2c8b926ff8156eaadfe60678866d74bdfcf1",
"title": "A File Verification Scheme Based on Verkle Trees"
},
{
"paperId": "29f6cc329fe4f9b6f41a8b7be2b4ab283d253f74",
"title": "Hardware Accelerated Reusable Merkle Tree Generation for Bitcoin Blockchain Headers"
},
{
"paperId": "e6a079c2e64cd246c67001cf8c95fa211d59e0e9",
"title": "BLAST-IoT: BLockchain Assisted Scalable Trust in Internet of Things"
},
{
"paperId": "26b26ca0b95d865603889313e64ace6617c112b3",
"title": "PRI: PCH-based privacy-preserving with reusability and interoperability for enhancing blockchain scalability"
},
{
"paperId": "78bb887d4dc5609e3603468e89b0179682e2a9bd",
"title": "Blockchain for Modern Applications: A Survey"
},
{
"paperId": "57562777be3443b7bde45689a29c02fa03fba2ee",
"title": "A systematic review of blockchain scalability: Issues, solutions, analysis and future research"
},
{
"paperId": "9006824870cd8fd818d2c42d1d7bc3db38751314",
"title": "Merkle Tree: A Fundamental Component of Blockchains"
},
{
"paperId": "0f33216c3f5ab6a215d553d4258980a0617b650c",
"title": "Review and Investigation of Merkle Tree’s Technical Principles and Related Application Fields"
},
{
"paperId": "b2b434ad786220ef5fb4b10b99bab0a5fd2a5b9e",
"title": "Types of Blockchain"
},
{
"paperId": "34689080e758df6ba9ff41098cbba734be5470ec",
"title": "Cryptography in Blockchain"
},
{
"paperId": "0d63e5bf40c61aeb51608567f181135ba1a8bb0c",
"title": "A Self-Stabilizing Hashed Patricia Trie"
},
{
"paperId": "680b11975830d4f46afb142387e373de2324524b",
"title": "CE-PBFT: A high availability consensus algorithm for large-scale consortium blockchain"
},
{
"paperId": "5aed6b0c6ca5b6bbe62557b9631e331ab99d99c2",
"title": "Multi-State Merkle Patricia Trie (MSMPT): High-Performance Data Structures for Multi-Query Processing Based on Lightweight Blockchain"
},
{
"paperId": "d4ed93fa85263bfaf0998cd5e6e12dd6f32a5bc5",
"title": "Graph Coded Merkle Tree: Mitigating Data Availability Attacks in Blockchain Systems Using Informed Design of Polar Factor Graphs"
},
{
"paperId": null,
"title": "IEEE Draft Standard for Blockchain-based Digital Asset Classification"
},
{
"paperId": "afb9d253720281f261bb967395f613f637b7a81e",
"title": "Scalable blockchains - A systematic review"
},
{
"paperId": "eb2225e14bc47053b9433cc352d31876bb6950b4",
"title": "Blockchain Technology for IoT Applications"
},
{
"paperId": null,
"title": "Blockchain Technology and Application: Third CCF China Blockchain Conference, CBCC 2020, Jinan, China, December 18-20, 2020, Revised Selected Papers. vol. 1305. Singapore"
},
{
"paperId": null,
"title": "Daily Active Addresses"
}
] | 18,465
|
en
|
[
{
"category": "Business",
"source": "external"
},
{
"category": "Environmental Science",
"source": "s2-fos-model"
},
{
"category": "Business",
"source": "s2-fos-model"
},
{
"category": "Economics",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/0263cf4e40d6bc60fe421ed73763e4fbed70c0fd
|
[
"Business"
] | 0.838219
|
Can Decentralization Drive Green Innovation? A Game Theoretical Analysis of Manufacturer Encroachment Selection with Consumer Green Awareness
|
0263cf4e40d6bc60fe421ed73763e4fbed70c0fd
|
[
{
"authorId": "33586843",
"name": "Dan-Ni Cao"
},
{
"authorId": "2118398094",
"name": "Jin Li"
},
{
"authorId": "2109615284",
"name": "Gege Liu"
},
{
"authorId": "37698352",
"name": "R. Mei"
}
] |
{
"alternate_issns": null,
"alternate_names": null,
"alternate_urls": null,
"id": null,
"issn": null,
"name": null,
"type": null,
"url": null
}
|
With the increase of public environmental awareness and the growth of e-commerce, sustainable development promotes the manufacturer to increasingly participate in green innovation and make full use of the online sales channel to enhance competitiveness. Despite decentralized encroachment being widely adopted in business reality, the current literature has commonly paid more attention to centralized encroachment. To complement related research, a dual-channel green supply chain composed of a manufacturer (its retail subsidiary) and a retailer is investigated. We focus on what encroachment strategy (centralization vs. decentralization) drives the green innovation and analyze the impact of consumer green awareness and product substitutability on the manufacturer’s encroachment strategy, green innovation efforts and supply chain performance. Under each encroachment strategy, we build a Stackelberg game model and derive the equilibrium outcome. Then, we theoretically analyze the effects of consumer green awareness and product substitutability on green innovation and each party’s profitability. Our comparative analysis shows what encroachment strategy drives green innovation and what encroachment strategy benefits both parties and social welfare. Numerical studies are also conducted to support the analytical results. Our key findings reveal that decentralization improves the green innovation and achieves a both-win situation for the manufacturer and the retailer. Besides that, decentralization can reduce the environmental damage and increase social welfare as well.
|
# processes
_Article_
## Can Decentralization Drive Green Innovation? A Game Theoretical Analysis of Manufacturer Encroachment Selection with Consumer Green Awareness
**Dan Cao** **[1], Jin Li** **[2], Gege Liu** **[2]** **and Ran Mei** **[3,]***
1 School of Business Administration, Zhejiang Gongshang University, Hangzhou 310018, China;
cd@mail.zjgsu.edu.cn
2 School of Management and E-Business, Modern Business Research Center, Key Research Institute,
Zhejiang Gongshang University, Hangzhou 310018, China; jinli@mail.zjgsu.edu.cn (J.L.);
lgg979492@163.com (G.L.)
3 School of Economics and Management, Tongji University, Shanghai 200092, China
***** Correspondence: 1610334@tongji.edu.cn
[����������](https://www.mdpi.com/article/10.3390/pr9060990?type=check_update&version=2)
**�������**
**Citation: Cao, D.; Li, J.; Liu, G.; Mei,**
R. Can Decentralization Drive Green
Innovation? A Game Theoretical
Analysis of Manufacturer
Encroachment Selection with
Consumer Green Awareness.
_[Processes 2021, 9, 990. https://](https://doi.org/10.3390/pr9060990)_
[doi.org/10.3390/pr9060990](https://doi.org/10.3390/pr9060990)
Academic Editor: Anet Režek
Režek Jambrak
Received: 10 May 2021
Accepted: 31 May 2021
Published: 3 June 2021
**Publisher’s Note: MDPI stays neutral**
with regard to jurisdictional claims in
published maps and institutional affil
iations.
**Copyright: © 2021 by the authors.**
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
[Attribution (CC BY) license (https://](https://creativecommons.org/licenses/by/4.0/)
[creativecommons.org/licenses/by/](https://creativecommons.org/licenses/by/4.0/)
4.0/).
**Abstract: With the increase of public environmental awareness and the growth of e-commerce,**
sustainable development promotes the manufacturer to increasingly participate in green innovation
and make full use of the online sales channel to enhance competitiveness. Despite decentralized encroachment being widely adopted in business reality, the current literature has commonly paid more
attention to centralized encroachment. To complement related research, a dual-channel green supply
chain composed of a manufacturer (its retail subsidiary) and a retailer is investigated. We focus on
what encroachment strategy (centralization vs. decentralization) drives the green innovation and
analyze the impact of consumer green awareness and product substitutability on the manufacturer’s
encroachment strategy, green innovation efforts and supply chain performance. Under each encroachment strategy, we build a Stackelberg game model and derive the equilibrium outcome. Then,
we theoretically analyze the effects of consumer green awareness and product substitutability on
green innovation and each party’s profitability. Our comparative analysis shows what encroachment
strategy drives green innovation and what encroachment strategy benefits both parties and social
welfare. Numerical studies are also conducted to support the analytical results. Our key findings
reveal that decentralization improves the green innovation and achieves a both-win situation for the
manufacturer and the retailer. Besides that, decentralization can reduce the environmental damage
and increase social welfare as well.
**Keywords: green innovation; manufacturer encroachment; consumer green awareness; substitutability**
**1. Introduction**
In recent years, environmental damage and deterioration issues, such as air and water
pollution, greenhouse effect, climate change, landfill waste, acidic rains and noise pollution,
have been widely concerned by lots of the countries in the world. These environmental
hazards are increasingly threatening human health and survival. To address these environmental issues, both governments and firms all over the world have taken action. According
to UK Climate Change Act schedules, the target of 80% reduction by 2050 in greenhouse
gas emissions has been set up. The government in China plans for a 40–45% reduction in
its carbon emissions per-unit GDP by 2025 compared to its 2005 level [1]. Many famous
firms including Wal-Mart, M&S, Tesco and Debenhams take various measures to reduce
their own and suppliers’ carbon footprint in raw material sourcing, production, logistics
and retailing [2].
Green supply chain can be defined as enhancing the overall environmental protection
awareness of the supply chain and promoting the improvement of economic and social
-----
_Processes 2021, 9, 990_ 2 of 24
benefits by utilizing resources in the whole process from production to sales [3]. In addition to traditional supply chain management activities, green supply chain management
(GSCM) puts more emphasis on reducing the harm of its operation to environment and
human health [4]. Many researchers all over the world are devoted to investigating GSCM
and green innovation involving various factors, such as green planning, green product
design, green manufacturing, by-products use and reverse logistics [5,6]. In the context of
environmental protection, green (environmental-friendly) products with appropriate quality and lowest negative effect on the environment have been manufactured [7]. In addition,
lots of manufacturers are willing to make green innovation efforts to improve the supply
chain greenness, e.g., producing new green products, developing green technologies and
using new clean energy to reduce pollutants. For example, Pepsi Cola, a giant beverage
manufacturer, makes use of advanced green technology to replace corrugated materials
with reusable plastic containers to reduce environmental pollution [8]. Green innovation
activities can establish better social image for supply chain members and improve their
core competitiveness by reducing its harm on the environment.
Compared with non-green products, green products are environmentally friendly but
paid at a higher price as well. The reason is that, for a manufacturer, it will cost more to
produce green products, which makes green products more expensive [9]. Intuitively, the
manufacturer implements green innovation activities in the production of green products
only when their benefits exceed the production costs. The key to the problem is whether
consumers are willing to pay a sufficiently high premium to offset the additional costs.
Fortunately, there is plenty of evidence showing that environmentally aware consumers
prefer green products and are willing to pay higher price than non-green ones. StarKist
tuna reports that consumers are willing to buy dolphin safe tuna at a $0.21 higher price
per can than regular one [10]. A survey by Bureau of Energy in Taiwan reveals that 50%
of the respondents in nine developed countries prefer eco-labelled products, which are
produced with the aim of supporting consumer decision-making for environmentally
friendly products by providing transparency and enhancing trust in the environmental
identities of products [11], and 24% of them would like to pay a premium price for these
green products [12]. In this sense, consumer’s stronger willingness to pay for green
products will enhance the demand and then incentive the manufacturer to adopt more
green innovations.
With the fierce market competition and rapid growth of e-commerce business, to
stay competitive and increase demand and profits, many traditional manufacturers have
established online direct sale channels. For example, Samsung and LG sell their mobile
phones through online stores as well as their offline retailers, which is known as a dualchannel supply chain [13]. Specifically, the direct channel often causes a conflict between
the manufacturer and the retailer, which is referred as manufacturer encroachment [14].
So far, there are two encroachment strategies: centralized encroachment and decentralized encroachment. Under centralized encroachment, the manufacturer centrally
makes all decisions (e.g., pricing) for her subsidiary in direct channel. In contrast, the
manufacturer under decentralized encroachment will grant more decision-making power
to her retail subsidiary. As an example, Sony Corporation allows its retail subsidiary,
StylingLife Holdings (a holding company for Sony’s group of retail businesses), to manage
their own retail businesses independently [15]. Centralized encroachment usually leads
to excessive vertical competition and inflexible trade between the manufacturer and the
retailer [16,17]. Decentralized encroachment is hoped to improve the dual-channel interactions and address these concerns. Arya et al. [18] show that decentralized encroachment for
a manufacturer can soften the retail competition by setting a transfer price above marginal
cost to her downstream subsidiary, which increases the wholesale price and eventually
benefits the manufacturer. In this paper, we will focus on these two types of manufacturer
encroachment strategies in the presence of green innovation.
An incentive has arisen to incorporate green innovation into the study of manufacturer
encroachment, where a manufacturer makes green innovation efforts to produce green
-----
_Processes 2021, 9, 990_ 3 of 24
products and sells them in dual channels. In the mobile phones industry, Nokia uses
materials with no toxic flame retardants to produce mobile phones and accessories and
then sells them through both retailers and their own physical/online stores [19]. For green
products, some key factors should be investigated, such as consumer green awareness and
manufacturer’s green innovation efforts. As mentioned earlier, consumer green awareness
will affect the demand and pricing decisions. The manufacturer makes green innovation
efforts to improve the environment performance and attract more customers, but it also
incurs more manufacturing and design/production costs. Therefore, the manufacturer
attempts to decide an optimal level of green innovation efforts to balance the gains from
the consumer’s demand with green awareness and the losses from green innovation
investment. On the other hand, since the manufacturer’s encroachment usually causes
the channel conflicts, the manufacturer also needs to determine the pricing through a
deliberate trade-off between the profits in both channels. Moreover, to examine the effect
of the channel conflicts between the manufacturer and the retailer, we will consider the
product substitutability across the two channels.
The primary goal of this paper is to study what encroachment strategy (centralization
vs. decentralization) drives the green innovation and analyze the impact of consumer
green awareness and product substitutability on the manufacturer’s encroachment strategy,
green innovation efforts and supply chain performance. Based on extant literature, these
issues are not yet fully addressed. To fill this important research gap, we build a Stackelberg
game model in a two-echelon supply chain with a green manufacturer and a retailer. The
manufacturer produces green products and directly sells them to end consumers through
her subsidiary, which may be centralized or decentralized with the manufacturer. In these
settings, the subsidiary will compete with the retailer selling substitutable green products.
We contribute to the literature in the following three ways. First, in terms of modelling, we contribute to the dual-channel literature by incorporating green innovation,
competition and consumer green awareness into game-theoretic models of manufacturer’s
encroachment. The analytical results show that as consumer green awareness increases,
the manufacturer is motivated to improve green innovation efforts, which in turn benefits
both the manufacturer and the retailer. Nevertheless, the higher product substitutability
will reduce the level of the manufacturer’s green innovation efforts. Accordingly, each
firm’s profit will decrease unless the encroachment cost exceeds a threshold. Second, unlike
most existing studies considering centralized encroachment only, this paper contributes
to the GSCM literature by allowing the manufacturer to choose between centralized and
decentralized encroachment. Third, our research of this paper shows how the manufacturer
makes a choice between centralized and decentralized encroachments when making green
innovation efforts in dual channels. Our main findings reveal that decentralized encroachment outperforms centralized encroachment in the level of green innovation efforts and
both members’ profitability. Moreover, decentralized encroachment can also reduce the
environmental damage and improve the social welfare.
The structure of this paper is organized as follows. Section 2 reviews relevant literature.
Section 3 describes the basic assumptions and supply chain model for a manufacturer using
centralized and decentralized encroachment. In Section 4, we derive the equilibriums under
centralized and decentralized encroachments. Section 5 compares these two strategies. In
Section 6, numerical studies are performed to verify the main findings. Finally, Section 7
summarizes the paper and puts forward suggestions for operations management and
future research.
**2. Literature Review**
There are three streams of literature related to our study: green supply chain management, dual-channel supply chain and manufacturer encroachment. These are briefly
reviewed below.
With the development of green supply chain management, many scholars have begun
to study green supply chains from different perspectives. For example, Liu et al. [20] used
-----
_Processes 2021, 9, 990_ 4 of 24
game theory to study the influence of green preference coefficient and competitiveness on
the decision-making of supply chain members in the situations of no competition, manufacturer competition and manufacturer-retailer competition. Green et al. [21] discovered
that green supply chain management could promote economic development, improve
ecological environment and enhance the competitiveness of manufacturer to some extent.
Ghosh and Shah [22] studied the influence of the consumer green preference coefficient
on the decision-making of supply chain members in the two-level green supply chain
under the conditions of manufacturer led, retailer led and Nash equilibrium. He et al. [23]
explored the effects of consumer preference characteristics on the green innovation efforts of the food supply chain, and they found that the change in consumer preference
characteristics is an important factor to motivate supply chain members to make green
innovation efforts. Liu et al. [24] pointed that the market demand for green products is
not only related to product price but also to consumers’ low-carbon preference. Lee [25]
suggested that the supply chain members must participate in green innovation activities at
the same time to achieve a win-win scenario in the CLSC. However, the above literature
shows that the studies on the green supply chain mainly concentrate on product pricing,
green innovation and consumers’ green preference. Moreover, in the above literature,
dual-channel is not involved.
The existing dual-channel green supply chain research mainly focused on the pricing,
green production issues and channel competition, rarely considering the consumer green
awareness. Cai [26] studied the channel selection and coordination of dual-channel supply
chains and concluded that the operating costs of channels, the substitutability of channels and the overall profit of the supply chain will affect the channel selection decisions
of suppliers and retailers. Heydari et al. [27] employed a Stackelberg game method to
research the optimal pricing decision and coordination strategies for a green supply chain
considering the introduction of online channel. Li et al. [8] presented a model to analyze
green production and the pricing of the members in decentralized and centralized decision
scenarios employing the Stackelberg game method. The results demonstrated when customers’ loyalty to retail channel and green cost meet certain conditions, it is beneficial for
both sides to develop direct channel.
Different from the above research, our work will focus on the analysis of consumers’
green preference and products’ substitutability of a dual-channel green supply chain under
different manufacturer’s encroachment strategies. Many researchers claimed that due to the
aggravation of market competition, encroachment will reduce the retailers’ profit [28], but
some other scholars disagreed with this. For example, by reducing double marginalization,
both the manufacturers and the retailers will benefit [29,30]. In the situation of quantity
competition, Arya et al. [18] pointed out that a right transfer price set by the manufacturers
can convey a less aggressive message to the retailers, which can effectively reduce the
retailers’ profit loss. Yoon [13] pointed out that the manufacturer encroachments do
not always pose a threat to the retailers; when the manufacturers participate in retail,
the manufacturers will lower their costs and set lower wholesale prices to the retailer.
However, manufacturer encroachment did cause channel conflict. When the demand
information is asymmetric [31], the retailers can hardly benefit from the direct sales channel
of the suppliers. Ha et al. [32] pointed out that when the quality is endogenous and
the manufacturers have enough flexibility in adjusting the quality, the encroachment is
unfavorable for the retailers.
This paper is also closely related to the study of Arya et al. [18]. From their model, the
retailer has advantages in sales because the retailer is more direct contact with consumers
than the manufacturer. The manufacturer will choose a strategy without encroachment
when the cost is too high. Besides, we study the decision-makings of green manufacturer
under the encroachment. We focus on what encroachment drives green innovation and the
analysis of the impacts of consumer green awareness and channel competitiveness on each
party’s profitability and total social welfare.
-----
_Processes 2021, 9, 990_ 5 of 24
In summary, this paper is different from the previous literature in three aspects. First,
we take the consumer green awareness and channel competitiveness into account in the
dual-channel green supply chain model. Second, we analyze whether the decentralization
strategy with transfer pricing can alleviate the retailer’s revenue loss under the manufacturer’s encroachment. The results show that decentralization with an appropriate transfer
pricing in a dual-channel green supply chains can drive green innovations and benefit both
the retailer and the manufacturer. Third, we analyze and compare the two encroachment
strategies’ decision-making, profits, environmental damage and social welfare.
**3. Model Formulation**
_3.1. Model Description_
We consider a dual-channel green supply chain consisting of a manufacturer (he/him)
and a retailer (she/her). The manufacturer has one upstream (manufacturing and wholesale) subsidiary and a downstream (retail) subsidiary. The upstream manufacturing subsidiary makes green innovation (such as developing and using green technologies) and
provides differentiated green products, namely, product d and product r, to the downstream subsidiary and the retailer, respectively. The two downstream parties then sell final
products to end markets. They engage in a Cournot competition in the retail market. The
manufacturer can choose between two encroachment strategies: centralized and decentralized encroachments. Under centralized encroachment, the manufacturer produces green
products and sells them directly to the consumers in addition to wholesaling them to the
retailer. Under decentralized encroachment, the manufacturer sells green products to its
subsidiary with a transfer price, where the subsidiary can make its own sale decisions
for maximizing its profit. This paper focuses on the impact of consumer green awareness and product competition on the manufacturer’s green innovation and the choice of
encroachment strategies.
To facilitate the expression, the superscripts C and D stand for equilibriums under
centralized and decentralized encroachments. We use the subscripts r, d and m to represent the retailer, subsidiary, and manufacturer, respectively. The decision variables and
parameters used in this paper are shown in Table 1.
**Table 1. Notations list.**
**Notations** **Description**
_i_ Index of firms: manufacturer (i = m), subsidiary (i = d) or retailer (i = r)
_j_ Index of centralized (j = C) or decentralized (j = D) encroachment
_a_ Market base (the intercept of demand function)
_λ_ The coefficient of demand sensitivity for green innovation per unit product
_c_ The manufacturer’s direct sales cost for per unit product, 0 ≤ _c < a_
_θ_ Green innovation effort made by the manufacturer
_wr_ Manufacturer’s wholesale price for the retailer
_wd_ Manufacturer’s transfer price for her subsidiary
_qi_ The product quantities of firm i = d, r and qi > 0
_pi_ The product’s retail price of firm i = d, r and pi > wi > 0
Πi Profit function of firm i = m, d, r
Under centralized encroachment, the manufacturer and the retailer compete in quantities in the market, both of which are rational and pursue the maximization of their interests.
The game sequence of centralized encroachment is as follows: At stage 1, the manufacturer
determines the wholesale price and the level of her green innovation efforts. At stage 2, the
manufacturer and the retailer jointly choose their respective sales volumes. With decentralized encroachment, the manufacturer leaves the retail decision to its sale subsidiary,
and the subsidiary and retailer compete quantitatively in the market to maximize their
own interests. The game order of decentralized encroachment is as follows: At stage 1,
the manufacturer determines the wholesale price of the retailer, the transfer price of the
-----
_Processes 2021, 9, 990_ 6 of 24
subsidiary and the level of green innovation efforts. Then, at stage 2, the subsidiary and
the retailer decide on their respective sales quantities at the same time.
_3.2. Assumptions_
To establish the appropriate models, we make the following assumptions:
The manufacturer and the retailer compete in the market for quantities while the
_•_
environmentally aware consumers are willing to pay a higher price for products with
higher green quality. To characterize these features, we follow similar widely adopted
demand functions [8,33–35] to depict the retail prices for products d and r as follows:
_pd = a −_ _qd −_ _kqr + λθ, pr = a −_ _qr −_ _kqd + λθ_ (1)
where a is the initial market potential, θ stands for the manufacturer’s efforts in green
innovation, pi is the retail price for per unit green product i = d, r, qi is the order
quantities for per unit green product i = d, r. The cross-price sensitivity coefficient k
is less than its price-sensitivity coefficient, i.e., 0 ≤ _k ≤_ 1 [33,34]. In extreme cases, the
value of k = 0 reflects that the two competing products are completely independent.
_λ represents the green preference coefficient of the consumers. The larger it is, the_
higher the price that consumers are willing to pay for green products.
Compared with non-green products, the manufacturer will invest more in green
_•_
innovation, such as low-carbon storage technology, solar technology and new energy
vehicle technology. This is often in the form of fixed costs. For tractability, we consider
_uθ[2]_
a second-order cost function h = 2 [to represent the green innovation investment,]
where u > 0 is the cost coefficient of the investment. It can be seen that the investment
cost of the green product is a convex function of green innovation efforts, that is, the
cost of green input increases with the level of the green innovation effort, which is
consistent with the practical industry operations. This function is commonly used in
the literatures [35].
In addition to the cost of green investment, the manufacturer will also incur the
_•_
cost of producing products. To avoid the trivial, we assume the manufacturer’s unit
production cost is not related with product green innovation and set it to zero. We
use c (0 _c < a) to represent the manufacturer’s unit selling cost while the retailer’s_
_≤_
unit selling cost is normalized to zero, indicating the retailer has cost advantage
in retailing. The retailer’s sales advantage comes from a better understanding of
customer’s preferences and more direct contact with the customers [30].
To make our paper realistic and without loss of generality, we assume the green
_•_
investment cost is sufficiently high, i.e., u > 0 (see the Proofs of Lemmas 1 and 2).
_•_ We assume that the manufacturers and retailers are completely rational, pursuing the
maximization of their respective interests. In the research of many scholars, firms
often seek to maximize their profits as their objectives [36].
**4. The Equilibrium Results**
_4.1. Centralized Encroachment_
Under centralized encroachment, the manufacturer first decides the retailer’s wholesale price (wr) and the level of green innovation efforts (θ), and then, the manufacturer
and the retailer choose their retail quantities (qd and qr) to maximize their own profits.
Given this decision sequence, to ensure sub-game perfection, the game is solved using
backward induction.
Given the manufacturer’s wholesale price and green innovation efforts, the manufacturer solves the following profit-maximization problem:
Max (qd, qr, θ, wr) = wrqr + (pd _c)qd_ (2)
_qd_ [∏]m _−_ _−_ _[u]2[θ][2]_
-----
_Processes 2021, 9, 990_ 7 of 24
In Equation (2), wrqr represents the manufacturer’s wholesale profit and (pd − _c)qd_
denotes the manufacturer’s retail profit. Similarly, the profit-maximization problem of the
retailer can be given by
Maxqr ∏r (qd, qr, wr) = (pr − _wr)qr_ (3)
Solving Equations (2) and (3) simultaneously, the retail quantities of the manufacturer
(qd) and the retailer (qr) can be obtained as follows:
_q[C]d_ [(][θ][,][ w][r][) =][ 2][(][a][ −] _[c][)][ −]_ _[ak][+(][2][ −]_ _[k][)][λθ][ +][ kw][r]_ (4)
4 _k[2]_
_−_
_qr[C][(][θ][,][ w][r][) =][ 2][a][ −]_ [2][w][r][ −] [(][a][ −] _[c][)][k][ + (][2][ −]_ _[k][)][λθ]_ (5)
4 _k[2]_
_−_
Intuitively, Equations (4) and (5) reveal that the direct selling cost depresses the
quantities in the direct channel but stimulates the quantities in the wholesale channel. Thus,
the existence of direct selling cost is helpful to alleviate manufacturer encroachment to
some extent. On the other hand, it implies that the retailer can still survive in the market
despite the manufacturer reaching the end consumers through direct sale channel. As a
matter of fact, instead of replacing the retailer, the manufacturer seeks to maximize the
total profits in two channels.
Anticipating Equations (4) and (5), the manufacturer decides the wholesale price
(wr) and green innovation efforts (θ) simultaneously to maximize the sum of retail and
wholesale profits:
Max (q[C]d [(][θ][,][ w][r][)][,][ q]r[C][(][θ][,][ w][r][)][,][ θ][,][ w][r][)] (6)
_wr,θ_ [∏]m
Substituting Equations (4) and (5) into Equation (6), solving the first-order conditions
of Equation (6) reveals the manufacturer’s equilibrium wholesale price and the green
innovation efforts. Then, we put them back to get the equilibrium results under centralized
encroachment, which are summarized as follows.
**Lemma 1. Under centralized encroachment, when c <** _⌢c =_ _au(2−k)(4+k)_
(8−k[2])u−(2−k)λ[2][, the equilibrium]
_green innovation efforts, price, retail quantities and profits are the following:_
_θ[C]_ = _λ[(2u2 −(8−k)(3k6[2]−)−k)(a2−−(k8)(−64−k+k)kλ[2][2])c]_, _wr[C]_ [=] [4(2−2(k8[2]−)a3+(k[2])au−−c)(k2[3]−] uk)(−62−(2k −)λk[2])cλ[2],
_q[C]d_ [=] [(2−k2)((84−+3kk)[2]a)−u−(8(−2−k[2]k))(c]u6−+(k2)λ−[2]k)cλ[2], qr[C] [=] 2(84−[( 13k−[2]k)u)a−+(ck2−)]ku)(−62−cλk)[2]λ[2][,]
Π[C]m [=] [(2 −k)(2 −26[)2 a([2]8−−23(k8[2]−)u4−k+(2k−[2])kca)(+(6−8k+)λk[2][2]])c[2]]u−2c[2]λ[2] _, and_
2
Πr[C] [=] [24(8{−2u3[(k[2] 1)−u−k)(a2+−ckk)(] −6−cλk[2])}λ[2]][2][ .]
_⌢_
_Otherwise, when c ≥_ _c, the manufacturer will not use this encroachment strategy._
Lemma 1 shows that whether the strategy of centralized encroachment can be used
depends on the encroachment cost. Only when the encroachment cost is less than a
threshold, the manufacturer can encroach into the end market. Otherwise, she has to rely
on the retailer to sell her green products. Next, we shall discuss the effect of consumer
green awareness.
**Proposition 1. Under centralized encroachment, the equilibrium green innovation efforts (θ[C]),**
_wholesale price (wr[C][), sales quantities (][q]i[C][) and profits (][Π][C]m_ _[and][ Π]r[C][) all increase with the consumer]_
_green awareness._
From Proposition 1, the greater the green preference of the consumers, i.e., the stronger
the consumer environmental awareness, they are more inclined to buy green products, and
-----
_Processes 2021, 9, 990_ 8 of 24
it stimulates the demand for products with higher greenness. Certainly, the manufacturer is
incentivized to make more green innovation efforts. Then, the increasing green innovation
efforts incur higher research and development costs, and thus, the manufacturer will raise
wholesale prices for green products to maximize her profits, and the retailer will set higher
retail price accordingly. As a result, both the manufacturer and the retailer will benefit from
increasing consumer green awareness.
**Proposition 2. Under centralized encroachment, the level of green innovation efforts (θ[C]) decreases**
_with the product substitutability (k)._
Proposition 2 suggests that as the competition between the manufacturer and the
retailer increases, the manufacturer tends to reduce green innovation efforts. The reason is
that increasing competition will force the manufacturer to reduce her prices. To maximize
the profits, the manufacturer will choose to reduce the green innovation investment to save
the producing costs. In other words, the costs of increasing green innovation are greater
than the benefits, which thereby curb the manufacturer’s green innovation efforts.
**Proposition 3. Under centralized encroachment, as the product substitutability, k, increases:**
_(a)_ _If c < c1, both players’ profits will decrease;_
_(b)_ _If c1 < c < c2, the manufacturer’s profit will decrease while the retailer’s profit will_
_increase, where c1 and c2 are provided in Appendix A;_
_⌢_
_(c)_ _If c2 < c <_ _c, both players’ profits will increase._
It can be seen from Proposition 3 that as the product substitutability increases, the
change of both players’ profits critically depends on the unit encroachment cost. The threshold of the encroachment cost is affected by several factors such as the consumer green
awareness, the cost-coefficient of green innovation efforts, the product substitutability, and
market demand. From Proposition 3(a), when the manufacturer’s encroachment is sufficiently cost-efficient (c < c1), both players suffer from increasing product substitutability.
This is intuitive because the manufacturer is easy to fall into face-to-face competition with
the retailer for low encroachment cost.
From Proposition 3(b), when the encroachment cost is in an intermediate range
(c1 _<_ _c <_ _c2), increasing product substitutability hurts the manufacturer but bene-_
fits the retailer. The reason is that compared with the low encroachment cost, the higher
encroachment cost will curb the manufacturer’s direct sales and soften the competition
with the retailer, which boosts the demand in the wholesale market. As the market becomes
more competitive, the retailer benefits more from the increasing wholesale demand. However, from the manufacturer’s perspective, she still suffers from the increasing competition
because the growing benefit from the wholesale market is not enough to cover the loss
caused by the reduction of direct channel sales.
Proposition 3(c) suggests that both players always prefer to the higher product substitutability when the encroachment is highly costly. The reason, again, is that the manufacturer’s encroachment in this case is so difficult that she turns to extend the wholesale
market. This effect benefits both the manufacturer and the retailer especially when the
products between the two channels become more substitutable and more competitive.
_4.2. Decentralized Encroachment_
We now move on to study the decentralized encroachment. Under decentralization,
the manufacturer first sets the wholesale price for the retailer, the transfer price for the
subsidiary and the green innovation efforts. Then, the retailer and the subsidiary compete in the market to determine their respective retail quantities. We solve the game by
backward induction.
-----
_Processes 2021, 9, 990_ 9 of 24
Given the manufacturer’s wholesale and transfer prices, the retailer’s retail decision
under decentralized encroachment is the same as that in Equation (3) under centralized
encroachment. The subsidiary solves the following profit maximization problem:
Maxqd [∏]d (qd, qr, wd) = (pd − _wd −_ _c)qd_ (7)
By solving Equations (3) and (7) simultaneously, the retail quantities, qd and qr, of the
subsidiary and the retailer can be given by
_qd[D][(][θ][,][ w][r][,][ w][d][) =][ 2][(][a][ −]_ _[c][)][ −]_ [2][w][d][ −] _[ak][ +][ k][w][r][ + (][2][ −]_ _[k][)][λθ]_, (8)
4 _k[2]_
_−_
_qr[D][(][θ][,][ w][r][,][ w]d[) = (][2][ −]_ _[k][)][a][ −]_ [2][w][r][ +][ ck][ +][ kw][d][+(][2][ −] _[k][)][λθ]_ . (9)
4 _k[2]_
_−_
As expected, the subsidiary’s retail quantities decrease with the transfer price and
the encroachment cost while the retailer’s retail quantities increase with these parameters.
Compared with Equations (4) and (5) under centralized encroachment, in addition to the
encroachment cost, the transfer price imposes a double-marginalization effect on the direct
channel, which conveys a less aggressive posture to the retailer and thus may increase his
market share.
Given the responses in Equations (8) and (9), the manufacturer decides the wholesale
price (wr), transfer price (wd) and green innovation efforts to maximize the sum of retail
and wholesale profits:
Max (qd[D][(][θ][,][ w][r][,][ w][d][)][,][ q]r[D][(][θ][,][ w][r][,][ w]d[)][,][ θ][,][ w][r][,][ w]d[)][.] (10)
_wr,wd,θ_ [∏]m
Substituting Equations (8) and (9) into Equation (10), the first-order conditions of
Equation (10) are solved to obtain the optimal wholesale price, transfer price and green
innovation efforts. Then, we substitute them back and derive the equilibrium results for
decentralized encroachment, as indicated in the following lemma.
**Lemma 2. Under decentralized encroachment, when c <** _⌣c =_ 2au(2−k) _, the equilibrium green_
4u−λ[2]
_innovation efforts, prices, quantities and profits are the following:_
_θ[D]_ = _λ[(3 −2k)a−(2−k)c ]_
2u(2 −k[2]) −λ[2](3−2k) [,]
_wr[D]_ = 22[2uau((22 − −kk[2][2])) − −λcλ[2][2](3(2−−2kk))] [,][ w]d[D] = 2[k2[u2u(2( −a−k[2]ak) −+ckλ[2])(−3c−λ2[2]k])] [,]
_qd[D]_ = 2[22uu((22 − a−k[2]2)c −−akλ[2])+(3−cλ2[2]k)] [,] _qr[D]_ = 2[22uu(2( −a−kak[2])+ −ckλ)[2]−(3c−λ[2]2k)] [,]
Πm[D] [=] 2u[(3 −4[22uk)(2a[2] −−k2[2]ac) −(2λ−[2]k()+3−22ck[2])]]−c[2]λ[2] _, and Πr[D]_ = 4[2[2uu(2( −a−kak[2])+ −ckλ)[2]−(3c−λ[2]2]k[2])][2][ .]
_⌣_
_Otherwise, when c ≥_ _c, the manufacturer will not use the decentralized encroachment strategy._
Similar to the equilibriums in Lemma 1 under centralized encroachment, Lemma 2
clearly demonstrates that the manufacturer still chooses to encroach into the end market
only when the direct selling cost is below a threshold. A difference lies in the threshold
_⌢_ _⌣_
value shifting from _c to_ _c . We can verify:_
_⌣_ _⌢_ _auk(2_ _k)[2u(k + 2)_ 3λ[2]]
_−_ _−_
_c_ _c =_ (11)
_−_ _−_
(λ[2] 4u)[u(k[2] 8) + (2 _k)λ[2]]_ _[<][ 0.]_
_−_ _−_ _−_
This comparison implies that the manufacturer can use the centralized encroachment
strategy in a greater range than the decentralized one. This is because the transfer price
between subsidiaries under decentralization strengthens the double marginalization in
direct channel, which in turn increases the difficulty of the manufacturer encroachment.
-----
_Processes 2021, 9, 990_ 10 of 24
**Proposition 4. Under decentralized encroachment, the green innovation efforts (θ[D]), wholesale**
_price (wr[D][), transfer price (][w]d[D][), retail quantities (][q]r[D]_ _[and][ q]d[D][) and profits (][Π]r[D]_ _[and][ Π]m[D][) all increase]_
_with the consumer green awareness._
Proposition 4 shows that consumer green awareness has a positive effect on the decisions under decentralized encroachment, which is consistent with the results of centralized
encroachment. More specifically, as consumer green awareness increases, the manufacturer
will make more green innovation efforts to produce greener products and charge higher
prices for the retailer and the subsidiary. Moreover, both the retailer and the subsidiary
can sell greener products in the end markets. Because of aforesaid reasons, both the
manufacturer and the retailer end up earning more profits.
We next discuss the effect of product substitutability and have the following propositions.
**Proposition 5. Under decentralized encroachment, the green innovation efforts (θ[D]) decrease with**
_the products substitutability (k)._
This proposition is similar to Proposition 2, indicating that the more intense the
competition, the less is the green innovation effort made by the manufacturer. As such,
increasing product competition will induce the manufacturer to reduce green innovation
and produce products with lower greenness to save the costs.
**Proposition 6. Under decentralized encroachment, as the product substitutability, k, increases:**
_⌣_
_(a)_ _If c <_ _c, the manufacturer’s profit will decrease;_
_⌣_
_(b)_ _If c < c3, the retailer’s profit will decrease; if c3 < c <_ _c, it will increase, where c3 is_
_provided in Appendix A._
Proposition 6 exhibits that the effect of product substitutability on each player’s
profitability is related with the encroachment cost. These results under decentralized
encroachment are similar to those in Proposition 3 under centralized encroachment except
for different thresholds. In particular, as the product substitutability increases, when the
encroachment cost is small enough (c < c3), the manufacturer can make full use of the
direct channel to initiate stronger competition with the retailer so that both parties are
_⌣_
worse off. When the encroachment cost is in the middle range (c3 < c < _c ), it is not_
easier for the manufacturer to sell through direct channel, and she in turn relies more on the
retailer to rake in more profit, which instead benefits the retailer. When the encroachment
_⌢_
cost is sufficiently large (c3 < c < _c ), the direct channel is suppressed, but the wholesale_
channel is efficiently expended. In this case, the retailer achieves higher profitability from
extending the wholesale market. It is worthy to note that the manufacturer is always
hurt due to increasing competition. The reasons are as follows: First, for the case of low
encroachment cost, when the product substitutability increases, as mentioned above, the
manufacturer will rely more on the direct channel but limit the wholesale market. This
channel conflict leads to the manufacturer’s profit loss. Second, as shown in Equation (11),
the manufacturer’s decentralized encroachment needs to lower the threshold of the encroachment cost, which makes it impossible for the manufacturer to increase her profit by
softening the channel competition.
Interestingly, we find that the threshold of the retailer’s benefit under decentralized
encroachment is lower than that under centralized encroachment, i.e., c3 < c1. It indicates
that the retailer under decentralized encroachment can benefit from stronger competition
in a more efficient way in comparison with centralized encroachment. The underlying
reason is again the double marginalization caused by the transfer price under decentralized encroachment. This double-marginalization effect in direct channel acts as a part of
encroachment costs and thus reduces the encroachment threshold values.
-----
_Processes 2021, 9, 990_ 11 of 24
**5. Centralized vs. Decentralized Encroachment**
With the equilibrium outcomes under centralized and decentralized encroachments
on hand, we compare them and have the following proposition.
**Proposition 7. (a) The transfer price under decentralized encroachment is set above marginal cost,**
_i.e., wd[D]_ _[>][ 0;]_
_(b) The manufacturer’s retail quantity is lower under decentralized encroachment, i.e., qd[D]_ _< q[C]d_ _[,]_
_while the retailer’s retail quantity is higher under decentralized encroachment, i.e., qr[D]_ _[>][ q]r[C][;]_
_(c) The level of green innovation efforts is higher under decentralized encroachment, i.e., θ[D]_ _> θ[C];_
_(d) The wholesale price is higher under decentralized encroachment, i.e.,wr[D]_ _[>][ w]r[C][.]_
Proposition 7(a) demonstrates that the transfer price under decentralized encroachment is set above marginal cost. In fact, the manufacturer signals to the retailer that
she is less aggressive in the retail competition. As reflected in Proposition 7(b), this posture is detrimental to the manufacturer’s retail market, but it enhances the demand in
the wholesale realm. By convincing the retailer that the manufacturer will not take a
more competitive strategy, decentralized encroachment induces the manufacturer to make
more green innovation efforts in Proposition 7(c) and charge higher wholesale price in
Proposition 7(d). Due to the greener products, environmentally aware consumers are
willing to pay higher prices (we can check that pd[D] _[>][ p][C]d_ [and][ p]r[D] _[>][ p]r[C][). Although higher]_
pricing reduces the loss imposed in retail profits, there are two significant aforementioned
benefits in the wholesale market: higher selling price and more retail quantities. Thus, a
stronger boost in wholesale profit (wr[D][q]r[D] _[>][ w]r[C][q]r[C][) may provide a benefit for both players.]_
We have the following proposition.
**Proposition 8. Under decentralized encroachment, both the manufacturer and the retailer benefit,**
_i.e., Πm[D]_ _[>][ Π][C]m[,][ Π]r[D]_ _[>][ Π]r[C][.]_
Proposition 8 shows that the strategy of decentralized encroachment is beneficial to
both the manufacturer and the retailer. From Proposition 7, the manufacturer’s wholesale
quantity increases while her retail quantity decreases under decentralized encroachment.
The resulting total profit increases, indicating that the manufacturer’s profit in the wholesale
channel is enough to cover the loss of the retail channel. Therefore, when producing green
products, the manufacturer prefers to adopt the strategy of decentralized encroachment.
In addition to evaluate the economic goals, we further investigate the societal and environmental performance. In this paper, the social welfare is considered in the comparisons
of centralized and decentralized encroachments. Typically, consumer surplus, known as
net income of consumers, refers to the difference between the willingness of all consumers
to pay for a certain number of products and the actual total price paid. Consumer surplus
measures the extra benefits that consumers believe they have gained [37]. Based on the
above explanation, firms (the manufacturer and the retailer) in supply chain seek economic
benefits and provide green products to increase consumer surplus. However, their operations also produce emissions and pollutions, which are detrimental to the environment. To
reflect these effects and consistent with previous literature [37–39], social welfare consists
of total profits of all supply chain parties, consumer surplus and environmental damage
impact. Under the encroachment structure j = C, D, each element is calculated as follows.
(1) Supply chain profit SC[j]. The supply chain profit in this study is equal to the total
profits of the two stakeholders, i.e., the manufacturer and the retailer. It is
2
_SC[j]_ = Πm[j] [+][ Π]r[j] [= (][a][ −] _[q]d[j]_ _[−]_ _[kq]r[j]_ [+][ λθ] _[j][ −]_ _[c][)][q]d[j]_ [+ (][a][ −] _[q]r[j]_ _[−]_ _[kq]d[j]_ [+][ λθ] _[j][)][q]r[j]_ _[−]_ _[u][(][θ]2[j][)]_ . (12)
-----
_Processes 2021, 9, 990_ 12 of 24
(2) Consumer Surplus CS[j]. Consumers buy green products and enjoy the surpluses,
which equal to the price increment between the maximum acceptable price and the actual
price. It can be given by
_CS[j]_ = [1]2 [(][q]d[j] [+][ q]r[j] [)]2. (13)
(3) Environmental damage ED[j]. From Proposition 7(c), decentralized encroachment
provides the higher level of green innovation efforts than centralized one. To avoid trivial
results and similar to [5], we take the green innovation efforts (θ[D]) under decentralized
encroachment as a benchmark, where the environmental damage is normalized to zero
(ED[D] = 0) relative to that under centralized encroachment. We use the parameter d to
denote the environmental damage cost of per unit product reduction of green innovation
efforts made by the manufacturer, which measures the level of environmental impact
because of greenness decline. A larger d shows a higher degree of the manufacturer’s
production damage to the environment. The environmental damage under centralized
encroachment can be formulated by
_ED[C]_ = [1] _d_ [+][ q]r[C][)]2. (14)
2 _[d][(][θ][D][ −]_ _[θ][C][)(][q][C]_
We use a quadratic damage function to characterize decreasing marginal returns due
to the fact that additional production will produce more environmental pollution. This
treatment is also widely used in the extant literature [37–40]. To sum up, social welfare
under each encroachment strategy j = C, D can be given as follows.
_SW_ _[j]_ = SC[j] + CS[j] _ED[j]._ (15)
_−_
By comparing the social welfare under centralized and decentralized encroachments,
we have the following proposition.
**Proposition 9. Compared with centralized encroachment:**
_(a) Environmental damage under decentralized encroachment is lower, i.e., ED[D]_ _< ED[C];_
_(b) If d ≤_ _d[∗], the social welfare under decentralized encroachment is lower, i.e., SW_ _[D]_ _≤_ _SW[C]; if_
_d > d[∗], the social welfare under decentralized encroachment is higher, i.e., SW_ _[D]_ _> SW[C], where d[∗]_
_is provided in Appendix A._
Proposition 9(a) indicates that the environmental damage under decentralized encroachment is lower than that under centralized encroachment. This thanks to the higher
level of green innovation efforts led by decentralized encroachment. Proposition 9(b) shows
that when the cost coefficient of environmental damage is sufficiently large, decentralized
encroachment has the potential to prominently reduce the environmental damage and
thereby leads to higher total social welfare. Therefore, from the perspective of environmental and societal performances, the government can make policies, such as reward and
punishment mechanisms, to incentive the manufacturers to adopt decentralized encroachment to produce greener products.
**6. Numerical Studies**
In this section, we use some numerical examples to discuss the impacts of some
key factors on the member’s profit improvement and social welfare due to centralized
vs. decentralized encroachment. These key factors include consumer green awareness
(λ), product substitutability (k) and cost factor of environment damage (d). Similar to the
extant literature [7,20,41,42], in order to simplify the calculation and make the value in the
corresponding interval, we set the parameter values as follows: a = 10, µ = 4, c = 2.
We can verify that the setting of these parameter values meets the basic assumptions of our
model, such as positive demands and profits.
-----
_Processes 2021, 9, 990_ 13 of 24
_6.1. Effect of Consumer Green Awareness (λ) on Profits_
We first analyze the effect of consumer green awareness on profits. To this end, we set
_k = 0.8 and vary λ from 0–1. For ease of expressions, we denote firm i’s profit improvement_
between decentralized and centralized encroachments by ∆Πi = Πi[D] _i_ [,][ i][ =][ m][,][ r][. In]
_[−]_ [Π][C]
this setting, we illustrate the effect of consumer green awareness on each player’s profit
improvement in Figure 1 where (a) and (b) show the effect of consumer green awareness
on the manufacturer’s and the retailer’s profit, respectively.
**Figure 1. The effect of λ on each party’s profit improvement.**
From Figure 1, we can see that as λ increases, the profits of both the manufacturer
and the retailer under either centralized or decentralized encroachment will increase.
These results verify the rightness of Propositions 1 and 4. More importantly, the profit
improvements of both the manufacturer and the retailer due to decentralized encroachment
increase with λ. In other words, both supply chain members can benefit more from
decentralized encroachment when consumer green awareness grows. Therefore, firms are
incentivized to adopt the preferred decentralization strategy under the environment of
advocating green consumptions.
_6.2. Effect of Product Substitutability (k) on Profits_
To examine the effects of the product substitutability (k), we choose c = 0.1 and 5
to represent the low and high levels of the encroachment cost. Let λ = 1 and fix the
other parameters as before. Figures 2 and 3 plot the results, where (a) and (b) show the
results when c = 0.1 and c = 5, respectivelyFirst, for the case of low encroachment cost
(c = 0.1), we observe that the profits of both the manufacturer and the retailer under each
regime decrease with k. These results are consistent with Propositions 3 and 6. For high
encroachment cost (c = 5), the manufacturer’s profit under each regime also decreases with
_⌣_
_k. The reason is_ _c < c2, indicating that when the encroachment cost is sufficiently high,_
the manufacturer will not choose to use the direct channel. To ensure the manufacturer’s
_⌣_
encroachment, we still set the high encroachment cost lower than the threshold, i.e., _c > 5._
Figure 3b shows that the retailer’s profit under decentralized encroachment will increase
as k becomes sufficiently large. This implies that the encroachment cost is very high such
that c > c3. These results are also in line with Proposition 6.
-----
_Processes 2021, 9, 990_ 14 of 24
**Figure 2. The effect of k on the manufacturer’s profit improvement.**
**Figure 3. The effect of k on the retailer’s profit improvement.**
Second, for the case of low encroachment cost (c = 0.1), it can be seen that as k
increases, both players’ profit improvements due to decentralized encroachment first
increase and then decrease. To explain this, note that the manufacturer can efficiently
use her direct channel in this setting. The benefit of decentralized encroachment lies in
allowing the manufacturer to convey reduced competitiveness to increase the wholesale
demand. When k = 0, the two channels are independent, and thus, there is no difference
between these two strategies. As k increases, the competition becomes fiercer, and thus,
the decentralized encroachment can reap the wholesale profit more. However, if k is
very large, the manufacturer under decentralized encroachment will reduce the transfer
price to deal with the stiff competition, which makes decentralized encroachment close
to the centralized one. Specifically, in the extreme case of k = 1, strong competition will
induce the manufacturer to exclude the retailer, which leads to equal profits between these
two strategies.
Third, when the encroachment cost is high (c = 5), it is not easy for the manufacturer
to encroach into the retail market, and thus, the manufacturer will rely more on the
wholesale market. Figures 2b and 3b show that both players’ profit improvements due to
decentralized encroachment increase with k. This means that due to the role of softened
competition played by decentralized encroachment, when the retailer benefits from the
increasing wholesale market, the manufacturer can also benefit from increasing dependence
on it for the case of high encroachment cost. In summary, we observe that each member’s
profit improvement due to decentralized encroachment is always positive. These results
verify that Proposition 8 holds as well.
-----
_Processes 2021, 9, 990_ 15 of 24
_6.3. Effect of λ, k and d on Social Welfares_
We next compare the social welfares between centralized and decentralized encroachments and study how the parameters affect their difference. For convenience, we define
∆SW = SW _[D]_ _SW[C]. Figure 4 is drawn to show these results, where (a) shows the effect_
_−_
of λ when k = 0.8 and d = 10, (b) shows the effect of k when λ = 2 and d = 10, and (c)
illustrates the effect of d when λ = 2 and k = 0.8. First, as λ increases, social welfare under
either centralized or decentralized encroachment also increases. This implies that higher
consumer green awareness not only benefits the firms in the supply chain but only benefits
both the environment and the society. Moreover, the higher the consumer green awareness,
the more the social welfare improvement due to decentralized encroachment. Therefore,
in addition to promoting green consumption, policy makers also need to encourage the
manufacturer to adopt decentralized direct channels in sales.
**Figure 4. The effect of λ, k and d on social welfare’s improvement.**
Second, Figure 4b shows increasing product substitutability (k) has a negative effect on
social welfare under each regime. This is because channel competition reduces both supply
chain members’ profits as indicated in Figures 2 and 3. However, it is interesting to find
that as k increases, the social welfare’s improvement under decentralized encroachment
first increases and then decreases. We have the following reasons. When the encroachment
cost is not high (c = 2), as reflected in Figures 2 and 3, the profit improvement of both
players also first increases and then decreases with k. For k sufficiently large, the level
of green innovation efforts under decentralized encroachment will decrease and closely
reach to that under centralized encroachment. Moreover, increasing competition under
centralized encroachment is helpful to restore the demand, which in turn leads to more
consumer surplus. That is why the social welfare’s improvement will decrease as k is large
-----
_Processes 2021, 9, 990_ 16 of 24
enough. At the extremes, when k is close to 1, centralized encroachment is preferred in
social welfare.
Finally, the environmental damage coefficient has a positive effect on social welfare’s
improvement in a linear way. Consistent with Proposition 9, when d exceeds a threshold,
the manufacturer under decentralized encroachment will make more green innovation
efforts to increase the product greenness, and subsequently reduce the harm to the environment. As a result, the total social welfare under decentralized encroachment will also
increase accordingly.
**7. Conclusions**
The rapid development of green supply chain management and e-commerce enables
a large number of supply chain enterprises to sell green products. In addition, many traditional manufacturers have established online direct sale channels to increase demand and
profits. In this dual-channel supply chain, the upstream manufacturer’s direct sale channel
encroaches into the retailer’s markets, which inevitably affects the downstream retailer’s
and consumers’ decision-making. As a result, manufacturer encroachment may reduce
the retailer’s market and hurt the retailer [32,43,44]. Because of induced lower wholesale
price and increased demand, the encroachment may also benefit the manufacturer and
the retailer [18,45,46]. So far, the manufacturer has two types of encroachment strategies,
namely, centralized and decentralized encroachments. Prior studies commonly pay more
attention on centralized encroachment, where the manufacturer makes centralized retail
decisions on behalf of his subsidiary. In contrast, under decentralized encroachment, the
manufacturer charges a transfer price to his subsidiary and permits the subsidiary to make
its own retail decisions. Despite decentralized encroachment being widely adopted in
business reality, the related studies on it are still limited.
We make a major contribution to studying what encroachment strategy drives green
innovation and analyze the effects of consumer green awareness and product substitutability on the manufacturer’s choice between centralized and decentralized encroachments.
Our analysis and main findings are summarized as follows: (1) Under each encroachment
strategy, increasing consumer green awareness incentives the manufacturer to put in more
efforts in green innovation. This also benefits both the manufacturer and the retailer. (2) Under each encroachment strategy, as the channel competition between the manufacturer and
the retailer intensifies, the manufacturer will reduce her green innovation efforts. (3) Under
centralized encroachment, when the encroachment cost is relatively low, higher product
substitutability will hurt both the manufacturer and the retailer. In contrast, when the encroachment cost is high, both parties can benefit from the increasing product substitutability.
(4) The effect of product substitutability on profits under decentralized encroachment has a
similar pattern to that under centralized encroachment. A difference is that due to the less
threshold value of encroachment cost, the manufacturer under decentralization is always
worse off as the product substitutability increases. (5) Decentralization is more efficient to
drive the manufacturer’s green innovation than centralization. Moreover, decentralization
benefits both the manufacturer and the retailer in profitability. (6) Because of a higher
level of green innovation efforts, decentralization reduces the environmental damage in
comparison with centralization. When the environment damage cost is sufficiently high,
decentralization is also preferred in social welfare.
According to the above results, we can provide meaningful managerial insights and
policy suggestions for firms and policymakers. In the context of supply chain sustainability
and profitability, when the firms participate in green innovation, decentralization is a
promising alternative to conventional centralization strategy. Although centralization is
proved to be perfect for eliminating double marginalization [12,23,47], our study demonstrates that decentralization with an appropriate transfer price above marginal cost can
convey softened retail competition between the manufacturer and the retailer, which best
balances the profitability from the wholesale and retail markets. This leads to higher green
innovation and consequently benefits both parties. Therefore, the government can devise
-----
_Processes 2021, 9, 990_ 17 of 24
various policies, such as monetary subsidies and/or taxations for environmentally friendly
activities [38,48–50], to inspire the adoption of decentralization among supply chain firms
in green innovations.
We acknowledge that our models have some limitations. Therefore, the following
studies may need to be further explored in the future. First, the competition between green
products and ordinary products can be considered in the multi-channel supply chains.
Secondly, research under asymmetric information can be incorporated in the model. Third,
contracts/policies can be considered, such as government subsidies and revenue sharing.
**Author Contributions: Conceptualization and methodology, J.L. and G.L.; formal analysis, D.C. and**
R.M.; writing—original draft preparation, J.L., G.L. and R.M.; writing—review and editing, D.C., J.L.
and G.L.; supervision, D.C., J.L. and R.M. All authors have read and agreed to the published version
of the manuscript.
**Funding: This research received no external funding.**
**Institutional Review Board Statement: Not applicable.**
**Informed Consent Statement: Not applicable.**
**Data Availability Statement: The data used to support the findings of this study are available**
upon request.
**Acknowledgments: This project was co-sponsored by the National Social Science Foundation of**
China (19BGL194), the Zhejiang Provincial Natural Science Foundation of China (LY20G020006 and
LQ19G030007) and the Zhejiang Gongshang University on-line and off-line Hybrid Teaching Reform
Project (1010XJ2919103).
**Conflicts of Interest: The authors declare no conflict of interest.**
**Appendix A. Proofs**
**Proof of Lemma 1. Under centralized encroachment, the game is solved by backward**
induction. At the second stage, the second derivatives of the manufacturer and the retailer
are given by
_∂_ ∏[2]m = _[∂]_ [∏][2]r = 2 < 0.
_−_
_∂[2]qd_ _∂[2]qr_
Hence, ∏m and ∏r are concave in qd and qr, respectively. Then, the first-order conditions associated with Equations (2) and (3) jointly yield unique equilibrium quantities as a
function of the wholesale price and product green degree, as those in Equations (4) and (5).
At the first stage, the Hessian matrix of the manufacturer’s profit function in Equation (6)
is given as follows:
6k[2]−16 _λ(k[3]−4k[2]+8)_
(k[2]−4)[2] (k[2]−4)[2]
_λ(k[2]−2k−4)_ _u(4k+8−2k[2]−k[3])+λ[2](2k−4)_
(k−2)(k+2)[2] (k−2)(k+2)[2]
.
_H =_
The leading principal minors are the following:
_| H1| =_ (6kk[2][2]−−416)[2][ and][ |][ H][2][|][ =][ −] _[λ][2][(][k][2][−][8][k]([+]k[2][12]−[)+]4)[2][2][u][(][3][k][2][−][8][)]_ .
The Hessian matrix is negative definite if | H1| < 0 and | H2| > 0. In order to satisfy
the two aforementioned conditions, the equation u > [3]4 _[λ][2][ must be ensured. Therefore,]_
the condition u > 34 _[λ][2][ is considered to maintain the concavity of the manufacturer’s]_
profit function.
By solving the first-order derivatives of _[∂]∂[Π]w[m]r_ = 0 and _[∂][Π]∂θ[m]_ = 0, we obtain the unique
optimal wr[C] [and][ θ][C][. Substituting them into Equations (4) and (5), we have the quantities of]
_q[C]d_ [and][ q]r[C][.]
-----
_Processes 2021, 9, 990_ 18 of 24
_q[C]d_ [=] [(2−k2)((84−+3kk)[2]a)−u−(8(−2−k[2]k))(c]u6−+(k2)λ−[2]k)cλ[2], qr[C] [=] 2(84−[( 13k−[2]k)u)a−+(ck2−)]ku)(−62−cλk)[2]λ[2][ .]
We define:
_⌢_ _au(2_ _k)(4 + k)_
_−_
_c =_
(8 _k[2])u_ (2 _k)λ[2][ .]_
_−_ _−_ _−_
_⌢_
Note that q[C]d [(] _c ) = 0 and_
_∂q[C]d_ (8 − _k[2])u −_ (2 − _k)λ[2]_
=
_−_
_∂c_ 2(8 3k[2])u (2 _k)(6_ _k)λ[2][ <][ 0.]_
_−_ _−_ _−_ _−_
_⌢_ _⌢_
Hence, we know that q[C]d _[>][ 0 for][ c][ <]_ _c . Similarly, we can verify that when c <_ _c,_
demand and profits in equilibrium are positive, which is the necessary conditions for
Nash equilibrium. Substituting them back, we can obtain all the equilibrium outcomes as
_⌢_
indicated in Lemma 1, for c < _c ._
_⌢_
When c ≥ _c, the direct selling cost is so high that the manufacturer chooses to close_
her direct channel. In this case, the manufacturer just relies on the retailer to sell her
products and does not use the encroachment strategy. This completes the proof. □
**Proof of Proposition 1. The first order conditions of q[C]d** [,][ q]r[C][,][ w]r[C][,][ θ][C][,][ Π][C]m[,][ Π]r[C] [with respect]
to λ are respectively shown as follows:
_dq[C]d_ �k[2]+2k − 8)(8 _c −_ 12a+8ak − 4ck − _ak[2]_ + ck[2][�]
= [2][λ][u],
_dλ_ [(2 _k)(6_ _k)λ[2]_ 2u(8 3k[2])][2]
_−_ _−_ _−_ _−_
_dqr[C]_ �k − 1)(8 _c −_ 12a+8ak − 4ck − _ak[2]_ + ck[2][�]
= [8][λ][u],
_dλ_ [(2 _k)(6_ _k)λ[2]_ 2u(8 3k[2])][2]
_−_ _−_ _−_ _−_
_dwr[C]_
= [2][λ][u][(][4] _[k][2][ −]_ _[k][3][ −]_ [8][)(][8] _[c][ −]_ [12][a][+][8][ak][ −] [4][ck][ −] _[ak][2][ +][ ck][2][�]_,
_dλ_ [(2 _k)(6_ _k)λ[2]_ 2u(8 3k[2])][2]
_−_ _−_ _−_ _−_
_dθ[C]_
=,
_[−][(][8]_ _[c][ −]_ [12][a][+][8][ak][ −] [4][ck][ −] _[ak][2][ +][ ck][2][)[][k][2][λ][2][ +][ u][(][16][ −]_ [6][k][2][)+][λ][2][(][12][ −] [8][k][)]]
_dλ_ [(2 _k)(6_ _k)λ[2]_ 2u(8 3k[2])][2]
_−_ _−_ _−_ _−_
_dΠ[C]m_
= [2][λ][u] [(][8] _[c][ −]_ [12][a][+][8][ak][ −] [4][ck][ −] _[ak][2][ +][ ck][2][�][2]_,
_dλ_ [(2 _k)(6_ _k)λ[2]_ 2u(8 3k[2])][2]
_−_ _−_ _−_ _−_
_dΠr[C]_ �a − _ak + ck)](8_ _c −_ 12a+8ak − 4ck − _ak[2]_ + ck[2][�]
= [32][λ][u][(][k][ −] [1][)[][c][λ][2][ −] [2][u] .
_dλ_ [(2 _k)(6_ _k)λ[2]_ 2u(8 3k[2])][3]
_−_ _−_ _−_ _−_
As 0 < k < 1 and u > [3][λ]4 [2] [, the following conditions are satisfied:]
8c 12a+8ak 4ck _ak[2]_ + ck[2] _< 0,_ _cλ[2]_ 2u(a _ak + ck) < 0_, and
_−_ _−_ _−_ _−_ _−_
(2 _k)(6_ _k)λ[2]_ 2u(8 3k[2]) < 0.
_−_ _−_ _−_ _−_
_dq[C]d_ _m_ _r_
Hence, we have _dλ_ _[>][ 0,][ dq]dλ[Cr]_ _[>][ 0,][ dw]dλ[Cr]_ _[>][ 0,][ d]d[θ]λ[C]_ _[>][ 0,][ d]d[Π]λ[C]_ _> 0 and_ _[d]d[Π]λ[C]_ _> 0. This_
completes the proof. □
**Proof of Proposition 2. The first order derivative of θ[C]** with respect to k are shown
as follows:
_dθ[C]_ = 4λH1
_dk_ [(2 _k)(6_ _k)λ[2]_ 2u(8 3k[2])][2][,]
_−_ _−_ _−_ _−_
where H1 = 16cu − 32αu+4cλ[2]+6ck[2]u − _ck[2]λ[2]+44αku −_ 32cku+2ckλ[2] _−_ 12αk[2]u.
The first and second order derivative of H1 with respect to k are shown as follows:
-----
_Processes 2021, 9, 990_ 19 of 24
_dH1_
= 4u(11 _a_ 8c 6ak + 3ck) + 2cλ[2](1 _k),_
_−_ _−_ _−_
_dk_
_d[2]_ _H1_
= [2cλ[2] + 12u(2 _a_ _c)] < 0._
_−_ _−_
_dk[2]_
So, as k increases, the first order derivative of H1 will decrease. When k = 1,
_dHdk1_ = 20u(a − _c) > 0. Hence,_ _[dH]dk[1]_ _[>][ 0, which means that as][ k][ increase,][ H][1][ will increase.]_
Specifically, when k = 1, H1 = 5c(λ[2] _−_ 2cu)< 0. Therefore, we conclude that H1 < 0
and _[d]dk[θ][C]_ _< 0. This completes the proof. □_
**Proof of Proposition 3. The first order derivative of Πr[C]** [with respect to][ k][ are shown]
as follows:
_dΠr[C]_ _−16[cλ[2]_ _−_ 2u(a − _αk + ck)]H2_
=
_dk_ [(2 _k)(6_ _k)λ[2]_ 2u(8 3k[2])][3][,]
_−_ _−_ _−_ _−_
where H2 = 16au[2] _−_ 4cλ[4] _−_ 16cu[2] _−_ 4aλ[2]u+12cλ[2]u+6ak[2]u[2] _−_ 6ck[2]u[2] + ckλ[4] _−_ 12aku[2] _−_ 2akλ[2]u
+6ckλ[2]u + ak[2]λ[2]u _ck[2]λ[2]u._
_−_
We can see that (2 − _k)(6 −_ _k)λ[2]_ _−_ 2u(8 − 3k[2]) < 0 and cλ[2] _−_ 2u(a − _αk + ck�_ _< 0._
_r_
So, the sign of _[d][Π][C]_
_dk_ [is opposite to that of][ H][2][.]
The first-order condition of H2 with respect to c is:
_dH2_
= 2(8 + 3k[2]) u[2] + (12 + 6k _k[2])u_ (4 _k)λ[4]_ _< 0._
_−_ _−_ _−_ _−_
_dc_
It means H2 decreases with c. We define
_au[2(3 k[2]_ 6k + 8�u (4 + 2k _k[2])λ[2]]_
_−_ _−_ _−_
_c1 =_
2(3k[2] + 8)u[2] (12 + 6k _k[2])λ[2]u+(4_ _k)λ[4][ .]_
_−_ _−_ _−_
When c = c1, H2 = 0.
_⌢_
Comparing _c and c1, we have_
_⌢_
_c −_ _c1 =_
� 2au[4k(3 _k2 −_ 8)u2+2(k3 + k2λ2 − 12k2+12k − 8λ2 +16)λ2u �
(k[2] + 2kλ[2] 10k 4λ[2] + 16)λ[4]]
_−_ _−_ _−_
[u(k[2] 8) + (2 _k)λ[2]][2(3_ _k[2]_ +8)u[2]+(k[2] 6k 12)λ[2]u+(4 _k)λ[4]]_ _[>][ 0.]_
_−_ _−_ _−_ _−_ _−_
_⌢_
It follows that if c < c1, H2 > 0; if c1 < c < _c, H3 < 0. Hence, we know if c < c1,_
_dΠr[C]_ _⌢_ Πr[C]
_dk_ _< 0; if c1 < c <_ _c, ddk_ _[>][ 0.]_
The first-order derivative of Π[C]m [with respect to][ k][ are shown as follows:]
_dΠ[C]m_ =
_[−][2][[][c][λ][2][ −]_ [2][u][(][a][ −] _[ak][ +][ ck][)]][H][3]_
_dk_ 16[2u(k[2] 2)+λ[2](3 2k)] [2][,]
_−_ _−_
where H3 = 16cu − 16au − 4cλ[2]+6aku + ckλ[2].
We can see that the sign of _[d][Π][C]m_ 2au(8−3k)
_dk_ [depends on that of][ H][3][. When][ c][ =][ c][2][ =] 16u+(k−4)λ[2][,]
_⌢_
_H3 = 0. Next, we compare_ _c and c2 in the following:_
_⌢_ _auk[2(3k[2]_ 8)u + (k[2] 8k + 12)λ[2]]
_−_ _−_
_c −_ _c2 =_
(16u + kλ[2] 4λ[2])[(k[2] 8)u + (2 _k)λ[2]]_ _[>][ 0.]_
_−_ _−_ _−_
Moreover, _[dH]dc[3]_ = 16u − _λ[2](4 −_ _k) > 0. This indicates that H3 increases with c. When_
_⌢_ Π[C]m
_c < c2, H3 < 0; when c2 < c <_ _c, H3 > 0. Hence, when c < c2, ddk_ _< 0; when_
_⌢_ Π[C]m
_c2 < c <_ _c, ddk_ _> 0._
-----
_Processes 2021, 9, 990_ 20 of 24
Now, comparing c1 and c2, we have
_au[6ku_ (4 _k)λ[2]][2(3_ _k[2]_ 8)u+(k[2] 8k + 12�λ[2]]
_−_ _−_ _−_ _−_
_c1_ _c2 =_
_−_ [16u + (k 4)λ[2]][2(3 _k[2]_ +8)u[2]+(k[2] 6k 12)λ[2]u+(4 _k)λ[4]]_ _[<][ 0.]_
_−_ _−_ _−_ _−_
_⌢_
This follows that c1 < c2 < _c ._
_m_ _r_
In summary, based on above results, we have: if c < c1, _[d][Π]dk[C]_ _< 0 and_ _[d]dk[Π][C]_ _< 0;_
_m_ _r_ _⌢_ Π[C]m _r_
if c1 < c < c2, _[d][Π]dk[C]_ _< 0 and_ _[d]dk[Π][C]_ _> 0; if c2 < c <_ _c, ddk_ _> 0 and_ _[d]dk[Π][C]_ _> 0. This_
completes the proof. □
**Proof of Lemma 2. With decentralized encroachment, we again use backward induction to**
solve the game. At the second stage, the second conditions of the profit functions are
_∂_ ∏[2]m = _[∂]_ [∏][2]r = 2 < 0.
_−_
_∂[2]qd_ _∂[2]qr_
Therefore, ∏m and ∏r are concave in qd and qr, respectively, which guarantees the
uniqueness of the optimal retail quantities as shown in Equations (8) and (9).
Then, at the first stage, the Hessian matrix obtained from the manufacturer’s profit
function is calculated as follows:
2(3k[2]−8) 2k(2−k[2]) _λ(k[2]−2k−4)_
(k[2]−4)[2] (k[2]−4)[2] (k−2)(k+2)[2]
2k(2−k[2]) 4(k[2]−2) _k[2]λ_
(k[2]−4)[2] (k[2]−4)[2] (k−2)(k+2)[2]
_λ(k[2]−2k−4)_ _k[2]λ_ _u(−k[2]−4k−4)+2λ[2]_
(k−2)(k+2)[2] (k−2)(k+2)[2] (k+2)[2]
.
_H =_
The leading principal minors are _| H1| =_ 2((k3[2]k−[2]−4)8[2])[,] _|H2| =_ 4((k2[2] −−4k)[2][2]) and
_|H3|_ = 2[2 u((kk[2]−−22)()+k+λ[2]2()3[2]−2k)] . The Hessian matrix is negative definite if | H1| _<_ 0,
_| H2| > 0, |H3|_ _<_ 0. In order to meet these mentioned conditions, the equation
2[2 _u(k[2]_ 2) + λ[2](3 2k)] < 0 must be established. Therefore, the condition
_−_ _−_
2[2 _u(k[2]_ 2) + λ[2](3 2k)] < 0 is considered to maintain the concavity of the profit func_−_ _−_
tion for manufacturer. Since u > [3][λ]4 [2] [,][ ∏][m][ is jointly concave in][ w][r][ and][ w][d][ and][ θ][. Solving]
the first order derivatives _[∂]∂[Π]w[m]r_ = 0, _[∂]∂[Π]w[m]d_ = 0, _[∂][Π]∂θ[m]_ = 0, we derive the corresponding
equilibrium prices and product green degree as follows.
_wr[D]_ = 22[2uau((22 − −kk[2][2])) − −λcλ[2][2](3(2−−2kk))] [,][ w]d[D] = 2[k2[u2u(2( −a−k[2]ak) −+ckλ[2])(−3c−λ2[2]k])] [,]
_θ[D]_ = _λ[(3 −2k)a−(2−k)c ]_
2u(2 −k[2]) −λ[2](3−2k) [.]
Substituting them into Equations (8) and (9), the equilibrium quantities are given by:
2u(2 _a_ 2c _ak) + cλ[2]_ 2u(a _ak + ck)_ _cλ[2]_
_−_ _−_ _−_ _−_
_qd[D]_ = 2[2u(2 _k[2])_ _λ[2](3_ 2k)] [,][ q]r[D] = 2[2u(2 _k[2])_ _λ[2](3_ 2k)] [.]
_−_ _−_ _−_ _−_ _−_ _−_
We define a threshold _⌣c so that qdD[(][c][ =]_ _⌣c ) = 0, where_ _⌣c =_ 2au4u(−2−λ[2]k[ .])
Note that _[∂]∂[q]cd[D]_ = − 2[2u(2−k4[2]u)−−λλ[2][2](3−2k)] _[<][ 0.]_
_⌣_ _⌣_
It follows that qd[D] _[>][ 0 for][ c][ <]_ _c . In a similar way, we can prove that when c <_ _c,_
demand and profits in equilibrium are positive, which is the necessary conditions for Nash
_⌣_
equilibrium. Using substitution, for c < _c, the equilibrium outcomes are derived as_
shown in Lemma 2.
-----
_Processes 2021, 9, 990_ 21 of 24
_⌣_
When c ≥ _c, the manufacturer will close the direct channel because of very high_
encroachment cost. In other words, the manufacturer will turn to the single wholesale
channel and not use decentralized encroachment strategy. This completes the proof. □
**Proof of Proposition 4. The first order derivatives of qd[D][,][ q]r[D][,][ w]r[D][,][ θ][D][,][ Π]m[D]** [and][ Π]r[D] [with]
respect to λ are shown as follows:
_dqdλd[D]_ = 2λu[2(u2(−2k−)[(k[2]3) −−λ2[2]k()3a−−2(k2)]−[2]k)c], _[dq]dλr[D]_ = 2λu[2(u1(−2k−)[(k[2]3) −−λ2[2]k()3a−−2(k2)]−[2]k)c],
_dwdλd[D]_ = 2kλ[u2(u1(−2−k)[(k[2])3− −λ2[2]k()3a−−2(k2)]−[2]k)c], _[dw]dλr[D]_ = 2λu[(22u −(2k−[2])[(k[2])3− −λ2[2]k(3)a−−2(k2)]−[2]k)c],
_dθ[D]_ [(3 −2k)a−(2−k)c][(3 −2k)λ[2]+2(2 −k[2])u]
_dλ_ = [2u(2−k[2])−λ[2](3−2k)] [2],
_ddΠλm[D]_ = [λ2uu [((23− −k[2]2)k−)aλ−[2]((32−−2kk))]c][2][2][,]
_ddΠλr[D]_ = 2λu(1 −k)[(3 −[2u2(k2)−a−k[2]()2−−λk)[2]c(][3−2 u2(ka)]−[3]ak+ck) −cλ[2]] .
_d_ _r_ _r_ _m_
Similar to the Proposition 1, we have _[dq][D]_
_dλ_ _[>][ 0,][ dq]dλ[D]_ _[>][ 0,][ dw]dλ[D]_ _[>][ 0,][ d]d[θ]λ[D]_ _[>][ 0,][ d]d[Π]λ[D]_ _[>][ 0,]_
_dΠr[D]_
_dλ_ _[>][ 0. This completes the proof.][ □]_
**Proof of Proposition 5. The first order derivative of θ[D]** with respect to k are shown
as follows:
_dθ[D]_ _−λH3_
=
_dk_ [2u(2 _k[2])_ _λ[2](3_ 2k)] [2][,]
_−_ _−_ _−_
where H3= 8au − 4cu − _cλ[2]_ _−_ 2ck[2]u − 12aku+8cku+4ak[2]u.
We get: _[dH]dk[3]_ = 4u(2c − 3a+2ak − _ck),_ _[d]dk[2]_ _[H][2][3]_ = 4u(2a − _c) > 0._
Similar to the Proposition 2, we have H3 > 0 and _[d]dk[θ][D]_ _< 0. This completes the proof._
**Proof of Proposition 6. The first-order condition of Πm[D]** [with respect to][ k][ is the following:]
�
_dΠm[D]_ = .
_−_ [[][2] _[u][(][a][ −]_ _[ak][ +][ ck][)]_ _[−][c][λ][2][][][2]_ _[au][(][2][ −]_ _[k][)][ −]_ [(][4][u][ −] _[λ][2][)][c]_
_dk_ 2[2u(2 _k[2])_ _λ[2](3_ 2k)][2]
_−_ _−_ _−_
It follows that _[d][Π]dkm[D]_ _< 0, for c <_ _⌣c =_ 2au4u(−2−λ[2]k[ .])
Similarly, the first-order condition of Πr[D] [with respect to][ k][ is in the following.]
_dΠr[D]_ = [[][2] _[u][(][a][ −]_ _[ak][ +][ ck][)][ −]_ _[c][λ][2][]][ H][4]_
_dk_ [2u(2 _k[2])_ _λ[2](3_ 2k)] [3][,]
_−_ _−_ _−_
where H4 = cλ[4] _−_ 4au[2]+4cu[2] + aλ[2]u − 3cλ[2]u − 2ak[2]u[2]+2ck[2]u[2]+4aku[2] _−_ 2ckλ[2]u.
_r_
We can see that the sign of _[d][Π]dk[D]_ depends on that of H4. We define a threshold c3 so
that H5(c = c3) = 0, where c3 = 2(k[2]au+[22) uu([2]k−[2]−(22kk++32))λ−[2]uλ+[2]]λ[4][ .]
Comparing c3 and _⌣c, we have c3 −_ _⌣c =_ (4auu−(2λku[2])[−2λ(k[2][2])[+22u)(uk[2][2]−−(22)+(k+33−)λ2[2]ku)+λ[2]λ][4]] _[<][ 0.]_
The first-order condition of H4 with respect to c is:
_dH4_
= 2(k[2] + 2)u[2] (2k + 3)uλ[2] + λ[4] _> 0._
_−_
_dc_
-----
_Processes 2021, 9, 990_ 22 of 24
_⌣_
This indicates H4 increases with c. Thus, if c < c3, H4 < 0; if c3 < c < _c, H4 > 0._
_r_ _⌣_ Πr[D]
It follows that if c < c3, _[d][Π]dk[D]_ _< 0; if c3 < c <_ _c, ddk_ _> 0. This completes the proof. □_
**Proof of Proposition 7. (a) From lemma 2, wd[D]** = 2[k2[u2u(2( −a−k[2]ak) −+ckλ[2])(−3c−λ2[2]k])] _[>][ 0.]_
(b) Using qd[D] [and][ q]r[D] [from Lemma 2 and][ q][C]d [and][ q]r[C] [from Lemma 1,]
_k(2_ _k)[_ 3λ[2]+2u(2 + k)][cλ[2] 2u(a _ak + ck)]_
_−_ _−_ _−_ _−_
_qd[D]_ _d_ [=]
_[−]_ _[q][C]_ 2[λ[2](k[2] 8k + 12) + 2u(3k[2] 8)][2u(k[2] 2)+λ[2](3 2k)] _[<][ 0,]_
_−_ _−_ _−_ _−_
_k[2][�]λ[2]_ 2u)[cλ[2] 2u(a _ak + ck)]_
_−_ _−_ _−_
_qr[D]_ _r_ [=]
_[−]_ _[q][C]_ 2[λ[2](k[2] 8k + 12) + 2u(3k[2] 8)][2u(k[2] 2)+λ[2](3 2k)] _[>][ 0.]_
_−_ _−_ _−_ _−_
Hence, qd[D] _< q[C]d_ [and][ q]r[D] _[>][ q]r[C][.]_
(c) Using θ[D] from Lemma 2 and θ[C] from Lemma 1,
_k[2]λ�k_ 1)[cλ[2] 2u(a _ak + ck)]_
_−_ _−_ _−_
_θ[D]_ _θ[C]_ =
_−_
[λ[2](k[2] 8k + 12) + 2u(3k[2] 8)][2u(k[2] 2)+λ[2](3 2k)] _[>][ 0.]_
_−_ _−_ _−_ _−_
Hence, θ[D] _> θ[C]._
(d) Using wr[D] [from Lemma 2 and][ w]r[C] [from Lemma 1,]
_k[2][2u�k[2]_ 2)+λ[2](2 _k)][cλ[2]_ 2u(a _ak + ck)]_
_−_ _−_ _−_ _−_
_wr[D]_ _r_ [=]
_[−]_ _[w][C]_ 2[λ[2](k[2] 8k + 12) + 2u(3k[2] 8)][2u(k[2] 2)+λ[2](3 2k)] _[>][ 0.]_
_−_ _−_ _−_ _−_
Hence, wr[D] _[>][ w]r[C][. This completes the proof.][ □]_
**Proof of Proposition 8. Using Πm[D]** [and][ Π]r[D] [from Lemma 2 and][ Π][C]m [and][ Π]r[C] [from Lemma 1,]
_k[2][cλ[2]_ 2u(a _ak + ck)]_ [2]
_−_ _−_
Πm[D] _m_ [=]
_[−]_ [Π][C] 4[λ[2](k[2] 8k + 12) + 2u(3k[2] 8)][2u(k[2] 2)+λ[2](3 2k)] _[>][ 0,]_
_−_ _−_ _−_ _−_
Πr[D] _r_ [=][ k][2][(][λ][2][ −] [2][u][)[][c][λ][2][ −] [2][u][(][a][ −] _[ak][ +][ ck][)]][2][[(]_ _[k][2][ −]_ [16][k][ +][ 24][)][λ][2][+][2][u][(][7][k][2][ −] [16][)]] _> 0._
_[−]_ [Π][C] 4[λ[2](k[2] 8k + 12) + 2u(3k[2] 8)][2][2u(k[2] 2)+λ[2](3 2k)] [2]
_−_ _−_ _−_ _−_
Hence, Πm[D] _[>][ Π][C]m_ [and][ Π]r[D] _[>][ Π]r[C][. This completes the proof.][ □]_
**Proof of Proposition 9. (a) From (14), we know: ED[D]** = 0, ED[C] = [1]2 _[d][(][θ][D][ −]_ _[θ][C][)(][q][C]d_ [+][ q]r[C][)][2][.]
Since θ[D] _θ[C]_ _> 0 from Proposition 7(c), we obtain ED[D]_ _< ED[C]._
_−_
(b) From (15), for j = C, D, SW _[j]_ = SC[j] + CS[j] _ED[j]. From Proposition 8, we can_
_−_
see SC[D] = Πm[D] [+][ Π]r[D] _[>][ SC][C][ =][ Π][C]m_ [+][ Π]r[C][. Using][ q]d[D] [and][ q]r[D] [from Lemma 2 and][ q][C]d [and]
_qr[C]_ [from Lemma 1,]
_k�cλ[2]_ 2u�a _ak + ck)][u(k[2]_ + k 4) + (3 2k)λ[2][�]
_−_ _−_ _−_ _−_ _−_
_qm[D]_ [+][ q]r[D] _m_ _r_ [=]
_[−]_ _[q][C]_ _[−]_ _[q][C]_ [λ[2](k[2] 8k + 12) + 2u(3k[2] 8)][2u(k[2] 2)+λ[2](3 2k)] _[<][ 0.]_
_−_ _−_ _−_ _−_
Hence, from Equation (13), we have CS[D] _< CS[C]. Comparing the social welfare_
under centralized and decentralized encroachments,
2
_SW_ _[D]_ _−_ _SW[C]_ = SC[D] _−_ _SC[C]_ + CS[D] _−_ _CS[C]_ + [1] _d_ [+][ q]r[C][)] .
2 _[d][(][θ][D][ −]_ _[θ][C][)(][q][C]_
We define a threshold d[∗] such that SW _[D]_ = SW[C] for d = d[∗], where
_d[∗]_ = [2][(][CS][C][ −] _[CS][D][ +][ SC][C][ −]_ _[SC][D][)]_ .
(θ[D] _−_ _θ[C])(q[C]d_ [+][ q][Cr][ )][2]
-----
_Processes 2021, 9, 990_ 23 of 24
Therefore, we have if d _d[∗], SW_ _[D]_ _SW[C]; if d > d[∗], SW_ _[D]_ _> SW[C]. This completes_
_≤_ _≤_
the proof. □
**References**
1. Wang, K.; Wei, Y.M.; Huang, Z. Potential Gains from Carbon Emissions Trading in China: A DEA Based Estimation on Abatement
Cost Savings. Omega 2016, 63, 48–59.
2. Ramanathan, U.; Bentley, Y.; Pang, G. The Role of Collaboration in the UK Green Supply Chains: An Exploratory Study of the
[Perspectives of Suppliers, Logistics and Retailers. J. Clean. Prod. 2014, 70, 231–241. [CrossRef]](http://doi.org/10.1016/j.jclepro.2014.02.026)
3. [Sarkis, J. A Strategic Decision Framework for Green Supply Chain Management. J. Clean. Prod. 2003, 11, 397–440. [CrossRef]](http://doi.org/10.1016/S0959-6526(02)00062-8)
4. Linton, J.D.; Klassen, R.; Jayaraman, V. Sustainable Supply Chains: An Introduction. J. Oper. Manag. 2007, 25, 1075–1082.
[[CrossRef]](http://doi.org/10.1016/j.jom.2007.01.012)
5. Hong, Z.; Guo, X. Green Product Supply Chain Contracts Considering Environmental Responsibilities. Omega 2019, 83, 155–166.
[[CrossRef]](http://doi.org/10.1016/j.omega.2018.02.010)
6. Basiri, Z.; Heydari, J. A Mathematical Model for Green Supply Chain Coordination with Substitutable Products. J. Clean. Prod.
**[2017, 145, 232–249. [CrossRef]](http://doi.org/10.1016/j.jclepro.2017.01.060)**
7. Hong, P.; Jagani, S.; Kim, J.; Youn, S.H. Managing Sustainability Orientation: An Empirical Investigation of Manufacturing Firms.
_[Int. J. Prod. Econ. 2019, 211, 71–81. [CrossRef]](http://doi.org/10.1016/j.ijpe.2019.01.035)_
8. Li, B.; Hou, P.-W.; Chen, P.; Li, Q.-H. Pricing Strategy and Coordination in a Dual Channel Supply Chain with a Risk-Averse
[Retailer. Int. J. Prod. Econ. 2016, 178, 154–168. [CrossRef]](http://doi.org/10.1016/j.ijpe.2016.05.010)
9. Conrad, K. Price Competition and Product Differentiation when Consumers Care for the Environment. Environ. Resour. Econ.
**[2005, 31, 1–19. [CrossRef]](http://doi.org/10.1007/s10640-004-6977-8)**
10. Reinhardt, F.L. Environmental Product Differentiation: Implications for Corporate Strategy. Calif. Manag. Rev. 1998, 40, 43–73.
[[CrossRef]](http://doi.org/10.2307/41165964)
11. Thøgersen, J. Eco-labeling is one among a number of policy tools that are used in what. In New Tools for Environmental Protection:
_Education, Information, and Voluntary Measures; Dietz, T., Stern, P.C., Eds.; National Academy Press: Washington, DC, USA, 2002;_
pp. 83–104.
12. Ranjan, A.; Jha, J.K. Pricing and Coordination Strategies of a Dual-Channel Supply Chain Considering Green Quality and Sales
[Effort. J. Clean. Prod. 2019, 218, 409–424. [CrossRef]](http://doi.org/10.1016/j.jclepro.2019.01.297)
13. [Yoon, D. Supplier Encroachment and Investment Spillovers. Prod. Oper. Manag. 2016, 25, 1839–1854. [CrossRef]](http://doi.org/10.1111/poms.12580)
14. Li, T.T.; Xie, J.X.; Zhao, X.B. Supplier Encroachment in Competitive Supply Chains. Int. J. Prod. Econ. 2015, 165, 120–131.
[[CrossRef]](http://doi.org/10.1016/j.ijpe.2015.03.023)
15. [Sony Sells a Portion of Its Shares in StylingLife Holdings. Available online: https://www.sony.net/SonyInfo/News/Press/2006](https://www.sony.net/SonyInfo/News/Press/200612/06-114E/index.html)
[12/06-114E/index.html (accessed on 27 April 2021).](https://www.sony.net/SonyInfo/News/Press/200612/06-114E/index.html)
16. Kalnins, A. An Empirical Analysis of Territorial Encroachment Within Franchised and Company-Owned Branded Chains.
_[Mark. Sci. 2004, 23, 476–489. [CrossRef]](http://doi.org/10.1287/mksc.1040.0082)_
17. Vinhas, A.S.; Anderson, E. How Potential Conflict Drives Channel Structure: Concurrent (Direct and Indirect) Channels.
_[J. Mark. Res. 2005, 42, 507–515. [CrossRef]](http://doi.org/10.1509/jmkr.2005.42.4.507)_
18. Arya, A.; Mittendorf, B.; Yoon, D.-H. Friction in Related-Party Trade When a Rival Is Also a Customer. Manag. Sci. 2008, 54,
[1850–1860. [CrossRef]](http://doi.org/10.1287/mnsc.1080.0906)
19. Patra, P. Distribution of Profit in a Smart Phone Supply Chain under Green Sensitive Consumer Demand. J. Clean. Prod. 2018, 192,
[608–620. [CrossRef]](http://doi.org/10.1016/j.jclepro.2018.04.144)
20. Liu, Z.L.; Anderson, T.D.; Cruz, J.M. Consumer Environmental Awareness and Competition in Two-Stage Supply Chains.
_[Eur. J. Oper. Res. 2012, 218, 602–613. [CrossRef]](http://doi.org/10.1016/j.ejor.2011.11.027)_
21. Green, K.W.G., Jr.; Zelbst, P.J.; Meacham, J.; Bhadauria, V.S. Green Supply Chain Management Practices: Impact on Performance.
_[Supply Chain Manag. Int. J. 2012, 17, 290–305. [CrossRef]](http://doi.org/10.1108/13598541211227126)_
22. Ghosh, D.; Shah, J. A Comparative Analysis of Greening Policies Across Supply Chain Structures. Int. J. Prod. Econ. 2012, 135,
[568–583. [CrossRef]](http://doi.org/10.1016/j.ijpe.2011.05.027)
23. He, J.; Lei, Y.; Fu, X. Do Consumer’s Green Preference and the Reference Price Effect Improve Green Innovation? A Theoretical
[Model Using the Food Supply Chain as a Case. Int. J. Environ. Res. Public Health 2019, 16, 5007. [CrossRef] [PubMed]](http://doi.org/10.3390/ijerph16245007)
24. Liu, X.; Du, W.; Sun, Y. Green Supply Chain Decisions Under Different Power Structures: Wholesale Price vs. Revenue Sharing
[Contract. Int. J. Environ. Res. Public Health 2020, 17, 7737. [CrossRef] [PubMed]](http://doi.org/10.3390/ijerph17217737)
25. Lee, D. Who Drives Green Innovation? A Game Theoretical Analysis of a Closed-Loop Supply Chain under Different Power
[Structures. Int. J. Environ. Res. Public Health 2020, 17, 2274. [CrossRef]](http://doi.org/10.3390/ijerph17072274)
26. [Cai, G. Channel Selection and Coordination in Dual-Channel Supply Chains. J. Retail. 2010, 86, 22–36. [CrossRef]](http://doi.org/10.1016/j.jretai.2009.11.002)
27. Heydari, J.; Govindan, K.; Aslani, A. Pricing and Greening Decisions in a Three-Tier Dual Channel Supply Chain. Int. J. Prod. Econ.
**[2019, 217, 185–196. [CrossRef]](http://doi.org/10.1016/j.ijpe.2018.11.012)**
28. Fein, A.J.; Anderson, E. Patterns of Credible Commitments: Territory and Brand Selectivity in Industrial Distribution Channels.
_[J. Mark. 1997, 61, 19–34. [CrossRef]](http://doi.org/10.1177/002224299706100202)_
-----
_Processes 2021, 9, 990_ 24 of 24
29. Tsay, A.A.; Agrawal, N. Channel Conflict and Coordination in the E-Commerce Age. Production and Operations Management.
_[Prod. Oper. Manag. 2009, 13, 93–110. [CrossRef]](http://doi.org/10.1111/j.1937-5956.2004.tb00147.x)_
30. [Arya, A.; Mittendorf, B.; Sappington, D.E.M. The Bright Side of Supplier Encroachment. Mark. Sci. 2007, 26, 651–659. [CrossRef]](http://doi.org/10.1287/mksc.1070.0280)
31. [Li, Z.; Gilbert, S.M.; Lai, G. Supplier Encroachment Under Asymmetric Information. Manag. Sci. 2014, 60, 449–462. [CrossRef]](http://doi.org/10.1287/mnsc.2013.1780)
32. [Ha, A.; Long, X.; Nasiry, J. Quality in Supply Chain Encroachment. Manuf. Serv. Oper. Manag. 2015, 18, 280–298. [CrossRef]](http://doi.org/10.1287/msom.2015.0562)
33. Cai, G.G.; Zhang, Z.G.; Zhang, M. Game Theoretical Perspectives on Dual Channel Supply Chain Competition with Price
[Dis-Counts and Pricing Schemes. Int. J. Prod. Econ. 2009, 117, 80–96. [CrossRef]](http://doi.org/10.1016/j.ijpe.2008.08.053)
34. Chen, J.; Zhang, H.; Sun, Y. Implementing Coordination Contracts in a Manufacturer Stackelberg Dual-Channel Supply Chain.
_[Omega 2012, 40, 571–583. [CrossRef]](http://doi.org/10.1016/j.omega.2011.11.005)_
35. [Swami, S.; Shah, J. Channel Coordination in Green Supply Chain Management. J. Oper. Res. Soc. 2013, 64, f336–f351. [CrossRef]](http://doi.org/10.1057/jors.2012.44)
36. Li, W.N.; Elsadany, A.A.; Zhou, W.; Zhu, Y.L. Global Analysis, Multi-Stability and Synchronization in a Competition Model of
[Public Enterprises with Consumer Surplus. Chaos Solitons Fractals 2021, 143, 110604. [CrossRef]](http://doi.org/10.1016/j.chaos.2020.110604)
37. [Pal, R.; Saha, B. Pollution Tax, Partial Privatization and Environment. Resour. Energy Econ. 2015, 40, 19–35. [CrossRef]](http://doi.org/10.1016/j.reseneeco.2015.01.004)
38. Bian, J.; Zhao, X. Tax or Subsidy? An analysis of Environmental Policies in Supply Chains with Retail Competition.
_[Eur. J. Oper. Res. 2020, 283, 901–914. [CrossRef]](http://doi.org/10.1016/j.ejor.2019.11.052)_
39. Spence, M. Product Differentiation and Welfare. Am. Econ. Rev. 1976, 66, 407–414.
40. Krass, D.; Nedorezov, T.; Ovchinnikov, A. Environmental Taxes and the Choice of Green Technology. Prod. Oper. Manag. 2013, 22,
[1035–1055. [CrossRef]](http://doi.org/10.1111/poms.12023)
41. Poyago-Theotoky, J. The Organization of R&D and Environmental Policy. J. Econ. Behav. Organ. 2007, 62, 63–75.
42. Ouchida, Y.; Goto, D. Do Emission Subsidies Reduce Emission? In the Context of Environmental R&D Organization. Econ. Model.
**2014, 36, 511–516.**
43. Bian, J.; Guo, X.; Li, K.W. Decentralization or Integration: Distribution Channel Selection under Environmental Taxation.
_[Transp. Res. Part E Logist. Transp. Rev. 2018, 113, 170–193. [CrossRef]](http://doi.org/10.1016/j.tre.2017.09.011)_
44. Zhang, J.; Li, S.; Zhang, S.; Dai, R. Manufacturer Encroachment with Quality Decision under Asymmetric Demand Information.
_[Eur. J. Oper. Res. 2019, 273, 217–236. [CrossRef]](http://doi.org/10.1016/j.ejor.2018.08.002)_
45. Chiang, W.K.; Chhajed, D.; Hess, J.D. Direct Marketing, Indirect Profits: A Strategic Analysis of Dual-Channel Supply-Chain
[Design. Manag. Sci. 2003, 49, 1–20. [CrossRef]](http://doi.org/10.1287/mnsc.49.1.1.12749)
46. Sun, X.; Tang, W.; Chen, J.; Li, S.; Zhang, J. Manufacturer Encroachment with Production Cost Reduction under Asymmetric
[Information. Transp. Res. Part E Logist. Transp. Rev. 2019, 128, 191–211. [CrossRef]](http://doi.org/10.1016/j.tre.2019.05.018)
47. Li, J.; Liang, J.; Shi, V.; Wang, Q. The Benefit of Manufacturer Encroachment Considering Consumer’s Environmental Awareness
and Product Competition. Ann. Oper. Res. 2021, 3, 1–19.
48. Li, J.; Wang, F.; He, Y. Electric Vehicle Routing Problem with Battery Swapping Considering Energy Consumption and Carbon
[Emissions. Sustainability 2020, 12, 10537. [CrossRef]](http://doi.org/10.3390/su122410537)
49. Zhang, C.; Su, W.; Zeng, S.; Balezentis, T.; Herrera-Viedma, E. A Two-Stage Subgroup Decision-Making Method for Processing
[Large-Scale Information. Expert Syst. Appl. 2021, 171, 114586. [CrossRef]](http://doi.org/10.1016/j.eswa.2021.114586)
50. Li, J.; Yi, L.; Shi, V.; Chen, X. Supplier Encroachment Strategy in the Presence of Retail Strategic Inventory: Centralization or
[Decentralization? Omega 2021, 98, 102213. [CrossRef]](http://doi.org/10.1016/j.omega.2020.102213)
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3390/PR9060990?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/PR9060990, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "GOLD",
"url": "https://www.mdpi.com/2227-9717/9/6/990/pdf?version=1622787550"
}
| 2,021
|
[] | true
| 2021-06-03T00:00:00
|
[
{
"paperId": "2328b57078cd120ae0952b2a79358a231ac81f63",
"title": "The benefit of manufacturer encroachment considering consumer's environmental awareness and product competition"
},
{
"paperId": "697d5b1b14ff52f2fbc1266207d12bbff06fb3a5",
"title": "Global Analysis, Multi-stability and Synchronization in a Competition Model of Public Enterprises with Consumer Surplus"
},
{
"paperId": "1e872ff46f1fc62612f68804c6fea6c821ae4477",
"title": "Electric Vehicle Routing Problem with Battery Swapping Considering Energy Consumption and Carbon Emissions"
},
{
"paperId": "cec3687b05da9e0c4887c07546999e0c906541b9",
"title": "Green Supply Chain Decisions Under Different Power Structures: Wholesale Price vs. Revenue Sharing Contract"
},
{
"paperId": "55323624dc86d421d26fd240328d4769dda4b2fe",
"title": "Tax or subsidy? An analysis of environmental policies in supply chains with retail competition"
},
{
"paperId": "9f5334c853d025fa71c7f04fb331e79ee083d847",
"title": "Who Drives Green Innovation? A Game Theoretical Analysis of a Closed-Loop Supply Chain under Different Power Structures"
},
{
"paperId": "562d9341a20626203ea3b8afdc528387eb16b103",
"title": "Supplier encroachment strategy in the presence of retail strategic inventory: Centralization or decentralization?"
},
{
"paperId": "ff7b2373c8a96f4e84e3bb422486e9ca76d60eab",
"title": "Do Consumer’s Green Preference and the Reference Price Effect Improve Green Innovation? A Theoretical Model Using the Food Supply Chain as a Case"
},
{
"paperId": "d11c4bbe06e7c4264be4c71180254a85b2942363",
"title": "Pricing and greening decisions in a three-tier dual channel supply chain"
},
{
"paperId": "bad44789619df2346a1661008e11f89c4b18bfd4",
"title": "Manufacturer encroachment with production cost reduction under asymmetric information"
},
{
"paperId": "af68cedfaae647a95e395960dd491680b31b40be",
"title": "Managing sustainability orientation: An empirical investigation of manufacturing firms"
},
{
"paperId": "e3eeeca6af78d9f50427868192c51f189ed16fc7",
"title": "Pricing and coordination strategies of a dual-channel supply chain considering green quality and sales effort"
},
{
"paperId": "b4d84193bf2dcee743b7dee208839d4677b19608",
"title": "Manufacturer encroachment with quality decision under asymmetric demand information"
},
{
"paperId": "374c856e57c2813fd31ed8eef2cab8a0c7dba387",
"title": "Distribution of profit in a smart phone supply chain under Green sensitive consumer demand"
},
{
"paperId": "408cdd8b64c365401d4cdd6ce9018e3e9a2d2e69",
"title": "Green product supply chain contracts considering environmental responsibilities"
},
{
"paperId": "4179435871c7b97d5aa55d08705771e0655c4aae",
"title": "Decentralization or integration: Distribution channel selection under environmental taxation"
},
{
"paperId": "4d00dd9b5d236a3d2027d035ea23ff641f7eb967",
"title": "A mathematical model for green supply chain coordination with substitutable products"
},
{
"paperId": "4a8d537bd58dbe6e9b6f61d841acf5c3a4b86d15",
"title": "Supplier Encroachment and Investment Spillovers"
},
{
"paperId": "36e7f367ad73e3e07cb541d2ac1050d119341348",
"title": "Potential gains from carbon emissions trading in China: A DEA based estimation on abatement cost savings"
},
{
"paperId": "c6a30ed4b957f678d8945133efa37925192be71e",
"title": "Pricing strategy and coordination in a dual channel supply chain with a risk-averse retailer"
},
{
"paperId": "2062063fd5e93e03728417cf74dd06fb869fe6e4",
"title": "Quality in Supply Chain Encroachment"
},
{
"paperId": "68b468872c453007dc54da5b808d48220cb125aa",
"title": "Supplier encroachment in competitive supply chains"
},
{
"paperId": "435fd655139ea20bb0465a1ecdd7472930a2d7d4",
"title": "Pollution tax, partial privatization and environment"
},
{
"paperId": "44773d46e2e2476f0a16b5e1c847aa335a975b80",
"title": "The role of collaboration in the UK green supply chains: an exploratory study of the perspectives of suppliers, logistics and retailers"
},
{
"paperId": "9fc1c3f111afb58f53f5f5f9cc24a41d8cd40a63",
"title": "Environmental Taxes and the Choice of Green Technology"
},
{
"paperId": "b61386529f4b92796bd56c4d397af0332a5c9192",
"title": "Channel coordination in green supply chain management"
},
{
"paperId": "db5d7918993d624eee5424c743c09a75d5e011a9",
"title": "Implementing coordination contracts in a manufacturer Stackelberg dual-channel supply chain"
},
{
"paperId": "500c00519cd6429b0ab241d437544b74745797aa",
"title": "Consumer environmental awareness and competition in two-stage supply chains"
},
{
"paperId": "5fd91d42c57e11c6a87f17d61291146e87c239d3",
"title": "Green supply chain management practices: impact on performance"
},
{
"paperId": "e428991eac03f67729f8fa613df80693958ce18d",
"title": "Supplier Encroachment under Asymmetric Information"
},
{
"paperId": "17b8835b6849aea0c6117eeba5d272d45e8006ed",
"title": "A comparative analysis of greening policies across supply chain structures"
},
{
"paperId": "9a9eb26e86328afd7c9d5e3285bb70f1714223d3",
"title": "Channel Selection and Coordination in Dual-Channel Supply Chains"
},
{
"paperId": "d07a4201f6f57560ef6f8e46899a63d42c370f9a",
"title": "Friction in Related-Party Trade When a Rival Is Also a Customer"
},
{
"paperId": "b86c6dcc9b6a211dc7cfacb102ff8cd61a11dab9",
"title": "Sustainable supply chains: An introduction"
},
{
"paperId": "764a46f67285c2bda64f1e86c0b6dcf579700d50",
"title": "The Bright Side of Supplier Encroachment"
},
{
"paperId": "d221bb2443c9d92f23f7938a1ed4e167b126b5a5",
"title": "How Potential Conflict Drives Channel Structure: Concurrent (Direct and Indirect) Channels"
},
{
"paperId": "a55066d3dd052c90f762378484cc6c81668503c0",
"title": "An Empirical Analysis of Territorial Encroachment Within Franchised and Company-Owned Branded Chains"
},
{
"paperId": "f7a52ca6bff0a790955828d415b701ca55baf0fb",
"title": "Channel Conflict and Coordination in the E‐Commerce Age"
},
{
"paperId": "035438e4dea47bb61cdf1285ed153013167bdf66",
"title": "New Tools for Environmental Protection: Education, Information and Voluntary Measures"
},
{
"paperId": "368163bc8d31fbfe3ab258c495b567c3d5ad6ed2",
"title": "Price Competition and Product Differentiation When Consumers Care for the Environment"
},
{
"paperId": "0a649f841e000b45fbdb5ade06b04d7872a6af24",
"title": "A STRATEGIC DECISION FRAMEWORK FOR GREEN SUPPLY CHAIN MANAGEMENT"
},
{
"paperId": "9b1ff8cb50ebd89d1fa394da441636b952b3f3b4",
"title": "Environmental Product Differentiation: Implications for Corporate Strategy"
},
{
"paperId": "13b4322bbb2d804e119dcdf7e08513b8e2ec4cf2",
"title": "Patterns of Credible Commitments: Territory and Brand Selectivity in Industrial Distribution Channels"
},
{
"paperId": "ab147883440385cbd5cb5062b515b98e6b777203",
"title": "A Two-stage subgroup Decision-making method for processing Large-scale information"
},
{
"paperId": "2aa7d84d4ab04ffe84c800ba8ebcef3ba562111c",
"title": "Do emission subsidies reduce emission? In the context of environmental R&D organization"
},
{
"paperId": "cd140febf18fb013c6979d02264c0a56fc864190",
"title": "Game theoretical perspectives on dual-channel supply chain competition with price discounts and pricing schemes"
},
{
"paperId": "de6c2ac9e27039effd09ed4a95824e0295ee7d2c",
"title": "The organization of R&D and environmental policy"
},
{
"paperId": "c0d90a6b6b2e9701cc9900daf0a2dcfa9c6ed40b",
"title": "Direct Marketing, Indirect Profits: A Strategic Analysis of Dual - Channel Supply - Chain Design"
},
{
"paperId": "0453ea6c2dbf048112d6b216b66cf437be6fc568",
"title": "Design, planning, scheduling, and control problems of flexible manufacturing systems"
},
{
"paperId": "6bb149a41f6133ceca1d9983eca918cd2abead4a",
"title": "Product Differentiation and Welfare"
},
{
"paperId": null,
"title": "Sony Sells a Portion of Its Shares in StylingLife Holdings"
},
{
"paperId": null,
"title": "Do Emission Subsidies Reduce Emission ?"
},
{
"paperId": null,
"title": "Supply Chain Management Practices: Impact on Performance. Supply Chain Manag"
},
{
"paperId": null,
"title": "Indirect Profits: A Strategic Analysis of Dual-Channel Supply-Chain Design"
},
{
"paperId": null,
"title": "Eco-labeling is one among a number of policy tools that are used in what"
}
] | 27,675
|
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Engineering",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/0265955b9565abd870ab9d874305a84ef397c743
|
[
"Computer Science"
] | 0.819012
|
Fast Multi-precision Multiplication for Public-Key Cryptography on Embedded Microprocessors
|
0265955b9565abd870ab9d874305a84ef397c743
|
Journal of Cryptology
|
[
{
"authorId": "70931341",
"name": "M. Hutter"
},
{
"authorId": "1764501",
"name": "Erich Wenger"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"J Cryptol"
],
"alternate_urls": [
"https://www.iacr.org/jofc/jofc.html",
"http://www.springeronline.com/sgw/cda/frontpage/0,11855,5-0-70-1009426-detailsPage=journal|description|description,00.html?referer=www.springeronline.com/journal/00145/about"
],
"id": "de5467ac-3f75-47f8-8397-1c10f6f9fc09",
"issn": "0933-2790",
"name": "Journal of Cryptology",
"type": "journal",
"url": "https://link.springer.com/journal/145"
}
| null |
# Fast Multi-precision Multiplication for
Public-Key Cryptography on Embedded
Microprocessors
Michael Hutter and Erich Wenger
Institute for Applied Information Processing and Communications (IAIK),
Graz University of Technology, Inffeldgasse 16a, 8010 Graz, Austria
_{Michael.Hutter,Erich.Wenger}@iaik.tugraz.at_
**Abstract. Multi-precision multiplication is one of the most fundamen-**
tal operations on microprocessors to allow public-key cryptography such
as RSA and Elliptic Curve Cryptography (ECC). In this paper, we
present a novel multiplication technique that increases the performance
of multiplication by sophisticated caching of operands. Our method significantly reduces the number of needed load instructions which is usually
one of the most expensive operation on modern processors. We evaluate
our new technique on an 8-bit ATmega128 microcontroller and compare
the result with existing solutions. Our implementation needs only 2, 395
clock cycles for a 160-bit multiplication which outperforms related work
by a factor of 10 % to 23 %. The number of required load instructions is
reduced from 167 (needed for the best known hybrid multiplication) to
only 80. Our implementation scales very well even for larger Integer sizes
(required for RSA) and limited register sets. It further fully complies to
existing multiply-accumulate instructions that are integrated in most of
the available processors.
**Keywords: Multi-precision Arithmetic, Microprocessors, Elliptic Curve**
Cryptography, RSA, Embedded Devices.
## 1 Introduction
Multiplication is one of the most important arithmetic operation in public-key
cryptography. It engross most of the resources and execution time of modern
microprocessors (up to 80 % for Elliptic Curve Cryptography (ECC) and RSA
implementations [6]). In order to increase the performance of multiplication, most
effort has been put by researchers and developers to reduce the number of instructions or minimize the amount of memory-access operations.
Common multiplication methods are the schoolbook or Comba [4] technique
which are widely used in practice. They require at least 2n[2] _load instructions_
to process all operands and to calculate the necessary partial products. In 2004,
Gura et al. [6] presented a new method that combines the advantages of these
methods (hybrid multiplication). They reduced the number of load instructions
-----
to only 2 _n[2]/d_ where the parameter d depends on the number of available reg_⌈_ _⌉_
isters of the underlying architecture. They reported a performance gain of about
25 % compared to the classical Comba multiplication. Their 160-bit implementation needs 3,106 clock cycles on an 8-bit ATmega128 microcontroller. Since
then, several authors applied this method [7,12,14,15,17] and proposed various
enhancements to further improve the performance. Most of the related work
reported between 2,593 and 2,881 clock cycles on the same platform.
In this paper, we present a novel multiplication technique that reduces the
number of needed load instructions to only 2n[2]/e where e > d. We propose a
new way to process the operands which allows efficiently caching of required
operands. In order to evaluate the performance, we use the ATmega128 microcontroller and compare the results with related work. For a 160-bit multiplication, 2,395 clock cycles are necessary which is an improvement by a factor of 10 %
compared to the best reported implementation of Scott et al. [14] (which need
2,651 clock cycles) and by a factor of about 23 % compared to the work of Gura
et al. [6]. We further compare our solution with different Integer sizes (160, 192,
256, 512, 1,024, and 2,048) and register sizes (e = 2, 4, 8, 10, and 20). It shows
that our solution needs about 15 % less clock cycles for any chosen Integer size.
Our solution also scales very well for different register sizes without significant
loss of performance. Besides this, the method fully complies with common architectures that support multiply-accumulate instructions using a (Comba-like)
triple-register accumulator.
The paper is organized as follows. In Section 2, we describe related work on
that topic and give performance numbers for different multiplication techniques.
Section 3 describes different multi-precision multiplication techniques used in
practice. We describe the operand scanning, product scanning, and the hybrid
method and compare them with our solution. In Section 4, we present the results
of our evaluations. We describe the ATmega128 architecture and give details
about the implementation. Summary and conclusions are given in Section 5.
## 2 Related Work
In this section, we describe related work on multi-precision multiplication over
prime fields. Most of the work given in literature make use of the hybridmultiplication technique [6] which provides best performance on most microprocessors. This technique was first presented at CHES 2004 where the authors
reported a speed improvement of up to 25 % compared to the classical Combamultiplication technique [4] on 8-bit platforms. Their implementation requires
3,106 clock cycles for a 160-bit multiplication on an ATmega128 [1]. Several
authors adopted the idea and applied the method for different devices and environments, e.g. sensor nodes. Wang et al. [18] and Ugus et al. [16] made use
of this technique and implemented it on the MICAz motes which feature an
ATmega128 microcontroller. Results for the same platform have been also reported by Liu et al. [11] and Szczechowiak et al. [15] in 2008 who provide software
libraries (TinyECC and NanoECC) for various sensor-mote platforms. One of
the first who improved the implementation of Gura has been due to Uhsadel et
-----
al. [17]. They have been able to reduce the number of needed clock cycles to only
2,881. Further improvements have been also reported by Scott et al. [14]. They
introduced additional registers (so-called carry catchers) and could increase the
performance to 2,651 clock cycles. Note that they fully unrolled the execution
sequence to avoid additional clock cycles for loop instructions. Similar results
have been also obtained by Kargl et al. [7] in 2008 which reported 2,593 clock
cycles for an un-rolled 160-bit multiplication on the ATmega128.
In 2009, Lederer et al. [9] showed that the needed number of addition and
move instructions can be reduced by simply rearranging the instructions during
execution of the hybrid-multiplication method. Similar findings have been also
reported recently by Liu et al. [12] who reported the fastest looped version of the
hybrid multiplication needing 2,865 clock cycles in total.
## 3 Multi-precision Multiplication Techniques
In the following subsections, we describe common multiplication techniques that
are often used in practice. We describe the operand scanning, product scanning,
and hybrid multiplication method[1]. The methods differ in several ways how to
process the operands and how many load and store instructions are necessary to
perform the calculation. Most of these methods lack in the fact that they load
the same operands not only once but several times throughout the algorithm
which results in additional and unnecessary clock cycles. We present a new multiplication technique that improves existing solutions by efficiently reducing the
_load instructions through sophisticated caching of operands._
Throughout the paper, we use the following notation. Let a and b be two
_m-bit large Integers that can be written as multiple-word array structures A =_
(A[n 1], . . ., A[2], A[1], A[0]) and B = (B[n 1], . . ., B[2], B[1], B[0]). Further let
_−_ _−_
_W be the word size of the processor (e.g. 8, 16, 32, or 64 bits) and n =_ _m/W_
_⌈_ _⌉_
the number of needed words to represent the Integers a or b. We denote the
result of the multiplication by c = ab and represent it in a double-size word
array C = (C[2n 1], . . ., C[2], C[1], C[0]).
_−_
**3.1** **Operand-Scanning Method**
Among the most simplest way to perform large Integer multiplication is the
operand-scanning method (or often referred as schoolbook or row-wise multiplication method). The multiplication can be implemented using two nested loop
operations. The outer loop loads the operand A[i] at index i = 0 . . . n 1 and
_−_
keeps the value constant inside the inner loop of the algorithm. Within the inner loop, the multiplicand B[j] is loaded word by word and multiplied with the
operand A[i]. The partial product is then added to the intermediate result of the
same column which is usually buffered in a register or stored in data memory.
1 Note that we do not consider multiplications methods such as Karatsuba-Ofman or
FFT in this paper since they are considered to require more resources and memory
accesses on common microcontrollers than the given methods [8].
-----
C[14] C[7] C[0]
A[7]B[7]
A[0]B[0]
A[0]B[7]
**Fig. 1. Operand-scanning multiplication of 8-word large Integers a and b**
Figure 1 shows the structure of the algorithm on the left side. The individual
row levels can be clearly discerned. On the right side of the figure, all n[2] partial
products are displayed in form of a rhombus. Each point in the rhombus represents a multiplication A[i] _B[j]. The most right-sided corner of the rhombus_
_×_
starts with the lowest indices i, j = 0 and the most left-sided corner ends with
the highest indices i, j = n 1. By following all multiplications from the right
_−_
to the lower-mid corner of the rhombus, it can be observed that the operand
_A[i] keeps constant for any index i_ [0, n). The same holds true for the operand
_∈_
_B[j] and j_ [0, n) by following all multiplications from right to the upper-mid
_∈_
corner of the rhombus. Note that this is also valid for the left-handed side of the
rhombus.
For the operand-scanning method, it can be seen that the partial products
are calculated from the upper-right side to the lower-left side of the rhombus (we
marked the processing of the partial products with a black arrow). In each row,
_n multiplications have to be performed. Furthermore, 2n load operations and n_
_store operations are required to load the multiplicand and the intermediate result_
_C[i + j] and to store the result C[i + j]_ _C[i + j] + A[i]_ _B[j]. Thus, 3n[2]_ + 2n
_←_ _×_
memory operations are necessary for the entire multi-precision multiplication.
Note that this number decreases to n[2] + 3n for architectures that can maintain
the intermediate result in available working registers.
**3.2** **Product-Scanning Method**
Another way to perform a multi-precision multiplication is the product-scanning
method (also referred as Comba [4] or column-wise multiplication method). There,
each partial product is processed in a column-wise approach. This has several
advantages. First, since all operands of each column are multiplied and added
consecutively (within a multiply-accumulate approach), a final word of the result
is obtained for each column. Thus, no intermediate results have to be stored or
loaded throughout the algorithm. In addition, the handling of carry propagation
-----
C[14] C[7] C[0]
A[7]B[7]
A[0]B[0]
A[0]B[7]
**Fig. 2. Product-scanning multiplication of 8-word large Integers a and b**
is very easy because the carry can be simply added to the result of the next
column using a simple register-copy operation. Second, only five working registers are needed to perform the multiplication: two registers for the operand and
multiplicand and three registers for accumulation[2]. This makes the method very
suitable for low-resource devices with limited registers.
Figure 2 shows the structure of the product-scanning method. By having a
look at the rhombus, it shows that by processing the partial products in a
column-wise instead of a row-wise approach, only one store operation is needed
to store the final word of the result. For the entire multi-precision operation, 2n[2]
_load operations are necessary to load the operands A[i] and B[j] and 2n store_
operations are needed to store the result. Therefore, 2n[2] +2n memory operations
are needed.
**3.3** **Hybrid Method**
The hybrid multiplication method [6] combines the advantages of the operandscanning and product-scanning method. It can be implemented using two nested
loop structures where the outer loop follows a product-scanning approach and the
inner loop performs a multiplication according to the operand-scanning method.
The main idea is to minimize the number of load instructions within the inner
loop. For this, the accumulator has to be increased to a size of 2d + 1 registers.
The parameter d defines the number of rows within a processed block. Note that
the hybrid multiplication is equals to the product-scanning method if parameter
_d is chosen as d = 1 and it is equal to the operand-scanning method if d = n._
Figure 3 shows the structure of the hybrid multiplication for d = 4. It shows
that the partial products are processed in form of individual blocks (we marked
the processing sequence of the blocks from 1 to 4). Within one block, all operands
are processed row by row according to the operand-scanning approach. Note that
2 We assume the allocation of three registers for the accumulator register whereas 2 +
_⌈log2(n)/W ⌉_ registers are actually needed to maintain the sum of partial products.
-----
C[14] C[7] C[0]
A[7]B[7]
A[0]B[0]
A[0]B[7]
**Fig. 3. Hybrid multiplication of 8-word large Integers a and b (d = 4)**
these blocks use operands with a very limited range of indices. Thus, several load
instructions can be saved in cases where enough working registers are available.
However, the outer loop of the hybrid method processes the blocks in a columnwise approach. So between two consecutive blocks no operands can be shared
and all operands have to be loaded from memory again. This becomes clear by
having a look at the processing of Block 1-3. Block 2 and 3 do not share any
operands that possess the same indices. Therefore, all operands that have already
been loaded for Block 1 and that can be reused in Block 3 have to be loaded
again after processing of Block 2 which requires additional and unnecessary load
instructions. However, in total, the hybrid method needs 2 _n[2]/d_ + 2n memory_⌈_ _⌉_
access instructions which provides good performances on devices that feature a
large register set.
**3.4** **Operand-Caching Method**
We present a new method to perform multi-precision multiplication. The main
idea is to reduce the number of memory accesses to a minimum by efficiently
caching of operands. We show that by spending a certain amount of store operations, a significant amount of load instructions can be saved by reusing operands
that have been already loaded in working registers.
The method basically follows the product-scanning approach but divides the
calculation into several rows. In fact, the product-scanning method provides best
performance if all needed operands can be maintained in working registers. In
such a case, only 2n load instructions and 2n store instructions would be necessary. However, the product-scanning method becomes inefficient if not enough
registers are available or if the Integer size is too large to cache a significant
amount of operands. Hence, several load instructions are necessary to reload
and overwrite the operands in registers.
In the light of this fact, we propose to separate the product-scanning method
into individual rows r = _n/e_ . The size e of each row is chosen in a way that all
_⌊_ _⌋_
-----
C[14] C[7] C[0]
1
2
3
4
1
2
3
4
binit
r0
r1
A[7]B[7]
A[0]B[0]
A[0]B[7]
**Fig. 4. Operand-caching multiplication of 8-word large Integers a and b (e = 3)**
needed words of one operand can be cached in the available working registers.
Figure 4 shows the structure of the proposed method for parameter e = 3. That
means, 3 registers are reserved to store 3 words of operand a and 3 registers
are reserved to store 3 words of operand b. Thus, we assume f = 2e + 3 = 9
available registers including a triple-word accumulator. The calculation is now
separated into r = ⌊8/3⌋ = 2 rows, i.e. r0 and r1, and consists of one remaining
block which we further denote as initialization block binit. This block calculates
the partial products which are not processed by the rows.
All rows are further separated into four parts. Part 1 and 4 use the classical product-scanning approach. Part 2 and 3 perform an efficient multiplyaccumulate operation of already cached operands.
The algorithm starts with the calculation of binit and processes the individual
rows afterwards (starting from the the smallest to the largest row, i.e. from
the top to the bottom of the rhombus). Furthermore, all partial products are
generated from right to left. In the following, we describe the algorithm in a
more detail.
**Initialization Block binit. This block (located in the upper-mid of the rhom-**
bus) performs the multiplication according to the classical product-scanning
method. The Integer size of the binit multiplication is (n − _re), i.e. 8 −_ 6 = 2
in our example, which is by definition smaller than e. Because of that, all
operands can be loaded and maintained within the available registers resulting in only 4(n _re) memory-access operations. Note that the calculation_
_−_
of binit is only required if there exist remaining partial products, i.e. n mod
_e ̸= 0. If n mod e = 0, the calculation of binit is skipped. Furthermore,_
consider the special case when n < e where only binit has to be performed
skipping the processing of rows (trivial case).
**Processing of Rows. In the following, we describe the processing of each row**
_p = r_ 1 . . . 0. Each row consists of four parts.
_−_
**Part 1. This part starts with a product-scanning multiplication. All operands**
for that row are first loaded into registers, i.e. A[i] with i = pe . . . e(p + 1) 1
_−_
-----
|3|Col2|
|---|---|
|Col1|A[3] x B[5] … … A[1] x B[7]|Col3|
|---|---|---|
|||C[8]|
|A[4] x B[5] … … A[2] x B[7]|||
||C[9]||
|A[2] x B[1] 2 … … A[0] x B[3] + C[3]|A[2] x B[1] … … A[0] x B[3]|Col3|
|---|---|---|
|||C[3]|
|A[2] x B[2] … … A[0] x B[4]|||
|+ C[4]|C[4]||
+ C[12]
3 2
A[1] x B[7]
+ C[8]
A[4] x B[5]
A[2] x B[7]
+ C[9]
A[2] x B[5]
A[0] x B[7]
+ C[7]
|Col1|+|Col3|Col4|
|---|---|---|---|
|ACC2|ACC1|ACC0||
|||||
|A[7] x B[5] … … A[5] x B[7]|Col2|
|---|---|
|+|C[12]|
**Fig. 5. Processing of Part 2 and 3 of the row r1**
and B[j] with j = 0 . . . e 1. The sum of all partial products A[i] _B[j] is_
_−_ _×_
then stored as intermediate result to the memory location C[i] (same index
range as A[i]). Therefore, 2e load instructions and e store instructions are
needed.
**Part 2. The second part, processes n** _e(p + 1) columns using a multiply-_
_−_
accumulate approach. Since all operands of A[i] were already loaded and
used in Part 1, only one word B[j] has to be loaded from one column to
the next. The operands A[i] are kept constant throughout the processing of
Part 2. Next to the needed load instructions for B[j], we have to load and
update the intermediate result of Part 1 with the result obtained in Part 2.
Thus, 2(n _e(p + 1)) load and n_ _e(p + 1) store instructions are required_
_−_ _−_
for that part.
**Part 3. The third part performs the same operation as described in Part 2**
except that the already loaded operands B[j] are kept constant and that
one word A[i] is loaded for each column. Figure 5 shows the processing of
Part 2 and 3 of row r1 (p = 0). For each column, two load instructions are
necessary (marked in grey). All other operands have been loaded and cached
in previous parts. Operands which are not required for further processing
are overwritten by new operands, e.g. B[1] . . . B[4] in Part 2 of our example.
**Part 4. The last part calculates the remaining partial products. In contrast to**
Part 1, no load instructions are required since all operands have been already
loaded in Part 3. Hence, only e memory-access operations are needed to store
the remaining words of the (intermediate) result c.
Table 1 summaries the memory-access complexity of the initialization block and
the individual parts of a row p. By summing up all load instructions, we get
_r−1_
�
2(n _re) +_ (4n 4pe 2e) = 2n + 4rn 2er[2] 2er (1)
_−_ _−_ _−_ _−_ _−_ _≤_ [2][n][2]
_e [.]_
_p=0_
The total number of store operations can be evaluated by
2(n _re) +_
_−_
_r−1_
�
(2n 2pe) = 2n + 2rn _er[2]_ _er_ (2)
_−_ _−_ _−_ _≤_ _[n][2]_
_e_ [+][ n.]
_p=0_
-----
**Table 1. Memory-access complexity of binit and each part of row p = 0 . . . r −** 1
**Component** **Load Instr.** **Store Instr.** **Total**
_binit_ 2(n − _re)_ 2(n − _re)_ 4(n − _re)_
Part 1 2e _e_ 3e
Part 2 2(n − _e(p + 1))_ _n −_ _e(p + 1)_ 3(n − _e(p + 1))_
Part 3 2(n − _e(p + 1))_ _n −_ _e(p + 1)_ 3(n − _e(p + 1))_
Part 4 0 _e_ _e_
Table 2 lists the complexity of different multi-precision multiplication tech
niques. It shows that the hybrid method needs 2
_⌈_ _[n]d[2]_
_[⌉]_ _[load][ instructions whereas]_
the operand-caching technique needs about [2][n][2]
_e_ [. Since the total number of avail-]
able registers f equals to 2e + 3 for the operand-caching technique (2e registers
for the operand registers and three registers for the accumulator) and 3d + 2 for
the hybrid method (d + 1 registers for the operands and 2d + 1 registers for the
accumulator), we obtain
2e + 3 = 3d + 2 = _e = [3][d][ −]_ [1] and _e > d._ (3)
_⇒_
2
If we compare the total number of memory-access instructions for the hybrid
and the operand-caching method and express both runtimes using f, we get
+ 2n > [6][n][2] (4)
_f_ 3 [+][ n]
_−_
2
� 3n2
_f_ 2
_−_
�
Note that there are more parameters to consider. The number of additions of
the operand-caching method is 3n[2] and the number of additions of the hybrid
method is n[2](2 + d/2) (upper bound). Also the pseudocode of Gura et al. [6] for
the hybrid multiplication method is inefficient in the special case of n mod d = 0.
_̸_
**Table 2. Memory-access complexity of different multiplication techniques**
**Method** **Load** **Store** **Memory**
**Instructions** **Instructions** **Instructions**
Operand Scanning 2n[2] + n _n[2]_ + n 3n[2] + 2n
Product Scanning [4] 2n[2] 2n 2n[2] + 2n
Hybrid [6] 2⌈n[2]/d⌉ 2n 2⌈n[2]/d⌉ + 2n
**Operand Caching** **2n[2]/e** **_n[2]/e + n_** **3n[2]/e + n**
## 4 Results
We used the 8-bit ATmega128 microcontroller for evaluating the new multiplication technique. The ATmega128 is part of the megaAVR family from Atmel [1].
It has been widely used in embedded systems, automotive environments, and
-----
**Table 3. Unrolled instruction counts for a 160-bit multiplication on the ATmega128**
**Method** **Instruction** **Clock**
LD ST MUL ADD MOVW Others **Cycles**
Operand Scanning 820 440 400 1,600 2 464 5,427
Product Scanning 800 40 400 1,200 2 159 3,957
Hybrid (d=4) 200 40 400 1,250 202 109 2,904
**Operand Caching (e=10)** **80** **60** **400** **1,240** **2** **68** **2,395**
sensor-node applications. The ATmega128 is based on a RISC architecture and
provides 133 instructions [2]. The maximum operating frequency is 16 MHz. The
device features 128 kB of flash memory and 4 kB of internal SRAM. There exist 32 8-bit general-purpose registers (R0 to R31). Three 16-bit registers can
be used for memory addressing, i.e. R26:R27, R28:R29, and R30:R31 which
are denoted as X, Y, and Z. Note that the processor also allows pre-decrement
and post-increment functionalities that can be used for efficient addressing of
operands. The ATmega128 further provides an hardware multiplier that performs an 8 8-bit multiplication within two clock cycles. The 16-bit result is
_×_
stored in the registers R0 (lower word) and R1 (higher word).
We used register R25 to store a zero value. Furthermore, we reserved R23,
R24, and R25 as accumulator register. Thus, 20 registers, i.e. R2...R21, can be
used to store and cache the words of the operands (e = 10 registers for each
operand a and b). All implementations have been done by using a self-written
code generator that allows the generation of (looped & unrolled) assembly code.
In order to demonstrate the performance of our method, we implemented
all multiplication techniques described in Section 3. For comparison reasons, we
decided to implement a 160 160-bit multiplication as it has been done by
_×_
most of the related work. Note that for RSA and ECC, larger Integer sizes
are recommended in practice [10,13]. The Standards for Efficient Cryptography
(SEC) already removed the recommended secp160r1 elliptic curve from their
standard since SEC version 2 of 2010 [3].
Table 3 summarizes the instruction counts for the operand scanning, product
scanning, hybrid, and operand-caching implementation. The operand-scanning
and product-scanning methods have been implemented without using all the
available registers (as it usually would be implemented). For hybrid multiplication, we applied d = 4 because it allows a better optimization regarding necessary
addition operations compared to a multiplication with d = 5. The carry propagation problem has been solved by implementing a similar approach as proposed
by Liu et al. [12]. Thus, 200 MOVW instructions have been necessary to handle
the carry propagation accordingly. For a fair comparison, all methods have been
optimized for speed and provide unrolled instruction sequences. Furthermore,
we implemented all accumulators as ring buffers to reduce necessary MOV instructions. After each partial-product generation, the indices of the accumulator
registers are shifted so that no MOV instructions are necessary to copy the carry.
-----
**Table 4. Comparison of multiplication methods**
for different Integer sizes
**Size** **Op.** **Prod.** **Hybrid** **Operand**
**[bit]** **Scan.** **Scan.** **Method** **Caching**
160 5,427 3,957 2,904 2,395
192 7,759 5,613 4,144 3,469
256 13,671 9,789 7,284 6,123
512 53,959 38,013 28,644 24,317
1,024 214,407 149,757 113,604 96,933
2,048 854,791 594,429 452,484 387,195
10[6]
10[5]
10[4]
10[3]
160 256 512 1024 2048
Integer size
**Fig. 6. Comparison chart**
Best results have been obtained for the operand-caching technique. By trading
additional 20 store instructions, up to 120 load instructions could be saved when
we compare the result with the best reference values (hybrid implementation).
Note that load, store, and multiply instructions on the ATmega128 are more
expensive than other instructions since they require two clock cycles instead of
only one. For operand-caching multiplication, almost the same amount of load
and store instructions are required. In total 2,395 clock cycles are needed to
perform the multiplication. Compared to the hybrid implementation, a speed
improvement of about 18 % could be achieved.
We also compare the performance of the implemented multi-precision methods
for different Integer sizes. Table 4 shows the result for Integer sizes from 160 up
to 2,048 bits[3]. The operand-caching technique provides the best performance
for any Integer size. It is therefore well suited for large Integer sizes such as it
is in the case of RSA. In average, a speed improvement of about 15 % could
be achieved compared to the hybrid method. Figure 6 shows the appropriate
performance chart in a double logarithmic scale.
**Table 5. Performance of operand-caching multi-**
plication for different Integer sizes and available
registers
**Size** **_e=2_** **_e=4_** **_e=8_** **_e=10_** **_e=20_**
160 3,915 2,965 2,513 2,395 2,205
192 5,611 4,255 3,577 3,469 3,207
256 9,915 7,531 6,339 6,123 5,671
512 39,291 29,915 25,227 24,317 22,451
1,024 156,411 119,227 100,635 96,933 89,529
2,048 624,123 476,027 401,979 387,195 357,581
10[6]
10[5]
10[4]
10[3]
2 4 8 10 20
Available registers e
**Fig. 7. Performance chart**
3 Note that due to a fully unrolled implementation such large Integer multiplications
might be impractical due to the huge amount of code.
-----
**Table 6. Comparison with related work**
**Method** **Instruction** **Clock**
LD ST MUL ADD MOVW Others **Cycles**
**Hybrid**
Gura et al. [6] (d=5) 167 40 400 1,360 355 197 3,106
Uhsadel et al. [17] (d=5) 238 40 400 986 355 184 2,881
Scott et al. [14] (d=4)[a] 200 40 400 1,263 70 38 2,651
Liu et al. [12] (d=4) 200 40 400 1,194 212 179 2,865
**Operand Caching**
**with looping[a,c]** (e=9) **92** **66** **400** **1,252** **41** **276** **2,685**
**unrolled[b,c]** (e=10) **80** **60** **400** **1,240** **2** **68** **2,395**
_a binit, Part 1, and Part 4 unrolled. Part 2 and Part 3 looped._
_b Fully unrolled implementation without overhead of loop instructions._
_c w/o PUSH/POP/CALL/RET._
Table 5 and Figure 7 show the performance for different Integer sizes in re
lation to parameter e. The parameter e is defined by the number of available
registers to store words of one operand, i.e. e = _[f]_ _[−]2_ [3] [, where][ f][ = 2][e][ + 3 denotes]
the number of available registers in total (including the triple-size register for
the accumulator). It shows that for e > 10 no significant improvement in speed
is obtained. The performance decrease for smaller e and higher Integer sizes.
However, if we compare our solution (160-bit multiplication with smallest parameter e = 2 _f = 7 registers) with the product-scanning method (needing_
_→_
_f = 5 registers), we obtain 3,915 clock cycles for the operand-caching method_
and 3,957 clock cycles for the product scanning method. It therefore provides
a good performance even for a smaller set of available registers. For the special
case e = 20, where all 20 words of one 160-bit operand can be maintained in registers (ideal case for product scanning), it shows that the number of clock cycles
reaches nearly the optimum of 2,160 clock cycles, i.e. 4n = 80 memory-access
instructions, n[2] = 400 multiplications, and 3n[2] = 1, 200 additions.
We compare our result with related work in Table 6. For a fair comparison, we
also implemented a operand-caching version that does not unroll the algorithm
but includes additional loop instructions. It shows that the operand-caching
method provides best performance. Compared to Gura et al. [6] 23 % less clock
cycles are needed for a 160-bit multiplication. A 10 % improvement could be
achieved compared to the best solution reported in literature [14]. Note that
most of the related work need between 167 to 238 load instructions which mostly
explains the higher amount of needed clock cycles.
## 5 Conclusions
We presented a novel multiplication technique for embedded microprocessors.
The multiplication method reduces the number of necessary load instructions
-----
through sophisticated caching of operands. Our solution follows the productscanning approach but divides the processing into several parts. This allows
the scanning of sub-products where most of the operands are kept within the
register-set throughout the algorithm.
In order to evaluate our solution, we implemented several multiplication techniques using different Integer sizes on the ATmega128 microcontroller. Using
operand-caching multiplication, we require 2,395 clock cycles for a 160-bit multiplication. This result improves the best reported solution by a factor of 10 % [14].
Compared to the hybrid multiplication of Gura et al. [6], we achieved a speed
up of 23 %. Our evaluation further showed that our solution scales very well for
different Integer sizes used for ECC and RSA. We obtained an improvement of
about 15 % for bit sizes between 256 and 2,048 bits compared to a reference
implementation of the hybrid multiplication.
It is also worth to note that our multiplication method is perfectly suitable for
processors that support multiply-accumulate (MULACC) instructions such as ARM
or the dsPIC family of microcontrollers. It also fully complies to architectures
which support instruction-set extensions for MULACC operations such as proposed
by Großsch¨adl and Sava¸s [5].
**Acknowledgements. The work has been supported by the European Com-**
mission through the ICT program under contract ICT-2007-216646 (European
Network of Excellence in Cryptology - ECRYPT II) and under contract ICTSEC-2009-5-258754 (Tamper Resistant Sensor Node - TAMPRES).
## References
1. Atmel Corporation. 8-bit AVR Microcontroller with 128K Bytes In-System Pro[grammable Flash (August 2007), http://www.atmel.com/dyn/resources/prod_](http://www.atmel.com/dyn/resources/prod_documents/doc2467.pdf)
[documents/doc2467.pdf](http://www.atmel.com/dyn/resources/prod_documents/doc2467.pdf)
[2. Atmel Corporation. 8-bit AVR Instruction Set (May 2008), http://www.atmel.](http://www.atmel.com/dyn/resources/prod_documents/doc0856.pdf)
[com/dyn/resources/prod_documents/doc0856.pdf](http://www.atmel.com/dyn/resources/prod_documents/doc0856.pdf)
3. Certicom Research. Standards for Efficient Cryptography, SEC 2: Recommended
[Elliptic Curve Domain Parameters, Version 2.0. (January 2010), http://www.](http://www.secg.org/)
[secg.org/](http://www.secg.org/)
4. Comba, P.: Exponentiation cryptosystems on the IBM PC. IBM Systems Journal 29(4), 526–538 (1990)
5. Großsch¨adl, J., Sava¸s, E.: Instruction Set Extensions for Fast Arithmetic in Finite
Fields GF(p) and GF(2[m]). In: Joye, M., Quisquater, J.-J. (eds.) CHES 2004. LNCS,
vol. 3156, pp. 133–147. Springer, Heidelberg (2004)
6. Gura, N., Patel, A., Wander, A., Eberle, H., Shantz, S.C.: Comparing Elliptic Curve
Cryptography and RSA on 8-bit CPUs. In: Joye, M., Quisquater, J.-J. (eds.) CHES
2004. LNCS, vol. 3156, pp. 119–132. Springer, Heidelberg (2004)
7. Kargl, A., Pyka, S., Seuschek, H.: Fast Arithmetic on ATmega128 for Elliptic
Curve Cryptography. Cryptology ePrint Archive Report 2008/442 (October 2008),
[http://eprint.iacr.org/](http://eprint.iacr.org/)
8. Ko¸c, C¸.K.: High Speed RSA Implementation. Technical report, RSA Laboratories,
RSA Data Security, Inc. 100 Marine Parkway, Suite 500 Redwood City (1994)
-----
9. Lederer, C., Mader, R., Koschuch, M., Großsch¨adl, J., Szekely, A., Tillich, S.:
Energy-Efficient Implementation of ECDH Key Exchange for Wireless Sensor Networks. In: Markowitch, O., Bilas, A., Hoepman, J.-H., Mitchell, C.J., Quisquater,
J.-J. (eds.) Information Security Theory and Practice. LNCS, vol. 5746, pp. 112–
127. Springer, Heidelberg (2009)
10. Lenstra, A., Verheul, E.: Selecting Cryptographic Key Sizes. Journal of Cryptology 14(4), 255–293 (2001)
11. Liu, A., Ning, P.: TinyECC: A Configurable Library for Elliptic Curve Cryptography in Wireless Sensor Networks. In: International Conference on Information
Processing in Sensor Networks - IPSN 2008, St. Louis, Missouri, USA, Mo, April
22-24, pp. 245–256 (2008)
12. Liu, Z., Großsch¨adl, J., Kizhvatov, I.: Efficient and Side-Channel Resistant RSA
Implementation for 8-bit AVR Microcontrollers. In: Workshop on the Security of
the Internet of Things - SOCIOT 2010, 1st International Workshop, Tokyo, Japan,
November 29. IEEE Computer Society, Los Alamitos (2010)
13. National Institute of Standards and Technology (NIST). SP800-57 Part 1: DRAFT
[Recommendation for Key Management: Part 1: General (May 2011), http://](http://csrc.nist.gov/publications/drafts/800-57/Draft_SP800-57-Part1-Rev3_May2011.pdf)
[csrc.nist.gov/publications/drafts/800-57/Draft_SP800-57-Part1-Rev3_](http://csrc.nist.gov/publications/drafts/800-57/Draft_SP800-57-Part1-Rev3_May2011.pdf)
[May2011.pdf](http://csrc.nist.gov/publications/drafts/800-57/Draft_SP800-57-Part1-Rev3_May2011.pdf)
14. Scott, M., Szczechowiak, P.: Optimizing Multiprecision Multiplication for Public
[Key Cryptography. Cryptology ePrint Archive, Report 2007/299 (2007), http://](http://eprint.iacr.org/)
[eprint.iacr.org/](http://eprint.iacr.org/)
15. Szczechowiak, P., Oliveira, L.B., Scott, M., Collier, M., Dahab, R.: NanoECC:
Testing the Limits of Elliptic Curve Cryptography in Sensor Networks. In: Verdone,
R. (ed.) EWSN 2008. LNCS, vol. 4913, pp. 305–320. Springer, Heidelberg (2008)
16. Ugus, O., Hessler, A., Westhoff, D.: Performance of Additive Homomorphic ECElGamal Encryption for TinyPEDS. In: GI/ITG KuVS Fachgespr¨ach Drahtlose
Sensornetze, RWTH Aachen, UbiSec 2007 (July 2007)
17. Uhsadel, L., Poschmann, A., Paar, C.: Enabling Full-Size Public-Key Algorithms
on 8-bit Sensor Nodes. In: 4th European Workshop on Security and Privacy in
Ad-hoc and Sensor Networks, ESAS 2007, Cambridge, UK, July 2-3 (2007)
18. Wang, H., Li, Q.: Efficient Implementation of Public Key Cryptosystems on Mote
Sensors (Short Paper). In: Ning, P., Qing, S., Li, N. (eds.) ICICS 2006. LNCS,
vol. 4307, pp. 519–528. Springer, Heidelberg (2006)
## A Algorithm for Operand-Caching Multiplication
The following pseudo code shows the algorithm for multi-precision multiplication using the operand-caching method. Variables that are located in data
memory are denoted by Mx where x represents the name of the Integer a or
_b. The parameter e describes the number of locally usable registers Ra[e −_
1, . . ., 0] and Rb[e _−_ 1, . . ., 0]. The triple-word accumulator is denoted by ACC =
(ACC2, ACC1, ACC0).
-----
**Require: word size n, parameter e, n** _e, Integers a, b_
_≥_ _∈_
[0, n), c [0, 2n).
_∈_
**Ensure: c = ab.**
_r =_ _n/e_ .
_⌊_ _⌋_
⎫
_ACCforRRABforACCend forM(ACC i[[eeCACC = 0 − − ← j[re = 02111 ← +0., . . ., to, . . .,, ACC ← i to0. n] ←ACC 0] 0] − i0 ← ← do)ACCre ← + −MM( RAB0ACC1.[ do[Ann[ − −j]2 ∗1re, ACC, . . ., reR −B[1i −, . . .,1).].j]. 0]._ ⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎬
**end for** **_binit_**
**forend forMCend forMforACC(ACC i[2CACC = 0 jn[n − =2 +1 ← to, ACC ←re i i + 1] −0. n ←ACC −1] to0ACC ←)re ← n + −ACC −0( RACC2.** **doreA[0 −j.]2 ∗, ACC1 doRB[n1 −).** _re −_ _j + i]._ ⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎭
_ACC0 ←_ 0.
**for p = r** 1 to 0 do **Row Loop:**
_−_ _}_
_RA[e −_ 1, . . ., 0] ← _MA[(p + 1)e −_ 1, . . ., pe]. ⎫
**forRforBACC i[e j = 0 − = 01 ← to, . . ., toACC e 0] i − do ←1 + doM RBA[[ej −] ∗** 1R, . . .,B[i − 0].j]. ⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎬ **Part 1**
**end for**
(ACCMACCC[pe21 ← +, ACC i0.] ←0)ACC ← (0ACC. 2, ACC1). ⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎭
**end for**
**for i = 0 to n** (p + 1)e 1 do ⎫
_−_ _−_
**end forRforACCB j[e = 0 − ←1, . . ., toACC e 0] − + ←1 R doMAB[j[]e ∗ +R iB], R[e −B[1e − −** 2j]., . . ., 1]. ⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎬
**Part 2**
_ACCM(ACCACCC[(2 ←p1 ← + 1), ACCACC0.e +0 +) i ←] M ←(CACCACC[(p + 1)20, ACC._ _e + i1].)._ ⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎭
**end for**
-----
**for i = 0 to n** (p + 1)e 1 do
_−_ _−_
_RA[e −_ 1, . . ., 0] ← _MA[(p + 1)e + i], RA[e −_ 2, . . ., 1].
**for j = 0 to e** 1 do
_−_
_ACC ←_ _ACC + RA[j] ∗_ _RB[e −_ 1 − _j]._
**end for**
_ACC ←_ _ACC + MC_ [(n + i].
_MC[n + i] ←_ _ACC0._
(ACC1, ACC0) ← (ACC2, ACC1).
_ACC2 ←_ 0.
**end for**
**for i = 0 to e** 2 do
_−_
**for j = i + 1 to e** 1 do
_−_
_ACC ←_ _ACC + RA[j] ∗_ _RB[e −_ _j + i]._
**end for**
_MC[2n −_ (p + 1)e + i] ← _ACC0._
(ACC1, ACC0) ← (ACC2, ACC1).
_ACC2 ←_ 0.
**end for**
_MC[2n −_ 1 − _pe] ←_ _ACC0._
_ACC0 ←_ 0.
**end for**
**Return c.**
⎫
⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎬
⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎭
⎫
⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎬
⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎭
**Part 3**
**Part 4**
## B Example: 160-Bit Operand-Caching Multiplication
C[38] C[19] C[0]
1
A[19]B[19]
4
A[0]B[0]
2
2
2
1
A[0]B[19]
**Fig. 8. Operand-caching multiplication for n = 20 and e = 7**
3
3
4
-----
|
{
"disclaimer": "Notice: The following paper fields have been elided by the publisher: {'abstract', 'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1007/s00145-020-09351-2?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1007/s00145-020-09351-2, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "BRONZE",
"url": "https://link.springer.com/content/pdf/10.1007/978-3-642-23951-9_30.pdf"
}
| 2,011
|
[
"JournalArticle"
] | true
| 2011-09-28T00:00:00
|
[] | 12,244
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Engineering",
"source": "s2-fos-model"
},
{
"category": "Environmental Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/0266f264f35f3c0049ffe3b517170d081a1824e1
|
[
"Computer Science"
] | 0.865562
|
Leveraging User-Diversity in Energy-Efficient Edge-Facilitated Collaborative Fog Computing
|
0266f264f35f3c0049ffe3b517170d081a1824e1
|
IEEE Access
|
[
{
"authorId": "2055852524",
"name": "Antoine Paris"
},
{
"authorId": "1423728195",
"name": "Hamed Mirghasemi"
},
{
"authorId": "1798136",
"name": "I. Stupia"
},
{
"authorId": "1698047",
"name": "L. Vandendorpe"
}
] |
{
"alternate_issns": null,
"alternate_names": null,
"alternate_urls": [
"http://ieeexplore.ieee.org/servlet/opac?punumber=6287639"
],
"id": "2633f5b2-c15c-49fe-80f5-07523e770c26",
"issn": "2169-3536",
"name": "IEEE Access",
"type": "journal",
"url": "http://www.ieee.org/publications_standards/publications/ieee_access.html"
}
|
With the increasing number of heterogeneous resource-constrained devices populating the current wireless ecosystem, enabling ubiquitous computing at the edge of the network requires moving part of the computing burden back to the edge to reduce user-side latency and relieve the backhaul network. Motivated by this challenge, this work investigates edge-facilitated collaborative fog computing to augment the computing capabilities of individual devices while optimizing for energy-efficiency. Collaborative-computing is modeled using the Map-Reduce framework, consisting in two computing rounds and a communication round. The computing load is optimally distributed among devices, taking into account their diversity in terms of computing and communication capabilities. Devices local parameters such as CPU frequency and RF transmit power are also optimized for energy-efficiency. The corresponding optimization problem is shown to be convex and optimality conditions are obtained through Lagrange duality theory. A waterfilling-like interpretation for the size of the computing load assigned to each device is given. Numerical experiments demonstrate the benefits of the proposed collaborative-computing scheme over various other schemes in several respects. Most notably, the proposed scheme exhibits increased probability of successfully dealing with more demanding computations in time, along with significant energy-efficiency gains. Both improvements come from the scheme ability to advantageously leverage devices diversity.
|
Received June 11, 2021, accepted July 1, 2021, date of publication July 5, 2021, date of current version July 13, 2021.
_Digital Object Identifier 10.1109/ACCESS.2021.3094888_
# Leveraging User-Diversity in Energy-Efficient Edge-Facilitated Collaborative Fog Computing
ANTOINE PARIS, (Member, IEEE), HAMED MIRGHASEMI, (Member, IEEE),
IVAN STUPIA, (Member, IEEE), AND LUC VANDENDORPE, (Fellow, IEEE)
Institute for Information and Communication Technologies, Electronics, and Applied Mathematics (ICTEAM), Catholic University of Louvain (UCLouvain),
1348 Louvain-la-Neuve, Belgium
Department of Electrical Engineering (ELEN), Catholic University of Louvain (UCLouvain), 1348 Louvain-la-Neuve, Belgium
Communication System Group (CoSy), Catholic University of Louvain (UCLouvain), 1348 Louvain-la-Neuve, Belgium
Corresponding author: Antoine Paris (antoine.paris@uclouvain.be)
This work was supported by the Fonds de la Recherche Scientifique (F.R.S.-FNRS) through the Excellence Of Science (EOS) Program
(MUlti-SErvice WIreless NETwork) under Project 30452698. The work of Antoine Paris was supported by the F.R.S.-FNRS.
**ABSTRACT With the increasing number of heterogeneous resource-constrained devices populating the**
current wireless ecosystem, enabling ubiquitous computing at the edge of the network requires moving
part of the computing burden back to the edge to reduce user-side latency and relieve the backhaul
network. Motivated by this challenge, this work investigates edge-facilitated collaborative fog computing to augment the computing capabilities of individual devices while optimizing for energy-efficiency.
Collaborative-computing is modeled using the Map-Reduce framework, consisting in two computing rounds
and a communication round. The computing load is optimally distributed among devices, taking into account
their diversity in terms of computing and communication capabilities. Devices local parameters such as CPU
frequency and RF transmit power are also optimized for energy-efficiency. The corresponding optimization
problem is shown to be convex and optimality conditions are obtained through Lagrange duality theory.
A waterfilling-like interpretation for the size of the computing load assigned to each device is given.
Numerical experiments demonstrate the benefits of the proposed collaborative-computing scheme over
various other schemes in several respects. Most notably, the proposed scheme exhibits increased probability
of successfully dealing with more demanding computations in time, along with significant energy-efficiency
gains. Both improvements come from the scheme ability to advantageously leverage devices diversity.
**INDEX TERMS Wireless collaborative computing, map-reduce, energy-efficiency, joint computation and**
communications optimization, fog computing.
**I. INTRODUCTION**
The current trends in communication and networking suggest
that the future wireless ecosystem will be populated by a
massive number of heterogeneous devices (i.e., in terms of
computing and communication capabilities): from relatively
powerful smartphones and laptops to ultra-low-power sensors, actuators and other connected ‘‘things’’ [1]–[3]. At the
same time, emerging applications like virtual and augmented
reality, context-aware computing, autonomous driving, Internet of Things (IoT) and so forth, require more and more
computing capabilities while aiming for smaller and smaller
latency [4]. All in all, recent years have seen the focus moving
The associate editor coordinating the review of this manuscript and
approving it for publication was Mahdi Zareei .
from communications as an objective per se, to communications as a way to enhance computing capabilities of energy
limited devices [5]–[7].
This paradigm shift started with Mobile Cloud Computing (MCC) [8], [9] first, and with Multi-access Edge
Computing (MEC) [10]–[13] later on. While MCC proved
itself to be effective to enable ubiquitous computing on
resource-constrained devices while prolonging their battery
life, MEC has the advantage of offering smaller computing latencies while reducing the pressure on the backhaul
network. This makes MEC both more suitable for ultra-lowlatency applications emerging from the recent 5G developments and more able to cope with the ever growing number
of connected devices and their ever growing computing
demand. Compared to MCC, the inherent spatial distribution
-----
of MEC also has the advantage of offering some level of
decentralization.
In contexts where MCC latency is unacceptable and in
the absence of MEC servers nearby, or when the use of
third-party owned MCC/MEC is deliberately ruled-out for
privacy reasons, fog computing offers an even more decentralized alternative. Fog computing is formally defined in [14]
as ‘‘a huge number of heterogeneous wireless devices that
**_communicate and potentially cooperate among them and_**
**_with the network to perform processing tasks without the_**
_intervention of third parties’’. Those distributed computing_
resources can also be exploited to enhance the computing
capabilities of individual devices. As MEC, fog computing benefits from reduced user-side latency and reduces the
pressure on the backhaul network: two features recognized
as key enablers for ubiquitous artificial intelligence (AI)
at the edge of the network [15]. Achieving this, however,
requires to move part of the computing burden from the
powerful server-side to the resource-constrained user-side. To
accommodate for relatively complex processing tasks on such
limited devices requires to (i) enable devices collaboration to
pool their computing capabilities and (ii) take care of devices
resources management, both in terms of computing resources
and communication resources. It is also worth noting that fog
computing can be integrated in a multi-tier architecture along
with MEC and MCC [16]. In this paper, we jointly optimize
for energy-efficiency the computation and communication
resources of a set of heterogeneous and resource-constrained
mobile devices taking part in collaborative fog computing.
_A. APPLICATION SCENARIO_
As already mentioned, enabling computationally demanding
intelligent mobile systems at the edge of the network requires
to offload part of the computing burden to mobile devices
to reduce user-side latency and relieve the backhaul network [15]. A recent comprehensive survey on ‘‘Deep Learning in Mobile and Wireless Networking’’ [17] discusses fog
computing to support those two objectives in the context of
machine learning (ML) or deep learning (DL) inference; two
key components of ubiquitous AI. ML/DL models, however,
often contain tens to hundreds of millions of parameters. As
such, it might be prohibitive or even impossible for a single
mobile device limited in computing, memory and battery
capacity to process (or even simply store) the full ML/DL
model needed to perform the inference with a reasonable
latency [18]. Though it is possible to reduce the size of
a ML/DL model using various model compression techniques such as pruning, weights quantization and so forth,
this always comes at the cost of accuracy [19]. Enabling
devices collaboration to augment their individual processing
capabilities is another solution that does not sacrifice accuracy [18]. Combined with proper resources management and
allocation to preserve devices battery life, this collaborative
inference approach should allow the processing of reasonably large ML/DL models. It is worth noting that this kind
of distributed/collaborative inference is envisioned as a key
enabler to ubiquitous AI in future 6G networks [20]. This
example application scenario, also known as ‘‘model-split’’
inference [21] and illustrated and detailed in Fig. 1, thus
consists in distributing a ML/DL model pre-trained off-device
in the cloud to perform collaborative on-device inferences on
local input data later on.
_B. RELATED WORKS AND MOTIVATIONS_
This paper extends our previous work [22], primarily with
more realistic energy consumption models, both for devices
local computations and for communications between devices.
Although this change might appear subtle at first, it more
than doubles the number of degrees of freedom in the
collaborative-computing scheme, making it both more challenging to optimize and more interesting. While the optimization problem in [22] was relatively easy to solve and
even admitted a semi-closed form solution, it only offered
little insights into the structure of the optimal solution. At the
opposite, the optimization problem in this work is more
challenging and doesn’t admit a closed-form solution. Yet,
we are now able to offer a waterfilling-like interpretation of
the optimal computing load assigned to each device, thus
gaining some insights on the structure of the optimal solution
that we were not able to offer in our previous work. Naturally,
the additional degrees of freedom also allow to improve the
performances of the collaborative-computing scheme (see
numerical experiments in Sec. V for comparison).
The system model used in this paper is inspired by
previous works on wireless distributed computing (WDC)
[23]–[28]. Most notably, we also use the Map-Reduce distributed computing framework [29] with an access point (AP)
or base station (BS) facilitating the communications between
devices (i.e., communications are edge-facilitated). While
Map-Reduce was originally introduced by Google as a programming model for processing very large data sets in parallel on several hundreds or thousands of machines in a
reasonable amount of time, the framework quickly attracted
a wider variety of applications, e.g., machine learning, physical simulations, digital media processing tasks, etc [30].
Around the same time, and with the advent of mobile
devices, Map-Reduce also started to be considered as a
viable framework to distribute computing tasks in mobile
systems [31], [32]. Together, these evolutions of the
Map-Reduce framework field of applications make it a very
good choice for the collaborative ML/DL inference application scenario discussed in the last section, and, more
generally, for the development of ubiquitous AI in future
6G networks [15], [20]. The vast majority of these works,
however, study WDC from a network-coding theory point
of view, focusing on coded distributed computing (CDC)
and discussing the trade-offs between the computation and
communication loads incurred by the collaboration. In short,
the conclusion is that increasing devices computing load
makes it possible to leverage network coding opportunities
during the exchange of intermediate computation results,
hence reducing the communication load. Considering the
-----
**FIGURE 1. Collaborative fog computing ML/DL inference scenario (adapted from the**
edge-based app-level mobile data analysis approach illustrated in [17]). The ML/DL
model is first trained off-device in the cloud using offline datasets. The ML/DL
model weights, noted w, are then offloaded to the edge of the network for future
on-device inference. Each device n ∈ [N] wants to perform some inference φ(dn, w )
on its local input data dn using the model weights w . However, this operation might
be prohibitive or even impossible to carry on a single mobile device due to memory,
computing capabilities or battery limitations. Distributing the ML/DL model weights
_w across multiple devices to enable collaborative inference is envisioned as a_
potential solution to this issue [18]. It is also worth noting again that this fog
computing inference scenario significantly reduces backhaul network traffic and
user-side latency compared to the cloud-based scenario in which inferences are
performed in the cloud [17].
_inherent energy-limited nature of mobile devices as the_
_main bottleneck to WDC, the focus in this work is shifted_
_towards the allocation of computing and communication_
_resources and optimization of the collaboration to mini-_
_mize devices energy consumption. The two approaches are_
however not mutually exclusive and could be combined in
future collaborative-computing schemes to further improve
the global energy-efficiency of the system.
Most existing works on WDC consider the set of collaborating devices to be homogeneous in terms of computing and
communication capabilities (with recent notable exceptions
– still focused on CDC – like [24] and [28]). Under these
conditions, the computing load is thus uniformly distributed
across all devices. Motivated by current trends in the wire_less ecosystem, we here consider the set of devices to be_
_heterogeneous instead. As a consequence, it might no longer_
be optimal to uniformly distribute the computing load (e.g.,
the ML/DL model weights, see Fig. 1) across mobile devices.
To allow our collaborative-computing scheme to take into
account – and leverage – devices diversity, our model thus
allows arbitrary partition of the computing load. Compared
to previous works focused on CDC, we also make the latency
constraint accompanying the computing task an explicit
constraint.
Various other cooperative-computing schemes were also
studied in the literature, see e.g., [33]–[38]. [33] discusses
cooperative-computing and cooperative communications in
the context of MEC systems wherein a user can partially
(or totally) offload a computing task to both a MEC server
and a so-called helper device that can then (i) perform
some local computations for the user device (i.e., cooperative computing), (ii) further offload part or all the task
to the MEC server (i.e., cooperative communication), or
(iii) both. The system model and problem formulation used
in this work also owes a lot to [33], especially with regards
to devices energy consumption models. Reference [34] also
devises an energy-efficient cooperative-computing scheme
in which a mobile device can partially or totally offload
a computing task to a surrounding idle device acting as
a helper. In the context of Mobile Wireless Sensors Networks (MWSNs), [35] augments this framework by optimally
selecting the helper device among a set of N surrounding
devices. In [36], a wireless powered cooperative-computing
scheme wherein a user device can offload computations to N
helper devices is described. In [37] authors describe Mobile
Device Cloud (MDC), i.e., a framework in which power balancing is performed among a cluster of mobile devices, and
empirically optimize the collaboration to maximize the lifetime of the set of devices. Finally, in [38], an energy-efficient
and incentive-aware network-assisted (i.e., coordinated by
the edge of the network), device-to-device (D2D) collaboration framework is presented.
_C. OBJECTIVE AND CONTRIBUTIONS_
Given the heterogeneous nature of mobile devices and their
limitations in terms of memory, computing and communication capabilities and battery life, this work aims to provide
insights into the following question: how to distribute a
**computing load (e.g., ML/DL model) across an hetero-**
**geneous set of resource-constrained wireless devices to**
**complete a given set of computing tasks (e.g., ML/DL**
**inferences) in the most energy-efficient way, while satis-**
**fying a given deadline?**
-----
More precisely, the contributions of this paper can be
summarized as follows
:
- we propose an N -devices edge-facilitated collaborative
fog computing scheme based on the Map-Reduce distributed computing framework [29] and formulate a joint
computation and communication resources optimization
problem with energy-efficiency as objective ;
- we gain engineering insights into the structure of the
optimal solution by leveraging Lagrange duality theory
and offer a waterfilling-like interpretation for the size of
the computing load assigned to each device ;
- through numerical experiments, we compare the performance of the proposed scheme with various other
schemes using less degrees-of-freedom in the optimization (such as the one proposed in [22]) to analyze the
relative benefits of each set of variables being optimized
and to show that the proposed scheme advantageously
exploits devices diversity.
_D. ORGANIZATION OF THE PAPER_
Section II starts by describing in details the collaborative
computing model and the energy and time consumption models for both local computation and edge-facilitated communications. Next, Sec. III formulates the joint computation and
communication resources allocation problem and analyzes its
feasibility. Section IV then reformulates the problem, proves
its convexity in this new formulation and leverages Lagrange
duality theory to gain some insights on the optimal solution
of the problem. Section V benchmarks the performances
of the optimal collaborative-computing scheme against various other schemes through numerical experiments. Finally,
Sec. VI discusses the results obtained in this work, their
limits, and opportunities for future research.
**II. SYSTEM MODEL**
As already illustrated in Fig. 1, we consider an heterogeneous
set of N wireless devices indexed by n [N ] sharing a
∈
common AP or BS. Each device n wants to perform a given
computing task φ(dn, w) within a given latency τ, with dn
some D-bit local input data to device n and w some L-bit
data common to all N devices. As detailed in Sec. I-A and
in Fig. 1, w could be a ML/DL model that was pre-trained
in the cloud, while φ(dn, w) could represent an inference
performed using this model w on an input dn. Motivated by
this example application, it is assumed that L _D. As a_
≫
consequence of the large size of w, it might be impossible or
prohibitive in terms of energy consumption for an individual
device to complete the computing task φ(dn, w) within the
deadline τ . Devices thus pool together and collaborate to
augment their individual computing capabilities. The collaborative computing model used in this work, i.e., MapReduce [29] is described in Sec. II-A. Next, and because
we are here concerned by optimizing the energy-efficiency
of the collaboration, Sec. II-B and II-C describe the models
used to quantify the time needed and energy consumed by the
different phases of the collaboration.
The AP/BS is responsible for coordinating and optimizing
the collaboration. This makes our collaborative-computing
scheme edge-facilitated (or network-assisted) and fits with
the fog computing definition given in the introduction. To
allow for offline optimization of the collaborative-computing
scheme, we assume that the AP/BS has perfect non-causal
knowledge of the uplink channels (i.e., devices to AP/BS)
during the communication phase, and perfect knowledge of
the computing and communication capabilities of all devices.
Although unrealistic with regard to the channels, this simplifying assumption allows us to provide a first performance evaluation of the proposed collaborative fog computing scheme in a best-case scenario.
_A. COLLABORATIVE COMPUTING MODEL_
The computing tasks {φ(dn, w)}n[N]=1 [(e.g., ML/DL inferences)]
are shared between N devices according to the Map-Reduce
framework [29]. First, we assume that the L-bit data w (e.g.,
ML/DL model weights) can be arbitrarily partitioned in N
smaller ln-bit data wn (one for each device n) with ln ∈ R≥0[1]
and
�N
(1)
_n=1_ _[l][n][ =][ L][.]_
As opposed to previous works focusing on CDC [23]–[28],
we are not assuming any redundancy in the computing loads
{wn}n[N]=1 [assigned to each device, that is][ w][i][ ∩] _[w][j][ = ∅]_ [for]
all i ̸= j.[2] Also, the sizes {ln}n[N]=1 [of the assigned computing]
loads {wn}n[N]=1 [are optimized for energy-efficiency taking into]
account the diversity of devices instead of being uniform
(e.g., ln = L/N or a multiple for all n) and fixed ahead of
time. Assuming relatively large downlink rates, and because
the focus is on the energy consumption of mobile devices
(rather than the energy consumption of the AP), we neglect
the time and energy needed to transmit wn from the AP to
device n, for all n [N ]. To make collaborative-computing
∈
possible, we also assume that the D-bit local input data
{dn}n[N]=1 [were shared between mobile devices through the AP]
in a prior phase that we neglect in this work because D is
assumed to be relatively small compared to L [23], [24].
1) MAP
During the first phase of the Map-Reduce framework, namely
the Map phase, each device n produces intermediate computation results (e.g., partial inference results using a subset wn
of the ML/DL model weights w)
_gn(d1, wn), gn(d2, wn), . . ., gn(dN_ _, wn),_
1In practice, ln should be an integer multiple of the size of the smallest
possible division of w. In this work, we relax this practical consideration to
avoid dealing with integer programming later on. Note that ln = 0 is also
possible, in which case device n does not participate to the collaboration.
2 In addition to creating network-coding opportunities during the communication phase, thereby decreasing the communication load, redundant
computing loads can also provide some level of protection against straggler
and faulty devices. Nevertheless, to keep the focus on the fundamental nature
of the energy-efficient collaborative-computing problem, we assume (i) no
straggler devices, (ii) no faulty devices and (iii) no network-coding during
the communication phase.
-----
where gn is the Map function executed at device n. The size
of the intermediate computation result gn(dm, wn) produced
by device n for device m is assumed to be proportional to
the size ln of its assigned computing load wn and is given by
_βln. Each device n thus computes intermediate computation_
results for all the other devices, i.e., gn(dm, wn) for all m ̸= n,
and for itself, i.e., gn(dn, wn), using the part wn of w received
from the AP. The Map phase is illustrated in the colored and
framed columns of Fig. 2.
2) SHUFFLE
Next, devices exchange intermediate computation results
with each other in the so-called Shuffle phase. As already
mentioned multiple times, coded shuffling [23]–[28] is not
considered in this work. In this simplified Shuffle phase, each
device n thus directly transmits the intermediate computation
results gn(dm, wn) to device m via the AP, for all m ̸= n.
Device n thus needs to transmit a total of (N − 1)βln bits of
intermediate computation results to the AP. To ease notations
in the rest of the paper, we define α = (N − 1)β. The
Map phase can thus be seen as a data compression phase,
reducing the size of wn from ln bits to αln bits of intermediate
computation results before transmission in the Shuffle phase.
The intermediate computation results exchanged during the
Shuffle phase are indicated in bold on Fig. 2.
3) REDUCE
Finally, during the Reduce phase, each device m combines a
total of [�]n[N]=1 _[β][l][n][ =][ β][L][ bits of intermediate computation]_
results {gn(dm, wn)}n[N]=1 [produced by all the collaborating]
devices to obtain φ(dm, w) as
_hm�g1(dm, w1), g2(dm, w2), . . ., gN_ (dm, wN )�,
where hm is the Reduce function executed at device m. This
last operation, which could be thought of as combining all the
partial inference results produced by all the devices to get the
final inference φ(dm, w), is illustrated in the colored rows on
Fig. 2.
We note tn[MAP], tn[SHU] and tn[RED] the amount of time needed
to perform the Map, Shuffle and Reduce phases, respectively,
at device n. Because the Map and Shuffle phases must be
over at every device before the Reduce phase starts (as all
the intermediate computation results need to be available),
we have the following constraint
_tn[MAP]_ + tn[SHU] ≤ _τ −_ maxn _n_ }, _n ∈_ [N ]. (2)
[{][t] [RED]
_B. LOCAL COMPUTING MODEL_
During the Map phase, each device n receives ln bits to process. The number of CPU cycles needed to process one bit of
input data at device n is assumed to be given by a constant cn.
At the opposite of our previous work [22], devices are now
assumed to be able to perform dynamic frequency scaling
(DFS), i.e., a device can adjust its CPU frequency on the
fly depending on the needs. Then, noting κn the effective
capacitance coefficient (that depends on the chip architecture
while constraint (2) becomes
_tn[MAP]_ + tn[SHU] ≤ _τ −_ _t_ [RED], _n ∈_ [N ]. (8)
**FIGURE 2. Illustration of the Map-Reduce collaborative-computing**
model. The computing tasksDuring the Map phase, each device {φ(dn, w n produces intermediate computation)}[N]n=1 [are shared across][ N][ devices.]
resultsphase, the intermediate computation results in bold on the figure are {gn(dm, wn)}[N]m=1 [(see framed columns). Next, during the Shuffle]
transmitted via the AP to the devices for which they have been computed.
The AP is said to facilitate the communications. Finally, during the Reduce
to obtainphase, each device φ(dm, w ) (see colored rows). m combines the intermediate values {gn(dm, wn)}[N]n=1
of each device), the energy needed for computation during the
Map phase can be modeled as [11], [12], [33]
_En[MAP]_ = _[κ][n][c]n[3][l]n[3]_ _n ∈_ [N ] (3)
(tn[MAP])[2][,]
with the following constraint
_cnln ≤_ _tn[MAP]fn[max],_ _n ∈_ [N ] (4)
where fn[max] is the maximum CPU frequency of device n.
Motivated by the fact that D _L and to avoid integer_
≪
variables in our optimization problem later on, the energy and
time to process local input data {dn}n[N]=1 [during the Map phase]
have been neglected in both (3) and (4).
Similarly, the energy needed at device n to combine the βL
bits of intermediate computation results during the Reduce
phase can be modeled as
_n[(][β][L][)][3]_
_En[RED]_ = _[κ][n][c][3]_ _n ∈_ [N ] (5)
(tn[RED])[2][,]
with the following constraint
_cnβL ≤_ _tn[RED]fn[max],_ _n ∈_ [N ]. (6)
Because increasing tn[RED] is always favorable for energyefficiency and because the Reduce phase cannot start before
the Map and Shuffle phases are over, one can see that we will
always have the same tn[RED] = t [RED] across all N devices. As a
consequence, constraint (6) becomes
�
_t_ [RED] (7)
≤
_βL max_
_n_
� _cn_
_fn[max]_
-----
_C. EDGE-FACILITATED DEVICES COMMUNICATIONS_
During the Shuffle phase, devices exchange intermediate
computation results through the AP. This exchange thus
involves both an uplink communication (devices to AP) and
a downlink communication (AP to devices). In most applications however, it is reasonable to assume that the downlink
rates are much larger than the uplink rates. For this reason,
and because we are primarily interested by the energy consumed by the resource-constrained devices, we neglect the
time needed for the downlink communications in this work.
We also assume that all the devices can communicate in
an orthogonal manner to the AP (e.g., through frequency
division multiple access techniques, or through interference
alignment [39]). Let hn denote the wireless channel power
gain from device n to the AP during the Shuffle phase, pn
the RF transmit power of device n, B the communication
bandwidth and N0 the noise power spectral density at the AP.
The achievable uplink rate of device n is then given by[3]
**FIGURE 3. Another interpretation of the energy-efficient collaborative fog**
computing problem: how can we send a total of L bits at a given rate of
_L/τ bits/second through N parallel special channels consisting of a_
‘‘computing channel’’ in series with a wireless communication channel in
the most energy-efficient way?
wireless communication channel in the most energy-efficient
way? This is illustrated in Fig. 3. This interpretation was
already mentioned in [34] for a single channel (i.e., for
_N_ 1) and is here generalized for multiple parallel channels.
=
_A. FEASIBILITY_
Before solving Problem (P1), we first seek to determine
its feasibility condition, i.e., condition that ensures that the
system is able to meet the deadline.
_Lemma 1 (Feasibility): Problem (P1) is feasible if and_
_only if the task size L satisfies_
�
_rn(pn) = B ln_ 1 + _[p][n][h][n]_
_N0B_
�
(9)
in nats/second.[4] Noting P[c]n [the constant energy consumption]
of the communication circuits at device n, the energy consumed during the Shuffle phase is thus given by
_En[SHU]_ = tn[SHU](pn + P[c]n[)] (10)
with the following constraints
_αln ≤_ _tn[SHU]rn(pn),_ _n ∈_ [N ] (11)
and
_pn ≤_ _p[max]n_ _,_ _n ∈_ [N ] (12)
where p[max]n is the maximum RF transmit power at device n.
**III. PROBLEM FORMULATION**
Putting everything together, the energy-efficient collaborative
fog computing problem can be formulated as follows
_N_
�
(P1) : minimize _En[MAP]_ + En[SHU] + En[RED]
**_l,t[MAP],t[SHU],t_** [RED],p
_n=1_
subject to (1), (4), (7), (8), (11), (12)
_ln, tn[MAP], tn[SHU], pn,_ _t_ [RED] ≥ 0,
_n_ [N ]
∈
_fn[max]_ _._ (13)
_cn_
_τ −_ _βL maxn_ � _fnc[max]n_
_n_ _/cn_
1 _[α][f][ max]_
+ _rn(p[max]n_ )
�
_L ≤_ _Lmax =_
_N_
�
_n=1_
_Proof: The maximum computing capacity of the sys-_
tem Lmax is obtained by solving the following optimization
problem
_N_
�
_Lmax ≜_ maximize _ln_
**_l,t[MAP],t[SHU],t_** [RED],p
_n=1_
subject to (4), (7), (8), (11), (12)
_ln, tn[MAP], tn[SHU], pn,_ _t_ [RED] ≥ 0,
_n ∈_ [N ].
where l, t[MAP], t[SHU] and p are N -length vectors containing
the corresponding variables. Interestingly, this problem can
be reformulated as follows: how can we send a total of L bits
at a given rate of L/τ bits/second through N parallel special
channels consisting of a ‘‘computing channel’’ in series with a
3The noise power N0B can be multiplied by the SNR gap � to account for
practical modulation and coding schemes. This additional factor is left out
here for the sake of clarity.
4Nats and bits are used interchangeably in this paper (with the proper
factor correction applied when needed) to avoid carrying ln(2) factors in the
derivations later on.
For the maximum computing capacity to be achieved,
constraints (8), (12) and (7) must be met, that is, the entire
time τ is used by all devices, all devices transmit at their
maximum RF transmit power p[max] and the Reduce phase
executes as fast as possible. Next, the two constraints (4)
and (11) on ln can be re-written in a single constraint as
follows
_ln ≤_ min � _tnMAPcnfn[max]_ _,_ _α[1]_ _[t]n[SHU]rn(p[max]n_ )� _._ (14)
At the optimum, this constraint is satisfied and, given the
relationship between tn[MAP] and tn[SHU], we have
_α_ _[t]n[MAP]fn[max]_ = tn[SHU]rn(p[max]n ), (15)
_cn_
which intuitively means that the number of bits of intermediate values produced by the Map phase at full speed must equal
the number of bits that can be transmitted at full speed during
-----
the Shuffle phase. Then using the satisfied constraints (8)
and (7) together, we have
�
_,_ (16)
_x[3]_ is a convex function for x 0. Its perspective function,
≥
_x[3]/y[2], is thus also a convex function for y > 0. The term_
associated to the energy consumed by the Map phase is thus
jointly convex with respect to ln ≥ 0 and tn[MAP] _> 0. Next,_
the terms associated to the energy consumed by the Shuffle
phase are linear and hence convex by definition. Finally,
the function 1/x[2] is a convex function with respect to x > 0
which makes the term associated to the energy consumed
by the Reduce phase a convex function as well. As convexity is preserved under addition, the objective function of
Problem (P2) is a convex function. We then show that the
set defined by the constraints is a convex set. The equality
constraint (1) is affine and thus defines a hyperplane. Next,
inequalities (4), (7), (8), and (21) are either linear or affine
and thus define a polyhedron. The only remaining constraint
(omitting trivial positivity constraints on all variables) is then
constraint (20). For constraint (20) to define a convex set,
its right-hand side term must be a concave function. The
function rn(x) is a concave function with respect to x ≥ 0. Its
perspective function, yrn(x/y) is thus also a concave function
with respect to x ≥ 0 and y > 0. Because the intersection of a
hyperplane, a polyhedron and a convex sublevel set remains a
convex set, the set defined by the constraints of Problem (P2)
is also convex. Problem (P2) can easily be solved using a software for convex optimization like cvxopt [40]. This wouldn’t however
offer any interpretation of the results. To this effect, we seek to
gain some insights into the optimal solution to Problem (P2)
mathematically using Lagrange duality theory.
We thus let λ ∈ R, βn ≥ 0, µn ≥ 0 be the Lagrange multipliers associated with constraints (1), (8) and (20) respectively. The partial Lagrangian is then given by
�N _κnc[3]n[l]n[3]_ _n[(][β][L][)][3]_
_L(x, µ, β, λ) =_ _n=1_ (tn[MAP])[2][ +][E][n] [+][t]n[SHU]P[c]n [+][ κ][n](t[c][RED][3] )[2]
�N � � _En_ ��
+ _n=1_ _[µ][n]_ _αln −_ _tn[SHU]rn_ _tn[SHU]_
+ βn(tn[MAP] + tn[SHU] + t [RED] − _τ_ )
� �N �
+ λ _L −_ _n=1_ _[l][n]_ _,_ (22)
where optimization variables and Lagrange multipliers have
been aggregated in the corresponding vectors to ease notations. The dual function is then given by
(DF) : g(µ, β, λ) = minx _L(x, µ, β, λ)_
s.t. (4), (7), (21)
_tn[MAP], tn[SHU],_ _t_ [RED] ∈ [0, τ ],
_n_ [N ]
∈
_ln,_ _En ≥_ 0, _n ∈_ [N ].
As the dual function provides a lower bound to the optimal
value of the primal problem, we then seek to maximize it to
obtain the best possible lower bound. The dual problem is
_tn[MAP]_ + tn[SHU] = τ − _βL maxn_
which allows us to finally obtain
� _cn_
_fn[max]_
_tn[MAP]_ = _τ −_ _βL maxn_ _n_ _/�cnfn[max]cn_ � _,_ (17)
1 _[α][f][ max]_
+ _rn(p[max]n_ )
and
_tn[SHU]_ = _τ −_ _βL maxnn_ �)fn[max]cn � _._ (18)
1 + _α[r]f[n]n[(][max][p][max]/cn_
The maximum computing load Lmax is thus given by
_τ −_ _βL maxn_ � _fn[max]cn_
_n_ _/cn_
1 _[α][f][ max]_
+ _rn(p[max]n_ )
�
_fn[max]_ _._ (19)
_cn_
_Lmax =_
_N_
�
_n=1_
We see through (17) and (18) that, at full capacity, the time
for the Map and Shuffle phases is shared according to the
ratio of (i) the maximum rate at which the Map phase can
produce intermediate computation results αfn[max]/cn, and (ii)
the maximum rate at which the Shuffle phase can transmit
intermediate computation results rn(p[max]n ). At lower than full
capacity, these time intervals will be able to adjust taking into
account the energy-efficiency of both phases.
**IV. OPTIMAL SOLUTION**
Inspired by [33], we then introduce a new set of variables
_En = tn[SHU]pn, i.e., the RF energy consumed by the Shuffle_
phase, and substitute pn for En/tn[SHU] to convexify Problem (P1). With this new variable, constraints (11) and (12)
can be re-written as
_αln ≤_ _tn[SHU]rn_
� _En_
_tn[SHU]_
�
(20)
and
_En ≤_ _tn[SHU]p[max]n_ (21)
respectively, for all n [N ]. Problem (P1) thus becomes
∈
�N _κnc[3]n[l]n[3]_
(P2) : **_l,t[MAP]minimize,t[SHU],t_** [RED],E _n=1_ (tn[MAP])[2][ +][ E][n][ +][ t]n[SHU]P[c]n
+ _[κ][n][c]n[3][(][β][L][)][3]_
(t [RED])[2]
subject to (1), (4), (7), (8), (20), (21)
_ln, tn[MAP], tn[SHU], En,_ _t_ [RED] ≥ 0,
_n ∈_ [N ].
We now prove the convexity of this new formulation.
_Lemma 2 (Convexity): Problem (P2) is convex._
_Proof: As this is a minimization problem, we start by_
showing the convexity of the objective function. The function
-----
given by
(D1) : maximize _g(µ, β, λ)_
**_µ,β,λ_**
subject to µn, _βn ≥_ 0, _n ∈_ [N ].
Problem (P2) is convex (Lemma 2) and satisfies Slater’s
condition if it is strictly feasible (in the sense given in
Lemma 1). Strong duality thus holds and Problem (P2) can
be solved by solving the dual problem (D1).
_A. DERIVATION OF THE DUAL FUNCTION_
Before solving the dual problem (D1), we seek to evaluate
the dual function g(µ, β, λ) for all µ, β, λ by solving Problem (DF). To this effect, we first decompose Problem (DF) in
2N 1 sub-problems as follows
+
minimizeln,tn[MAP] (κtn[MAP]nc[3]n[l]n)[3][2][ +][ (][αµ][n][ −] _[λ][)][l][n][ +][ β][n][t]n[MAP]_
subject to 0 ≤ _ln ≤_ _tn[MAP]fn[max]/cn_
_tn[MAP]_ ≤ _τ_ (23)
� _En_ �
minimizeEn,tn[SHU] _En + (P[c]n_ [+][ β][n][)][t]n[SHU] − _µntn[SHU]rn_ _tn[SHU]_
subject to 0 ≤ _En ≤_ _tn[SHU]p[max]n_
_tn[SHU]_ ≤ _τ_ (24)
�N _κnc[3]n[(][β][L][)][3]_
minimizet [RED] _n=1_ (t [RED])[2] + βnt [RED]
� _cn_ �
subject to βL maxn _fn[max]_ ≤ _t_ [RED] ≤ _τ._ (25)
It is interesting to note that Problems (23) and (24) correspond to the Map and Shuffle phases at device n respectively
while Problem (25) corresponds to the Reduce phase.
_Lemma 3 (Solution of Problem (23)): For any µn, βn ≥_ 0
_and λ ∈_ R, the optimal solution of Problem (23) satisfies
_ln[∗]_ [=][ M]n[∗][t]n[MAP][∗] (26)
_with Mn[∗][, the effective processing rate (in bits/second) of]_
_device n defined as_
_Lemma 4 (Solution of Problem (24)): For any µn, βn ≥_ 0,
_the optimal solution of Problem (24) satisfies_
_En[∗]_ [=][ p]n[∗][t]n[SHU][∗] (30)
_with p[∗]n[, the RF transmit power used during the Shuffle phase]_
_at device n defined as_
0 _Bµn ≤_ _[BN]hn[0]_
_p[∗]n_ [≜] B �µn − _[N]hn[0]_ � _Bµn ∈_ � _BNhn0_ _[,][ BN]hn[0]_ [+][ p]n[max]� (31)
p[max]n _Bµn ≥_ _[BN]hn[0]_ _n_
[+][ p][max]
_and tn[SHU][∗]_ _given by_
= 0 _ρ2,n < 0_
_tn[SHU][∗]_ ∈ [0, τ ] _ρ2,n = 0_ (32)
= τ _ρ2,n > 0_
_with ρ2,n = µnrn(p[∗]n[)]_ [−] _[P][c]n_ [−] _[β][n][ −]_ _[µ][n]_ 1+pp[∗]n[∗]nNhnN0hn0B + _δ2,np[max]n_ _and_
_δ2,n =_
0 _p[∗]n_ _[<][ p]n[max]_
µn 1+p[max]nNhn0 _Nhn0B_ − 1 _p[∗]n_ [=][ p]n[max]. (33)
_Proof: See Appendix B._ _Lemma 5 (Solution of Problem (25)): For any β1, . . ., βN_
0, the optimal solution of Problem (25) satisfies
≥
_t_ _[RED][∗]_ =
_N_
βL maxn{ _fn[c][max][n]_ } _n�=1_ _βn >_ �max2 [�]nn[N]�=f 1n[max]cn[κ][n][c]��n[3] 3 (34)
βL�[3] 2 [�]�n[N]Nn==11 _[κ][β][n][n][c]n[3]_ _n�N=1_ _βn ≤_ �max2 [�]nn[N]�=f 1n[max]cn[κ][n][c]��n[3] 3 .
_Mn[∗]_ [≜]
0 _λ −_ _αµn ≤_ 0
� _λ3−καµnc[3]nn_ _λ −_ _αµn ∈_ �0, 3κncn(fn[max])[2][�] (27)
_fnc[max]n_ _λ −_ _αµn ≥_ 3κncn(fn[max])[2]
_and tn[MAP][∗]_ _given by_
_tn[MAP][∗]_
= 0 _ρ1,n < 0_
∈ [0, τ ] _ρ1,n = 0_ (28)
= τ _ρ1,n > 0_
_Proof: See Appendix C._
_B. MAXIMIZATION OF THE DUAL FUNCTION AND_
_INTERPRETATION_
The dual function being concave but non-differentiable,
we could now maximize it using the subgradient-based ellipsoid method, as was done for example in [33]. However,
in addition to being unpractical to solve the actual problem
(when compared to the use of a convex optimization solver
like cvxopt [40]), this method doesn’t offer any additional
insight into the structure of the optimal solution.
Instead, we intuitively look at what happens if we maximize the dual function g(µ, β, λ) taking into account the
results of Lemmas 3, 4 and 5. To ease the analysis, we start
with λ = 0 and µn = 0 for all n. In this case, ln[∗] [=][ 0]
for all devices (see (26) and (27)) and the penalty term L
−
�N
_n=1_ _[l]n[∗]_ [=][ L][ associated with][ λ][ appearing in the dual function]
is thus strictly positive. Intuitively, this implies that the task
has not been fully distributed across the devices, violating
constraint (1). It is thus possible to increase the value of the
dual function through this positive penalty term by increasing
the value of λ. Because ln[∗] [is proportional to][ √][λ][ −] _[αµ][n]_
through Mn[∗][, this increases the number of bits][ l]n[∗] [processed]
_with ρ1,n = 2κn(cnMn[∗][)][3][ −]_ _[β][n][ +][ γ][2][,][n]_ _fnc[max]n_ _[and]_
_γ2,n =_
� _n_
0 _Mn[∗]_ _[<][ f][ max]cn_ (29)
_λ −_ _αµn −_ 3κncn(fn[max])[2] _Mn[∗]_ [=][ f]n[ max]cn _[.]_
_Proof: See Appendix ??._
-----
by each device. Moreover, because ln[∗] [is also inversely pro-]
portional to �κnc[3]n through Mn[∗][, less energy-efficient devices]
(i.e., the ones with larger values of κnc[3]n[) get fewer bits to]
process. The value of λ can be increased in this way until the
penalty term L − [�]n[N]=1 _[l]n[∗]_ [equals 0 (i.e., until the task is fully]
distributed across the devices). Next, because µn = 0 for all
devices as of now, p[∗]n [=][ 0 (see (][31][)) and the penalty term]
_αln[∗]_ _n_ _rn(p[∗]n[)][ =][ α][l]n[∗]_ [associated with][ µ][n][ appearing in the]
[−] _[t]_ [SHU*]
dual function is strictly positive for all devices. Intuitively,
this implies that the rate constraint (20) is violated for all
devices. It is thus possible to increase the value of the dual
function through this penalty term by increasing the value
of µn. Increasing µn has a double effect: (i) it decreases the
value of ln[∗] [because][ l]n[∗] [is proportional to][ √][λ][ −] _[αµ][n][, and (ii)]_
it increases the value of p[∗]n [because][ p]n[∗] [is directly proportional]
to µn. Combined, these two effects work together towards
satisfying the rate constraint (20). For devices with very bad
channel or very low maximum RF transmit power p[max]n, µn
could increase so much that λ − _αµn would become negative,_
meaning that the number of bits to be processed ln[∗] [would]
fall to 0 (see (26) and (27) again). At this point, there is
an iterative interplay between λ and {µn}n[N]=1 [in which both]
successively increase to maximize the dual function until both
constraints (1) and (20) are satisfied and a maximum has been
reached.
It is now possible to give a waterfilling-like interpretation
of the structure of the optimal computing load partition, i.e.,
{ln[∗][}][N]n=1[, through the effective processing rate][ {][M]n[∗][}]n[N]=1 [given]
in (27). First, λ acts as a kind of global (i.e., across all
devices) water level for {ln[∗][}][N]n=1 [through the effective process-]
ing rate Mn[∗][. This water level has to be sufficiently high for the]
tasks to be fully executed in time. Then, αµn can be seen as
the base of the water vessel of device n. Following the above
discussion, this base αµn mainly depends on the communication capabilities and energy-efficiency of device n (i.e., hn
and p[max]n ). This is illustrated in Fig. 4. Finally, the actual water
content of each vessel, i.e., λ − _αµn is divided by 3κnc[3]n[. This]_
term, related to the computing energy-efficiency of device n
can be interpreted as a ‘‘pressure’’ applied to the water vessel
of each device. The less energy-efficient device n is, the larger
3κnc[3]n [becomes and the more pressure is applied to its water]
vessel, hence reducing the corresponding water level and ln[∗][.]
This is illustrated in Fig. 5.
**V. NUMERICAL RESULTS**
In this section, the performances of the optimal collaborativecomputing scheme (denoted Opt in what follows) are benchmarked against various other schemes through numerical
experiments. The schemes used for comparison are
- Blind: the task allocation (i.e., choosing the value of
_ln for each device n) doesn’t take into account the het-_
erogeneity of the devices; the scheme is blind to device
diversity (both in terms of computing and communicating capabilities). In this case, the variable ln is set
to L/N for each device n. This corresponds to what is
**FIGURE 4.** First part of the waterfilling-like interpretation of the optimal
effective processing rate Mn[∗] for a single device n. λ acts as a global (i.e.,
shared by all devices) water level that has to be sufficiently high for the
tasks to be fully executed in time while αµn acts as the base of the water
vessel of device n and depends on device n communication capabilities
and energy-efficiency (i.e., channel conditions and maximum RF transmit
power).
**FIGURE 5.** Second part of the waterfilling-like interpretation of the
optimal effective processing rate Mn[∗] for a single device n. Each water
vessel is ‘‘compressed’’ by an applied ‘‘pressure’’ 3κncn[3] that represents
the computing efficiency of device n. Less efficient devices will then see
their effective processing rate reduced by a factor that depends on their
computing energy-efficiency.
done in most works on CDC assuming homogeneous
devices [23], [25], [26].
- NoDFS: the CPU frequency of each device n is fixed to
its maximum value fn[max] rather than being optimized for
energy-efficiency. In this case, the variable tn[MAP] is set
to cnln/fn[max] for each device n while tn[RED] (now different
for each device) is set to cnβL/fn[max]. This scheme is
close to the one proposed in our previous work [22].
- Blind-NoDFS: this scheme combines the two previous cases. In this case, ln = L/N and tn[MAP] = _fn[max]cn_ _NL_
for each device n while tn[RED] = cnβL/fn[max]. The only
optimization left here concerns the Shuffle phase and the
variables tn[SHU] and En.
- NoOpt: in this scheme, nothing is optimized. This is
basically Blind-NoDFS with α _N[L]_ [=][ t]n[SHU]rn � _tn[SHU]En_ �
and En = tn[SHU]p[max]n .
The parameters used in the following numerical experiments are given in Table 1. The ranges for the parameters
were selected to comply with devices consumption models
used in the literature on MEC systems, e.g., [11]–[13], [33].
_A. MAXIMUM COMPUTING LOAD AND OUTAGE_
_PROBABILITY_
To show that the proposed scheme indeed enhances the computing capabilities of individual devices, we start by comparing the maximum computing load of Opt and Blind,
-----
**TABLE 1. Parameters used in the numerical experiments.**
noted Lmax[Opt] [and][ L]max[Blind][, respectively. Other schemes are not]
included here as Lmax[NoDFS] = Lmax[Opt] [and][ L]max[Blind-NoDFS] = Lmax[Blind] [=]
_Lmax[NoOpt]. For Opt, the maximum computing load Lmax[Opt]_ [can be]
readily obtained using Lemma 1. For Blind, we introduce
the following Lemma.
_Lemma 6 (Maximum Computing Load of Blind): The_
_maximum computing load achievable by the Blind scheme_
_is given by_
_fn[max]_
_cn_
_τ −_ _βL maxn_ � _fnc[max]n_
_n_ _/cn_
1 _[α][f][ max]_
+ _rn(p[max]n_ )
�
**FIGURE 6. Maximum computing loads L[Opt]max** [and][ L]max[Blind] [averaged over]
1.000.000 random instances of the problem. Note that L[Opt]max [for][ N][ =][ 10 is]
hidden by L[Blind]max [for][ N][ =][ 30][,][ 40 and 50.]
**FIGURE 7. Empirical outage probability Pout[Opt]** [and][ P]out[Blind] for L = 10 Mb,
averaged over 1.000.000 random instances of the problem.
probability is defined, for a random heterogeneous set of
devices and a given allowed latency τ, as the probability that
the maximum computing load that can be processed by the
system is lower than the actual computing load L, i.e.,
_P[∗]out_ [=][ Pr] �L ≥ _Lmax[∗]_ � _._ (36)
For a given task size L, this probability can be empirically
computed by averaging over a large number of randomly generated sets of devices. For L = 10 Mb, both P[Opt]out [and][ P]out[Blind]
are depicted in Fig. 7 as a function of the allowed latency
_τ and for several values of N_ . This plot again demonstrates
the benefits of leveraging devices diversity to distribute the
task among the devices. At the opposite, we see that Blind
suffers from devices diversity. Indeed, for larger values of the
allowed latency τ, increasing the number of devices N penalizes Blind by increasing its outage probability P[Blind]out [. Intu-]
itively, this comes from the fact that increasing the number
of devices N increases the probability of having a very weak
device limiting the whole system. Mathematically, the lower
tail of the distribution of Lmax[Blind] [grows larger and larger with]
_N_, making the distribution more and more skewed towards
small values of Lmax[Blind][. This also explains why this trend]
(35)
_[.]_
_Lmax[Blind]_ [=][ N][ min]
_Proof: Obtaining Lmax[Blind]_ [requires solving the following]
linear program
_Lmax[Blind]_ ≜ maximize
_l,t[MAP],t[SHU][ Nl]_
subject to cnl ≤ _tn[MAP]fn[max]_ ∀n
_αl ≤_ _tn[SHU]rn(p[max]n_ ) ∀n
� _cn_ �
_t_ [RED] ≥ _βL maxn_ _fn[max]_
_tn[MAP]_ + tn[SHU] ≤ _τ −_ _t_ [RED], _n ∈_ [N ]
_l, tn[MAP],_ _tn[SHU]_ ≥ 0, _n ∈_ [N ].
A reasoning similar to the one used in Lemma 1 – and
omitted here for the sake of space – can then be used to obtain
the analytical expression given above. Values of Lmax[Opt] [and][ L]max[Blind] [for different values of the allowed]
latency τ and various numbers of devices N are plotted in
Fig. 6. As expected, both Lmax[Opt] [and][ L]max[Blind] grow with the
allowed latency τ . However, Lmax[Opt] [grows with][ τ][ much faster]
than Lmax[Blind] [does. Next, one can see that increasing the number]
of devices N for a given allowed latency τ is always more
profitable for Opt than for Blind. Furthermore, the benefits
of further increasing the number of devices N remain constant
for Opt but quickly saturates for Blind. Both observations
can be explained by the fact that Opt is able to leverage
devices diversity by optimally exploiting the different computing and communicating capabilities of the devices while
Blind, as per its name, is not.
Another way of looking at the maximum computing loads
of the different schemes is through what we define as the
‘‘outage probability’’ of the system. In this context, the outage
-----
was not visible on Fig. 6 as it only shows the mean of the
distribution of Lmax[Blind][. In addition, it appears that the benefits]
on P[Blind]out of allowing a looser deadline (for a given N ) saturate
as the value of τ increases beyond a certain point that varies
with the number of devices N . Again, and for the same reason,
this trend was not visible on Fig. 6 and cannot be explained
by looking at the mean of Lmax[Blind] [only. This saturation effect]
appears when the mode of the distribution of Lmax[Blind] [becomes]
larger than the value of the actual computing load L used
to compute P[Blind]out [. Passed this point, the benefits on][ P]out[Blind]
of further pushing the mode to larger values by increasing τ
become smaller and smaller.
Coming back to the example application of collaborative on-device ML/DL inference, this indicates that Opt
enables inferences with larger ML/DL models (i.e., more
accurate/complex inferences) for the same latency, or the
other way around, similar inferences for a smaller latency.
_B. ENERGY CONSUMPTION_
We now look at the energy consumed by the different
collaborative-computing schemes. Fig. 8 depicts the total
energy consumed per bit processed for different numbers of
devices N, while Fig. 9 depicts the energy consumed by each
phase of the collaboration (i.e, Map, Shuffle and Reduce).
In both figures, L = 1 Mb, βL = 0.1 kb and τ = 100 ms.
Each point is the result of an average over 100 feasible
(for each scheme) instances of the problem, i.e., instances
for which L ≤ _Lmax[Opt]_ _[,][ L]max[Blind][. Note that the parameters have]_
also been chosen to allow comparison between the schemes,
i.e., to ensure that feasible instances arise with reasonable
probability for all schemes.
First, one can observe in Fig. 8 that the energy consumed
by both Blind-NoDFS and NoOpt is actually the same.
This stems from the fact that it is always optimal (from an
energy-efficiency point of view) for constraint (20) to be met.
Indeed, the opposite would mean that the device is investing either too much time tn[SHU] (hence increasing the energy
consumption of the communications circuits tn[SHU]P[c]n[) or too]
much RF energy En with regards to the number of bits αln that
needs to be transmitted in the Shuffle phase. For the same
reason, constraint (21) is almost always satisfied as well,
meaning that devices participating to the Shuffle phase transmit at the maximum RF power allowed, i.e., p[max]n . Schemes
Blind-NoDFS and NoOpt are thus equivalent and both
transmit at the maximum RF transmit power and at the maximum rate. These two observations are valid for all the other
schemes as well. In addition, the energy per bit consumed by
both Blind-NoDFS and NoOpt is roughly constant with
the number of devices. At the opposite, the energy consumed
by the other schemes decreases with N as diversity across
the devices is exploited for energy-efficiency. Interestingly,
optimizing {tn[MAP]}n[N]=1 [and][ t] [RED][ only (in][ Blind][) is more]
beneficial than optimizing ln only (in NoDFS), even though
the number of bits assigned to each device for processing
by Blind is uniform across the devices, and thus blind to
**FIGURE 8.** Comparison of the total energy consumed by the different
schemes as a function of the number of devices N. Note that the energy
consumption is the same for both NoOpt (yellow curve) and
Blind-NoDFS (black curve).
diversity. Combining both schemes in Opt leads to a gain in
energy-efficiency with respect to NoOpt reaching two orders
of magnitude for large values of N .
Fig. 9 breaks down the energy consumption of the different
schemes in 3 components: E [MAP], E [SHU] and E [RED]. Note
that NoOpt, being equivalent to Blind-NoDFS, has been
omitted to avoid cluttering the plot. First, it appears that the
energy consumption of the Map phase largely dominates the
energy consumption of the Shuffle and Reduce phases for
small values of N .[5] As the number of devices N increases,
this difference decreases for all schemes leveraging diversity across the devices (i.e., all but Blind-NoDFS). At the
opposite, the energy consumed by the Shuffle phase increases
with the number of devices N, no matter the scheme used.
This figure also shows that there is not much (if anything) to
gain from optimization in the Shuffle phase. Next, one can
see that the energy efficiency of the Reduce phase increases
with N when t [RED] can be optimized (i.e., when devices can
perform DFS). This decrease with N is however slower than
what we observed for the Map phase. For Blind, this can
be explained by the fact that priority in the optimization is
given to the more energy intensive Map phase. For Opt, this
comes from the fact that, at the opposite of the Map phase,
all devices have to perform the Reduce phase. Finally, for
NoDFS and Blind-NoDFS, each device has to perform the
Reduce phase at full speed causing E [RED] to increase with N .
_C. ENERGY-LATENCY TRADE-OFF_
Figs. 10 and 11 depict the total energy consumption per bit
and the energy consumed per bit by each phase, respectively,
for the different schemes and for different values of the
allowed latency τ . For both figures, L = 1 Mb, βL = 0.1 kb
and N 10. Each point is the result of an average over
=
100 feasible (for each scheme) instances of the problem,
i.e., instances for which L ≤ _Lmax[Opt]_ _[,][ L]max[Blind][. Again, note that]_
5Note that this statement is strongly dependent on the energy consumption model and the parameters used for the numerical experiments. As an
example, increasing the number of bits transmitted during the Shuffle phase,
_αln, through the total size of the intermediate computation results βL would_
directly result in an increase of E [SHU] by the same factor.
-----
**FIGURE 9.** Breakdown of the energy consumed by the three phases of
the collaboration as a function of the number of nodes N. Note that the
energy consumption for the Reduce phase is the same for both NoDFS
and Blind-NoDFS.
**FIGURE 10.** Comparison of the total energy consumed by the different
schemes as a function of the allowed latency τ . Note that the energy
consumption is the same for both NoOpt (yellow curve) and
Blind-NoDFS (black curve).
the parameters have also been chosen to allow comparison
between the schemes, i.e., to ensure that feasible instances
arise with reasonable probability for all schemes.
Interestingly, Figs. 10 and 11 closely resemble
Figs. 8 and 9, implying that the effect of increasing the
number of devices N is roughly equivalent to the effect of
increasing the allowed latency τ . The underlying mechanisms, however, are different. For schemes where devices are
able to perform DFS (i.e., Opt and Blind), increasing τ
enables the devices to further decrease their CPU frequency,
hence saving energy. For NoDFS, increasing τ enables the
system to increase the number of bits assigned to the most
energy-efficient devices, hence reducing the load on less
energy-efficient devices and again saving energy.
_D. NUMBER OF PARTICIPATING DEVICES_
Finally, Fig. 12 shows the average fraction of devices participating to the collaboration, i.e., devices with ln > 0 that thus
participate to the Map and Shuffle phases, as a function of the
computing load L/Lmax. For Blind (and Blind-NoDFS),
this fraction is of course constant and equal to 1 as ln = L/N
for all n. For Opt, this fraction starts at around 0.6 for very
small computing loads and quickly reaches 1 for computing
loads >0.2. At the opposite, for NoDFS, the fraction of
devices participating to the Map and Shuffles phases closely
**FIGURE 11.** Breakdown of the energy consumed by the three phases of
the collaboration as a function of the allowed latency τ . Note that the
energy consumption for the Reduce phase is the same for both NoDFS
and Blind-NoDFS.
**FIGURE 12.** Average fraction of devices participating to the
collaboration, i.e., devices with ln[∗] > 0 that thus participate to the Map
and Shuffle phases, as a function of the computing load L/Lmax.
follows the fraction L/Lmax. To explain these radically different behaviors, we look at the energy consumed by the Map
phase at each device n for both schemes. For Opt first, Eq. (3)
indicates that En[MAP] is a cubic function of ln. For NoDFS,
injecting tn[MAP] = cnln/fn[max] in (3) shows that En[MAP] becomes
a linear function of ln. This explains why the computing load
is more evenly spread across devices for Opt than for NoDFS.
**VI. DISCUSSION AND FUTURE WORKS**
This work built upon our previous work [22] to further highlight the benefits of leveraging devices diversity – whether
in terms of computing or communication capabilities – to
enhance individual computing capabilities of the devices
while increasing energy-efficiency of the system as a whole.
It also provides new insights on the structure of the optimal solution through a waterfilling-like interpretation. As
mentioned in the introduction, this makes collaborativecomputing another potential viable architecture to be used in
conjunction with MEC and MCC to enable ubiquitous computing on heterogeneous devices. However, further validation with more realistic and practical assumptions is needed.
Interferences between devices during the Shuffle phase, for
example, were neglected in this work. As the interference
level is expected to increase with the number of devices
participating to the Shuffle phase, taking into account interference in the communication model could have a significant
-----
impact on the number of devices participating to the collaboration. Non-causal knowledge of the uplink channels was also
assumed to allow for offline optimization of the collaboration.
To get rid of this unrealistic assumption, one could instead
consider the expectation taken over the channel gain hn of
_rn(pn) in constraint (11) for a given channel gain distribution._
Adaptation to the actual channel condition observed during
the Shuffle phase could then be performed on-the-fly by each
device. Downlink communications were also neglected in this
work. While this makes sense in a scenario where optimizing the energy-consumption of end devices is the primary
objective, care should be taken to avoid simply ignoring the
energy burden imposed to the edge of the network. On the
other hand, the system could also optimize channel and bandwidth allocation across devices, considered to be given in
this work. The Shuffle phase could also be further optimized
by integrating results from CDC [23]–[28]. These additional
degrees of freedom could enable additional energy savings
and increased system-wise performance. This would however
come at the cost of a complexified optimization problem, and
a sweet spot between optimization complexity and efficiency
gains should thus be found.
**APPENDIX A**
**PROOF OF LEMMA 3**
Problem (23) being convex, the optimal solution satisfies the
KKT conditions. The Lagrangian of problem (23) is given by
_L1,n =_ ([κ]tn[MAP][n][c]n[3][l]n)[3][2][ +][ (][αµ][n][ −] _[λ][)][l][n][ +][ β][n][t]n[MAP]_ − _γ1,nln_
+ γ2,n �ln − _[t]n[MAP]cnfn[max]_ � − _γ3,ntn[MAP]_
� �
+ γ4,n _tn[MAP]_ − _τ_ (37)
with γ1,n, γ2,n, γ3,n, γ4,n ≥ 0 the Lagrange multipliers. The
KKT conditions are then given by
_∂L1,n_ _n[l]n[2]_
= 3 _[κ][n][c][3]_ (38)
_∂ln_ (tn[MAP])[2][ +][ αµ][n][ −] _[λ][ −]_ _[γ][1][,][n][ +][ γ][2][,][n][ =][ 0]_
_∂L1,n_ _n[l]n[3]_ _fn[max]_
_∂tn[MAP]_ = −2 ([κ]tn[MAP][n][c][3] )[3][ +][ β][n][ −] _[γ][2][,][n]_ _cn_ − _γ3,n + γ4,n = 0_
(39)
_δ1,nEn = 0_ (47)
� �
_δ2,n_ _En −_ _tn[SHU]p[max]n_ = 0 (48)
_δ3,ntn[SHU]_ = 0 (49)
� �
_δ4,n_ _tn[SHU]_ − _τ_ = 0. (50)
We first obtain (30), (31) and (33) using condition (45) and
complementary slackness conditions (47) and (48). Substituting (30) in (46) and defining ρ2,n = δ4,n − _δ3,n, we then_
obtain (32) using complementary slackness conditions (49)
and (50).
**APPENDIX C**
**PROOF OF LEMMA 5**
Problem (25) being convex, the optimal solution satisfies the
KKT conditions. The Lagrangian of problem (25) is given by
�N _κnc[3]n[T][ 3]_
_L3 =_ _n=1_ (t [RED])[2][ +][ β][n][t] [RED]
� � _cnβL_ � � � �
+ ϵ1 maxn _fn[max]_ − _t_ [RED] + ϵ2 _t_ [RED] − _τ_ (51)
with ϵ1, ϵ2 ≥ 0 the Lagrange multipliers. The KKT conditions are then given by
**APPENDIX B**
**PROOF OF LEMMA 4**
Problem (24) being convex, the optimal solution satisfies the
KKT conditions. The Lagrangian of problem (24) is given by
� _En_ �
_L2,n = En + tn[SHU]P[c]n_ [−] _[µ][n][t]n[SHU]rn_ _tn[SHU]_ + βntn[SHU]
− _δ1,nEn + δ2,n(En −_ _tn[SHU]p[max]n_ ) − _δ3,ntn[SHU]_
� �
+ δ4,n _tn[SHU]_ − _τ_ (44)
with δ1,n, δ2,n, δ3,n, δ4,n ≥ 0 the Lagrange multipliers. The
KKT conditions are then given by
_∂L2,n_
_∂En_ = 1 − _δ1,n + δ2,n −_ _µn_
_hn_
_N0_ 0 (45)
1 _En_ _hn_ =
+ _tn[SHU]_ _BN0_
_En_ _hn_
_∂L2,n_ _tN[SHU]_ _N0_ � _En_ �
_∂tn[SHU]_ = µn 1 + _tn[SHU]En_ _BNhn0_ − _µnrn_ _tn[SHU]_ + P[c]n [+][ β][n]
− _δ2,np[max]n_ − _δ3,n + δ4,n = 0_ (46)
with the complementary slack conditions
with the complementary slackness conditions
_N_
�
_βn = 0_
_n=1_
(52)
�3 _N_
�
_κnc[3]n_ [+]
_n=1_
_∂L3_ � _T_
_∂t_ [RED][ =][ ϵ][2][ −] _[ϵ][1][ −]_ [2] _t_ [RED]
_ϵ1_
_γ2,n_
_γ1,nln = 0,_ (40)
�ln − _[t]n[MAP]cnfn[max]_ � = 0, (41)
_γ3,ntn[MAP]_ = 0, (42)
� �
_γ4,n_ _tn[MAP]_ − _τ_ = 0. (43)
with the complementary slackness conditions
� _cnβL_
_fn[max]_ − =
� �
_ϵ2_ _t_ [RED] − _τ_ = 0. (54)
� �
_t_ [RED] 0 (53)
− =
�
max
_n_
We first obtain (26), (27) and (29) using condition (38) and
complementary slackness conditions (40) and (41). Substituting (26) in (39) and defining ρ1,n = γ4,n − _γ3,n, we then_
obtain (28) using complementary slackness conditions (42)
and (43).
Condition (52) together with complementary slackness
conditions (53) (54) allow us to obtain (34).
-----
**ACKNOWLEDGMENT**
The authors would also like to thank their colleague Emre
Kilcioglu for proofreading and comments, and the anonymous reviewers for their constructive criticism.
**REFERENCES**
[1] M. Chiang and T. Zhang, ‘‘Fog and IoT: An overview of research opportunities,’’ IEEE Internet Things J., vol. 3, no. 6, pp. 854–864, Dec. 2016.
[2] T. Qiu, N. Chen, K. Li, M. Atiquzzaman, and W. Zhao, ‘‘How can heterogeneous Internet of Things build our future: A survey,’’ IEEE Commun.
_Surveys Tuts., vol. 20, no. 3, pp. 2011–2027, 3rd Quart., 2018._
[3] B. M. Rashma, S. Macherla, A. Jaiswal, and G. Poornima, ‘‘Handling
heterogeneity in an IoT infrastructure,’’ in Advances in Machine Learning
_and Computational Intelligence, S. Patnaik, X.-S. Yang, and I. K. Sethi,_
Eds. Singapore: Springer, 2021, pp. 635–643.
[4] S. Kumar, P. Tiwari, and M. Zymbler, ‘‘Internet of Things is a revolutionary
approach for future technology enhancement: A review,’’ J. Big Data,
vol. 6, no. 1, pp. 1–21, Dec. 2019.
[5] A. Mammela and A. Anttonen, ‘‘Why will computing power need particular attention in future wireless devices?’’ IEEE Circuits Syst. Mag., vol. 17,
no. 1, pp. 12–26, 1st Quart., 2017.
[6] F. Pereira, R. Correia, P. Pinho, S. I. Lopes, and N. B. Carvalho, ‘‘Challenges in resource-constrained IoT devices: Energy and communication as
critical success factors for future IoT deployment,’’ Sensors, vol. 20, no. 22,
p. 6420, Nov. 2020.
[7] S. Barbarossa, S. Sardellitti, and P. D. Lorenzo, ‘‘Communicating while
computing: Distributed mobile cloud computing over 5G heterogeneous
networks,’’ IEEE Signal Process. Mag., vol. 31, no. 6, pp. 45–55,
Nov. 2014.
[8] H. T. Dinh, C. Lee, D. Niyato, and P. Wang, ‘‘A survey of mobile cloud
computing: Architecture, applications, and approaches,’’ Wireless Com_mun. Mobile Comput., vol. 13, no. 18, pp. 1587–1611, Dec. 2013._
[9] C. Arun and K. Prabu, ‘‘Applications of mobile cloud computing: A survey,’’ in Proc. Int. Conf. Intell. Comput. Control Syst. (ICICCS), 2017,
pp. 1037–1041.
[10] Y. C. Hu, M. Patel, D. Sabella, N. Sprecher, and V. Young, ‘‘Mobile edge
computing—A key technology towards 5G,’’ ETSI White Paper, vol. 11,
no. 11, pp. 1–16, 2015.
[11] Y. Mao, C. You, J. Zhang, K. Huang, and K. B. Letaief, ‘‘A survey on
mobile edge computing: The communication perspective,’’ IEEE Commun.
_Surveys Tuts., vol. 19, no. 4, pp. 2322–2358, 4th Quart., 2017._
[12] P. Mach and Z. Becvar, ‘‘Mobile edge computing: A survey on architecture
and computation offloading,’’ IEEE Commun. Surveys Tuts., vol. 19, no. 3,
pp. 1628–1656, 3rd Quart., 2017.
[13] A. Filali, A. Abouaomar, S. Cherkaoui, A. Kobbane, and M. Guizani,
‘‘Multi-access edge computing: A survey,’’ IEEE Access, vol. 8,
pp. 197017–197046, 2020.
[14] L. M. Vaquero and L. Rodero-Merino, ‘‘Finding your way in the fog:
Towards a comprehensive definition of fog computing,’’ ACM SIGCOMM
_Comput. Commun. Rev., vol. 44, no. 5, pp. 27–32, Oct. 2014._
[15] Y. Shi, K. Yang, T. Jiang, J. Zhang, and K. B. Letaief, ‘‘Communicationefficient edge AI: Algorithms and systems,’’ IEEE Commun. Surveys Tuts.,
vol. 22, no. 4, pp. 2167–2191, Jul. 2020.
[16] E. El Haber, T. M. Nguyen, and C. Assi, ‘‘Joint optimization of computational cost and devices energy for task offloading in multi-tier edgeclouds,’’ IEEE Trans. Commun., vol. 67, no. 5, pp. 3407–3421, May 2019.
[17] C. Zhang, P. Patras, and H. Haddadi, ‘‘Deep learning in mobile and wireless
networking: A survey,’’ IEEE Commun. Surveys Tuts., vol. 21, no. 3,
pp. 2224–2287, 3rd Quart., 2019.
[18] R. Stahl, Z. Zhao, D. Mueller-Gritschneder, A. Gerstlauer, and
U. Schlichtmann, ‘‘Fully distributed deep learning inference on resourceconstrained edge devices,’’ in Embedded Computer Systems: Architectures,
_Modeling, and Simulation, D. Pnevmatikatos, M. Pelcat, and M. Jung,_
Eds. Cham, Switzerland: Springer, 2019, pp. 77–90.
[19] N. D. Lane, S. Bhattacharya, A. Mathur, P. Georgiev, C. Forlivesi, and
F. Kawsar, ‘‘Squeezing deep learning into mobile and embedded devices,’’
_IEEE Pervas. Comput., vol. 16, no. 3, pp. 82–88, Jul. 2017._
[20] K. B. Letaief, W. Chen, Y. Shi, J. Zhang, and Y.-J.-A. Zhang, ‘‘The roadmap
to 6G: AI empowered wireless networks,’’ IEEE Commun. Mag., vol. 57,
no. 8, pp. 84–90, Aug. 2019.
[21] J. Park, S. Samarakoon, A. Elgabli, J. Kim, M. Bennis, S.-L. Kim,
and M. Debbah, ‘‘Communication-efficient and distributed learning over
wireless networks: Principles and applications,’’ 2020, arXiv:2008.02608.
[Online]. Available: http://arxiv.org/abs/2008.02608
[22] A. Paris, H. Mirghasemi, I. Stupia, and L. Vandendorpe, ‘‘Energy-efficient
edge-facilitated wireless collaborative computing using map-reduce,’’ in
_Proc. IEEE 20th Int. Workshop Signal Process. Adv. Wireless Commun._
_(SPAWC), Jul. 2019, pp. 1–5._
[23] S. Li, Q. Yu, M. A. Maddah-Ali, and A. S. Avestimehr, ‘‘A scalable
framework for wireless distributed computing,’’ IEEE/ACM Trans. Netw.,
vol. 25, no. 5, pp. 2643–2654, Oct. 2017.
[24] M. Kiamari, C. Wang, and A. S. Avestimehr, ‘‘Coding for edge-facilitated
wireless distributed computing with heterogeneous users,’’ in Proc. 51st
_Asilomar Conf. Signals, Syst., Comput., Oct. 2017, pp. 536–540._
[25] S. Li, M. A. Maddah-Ali, Q. Yu, and A. S. Avestimehr, ‘‘A fundamental
tradeoff between computation and communication in distributed computing,’’ IEEE Trans. Inf. Theory, vol. 64, no. 1, pp. 109–128, Jan. 2018.
[26] F. Li, J. Chen, and Z. Wang, ‘‘Wireless MapReduce distributed
computing,’’ 2018, arXiv:1802.00894. [Online]. Available: http://arxiv.
org/abs/1802.00894
[27] S. Li, M. A. Maddah-Ali, and A. S. Avestimehr, ‘‘Coding for distributed fog computing,’’ IEEE Commun. Mag., vol. 55, no. 4, pp. 34–40,
Apr. 2017.
[28] F. Xu and M. Tao, ‘‘Heterogeneous coded distributed computing: Joint
design of file allocation and function assignment,’’ in Proc. IEEE Global
_Commun. Conf. (GLOBECOM), Dec. 2019, pp. 1–6._
[29] J. Dean and S. Ghemawat, ‘‘MapReduce: Simplified data processing on
large clusters,’’ in Proc. 6th Symp. Operating Syst. Design Implement.
_(OSDI), San Francisco, CA, USA, 2004, pp. 137–150._
[30] S. Chen and W. S. Schlosser, ‘‘Map-reduce meets wider varieties of applications,’’ Intel, Pittsburgh, PA, USA, Tech. Rep. IRP-TR-08-05, 2008.
[31] R. P. Elespuru, S. Shakya, and S. Mishra, ‘‘Mapreduce system over heterogeneous mobile devices,’’ in Software Technologies for Embedded and
_Ubiquitous Systems, S. Lee and P. Narasimhan, Eds. Berlin, Germany:_
Springer, 2009, pp. 168–179.
[32] A. Dou, V. Kalogeraki, D. Gunopulos, T. Mielikainen, and V. H. Tuulos,
‘‘Misco: A MapReduce framework for mobile systems,’’ in Proc. 3rd Int.
_Conf. Pervas. Technol. Rel. Assistive Environ. (PETRA). New York, NY,_
USA: ACM, 2010, pp. 1–8.
[33] X. Cao, F. Wang, J. Xu, R. Zhang, and S. Cui, ‘‘Joint computation and
communication cooperation for energy-efficient mobile edge computing,’’
_IEEE Internet Things J., vol. 6, no. 3, pp. 4188–4200, Jun. 2019._
[34] C. You and K. Huang, ‘‘Exploiting non-causal CPU-state information for
energy-efficient mobile cooperative computing,’’ IEEE Trans. Wireless
_Commun., vol. 17, no. 6, pp. 4104–4117, Jun. 2018._
[35] Z. Sheng, C. Mahapatra, V. C. M. Leung, M. Chen, and P. K. Sahu, ‘‘Energy
efficient cooperative computing in mobile wireless sensor networks,’’
_IEEE Trans. Cloud Comput., vol. 6, no. 1, pp. 114–126, Jan. 2018._
[36] D. Wu, F. Wang, X. Cao, and J. Xu, ‘‘Wireless powered user
cooperative computation in mobile edge computing systems,’’ 2018,
_arXiv:1809.01430. [Online]. Available: http://arxiv.org/abs/1809.01430_
[37] A. Mtibaa, A. Fahim, K. A. Harras, and M. H. Ammar, ‘‘Towards resource
sharing in mobile device clouds: Power balancing across mobile devices,’’
_SIGCOMM Comput. Commun. Rev., vol. 43, no. 4, pp. 51–56, Aug. 2013._
[38] L. Pu, X. Chen, J. Xu, and X. Fu, ‘‘D2D fogging: An energy-efficient and
incentive-aware task offloading framework via network-assisted D2D collaboration,’’ IEEE J. Sel. Areas Commun., vol. 34, no. 12, pp. 3887–3901,
Dec. 2016.
[39] K. Yang, Y. Shi, and Z. Ding, ‘‘Low-rank optimization for data shuffling in
wireless distributed computing,’’ in Proc. IEEE Int. Conf. Acoust., Speech
_Signal Process. (ICASSP), Apr. 2018, pp. 6343–6347._
[40] M. S. Andersen, J. Dahl, and L. Vandenberghe. CVXOPT: A Python
_Package for Convex Optimization._
ANTOINE PARIS (Member, IEEE) received the
B.Sc. and M.Sc. degrees in electrical engineering
from UCLouvain, Louvain-la-Neuve, Belgium,
in 2016 and 2018, respectively. He is currently
a F.R.S.-FNRS Research Fellow at the Institute
of Information and Communication Technologies,
Electronics and Applied Mathematics (ICTEAM),
University of Louvain. His research interests
include collaborative computing, fog computing,
and wireless sensors networks.
-----
HAMED MIRGHASEMI (Member, IEEE)
received the B.Sc. and M.Sc. degrees in electrical
engineering from the Sharif University of Technology, Tehran, Iran, in 2006 and 2009, respectively,
and the Ph.D. degree from the Telecom ParisTech,
Paris, France, in 2014. He is currently a Postdoctoral Researcher with UCLouvain, Louvainla-Neuve, Belgium. His research interests include
information theory, stochastic optimization, and
deep learning.
IVAN STUPIA (Member, IEEE) received the
Ph.D. degree from the University of Pisa, Italy,
in 2009. In 2011, he joined the Institute of
Information and Communication Technologies,
Electronics and Applied Mathematics (ICTEAM),
University of Louvain. He was involved in various European and national projects on wireless
communications in different fields of application,
such as cellular systems, wireless sensor networks,
security, and aeronautical communications. His
academic experience is corroborated by more than 50 publications in international journals and proceedings of international conferences. His general and
research interests include the areas of wireless communications and signal
processing with special emphasis on application of advanced mathematical
tools to the design of self-adaptive/self-organizing wireless networks, energy
harvesting and wireless power transfer for the Internet of Things (IoT)
services, and green and cost-effective design of wireless networks.
LUC VANDENDORPE (Fellow, IEEE) was born
in Mouscron, Belgium, in 1962. He received the
degree (summa cum laude) in electrical engineering and the Ph.D. degree in applied science from
the Catholic University of Louvain (UCLouvain),
Louvain La Neuve, Belgium, in 1985 and 1991,
respectively. Since 1985, he has been with the
Communications and Remote Sensing Laboratory,
UCL, where he first worked in the field of bit rate
reduction techniques for video coding. In 1992,
he was a Visiting Scientist and a Research Fellow with the Telecommunications and Traffic Control Systems Group, Delft University of Technology,
The Netherlands, where he worked on spread spectrum techniques for personal communications systems. From October 1992 to August 1997, he was
a Senior Research Associate with the Belgian NSF, UCL, and an invited
Assistant Professor. He is currently a Full Professor with the Institute for
Information and Communication Technologies, Electronics, and Applied
Mathematics, UCLouvain. His research interests include digital communication systems and more precisely resource allocation for OFDM(A)based multicell systems, MIMO and distributed MIMO, sensor networks,
UWB-based positioning, and wireless power transfer. He is or has been a
TPC Member for numerous IEEE conferences, such as VTC, GLOBECOM,
SPAWC, ICC, PIMRC, and WCNC. He was an Elected Member of the Signal
Processing for Communications Committee, from 2000 to 2005, and the
Sensor Array and Multichannel Signal Processing Committee of the Signal
Processing Society, from 2006 to 2008 and from 2009 to 2011. He was the
Chair of the IEEE Benelux Joint Chapter on communications and vehicular
technology, from 1999 to 2003. He was the Co-Technical Chair for IEEE
ICASSP 2006. He served as an Editor for synchronization and equalization of
IEEE TRANSACTIONS ON COMMUNICATIONS, from 2000 to 2002, and an Associate
Editor for IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, from 2003 to
2005, and IEEE TRANSACTIONS ON SIGNAL PROCESSING, from 2004 to 2006.
HAMED
deep learning.
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.1109/ACCESS.2021.3094888?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/ACCESS.2021.3094888, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBYNCND",
"status": "GOLD",
"url": "https://ieeexplore.ieee.org/ielx7/6287639/9312710/09475030.pdf"
}
| 2,020
|
[
"JournalArticle"
] | true
| 2020-03-31T00:00:00
|
[
{
"paperId": "41703df035d404823d4d6833030c0111c528309e",
"title": "Challenges in Resource-Constrained IoT Devices: Energy and Communication as Critical Success Factors for Future IoT Deployment"
},
{
"paperId": "186b8261221e8a87af1bad44fcb7c3e159be3e0c",
"title": "Communication-Efficient and Distributed Learning Over Wireless Networks: Principles and Applications"
},
{
"paperId": "47a8355e76c3675c481f135cce3a5911c74aeac3",
"title": "Communication-Efficient Edge AI: Algorithms and Systems"
},
{
"paperId": "3edf3fb8f8296374ed815a6f742ec79bb12379b0",
"title": "Internet of Things is a revolutionary approach for future technology enhancement: a review"
},
{
"paperId": "0e09d91a2a2f2745cd1161ab4c03b61dbd8b72ae",
"title": "Heterogeneous Coded Distributed Computing: Joint Design of File Allocation and Function Assignment"
},
{
"paperId": "11650dd5388f32372cdf30b6907f73696beb52bc",
"title": "The Roadmap to 6G: AI Empowered Wireless Networks"
},
{
"paperId": "e34825220509aa620fcee5d467ae599522747a67",
"title": "Fully Distributed Deep Learning Inference on Resource-Constrained Edge Devices"
},
{
"paperId": "cdc609c16afd59c7a60c23e580d13fdb7d92a24f",
"title": "Energy-Efficient Edge-Facilitated Wireless Collaborative Computing using Map-Reduce"
},
{
"paperId": "ab480bcce180003181726e8b8ee54102171f63eb",
"title": "Joint Optimization of Computational Cost and Devices Energy for Task Offloading in Multi-Tier Edge-Clouds"
},
{
"paperId": "9e568a6b04a4381999a606a0ab542ee726a2336b",
"title": "Wireless Powered User Cooperative Computation in Mobile Edge Computing Systems"
},
{
"paperId": "cdb2a32822a21dba84568038a817e71979036766",
"title": "Joint Computation and Communication Cooperation for Energy-Efficient Mobile Edge Computing"
},
{
"paperId": "4fc5a04ffa85a9be13779b2f8bfbe5298189adbd",
"title": "Low-Rank Optimization for Data Shuffling in Wireless Distributed Computing"
},
{
"paperId": "8a5d0579590465494c9aba58a857af43b190b6a6",
"title": "Deep Learning in Mobile and Wireless Networking: A Survey"
},
{
"paperId": "3c1d360b06f1273163985e72ab9a45bae36cee4a",
"title": "How Can Heterogeneous Internet of Things Build Our Future: A Survey"
},
{
"paperId": "318a48935f0036db36520ce1c1f78798ee9793af",
"title": "Wireless MapReduce Distributed Computing"
},
{
"paperId": "5e00a8bd1bb34c809688b3ec76975046b15e45a1",
"title": "Coding for edge-facilitated wireless distributed computing with heterogeneous users"
},
{
"paperId": "d53533b4f24504ecb04612a0d08c46db966caa2f",
"title": "Squeezing Deep Learning into Mobile and Embedded Devices"
},
{
"paperId": "6e2ecd04dd799b2b0cef1627eeab0347d320d120",
"title": "Applications of mobile cloud computing: A survey"
},
{
"paperId": "90b31e90d4b7248580a4d41c1a504a16b3834702",
"title": "Exploiting Non-Causal CPU-State Information for Energy-Efficient Mobile Cooperative Computing"
},
{
"paperId": "51aa27a83c888718e30374369ac9cd54927dd0e0",
"title": "Coding for Distributed Fog Computing"
},
{
"paperId": "afd9dadac8d3354615e26b2038887ecdbc9e33e5",
"title": "Mobile Edge Computing: A Survey on Architecture and Computation Offloading"
},
{
"paperId": "872734bfa9f84d4e06b16911ba9ca4b9456d7a52",
"title": "Why Will Computing Power Need Particular Attention in Future Wireless Devices?"
},
{
"paperId": "8c5293da3ad1a463cb9694edfbf1bf19b8cbd698",
"title": "A Survey on Mobile Edge Computing: The Communication Perspective"
},
{
"paperId": "c141e32b6b3b52583cc78091550548c7a6222f8f",
"title": "D2D Fogging: An Energy-Efficient and Incentive-Aware Task Offloading Framework via Network-assisted D2D Collaboration"
},
{
"paperId": "a93e7aa3705a6b02306dfd113274ff2c006497d1",
"title": "Caching at the wireless edge: design aspects, challenges, and future directions"
},
{
"paperId": "4562404f32057d884a1a04e132513bba0871ce87",
"title": "A Scalable Framework for Wireless Distributed Computing"
},
{
"paperId": "aa245959d84734f92d1e8f179417eb7226868e62",
"title": "Fog and IoT: An Overview of Research Opportunities"
},
{
"paperId": "d19059600752f284f5f715675cb401ecc3ca7879",
"title": "A Fundamental Tradeoff Between Computation and Communication in Distributed Computing"
},
{
"paperId": "d588c79ff7a6dd0377f44ca78534869a9a5b2c5d",
"title": "Communicating While Computing: Distributed mobile cloud computing over 5G heterogeneous networks"
},
{
"paperId": "ede4ffff2968c84ea26bf64f6f26f670b8ab3824",
"title": "Finding your Way in the Fog: Towards a Comprehensive Definition of Fog Computing"
},
{
"paperId": "6da3d71dc601fd9cd6b4e84bc947de5474c5873b",
"title": "A survey of mobile cloud computing: architecture, applications, and approaches"
},
{
"paperId": "329ec922c147bd64ec3cb8d06261d9ca800b0cdc",
"title": "Towards resource sharing in mobile device clouds: power balancing across mobile devices"
},
{
"paperId": "f72847ed440f0ce372880066b24544dab3d6dbe5",
"title": "Misco: a MapReduce framework for mobile systems"
},
{
"paperId": "54c34971dc27ab3bb03bf4adaba182b71cc72125",
"title": "MapReduce System over Heterogeneous Mobile Devices"
},
{
"paperId": "c35bc2704272dda5fe58f53fe2c134839016d6d9",
"title": "Handling Heterogeneity in an IoT Infrastructure"
},
{
"paperId": "066afd1ed7940d794ae19e62645ca66053ce6678",
"title": "Energy Efficient Cooperative Computing in Mobile Wireless Sensor Networks"
},
{
"paperId": "a8e0b0b2fd084b9c42438fc3c9dbdf497fc4ea66",
"title": "Multi-access Edge Computing: A Survey"
},
{
"paperId": null,
"title": "-FNRS Research Fellow at the Institute of Information and Communication Technologies"
},
{
"paperId": null,
"title": "‘‘Mobile edge computing—A key technology towards 5G,’’"
},
{
"paperId": "627be67feb084f1266cfc36e5aed3c3e7e6ce5f0",
"title": "MapReduce: simplified data processing on large clusters"
},
{
"paperId": null,
"title": "Map-Reduce Meets Wider Varieties of Applications"
},
{
"paperId": "3239ccaf508fa33cb22439e89cb746a55e395ee3",
"title": "Software Technologies for Embedded and Ubiquitous Systems"
},
{
"paperId": null,
"title": "CVXOPT: A Python Package for Convex Optimization"
}
] | 22,289
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/026730ead96211ee8e954e9df5c220191fa2d329
|
[
"Computer Science"
] | 0.883916
|
Comparative Performance of Machine Learning Algorithms for Cryptocurrency Forecasting
|
026730ead96211ee8e954e9df5c220191fa2d329
|
Indonesian Journal of Electrical Engineering and Computer Science
|
[
{
"authorId": "101995247",
"name": "Nor Azizah Hitam"
},
{
"authorId": "2717236",
"name": "A. R. Ismail"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"Indones J Electr Eng Comput Sci"
],
"alternate_urls": null,
"id": "bb21160f-0f31-4e34-abab-715b95a870a2",
"issn": "2502-4752",
"name": "Indonesian Journal of Electrical Engineering and Computer Science",
"type": "journal",
"url": "http://www.iaescore.com/journals/index.php/IJEECS"
}
|
Machine Learning is part of Artificial Intelligence that has the ability to make future forecastings based on the previous experience. Methods has been proposed to construct models including machine learning algorithms such as Neural Networks (NN), Support Vector Machines (SVM) and Deep Learning. This paper presents a comparative performance of Machine Learning algorithms for cryptocurrency forecasting. Specifically, this paper concentrates on forecasting of time series data. SVM has several advantages over the other models in forecasting, and previous research revealed that SVM provides a result that is almost or close to actual result yet also improve the accuracy of the result itself. However, recent research has showed that due to small range of samples and data manipulation by inadequate evidence and professional analyzers, overall status and accuracy rate of the forecasting needs to be improved in further studies. Thus, advanced research on the accuracy rate of the forecasted price has to be done.
|
**Indonesian Journal of Electrical Engineering and Computer Science**
Vol. 11, No. 3, September 2018, pp.1121~ 1128
ISSN: 2502-4752, DOI: 10.11591/ijeecs.v11.i3.pp1121-1128 1121
# Comparative Performance of Machine Learning Algorithms for
Cryptocurrency Forecasting
**Nor Azizah Hitam, Amelia Ritahani Ismail**
Department of Computer Science, International Islamic University Malaysia (IIUM),
Kuala Lumpur, Malaysia
**Article Info** **ABSTRACT**
**_Article history:_**
Received May 28, 2018
Revised Jun 5, 2018
Accepted Jun 11, 2018
**_Keywords:_**
Artificial Intelligence
Machine Learning
Support Vector Machines
Neural Networks
Deep Learning
**_Corresponding Author:_**
Machine Learning is part of Artificial Intelligence that has the ability to
make future forecastings based on the previous experience. Methods has
been proposed to construct models including machine learning algorithms
such as Neural Networks (NN), Support Vector Machines (SVM) and Deep
Learning. This paper presents a comparative performance of Machine
Learning algorithms for cryptocurrency forecasting. Specifically, this paper
concentrates on forecasting of time series data. SVM has several advantages
over the other models in forecasting, and previous research revealed that
SVM provides a result that is almost or close to actual result yet also improve
the accuracy of the result itself. However, recent research has showed that
due to small range of samples and data manipulation by inadequate evidence
and professional analyzers, overall status and accuracy rate of the forecasting
needs to be improved in further studies. Thus, advanced research on the
accuracy rate of the forecasted price has to be done.
_Copyright © 2018 Institute of Advanced Engineering and Science._
_All rights reserved._
Amelia Ritahani Ismail
Department of Computer Science,
International Islamic University Malaysia (IIUM),
Kuala Lumpur, Malaysia.
E-mail: amelia@iium.edu.my
**1.** **INTRODUCTION**
Forecasting future values or price of experimental time series plays a vital role in almost all fields of
studies including economics, science and engineering, finance, business, meteorology and
telecommunication [1]. Cryptocurrency, an alternative medium of exchange consisting of over 1441 (as of
January 2018) decentralized crypto coin types. Relating machine learning algorithms to cryptocurrency is
considered as a new field with limited research studies. In general, system can be used to any directive
machine learning problem, in return the system will provide a description relevant to samples both in and out
of the dataset.
There are numerous type of cryptocurrency including Bitcoin, Litecoin, Ethereum, Nem, Ripple,
Iota, Stellar and others. The cryptographic foundation of each crypto coin makes them vital. Considering the
exchange rates of cryptocurrencies are notorious for being volatile, we attempt to model an algorithm that
can be used in trading of numerous cryptocurrencies. In order to show the accuracy rate of the predicted price
of the proposed methodology, two different data are used as explanatory examples. The comparative
cryptocurrencies are Litecoin and Ethereum, Bitcoin, Stellar, Ripple and Nem. This paper uses the mean
absolute percentage error (MAPE) calculation to evaluate the proposed models.
The outline of this paper is as follows. Section 1 introduces some basic notions of cryptocurrencies
and machine learning algorithms. Section 2 discusses the type of cryptocurrency and two largest alternative
blockchain technologies, Litecoin (LTC) and Ethereum (XRP) and the purposes of each development.
Section 3 presents about machine learning algorithms and three most widely used algorithms, Artificial
**_Journal homepage: http://iaescore.com/journals/index.php/ijeecs_**
-----
1122
ISSN: 2502-4752
Neural Networks (ANN) and Support Vector Machines (SVM) and Deep Learning. Section 4 explains the
experiments and results of experiments using all models.
**1.1.** **Cryptocurrency**
Litecoin (LTC) and Ethereum (XRP) are among the largest alternative blockchain technologies,
known as altcoins and were invented after Bitcoin (BTC). Altcoins may have different purposes of
development but are using general methodology based on decentralized P2P network, with the assumption of
no network failure and no Internet interruption [2-5]. Research on the cryptocurrency field is still limited.
Mostly, research in this field is focusing on a single cryptocurrency rather than broader areas such as
technological advancement, government participation in market regulations as well as market
development [6]. This section will focus on six types of cryptocurrency begins with Bitcoin, Ethereum,
Litecoin, Nem, Ripple followed by Stellar. In the succeeding section, we focus the review of previous studies
on Machine Learning, Support Vector Machines (SVM), Artificial Neural Networks (ANNs) and Deep
Learning applied in forecasting.
A peer to peer (p2p) payment cash system, non regulated digical currency and introduced in 2008
with no legal status tendered is known as Bitcoin. It is called as one type of cryptocurrencies with its
cryptographic function in its security of creation and money transfer. In recent years, bitcoin turns out to be
the most well known currency in the area of volume trading, thus makes a Bitcoin as the most potential
financial medium for investors [7]. It locks the transaction as the individualities of the sender, receiver and
the volume of transaction are all encrypted [6].
Ethereum (XRP) is a decentralized block-chain based technology that runs Turing-complete to build
and execute smart contracts or circulated systems [8-9]. The value of its coin is called ether. It was
introduced by Vitalik Buterin in 2013 and funded a year later amounted US$18 million worth of bitcoins,
raised through online public crowd sale [8]. Ether has no boundaries on its circulation, can be traded in
cryptocurrency exchanges, not to be one of the payment system but it‘s intention is merely to be used in the
Ethereum network [1, 9].
Litecoin (LTC) was released in October 2011 using a similar technology to Bitcoin, and invented by
Charles Lee. The block generation time is decreased as much as 4 times per block (from 10 minutes to 2.5
minutes per block) 84 million of maximum limit, it is equivalent to 4 times higher than Bitcoin and has
adopted a different hashing algorithm [9-10]. Litecoin is considered as the ‗silver standard‘ of crypto coin
and turn into a second most accepted by both miners and exchanges [9]. It uses Scrypt encryption algorithm
and contradicts to SHA-256 and developed to bid the Bitcoin network transaction confirmation speed and
uses an algorithm that was resilient to the advancement of hardware mining technologies.
NEM is a blockchain notarization also known as a peer-to-peer platform that provides services like
online payment and messaging system. Having a conjointly owned notarization, it then makes NEM to
become as the first public/private blockchain combination [8].
Ripple, an open source digital currency, produced by Jed McCaleb and partner, Chris Larsen, a
distributed peer-to-peer network payment medium controlled and managed by a single organization and
offers another medium of security mechanism [6, 8]. The development of Ripple is based on Byzantine
Consensus Protocol and maximum number of Ripple is 100 million [8].
Stellar, like Ripple offers and entire substitute of security instrument and implemented based on
Byzantine Consensus Protocol. Stellar has implemented a new technology to process the financial
transactions including open source, scattered and unlimited ownership [6, 11].
**1.2.** **Machine Learning**
To succeed on trading, mastering analysis is very important. Future value can be analyzed in two
different ways, technical analysis and fundamental analysis. Technical analysis uses trading information from
the market information, such as price, trading volume to forecast future price while other uses the
information outside the market like economic situation, interest rate and geopolitical issues to forecast future
direction [11]. Many investors focus on technical while some focus fundamental. However, there are some
investors who focus on overlaps between fundamental as well as technical. This paper will present about
technical analysis by applying the machine learning algorithms. Machine learning has been established as a
serious model in classical statistics in the forecasting world for over more than two decades [1], [12]. Two
most widely used algorithms for forecasting price movement are known as Artificial Neural Networks
(ANNs) and Support Vector Machine (SVM) and both has own patterns of learning [11, 13]. ANNs has been
widely used for prediction in securities. Number of issues in ANNs has been discussed by researchers
including the selection of parameters and training set [14]. According to [1], the embedding formu lation
recommends that when a historical dataset S is available, the one-step forecasting can be considered as
supervised learning. Supervised learning is the task of deriving a function from training data consist of a set
Indonesian J Elec Eng & Comp Sci, Vol. 11, No. 3, September 2018 : 1121 – 1128
-----
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752
1123
of training dataset. It comes in a set of input and output variables that is also considered as dependent on the
inputs. One-step forecasting can be applied when a mapping model is exist [1]. In one-step forecasting, the
previous values of the series, n are available, thus forecasting can be performed as a generic regression
problem as Figure 1. General approach to model an input/output sense, relies on the accessibility of
experimental pairs and denoted as training set. Training set is initiated by the historical series S by creating
the [(N – n -1) x n] input data matrix.
In one step forecasting, the approximator ˆf returns the prediction of the value of the time series at
time t + 1 as a function of the n previous values (the rectangular box containing z-1 represents a unit delay
operator, i.e., yt-1 = z-1 yt) [1].
And the [(N – n 1) x 1] output vector
(1)
For the sake of simplicity, a is assume as d = 0 lag time. Henceforth, in this chapter we will refer to
the ith row of X, which is essentially a temporal pattern of the series, as to the (reconstructed) state of the
series at time t – i + 1.
Figure 1. Proposed Methodology
**1.2.1.** **Support Vector Machine (SVM)**
Support Vector Machine (SVM) method or classifier was introduced as an induction principle that
can avoid over-fitting the data at the assimilation of the training dataset [15] and is known as the most
flexible technique to construct the explicit and accurate boundaries [16], [17]. SVM works very well in
various applications, provide fast training result and easy to use [18]. Eventually, SVM has been invented to
answer pattern recognition problems to fault diagnosis problems [15, 19]. It gives nonlinear and solid
solution by applying kernel functions to map the input space into a higher dimensional feature [20]. There are
many benefits of the SVM including outperforms in generalization model and perform well with small
datasets. SVM creates a lot of benefits in many fields including pattern classification problem [14]. Besides,
SVM is to produce a classification hyper-plane that differentiate two classes of data with maximum margin.
Standard SVM model is as follows:
_Comparative Performance of Machine Learning Algorithms for Cryptocurrency… (Amelia Ritahani Ismail)_
-----
1124 ISSN: 2502-4752
(2)
Another important point of discussion is the options offered by type of SVM. SVM offers linear and
nonlinear type of models. Linear SVMs outperforms the nonlinear in terms of speed and execution time, but
underperform dealing with complex datasets contains many training examples but less features. While
nonlinear SVMs although losing its explanatory power, seems to perform steadily across various problems,
and becomes most preferred choice compared to linear SVMs [18].
**1.2.2.** **Artificial Neural Networks (ANNs)**
A common neural network that is doing the deep learning at its hidden layers is called an artificial
neural networks [21]. Standard ANNs comprises of input layer, hidden layers and output layer [22]. It is an
extremely similar system consisting interrelated and interacting processing nodes or neurons [23, 23], works
like a human brain and process the information by interacting with a numbers of straightforward processing
features [23]. There are input and output neurons in this environment where input neurons will be triggered
upon instruments sensing the environment. While other neurons trigger through weighted connections from
neurons which was activated earlier, some neurons could effect the environment by activating actions [24].
Depending on the issue and how neurons are linked, such behavior may need a long connecting chains of
computational phases where each phase revises the aggregate activation of the network.
**1.2.3.** **Deep Learning (DL)**
Deep Learning is considered as a diverse methods in neural networks [25] and primarily to get the
most precise result across many phases, as shown in Table 1 [24]. DL is capable to produce influencing
results based on multiple layer extraction [25]. Models explained in this section applies a non-linear function
on the hidden units and enables a more lavish model that is capable to learn more abstract illustrations to
form a deep network when modules are arranged on top of each other [26]. The goal of deep network is to
design structures at the lower layers that will separate the variation factors in the input data ad chain the
representations at the higher layers, but the drawbacks of the training with multiple hidden layer units lies in
the event of the error signal being backpropagated [26].
Table 1. Variable Description
Variable Description
Open Price The first price of a given cryptocurrency in a daily trading
Close Price The price of the last transaction for a given cryptocurrency at the end of a daily trading
High Price The highest price that was paid for a cryptocurrency during a daily trading
Low Price The lowest price of a cryptocurrency reached in a daily trading
**2.** **PROPOSED METHODOLOGY**
In this paper, we consider time series data based on 5 years of daily history, as inputs for all models
and may vary based on the availability of datasets from the source. The data is prepared from daily open,
close, high and low price of a daily trading for all total of six types of cryptocurrencies and are downloaded
from the market capitalization database and range from 2013 through 2018.
**2.1.** **Data Description**
Our main purpose of this paper is to get the most accurate forecasting price, based on the above
mentioned methods. Bitcoin, BTC is the first digital currency in market capitalization list and begins since
March 2013 through January 2018. Training data for bitcoin starts from 28th March 2013 to 16th until
January 2017, followed by Ethereum from 7th August, 2015 to 16th January, 2017, Litecoin from 28th April
2013 to 16th January 2017, Nem 1st April 2015 to 16th January 2017, Ripple 4th August, 2015 through 16th
January, 2017 and Stellar from 4th August 2013 to 16th January 2017. While testing data starts for all
selected type of cryptocurrencies start from 17th January, 2017 through 16th January 2018 subsequently.
Table 2 The training and testing dataset in our time series data. The first part is the training set
(number of values as per #Observations) in the first segment, accordingly. Several classifiers are then used to
predict the test data (number of values in the testing set is = 364) in the second segment.
Indonesian J Elec Eng & Comp Sci, Vol. 11, No. 3, September 2018 : 1121 – 1128
-----
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 1125
Table 2. The training and testing dataset in our time series data
Training Data Test Data
Cryptocurrency Name
From To #Observations From To #Observations
28-MarBitcoin, BTC, XBT
13 16-Jan-17 1388 17-Jan-17 16-Jan-18 364
Ether or
―Ethereum‖, ETH 7-Aug-15 16-Jan-17 526 17-Jan-17 16-Jan-18 364
Litecoin, LTC 28-Apr-13 16-Jan-17 1358 17-Jan-17 16-Jan-18 364
Nem, XEM 1-Apr-15 16-Jan-17 657 17-Jan-17 16-Jan-18 364
Ripple, XRP 4-Aug-13 16-Jan-17 1262 17-Jan-17 16-Jan-18 364
Stellar, XLM 5-Aug-14 16-Jan-17 896 17-Jan-17 16-Jan-18 364
**3.** **RESULTS AND ANALYSIS**
The result section begins by showing performance measures for each cryptocurrency types
according to classifiers. These serve as a control for the rest of the discussion. The analysis is separated into
two different experiments: i) Performance measures by various classifiers ii) Forecasted cryptocurrency value
by machine learning algorithms vs actual value. Table 3 shows the performance accuracy in correspondence
to four classifiers on the cryptocurrency market capitalization. The maximum value is 95.5%, which means
that any alphas over 95.5% have p-value of 0.01 or less.
Table 3. Performance Measures by various classifiers
Performance Accuracy (%)
Classifiers
Bitcoin Ethereum Litecoin Nem Ripple Stellar
SVM 78.90 **95.50** **82.40** 47.70 70.00 58.70
ANNs **79.40** 78.00 75.80 **77.80** 81.40 89.80
DL 61.90 69.40 62.80 57.20 60.90 70.70
BoostedNN 81.20 81.60 72.20 77.40 **81.50** **92.80**
Several different classifiers were trained with the same set of features. In this case, the datasets were
evaluated using classification accuracy. The comparison of all classifiers generated by different methods are
based on the same dataset. Thus it will be fair for all classifiers to perform the testing and training.
The results for the classifiers with the best performance on the test set are testified. The results show
that SVM classifier works well for Ethereum followed by Litecoin. While, ANN is seen works best for
Bitcoin followed by Nem. Ripple and Stellar has the best performance accuracy for BoostedNN. However,
among all, SVM classifier performs the best compared to the other classifiers with the performance accuracy
of 95.5%.
For comparability, same data sets and period of 364 days were chosen for all classifiers.
Performance can be seen in Figure 2-7. The SVM significantly outperformed the other classifiers. This result
is further explored using mean absolute percentage error (MAPE) calculation. SVM mean absolute
percentage error is 0.31% and is the lowest MAPE. Thus, the SVM is considered as reliable forecasting
model for these six selected cryptocurrency.
Figure 2. SVM value is comparable to actual Bitcoin for the period from 17/1/2017 to
16/1/2018
_Comparative Performance of Machine Learning Algorithms for Cryptocurrency… (Amelia Ritahani Ismail)_
-----
1126
ISSN: 2502-4752
Figure 3. SVM value is comparable to actual Litecoin for the period from 17/1/2017 to
16/1/2018
Figure 4. SVM value is comparable to actual Ripple for the period from 17/1/2017 to
16/1/2018
Figure 5. SVM value is comparable to actual Ethereum for the period from 17/1/2017 to
16/1/2018
Indonesian J Elec Eng & Comp Sci, Vol. 11, No. 3, September 2018 : 1121 – 1128
-----
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 1127
Figure 6. SVM value is comparable to actual Nem for the period from 17/1/2017 to 16/1/2018
Figure 7. SVM value is comparable to actual Stellar for the period from 17/1/2017 to
16/1/2018
**4.** **CONCLUSION**
The paper is highly focuses on the comparative performance of machine learning algorithms of six
cryptocurrencies. To begin with, the review of cryptocurrency has covered six major cryptocurrency, there
are Bitcoin, Ethereum, Litecoin, Nem, Ripple and Stellar. Further, previous studies on Machine Learning,
Support Vector Machines (SVM), Artificial Neural Networks (ANNs) and Deep Learning forecasting has
been explored.
Firstly, the performance measures were done to get the accuracy of classifiers over the selected
cryptocurrency and obtained the result as in Figure 3. Result shows that SVM outperformed other classifiers
with the accuracy of 95.5%. It is realized, that the quality of training data and population of dataset plays an
important role for a successful prediction.
Secondly, the forecasted cryptocurrency value by Machine Learning vs actual value of
cryptocurrency were then analyzed. From the comparative analysis done in this section, SVM has a
comparable values for all cryptocurrency for the period from 17/1/2017 to 16/1/2018.
Moreover, the result is further explored using mean absolute percentage error (MAPE) calculation.
The results show that SVM has the lowest value of MAPE. Thus, the SVM is considered as a reliable
forecasting model for the selected cryptocurrency.
In future, the algorithm will be improved on the accuracy rate of the forecasted price. Besides, with
the power of SVM, future work will be done to further optimize the SVM to get the most accurate result as
per actual value of cryptocurrency.
_Comparative Performance of Machine Learning Algorithms for Cryptocurrency… (Amelia Ritahani Ismail)_
-----
1128
ISSN: 2502-4752
**REFERENCES**
[1] Bontempi, G., Taieb, S. Ben, & Borgne, L. (2013). ―Machine Learning Strategies for Time Series Forecasting‖,
62–77.
[2] Huckle, S., & White, M. (2016). ―Socialism and the blockchain‖. _Future_ _Internet,_ _8(4)._
https://doi.org/10.3390/fi8040049
[3] Bitcoin. Bitcoin Developer Guide. Available online: https://bitcoin.org/en/developer-guide#block-chain (accessed
on 24 January 2018).
[4] Ethereum. Ethereum Project. Available online: https://www.ethereum.org/ (accessed on 24 January 2018).
[5] Litecoin. Litecoin—Open Source P2P Digital Currency. Available online: https://litecoin.org/ (accessed on 24
January 2018).
[6] Farell, R. (2015). ―An Analysis of the Cryptocurrency Industry‖. Wharton Research Scholars Journal. Paper, 130.
Retrieved from http://repository.upenn.edu/wharton_research_scholars%0Ahttp://repository.upenn.edu/wharton
_research
[7] _scholars/130
[8] Krause, D. (2017). Bitcoin - A Favourable Instrument For Diversification? A Quantitative Study On The Relations
Between Bitcoin
[9] Lee, D., Chuen, K., Guo, L., Wang, Y., & Chian, L. K. (2017). Cryptocurrency: A New Investment Opportunity?,
1–54.
[10] Heid, A. (2013). ―Analysis of the Cryptocurrency Marketplace‖. Retrieved February, 15, 2014.
[11] Application, F. A., & Guidelines, G. (2013). Ashesi University College. Office, 1–4.
[12] Chaigusin, S. (2014). An Application of Decision Tree for Stock Trading Rules : A Case ofthe Stock Exchange of
Thailand Proceedings of Eurasia Business Research Conference,(June).
[13] Ahmed, N. K., Atiya, A. F., El Gayar, N., & El-Shishiny, H. (2010). ―An empirical comparisonof machine learning
models for time series forecasting‖. _Econometric_ _Reviews,_ _29(5),_ 594–621.
https://doi.org/10.1080/07474938.2010.481556
[14] Patel, J., Shah, S., Thakkar, P., & Kotecha, K. (2015). ―Predicting stock and stock priceindex movement using
Trend Deterministic Data Preparation and machine learningtechniques‖. _Expert Systems with Applications, 42(1),_
259–268.https://doi.org/10.1016/j.eswa.2014.07.040
[15] Kongsilp, W., Mateus, C., Huang, M., Ting-ting, Z., Wan-yi, C., Maita, A. R. C., … deCarvalho, A. F. (2015).
―Prediction of Stock Trading Signal Based on Support VectorMachine‖. _Engineering Computations, 32(1), 445–_
463.https://doi.org/10.1108/02644401311286099
[16] Zhang, L., & Wang, J. (2015). ―Optimizing parameters of support vector machines using team-search-based
particle swarm optimization‖. _Engineering Computations, 32(5), 1194–1213. https://doi.org/10.1108/EC-12-2013-_
0310
[17] Basudhar, A. and Missoum, S. (2010), ―An improved adaptive sampling scheme for the construction of explicit
boundaries‖, Structural and Multidisciplinary Optimization, Vol. 42 No. 4, pp. 1-13.
[18] Lin, K., Basudhar, A., & Missoum, S. (2012). ―Parallel construction of explicit boundaries using support vector
machines‖. _Engineering Computations, 30(1), 132–148. https://doi.org/10.1108/02644401311286099_
[19] Huerta, R., Corbacho, F., & Elkan, C. (2013). ―Nonlinear support vector machines can systematically identify
stocks with high and low future returns‖. Algorithmic Finance, 2(1), 45–58. https://doi.org/10.3233/AF-13016
[20] Baccarini, L.M.R., Rocha e Silva, V.V., de Menezes, B.R. and Caminhas, W.M. (2011), ―SVM practical industrial
application for mechanical faults diagnostic‖, _Expert Systems with Applications, Vol. 38 No. 6, pp. 6980-6984._
[21] Hacib, T., Acikgoz, H., Bihan, Y. Le, Mekideche, M. R., Meyer, O., & Pichon, L. (2010). ―Support vector
machines for measuring dielectric properties of materials‖. COMPEL: The International Journal for Computation
_and_ _Mathematics_ _in_ _Electrical_ _and_ _Electronic_ _Engineering,_ _29(4),_ 1081–1089.
https://doi.org/10.1108/03321641011044497
[22] Nivetha, R. Y. (2017). ―Developing a Prediction Model for Stock Analysis‖, 4–6.
https://doi.org/10.1109/ICTACC.2017.11
[23] Borodo, S. M., Shamsuddin, S. M., & Hasan, S. (2016). ―Big data platforms and techniques‖. Indonesian Journal
_of_ _Electrical_ _Engineering_ _and_ _Computer_ _Science_ _(IJEECS),_ _1(1),_ 191–200.
https://doi.org/10.11591/ijeecs.v1.i1.pp191-200
[24] Lu, C. J. (2010). ―Integrating independent component analysis-based denoising scheme with neural network for
stock price prediction‖. _Expert_ _Systems_ _with_ _Applications,_ _37(10),_ 7056–7064.
https://doi.org/10.1016/j.eswa.2010.03.012
[25] The, S., Ai, S., Dalle, I., & Galleria, S. (2014). ―Deep Learning in Neural Networks : An Overview‖, 1–88.
[26] Borodo, S. M., Shamsuddin, S. M., & Hasan, S. (2016). ―Big data platforms and techniques‖. Indonesian Journal
_of_ _Electrical_ _Engineering_ _and_ _Computer_ _Science_ _(IJEECS),_ _1(1),_ 191–200.
https://doi.org/10.11591/ijeecs.v1.i1.pp191-200
[27] M. L. Ã., Karlsson, L., & Loutfi, A. (2014). ―A review of unsupervised feature learning and deep learning for time
series modeling‖. _Pattern Recognition Letters, 42(C), 11–24. https://doi.org/10.1016/j.patrec.2014.01.008_
Indonesian J Elec Eng & Comp Sci, Vol. 11, No. 3, September 2018 : 1121 – 1128
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.11591/IJEECS.V11.I3.PP1121-1128?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.11591/IJEECS.V11.I3.PP1121-1128, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBYNC",
"status": "HYBRID",
"url": "http://ijeecs.iaescore.com/index.php/IJEECS/article/download/13469/9222"
}
| 2,018
|
[] | true
| 2018-09-01T00:00:00
|
[
{
"paperId": "06ab46ea8d8f29abcf5e910b321d8ab69a1207fa",
"title": "Socialism and the Blockchain"
},
{
"paperId": "0a7f19c6eccef7523acbf204f91a60e97f95f60d",
"title": "Optimizing parameters of support vector machines using team-search-based particle swarm optimization"
},
{
"paperId": "0ac9579e715aded4f20caf1485a1e8c21678668d",
"title": "Prediction of Stock Trading Signal Based on Support Vector Machine"
},
{
"paperId": "af3b7e1e2921aa68f6e00e07589c0d0f585e9f76",
"title": "A review of unsupervised feature learning and deep learning for time-series modeling"
},
{
"paperId": "193edd20cae92c6759c18ce93eeea96afd9528eb",
"title": "Deep learning in neural networks: An overview"
},
{
"paperId": "78852ae4a9b2a14252999032e379e30d014971b3",
"title": "Parallel construction of explicit boundaries using support vector machines"
},
{
"paperId": "ee7dfe9335163564555d270baf538093ec407a02",
"title": "Nonlinear support vector machines can systematically identify stocks with high and low future returns"
},
{
"paperId": "bf77e0c9f8c30684be7b5e31183d373327f977eb",
"title": "SVM practical industrial application for mechanical faults diagnostic"
},
{
"paperId": "7afcb40f4969ba18d756b74ff61ddbbc6a40b180",
"title": "Integrating independent component analysis-based denoising scheme with neural network for stock price prediction"
},
{
"paperId": "38bcea3941f7264b9b9bad0eb10c700d4bc02904",
"title": "Support vector machines for measuring dielectric properties of materials"
},
{
"paperId": "ecab2d2946b26988a208c4593a41dd01b0da15e4",
"title": "An improved adaptive sampling scheme for the construction of explicit boundaries"
},
{
"paperId": "1cd54d46f6fad73a61d961765a0e962ca9bc9fcb",
"title": "Bitcoin a favourable instrument for diversification? : A quantitative study on the relations between Bitcoin and global stock markets"
},
{
"paperId": null,
"title": "Developing a Prediction"
},
{
"paperId": "ec5f9fa3a1422df6cf7df431efc502d39a29e2db",
"title": "Big Data Platforms and Techniques"
},
{
"paperId": "a68100519819845d990d496e9d8827de1394825b",
"title": "An Analysis of the Cryptocurrency Industry"
},
{
"paperId": "81be22bd4cf4e7ebb6ff5aa2a415e6b9aaf712fc",
"title": "Predicting stock and stock price index movement using Trend Deterministic Data Preparation and machine learning techniques"
},
{
"paperId": null,
"title": "Bitcoin Developer Guide Ethereum Project Litecoin — Open Source P 2 P Digital Currency"
},
{
"paperId": null,
"title": "An Application of Decision Tree for Stock Trading Rules : A Case ofthe Stock Exchange of Thailand"
},
{
"paperId": "99a39ea1b9bf4e3afc422329c1d4d77446f060b8",
"title": "Machine Learning Strategies for Time Series Forecasting"
}
] | 6,811
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/026a1ec02e15aacd176148eedb487bbc08edf905
|
[
"Computer Science"
] | 0.864079
|
Binding and Normalization of Binary Sparse Distributed Representations by Context-Dependent Thinning
|
026a1ec02e15aacd176148eedb487bbc08edf905
|
Neural Computation
|
[
{
"authorId": "2851989",
"name": "D. Rachkovskij"
},
{
"authorId": "33568664",
"name": "E. Kussul"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"Neural Comput"
],
"alternate_urls": [
"http://ieeexplore.ieee.org/servlet/opac?punumber=6720226",
"http://www.mitpressjournals.org/loi/neco",
"https://www.mitpressjournals.org/loi/neco"
],
"id": "69b9bcdd-8229-4a00-a6e0-00f0e99a2bf3",
"issn": "0899-7667",
"name": "Neural Computation",
"type": "journal",
"url": "http://cognet.mit.edu/library/journals/journal?issn=08997667"
}
| null |
```
Neural Computation (2001) v. 13 n. 2, pp. 411-452
# Binding and Normalization of Binary Sparse Distributed Representations by Context-Dependent Thinning
```
Dmitri A. Rachkovskij
V. M. Glushkov Cybernetics Center
Pr. Acad. Glushkova 40
Kiev 03680
Ukraine
dar@infrm.kiev.ua
(Preferable contact method)
Ernst M. Kussul
Centro de Instrumentos
Universidad Nacional Autonoma de
Mexico
Apartado Postal 70186
04510 Mexico D.F.
Mexico
ekussul@servidor.unam.mx
Keywords: distributed representation, sparse coding, binary coding, binding, variable binding,
representation of structure, structured representation, recursive representation, nested representation,
compositional distributed representations, connectionist symbol processing.
**Abstract**
Distributed representations were often criticized as inappropriate for encoding of data with a complex
structure. However Plate's Holographic Reduced Representations and Kanerva's Binary Spatter Codes
are recent schemes that allow on-the-fly encoding of nested compositional structures by real-valued or
dense binary vectors of fixed dimensionality.
In this paper we consider procedures of the Context-Dependent Thinning which were developed
for representation of complex hierarchical items in the architecture of Associative-Projective Neural
Networks. These procedures provide binding of items represented by sparse binary codevectors (with
low probability of 1s). Such an encoding is biologically plausible and allows a high storage capacity of
distributed associative memory where the codevectors may be stored.
In contrast to known binding procedures, Context-Dependent Thinning preserves the same low
density (or sparseness) of the bound codevector for varied number of component codevectors. Besides, a
bound codevector is not only similar to another one with similar component codevectors (as in other
schemes), but it is also similar to the component codevectors themselves. This allows the similarity of
structures to be estimated just by the overlap of their codevectors, without retrieval of the component
codevectors. This also allows an easy retrieval of the component codevectors.
Examples of algorithmic and neural-network implementations of the thinning procedures are
considered. We also present representation examples for various types of nested structured data
(propositions using role-filler and predicate-arguments representation schemes, trees, directed acyclic
graphs) using sparse codevectors of fixed dimension. Such representations may provide a fruitful
alternative to the symbolic representations of traditional AI, as well as to the localist and microfeaturebased connectionist representations.
-----
1. Introduction
The problem of representing nested compositional structures is important for connectionist systems,
because hierarchical structures are required for an adequate description of real-world objects and
situations.
In fully local representations, an item (entity, object) of any complexity level is represented by a
single unit (node, neuron) (or a set of units which has no common units with other items). Such
representations are similar to symbolic ones and share their drawbacks. These drawbacks include the
limitation of the number of representable items by the number of available units in the pool, and
therefore, the impossibility to represent the combinatorial variety of real-world objects. Besides, a unit
corresponding to a complex item only represents its name and pointers to its components (constituents).
Therefore in order to determine the similarity of complex items, they should be unfolded into the baselevel (indecomposable) items.
The attractiveness of distributed representations was emphasized by the paradigm of cell
assemblies (Hebb, 1949) that influenced the work of Marr (1969), Willshaw (1981), Palm (1980),
Hinton, McClelland & Rumelhart (1986), Kanerva (1988), and many others. In fully distributed
representations, an item of any complexity level is represented by its configuration pattern over the
whole pool of units. For binary units, this pattern is a subset of units which are in the active state. If the
subsets corresponding to various items intersect, then the number of these subsets is much more than the
number of units in the pool, providing an opportunity to solve the problem of information capacity of
representations. If similar items are represented by similar subsets of units, the degree of corresponding
subsets' intersection could be the measure of their similarity.
The potentially high information capacity of distributed representations provides hope for
solving the problem of representing combinatorially growing number of recursive compositional items in
a reasonable number of bits. Representing composite items by concatenation of activity patterns of their
component items would increase the dimensionality of the coding pool. If the component items are
encoded by pools of equal dimensionality, one could try to represent composite items as superposition of
activity patterns of their components. The resulting coding pattern would have the same dimensionality.
However another problem arises here, known as "superposition catastrophe" (e.g. von der
Malsburg, 1986) as well as "ghosts", "false” or “spurious” memory (e.g. Feldman & Ballard, 1982;
Hopfield, 1982; Hopfield, Feinstein, Palmer, 1983). A simple example looks as follows. Let there be
component items a, b, c and composite items ab, ac, cb. Let us represent any two of the composite items,
e.g. _ac or_ _cb. For this purpose, superimpose activity patterns corresponding to the component items_ _a_
and c, c and b. The ghost item ab also becomes represented in the result, though it is not needed. In the
"superposition catastrophe" formulation, the problem consists in no way of telling which two items (ab,
_ac, or ab, cb, or ac, cb) make up the representation of the composite pattern abc, where the patterns of_
all three component items are activated.
The supposition of no internal structure in distributed representations (or assemblies) (Legendy,
1970; von der Malsburg, 1986; Feldman, 1989) held back their use for representation of complex data
structures. The problem is to represent in the distributed fashion not only the information on the set of
base-level components making up a complex hierarchical item, but also the information on the
combinations in which they meet, the grouping of those combinations, etc. That is, some mechanisms
were needed for binding together the distributed representations of certain items at various hierarchical
levels.
One of the approaches to binding is based on temporal synchronization of constituent activation
(Milner, 1974; von der Malsburg, 1981, 1985; Shastri & Ajjanagadde, 1993; Hummel & Holyoak, 1997).
Though this mechanism may be useful inside single level of composition, its capabilities to represent and
store complex items with multiple levels of nesting are questionable. Here we will consider binding
mechanisms based on the activation of specific coding-unit subsets corresponding to a group
(combination) of items, the mechanisms that are closer to the so-called conjunctive coding approach
(Smolensky, 1990; Hummel & Holyoak, 1997).
"Extra units" considered by Hinton (1981) represent various combinations of active units of two
or more distributed patterns. The extra units can be considered as binding units encoding various
-----
combinations of distributedly encoded items by distinct distributed patterns. Such a representation of
bound items is generally described by tensor products (Smolensky, 1990) and requires an exponential
growth of the number of binding units with the number of bound items. However, as was already
mentioned, for recursive structures it is desired that the dimensionality of the patterns representing
composite items is the same as of the component items’ patterns. Besides, the most important property of
distributed representations - their similarity for similar structures - should be preserved.
The problem was discussed by Hinton (1990), and a number of mechanisms for construction of
his _reduced descriptions has been proposed. Hinton (1990), Pollack (1990), Sperduti (1994) get the_
reduced description of a composite pattern as a result of multilevel perceptron training using backpropagation algorithm. However, their patterns are low dimensional. Plate (1991, 1995) binds highdimensional patterns with gradual (real-valued) elements on the fly, without increase of the
dimensionality, using the operation of circular convolution. Kanerva (1996) uses bitwise XOR to bind
binary vectors with equal probability of 0s and 1s. Binary distributed representation are especially
attractive, because binary bitwise operations are enough to handle them, providing the opportunity for
significant simplification and acceleration of algorithmic implementations.
Sparse binary representations (with small fraction of 1s) are of special interest. The sparseness
of codevectors allows a high storage capacity of distributed associative memories (Willshaw, Buneman,
& Longuet-Higgins, 1969; Palm, 1980; Lansner & Ekeberg, 1985; Amari, 1989) which can be used for
their storage, and still further acceleration of software and hardware implementations (e.g. Palm &
Bonhoeffer, 1984; Kussul, Rachkovskij, & Baidyk, 1991a; Palm, 1993). Sparse encoding has also
neurophysiological correlates (Foldiak & Young, 1995). The procedure for binding of sparse distributed
representations ("normalization procedure") was proposed by Kussul as one of the features of the
Associative-Projective Neural Networks (Kussul, 1988, 1992; Kussul, Rachkovskij, & Baidyk, 1991). In
this paper, we describe various versions of such a procedure, its possible neural-network
implementations, and provide examples of its use for the encoding of complex structures.
In section 2 we discuss representational problems encountered in the Associative-Projective
Neural Networks and an approach for their solution. In section 3 the requirements on the ContextDependent Thinning procedure for binding and normalization of binary sparse codes are formulated. In
section 4 several versions of the thinning procedure along with their algorithmic and neural-network
implementations are described. Some generalizations and notations are given in section 5. In section 6
retrieval of individual constituent codes from the composite item code is considered. In section 7 the
similarity characteristics of codes obtained by Context-Dependent Thinning procedures are examined. In
section 8 we show examples of encoding various structures using Context-Dependent Thinning. Related
work and general discussion are presented in section 9, and conclusions are given in section 10.
2. Representation of composite items in the APNN: the problems and the answer
2.1. Features of the APNN
The Associative-Projective Neural Networks (APNN) is the name of a neural-network architecture
proposed by Kussul in 1983 for the AI problems which require efficient manipulation of hierarchical
data structures. Fragments of the architecture implemented in software and hardware were also used for
solution of pattern-recognition tasks (Kussul, 1992, 1993). APNN features of interest here are as follows
(Kussul, 1992; Kussul, Rachkovskij, & Baidyk, 1991a; Amosov et al., 1991):
- Items of any complexity (an elementary feature, a relation, a complex structure, etc.) are represented by
stochastic distributed activity patterns over the neuron field (pool of units);
- The neurons are binary and therefore patterns of activity are binary vectors;
- Items of any complexity are represented over the neural fields of the same high dimensionality N;
- The number M of active neurons in the representations of items of various complexity is approximately
(statistically) the same and small compared to the field dimensionality N. However M is large enough to
maintain its own statistical stability.
-----
- Items of various complexity level are stored in different distributed auto-associative neural-network
memories with the same number N of neurons.
Thus, items of any complexity are encoded by sparse distributed stochastic patterns of binary
neurons in the neural fields of the same dimensionality N. It is convenient to represent activity patterns
in the neural fields as binary vectors, where 1s correspond to active neurons. Let us use bold-face
lowercase font for codevectors to distinguish them from the items they represent denoted in italics.
The number of 1s in x is denoted |x|. We seek to make |x| ≈ _M for x of various complexity. The_
similarity of codevectors is determined by the number of 1s in their intersection or overlap: |x∧y|, where
∧ is elementwise conjunction of x and y. The probability of 1s in x (or density of 1s in x, or simply the
vector density) is p(x)=|x|/N.
Information encoding by stochastic binary vectors with a small number of 1s allows a high
capacity of correlation-type neural-network memory (known as Willshaw memory or Hopfield network)
to be reached using Hebbian learning rule (Wilshaw, Buneman, & Longuet-Higgins, 1969; Palm, 1980;
Lansner & Ekeberg, 1985; Frolov & Muraviev, 1987, 1988; Frolov, 1989; Amari, 1989; Tsodyks, 1989).
The codevectors we are talking about may be exemplified by vectors with _M=100...1000,_ _N=10_
000...100 000. Though the maximal storage capacity is reached at _M=logN (Willshaw, Buneman, &_
Longuet-Higgins, 1969; Palm, 1980), we use _M≈√N to get a network with a moderate number_ _N of_
neurons and N[2] connections at sufficient statistical stability of M (e.g. the standard deviation of M less
than 3%). Under this choice of the codevector density _p=M/N_ ≈ 1/√N, the information capacity holds
high enough and the number of stored items can exceed the number of neurons in the network
(Rachkovskij, 1990a, 1990b; Baidyk, Kussul, & Rachkovskij, 1990).
Let us consider the problems arising in the APNN and other neural-network architectures with
distributed representations when composite items are constructed.
2.2. Density of composite codevectors
The number H of component items (constituents) comprising a composite item grows exponentially with
the nesting level, that is, with going to the higher levels of the part-whole hierarchy. If S items of a level l
constitute an item of the adjacent higher level (l+1), then for level (l+L) the number H becomes
_H = S[L]_ . (2.1)
The presence of several items comprising a composite item is encoded by the concurrent activation of
their patterns, that is, by superposition of their codevectors. For binary vectors, we will use superposition
by bitwise disjunction. Let us denote composite items by concatenation of symbols denoting their
component items, e.g. abc. Corresponding composite codevectors (ai ∨ _bi_ ∨ _ci, i=1,...,N) will be denoted_
as a ∨ **b ∨** **c or simply abc.**
Construction of composite items will be accompanied by fast growth of density p' and respective
number M' of 1s in their codevectors. For H different superimposed codevectors of low density p:
_p'H = 1-(1-p)[H]_ ≈ 1-e[-][pH], (2.2)
_M'H_ ≈ _p'HN. (2.3)_
Equations 2.2 and 2.3 take into account the "absorption" of coincident 1s that prevents the exponential
growth of their number versus the composition level L. However it is important that p'>>p (see Figure 1)
and M' >> _M. Since the dimensionality N of codevectors representing items of various complexity is the_
same, the size of corresponding distributed auto-associative neural networks, where the codevectors are
stored and recalled, is also the same. Therefore at M' >> _M_ ≈ √N (at the higher levels of hierarchy) their
storage capacity in terms of the number of recallable codevectors will decrease dramatically. To
maintain high storage capacity at each level, M' should not substantially exceed M. However, due to the
requirement of statistical stability, the number of 1s in the code can not be reduced significantly.
Besides, the operation of distributed auto-associative neural-network memory usually implies the same
number of 1s in codevectors. Thus it is necessary to keep the number of 1s in the codevectors of
complex items approximately equal to M. (However, some variation of M between distinct hierarchical
levels may be tolerable and even desirable).
-----
These provide one of the reasons why composite items should be represented not by all 1s of
their component codevectors, but only by their fraction approximately equal to M (i.e. only by some M
representatives of active neurons encoding the components).
2.3. Ghost patterns and false memories
The well-known problem of ghost patterns or superposition catastrophe was mentioned in the
Introduction. It consists in losing the information on the membership of component codevectors in
particular composite codevector, when several composite codevectors are superimposed in their turn.
This problem is due to the essential property of superposition operation. The contribution of
each member to their superposition does not depend on the contributions of other members. For
superposition by elementwise disjunction, representation of a in a∨b and a∨c is the same. The result of
superposition of several base-level component codevectors contains only the information concerning
participating components and no information about the combinations in which they meet. Therefore if
common items are constituents of several composite items, then the combination of the latter generally
can not be inferred from their superposition codevector. For example, let a complex composite item
consist of base-level items a, b, c, d, e, f. Then how could one determine that it really consists of the
composite items abd, bce, caf, if there may be also other items, such as abc, def, etc. ? In the formulation
of "false” or “spurious patterns", superposition of composite patterns **abc** and **def generates false**
patterns ("ghosts") abd, bce, caf, etc.
The problem of introducing “false assemblies” or “spurious memories” (unforeseen attractors)
into a neural network (e.g. Kussul, 1980; Hopfield, Feinstein, & Palmer, 1983; Vedenov, 1987, 1988)
has the same origin as the problem of ghosts. Training of an associative memory of matrix type is usually
performed using some version of Hebbian learning rule implemented by superimposing in the weight
matrix the outer products of memorized codevectors. For binary connections, e.g.
_Wij' = Wij_ ∨ _xixj, (2.4)_
where _xi_ and _xj are the states of the_ _i-th and the_ _j-th neurons when the pattern_ **x to be memorized is**
presented (i.e. the values of the corresponding bits of x), Wij and Wij' are the connection weights between
the i-th and the j-th neurons before and after training, respectively, ∨ stands for disjunction.
When this learning rule is sequentially used to memorize several composite codevectors with
partially coinciding components, false assemblies (attractors) may appear - that is, memorized composite
codevectors that were not presented to the network. For example, when representations of items _abd,_
_bce, caf are memorized, the false assembly abc (unforeseen attractor) is formed in the network (Figure_
2A). Moreover, various two-item assemblies _ab,_ _ad, etc. are present, which also were not explicitly_
presented for storing.
The problem of introducing false assemblies can be avoided if non-distributed associative
memory is used, where the patterns are not superimposed when stored and each composite codevector is
placed into a separate memory word. However the problem of false patterns or superposition catastrophe
still persists.
2.4. An idea of the thinning procedure
A systematic use of distributed representations provides the prerequisite to solve both the problem of
codevector density growth and the superposition catastrophe. The idea of solution consists in including
into the representation of a composite item not full sets of 1s encoding its component items, but only
their subsets. If we choose the fraction of 1s from each component codevector so that the number of 1s
in the codevector of a composite item is equal to _M, then the density of 1s will be preserved in_
codevectors of various complexity. For example, if S=3 items of level l comprise an item of level l+1,
then approximately _M/3 of 1s should be preserved from each codevector of the_ _l-th level. Then the_
codevector of level l+1 will have approximately M of 1s. If two items of level l+1 comprise an item of
level l+2, then approximately M/2 of 1s should be preserved from each codevector of level l+1. Thus the
low number _M of 1s in the codevectors of composite items of various complexity is maintained, and_
-----
therefore high storage capacity of the distributed auto-associative memories where these low-density
codevectors are stored can be maintained as well (see also section 2.1).
Hence the component codevectors are represented in the codevector of the composite item in a
reduced form - by a fraction of their 1s. The idea that the items of higher hierarchical levels ("floors")
should contain their components in reduced, compressed, coarse form is well-accepted among those
concerned with diverse aspects of Artificial Intelligence research. Reduced representation of component
codevectors in the codevector of composite item realized in the APNN may be relevant to "coarsen
models" of Amosov (1968), "reduced descriptions" of Hinton (1990), and "conceptual chunking" of
Halford, Wilson, & Phillips (in press).
Reduced representation of component codevectors in the codevectors of composite items also
allows a solution of the superposition catastrophe. If the subset of 1s included in the codevector of a
composite item from each of the component codevectors depends on the composition of component
items, then different subsets of 1s from each component codevector will be found in the codevectors of
different composite items. For example, non-identical subsets of 1s will be incorporated into the
codevectors of items abc and acd from a. Therefore the component codevectors will be bound together
by the subsets of 1s delegated to the codevector of the composite item. It hinders the occurrence of false
patterns and assemblies.
For the example from the Introduction, when both _ac and_ _cb are present, we will get the_
following overall composite codevector: ac ∨ ca ∨ cb ∨ bc, where xy stands for the subset of 1s in x that
becomes incorporated into the composite codevector given y as the other component. Therefore if ac ≠
**ab,** **bc** ≠ **ba, we do not observe the ghost pattern ab** ∨ **ba in the resultant codevector.**
For the example of Figure 2A, where false assemblies emerge, they do not emerge under reduced
representation of items (Figure 2B). Now interassembly connections are formed between different
subsets of active neurons which have relatively small intersection. Therefore the connectivity of
assembly corresponding to the non-presented item abc is low.
That the codevector of a composite item contains the subsets of 1s from the component
codevectors preserves the information on the presence of component items in the composite item. That
the composition of each subset of 1s depends on the presence of other component items preserves the
information on the combinations in which the component items occurred. That the codevector of a
composite item has approximately the same number of 1s as its component codevectors allows the
combinations of such composite codevectors to be used for construction of still more complex
codevectors of higher hierarchical levels.
Thus an opportunity emerges to build up the codevectors of items of varied composition level
containing the information not only on the presence of their components, but on the structure of their
combinations as well. It provides the possibility to estimate the similarity of complex structures without
their unfolding but simply as overlap of their codevectors which is considered by many authors as a very
important property for AI systems (e.g. Kussul, 1992; Hinton, 1990; Plate, 1995, 1997).
Originally the procedure reducing the sets of coding 1s of each item from the group which
makes up a composite item was named "normalization" (Kussul, 1988; Kussul & Baidyk, 1990; Kussul,
1992). That name emphasized the property to maintain the number of 1s in the codes of composite items
equal to that of component items.. However in this paper we will call it "Context- Dependent Thinning"
(CDT) by its action mechanism, that reduces the number of 1s taking into account the context of other
items from their group.
3. Requirements on the Context-Dependent Thinning procedures
Let us summarize the requirements on the CDT procedures and on the characteristics of codevectors
produced by them. The procedures should process sparse binary codevectors. An important case of input
is superimposed component codevectors. The procedures should output the codevector of the composite
item where the component codevectors are bound and the density of the output codevector is comparable
to the density of component codevectors. Let us call the resulting (output) codevector as "thinned"
codevector. The requirements may be expressed as follows.
-----
3.1. Determinism
Repeated application of the CDT procedures to the same input should produce the same output.
3.2. Variable number of inputs
The procedure should process one, two, or several codevectors. One important case of input is a vector
in which several component codevectors are superimposed.
3.3. Sampling of inputs
Each component codevector of the input should be represented in the output codevector by a fraction of
its 1s (or their reversible permutation).
3.4. Proportional sampling
The number of 1s representing input component codevectors in the output codevector should be
proportional to their density. If the number of 1s in a and b is the same, then the number of 1s from a
and b in thinned ab should also be (approximately) the same.
3.5. Uniform low density
The CDT procedures should maintain (approximately) uniform low density of output codevectors (small
number M' of 1s) under varied number of input codevectors and their correlation degree.
3.6. Density control
The CDT procedures should be able to control the number M' of 1s in output codevectors within some
range around M (the number of 1s in the component codevectors). For one important special case, M'=M.
3.7. Unstructured similarity
An output codevector of the CDT procedures should be similar to each component codevector at the
input (or to its reversible permutation). Fulfillment of this requirement follows from fulfillment of the
sampling of inputs requirement (3.3). The thinned codevector for ab is similar to a and b. If the densities
of component codevectors are the same, the magnitude of similarity is the same (as follows from the
requirement of proportional sampling, 3.4).
3.8. Similarity of subsets
The reduced representations of a given component codevector should be similar to each other to a degree
that varies directly with the similarity of the set of other codevectors with which it is composed. The
representation of **a in the thinned** **abc should be more similar to its representation in the thinned** **abd**
than in thinned aef.
3.9. Structured similarity
If two sets (collections) of component items are similar, their thinned codevectors should be similar as
well. It follows from the similarity of subsets requirement (3.8). If **a and** **a' are similar,** **b** and **b' are**
similar, then thinned ab should be similar to thinned a'b'. Or, thinned abc should be similar to thinned
**abd.**
3.10. Binding
Representation of a given item in an output thinned codevector should be different for different sets
(collections) of component items. Representation of **a** in thinned **abc should be different from the**
representation of **a in thinned** **abd. Thus the representation of** **a** in the thinned composite codevector
contains information on the other components of a composite item.
4. Versions of the Context-Dependent Thinning procedures
-----
Let us consider some versions of the CDT procedure, their properties and implementations.
4.1. Direct conjunctive thinning of two or more codevectors
Direct conjunctive thinning of binary x and y is implemented as their element-wise conjunction:
**z = x** ∧ **y, (4.1)**
where z is thinned and bound result.
The requirement of determinism (section 3.1) holds for the direct conjunctive thinning
procedure. The requirement of variable number of inputs (3.2) is not met, since only two codevectors are
thinned. Overlapping 1s of **x and** **y go to** **z, therefore the sampling of inputs requirement (3.3) holds.**
Since equal number of 1s from **x and** **y enters into** **z even if** **x** and **y are of different density, the**
requirement of proportional sampling (3.4) is not fulfilled in general case.
For stochastically independent vectors x and y the density of the resulting vector z is:
_p(z) = p(x)p(y) < min(p(x),p(y)) < 1. (4.2)_
Here min() selects the smallest of its arguments. Let us note that for correlated x and y the density of 1s
in z depends on the degree of their correlation. Thus p(z) is maintained the same only for independent
codevectors of constant density, and the requirement of uniform low density (3.5) is generally not met.
Since p(z) for sparse vectors is substantially lower than p(x) and p(y), the requirement of density control
(3.6) is not met and recursive construction of bound codevectors is not supported (see also Kanerva,
1998; Sjödin, et. al., 1998). Similarity and binding requirements (3.7-3.10) may be considered as
partially satisfied for two codevectors (see also Table 1).
Table 1. Properties of various versions of thinning procedures. "Yes" means that the property is present,
"No" means that the property is not present, “No-Yes” and "Yes-No" mean that the property is partially
present. See text for details.
Properties of thinning procedures Direct conjunctive Permutive Additive (4.3) and
thinning (4.1) thinning (4.2) subtractive (4.4) CDT
Determinism (3.1) Yes Yes Yes
Variable number of inputs (3.2) No-Yes Yes Yes
Sampling of inputs (3.3) Yes Yes Yes
Proportional sampling (3.4) No Yes Yes
Uniform low density (3.5) No No Yes
Density control (3.6) No No Yes
Unstructured similarity (3.7) Yes-No Yes Yes
Similarity of subsets (3.8) Yes-No Yes Yes
Structured similarity (3.9) Yes-No Yes Yes
Binding (3.10) Yes-No Yes Yes
Though the operation of direct conjunctive thinning of two codevectors does not meet all
requirements on the CDT procedure, it has been applied by us for encoding of external information, in
particular, for binding of distributed binary codevectors of feature item and its numerical value (Kussul
& Baidyk, 1990; Rachkovskij & Fedoseyeva, 1990; Artykutsa et al., 1991; Kussul, Rachkovskij, &
Baidyk, 1991a, 1991b). The density p of the codevectors of features and numerical values was chosen so
as to provide a specified density p' of the resulting codevector (Table 2, K=2).
Table 2. The density p of K independent codevectors chosen to provide a specified density p' of
codevectors produced by their conjunction.
|Properties of thinning procedures|Direct conjunctive thinning (4.1)|Permutive thinning (4.2)|Additive (4.3) and subtractive (4.4) CDT|
|---|---|---|---|
|Determinism (3.1) Variable number of inputs (3.2) Sampling of inputs (3.3) Proportional sampling (3.4) Uniform low density (3.5) Density control (3.6) Unstructured similarity (3.7) Similarity of subsets (3.8) Structured similarity (3.9) Binding (3.10)|Yes No-Yes Yes No No No Yes-No Yes-No Yes-No Yes-No|Yes Yes Yes Yes No No Yes Yes Yes Yes|Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes|
-----
|p'|K 2 3 4 5 6 7 8 9 10 11 12|
|---|---|
|0.001 0.010 0.015|0.032 0.100 0.178 0.251 0.316 0.373 0.422 0.464 0.501 0.534 0.562 0.100 0.215 0.316 0.398 0.464 0.518 0.562 0.599 0.631 0.658 0.681 0.122 0.247 0.350 0.432 0.497 0.549 0.592 0.627 0.657 0.683 0.705|
To thin more than two codevectors, it is natural to generalize equation 4.1:
**z = ∧s xs, (4.3)**
where s = 1…S, S is the number of codevectors to be thinned. Though this operation allows binding of
two or more codevectors, a single vector can not be thinned. The density of resulting codevector **z**
depends on the densities of xs and on their number S. Therefore to meet the requirement of uniform low
density (3.5), the densities of **xs should be chosen depending on the number of thinned codevectors.**
Also, the requirement of density control (3.6) is not satisfied.
We applied this version of direct conjunctive thinning to encode positions of visual features on a
two-dimensional retina. Three codevectors were bound (S=3): the codevector of a feature, the codevector
of its _X-coordinate, and the codevector of its_ _Y-coordinate (unpublished work of 1991-1992 on_
recognition of handwritten digits, letters, and words in collaboration with WACOM Co., Japan). Also,
this technique was used to encode words and word combinations for text processing (Rachkovskij,
1996). In so doing, the codevectors of letters comprising words were made bound (S>10). The density of
codevectors to be bound by thinning was chosen so as to provide a specified density of the resulting
codevector (Table 2, K=3...12).
Neural-network implementations of direct conjunctive thinning procedures are rather
straightforward and will not be considered here.
4.2. Permutive conjunctive thinning
The codevectors to be bound by direct conjunctive thinning are not superimposed. Let us consider the
case where S codevectors are superimposed by disjunction:
**z = ∨sxs. (4.4)**
Conjunction of a vector with itself produces the same vector: **z** ∧ **z =** **z. So let us modify** **z** by
permutation of all its elements and make conjunction with the initial vector:
**z' = z** ∧ **z[~]. (4.5)**
Here, z[~] is the permuted vector. In vector-matrix notation, it can be rewritten as:
**z' = z** ∧ **Pz, (4.5a)**
where P is an N x N permutation matrix (each row and each column of P has a single 1, and the rest of
**P is 0; multiplying a vector by a permutation matrix permutes the elements of the vector).**
Proper permutations are those producing the permuted vector that is independent of the initial
vector, e.g. random permutations or shifts. Then the density of the result is
_p(z') = p(z)p(z[~]) = p(z) p(Pz). (4.6)_
Let us consider the composition of the resulting vector:
**z' = z** ∧ **z[~] = (x1** ∨... ∨ **xs) ∧** **z[~]**
= (x1 ∨ ... ∨ **xs) ∧ (x1~** ∨ ...∨ **xs~)**
= x1 ∧ (x1~ ∨ ... ∨ **xs~) ∨ ... ∨** **xs** ∧ (x1~ ∨ ... ∨ **xs~)**
= (x1 ∧ **x1~) ∨ ... ∨ (x1** ∧ **xs~) ∨ ... ∨ (xs** ∧ **x1~) ∨ ...∨ (xs** ∧ **xs~). (4.7)**
Thus the resulting codevector is the superposition of all possible pairs of bitwise codevector
conjunctions. Each pair includes certain component codevector and certain permuted component
codevector.
Because of initial disjunction of component codevectors, this procedure meets more
requirements on the CDT procedures than direct conjunctive thinning. The requirement of variable
number of inputs (3.2) is now fully satisfied. As follows from equation 4.7, each component codevector
**xs is thinned by conjunction with one and the same stochastic independent vector** **z[~]. Therefore**
-----
statistically the same fraction of 1s is left from each component **xs. Therefore the requirements of**
sampling of inputs (3.3) and of proportional sampling (3.4) hold.
For S sparse codevectors of equal density p(x)<<1
_p(z)≈Sp(x), (4.8)_
_p(z') = p(z)p(z~) ≈_ _S2p2(x). (4.9)_
To satisfy the requirements of density (3.5-3.6), p(z')=p(x) should hold for various S. It means that p(x)
should be equal to 1/S[2]. Therefore at fixed density _p(x) the density requirements (3.5-3.6) are not_
satisfied for variable number S of component items. The similarity and binding requirements (3.7-3.10)
hold. In particular, the requirement of similarity of subsets (3.8) holds because the higher the number of
identical items, the more identical conjunctions are superimposed in equation 4.7.
A neural-network implementation of permutive conjunctive thinning is shown in Figure 3. In the
neural network terms, units are called "neurons", and their pools are called "neural fields". There are two
input fields fin1 and fin2 and one output field fout consisting of N binary neurons each. fin1 is connected to
**fout by a bundle of** _N direct projective (1-to-1) connections. Each connection of this bundle connects_
neurons of the same number. fin2 is connected to fout by the bundle of N permutive projective connections.
Each connection of this bundle connects neurons of different numbers. "Synapses" of the neurons of the
output field have weights of +1. The same pattern of superimposed component codevectors is activated
in both input fields. Each neuron of the output field summarizes the excitations of its two inputs
(synapses). The output is determined by comparison of the excitation level with the threshold θ = 1.5.
Therefore each output neuron performs conjunction of its inputs. Thus the activity pattern of the output
field corresponds to bit conjunction of pattern present in the input fields and its permutation. Obviously,
there are a lot of different configurations of permutive connections. Permutation by shift is particularly
attractive as it is simple and fast to implement in computer simulations.
4.3. Additive CDT procedure
Though for permutive conjunctive thinning the density of resulting codevectors is closer to the density of
each component codevector than for direct conjunction, it varies with the number and density of
component codevectors.
Let us make a codevector z by the disjunction of S component codevectors xs, as in equation 4.4.
Since the density of component codevectors is low and their number is small, then the "absorption" of
common 1s is low and, according to equations 4.8 and 4.9, p(z') is approximately S[2]p(x) times p(x). For
example, if p(x)=0.01 and S=5, then p(z') ≈ (1/4)p(x).
Therefore, to make the density of the thinned codevector equal to the density of its component
codevectors, let us superimpose an appropriate number K of independent vectors with the density p(z'):
〈z〉 = ∨k (z ∧ **zk~) = z** ∧ (∨k **zk~) (4.10)**
Here 〈z〉 is the thinned output vector, zk~ is a unique (independent stochastic) permutation of elements of
vector z, fixed for each k. In vector-matrix notation, we can write:
〈z〉 = ∨k (z ∧ **Pkz) = z** ∧ ∨k (Pkz) (4.10a)
The number K of vectors to be superimposed by disjunction can be determined as follows. If the
density of the superposition of permuted versions of z is made
_p(∨k(Pkz)) = 1/S, (4.11)_
then after conjunction with z (equations 4.10, 4.10a) we will get the needed density of 〈z〉:
_p(〈z〉) = p(z)/S_ ≈ _Sp(x)/S = p(x). (4.12)_
Taking into account the "absorption" of 1s in disjunction of K permuted vectors z[~], equation (4.11) can
be rewritten:
1/S = 1-(1-pS)[K]. (4.13)
Then K=ln(1-1/S)/ln(1-pS). (4.14)
The dependence K(S) at different p is shown in Table 3.
-----
Table 3. The number K of permutations of an input codevector that produces the proper density of the
thinned output codevector in the additive version of Context-Dependent Thinning. (K should be rounded
to the nearest integer).
Number S of component codevectors in the
Density p of input codevector
component codevector 2 3 4 5 6 7
0.001 346.2 135.0 71.8 44.5 30.3 21.9
0.005 69.0 26.8 14.2 8.8 6.0 4.3
0.010 34.3 13.3 7.0 4.4 2.9 2.1
4.3.1. Meeting the requirements to the CDT procedures
Since the configuration of each k-th permutation is fixed, the procedure of additive CDT is deterministic
(3.1). The input vector z to be thinned is the superposition of component codevectors. The number of
these codevectors may be variable, therefore the requirement of variable number of inputs (3.2) holds.
The output vector is obtained by conjunction of **z (or its reversible permutation) with the**
independent vector ∨k(Pkz). Therefore the 1s of all codevectors superimposed in **z are equally**
represented in 〈z〉 and both the sampling of inputs and the proportional sampling requirements (3.3 - 3.4)
hold.
Density control of the output codevector for variable number and density of component
codevectors is realized by varying K (Table 3). Therefore the density requirements (3.5 - 3.6) hold. Since
the sampling of inputs and proportional sampling requirements (3.3-3.4) hold, the codevector 〈z〉 is
similar to all component codevectors xs, and the requirement of unstructured similarity (3.7) holds. The
more similar are the components of one composite item to those of another, the more similar are their
superimposed codevectors **z. Therefore the more similar are the vectors - disjunctions of** _K fixed_
permutations of z, and the more similar representations of each component codevector will remain after
conjunction (equation 4.10) with **z. Thus the similarity of subsets requirement (3.8) holds.**
Characteristics of this similarity will be considered in more detail in section 7.
Since different combinations of component codevectors produce different **z and therefore**
different codevectors of _K permutations of_ **z, representations of certain component codevector in the**
thinned codevector will be different for different combinations of component items, and the binding
requirement (3.10) holds. The more similar are representations of each component in the output vector,
the more similar are output codevectors (the requirement of structured similarity 3.9 holds).
4.3.2. An algorithmic implementation
As mentioned before, shift is an easily implementable permutation. Therefore an algorithmic
implementation of this CDT procedure may be as in Figure 4A. Another example of this procedure does
not require preliminary calculation of _K (Figure 4B). In this version, conjunctions of the initial and_
permuted vectors are superimposed until the number of 1s in the output vector becomes equal to M.
4.3.3. A neural-network implementation
A neural-network implementation of the first example of the additive CDT procedure (Figure 4A) is
shown in Figure 5.
To choose _K depending on the density of_ **z, the neural-network implementation should**
incorporate some structures not shown in the figure. They should determine the density of the initial
pattern **z** and "activate" (turn on) _K bundles of permutive connections from their total number_ _Kmax._
Alternatively, these structures should actuate the bundles of permutive connections one-by-one in the
|Density p of component codevector|Number S of component codevectors in the input codevector 2 3 4 5 6 7|
|---|---|
|0.001 0.005 0.010|346.2 135.0 71.8 44.5 30.3 21.9 69.0 26.8 14.2 8.8 6.0 4.3 34.3 13.3 7.0 4.4 2.9 2.1|
-----
fixed order[1] until the density of the output vector in **fout becomes M/N. Let us recall that bundles of shift**
permutive connections are used in algorithmic implementations.
4.4. Subtractive CDT procedure
Let us consider another version of the CDT procedure. Rather than masking **z** with the straight
disjunction of permuted versions of **z, as in additive thinning, let us mask it with the inverse of that**
disjunction:
〈z〉 = z ∧ ¬(∨k **zk~) = z** ∧ ¬∨k (Pkz). (4.15)
If we choose _K to make the number of 1s in_ 〈z〉 equal to _M, then this procedure will satisfy the_
requirements of section 3. Therefore the density of superimposed permuted versions of **z before**
inversion should be 1-1/S (compare to equation 4.13). Thus the number _K of permuted vectors to be_
superimposed in order to obtain the required density (taking into account "absorption" of 1s) is
determined from:
1-1/S = 1 - (1-pS)[K]. (4.16)
Then, for pS << 1
_K_ ≈ lnS/(pS). (4.17)
Algorithmic implementations of this subtractive CDT procedure (Kussul, 1988; Kussul &
Baidyk, 1990; Kussul, 1992) are analogous to those presented for the additive CDT procedure in section
4.3.2. A neural-network implementation is shown in Figure 6.
Since the value of lnS/S is approximately the same at S=2,3,4,5 (Table 4), one can choose the K
for a specified density of component codevectors p(x) as:
_K_ ≈ 0.34/p(x). (4.18)
At such K and S, p(〈z〉) ≈ _p(x). Therefore the number K of permutive connection bundles in Figure 6 can_
be fixed, and their sequential activation is not needed. So each neuron of _Fout may be considered as_
connected to an average of K randomly chosen neurons of Fin1 by inhibitory connections. More precise
values of K (obtained as exact solution of equation 4.16) for different values of p are presented in Table
4.
Table 4. The function lnS/S and the number K of permutations of an input codevector that produces the
proper density of the thinned output codevector in the subtractive version of Context-Dependent
Thinning procedure. (K should be rounded to the nearest integer).
Number S of component codevectors in the input
codevector
2 3 4 5 6 7
_p_ lnS/S
0.347 0.366 0.347 0.322 0.299 0.278
0.001 346.2 365.7 345.9 321.1 297.7 277.0
0.005 69.0 72.7 68.6 63.6 58.8 54.6
0.010 34.3 36.1 34.0 31.4 29.0 26.8
This version of the CDT procedure has been originally proposed under the name "normalization
procedure" (Kussul, 1988; Kussul & Baidyk, 1990; Amosov et al., 1991). We have used it in the
multilevel APNN for sequence processing (Rachkovskij, 1990b; Kussul & Rachkovskij, 1991). We have
also used it for binding of sparse codes in perceptron-like classifiers (Kussul, Baidyk, Lukovich, &
1As noted by Kanerva (personal communication), all Kmax bundles could be activated in parallel, if the
weight of the k-th bundle is set to be 2[-][k] and the common threshold of fint neurons is adjusted dynamically
so that fout has the desired density of 1s.
|p|Number S of component codevectors in the input codevector 2 3 4 5 6 7 lnS/S 0.347 0.366 0.347 0.322 0.299 0.278|
|---|---|
|0.001 0.005 0.010|346.2 365.7 345.9 321.1 297.7 277.0 69.0 72.7 68.6 63.6 58.8 54.6 34.3 36.1 34.0 31.4 29.0 26.8|
-----
Rachkovskij, 1993) and in one-level APNN applied for recognition of vowels (Rachkovskij &
Fedoseyeva, 1990, 1991), textures (Artykutsa et al., 1991; Kussul, Rachkovskij, & Baidyk, 1991b),
shapes (Kussul & Baidyk, 1990), handprinted characters (Lavrenyuk, 1995), logical inference (Kasatkin
& Kasatkina, 1991).
5. Procedures of auto-thinning, hetero-thinning, self-exclusive thinning and notation
In sections 4.2-4.4 we considered the versions of thinning procedures where a single vector
(superposition of component codevectors) was the input. The corresponding pattern of activity was
present both in the field fin1 and fin2 (Figures 3, 5, 6), and therefore the input vector thinned itself. Let us
call these procedures "auto-thinning" or "auto-CDT" and denote them as
label〈u〉. (5.1)
Here u is the codevector to be thinned (usually superposition of component codevectors) which
is in the input fields **fin1 and** **fin2 of Figures 3,5,6.** [label]〈...〉 denotes particular configuration of thinning
(particular realization of bundles of permutive connections). Let us note that angle brackets are used by
Plate to denote normalization operation in HRRs (e.g. Plate, 1995; see also section 9.1.5).
A lot of orthogonal configurations of permutive connections are possible. Differently labeled
CDT procedures implement different thinning. In the algorithmic implementations (Figure 4) different
labels will use different seeds. No label corresponds to some fixed configuration of thinning. Unless
otherwise specified, it is assumed that the number K of bundles is chosen to maintain the preset density
of the thinned vector 〈u〉, usually |〈u〉|≈M.
∨k(Pku) can be expressed as Ru thresholded at 1/2, where the matrix R is the disjunction, or it
can also be the sum, of K permutation matrices Pk. This, in turn, can be written as a function T(u), so that
we get
〈u〉 = u ∧ _T(u). (5.2)_
It is possible to thin one codevector by another one if the pattern to be thinned is activated in fin1,
and the pattern which thins is activated in fin2. Let us call such procedure hetero-CDT, hetero-thinning,
thinning u with w. We denote hetero-thinning as
label〈u〉w. (5.3)
Here w is the pattern that does the thinning. It is activated in fin2 of Figures 3,5,7. u is the pattern which
is thinned, it is activated in fin1. [label]〈...〉 is the configuration label of thinning. For auto-thinning, we may
write 〈u〉 **=** 〈u〉u.
For the additive hetero-thinning, equation 4.10 can be rewritten as
〈u〉w = u ∧ (∨k(Pkw)) = u ∧ _T(w). (5.4)_
For the subtractive hetero-thinning, equation 4.15 can be rewritten as
〈u〉w = u ∧ ¬(∨k(Pkw)) = u ∧ ¬T(w). (5.5)
_Examples._
As before, we denote composite codevector u to be thinned by its component codevectors, e.g. u
= a ∨ **b ∨** **c or simply u = abc.**
Auto-thinning of composite codevector u:
〈u〉u = 〈a∨b∨c〉a∨b∨c = 〈a∨b∨c〉 = 〈abc〉abc = 〈abc〉.
Hetero-thinning of composite codevector u with codevector d:
〈u〉d = 〈a∨b∨c〉d = 〈abc〉d.
For both additive and subtractive CDT procedures:
〈abc〉 = 〈a〉abc ∨ 〈b〉abc ∨ 〈c〉abc.
We can also write 〈abc〉 = (a ∧ _T(abc)) ∨ (b_ ∧ _T(abc))_ ∨ (c ∧ _T(abc)). Analogous expression can be_
written for a composite pattern with other numbers of components. Let us note that K should be the same
for thinning of the composite pattern as a whole or its individual components.
For the additive CDT procedure it is also true:
〈abc〉 = 〈a〉abc ∨ 〈b〉abc ∨ 〈c〉abc =
-----
〈a〉a ∨ 〈b〉a ∨ 〈c〉a ∨ 〈a〉b ∨ 〈b〉b ∨ 〈c〉b ∨ 〈a〉c ∨ 〈b〉c ∨ 〈c〉c.
For the subtractive CDT procedure we can write:
〈a〉bcd = 〈〈〈a〉b〉c〉d and
〈abc〉 = 〈〈〈a〉a〉b〉c ∨ 〈〈〈b〉a〉b〉c ∨ 〈〈〈c〉a〉b〉c.
Let us also consider a modification of the auto-CDT procedures which will be used in section
7.2. If we eliminate the thinning of a component codevector with itself, we obtain "self-exclusive" autothinning. Let us denote it as 〈abc〉\abc:
〈abc〉\abc = 〈a〉bc ∨ 〈b〉ac ∨ 〈c〉ab.
6. Retrieval of component codevectors
After thinning, the codevectors of component items are present in the thinned codevector of a composite
item in a reduced form. We must be able to retrieve complete component codevectors. Since the
requirement of the unstructured similarity (3.7) holds, the thinned composite codevector is similar to its
component codevectors. So if we have a full set (alphabet) of component codevectors of the preceding
(lower) level of compositional hierarchy, we can compare them with the thinned codevector. The
similarity degree is determined by the overlap of codevectors. The alphabet items corresponding to the
codevectors with maximum overlaps are the sought-for components.
The search of the most similar component codevectors can be performed by a sequential finding
of overlaps of the codevector to be decoded with all codevectors of the component alphabet. An
associative memory can be used to implement this operation in parallel. After retrieving of the full-sized
component codevectors of the lower hierarchical level, one can then retrieve their component
codevectors of still lower hierarchical level in an analogous way. For this purpose, the alphabet of the
latter should be known as well. If the order of component retrieval is important, some auxiliary
procedures can be used (Kussul, 1988; Amosov et al., 1991; Rachkovskij, 1990b; Kussul & Rachkovskij,
1991).
_Example. Let us consider the alphabet of six component items a, b, c, d, e, f. They are encoded_
by stochastic fixed vectors of _N=100000 bits with_ _M≈1000 bits set to 1. Let us obtain the thinned_
codevector 〈abc〉. The number of 1s in 〈abc〉 in our numerical example is |〈abc〉| = 1002. Let us find the
overlap of each component codevector with the thinned codevector: |a ∧ 〈abc〉| = 341; |b ∧ 〈abc〉| = 350;
|c ∧ 〈abc〉| = 334; |d ∧ 〈abc〉| = 12; |e ∧ 〈abc〉| = 7; |f ∧ 〈abc〉| =16. So the representation of the
component items a, b, c is substantially higher than the representation of the items d, e, f occurring due
to a stochastic overlap of independent binary codevectors. The numbers obtained are typical for the
additive and the subtractive versions of thinning, as well as for their self-exclusive versions.
7. Similarity preservation by the thinning procedures
In this section, let us consider the similarity of thinned composite codevectors as well as the similarity of
thinned representations of component codevectors in the thinned composite codevectors. These kinds of
similarity are considered under different combinations of component items and different versions of
thinning procedures.
Let us use the following Context-Dependent Thinning procedures:
- permutive conjunctive thinning, section 4.2 (Paired-M);
- additive auto-CDT, section 4.3 (CDTadd);
- additive self-exclusive auto-CDT, section 5 (CDTadd-sl);
- subtractive auto-CDT, section 4.4 (CDTsub);
- subtractive self-exclusive auto-CDT, section 5 (CDTsub-sl).
For these experiments, let us first obtain the composite codevectors which have 5 down to 0
component codevectors in common: **abcde, abcdf, abcfG, abfgh, afghi, fghij. For the component**
codevectors, N=100000 bits with M≈1000. Then, for each thinning procedure, let us thin the composite
-----
codevectors down to the density of their component codevectors. (For permutive conjunctive thinning,
the density of component codevectors was chosen to get approximately M of 1s in the result).
7.1. Similarity of thinned codevectors
Let us find an overlap of thinned codevector 〈abcde〉 with 〈abcde〉, 〈abcdf〉, 〈abcfg〉, 〈abfgh〉, 〈afghi〉,
and 〈fghij〉. Here 〈〉 is used to denote any thinning procedure. A normalized measure of the overlap of x
with various y is determined as |x∧y|/|x|.
The experimental results are presented in Figure 7A, where the normalized overlap of thinned
composite codevectors is shown versus the normalized overlap of corresponding unthinned composite
codevectors. It can be seen that the overlap of thinned codes for various versions of the CDT procedure
is approximately equal to the square of overlap of unthinned codes. For example, the similarity (overlap)
of **abcde and** **abfgh is approximately 0.4 (two common components of five total), and the overlap of**
their thinned codevectors is about 0.16.
7.2. Similarity of component codevector subsets included into thinned codevectors
Some experiments were conducted in order to investigate the similarity of subsets requirement (3.8). The
similarity of subsets of a component codevector incorporated into various thinned composite vectors was
obtained as follows. First, the intersections of various thinned five-component composite codevectors
with their component a were determined: **u=a∧〈abcde〉,** **v=a∧〈abcdf〉,** **w=a∧〈abcfg〉,** **x=a∧〈abfgh〉,**
**y=a∧〈afghi〉. Then, the normalized values of the overlap of intersections were obtained as |u∧v|/|u|,**
|u∧w|/|u|, |u∧x|/|u|, |u∧y|/|u|.
Figure 7B shows how the similarity (overlap) of component codevector subsets incorporated into
two thinned composite codevectors varies versus the similarity of corresponding unthinned composite
codevectors. It can be seen that these dependencies are different for different thinning procedures: for
the CDTadd and the CDTadd-sl they are close to linear, but for the CDTsub and CDTsub-sl they are
polynomial. Which is preferable, depends on the application.
7.3. The influence of the depth of thinning
By the depth of thinning we understand the density value of a thinned composite codevector. Before, we
considered it equal to the density of component codevectors. Here, we vary the density of the thinned
codevectors. The experimental results presented in Figure 8 are useful for the estimation of resulting
similarity of thinned codevectors in applications.
As in sections 7.1-7.2, composite codevectors of five components were used. Therefore
approximately 5M of 1s (actually, more close to 4.9M because of random overlaps) were in the input
vector before thinning. We varied the number of 1s in the thinned codevectors from 4M to M/4. Only the
additive and the subtractive CDT procedures were investigated.
The similarity of thinned codevectors is shown in Figure 8A. For a shallow thinning, where the
resulting density is near the density of input composite codevector, the similarity degree of resulting
vectors is close to that of input codevectors (the curve is close to linear). For a deep thinning, where the
density of thinned codevectors is much less than the density of input codevectors, the similarity function
behave as a power function, transforming from linear through quadratic to cubic (for subtractive
thinning).
The similarity of component subsets in the thinned codevector is shown in Figure 8B. For the
additive CDT procedure, the similarity function is linear, and its angle reaches approximately 45° for
“deep” thinning. For the subtractive CDT procedure, the function is similar to the additive one for the
“shallow” thinning, and becomes near-quadratic for the “deep” thinning.
8. Representation of structured expressions
-----
Let us consider representation of various kinds of structured data by binary sparse codevectors of fixed
dimensionality. In the examples below, the items of the base level for a given expression. In its turn, they
may, in their turn, represent complex structured data.
8.1. Transformation of symbolic bracketed expressions into representations by codevectors
Performing the CDT procedure can be viewed as analog of introducing brackets into symbolic
descriptions. As mentioned in section 5, the CDT procedures with different thinning configurations are
denoted by different labels at the opening thinning bracket: [1]〈 〉, [2]〈 〉, [3]〈 〉, [4]〈 〉, [5]〈 〉, etc.
Therefore in order to represent a complex symbolic structure by a distributed binary codevector,
one should
- map each symbol of the base-level item to the corresponding binary sparse codevector of fixed
dimensionality;
- replace conventional brackets in symbolic bracketed representation by "thinning" ones. Each
compositional level has its own label of thinning brackets, that is, thinning configuration;
- superimpose the codevectors inside the thinning brackets of the deepest nesting level by elementwise
disjunction;
- perform the CDT procedure on superimposed codevectors using the configuration of thinning
corresponding to particular "thinning" label;
- superimpose the resulting thinned vectors inside the thinning brackets of the next nesting level;
- perform the CDT procedure on superimposed codevectors using appropriate thinning configuration of
that nesting level;
- repeat the two previous steps until the whole structure is encoded.
8.2. Representation of ordered items
For many propositions, the order of arguments is essential. To represent the order of items encoded by
the codevectors, binding with appropriate roles is usually used.
One approach is to use explicit binding of role codevectors (agent-object, antecedentconsequent, or just ordinal number) with the item (filler) codevector. This binding can be realized by an
auto- or hetero-CDT procedure (Rachkovskij, 1990b). The item a which is #3 may be represented as 〈a ∨
**n3〉 or 〈a〉n3, where n3 is the codevector of the “third place” role.**
Another approach is to use implicit binding by providing different locations for different
positions of an item in a proposition. To preserve fixed dimensionality of codevectors, it was proposed
to encode different positions by the specific shifts of codevectors (Kussul & Baidyk, 1993). (Reversible
permutations can be also used). For our example, we have **a** shifted by the number _n3 of 1-bit shifts_
corresponding to the “third place” of an item.
These and other techniques to represent the order of items have their pros and cons. Thus a
specific technique should be chosen depending on the application. So we will not consider details here.
It is important that such techniques exist, and we will denote the codevector of item a at the n-th place
simply by a_n.
Let us note that generally the modification of an item codevector to encode its ordinal number
should be different for different nesting levels. It is analogous to having its own thinning configuration at
each level of nesting. Therefore **a and** **b should be modified in the same manner in** [1]〈...a_n...〉 and
1〈...b_n...〉, but a should generally be modified differently in 2〈...a_n...〉.
8.3. Examples
8.3.1. Role-filler structure
Representations of structures or propositions by the role-filler scheme are widely used (Smolensky,
1990; Pollack, 1990; Plate, 1991, 1995; Kanerva, 1996; Sperduti, 1994). Let us consider the relational
instance
-----
_knows(Sam, loves(John, Mary)). (8.1)_
Using the Holographic Reduced Representations of Plate, it can be represented as:
**L1 = love + loveagt∗john + loveobj∗mary,** (8.2)
**L2 = know + knowagt∗sam + knowobj∗L1,** (8.3)
where ∗ stands for binding operation, and + denotes addition. In our representation:
**L1 = [2]〈love ∨** [1]〈loveagt ∨ **john〉** ∨ [1]〈loveobj ∨ **mary〉〉, (8.4)**
**L2 = [4]〈know ∨** [3]〈knowagt ∨ sam〉 ∨ [3]〈knowobj ∨ L1〉〉. (8.5)
8.3.2. Predicate-arguments structure
Let us consider representation of relational instances loves(John, Mary) and loves(Tom, Wendy) by the
predicate-arguments (or symbol-argument-argument) structure (Halford, Wilson, & Phillips, in press):
**loves∗John∗Mary + loves∗Tom∗Wendy. (8.6)**
Using our representation, we obtain:
2〈1〈loves_0 ∨ **John_1 ∨** **Mary_2〉** ∨ 1〈loves_0 ∨ **Tom_1 ∨** **Wendy_2〉〉. (8.7)**
Let us note that this example may be represented using the role-filler scheme of HRRs as
**L1 = loves + lover∗Tom + loved∗Wendy, (8.8)**
**L2 = loves + lover∗John + loved∗Mary, (8.9)**
**L = L1 + L2. (8.10)**
Under such a representation, the information about who loves whom is lost in L (Plate, 1995; Halford,
Wilson, & Phillips, in press). In our representation, this information is preserved even using the rolefiller scheme:
**L1 = [2]〈loves** ∨ [1]〈lover ∨ **Tom〉** ∨ [1]〈loved ∨ **Wendy〉〉, (8.11)**
**L2 = [2]〈loves ∨** [1]〈lover ∨ **John〉** ∨ [1]〈loved ∨ **Mary〉〉, (8.12)**
**L = 〈L1 ∨** **L2〉. (8.13)**
Another example of relational instance from Halford, Wilson, & Phillips (in press):
_cause(shout-at(John,Tom),hit(Tom, John)). (8.14)_
Using our representation scheme, it may be represented as
2〈cause_0 ∨ 1〈shout-at_0 ∨ **John_1 ∨** **Tom_2〉_1 ∨** 1〈hit_0 ∨ **Tom_1 ∨** **John_2〉_2〉. (8.15)**
8.3.3. Tree-like structure
An example of bracketed binary tree adapted from Pollack (1990):
((d (a n)) (v (p (d n)))). (8.16)
If we do not take the order into account, but use only the information about the grouping of constituents,
our representation may look as simple as:
4〈3〈d ∨ 2〈a ∨ **n〉〉** ∨ 3〈v ∨ 2〈p ∨ 1〈d ∨ **n〉〉〉〉. (8.17)**
8.3.4 Labeled directed acyclic graph
Sperduti & Starita (1997), Frasconi, Gori, & Sperduti (1997) provide examples of labeled directed
acyclic graphs. Let us consider
_F( a, f(y), f(y, F(a, b)) ). (8.18)_
Using our representation, it may look as
3〈F_0 ∨ **a_1 ∨** 2〈f_0 ∨ **y_1〉_2 ∨** 2〈f_0 ∨ **y_1 ∨** 1〈F_0 ∨ **a_1 ∨** **b_2〉_2〉_3 〉. (8.19)**
9. Related work and discussion
The procedures of Context-Dependent Thinning allow construction of binary sparse representations of
complex data structures, including nested compositional structures or part-whole hierarchies. The basic
principles of such representations and their use for data handling were proposed in the context of the
-----
Associative-Projective Neural Networks paradigm (Kussul, 1988, 1992; Kussul, Rachkovskij, & Baidyk,
1991a).
9.1 Comparison to other representation schemes
Let us compare our scheme for representation of complex data structures using the CDT procedure (we
will call it "APNN-CDT" below) with other schemes using distributed representations. The best known
schemes are (L)RAAMs (Pollack, 1990; Blair, 1997; Sperduti 1994), Tensor Product representations
(Smolensky, 1990; Halford, Wilson, & Phillips, in press), Holographic Reduced Representations (HRRs)
(Plate, 1991, 1995), Binary Spatter Codes (BSCs) (Kanerva, 1994, 1996). For this comparison, we will
use the framework of Plate (1997) who proposes to distinguish these schemes using the following
features: the nature of distributed representation; the choice of superposition; the choice of binding
operation; how the binding operation is used to represent predicate structure; the use of other operations
and techniques.
9.1.1. The nature of distributed representation
Vectors of random real-valued elements with the Gaussian distribution are used in HRRs. Dense binary
random codes with the number of 1s equal to the number of 0s are used in BSCs. Vectors with real or
binary elements (without specified distributions) are used in other schemes.
In the APNN-CDT scheme, binary vectors with randomly distributed small number of 1s are
used to encode base-level items.
9.1.2. The choice of superposition
The operation of superposition is used for unstructured representation of an aggregate of codevectors.
In BSCs superposition is realized as a bitwise thresholded addition of codevectors. Schemes
with non-binary elements, such as HRRs, use elementwise summation. For tensors, superposition is
realized as adding up or ORing the corresponding elements.
In the APNN-CDT scheme, elementwise OR is used.
9.1.3. The choice of binding operation
Most schemes use special operations for binding of codevectors. The binding operations producing the
bound vector that has the same dimension as initial codevectors (or one of them in (L)RAAMs) are
convenient for representation of recursive structures. The binding operation is performed "on the fly" by
circular convolution (HRRs), elementwise multiplication (Gayler, 1998), or XOR (BSCs). In
(L)RAAMs, binding is realized through multiplication of input codevectors by the weight matrix of the
hidden layer formed by training of a multilayer perceptron using the codevectors to be bound.
The vector obtained by binding can be bound with another codevector in its turn. In Tensor
Models, binding of several codevectors is performed by their tensor product. The dimensionality of
resulting tensor grows with the number of bound codevectors.
In the APNN-CDT scheme, binding is performed by the Context-Dependent Thinning procedure.
Unlike the other schemes where the codevectors to be bound are not superimposed, they can be
superimposed by disjunction in the basic version of the CDT procedure. Superposition codevector z (as
in equation 4.4) makes the context codevector. The result of the CDT procedure may be considered as
superimposed bindings of each component codevector with the context codevector. Or, it may be
considered as superimposed paired bindings of all component codevectors with each other. (Note that in
the “self-exclusive” CDT version (section 5) the codevector of each component is not bound to itself. In
the hetero-CDT version, one codevector is bound to another codevector through thinning with the latter).
According to Plate's framework, CDT as a binding procedure can be considered as a special kind
of superposition (disjunction) of certain elements of the tensor product of **z by itself (i.e.** _N[2]_ scalar
products zizj). Actually, 〈zi〉 is disjunction of certain zizj=zi∧zj, where zj is the j-th element of permuted z
-----
(equation 4.10). CDT can be also considered as a hashing procedure: the subspace to where hashing is
performed is defined by 1s of z, and some 1s of z are mapped to that subspace.
Since the resulting bound codevector 〈z〉 is obtained in the CDT procedure by thinning the 1s of
**z, (where the component codevectors are superimposed),** 〈z〉 is similar to its component codevectors
(unstructured similarity is preserved). Therefore to retrieve the components bound in the thinned
codevector, we only need to choose the most similar component codevectors from their alphabet. This
can be done using an associative memory.
None of the mentioned binding operations, except for the CDT, preserves unstructured
similarity. Therefore to extract some component codevector from the bound codevector, they demand to
know the other component codevector(s). Then rebinding of the bound codevector with the inverses of
known component codevector(s) produces a noisy version of the sought component codevector. This
operation is known as decoding or "unbinding". To eliminate noise from the unbound codevector, a
"clean-up" memory with the full alphabet of component codevectors is also required in those schemes. If
some or all components of the bound codevector are not known, decoding in those schemes requires
exhaustive search (substitution, binding, and checking) through all combinations of codevectors from the
alphabet. Then the obtained bound codevector most similar to the bound codevector to be decoded
provides the information on its composition.
As in the other schemes, structured similarity is preserved by the CDT, i.e. bindings of similar
patterns are similar to each other. However the character of similarity is different. In most of the other
schemes the similarity of the bound codevectors is equal to the product of similarities of the component
codevectors (e.g. Plate, 1995). For example, the similarity of a∗b and a∗b' is equal to the similarity of b
and b'. Therefore if b and b' are not similar at all, the bound vectors also will not be similar.
The codevectors to be bound by the CDT procedure are initially superimposed component
codevectors. So their initial similarity is the mean of the components' similarities. Also, the thinning
itself preserves approximately the square of similarity of the input vectors. So, the similarity for
dissimilar b and b' will be >0.25 instead of 0 for the other schemes.
9.1.4. How the binding operation is used to represent predicate structure
In most of the schemes predicate structures are represented by role-filler bindings. Halford, Wilson, &
Phillips (in press) use predicate-argument bindings. The APNN-CDT scheme allows such
representations of predicate structures as role-filler bindings, predicate-argument bindings, and also
offers a potential for other possible representations. Both ordered and unordered arguments can be
represented.
9.1.5. The use of other operations and techniques.
_- Normalization. After superposition of codevectors, some normalizing transformation is used in various_
schemes to bring the individual elements or the total strength of the resulting codevector within certain
limits. In BSCs, it is the threshold operation that converts a non-binary codevector (the bitwise sum of
component codevectors) to a binary one. In HRRs, it is the scaling of codevectors to the unit length that
facilitates their comparison.
The CDT procedure performs a dual role: it not only binds superimposed codevectors of
components, but it also normalizes the density of the resulting codevector. It would be interesting to
check to what extent the normalization operations in other schemes provide the effect of binding as well.
_Clean-up memory._ Associative memory is used in various representation schemes for storage of
component codevectors and their recall (clean-up after finding their approximate noisy versions using
unbinding). After the CDT procedure, the resulting codevector is similar to its component codevectors,
however the latter are represented in the reduced form. Therefore it is natural to use associative
memories in the APNN-CDT scheme to store and retrieve the codevectors of component items of various
complexity levels. Since component codevectors of different complexity levels have approximately the
same and small number of 1s, an associative memory based on assembly neural networks with simple
Hebbian learning rule allows efficient storage and retrieval of a large number of codevectors.
-----
_Chunking. The problem of chunking remains one of the least developed issues in existing representation_
schemes.
In the HRRs and BSCs chunks are normalized superpositions of stand-alone component
codevectors and their bindings. In its turn, the codevector of a chunk can be used as one of the
components for binding. Thus, chunking allows structures of arbitrary nesting or composition level to be
built. Each chunk should be stored in a clean-up memory. When complex structures are decoded by
unbinding, noisy versions of chunk codevectors are obtained. They are used to retrieve pure versions
from the clean-up memory, which can be decoded in their turn.
In those schemes, the codevectors of chunks are not bound. Therefore they can not be
superimposed without the risk of structure loss, as it was repeatedly mentioned in this paper. In the
APNN-CDT scheme, any composite codevector after thinning represents a chunk. Since the component
codevectors are bound in the chunk codevector, the latter can be operated as a single whole (an entity)
without confusion of components belonging to different items.
When a compositional structure is constructed using HRRs or BSCs, the chunk codevector is
usually the filler which becomes bound with some role codevector. In this case, in distinction to the
APNN-CDT scheme, the components a, b, c of the chunk become bound with the role rather than with
each other:
**role∗(a + b + c) = role∗a + role∗b + role∗c. (9.1)**
Again, if the role is not unique, it can not be determined to which chunk the binding **role∗a belongs.**
Also, the role codevector should be known for unbinding and subsequent retrieving of the chunk.
Thus in the representation schemes of HRRs and Binary Spatter Codes each of the component
codevectors belonging to a chunk binds with (role) codevectors of other hierarchical levels not belonging
to that chunk. Therefore such bindings may be considered as "vertical". In the APNN-CDT scheme, a
"horizontal" binding is essential: the codevectors of the chunk components are bound with each other.
In the schemes of Plate, Kanerva, and Gayler, the vertical binding chain **role_upper_level** ∗
(role_lower_level ∗ **filler) is indistinguishable from** **role_lower_level** ∗ (role_upper_level ∗
**filler),because their binding operations are associative and commutative. For the CDT procedure, in**
contrast, [2]〈[1]〈a ∨ **b〉** ∨ **c〉** ≠ [2]〈a ∨ [1]〈b ∨ **c〉〉, and also 〈〈a** ∨ **b〉** ∨ **c〉** ≠ 〈a ∨ 〈b ∨ **c〉〉.**
Gayler (1998) proposes to bind a chunk codevector with its permuted version. It resembles the
version of thinning procedure from section 4.2, but for real-valued codevectors. Different codevector
permutations for different nesting levels allow the components of chunks from different levels to be
distinguished, in a similar fashion as using different configurations of thinning connections in the CDT.
However since the result of binding in the scheme of Gayler and in the other considered schemes (with
the exception of APNN-CDT) is not similar to the component codevectors, in those schemes decoding of
the chunk codevector created by binding with a permutation of itself will generally require exhaustion of
all combinations of component codevectors.
This problem with the vertical binding schemes of Plate, Kanerva, and Gayler can be rectified
by using a binding operation that, prior to a conventional binding operation, permutes its left and right
arguments differently (as discussed on p. 84 in Plate (1994)).
The obvious problem of Tensor Product representation is the growth of dimensionality of the
resulting pattern obtained by the binding of components. If it is not solved, the dimensionality will grow
exponentially with the nesting depth. Halford, Wilson, & Phillips (in press) consider chunking as the
means to reduce the rank of tensor representation. To realize chunking, they propose to use the
operations of convolution, concatenation, superposition, as well as some special function that associates
the outer product with the codevector of lower dimension. However the first three operations do not rule
out confusion of grouping or ordering of arguments inside chunk, (i.e., different composite items may
produce identical chunks). And the special function (and its inverse) requires concrete definition.
Probably it could be done using associative memory, e.g. of the sigma-pi type proposed by Plate (1998).
In (L)RAAMs the chunks of different nesting levels are encoded in the same weight matrix of
connections between the input layer and the hidden layer of a multilayer perceptron. It may be one of the
reasons for poor generalization. Probably if additional multilayer perceptrons are introduced for each
-----
nesting level (with the input for each following perceptron provided by the hidden layer of the preceding
one, similarly to Sperduti & Starita, 1997), generalization in those schemes would improve.
In the APNN, chunks (thinned composite codevectors) of different nesting levels are memorized
in different auto-associative neural networks. It allows an easy similarity-based decoding of a chunk
through its subchunks of the previous nesting level and decreases memory load at each nesting level (see
also Rachkovskij, accepted).
9.2. Sparse binary schemes
Indicating unknown areas where useful representation schemes for nested compositional structures can
be found, Plate (1997) notes that known schemes poorly handle sparse binary patterns, because known
binding and superposition operations change the density of sparse patterns.
From the work known to us, only Sjödin (1998) expresses the idea of "thinning" in an effort to
avoid the low associative memory capacity for dense binary patterns. He defines the thinning operation
as preserving the 1s corresponding to the maximums of some function defined over the binary vector.
The values of that function can be determined at cyclic shifts of the codevector by the number of steps
equal to the ordering number of 1s in that codevector. However it is not clear from such a description
what are the properties of the maximums and, therefore, what is the character of similarity.
The CDT procedure considered in this paper allows the density of codevectors to be preserved
while binding them. Coupled with the techniques for encoding of the pattern ordering, this procedure
allows implementation of various representation schemes of complex structured data. Approximately the
same low density of binary codevectors at different nesting levels permits the use of identical procedures
for construction, recognition, comparison, and decoding of patterns at different hierarchical levels of the
Associative-Projective Neural Network architecture (Kussul 1988, 1992; Kussul, Rachkovskij, &
Baidyk, 1991).
The CDT procedure preserves the similarity of encoded descriptions allowing the similarity of
complex structures to be determined by the overlap of their codevectors. Also, in the codevectors of
complex structures formed using the CDT procedure, representation of the component codevectors
(subset of their 1s) is reduced. Therefore, the APNN-CDT scheme can be considered another
implementation of Hinton's (1990) reduced descriptions, Amosov's (1967) item coarsing, or compression
of Halford, Wilson, & Phillips (in press). Besides, the CDT scheme is biologically relevant since it uses
sparse representations and allows simple neural-network implementation.
10. Conclusion
Procedures of Context-Dependent Thinning described in the paper perform binding of items represented
as sparse binary codevectors. They allow a variable number of superimposed patterns to be bound on the
fly while preserving the density of bound codevectors. The result of the CDT is of the same
dimensionality as the component codevectors. Using of the auto-CDT procedures as analogs of brackets
in the bracketed symbolic representation of various complex data structures permits easy transformation
of these representations to the binary codevectors of fixed dimensionality with a small number of 1s.
Unlike other binding procedures, binding by the auto-CDT preserves the similarity of bound
codevector with each of the component codevectors. It makes possible both to determine the similarity of
complex items with each other by the overlap of their codevectors and to retrieve in full size the
codevectors of their components. Such operations are efficiently implementable by distributed
associative memories which provide high storage capacity for the codevectors with small number of 1s.
The APNN-CDT style representations have been already used by us in applications (an earlier
work is reviewed in Kussul, 1992, 1993; more recent developments are described in Lavrenyuk, 1995;
Rachkovskij, 1996; Kussul & Kasatkina, 1999; Rachkovskij, accepted). We hope that the CDT
procedures will find their application for distributed representation and manipulation of complex
compositional data structures, contributing to the progress of connectionist symbol processing
(Touretzky, 1995, 1990; Touretzky & Hinton, 1988; Hinton, 1990; Plate, 1995). Fast (parallel)
-----
evaluation of similarity or finding the most similar compositional items allowed by such representations
are extremely useful for solution of a wide range of AI problems.
Acknowledgements: The authors are grateful to Pentti Kanerva and Tony Plate for their extensive and
helpful comments, valuable suggestions, and continuous support. This work was funded in part by the
International Science Foundation Grants U4M000 and U4M200.
**References**
Amari, S. (1989). Characteristics of sparsely encoded associative memory. Neural Networks, 2, 445-457.
Amosov, N. M. (1967). Modelling of thinking and the mind. New York: Spartan Books.
Amosov, N. M., Baidyk, T. N., Goltsev, A. D., Kasatkin, A. M., Kasatkina, L. M., Kussul, E. M., &
Rachkovskij, D. A. (1991). Neurocomputers and intelligent robots. Kiev: Naukova dumka. (In
Russian).
Artykutsa, S. Ya., Baidyk, T. N., Kussul, E. M., & Rachkovskij, D. A. (1991). Texture recognition using
neurocomputer. (Preprint 91-8). Kiev, Ukraine: V. M. Glushkov Institute of Cybernetics. (In
Russian).
Baidyk, T. N., Kussul, E. M., & Rachkovskij, D. A. (1990). Numerical-analytical method for neural
network investigation. In Proceedings of The International Symposium on Neural Networks and
_Neural Computing - NEURONET'90 (pp. 217-222). Prague, Czechoslovakia._
Blair, A. D. (1997). Scaling-up RAAMs. (Technical Report CS-97-192). Brandeis University,
Department of Computer Science.
Fedoseyeva, T. V. (1992). The problem of training neural network training to recognize word roots. In
_Neuron-like networks and neurocomputers (pp. 48-54). Kiev, Ukraine: V. M. Glushkov Institute of_
Cybernetics. (In Russian).
Feldman, J. A., & Ballard, D. H. (1982). Connectionist models and their properties. Cognitive Science, 6,
205-254.
Feldman, J. A. (1989). Neural Representation of Conceptual Knowledge. In L. Nadel, L. A. Cooper, P.
Culicover, & R. M. Harnish (Eds.), Neural connections, mental computation (pp. 68-103).
Cambridge, Massachusetts, London, England: A Bradford Book, The MIT Press.
Foldiak, P., & Young, M. P. (1995). Sparse Coding in the Primate Cortex. In M. A. Arbib (Ed.),
_Handbook of brain theory and neural networks (pp. 895-898). Cambridge, MA: MIT Press._
Frasconi, P., Gori, M., & Sperduti, A. (1997). A general framework for adaptive processing of data
structures. Technical Report DSI-RT-15/97. Firenze, Italy: Universita degli Studi di Firenze,
Dipartimento di Sistemi e Informatica.
Frolov, A. A., & Muraviev, I. P. (1987). Neural models of associative memory. Moscow: Nauka. (In
Russian).
Frolov, A. A., & Muraviev, I. P. (1988). Informational characteristics of neuronal and synaptic plasticity.
_Biophysics, 33, 708-715._
Frolov, A. A. (1989). Information properties of bilayer neuron nets with binary plastic synapses.
_Biophysics, 34, 868-876._
Gayler, R. W. (1998). Multiplicative binding, representation operators, and analogy. In K. Holyoak, D.
Gentner, & B. Kokinov (Eds.), Advances in analogy research: Integration of theory and data from
_the cognitive, computational, and neural sciences. (p. 405). Sofia, Bulgaria: New Bulgarian_
University. (Poster abstract. Full poster available at:
http://cogprints.soton.ac.uk/abs/comp/199807020).
Halford, G. S., Wilson W. H., & Phillips S. (in press). Processing capacity defined by relational
complexity: implications for comparative, developmental, and cognitive psychology. Behavioral and
_Brain Sciences._
Hebb, D. O. (1949). The organization of behavior. New York: Wiley.
-----
Hinton, G. E. (1981). Implementing semantic networks in parallel hardware. In G. E. Hinton & J. A.
Anderson (Eds.), Parallel models of associative memory (pp. 161-187). Hillside, NJ: Lawrence
Erlbaum Associates.
Hinton, G. E. (1990). Mapping part-whole hierarchies into connectionist networks. Artificial
_Intelligence, 46, 47-76._
Hinton, G. E., McClelland, J. L., & Rumelhart, D. E. (1986). Distributed representations. In D. E.
Rumelhart, J. L. McClelland, & the PDP research group (Eds.), Parallel distributed processing:
_Exploration in the microstructure of cognition 1: Foundations (pp. 77-109). Cambridge, MA: MIT_
Press.
Hopfield, J. J. (1982). Neural networks and physical systems with emergent collective computational
abilities. Proceedings of the National Academy of Sciences, USA, 79, 2554-2558.
Hopfield, J. J., Feinstein, D. I., & Palmer, R. G. (1983). "Unlearning" has a stabilizing effect in
collective memories. Nature, 304, 158-159.
Hummel, J. E., & Holyoak K. J. (1997). Distributed representations of structure: A theory of analogical
access and mapping. Psychological Review, 104, 427-466.
Kanerva, P. (1988). Sparse distributed memory. Cambridge, MA: MIT Press.
Kanerva, P. (1994). The Spatter Code for encoding concepts at many levels. In M. Marinaro and P.G.
Morasso (eds.), ICANN '94, Proceedings of International Conference on Artificial Neural Networks
(Sorrento, Italy), 1, pp. 226-229. London: Springer-Verlag.
Kanerva, P. (1996). Binary Spatter-Coding of Ordered K-tuples. In C. von der Malsburg, W. von Seelen,
J. C. Vorbruggen, & B. Sendhoff (Eds.). Proceedings of the International Conference on Artificial
_Neural Networks - ICANN'96, Bochum, Germany. Lecture Notes in Computer Science, 1112, 869-_
873. Berlin: Springer.
Kanerva, P. (1998). Encoding structure in Boolean space. In L. Niklasson, M. Boden, and T. Ziemke
(eds.), ICANN 98: Perspectives in Neural Computing (Proceedings of the 8th International
Conference on Artificial Neural Networks, Skoevde, Sweden), 1, pp. 387-392. London: Springer.
Kasatkin, A. M., & Kasatkina, L. M. (1991). A neural network expert system. In Neuron-like networks
_and neurocomputers (pp. 18-24). Kiev, Ukraine: V. M. Glushkov Institute of Cybernetics. (In_
Russian).
Kussul, E. M. (1980). Tools and techniques for development of neuron-like networks for robot control.
Unpublished Dr. Sci. dissertation. Kiev, Ukrainian SSR: V. M. Glushkov Institute of Cybernetics. (In
Russian).
Kussul, E. M. (1988). Elements of stochastic neuron-like network theory. In Internal Report "Kareta_UN" (pp. 10-95). Kiev, Ukraine: V. M. Glushkov Institute of Cybernetics. (In Russian)._
Kussul, E. M. (1992) Associative neuron-like structures. Kiev: Naukova Dumka. (In Russian).
Kussul, E. M. (1993). On some results and prospects of development of associative-projective
neurocomputers. In Neuron-like networks and neurocomputers (pp. 4-11). Kiev, Ukraine: V. M.
Glushkov Institute of Cybernetics. (In Russian).
Kussul, E. M., & Baidyk, T. N. (1990). Design of a neural-like network architecture for recognition of
object shapes in images. Soviet Journal of Automation and Information Sciences, 23(5), 53-58.
Kussul, E. M., & Baidyk, T. N. (1993). On information encoding in associative-projective neural
networks. (Preprint 93-3). Kiev, Ukraine: V. M. Glushkov Institute of Cybernetics. (In Russian).
Kussul, E. M., Baidyk, T. N., Lukovich, V. V., & Rachkovskij, D. A. (1993). Adaptive neural network
classifier with multifloat input coding. In Proceedings of NeuroNimes'93, Nimes, France, Oct. 25-29,
_1993. EC2-publishing._
Kussul, E. M., & Kasatkina, L. M. (1999). Neural network system for continuous handwritten words
recognition. (Submitted).
Kussul, E. M., & Rachkovskij, D. A. (1991). Multilevel assembly neural architecture and processing of
sequences. In A .V. Holden & V. I. Kryukov (Eds.), Neurocomputers and Attention: Vol. II.
_Connectionism and neurocomputers (pp. 577-590). Manchester and New York: Manchester_
University Press.
-----
Kussul, E. M., Rachkovskij, D. A., & Baidyk, T. N. (1991a). Associative-Projective Neural Networks:
architecture, implementation, applications. In Proceedings of the Fourth International Conference
_"Neural Networks & their Applications", Nimes, France, Nov. 4-8, 1991 (pp. 463-476)._
Kussul, E. M., Rachkovskij, D. A., & Baidyk, T. N. (1991b). On image texture recognition by
associative-projective neurocomputer. In C. H. Dagli, S. Kumara, & Y. C. Shin (Eds.), Proceedings
_of the ANNIE'91 conference "Intelligent engineering systems through artificial neural networks" (pp._
453-458). ASME Press.
Lansner, A., & Ekeberg, O. (1985). Reliability and speed of recall in an associative network. IEEE
_Trans. Pattern Analysis and Machine Intelligence, 7, 490-498._
Lavrenyuk, A. N. (1995). Application of neural networks for recognition of handwriting in drawings. In
_Neurocomputing: issues of theory and practice (pp. 24-31). Kiev, Ukraine: V. M. Glushkov Institute_
of Cybernetics. (In Russian).
Legendy, C. R. (1970). The brain and its information trapping device. In Rose J. (Ed.), Progress in
_cybernetics, vol. I. New York: Gordon and Breach._
Marr, D. (1969). A theory of cerebellar cortex. Journal of Physiology, 202, 437-470.
Milner, P. M. (1974). A model for visual shape recognition. Psychological Review, 81, 521-535.
Milner, P.M. (1996). Neural representations: some old problems revisited. Journal of Cognitive
_Neuroscience, 8, 69-77._
Palm, G. (1980). On associative memory. Biological Cybernetics, 36, 19-31.
Palm, G. (1993). The PAN system and the WINA project. In P. Spies (Ed.), Euro-Arch’93 (pp. 142-156).
Springer-Verlag
Palm, G. & Bonhoeffer, T. (1984). Parallel processing for associative and neuronal networks. Biological
_Cybernetics, 51, 201-204._
Plate, T. A. (1991). Holographic Reduced Representations: Convolution algebra for compositional
distributed representations. In J. Mylopoulos & R. Reiter (Eds.), Proceedings of the 12th
_International Joint Conference on Artificial Intelligence (pp. 30-35). San Mateo, CA: Morgan_
Kaufmann.
Plate, T. A. (1994). Distributed representations and nested compositional structure. PhD Thesis,
Department of Computer Science, University of Toronto.
Plate, T. A. (1995). Holographic Reduced Representations. IEEE Transactions on Neural Networks, 6,
623-641.
Plate, T. (1997). A common framework for distributed representation schemes for compositional
structure. In F. Maire, R. Hayward, & J. Diederich (Eds.), Connectionist systems for knowledge
_representation and deduction (pp. 15-34). Queensland University of Technology._
Plate, T.A. (to appear). Structured Operations with Vector Representations. Expert Systems: The
International Journal of Knowledge Engineering and Neural Networks, Special Issue on
Connectionist Symbol Processing.
Plate, T. (submitted). Randomly connected sigma-pi neurons can form associative memories.
Pollack, J. B. (1990). Recursive distributed representations. Artificial Intelligence, 46, 77-105.
Rachkovskij, D. A. (1990a). On numerical-analytical investigation of neural network characteristics. In
_Neuron-like networks and neurocomputers (pp. 13-23). Kiev, Ukraine: V. M. Glushkov Institute of_
Cybernetics. (In Russian).
Rachkovskij, D. A. (1990b). Development and investigation of multilevel assembly neural networks.
Unpublished Ph.D. dissertation. Kiev, Ukrainian SSR: V. M. Glushkov Institute of Cybernetics. (In
Russian).
Rachkovskij, D. A. (1996). Application of stochastic assembly neural networks in the problem of
interesting text selection. In Neural network systems for information processing (pp. 52-64). Kiev,
Ukraine: V. M. Glushkov Institute of Cybernetics. (In Russian).
Rachkovskij, D. A. (to appear). Representation and Processing of Structures with Binary Sparse
Distributed Codes. IEEE Transactions on Knowledge and Data Engineering, Special Issue on
Connectionist Models for Learning In Structured Domains.
-----
Rachkovskij, D. A. & Fedoseyeva, T. V. (1990). On audio signals recognition by multilevel neural
network. In Proceedings of The International Symposium on Neural Networks and Neural Computing
_- NEURONET'90 (pp. 281-283). Prague, Czechoslovakia._
Rachkovskij, D. A. & Fedoseyeva T. V. (1991). Hardware and software neurocomputer system for
recognition of acoustical signals. In Neuron-like networks and neurocomputers (pp. 62-68). Kiev,
Ukraine: V. M. Glushkov Institute of Cybernetics. (In Russian).
Shastri, L. & Ajjanagadde, V. (1993). From simple associations to systematic reasoning: connectionist
representation of rules, variables, and dynamic bindings using temporal synchrony. Behavioral and
_Brain Sciences, 16, 417-494._
Sjödin, G. (1998). The Sparchunk Code: a method to build higher-level structures in a sparsely encoded
SDM. In Proceedings of IJCNN'98 (pp. 1410-1415), IEEE, Piscataway, NJ: IEEE.
Sjödin, G., Kanerva, P., Levin, B., & Kristoferson, J. (1998). Holistic higher-level structure-forming
algorithms. In Proceedings of 1998 Real World Computing Symposium - RWC'98 (pp. 299-304).
Smolensky, P. (1990). Tensor product variable binding and the representation of symbolic structures in
connectionist systems. Artificial Intelligence, 46, 159-216.
Sperduti, A. (1994). Labeling RAAM. Connection Science, 6, 429-459.
Sperduti, A. & Starita, A. (1997). supervised neural networks for the classification of structures. IEEE
_Transactions on Neural Networks, 8, 714-735._
Tsodyks, M. V. (1989). Associative memory in neural networks with the Hebbian learning rule. Modern
_Physics Letters B, 3, 555-560._
Touretzky, D. S. (1990). BoltzCONS: Dynamic symbol structures in a connectionist network. Artificial
_Intelligence, 46, 5-46._
Touretzky, D. S. (1995). Connectionist and symbolic representations. In M. A. Arbib (Ed.), Handbook of
_brain theory and neural networks (pp. 243-247). Cambridge, MA: MIT Press._
Touretzky, D. S., & Hinton, G. E. (1988). A distributed connectionist production system. Cognitive
_Science, 12, 423-466._
Vedenov, A. A. (1987). "Spurious memory" in model neural networks. (Preprint IAE-4395/1). Moscow:
I. V. Kurchatov Institute of Atomic Energy.
Vedenov, A. A. (1988). Modeling of thinking elements. Moscow: Science. (In Russian).
von der Malsburg, C. (1981). The correlation theory of brain function. (Internal Report 81-2). Gottingen,
Germany: Max-Planck-Institute for Biophysical Chemistry, Department of Neurobiology.
von der Malsburg, C. (1985). Nervous structures with dynamical links. Ber. Bunsenges. Phys. Chem., 89,
703-710.
von der Malsburg, C. (1986) Am I thinking assemblies? In G. Palm & A. Aertsen (Eds.), Proceedings of
_the 1984 Trieste Meeting on Brain Theory (pp. 161-176). Heidelberg: Springer-Verlag._
Willshaw, D. (1981). Holography, associative memory, and inductive generalization. In G. E. Hinton &
J. A. Anderson (Eds.), Parallel models of associative memory (pp. 83-104). Hillside, NJ: Lawrence
Erlbaum Associates.
Willshaw, D. J., Buneman, O. P., & Longuet-Higgins, H. C. (1969). Non-holographic associative
memory. Nature, 222, 960-962.
-----
Figure captions
Figure 1. Growth of the density p' of 1s in the composite codevectors of higher hierarchical levels (see
equations 2.1 and 2.2). Here each codevector of the higher level is formed by bit disjunction of S
codevectors of the preceding level. Items at each level are uncorrelated. The codevectors of base-level
items are independent with the density of 1s equal to p. The number of hierarchical levels is L. (For any
given number of base-level items, the total number of 1s in the composite codevectors is obviously
limited by the number of 1s in the disjunction of base-level codevectors).
Figure 2. Hatched circles represent patterns of active units encoding items; formed connections are
plotted by arrowed lines.(A) Formation of a false assembly. When three assemblies (abd; bce; acf) are
consecutively formed in a neural network by connecting all active units of patterns encoding their items,
the fourth assembly abc (gridy hatch) is formed as well, though its pattern was not explicitly presented
to the network. (B) Preventing of a false assembly. If each of three assemblies is formed by connecting
only subsets of active units encoding the component items, then the connectivity of the false assembly is
weak. xyz denotes the subset of units encoding item x when it is combined with items y and _z.. The_
pairwise intersections of the small circles represent the false assembly.
Figure 3. A neural-network implementation of permutive conjunctive thinning. The same N-dimensional
binary pattern is activated in the input neural fields fin1 and fin2. It is a superposition of several component
codevectors. fin1 is connected to fout by a bundle of direct projective connections. fin2 is connected to fout
by a bundle of permutive connections. Conjunction of the superimposed component codevectors and
their permutation is obtained in the output neural field fout, where the neural threshold θ=1.5.
Figure 4. (A), (B). Algorithmic implementations of the additive version of the Context- Dependent
Thinning procedure. Parameter seed defines a configuration of shift permutations. For small K,
checking that r is unique would be useful.
Figure 5. A neural-network implementation of the additive version of Context-Dependent Thinning
procedure. There are four neural fields with the same number of neurons: two input fields fin1 and fin2, the
output field fout, the intermediate field fint. The neurons of fin1 and fout are connected by the bundle of
direct projective connections (1-to-1). fint and fout are also connected in the same manner. The same
binary pattern z (corresponding to superimposed component codevectors) is in the input fields fin1 and
**fin2. The intermediate field fint is connected to the input field fin2 by K bundles of permutive projective**
connections. The number K of required bundles is estimated in Table 3. Only two bundles are shown
here: one by solid lines and one by dotted lines. The threshold of fint neurons is 0.5. Therefore fint
accumulates (by bit disjunction) various permutations of the pattern z in fin2. The threshold of fout is equal
to 1.5. Hence this field performs conjunction of the pattern z from fin1 and the pattern of K permuted and
superimposed z from fint. z, 〈z〉, w correspond to the notation of Figure 4.
Figure 6. A neural-network implementation of the subtractive Context-Dependent Thinning procedure.
There are three neuron fields of the same number of neurons: two input fields fin1 and fin2, as well as the
output field fout. The copy of the input vector z is in both input fields. The neurons of fin1 and fout are
connected by the bundle of direct projective connections (1-to-1). The neurons of fin2 and **fout are**
connected by K bundles of independent permutive connections. (Only two bundles of permutive
connections are shown here: one by solid lines, and one by dotted lines). Unlike Figure 5, the synapses
of permutive connections are inhibitory (the weight is -1). The threshold of the output field neurons is
0.5. Therefore the neurons of z remaining active in fout are those for which none of the permutive
connections coming from z are active. As follows from Table 4, K is approximately the same for the
number S = 2,...,5 of component codevectors of certain density p.
-----
Figure 7: (A) Overlap of thinned composite codevectors and (B) Overlap of component subsets in
thinned composite codevectors - versus the overlap of the corresponding unthinned composite
codevectors for various versions of thinning procedures. CDTadd - the additive, CDTsub - the
subtractive, CDTadd-sl - the self-exclusive additive, CDTsub-sl - the self-exclusive subtractive CDT
procedure, Paired-M - permutive conjunctive thinning; the densities of component codevectors are
chosen to obtain M of 1s in the thinned codevector. For component codevectors, N=100000, M≈1000.
The number of component codevectors: 5. The results are averaged over 50 runs with different random
codevectors.
Figure 8: (A) Overlap of thinned composite codevectors and (B) Overlap of component subsets in
thinned composite codevectors for various "depth" of additive (CDTadd) and subtractive (CDTsub)
procedures of Context-Dependent Thinning. There are five components in the composite item. Therefore
the input composite codevector includes approximately 5M of 1s. The composite codevector is thinned
to have from 4M to M/4 of 1s. Two curves for thinning depth M are consistent with the corresponding
curves in Figure 7. For all component codevectors, N=100000, M≈1000. The results are averaged over
50 runs with different random codevectors.
-----
1,0
0,9
0,8
0,7
0,6
0,5
0,4
0,3
0,2
0,1
0,0
0 1 2 3 4 5 6 7 8 9 10
Number L of hierarchical levels
Figure 1.
�������������������������
�������������������������
�������������������������d
�������������������������
�����������������������
��������������������������
��������������������������
��������������������������e
��������������������������
���������������������������������������������������
���������������������������������������������������
��������������������������
��������������������������a
��������������������������
��������������������������
## (A)
Figure 2.
-----
.
## # fin1 fout fin2 #
1 0 0 0 1
2 1 0 1 2
3 0 0 0 3
4 1 1 1 4
5 1 0 1 5
6 1 1 1 6
7 0 0 0 7
8 0 0 0 8
... ... ... ... ...
N 1 1 1 N
## θθθθ=1.5
Figure 3.
The input codevector z = x1 ∨ **x2** ∨ ... ∨ **xS** **(A)**
The output (thinned) codevector is 〈z〉
Calculate K=ln(1-1/S)/ln(1-pS)
Set auxiliary vector w to 0
Seed the random-number generator: randomize(seed).
for (k=1, 2,..., K)
if ((r=rand()) ≠ 0)
for (i=1, 2,..., N)
wi = wi ∨ zi+r modulo N
for (i=1, 2,..., N)
〈z〉i = zi ∧ wi
The input codevector z = x1 ∨ **x2** ∨ ... ∨ **xS** **(B)**
Set the output (thinned) codevector 〈z〉 to 0.
Seed the random-number generator: randomize(seed).
while(|〈z〉| < M)
if ((r=rand()) ≠ 0)
for (i=1, 2,..., N)
〈z〉i = 〈z〉i ∨ (zi ∧ zi+r modulo N)
Figure 4. (A), (B).
-----
## # fin1 fout fint fin2 #
# z 〈〈〈〈z〉〉〉〉 w z
1 0 0 1 0 1
2 1 0 0 1 2
3 0 0 0 0 3
4 1 1 1 1 4
5 1 1 1 1 5
6 0 0 1 0 6
7 0 0 1 0 7
8 0 0 1 0 8
9 1 1 1 1 9
... ... ... ... ... ...
N 1 0 0 1 N
## θθθθ=1.5 θθθθ=0.5
Figure 5.
## # fin1 fout fin2 #
# z 〈〈〈〈z〉〉〉〉 z
1 0 0 0 1
2 1 0 1 2
3 0 0 0 3
4 1 1 1 4
5 1 1 1 5
6 0 0 0 6
7 0 0 0 7
8 0 0 0 8
9 1 0 1 9
... ... ... ... ...
N 1 1 1 N
Figure 6.
-----
1
0,9
0,8
0,7
0,6
0,5
0,4
0,3
0,2
0,1
0
## (A)
1
0,9
0,8
0,7
0,6
0,5
0,4
0,3
0,2
0,1
0
0 0,2 0,4 0,6 0,8 1
Overlap of composite codevectors
CDTadd
CDTsub
CDTadd-sl
CDTsub-sl
Paired-M
0 0,2 0,4 0,6 0,8 1
Overlap of composite codevectors
## (B)
Figure 7.
-----
1
0,9
0,8
0,7
0,6
0,5
0,4
0,3
0,2
0,1
0
## (A)
0 0,2 0,4 0,6 0,8 1
Overlap of composite codevectors
1
0,9
0,8
0,7
0,6
0,5
0,4
0,3
0,2
0,1
0
## (B)
0 0,2 0,4 0,6 0,8 1
Overlap of composite codevectors
Figure 8.
-----
|
{
"disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references', 'abstract'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1162/089976601300014592?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1162/089976601300014592, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "GREEN",
"url": "http://cogprints.org/1240/1/cdt-1sp.pdf"
}
| 2,001
|
[
"JournalArticle"
] | true
| 2001-02-01T00:00:00
|
[] | 27,966
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Economics",
"source": "s2-fos-model"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/026e0df72f523f1b772e9295344b7ba9877e1098
|
[
"Computer Science"
] | 0.854808
|
Incentives for Crypto-Collateralized Digital Assets
|
026e0df72f523f1b772e9295344b7ba9877e1098
|
Proceedings
|
[
{
"authorId": "31520104",
"name": "Philip N. Brown"
}
] |
{
"alternate_issns": [
"0377-2969"
],
"alternate_names": null,
"alternate_urls": null,
"id": "efad5e01-e2cf-44e8-9cfa-c5d9634707fc",
"issn": "2198-7432",
"name": "Proceedings",
"type": null,
"url": "http://paspk.org/Proceedings-of-PAS-39"
}
|
Digital currencies such as Bitcoin frequently suffer from high price volatility, limiting their utility as a means of purchasing power. Hence, a popular topic among cryptocurrency researchers is a digital currency design which inherits the decentralization of Bitcoin while somehow mitigating its violent price swings. One such system which attempts to establish a price-stable cryptocurrency is the BitShares market-pegged-asset protocol. In this paper, we present a simple mathematical model of the BitShares protocol, and analyze it theoretically and numerically for incentive effects. In particular, we investigate how the selection of two key design parameters function as incentive mechanisms to encourage token holders to commit their core BitShares tokens as collateral for the creation of new price-stabilized tokens. We show a pair of analytical results characterizing some simple facts regarding the interplay between these design parameters. Furthermore, we demonstrate numerically that in some settings, setting these design parameters is a complex, sensitive, and unintuitive task, prompting further work to more fully understand this design process.
|
## ***proceedings***
*Proceedings*
# **Incentives for Crypto-Collateralized Digital Assets [†]**
**Philip N. Brown**
Department of Computer Science at the University of Colorado, Colorado Springs, CO 80918, USA;
philip.brown@uccs.edu
- Presented at the 3rd annual Decentralized Conference, Athens, Greece, 30 October–1 November 2019.
���������
Published: 21 October 2019 **�������**
**Abstract:** Digital currencies such as Bitcoin frequently suffer from high price volatility, limiting their
utility as a means of purchasing power. Hence, a popular topic among cryptocurrency researchers is
a digital currency design which inherits the decentralization of Bitcoin while somehow mitigating its
violent price swings. One such system which attempts to establish a price-stable cryptocurrency is the
BitShares market-pegged-asset protocol. In this paper, we present a simple mathematical model of
the BitShares protocol, and analyze it theoretically and numerically for incentive effects. In particular,
we investigate how the selection of two key design parameters function as incentive mechanisms
to encourage token holders to commit their core BitShares tokens as collateral for the creation of
new price-stabilized tokens. We show a pair of analytical results characterizing some simple facts
regarding the interplay between these design parameters. Furthermore, we demonstrate numerically
that in some settings, setting these design parameters is a complex, sensitive, and unintuitive task,
prompting further work to more fully understand this design process.
**Keywords:** blockchain; cryptocurrency; decision theory; incentive design
**1. Introduction**
Since the introduction of the Bitcoin cryptocurrency in 2008 [ 1 ], digital currencies based on
blockchain technology have sprung into the public view in large part due to the dramatic price swings
they experience [ 2 ]. For example, in the two years from 1 January 2017 to 1 January 2019, the US-Dollar
denominated price of the Bitshares core token (BTS) gained approximately 7500%, then lost about 85%,
then gained 1600%, and then lost another 95% [ 3 ]. It is widely believed that this extreme volatility
hinders the consumer adoption of cryptocurrencies for ordinary payments, since the purchasing power
of an ordinary currency is expected to change very slowly, if at all [4].
Various methods have been proposed in recent years to create digital currencies which maintain
the decentralized and trustless nature of ordinary cryptocurrencies without suffering from the same
price volatility [ 5 – 9 ]. To date, the longest-running such project is the Bitshares market-pegged-asset
protocol (MPA), which has been nearly continuously operational in some form since mid-2014 [ 10 ].
The MPA system functions in a largely decentralized manner by allowing core token (BTS) holders to
lock away their BTS tokens as collateral in exchange for a loan of bitAsset tokens, which can then be
sold on the open market. The protocol has several mechanisms to regulate token supply and collateral
ratios for the purpose of maintaining the market price of the pegged assets at or near a price target.
While largely successful in maintaining a price peg (and considerably moreso than some early
competitors), two major problems have plagued the BitShares MPA system in recent years:
*•* **Loose price-pegging.** From 2014 to November 2018, the US Dollar-pegged smartcoin BitUSD
has maintained an average price *near* $1, but for extended periods has traded above $1.15,
with occasional drops to around $0.90. This is a considerably narrower range than typical
freely-floating cryptocurrencies, but stands as an important area for improvement.
*Proceedings* **2019**, *28* [, 2; doi:10.3390/proceedings2019028002](http://dx.doi.org/10.3390/proceedings2019028002) [www.mdpi.com/journal/proceedings](http://www.mdpi.com/journal/proceedings)
-----
*Proceedings* **2019**, *28*, 2 2 of 8
*•* **Undercollateralization.** To insulate collateral holders from each others’ risky behavior, the MPA
system has an in-built safety mechanism known as “global settlement.” In the event of
undercollateralization, the global settlement mechanism immediately closes all collateral positions
and establishes a fixed exchange rate between BTS and the bitAsset, effectively ceasing any form
of price-pegging. This mechanism was triggered in late 2018 on several BitShares-platform
smartcoins, including BitUSD (which at the time of writing has a market capitalization of
approximately $10 million).
In this paper, we present a highly-simplified model of the BitShares price-pegging mechanism,
and use it to perform an initial numerical study on some of the key design tradeoffs facing a smartcoin
protocol designer. In particular, we study how the selection of two protocol incentive parameters
impact the likelihood of undercollateralization events for the price-pegged smartcoins. Our central
contribution is to show that the incentive model in BitShares contains regions of high sensitivity; that is,
token holders’ decisions are discontinuous in incentive parameters in such a way that a small change
in parameter values can cause a large sudden jump in overall system risk levels.
**2. Model and Performance Metrics**
*2.1. Collateral Incentive Model*
Due to the complexity of modeling the full problem, in this paper we allow (with loss of
generality) the collateralization of a smartcoin to be contolled by a single centralized entity; we term
this entity the *agent* . This simplification allows us to closely examine some of the key incentive issues
involved in collateral management without any of the complications arising from a multi-agent setting.
Throughout, the reader may refer to Table 1 for a compact depiction of the agent’s perspective of
the protocol.
The agent is modeled as possessing an endowment of *Q* core tokens to be committed as collateral.
The agent must decide what *value* of stable tokens to be created using the endowment *Q* as collateral.
In other words, the agent must select a *collateral ratio R*, where
*⟨* Value of stable tokens created *⟩* = *[Q]* (1)
*R* [.]
The agent’s choice of *R* is lower-bounded by the blockchain *maintenance collateral ratio* parameter
*M* *≥* 1; that is,
*R* *≥* *M* . (2)
In the BitShares blockchain, a typical value of *M* at press time is 1.5 or 1.6.
Once the collateral is committed and stable tokens created, it is assumed that the agent trades the
stable tokens on the open market for their equivalent value in core tokens, resulting in the following
situation (in all of the following, the amounts are denominated in terms of their equivalent market
value of core tokens):
1. Core token collateral held by blockchain: *Q* .
2. Stable-token debt: *Q* / *R* .
3. Core tokens held freely: *Q* / *R* .
We consider the effect of a sudden shock *d* *>* 0 on the underlying core token price; this shock is
drawn from some distribution *D* . Here, we assume for simplicity that the shock occurs, and then the
agent’s collateralized position is immediately closed and all profits or losses are realized immediately.
We refer to a shock as “negative” if *d* *<* 1 (thus reducing the value of the collateral); otherwise we refer
to it as “positive.” According to the BitShares protocol, if the price shock causes the agent’s collateral
ratio to fall below 1, then all of the agent’s collateral is taken to cover the debt. This scenario is known
as *global settlement* .
-----
*Proceedings* **2019**, *28*, 2 3 of 8
If the shock is negative but less severe and the collateral ratio remains above 1 but falls below *M*,
then the agent’s debt is multiplied by a penalty factor known as the *maximum short-squeeze ratio* *S* *≥* 1;
this is known as a *margin call* .
Finally, if the shock is such that the agent’s collateral ratio remains above *M*, no penalty is assessed
and in our model the agent simply realizes the resulting loss or gain.
Mathematically, this penalty-free scenario results in a shock-adjusted debt of *dR* *[Q]* [, so the agent’s]
profit ratio *P* ( *R*, *d* ; *M*, *S* ) is
*P* ( *R*, *d* ; *M*, *S* ) = *[Q]* [/] *[R]* *[ −]* *[Q]* [/] [(] *[Rd]* [)] (3)
*Q*
= [1] 1 *−* [1] . (4)
*R* � *d* �
For moderate negative shocks, the agent must pay the additional *S* penalty for the margin call,
resulting in a profit ratio of
*P* ( *R*, *d* ; *M*, *S* ) = *[Q]* [/] *[R]* *[ −]* *[Q]* *[S]* [/] [(] *[Rd]* [)] (5)
*Q*
= [1] 1 *−* *[S]* . (6)
*R* � *d* �
Finally, for shocks that result in undercollateralization and global settlement, the agent must
surrender the entire collateral amount to cover the debt, resulting in
*P* ( *R*, *d* ; *M*, *S* ) = *[Q]* [/] *[R]* *[ −]* *[Q]* (7)
*Q*
= [1] (8)
*R* *[−]* [1.]
Thus, the full *ex post* profit ratio (that is, the profit given *d* ) is
*R* 1 *[−]* [1] if *d* *∈* [ 0, *S* / *R* )
1
*R* �1 *−* *[S]* *d* � if *d* *∈* [ *S* / *R*, *M* / *R* )
1
*R* �1 *−* [1] *d* � if *d* *∈* [ *M* / *R*, ∞ ) .
*P* ( *R*, *d* ; *M*, *S* ) =
(9)
When the dependence of *P* ( *R*, *d* ; *M*, *S* ) on *M* and *S* is clear, we shall sometimes simply write
*P* ( *R*, *d* ) .
**Table 1.** Tabular depiction of collateralization incentive model. In the first row, labeled *Start*, the agent
holds *Q* core blockchain tokens, no debt, and no collateral. In stage 2, the agent has committed *Q* tokens
as collateral and received a loan of *Q* / *R* price-stable tokens, which are then sold. In stage 3, the value of
the core token has changed by a factor of *d* ; since our accounting is being done in core tokens, the effect
of this is that the agent’s debt is modified by a factor of 1 / *d* . In stage 4, the debt position is closed,
resulting in either a profit or a loss for the agent depending on the *M* and *S* parameters.
**Stage** **Debt to Blockchain** **Locked Collateral** **Freely-Held Tokens**
Start 0 0 *Q*
Position Opened *Q* *R* *Q* *Q* *R*
After Price Shock *Rd* *Q* *Q* *Q* *R*
Position Closed 0 0 *P* ( *R*, *d* ; *M*, *S* ) Depends on *d* ; see (9)
-----
*Proceedings* **2019**, *28*, 2 4 of 8
*2.2. Decision Model*
Let the price shock *d* be drawn from some distribution *D* with probability density function
*f* : [ 0, ∞ ) *→* [ 0, ∞ ) . Given this distribution, *M*, and *S*, the agent’s goal is simple: select collateral ratio
*R* to maximize expected profit. Throughout, we use the notation *P* ( *R* ; *M*, *S* ) to denote the agent’s
expected profit given distribution *D* . That is, the agent’s optimal collateral ratio is
*R* *[∗]* ( *M*, *S* ) ≜ arg max E (10)
*R* *≥* *M* *d* *∼* *D* [[] *[P]* [(] *[R]* [,] *[ d]* [;] *[ M]* [,] *[ S]* [)]] [ .]
However, the system designer’s goal in selecting incentive parameters *M* and *S* is somewhat
more nuanced. On the one hand, the designer wishes to ensure that the agent selects a low-enough
collateral ratio *R* so that a reasonable number of price-stable tokens are created. On the other hand,
the designer wishes to ensure that the agent selects a *high* -enough *R* so that the price-stable token
remains fully collateralized in the event of a significant market downturn (i.e., the agent’s collateral is
sufficient to repay the debt even when a very low value of *d* is realized).
In this paper, we focus simply on the effect of *M* and *S* on the probability of an
undercollateralization event, and leave the study of the above nontrivial tradeoff for future work.
That is, we seek to characterize the probability of an undercollateralization event *given that the agent is*
*selecting the profit-maximizing R* *[∗]* ( *M*, *S* ) *.*
Let *F* : [ 0, ∞ ] *→* [ 0, 1 ] denote the cumulative distribution function of the price-shock distribution
*D* . Then the probability of undercollateralization in the presence of a profit-maximizing agent, denoted
*p* *u*, is given by
*S*
*p* *u* ( *M*, *S* ) ≜ *F* . (11)
� *R* *[∗]* ( *M*, *S* ) �
That is, undercollateralization occurs when *d* is realized *below* the threshold *S* / *R*, as in the
first condition of (9) ; Equation (11) is simply the cdf evaluated at this point with *R* = *R* *[∗]* ( *M*, *S* ) .
This threshold is triggered since all collateral is exhausted in paying the price-stable token debt *after*
the payment of the margin-call penalty *S* .
**3. Our Contributions**
The central goal of this paper is to characterize how the system designer’s choice of *M* and *S*
affect the emergent probability of undercollateralization *p* *u* . As a first step, in this paper we perform
a numerical study of the undercollateralization risk in the presence of a price disturbance that is
distributed lognormally. That is, we consider a situation in which the price disturbance *d* is given by
the formula
*d* = exp ( *s* ), (12)
where *s* is drawn from normal distribution *N* ( *µ*, *σ* [2] ) with mean *µ* and variance *σ* [2] . Under this
distribution, if *µ* = 0 we have the convenient fact for all *δ* *>* 0 that
*f* ( *δ* ) = *f* ( 1/ *δ* ) . (13)
That is, this distribution models symmetric *multiplicative* price shocks: when *µ* = 0, the price is
equally likely to double as it is to halve.
Mathematically, given *µ* and *σ* as above, the probability density function of this distribution is
given by
2 [�]
*f* ( *x* ) = 1 exp *−* [1] lo g ( *x* *−* *µ* ) . (14)
*σx* *√* 2 *π* � 2 � *s* �
-----
*Proceedings* **2019**, *28*, 2 5 of 8
*3.1. The Interplay between MCR and MSSR*
First, we note that regardless of the distribution *D* from which price shocks are drawn, the system
operator’s choice of *M* affects agent profits *only* if *S* *>* 1. That is, if the operator selects *S* = 1,
this effectively curtails their ability to have a nuanced effect on agent behavior.
**Proposition 1.** *If* *S* = 1 *, then for any fixed* *R* *, the agent’s expected profit* *P* ( *R* ; *M*, *S* ) *is a constant function*
*of M.*
**Proof.** Consider the *ex post* profit function given by (9), and let *S* = 1. Then the function collapses to
*P* ( *R*, *d* ; *M*, 1 ) =
=
*R* 1 *[−]* [1] if *d* *∈* [ 0, 1/ *R* )
1
*R* 1 �1 *−* [1] *d* � if *d* *∈* [ 1/ *R*, *M* / *R* )
*R* �1 *−* [1] *d* � if *d* *∈* [ *M* / *R*, ∞ )
*R* 1 *[−]* [1] if *d* *∈* [ 0, 1/ *R* )
1 (15)
� *R* �1 *−* [1] *d* � if *d* *∈* [ 1/ *R*, ∞ ),
which is clearly a constant function of *M* . Since the *ex post* profit is constant in *M*, then so must the
expected profit be as well.
Thus, it may be helpful to think of *S* as a “switch” which activates *M* . If *S* is not used
(i.e., no penalty is charged for margin calls), then *M* loses its effectiveness to influence agent behavior.
This suggests that the system operator should always maintain *S* *>* 1.
From here, let us assume that *S* *>* 1. Our next result considers the effect of *M* on agent profits.
Intuitively, one would expect that increasing *M* would strictly decrease the agent’s expected profit.
Proposition 1 demonstrates that this is not generally true (in particular, it fails when *S* = 1). However,
our next result shows that for a wide range of price shock distributions, the agent’s profit (for fixed *R* )
is strictly decreasing in *M* .
**Proposition 2.** *Let* *S* *>* 1 *, let* *R* *>* *M* *, and let price shock distribution* *D* *be such that its pdf has* *f* ( *M* / *R* ) *>* 0 *.*
*Then the agent’s expected profit is strictly decreasing in M:*
*∂* [(] *[M]* [/] *[R]* [)]
*[f]* *×* ( 1 *−* *S* ) *<* 0. (16)
*∂M* *[P]* [(] *[R]* [;] *[ M]* [,] *[ S]* [) =] *MR*
**Proof.** Let *f* ( *t* ) be the pdf of distribution *D* . Then when the agent selects collateral ratio *R*, the expected
profit ratio can be computed from (9) as
*S*
*P* ( *R* ; *M*, *S* ) ≜ E [1] *R* *f* ( *t* ) *dt* *−* *[S]*
*d* *∼* *D* [[] *[P]* [(] *[R]* [,] *[ d]* [;] *[ M]* [,] *[ S]* [)] =] *R* *[−]* � 0 *R*
*M*
*R*
� *S*
*R*
*f* ( *t* )
*dt* *−* [1]
*t* *R*
∞
� *M*
*R*
*f* ( *t* )
*dt* . (17)
*t*
The partial derivative of *P* ( *R* ; *M*, *S* ) can be computed from (17) as
*∂* [(] *[M]* [/] *[R]* [)] + *[f]* [(] *[M]* [/] *[R]* [)] (18)
*∂M* *[P]* [(] *[R]* [;] *[ M]* [,] *[ S]* [) =] *[ −]* *[S ]* *[f]* *MR* *MR*
which completes the proof of Proposition 2.
*3.2. Optimal Agent Behavior as a Function of MCR*
The complexities of the system render a full theoretical analysis challenging; accordingly,
we perform a simple numerical simulation study on the effect of the MCR parameter *M* on the
risk of undercollateralization events. To demonstrate how this parameter affects *p* *u*, we first plot the
agent’s expected profit ratio *P* ( *R* ; *M*, *S* ) as a function of collateral ratio *R* for a fixed distribution *D*
-----
*Proceedings* **2019**, *28*, 2 6 of 8
(with lognormal parameters *µ* = 0.033 and *σ* = 0.2, fixed *S* = 1.01, and varying *M* . See Figure 1 for a
depiction of this.
**Remark 1.** *The simulations depicted in Figure 1 demonstrate that the agent’s optimal choice of collateral ratio*
*can be extremely sensitive to the system operator’s choice of* *M* *. In particular, at a threshold of approximately*
*M* *≈* 1.53 *, P* ( *R* ; *M*, *S* ) *is discontinuous in M, increasing suddenly from about* 1.53 *to about* 1.85 *.*
**Figure 1.** Plots of agent’s expected profit ratio *P* ( *R* ; *M*, *S* ) as a function of *R* for fixed lognormal
distribution parameters of *µ* = 0.033 and *σ* = 0.2, and *S* = 1.01 and various values of *M* . Note that
when *M* is low, e.g., *M* = 1.4, the agent’s optimal decision (that is, the maximizer of the corresponding
trace on the chart, marked approximately by the red discs) is *R* *[∗]* ( *M*, *S* ) = *M* . However, increasing *M*
has the tendency to “pull down” the left-hand end of the trace. When *M* *≈* 1.5, this gives rise to a local
maximum in the profit function away from *R* = *M* ; around *M* *≈* 1.53, this local maximum becomes the
global maximum and the agent’s optimal collateral ratio “snaps” to the right, yielding *R* *[∗]* ( *M*, *S* ) *>* *M* .
*3.3. Undercollateralization Risk as a Function of MCR*
To understand how these agent decisions impact *p* *u* ( *M*, *S* ), we then compute the optimal collateral
ratio *R* *[∗]* ( *M*, *S* ) as a function of *M* for several values of *S*, and report these results in Figure 2. We find
several features of Figure 2 worthy of note.
**Remark 2.** *The probability of an undercollateralization event,* *p* *u* *, can be sharply (discontinuously) dependent*
*on* *M* *, but the presence of this discontinuity depends on the value of* *S* *. Considering the right-hand plot in*
*Figure 2, note that around* *M* *≈* 1.53 *, there is a discontinuity in the* *p* *u* *plot corresponding to* *S* = 1.01 *.*
*This indicates that* *p* *u* *is an extremely sensitive function of the incentive parameters* *M* *and* *S* *in complex ways*
*that are not* a priori *obvious and are deserving of further study.*
**Remark 3.** *The probability of undercollateralization,* *p* *u* *, is not monotone in* *S* *, and the nature of its*
*non-monotonicity is* not consistent over all values of *M* . *That is, consider* *M* *[′]* = 1.5 *; here, we have*
*that* *p* *u* *is locally increasing in* *S* *:* *p* *u* ( *M* *[′]*, 1.01 ) *>* *p* *u* ( *M* *[′]*, 1.005 ) *. Alternatively, consider* *M* *[′′]* = 1.6 *; here, we*
*have that* *p* *u* *is locally* decreasing *in* *S* *:* *p* *u* ( *M* *[′]*, 1.01 ) *<* *p* *u* ( *M* *[′]*, 1.005 ) *. This indicates that without careful*
*analysis, it may be nearly impossible to determine how to select incentive parameters*
**Remark 4.** *For low-enough* *S* *and very low* *M* *, the optimal collateral ratio appears to be* *R* *[∗]* ( *M*, *S* ) = *M* *.*
*Consider the left-hand plot of Figure 2: here, for* *S* *∈{* 1.005, 1.01 *}* *, the agent’s optimal value of* *R* *is the same.*
*That is, in this regime,* *S* *has no effect on collateral ratios. However, in the right-hand plot of Figure 2, in that*
-----
*Proceedings* **2019**, *28*, 2 7 of 8
*same regime, it is evident that the probability of undercollateralization actually* is *a function of* *S* *, and that higher*
*S leads to higher p* *u* *. That is, in this regime, this suggests that it is strictly better to set S* = 1 *.*
**Figure 2.** Plots of system behavior as a function of *M* for fixed lognormal distribution parameters of
*µ* = 0.033 and *σ* = 0.2 and various values of *S* . ( **Left** ) Agent’s optimal collateral ratio *R* *[∗]* ( *M*, *S* ) with
respect to *M* . Note that when *S* is low, the agent’s optimal action is to set *R* = *M*, but that larger values
of *S* render other, less-risky collateral ratios optimal. ( **Right** ) Probability of undercollateralization
*p* *u* ( *M*, *S* ) with respect to *M* . Several features are of note here: first, when *S* = 1.02, the probability of
undercollateralization is extremely low, despite the fact that only a 2% penalty is assessed the agent
in the event of a margin call. Second, when *S* *≤* 1.01, the probability of undercollateralization ( *p* *u* ) is
sharply dependent on the value of *M*, can be as high in these simulations as 3.5%, and for *S* = 1.01,
*p* *u* is discontinuous around *M* = 1.53. That is, the probability of undercollateralization is *extremely*
sensitive to the operator’s selection of *M* .
**4. Discussion and Future Work**
Our analytical results in Propositions 1 and 2 suggest that selecting *S* *>* 1 is a crucial element of
an effective incentive design; any value of *S* *>* 1 ensures that *M* is an effective tool for influencing the
behavior of agents.
Taken together, Remarks 1–4 illustrate that even on this carefully simplified model of the BitShares
incentive system, emergent behavior among agents can be extremely complex and unintuitive.
If behavior is challenging on a simple model, we expect that it is likely to be more challenging
as the model’s complexity increases to match the real-world functions of the system. In particular,
we recommend that careful attention be paid to the following key things:
*•* The behavior of token holders may be *extremely* sensitive to small changes in MCR. This may
be exploited to increase liquidity (i.e., by decreasing MCR slightly to incentivize the creation of
additional price-stable tokens), but it may be very difficult to predict its precise effects and can
easily increase the overall risk in the system.
*•* The effects of MSSR on token holder behavior may be unintuitive, and essential aspects of their
character may be highly dependent on MCR. In our simulations, we found that for low MCR, it is
strictly better to set MSSR very close to 1 rather than at a moderate value such as 1.01, as noted
in Remark 4. However, for a somewhat higher MCR, as we note in Remark 3, this situation is
reversed and risks decrease with *S* .
*Future Work*
Clearly, the most important priority is to adapt the present model to a multiagent dynamic
context which more accurately captures the complexities of the agents’ decisions. Modeling this as a
one-shot decision process has benefits in its simplicity, but clearly misses some important aspects of
-----
*Proceedings* **2019**, *28*, 2 8 of 8
the real-world system. In particular, in the real BitShares system, agents have the ability to update
their collateral ratios over time as the core token price evolves, and integrating this into our model will
be crucial to generate high-fidelity predictions.
Furthermore, there are intrinsic game-theoretic aspects to this system, due to the market dynamics
and the fact that in the real system, not all agents are charged the margin-call penalty every time. Lastly,
an important aspect that is missed in our analysis is that in the actual BitShares system, there is a much
more nuanced relationship between the external market and the actions of the agents committing
collateral to create price-stable tokens. That is, the market value of the price-stable tokens can actually
fluctuate somewhat, and this fluctuation likely dramatically impacts the incentives faced by core
token holders.
**Funding:** This research was funded by BitShares worker proposal 1.14.204.
**References**
1. Nakamoto, S. *Bitcoin: A Peer-to-Peer Electronic Cash System* [. Available online: https://bitcoin.org/bitcoin.pdf](https://bitcoin.org/ bitcoin.pdf)
(accessed on 30 June 2019).
2. Godsiff, P. *Bitcoin: Bubble or Blockchain* ; Smart Innovation, Systems and Technologies; Springer: Cham,
Switzerland, 2015; doi:10.1007/978-3-319-19728-9_16.
3. CoinMarketCap. Available online: [https://coinmarketcap.com/currencies/bitcoin/ (accessed on](https://coinmarketcap.com/currencies/bitcoin/)
30 June 2019).
4. Athey, S.; Parashkevov, I.; Sarukkai, V.; Xia, J. *Bitcoin Pricing, Adoption, and Usage: Theory and Evidence* ;
Stanford University Graduate School of Business Research Paper; Stanford University: Stanford, CA,
USA, 2016.
5. Lee, J. *Nu Whitepaper* [; 2014. Available online: https://nubits.com/assets/nu-whitepaper-23_sept_2014-en.](https://nubits.com/assets/nu-whitepaper-23_sept_2014-en.pdf)
[pdf (accessed on 30 June 2019).](https://nubits.com/assets/nu-whitepaper-23_sept_2014-en.pdf)
6. Chohan, U.W. *Are Stable Coins Stable?* Notes on the 21st Century (CBRi), Preprint Available at SSRN. Available
[online: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3326823. (accessed on 12 February 2019).](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3326823)
7. [Libra White Paper. Available online: https://libra.org/en-US/wp-content/uploads/sites/23/2019/06/](https://libra.org/en-US/wp-content/uploads/sites/23/2019/06/LibraWhitePaper_en_US.pdf)
[LibraWhitePaper_en_US.pdf (accessed on 30 June 2019).](https://libra.org/en-US/wp-content/uploads/sites/23/2019/06/LibraWhitePaper_en_US.pdf)
8. [The MakerDAO White Paper. Available online: https://makerdao.com/whitepaper/DaiDec17WP.pdf](https://makerdao.com/whitepaper/DaiDec17WP.pdf)
(accessed on 30 June 2019).
9. Klages-Mundt, A.; Minca, A. (In)Stability for the Blockchain: Deleveraging Spirals and Stablecoin Attacks.
*arXiv* **2019**, arXiv:1906.02152
10. The BitShares Blockchain Foundation. *The BitShares Blockchain* [. Available online: https://www.bitshares.](https://www.bitshares.foundation/papers/BitSharesBlockchain.pdf)
[foundation/papers/BitSharesBlockchain.pdf (accessed on 30 June 2019).](https://www.bitshares.foundation/papers/BitSharesBlockchain.pdf)
*⃝* c 2019 by the author. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
[(CC BY) license (http://creativecommons.org/licenses/by/4.0/).](http://creativecommons.org/licenses/by/4.0/.)
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3390/proceedings2019028002?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/proceedings2019028002, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "HYBRID",
"url": "https://www.mdpi.com/2504-3900/28/1/2/pdf?version=1571629186"
}
| 2,019
|
[] | true
| 2019-10-21T00:00:00
|
[
{
"paperId": "42e734c0a243f7d2cd837e8e8453251257b241f9",
"title": "Are Stable Coins Stable?"
},
{
"paperId": "4e430b64bac6fdf86a947ee81b9150ed1451704b",
"title": "(In)Stability for the Blockchain: Deleveraging Spirals and Stablecoin Attacks"
},
{
"paperId": "8a7c799b5734dfd961606ace218deab192e06ec5",
"title": "Bitcoin Pricing, Adoption, and Usage: Theory and Evidence"
},
{
"paperId": "2359ad394ab9a17e112c249a78eb5c5d3e55669c",
"title": "Bitcoin: Bubble or Blockchain"
},
{
"paperId": "433561f47f9416a6500c8350414fdd504acd2e5e",
"title": "Bitcoin Proof of Stake: A Peer-to-Peer Electronic Cash System"
},
{
"paperId": "ecdd0f2d494ea181792ed0eb40900a5d2786f9c4",
"title": "Bitcoin : A Peer-to-Peer Electronic Cash System"
},
{
"paperId": null,
"title": "CoinMarketCap"
}
] | 8,121
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Economics",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Economics",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/026f542ead119d0d53a6c7a35496d207b8b0b053
|
[
"Computer Science",
"Economics"
] | 0.873319
|
DeFi Risk Transfer: Towards A Fully Decentralized Insurance Protocol
|
026f542ead119d0d53a6c7a35496d207b8b0b053
|
International Conference on Blockchain
|
[
{
"authorId": "2038268968",
"name": "Matthias Nadler"
},
{
"authorId": "2129392146",
"name": "Felix Bekemeier"
},
{
"authorId": "2083839088",
"name": "Fabian Schär"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"ICBC",
"IEEE Int Conf Blockchain Cryptocurrency",
"IEEE International Conference on Blockchain and Cryptocurrency",
"Int Conf Blockchain"
],
"alternate_urls": null,
"id": "f1ab8d75-7f15-4bb4-ad88-e834ec6ed604",
"issn": null,
"name": "International Conference on Blockchain",
"type": "conference",
"url": null
}
|
In this paper, we propose a fully decentralized and smart contract-based insurance protocol. We identify various issues in the Decentralized Finance (DeFi) insurance context and propose a solution to overcome these shortcomings. We introduce an economic model that allows for risk transfer without any external dependencies or centralized intermediaries. In particular, our proposal does not need any sort of subjective claim assessment, community voting or external data providers (oracles). Moreover, it solves the problem of over-insurance and proposes various ways to mitigate the capital inefficiencies usually seen with DeFi collateral. The work takes inspiration from peer-to-peer (P2P) insurance and collateralized debt obligations (CDO). We formally describe the protocol, assess its efficiency and key properties and present a reference implementation. Finally, we address limitations, extensions and ideas for further research.
|
# DeFi Risk Transfer: Towards A Fully Decentralized Insurance Protocol
### Fabian Schär
_Center for Innovative Finance (CIF)_
_University of Basel_
Basel, Switzerland
f.schaer@unibas.ch
### Matthias Nadler
_Center for Innovative Finance (CIF)_
_University of Basel_
Basel, Switzerland
matthias.nadler@unibas.ch
### Felix Bekemeier
_Center for Innovative Finance (CIF)_
_University of Basel_
Basel, Switzerland
felix.bekemeier@unibas.ch
**_Abstract—In this paper, we propose a fully decentralized and_**
**smart contract-based insurance protocol. We identify various**
**issues in the Decentralized Finance (DeFi) insurance context**
**and propose a solution to overcome these shortcomings. We**
**introduce an economic model that allows for risk transfer**
**without any external dependencies or centralized intermediaries.**
**In particular, our proposal does not need any sort of subjective**
**claim assessment, community voting or external data providers**
**(oracles). Moreover, it solves the problem of over-insurance and**
**proposes various ways to mitigate the capital inefficiencies usually**
**seen with DeFi collateral. The work takes inspiration from**
**peer-to-peer (P2P) insurance and collateralized debt obligations**
**(CDO). We formally describe the protocol, assess its efficiency**
**and key properties and present a reference implementation.**
**Finally, we address limitations, extensions and ideas for further**
**research.**
**_Index Terms—Blockchain, DeFi, Decentralized Insurance, Risk_**
**Transfer, Smart Contracts**
I. INTRODUCTION
Decentralized Finance (DeFi) refers to public blockchainbased financial infrastructure that uses smart contracts to
replicate traditional financial services in a more open, interoperable, and transparent way [1]. These smart contract-based
services are usually referred to as protocols. They provide
basic building blocks such as the opportunity to swap assets or
allocate liquidity efficiently and can be reused and combined in
any way. While decentralized exchanges and lending markets
are arguably among the most prominent protocols and get a
lot of attention, there are other crucial building blocks that are
required for a well-functioning financial infrastructure. One of
these building blocks is the ability to transfer risks.
Consider the following general example: An economic agent
has an investment opportunity that may result in a small
loss or a large gain. Further assume that both outcomes have
the same probability. The expected return would be positive
and a risk-neutral (or risk-seeking) agent would be willing to
engage. However, if the same opportunity is instead presented
to a risk-averse agent, they may decline and forego a positive
expected return due to the cost of uncertainty. If a financial
market allows risk to be transferred, there is a simple solution.
The risk-averse person can approach an entity with a higher
risk tolerance and offer them a premium in return for their
willingness to bear the risk. They essentially share the positive
expected return and the risk would be borne by the entity with
the higher risk tolerance.
Similarly, a blockchain-based financial infrastructure becomes more efficient if smart contract risks are transferable.
Risk-averse investors could share some of their expected return
as a compensation for an insurance policy that covers the
smart contract risks of the respective liquidity pool. DeFi
users who are willing to bear additional risk could generate
a higher yield. The existence of a market for risk transfer
would be beneficial for everyone, as it allows all DeFi users
to structure their portfolio in accordance with their individual
risk preferences.
There already exists a relatively large number of smart
contract-based insurance protocols, including but not limited
to Nexus Mutual [2], Nsure [3], cozy.finance [4], Unslashed
Finance [5] and Risk Harbor [6]. While some of these protocols offer innovative solutions and have provided valuable
contributions to the DeFi protocol space, they are arguably not
fully decentralized and face various challenges.
_First, insurance requires that the insurer can credibly_
demonstrate its ability to cover potential losses at all times.
Centralized insurance is based on a combination of reputation
and regulation. Moreover, centralized insurance companies
rely on active asset and risk management to strike a balance
between liquidity and capital efficiency. DeFi, on the other
hand, is built on a pseudonymous system with little to no legal
recourse. It relies on transparency and (over-)collateralization.
Consequently, many implementations face trade-offs between
capital efficiency, security and special privileges that allow for
manual interventions.
_Second, DeFi insurance protocols usually struggle with_
claim assessment. Generally speaking there are two options.
(a) The insurance policy is parametric and relies on oracles
and (b) the outcome is decided through a vote, by so-called
claim assessors. Both approaches are quite subjective and
can easily lead to false outcomes. The former introduces
dependencies to external data providers and does not reflect
true damages due to its parametric nature. The latter relies on
a voting process among pseudonymous actors that can assume
various roles within (and outside) the system. Moreover, truly
decentralized voting will be either subject to sybil attacks [7]
or whale dominance with potentially problematic incentives.
-----
There are good arguments, why neither the oracle-based nor
the claim assessor-based approach should be considered fully
decentralized.
_Third, most protocols cannot prevent over-insurance. DeFi_
users can buy cover for protocols to which they have no exposure. This can create problematic incentives and – depending
on the jurisdiction – result in conflict with the law.
In this paper, we propose a novel DeFi insurance protocol
that solves these issues. To the best of our knowledge it is
the first proposal for a fully decentralized insurance protocol
with no external dependencies. As part of this research project,
we have also built a basic reference implementation of the
protocol. The implementation can be found in the appendix.
After this short introduction, we discuss related works from
the DeFi, insurance and finance literature. In Section III we
turn to the technical part, describe the protocol and perform
a gas efficiency analysis. In Section IV we study external
incentives for liquidity providers and derive the implicit cost of
liquidity provision for various pools involving our protocol’s
tranche tokens. In Section V we discuss our results, potential
extensions and limitations. Finally, we conclude in Section VI.
II. RELATED WORK
The motivation for a DeFi insurance protocol is closely
linked to discussions on smart contract and DeFi risks,
protocol failures and shock propagation. These issues have
received an increasing amount of research attention and are
an important part of the academic discourse on DeFi [8]–
[12]. Our protocol can mitigate some of the consequences
by allocating risk in a more efficient way. Moreover, market
prices for risk premiums can serve as an indication of the
perceived risk; similar to prediction markets. With regard to
yield-generating lending protocols, different authors discuss
the risks of illiquidity, dependencies and misaligned incentives
[9], [13]–[15]. Moreover, there are various papers discussing
oracle reliability and potential manipulation [16], [17]. Our
proposal does not have any dependencies, allows the insurant
to hedge against oracle exposure, and even works in situations
where the insured protocols become illiquid.
Existing DeFi insurance protocols are mostly based on
principles of mutual insurance, where users participate in
the commercial success of the protocol. In theory, mutuals
can have certain advantages for large risk pools [18], in the
presence of transactional costs and governance issues [19], and
in addressing problems of adverse selection [20]. However,
due to centralized economic value capture in most mutuals,
problems potentially remain with respect to default risks [21].
In a DeFi context, mutual-based insurance protocols usually
rely on centralized or vote-based claim assessment and may
depend on know your customer (KYC) principles or introduce
other forms of dependencies.
Our protocol is fundamentally different from a mutual
insurance. There is no centralized economic value capture and
the protocol does not accumulate reserves. The general concept
of our protocol is inspired by peer-to-peer (P2P) insurance and
financial instruments with tranches, such as collateralized debt
obligations (CDO).
In a P2P insurance model, individuals pool their insurance
premiums and use these funds to cover individual damages.
P2P risk transfer is still at a very early stage of research,
with seminal works including [22]–[27]. Several authors have
started to formally explore the organizational structure, optimality and pitfalls of P2P insurance [28]–[30].
Our protocol is based on similar principles. In particular, we
make use of different risk preferences and levels that allow
individuals to pool their risks without the explicit need for
an intermediary. However, there is an important difference
between P2P insurance and our approach: P2P insurance
usually covers individual risks. As such, P2P insurance is built
on the general assumption that damages within the collective
are uncorrelated and that premiums of the unaffected insurants
can be used to compensate the ones that have suffered losses.
Our protocol insures large scale risks that will affect all
insurance holders. Consequently, we need explicit roles in
accordance with the individuals’ risk preferences. This is
achieved by creating tranches with different seniorities and
security guarantees.
As such, our protocol incorporates some aspects of CDOs.
CDOs have been discussed extensively in the subject-related
literature [31]–[34]. They split cash flows among tranches
with different seniority. The most senior tranches are honored
first and the most junior tranches bear the losses. In addition
to traditional use cases, such as CDOs for bank refinancing,
insurance risk also appears to be a suitable use-case for CDOs
[35]. Likewise, CDOs are used widely in various applications,
also outside traditional financial markets. For example, CDOs
have already been discussed related to the support of microcredits [36].
This combination of P2P insurance, seniority-based
promises and DeFi specifics builds the foundation of our
protocol and allow us to propose a fully decentralized DeFi
insurance.
III. PROTOCOL
In this section, we present a decentralized risk hedging
protocol, based on tranched insurance. First, we provide a
quick overview and describe the core functionality of the
protocol. Second, we take a more technical perspective and describe individual function calls and state transitions. Third, we
discuss potential technical extensions and trade-offs. Fourth,
we provide a short efficiency analysis and discussion of the
protocol’s computational costs (gas fees).
_A. Protocol Overview_
The general idea of our insurance protocol is to pool assets
from two third-party protocols, and allow users to split the
pool redemption rights into two tranches: A and B. If any
of the third-party protocols suffer losses during the insurance
period, those losses will be primarily borne by the B-tranche
holders. A-tranche holders will only be negatively affected
if 50% or more of the pooled funds are irrecoverable, or if
-----
both protocols become temporarily illiquid and face (partial)
losses. We effectively split the redemption rights into a riskier
and less risky version and allow the market for A- and Btranches to determine the fair risk premium in line with the
users’ expectations.
The protocol consists of three main phases: risk splitting,
_investing/divesting and redemption._
In the risk splitting phase, anyone may allocate their
preferred number of C-tokens to the insurance protocol. These
_C-tokens represent the underlying asset, e.g., a stablecoin or_
Ether. In exchange, the users receive equal denominations in
_A- and B-tranches, thereby ensuring that an equal number of_
both tranches will be created. A and B are ERC-20 compliant
tokens and can be transferred separately. This allows the users
to swap the tokens on decentralized exchanges to obtain a
relative allocation of A- and B-tranches that reflects their risk
preferences.
At the beginning of the invest/divest phase, the insurance
protocol allocates the accumulated collateral of C equally into
two protocols. In return, it receives interest-bearing tokens
(wrapped liquidity shares) from each protocol. We denote
these shares as Cx and Cy.
To make things less abstract, consider the following example: A stablecoin (C) gets allocated to two distinct yieldgenerating lending protocols. In return, the insurance protocol
receives the respective interest-bearing tokens (Cx and Cy).
They are locked in the insurance contract, where they will
accumulate interest over time. At the end of the invest/divest
phase, the insurance protocol tries to liquidate the wrapped
shares. This is a necessary step in preparation for the redemption of the A- and B-tranches.
In a third step, the protocol enters the redemption phase.
The goal of this phase is to compute potential losses and allow
the A- and B-tranche holders to claim their respective share of
the underlying. It is important to understand that the redemption phase can be executed in one of two distinct modes. Mode
selection depends on the success of the liquidation at the end of
the invest/divest phase. If the liquidation of Cx and Cy works
as expected and the insurance protocol receives the collateral
tokens C, then redemption can be conducted in liquid mode. In
this mode, it is straightforward to distribute the interest equally
among all A- and B-tranche holders. Similarly, potential losses
can be computed and primarily allocated to B-tranche holders.
If the liquidation of Cx or Cy fails, the protocol enters fallback
_mode. This can happen if a third-party protocol suffers from a_
liquidity crunch or if an external contract changes the expected
behavior. In fallback mode, users redeem their tranche tokens
directly for their preferred mix of Cx and Cy tokens. The
higher tranche seniority of A-tranches is ensured through a
timelock-based redemption sequence. In a first step, A-tranche
holders get to choose if they want to claim their share in Cx,
_Cy or a mix of the two. After the timelock is over, B-tranche_
holders can claim what is left.
_B. Technical Implementation_
A reference implementation of the insurance contract is
available in the appendix and demonstrates how our protocol
can be used to provide insurance for two yield-generating
protocols that wrap the Maker DAO stablecoin Dai [37],
denoted as C. The two yield-generating protocols are Aave
version 2 [38] with aDai and Compound Finance [39] with
cDai, denoted as Cx and Cy respectively. The reference implementation includes the full Solidity code for the Ethereum
Virtual Machine-based (EVM) contract and can be used as
a starting point for developers who want to create their own
insurance contracts using a similar approach.
In this subsection we provide an overview of the reference
implementation’s technical specifications, including the functions, variables and states. We present this information in a
chronological order, following the timeline presented in Figure
2. The states are referred to as: ReadyToAccept, ReadyToInvest,
_MainCoverActive, ReadyToDivest, Liquid, FallbackOnlyA and_
_FallbackAll. Note that strictly speaking a smart contract cannot_
automatically transition from one state to another based on
the passage of time; this is a fundamental limitation of smart
contract technology. Any state change on the contract has
to be initiated by a function call. Our implementation works
around this by defining states as a set of successfully callable
functions and reverting function calls, if they are outside the
allowed time windows. Hence, the set may change based on
time conditions.
Before the first state, the initial parameters must be defined
and contract deployed. The parameters include the addresses
of the tokens involved in the contract, as well as the absolute
values for the timestamps when state transitions occur. These
forced state transitions are represented in Figures 1 and 2 as
_S, T1, T2 and T3, where S < T1 < T2 < T3. Furthermore, the_
constructor deploys two ERC-20 token contracts for A- and
_B-tranches, with the insurance contract as the sole, immutable_
owner. This means that only the insurance contract can mint
and burn the tranche tokens.
After deployment, the contract is in the ReadyToAccept
state and the public function splitRisk() is available
for anyone to call. The input parameter for the function is
an amount of C tokens. The splitRisk() function then
transfers this amount of C tokens from the caller to the
insurance contract and issues a number of A- and B-tranche
tokens equal to half that amount to the caller. For example, if
the input is 100, the function will transfer 100 C tokens from
the caller to the insurance contract and issue 50 tranche A
tokens and 50 tranche B tokens to the caller. It is important
to note that the act of calling the splitRisk() function
does not provide the user with any form of insurance cover.
In order to obtain insurance cover – or to assume more risk
– the user must sell or trade a portion of their tranche A or
tranche B tokens.
When time S is reached, the contract transitions to the
ReadyToInvest state and users can no longer mint new
tranche tokens. The invest() function is available during
-----
Fig. 1. State Transition Diagram: Represents state transitions and their respective function sets.
this state and it is tailored to the specific needs of the protocols
that are part of the insurance contract, with the goal of splitting
the deposited C tokens equally among the protocols. In the
reference implementation, the function will send half of the
available C to Aave and the other half to Compound in
exchange for their respective yield-bearing tokens, Cx and Cy.
After a successful invest() call, the insurance contact holds
_Cx and Cy of equal value and does no longer hold C. Calling_
the invest function incurs a transaction fee, paid by the caller
while the benefits of the call are shared among all participants.
To avoid the problem of a first mover disadvantage, to ensure
that the call is executed in a timely fashion and to split the
costs equally among all participants, the invest() function
should compensate the caller for executing the transaction.[1]
The unlikely case in which no successful invest() call is
made before the forced state transition at T1 will be covered
later in this subsection.
When a successful invest() call is made, the contract
transitions to the MainCoverActive state and sets the variable isInvested = true. The contract is now exposed to
the risks of the third-party protocols and the main period of
insurance cover for the A-tranches begins. In this state, no
functions can be called on the contract. However, the A- and
_B-tranches remain transferable._
At time _T1,_ the contract will transition from the
MainCoverActive state to the ReadyToDivest state,
where the divest() function can be invoked. It has a
similar structure to the invest() function, but instead of
depositing the underlying assets into the third-party protocols,
divest() tries to withdraw the underlying assets including
any accumulated yield from the protocols. A divest() call
is considered successful if no errors occur while withdrawing
the assets and if both Cx and Cy have been fully converted
back to C.
A successful divest() call immediately transitions the
protocol to the Liquid state by setting inLiquidMode =
true. In this state, the allocation of the redeemed assets to
the A- and B-tranches is deterministic and can be calculated
as part of the divest() call. Let us define CS as the total
initially invested amount, CT1 as the total redeemed amount
1We did not include a compensation mechanism in the reference implementation. When implemented, it should cover at least the base fee of the
transaction plus a fixed amount for the tip.
and i as the interest. We can then differentiate between three
cases and determine the payouts for each case, as shown in
Table I. The payout per A- and B-tranche token is stored on
the contract and can be accessed using the variables cPayoutA
and cPayoutB, respectively. During the liquid state, users can
call the claim() function which accepts an amount for Aand B-tranches as input. If the caller is in control of at least
the specified amount of tranches, the contract will burn these
tranches and transfer the payout to the caller. For convenience,
a claimAll() function is available and will internally call
the claim() function with the caller’s current balance of
tranches.
If no successful divest() call is made during the
ReadyToDivest state, a forced transition occurs at T2
and the protocol enters fallback mode, which starts in state
FallbackOnlyA. In fallback mode, the protocol has no
knowledge about the value of its interest-bearing tokens relative to the initial investment. Therefore, instead of assigning a
payout to the tranches, the tranche holders can choose which
of the two interest-bearing tokens they would like to redeem.
Based on the total amount of tranche tokens and the
remaining interest-bearing tokens, the contract determines a
fixed redeem-ratio for each of the two interest-bearing tokens.
These ratios are stored on the contract as cxPayout and
_cyPayout and are defined as the total amount of the respective_
asset, divided by half of the total amount of tranches. For
example, assume 50 A- and 50 B-token have been minted and
the contract holds 20 Cx and 1500 Cy. A tranche can now be
redeemed for 0.4 Cx or 30 Cy. Once all tranches are redeemed
there are no interest-bearing tokens left on the contract. A- and
_B-tranches can be redeemed for the same amount. However,_
during the FallbackOnlyA state, as the name suggests, only
_A-tranches can be redeemed for interest-bearing tokens with_
the function claimA(). As an input for this function, the
caller specifies how many of their A-tranches they want to
redeem for Cx and how many for Cy. The contract then burns
the tranches and transfers the assets according to the redeemratios.
At time T3, if the contract is in fallback mode, the final
transition happens to the FallbackAll state. This state is
identical to FallbackOnlyA with the only difference that
_B-tranches can now also be redeemed via the claimB()_
function.
-----
deployment
_D_
invest()
_S_
claimAll()
if inLiquidMode = TRUE
divest()
_T1_
_t_
_T2_ _T3_
if inLiquidMode = FALSE
claimA()
claimB()
Fig. 2. Sequential actions in liquid mode (top, divest() successful) and fallback mode (bottom, divest() unsuccessful).
Finally, to ensure we never end up in a state where the assets
cannot be recovered, we need to define a state transition from
ReadyToInvest to Liquid if the invest() function
was not successfully called. This transition happens after T1
if isInvested == false and allows the users to reclaim
their initially invested funds.
_C. Extensions and Trade-Offs_
To obtain insurance cover, a protocol user must sell their Btranches. A possible extension to the insurance contract would
be to use intra-transaction composability and connect it to a
decentralized exchange. This would allow users to sell their
_B-tranches in the same transaction as the splitRisk()_
function. However, note that any additions to the insurance
contract will introduce additional risk. Keeping the contract as
simple as possible and reducing dependencies to a minimum
will help to manage this risk. We argue that most extensions
which introduce new dependencies should be implemented at
the user interface level in a separate contract.
Consider the following example: Let us assume that we want
to create a function to insure an amount of C tokens. We create
a new contract with a function that uses a flash loan [15]
for twice the amount and calls splitRisk(). In the same
function, the B-tranches are sold to a decentralized exchange
and the A-tranches transferred to the caller. Finally, the flash
loan is repaid, using the proceeds from the sale and the funds
from the initial caller. The additional contract can be developed
and deployed independently of the insurance contract. This
separation offers more flexibility and introduces no additional
risks for other users. The trade-off here is that the transaction
fees might be slightly higher, as external calls are more costly
than internal ones.
_D. Transaction Costs_
Depositing funds into a protocol incurs a transaction fee,
which is imposed by the blockchain network and expressed in
units of computation – commonly called gas. This transaction
fee can vary slightly based on circumstantial parameters, but
it largely depends on the computational complexity of the
transaction. Depositing funds into our reference implementation via the splitRisk() function costs around 83,000
gas. Depositing to Aave or Compound directly incurs a fee
of 249,000 or 156,000 gas, respectively. While calling the
invest() function is expensive (488,000 gas), this cost can
be split among all users in the insurance contract. Similar
to yield aggregation protocols [40], the insurance contract
becomes more gas efficient, the more users participate and
even for just a few users, we expect the minting of insured
tokens to be cheaper than minting uninsured tokens.
IV. LP-INCENTIVES AND DIVERGENCE LOSS
Recall that users must mint A- and B-tranches in equal
proportions. Consequently, they will only be able to reach
token allocations in line with their risk preferences if there
is a liquid market. Insufficient liquidity would lead to large
price spreads (or slippage). Hence, there is a need for market
makers, or more generally liquidity providers.
In what follows, we analyze the incentives for liquidity
provision of A- and B-tranches on constant product market
makers (CPMM), a special form of automated market makers
(AMM) [41]. Note, that CPMMs are only one of many possibilities; tranche token markets could emerge on any trading
infrastructure. However, there are a few reasons why CPMMs
are of particular importance. First, they usually handle a large
part of the on-chain trading volume. Second, CPMMs allow
for composable calls and will always be able to quote a price
TABLE I
THE THREE POTENTIAL OUTCOMES FOR LIQUID MODE
**Case** **Payoff A** **Payoff B** **Description**
_CT1 ≥_ _CS_ _C2T1_ _C2T1_ Proceeds are split equally among all tranche token holders. Both
tranches are treated equally.
_CS > CT1 >_ _[C]2[S]_ _C2S_ + i _CT1 −_ � _C2S_ + i� _A-tranche holders get fully compensated and receive yield payment._
_B-tranche holders receive a proportion of their initial stake._
_CT1 ≤_ _[C]2[S]_ _CT1_ 0 Proceeds are used to partially compensate A-tranche holders. This can
only occur if both yield-generating protocols suffer losses.
-----
for any (input) amount. Third, CPMMs can be set up in a
completely decentralized way and are therefore in line with the
strict decentralization requirement of our insurance protocol.
In a CPMM setup, profitability for liquidity providers
is determined by two opposing effects. On the one hand,
the pool accumulates protocol fees. The gains are assigned
proportionally to all liquidity provision shares. The rate of
return depends on the pool’s trading volume relative to the
pool’s liquidity. On the other hand, liquidity providers are s.t.
divergence loss (also known as impermanent loss). Divergence
loss refers to the problem that liquidity providers lose value, if
the liquidity redemption price ratio differs from the liquidity
provision price ratio. Intuitively, this effect can be thought of
as negative arbitrage. Divergence loss is zero if the two pool
tokens maintain their initial price ratio and increases when the
relative price is shifting in one direction.
To assess the incentives for A- and B-tranche liquidity
providers we have to understand divergence loss in the context
of our tranche tokens. Let us assume a standard a _b = k setup,_
_·_
where a and b represent the initial amount of A and B tokens
in the pool and k is a constant product, that determines all
feasible combinations of a and b. Let us rearrange the equation
and take the partial derivative w.r.t. a. The absolute value of
the resulting slope can be reinterpreted as the relative price.
_pAB =_ _[k]_ (1)
_a[2]_
Trading activity may shift the token allocation to a[∗] and
_b[∗], with a[∗]_ _b[∗]_ = k. Using (1) we obtain the new price ratio
_·_
_p[∗]AB_ [. This allows us to express the post-trade quantities as a]
function of the new price ratio p[∗]AB [.]
_p[∗]_
_AB_ _AB_
_pAB_ _pAB_
_[−]_ _[p][∗]_ _[−]_ [1]
_p[∗]AB_
_pAB_ [+ 1]
�
2
_·_
��������
_D =_
��������
_._ (8)
We can now use this equation to analyze two distinct
outcomes and observe the effects on the pool and the liquidity
providers. First, assume the cover is not needed. The contract
enters Liquid state, and A- and B-tranches can be redeemed
for equal amounts of C. We refer to this case as the standard
_case. Second, assume one of the underlying yield-generating_
protocols suffers losses. These losses will be reflected in the
price of tranche B and therefore have an effect on the liquidity
pools that contain B. We refer to this as the benefit case.
_pA_
_pC_
_pB_
_S_
invest()
Interest
_t_
_T1_
divest()
�
_k_ �
, _b[∗]_ = _k · p[∗]AB_ (2)
_p[∗]AB_
_a[∗]_ =
We can now compute portfolio values Vp of a simple buy and
hold strategy (3) with the outcome of liquidity provision (4).
_VP (a, b) = p[∗]AB_ (3)
_[·][ a][ +][ b]_
_VP (a[∗], b[∗]) = p[∗]AB_ (4)
_[·][ a][∗]_ [+][ b][∗]
Using (2) to substitute quantities in (3) and (4) we get
�
_k_
�
+
_pAB_
_VP (a, b) = p[∗]AB_
_[·]_
�
_VP (a[∗], b[∗]) = 2 ·_
_k · p[∗]AB_ _[.]_ (6)
_k · pAB_ _,_ (5)
Divergence loss can be expressed as follows
_VP (a[∗], b[∗]) −_ _VP (a, b)_
_D :=_
���� _VP (a, b)_
(7)
����
Fig. 3. Relative price development of A and B shares between S and T1,
compared to the price of the underlying redeemable asset C.
_1) Standard Case: In the standard case A-tranches lose_
their cover value over time. Conversely, B-tranches become
less risky and will eventually be redeemable for an equal
amount of C as A-tranches. Hence, we know that p[∗]AB [= 1][.]
Making use of substitution in (8), the expected divergence loss
can be expressed as a function of the initial price ratio pAB .
The greater the initial risk premium, the higher the divergence
loss for liquidity provision in A/B-pools. Alternatively, a
liquidity provider could decide to contribute to an A/C- or
_B/C-pool. In T1, we know that pA = pB = pC_ (1 + i),
_·_
where i is the accumulated interest. Hence, we know that
_p[∗]AC_ [=][ p]BC[∗] [= 1 +][ i][. If we plug this value into (8), the]
expected divergence loss, for any expected interest rate, can be
expressed as a function of the initial price ratio pAB . Figure
3 shows the price relations of the three tokens. For A/Bpool liquidity provision considerations, interest rates can be
neglected. However, for A/C- and B/C-pools, interest plays
an important role. Note that B-tranche prices already have
a positive time trend. As such, interest will further increase
the price spread to C. Conversely, A-tranche prices have a
negative time trend and interest will therefore decrease the
spread. Consequently, any (positive) interest will create a
situation where the divergence loss of B/C-pools is greater
than the divergence loss of A/C-pools. This is shown in Figure
4.
While the extent of the divergence loss depends on various
factors, it is important to understand that the effect is relatively
small. Moreover, there are ways to mitigate a trend-based
From (7) we plug in (5) and (6). After rearranging we get
-----
Fig. 4. Divergence Loss (in line with equation (8)) for a/c-and b/c-Pools
with an expected interest of 5%. The two points marked in our graph represent
an example for an initial price spread between A and B. The initial valuation
of each a token starts at 1.02 c, and the valuation of each b token at 0.98 c.
divergence loss. Alternative pool models, such as the constant
_power sum invariant [42] can be used to design decentralized_
exchanges that are better suited for tokens with an inherent
price trend.
_2) Benefit Case: If any of the yield-generating protocols_
suffer a loss, A-tranche holders will be compensated at the
expense of B-tranche holders. In extreme scenarios, where
one of the yield-generating protocols loses its entire collateral, B-tranches become worthless. From (8) we know that
limp[∗]AB _[→∞]_ _[D][ =][ −][1][. Hence,][ A/B][- and][ B/C][-pool liquidity]_
providers are at risk of losing their entire stake. While this
constitutes an additional risk for providers of B-tranche liquidity, where they have to expose the B counterpart to an
additional risk and effectively stake twice the amount, they
receive trading fees in return. As such, the incentives depend
on the specifics and the risks of the insured protocols as well as
the relative trading volume. In extreme cases, where A/B and
_B/C liquidity provision would be prohibitively risky, liquidity_
providers could instead contribute to A/C-pools. Liquid A/Cpools would be sufficient, in the sense that anyone who is
interested in coverage could obtain it directly from the pool.
This scenario will be further discussed in Section V.
V. DISCUSSION
In the introduction we argued that current smart contractbased insurance protocols face various challenges and limitations. We will start our discussion by revisiting these points
and explain how our model addresses them.
First, the vast majority of existing insurance protocols
allows for over-insurance, where users can buy cover that
exceeds their exposure. This can create problematic incentives
and – depending on the jurisdiction – result in conflict with
the law. Our model does not allow for over-insurance. The
risk and capital are linked through our tranches and cannot be
separated without the use of another protocol.
Second, there are various challenges relating to claim assessment. All of the existing insurance protocols we have
examined have some form of dependency on external factors
during the claim assessment process. These dependencies can
be introduced through parametric triggers, oracles, community
voting or decisions by a predetermined expert council. All
of these approaches can lead to undesirable outcomes. The
incentives may not be aligned and create situations that can
result in deviations from the true outcome. In our model,
we do not rely on claim assessors, voting in a decentralized
autonomous organization (DAO), expert councils, oracles or
any trigger events. Instead, we use a deterministic distribution
schedule of a common underlying (Liquid Mode) and a
sequential choice model in accordance with the seniority of
the tranches (Fallback Mode). Consequently, payouts are not
conditional on any subjective decisions by an involved- or
third-party.
Third, we argued that many DeFi insurance protocols suffer
from capital inefficiencies and there certainly is a trade-off
between capital efficiency, security and special privileges. We
found that most existing protocols tend to be conservative
or cautious in their approach. The collateral is usually held
in low-risk, non-interest-bearing assets. As a result, these
protocols have at most 50% capital efficiency before leverage.
Some protocols are capable of increasing their efficiency by
covering multiple – ideally uncorrelated – risks with the
same collateral; however, they still require the collateral to
be in a low-risk, non-interest-bearing asset. In our model it
is possible to hold the collateral, i.e., the B-tranche, in a
interest-bearing asset without any significant drawbacks on
the security side, if the risks of the insured protocols are
indeed uncorrelated. Moreover, our approach is quite flexible
in the sense that further leverage, based on a larger number of
underlying protocols is feasible and could be implemented as
an extension.
In addition to these three initial points, there is another
advantage related to the risk premium that we came across in
the course of our research. As shown in Section IV, both our
cover and collateral (A- and B-tranches) are freely tradable.
The risk premium is simply determined by the relative price
between the two tranches. This allows us to create a marketbased price-finding mechanism for a fair risk premium. The
price can emerge naturally and does not depend on preset
parameters or statically implemented risk spreads that may
paralyze risk transfer activity.
In Section IV we show that there are greater incentives to
provide liquidity for the A-tranches than for the B-tranches.
Even in an extreme case, where the B liquidity would be very
low to non-existent, one could still obtain B-tranches. To do
so, they call the splitRisk() function to mint A- and Btranches in equal amounts and then sell the A-tranches, for
which the market can be assumed to be sufficiently liquid.
Anyone interested in the insurance cover could simply buy Atranches on the open market and would not have to interact
-----
with the protocol. Assuming a constant supply, greater demand
for A-tranches would increase the risk spread and therefore
incentivize the creation of additional A- (and B-) tranches.
There are many benefits to our proposal and we believe that
this paper significantly contributes towards the DeFi protocol
stack. However, every proposal also has its limitations and
drawbacks. In the remainder of this section, we discuss some
of these limitations and propose potential extensions and new
research avenues to mitigate these issues.
First, our model requires a common underlying among
all involved protocols. The reason for this is to eliminate
any reliance on external price sources, i.e., oracles. In liquid
mode, we redeem everything to denominations of a unified
underlying at the end of a predefined time period. While
it is theoretically possible to wrap tokens to give them an
arbitrary underlying, this will have one of two consequences:
either a dependency on external price sources has to be
introduced, or the fallback case in our model would introduce
an insurance against relative price movements of the assets and
the underlying. The latter may be desirable in some cases, but
it is not the default behaviour we want to achieve.
Second, our protocol has a fixed time span. Consequently,
insuring assets over a longer period of time requires regular
actions from all involved parties. A new contract has to be
deployed for each period and the assets need to be moved
over. This problem is exacerbated by shorter insurance periods.
Longer insurance periods on the other hand increase the time
that claimants have to wait for their compensation in case of
an incident and also increase the risk of both protocols failing
during the same period. We believe this limitation could be
mitigated with an extension to the protocol, which uses short
insurance periods and rolls over any non-redeemed tranches to
a new insurance period. However, an extension of this nature
could significantly increase the complexity of the protocol and
would require further research to determine the practicality and
potential consequences.
Third, in our model we specify minting and redeeming time
windows for the tranches. Consequently, the total supply of
_A- and B-tranches cannot change during the main insurance_
period. This can be an issue, especially if there is insufficient
liquidity for the B-tranches, as discussed in Section IV or if
the demand for cover changes significantly. Further research
into this topic is necessary, but we believe that under certain
circumstances, the minting window could be extended to allow
the creation of new shares during the active insurance phase.
One requirement for this would be a way to track the accrued
interest on the insurance protocol and to increase the costs of
the newly created tranches accordingly. Similar considerations
can be made for the redeeming window. Early redemption of
equal parts of A- and B-tranches should be possible without
large changes to the model. Even early redemption of just
_A-tranches is theoretically possible._
Finally, our model and the reference implementation use
two protocols. This is not a strict limitation. In fact, it can
be shown that the model works as described as long as the
number of tranches is equal to the number of insured protocols.
For example, an extension to three protocols is possible with
the introduction of a third tranche, without any fundamental
changes to the protocol.
A more challenging extension is the addition of further
protocols without any changes to the number of tranches. This
extension would severely increase the complexity of fallback
mode. Recall that A-tranche holders get to choose which of
the remaining interest-bearing tokens they want to redeem. In
a world where the number of tranches is equal to the number
of protocols, this is unproblematic, since there will always
be sufficient collateral of any type for A-tranche holders to
choose from. In a model where the number of protocols is
greater than the number of tranches, A-tranche holders might
compete with each other and race to redeem the more valuable
collateral. As such, models where the number of protocols is
greater than the number of tranches can create a first mover
advantage, where A-tranche holders are treated inconsistently.
A potential solution to solve this issue is a two-step approach,
that lets tranche holders choose and commit their redemption
preferences before the final redemption ratios are calculated.
VI. CONCLUSION
In this paper, we propose a fully decentralized DeFi insurance model that does not rely on any external information
sources, such as price feeds (oracles) or claim assessors. The
general idea of our insurance protocol is to pool assets from
two third-party protocols, and allow users to split the pool
redemption rights into two freely tradable tranche tokens: A
and B. Any losses are first absorbed by the B-tranche holders.
_A-tranche holders will only be negatively affected if 50% or_
more of the pooled funds are irrecoverable, or if both protocols
become temporarily illiquid and face (partial) losses.
The market for A- and B-tranches determines the fair risk
premium for the insurance.
Our approach has several advantages over other DeFi insurance solutions. In addition to being fully decentralized and
trustless, it also prevents over-insurance, does not rely on any
parametric triggers, and is highly capital-efficient.
We provide a complete reference implementation of the
insurance protocol in Solidity, with coverage for two popular
lending market protocols.
We believe that fully decentralized and trustless infrastructure is crucial and may create more transparent, open and
resilient financial markets. Our contribution should be seen
as a composable building block and a foundation for further
research and development efforts.
ACKNOWLEDGMENT
The authors would like to thank Tobias Bitterli, Mitchell
Goldberg, Emma Littlejohn, Katrin Schuler and Dario
Thürkauf.
APPENDIX
The full Solidity source code for our reference implementa[tion can be found in our github repository: https://github.com/](https://github.com/cifunibas/decentralized-insurance)
[cifunibas/decentralized-insurance](https://github.com/cifunibas/decentralized-insurance)
-----
REFERENCES
[1] F. Schaer, “Decentralized finance: On blockchain- and smart contractbased financial markets,” Fed. Reserve Bank St. Louis Rev., vol. Second
Quarter 2021, pp. pp. 153–74, 2021.
[2] H. Karp and R. Melbardis, “Nexus mutual whitepaper: A peer-to-peer
discretionary mutual on the ethereum blockchain,” 2017. [Online].
[Available: https://nexusmutual.io/assets/docs/nmx_white_paperv2_3.pdf](https://nexusmutual.io/assets/docs/nmx_white_paperv2_3.pdf)
[3] Nsure.Network, “Nsure.network - open insurance platform for open
[finance,” 2020. [Online]. Available: https://nsure.network/Nsure_WP_0.](https://nsure.network/Nsure_WP_0.7.pdf)
[7.pdf](https://nsure.network/Nsure_WP_0.7.pdf)
[4] Cozy.Finance, “Cozy finance developer docs,” 2020. [Online]. Available:
[https://docs.cozy.finance/](https://docs.cozy.finance/)
[5] Unslashed.Finance, “Insurance for decentralized finance,” 2021.
[[Online]. Available: https://documentation.unslashed.finance/](https://documentation.unslashed.finance/)
[6] M. Resnick, R. Ben-Har, D. Patel, and A. Bipin,
“Risk harbor v2,” 01 2022. [Online]. Available: [https://github.com/Risk-Harbor/RiskHarbor-Whitepaper/blob/](https://github.com/Risk-Harbor/RiskHarbor-Whitepaper/blob/main/Risk%20Harbor%20Core%20V2%20Whitepaper.pdf)
[main/Risk%20Harbor%20Core%20V2%20Whitepaper.pdf](https://github.com/Risk-Harbor/RiskHarbor-Whitepaper/blob/main/Risk%20Harbor%20Core%20V2%20Whitepaper.pdf)
[7] J. R. Douceur, “The sybil attack,” in International workshop on peer_to-peer systems._ Springer, 2002, pp. 251–260.
[8] N. Atzei, M. Bartoletti, and T. Cimoli, “A Survey of Attacks on
Ethereum Smart Contracts (SoK),” in Principles of security and trust,
ser. Lecture notes in computer science, 0302-9743, M. Maffei and
M. Ryan, Eds., vol. 10204. Springer, Berlin, Heidelberg, 2017, pp.
164–186.
[9] L. Gudgeon, D. Perez, D. Harz, B. Livshits, and A. Gervais,
“The decentralized financial crisis,” _2020_ _Crypto_ _Valley_ _Conf._
_Blockchain_ _Technol._ _(CVCBT),_ 2020. [Online]. Available: [https:](https://arxiv.org/pdf/2002.08099)
[//arxiv.org/pdf/2002.08099](https://arxiv.org/pdf/2002.08099)
[10] D. Macrinici, C. Cartofeanu, and S. Gao, “Smart contract applications
within blockchain technology: A systematic mapping study,” Telemat.
_Inform., vol. 35, no. 8, pp. 2337–2354, 2018. [Online]. Available:_
[http://www.sciencedirect.com/science/article/pii/S0736585318308013](http://www.sciencedirect.com/science/article/pii/S0736585318308013)
[11] Z. Zheng, S. Xie, H.-N. Dai, W. Chen, X. Chen, J. Weng, and
M. Imran, “An overview on smart contracts: Challenges, advances
and platforms,” Future Gener. Comput. Syst., vol. 105, pp. 475–491,
[2020. [Online]. Available: http://www.sciencedirect.com/science/article/](http://www.sciencedirect.com/science/article/pii/S0167739X19316280)
[pii/S0167739X19316280](http://www.sciencedirect.com/science/article/pii/S0167739X19316280)
[12] L. Zhou, X. Xiong, J. Ernstberger, S. Chaliasos, Z. Wang,
Y. Wang, K. Qin, R. Wattenhofer, D. Song, and A. Gervais, “SoK:
Decentralized Finance (DeFi) Attacks,” 8/27/2022. [Online]. Available:
[https://arxiv.org/pdf/2208.13035](https://arxiv.org/pdf/2208.13035)
[13] M. Bartoletti, J. H.-y. Chiang, and A. L. Lafuente, “Sok: Lending
pools in decentralized finance,” in _Financial_ _Cryptography_ _and_
_Data Security. FC 2021 Int. Workshops, vol. 12676._ Springer,
[Berlin, Heidelberg, 2021, pp. 553–578. [Online]. Available: https:](https://link.springer.com/chapter/10.1007/978-3-662-63958-0_40)
[//link.springer.com/chapter/10.1007/978-3-662-63958-0_40](https://link.springer.com/chapter/10.1007/978-3-662-63958-0_40)
[14] A. Lehar and C. A. Parlour, “Systemic fragility in decentralized
[markets,” 2022. [Online]. Available: https://papers.ssrn.com/sol3/papers.](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4164833)
[cfm?abstract_id=4164833](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4164833)
[15] K. Qin, L. Zhou, B. Livshits, and A. Gervais, “Attacking the
DeFi Ecosystem with Flash Loans for Fun and Profit,” in Financial
_Cryptography_ _and_ _Data_ _Security._ _FC_ _2021._ _Lecture_ _Notes_ _in_
_Computer Science, N. Borisov and C. Diaz, Eds., vol. 12674._
Springer, Berlin, Heidelberg, 2021, pp. 3–32. [Online]. Available:
[https://link.springer.com/chapter/10.1007/978-3-662-64322-8_1](https://link.springer.com/chapter/10.1007/978-3-662-64322-8_1)
[16] G. Angeris and T. Chitra, “Improved price oracles: Constant function
market makers,” in Proc. 2nd ACM Conf. Advances in Financial
_Technologies, ser. AFT ’20, no. 12._ New York, NY, USA: Association
for Computing Machinery, 2020, pp. 80–91. [Online]. Available:
[https://doi.org/10.1145/3419614.3423251](https://doi.org/10.1145/3419614.3423251)
[17] B. Liu, P. Szalachowski, and J. Zhou, “A first look into defi
oracles,” in _2021_ _IEEE_ _Int._ _Conf._ _Decentralized_ _Applications_
_and_ _Infrastructures_ _(DAPPS)._ Los Alamitos, CA, USA: IEEE
Computer Society, aug 2021, pp. 39–48. [Online]. Available:
[https://doi.ieeecomputersociety.org/10.1109/DAPPS52256.2021.00010](https://doi.ieeecomputersociety.org/10.1109/DAPPS52256.2021.00010)
[18] P. Albrecht and M. Huggenberger, “The fundamental theorem of
mutual insurance,” Insur.: Math. Econ., vol. 75, pp. 180–188, 2017.
[[Online]. Available: https://www.sciencedirect.com/science/article/pii/](https://www.sciencedirect.com/science/article/pii/S0167668716304978)
[S0167668716304978](https://www.sciencedirect.com/science/article/pii/S0167668716304978)
[19] C. Laux and A. Muermann, “Financing risk transfer under governance
problems: Mutual versus stock insurers,” J. Financ. Intermediation,
vol. 19, no. 3, pp. 333–354, 2010. [Online]. Available: [http:](http://www.sciencedirect.com/science/article/pii/S1042957309000266)
[//www.sciencedirect.com/science/article/pii/S1042957309000266](http://www.sciencedirect.com/science/article/pii/S1042957309000266)
[20] J. A. Ligon and P. D. Thistle, “The formation of mutual insurers
in markets with adverse selection,” J. Bus., vol. 78, no. 2, pp.
[529–556, 2005. [Online]. Available: https://ideas.repec.org/a/ucp/jnlbus/](https://ideas.repec.org/a/ucp/jnlbus/v78y2005i2p529-556.html)
[v78y2005i2p529-556.html](https://ideas.repec.org/a/ucp/jnlbus/v78y2005i2p529-556.html)
[21] C. S. Tapiero, Y. Kahane, and L. Jacque, “Insurance premiums and
default risk in mutual insurance,” Scand. Actuar. J., vol. 1986, no. 2,
pp. 82–97, 1986.
[22] M. Denuit, J. Dhaene, and C. Y. Robert, “Risk–sharing rules and
their properties, with applications to peer–to–peer insurance,” J.
_Risk Insur., vol. 89, no. 3, pp. 615–667, 2022. [Online]. Available:_
[https://onlinelibrary.wiley.com/doi/10.1111/jori.12385?af=R](https://onlinelibrary.wiley.com/doi/10.1111/jori.12385?af=R)
[23] M. Denuit and C. Y. Robert, “Large-loss behavior of conditional
mean risk sharing,” _ASTIN_ _Bulletin,_ vol. 50, no. 3, pp.
1093–1122, 2020. [Online]. Available: [https://www.cambridge.](https://www.cambridge.org/core/article/largeloss-behavior-of-conditional-mean-risk-sharing/B6AE93167BC6BD47C35BB050C292B05F)
[org/core/article/largeloss-behavior-of-conditional-mean-risk-sharing/](https://www.cambridge.org/core/article/largeloss-behavior-of-conditional-mean-risk-sharing/B6AE93167BC6BD47C35BB050C292B05F)
[B6AE93167BC6BD47C35BB050C292B05F](https://www.cambridge.org/core/article/largeloss-behavior-of-conditional-mean-risk-sharing/B6AE93167BC6BD47C35BB050C292B05F)
[24] R. Feng, M. Liu, and N. Zhang, “A unified theory of decentralized
insurance,” 2022. [Online]. Available: [https://papers.ssrn.com/sol3/](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4013729)
[papers.cfm?abstract_id=4013729](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4013729)
[25] R. Feng, C. Liu, and S. Taylor, “Peer-to-peer risk sharing with an
application to flood risk pooling,” Ann. Oper. Res., 2022. [Online].
[Available: https://doi.org/10.1007/s10479-022-04841-x](https://doi.org/10.1007/s10479-022-04841-x)
[26] M. Denuit, “Size-biased transform and conditional mean risk sharing,
with application to p2p insurance and tontines,” ASTIN Bulletin, vol. 49,
no. 3, pp. 591–617, 2019.
[27] ——, “Investing in your own and peers’risks: the simple analytics of
p2p insurance,” Eur. Actuar. J., vol. 10, no. 2, pp. 335–359, 2020.
[[Online]. Available: https://doi.org/10.1007/s13385-020-00238-x](https://doi.org/10.1007/s13385-020-00238-x)
[28] A. Charpentier, L. Kouakou, M. Loewe, P. Ratz, and F. Vermet,
“Collaborative insurance sustainability and network structure,” 2021.
[[Online]. Available: https://arxiv.org/pdf/2107.02764](https://arxiv.org/pdf/2107.02764)
[29] G. P. Clemente and P. Marano, “The broker model for peer-to-peer
insurance: an analysis of its value,” Geneva Pap. Risk Insur.: Issues
_Pract., vol. 45, no. 3, pp. 457–481, 2020. [Online]. Available:_
[https://link.springer.com/article/10.1057/s41288-020-00165-8](https://link.springer.com/article/10.1057/s41288-020-00165-8)
[30] S. Levantesi and G. Piscopo, “Mutual peer-to-peer insurance: The
allocation of risk,” J. Co-op. Organ. Manag., vol. 10, no. 1,
[p. 100154, 2022. [Online]. Available: https://www.sciencedirect.com/](https://www.sciencedirect.com/science/article/pii/S2213297X21000264)
[science/article/pii/S2213297X21000264](https://www.sciencedirect.com/science/article/pii/S2213297X21000264)
[31] D. Duffie and N. Gârleanu, “Risk and valuation of collateralized debt
obligations,” Financ. Anal. J., vol. 57, no. 1, pp. 41–59, 01 2001.
[[Online]. Available: https://doi.org/10.2469/faj.v57.n1.2418](https://doi.org/10.2469/faj.v57.n1.2418)
[32] J. Armstrong and J. Kiff, “Understanding the benefits and risks of
synthetic collateralized debt obligations,” Bank of Canada Financial
_System Review, pp. 53–61, 2005._
[33] D. J. Lucas, L. S. Goodman, and F. J. Fabozzi, Collateralized debt obli_gations: structures and analysis, 2nd ed., ser. Wiley finance._ Hoboken,
N.J.: John Wiley & Sons, 04 2006, vol. 140.
[34] C. Bluhm and C. Wagner, “Valuation and risk management of
collateralized debt obligations and related securities,” Annu. Rev.
_Financ. Econ., vol. 3, no. 1, pp. 193–222, 2022/12/15 2011. [Online]._
[Available: https://doi.org/10.1146/annurev-financial-102710-144835](https://doi.org/10.1146/annurev-financial-102710-144835)
[35] J. P. Forrester, “Insurance risk collateralized debt obligations,” J.
_Struct. Finance, vol. 14, no. 1, p. 28, 04 2008. [Online]. Available:_
[http://jsf.pm-research.com/content/14/1/28.abstract](http://jsf.pm-research.com/content/14/1/28.abstract)
[36] H. N. Bystroem, “The microfinance collateralized debt obligation: A
modern robin hood?” World Dev., vol. 36, no. 11, pp. 2109–2126, 2008.
[37] MakerDAO, “The dai stablecoin system whitepaper,” 12 2017. [Online].
[Available: https://makerdao.com/whitepaper/DaiDec17WP.pdf](https://makerdao.com/whitepaper/DaiDec17WP.pdf)
[38] Aave, “Aave protocol whitepaper v2.0,” 12 2020. [Online]. Available:
[https://github.com/aave/protocol-v2/blob/master/aave-v2-whitepaper.pdf](https://github.com/aave/protocol-v2/blob/master/aave-v2-whitepaper.pdf)
[39] R. Leshner and G. Hayes, “Compound: The money market protocol,”
02 2019. [Online]. Available: [https://compound.finance/documents/](https://compound.finance/documents/Compound.Whitepaper.pdf)
[Compound.Whitepaper.pdf](https://compound.finance/documents/Compound.Whitepaper.pdf)
[40] S. Cousaert, J. Xu, and T. Matsui, “Sok: Yield aggregators in defi,” in
_2022 IEEE Int. Conf. Blockchain and Cryptocurrency (ICBC), 2022, pp._
[1–14. [Online]. Available: https://ieeexplore.ieee.org/document/9805523](https://ieeexplore.ieee.org/document/9805523)
[41] V. Mohan, “Automated market makers and decentralized exchanges:
a DeFi primer,” _Financial_ _Innov.,_ vol. 8, no. 1, pp. 1–48,
2022. [Online]. Available: [https://link.springer.com/article/10.1186/](https://link.springer.com/article/10.1186/s40854-021-00314-5)
[s40854-021-00314-5](https://link.springer.com/article/10.1186/s40854-021-00314-5)
-----
[42] A. Niemerg, D. Robinson, and L. Livnev, “Yieldspace: An automated
liquidity provider for fixed yield tokens,” 2020. [Online]. Available:
[https://yield.is/YieldSpace.pdf](https://yield.is/YieldSpace.pdf)
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/2212.10308, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "GREEN",
"url": "https://arxiv.org/pdf/2212.10308"
}
| 2,022
|
[
"JournalArticle",
"Conference"
] | true
| 2022-12-20T00:00:00
|
[
{
"paperId": "d52befb1085c0539dfe833523dc88eb71e9f2613",
"title": "SoK: Decentralized Finance (DeFi) Attacks"
},
{
"paperId": "1ab554c549e8a81635c95a849651d319ba3df696",
"title": "Risk‐sharing rules and their properties, with applications to peer‐to‐peer insurance"
},
{
"paperId": "1a9997d381a041fecddd92bcd7a3caed6130d2b3",
"title": "Collaborative Insurance Sustainability and Network Structure"
},
{
"paperId": "15a789d417b0a1c9e8c5ed42eded026a85d51134",
"title": "SoK: Yield Aggregators in DeFi"
},
{
"paperId": "d102780639e4d91e6549b23f621da00e961447bf",
"title": "SoK: Lending Pools in Decentralized Finance"
},
{
"paperId": "ba90fb4781a57a14e45559b6607c106e2aea90a6",
"title": "Peer-to-peer risk sharing with an application to flood risk pooling"
},
{
"paperId": "1669088e855d088bd5021a767492aef35e1ad40d",
"title": "Automated market makers and decentralized exchanges: a DeFi primer"
},
{
"paperId": "b14b58dbce132a6d17c4dc346b68cf0673c00101",
"title": "LARGE-LOSS BEHAVIOR OF CONDITIONAL MEAN RISK SHARING"
},
{
"paperId": "ff7482eaaa4cef5d6076a98c674019324654a194",
"title": "The broker model for peer-to-peer insurance: an analysis of its value"
},
{
"paperId": "2ce49584a1fd0f34f37cf617d2027f1ca593ae1e",
"title": "Investing in your own and peers’ risks: the simple analytics of P2P insurance"
},
{
"paperId": "cd19976ac5207bf02a73b332fc4aa3946e6812b5",
"title": "A First Look into DeFi Oracles"
},
{
"paperId": "2c40956941634b92c16ccf3ba305abeeab8f8e55",
"title": "Improved Price Oracles: Constant Function Market Makers"
},
{
"paperId": "47072c24806046a9c4827467d7047af8c6a07b62",
"title": "Attacking the DeFi Ecosystem with Flash Loans for Fun and Profit"
},
{
"paperId": "082f7f6e1fda6358d47df5d26fe862ef6021a803",
"title": "Decentralized Finance: On Blockchain- and Smart Contract-Based Financial Markets"
},
{
"paperId": "c876d76a22b23ae28fe6d9c5589f857768d543fc",
"title": "The Decentralized Financial Crisis"
},
{
"paperId": "e0bc89f5776804bc2be27f1945f900d1ac8f1e7f",
"title": "An Overview on Smart Contracts: Challenges, Advances and Platforms"
},
{
"paperId": "cf38c389e415412bd8aa1af0f3e8abbc7f97b47c",
"title": "SIZE-BIASED TRANSFORM AND CONDITIONAL MEAN RISK SHARING, WITH APPLICATION TO P2P INSURANCE AND TONTINES"
},
{
"paperId": "f5a21fb87d88b4510dd4d42fbdc52a674a592ea6",
"title": "Smart contract applications within blockchain technology: A systematic mapping study"
},
{
"paperId": "4f23887df244122da1d087c65121a85b57a54b13",
"title": "The fundamental theorem of mutual insurance"
},
{
"paperId": "aec843c0f38aff6c7901391a75ec10114a3d60f8",
"title": "A Survey of Attacks on Ethereum Smart Contracts (SoK)"
},
{
"paperId": "790959fa05601e03d5f22646c40feb30a7917414",
"title": "Valuation and Risk Management of Collateralized Debt Obligations and Related Securities"
},
{
"paperId": "a905a4fdb5fa3da754eda0b7c2e7ff3da264ca98",
"title": "Financing Risk Transfer under Governance Problems: Mutual versus Stock Insurers"
},
{
"paperId": "f13cd1e7c2958841de80e60abe7c23589a0b0292",
"title": "The Microfinance Collateralized Debt Obligation: A Modern Robin Hood?"
},
{
"paperId": "28c587bbe661e4e8ea054d106d59efd7fa6d8f2c",
"title": "Insurance Risk Collateralized Debt Obligations"
},
{
"paperId": "b717d4e45d99795f1feb912dcd6d78c15d46c539",
"title": "The Formation of Mutual Insurers in Markets with Adverse Selection"
},
{
"paperId": "e2f09c5978b7413b9a92de6e2050b365814da235",
"title": "Collateralized Debt Obligations: Structures and Analysis"
},
{
"paperId": "35516916cd8840566acc05d0226f711bee1b563b",
"title": "The Sybil Attack"
},
{
"paperId": "d8a92d6adeff34a1acca31558390434cc7e2306b",
"title": "Risk and Valuation of Collateralized Debt Obligations"
},
{
"paperId": "7f2746713421f4178d66a7c65277632d0b494a88",
"title": "Insurance premiums and default risk in mutual insurance"
},
{
"paperId": "469f794c6094795371e4cd318f8a6d22766fc09f",
"title": "Systemic Fragility in Decentralized Markets"
},
{
"paperId": "172bc103350c37288d55362f10b23cdf845c3389",
"title": "A Unified Theory of Decentralized Insurance"
},
{
"paperId": "430c6c797ec9c05bd5f7edfe4158ab1590b52b17",
"title": "Mutual peer-to-peer insurance: The allocation of risk"
},
{
"paperId": null,
"title": "Risk harbor v2"
},
{
"paperId": null,
"title": "Unslashed.Finance, “Insurance for decentralized finance,”"
},
{
"paperId": "7298a315da0f62f95012f0f4ed748c2a455ec5d7",
"title": "YieldSpace: An Automated Liquidity Provider for Fixed Yield Tokens"
},
{
"paperId": null,
"title": "Cozy finance developer docs"
},
{
"paperId": null,
"title": "Aave protocol whitepaper v2.0"
},
{
"paperId": null,
"title": "Nsure.network - open insurance platform for open finance"
},
{
"paperId": null,
"title": "Nexus mutual whitepaper: A peer-to-peer discretionary mutual on the ethereum blockchain"
},
{
"paperId": "1fc2453a60f652893fb22517d5a44541746df67b",
"title": "Understanding the Benefits and Risks of Synthetic Collateralized Debt Obligations"
},
{
"paperId": null,
"title": "Risk harbor v 2 , ” 01 2022 . [ Online ] The sybil attack , ” in International workshop on peer - to - peer systems"
},
{
"paperId": null,
"title": "Compound : The money market protocol , ” 02 2019 . [ Online ]"
},
{
"paperId": null,
"title": "The dai stablecoin system whitepaper , ” 12 2017 . [ Online ] Aave , “ Aave protocol whitepaper v 2 . 0 , ” 12 2020 . [ Online ]"
}
] | 14,996
|
en
|
[
{
"category": "Medicine",
"source": "external"
},
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Medicine",
"source": "s2-fos-model"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/0272fb91bc0ca70d246268095be9d6138293babc
|
[
"Medicine",
"Computer Science"
] | 0.88378
|
MedCo: Enabling Secure and Privacy-Preserving Exploration of Distributed Clinical and Genomic Data
|
0272fb91bc0ca70d246268095be9d6138293babc
|
IEEE/ACM Transactions on Computational Biology & Bioinformatics
|
[
{
"authorId": "2201389",
"name": "J. Raisaro"
},
{
"authorId": "1398987134",
"name": "J. Troncoso-Pastoriza"
},
{
"authorId": "122581714",
"name": "Mickaël Misbach"
},
{
"authorId": "40381688",
"name": "João Sá Sousa"
},
{
"authorId": "5332114",
"name": "S. Pradervand"
},
{
"authorId": "6961679",
"name": "E. Missiaglia"
},
{
"authorId": "1702906",
"name": "O. Michielin"
},
{
"authorId": "144067653",
"name": "B. Ford"
},
{
"authorId": "1757221",
"name": "J. Hubaux"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"IEEE/ACM Trans Comput Biology Bioinform",
"IEEE/ACM Trans Comput Biology Bioinform",
"IEEE/ACM Transactions on Computational Biology and Bioinformatics"
],
"alternate_urls": [
"http://www.computer.org/portal/web/tcbb/home"
],
"id": "dc4a9aad-72db-4530-a183-eaa4bf1d4490",
"issn": "1545-5963",
"name": "IEEE/ACM Transactions on Computational Biology & Bioinformatics",
"type": "journal",
"url": "https://ieeexplore.ieee.org/servlet/opac?punumber=8857"
}
|
The increasing number of health-data breaches is creating a complicated environment for medical-data sharing and, consequently, for medical progress. Therefore, the development of new solutions that can reassure clinical sites by enabling privacy-preserving sharing of sensitive medical data in compliance with stringent regulations (e.g., HIPAA, GDPR) is now more urgent than ever. In this work, we introduce MedCo, the first operational system that enables a group of clinical sites to federate and collectively protect their data in order to share them with external investigators without worrying about security and privacy concerns. MedCo uses (a) collective homomorphic encryption to provide trust decentralization and end-to-end confidentiality protection, and (b) obfuscation techniques to achieve formal notions of privacy, such as differential privacy. A critical feature of MedCo is that it is fully integrated within the i2b2 (Informatics for Integrating Biology and the Bedside) framework, currently used in more than 300 hospitals worldwide. Therefore, it is easily adoptable by clinical sites. We demonstrate MedCo's practicality by testing it on data from The Cancer Genome Atlas in a simulated network of three institutions. Its performance is comparable to the ones of SHRINE (networked i2b2), which, in contrast, does not provide any data protection guarantee.
|
## MEDCO: Enabling Secure and Privacy-Preserving Exploration of Distributed Clinical and Genomic Data
#### Jean Louis Raisaro, Juan Ramón Troncoso-Pastoriza, Mickaël Misbach, João Sá Sousa, Sylvain Pradervand, Edoardo Missiaglia, Olivier Michielin, Bryan Ford and Jean-Pierre Hubaux
**Abstract—The increasing number of health-data breaches is creating a complicated environment for medical-data sharing and,**
consequently, for medical progress. Therefore, the development of new solutions that can reassure clinical sites by enabling
privacy-preserving sharing of sensitive medical data in compliance with stringent regulations (e.g., HIPAA, GDPR) is now more urgent
than ever. In this work, we introduce MedCo, the first operational system that enables a group of clinical sites to federate and
collectively protect their data in order to share them with external investigators without worrying about security and privacy concerns.
MedCo uses (a) collective homomorphic encryption to provide trust decentralization and end-to-end confidentiality protection, and (b)
obfuscation techniques to achieve formal notions of privacy, such as differential privacy. A critical feature of MedCo is that it is fully
integrated within the i2b2 (Informatics for Integrating Biology and the Bedside) framework, currently used in more than 300 hospitals
worldwide. Therefore, it is easily adoptable by clinical sites. We demonstrate MedCo’s practicality by testing it on data from The Cancer
Genome Atlas in a simulated network of three institutions. Its performance is comparable to the ones of SHRINE (networked i2b2),
which, in contrast, does not provide any data protection guarantee.
**Index Terms—Secure data-sharing, homomorphic encryption, differential privacy, i2b2, distributed data, decentralized trust, genomic**
privacy.
#### �
#### 1 INTRODUCTION
ITH the increasing digitalization of clinical and genomic information, data sharing is becoming the
# W
keystone for realizing the promise of personalized medicine.
Several initiatives, such as the Patient-Centered Clinical
Research Network (PCORNet) [1] in the USA, eTRIKS/TranSMART [2] in the EU, the Swiss Personalized Health
Network (SPHN) [3] in Switzerland, and the Global Alliance
for Genomics and Health (GA4GH) [4], are laying down
the foundations for new biomedical research infrastructures
aimed at interconnecting (so far) siloed repositories of clinical and genomic data. In this global ecosystem, the ability to
provide strong privacy and security guarantees in order to
comply with increasingly strict regulations (e.g., HIPAA [5]
in USA or the new GDPR [6] in EU) is crucial, yet extremely
challenging, to achieve.
Currently, there exist two main approaches for sharing
medical data. The first is the centralized approach (see
Figure 1(A)) typical of initiatives such as All of Us [7]
and Genomics England [8]. With this approach, data from
multiple institutions are brought together in a single and
centralized repository that can be accessed by researchers
- J.L. Raisaro, J.R. Troncoso-Pastoriza, M. Misbach, J. Sá Sousa, B. Ford
and J.-P. Hubaux are with the School of Computer and Communication
Sciences, EPFL, Lausanne, Switzerland.
E-mail: see https://people.epfl.ch/jean-pierre.hubaux
- S. Pradervand, E. Missiaglia and O. Michielin are with the Lausanne
University Hospital, CHUV, Lausanne, Switzerland.
- S. Pradervand is with Genomic Technologies Facility, University of Lausanne, UNIL, Lausanne, Switzerland, and with Vital-IT, Swiss Institute
of Bioinformatics, Lausanne, Switzerland.
willing to run analysis on a unified dataset. The second
is the decentralized approach (see Figure 1(B)), where the
different institutions keep the data at their premises and
form an interoperable peer-to-peer network accessible by
researchers. PCORNet [1] and the Beacon Project of the
GA4GH [9] are examples of this second approach. Unfortunately, both approaches to sharing medical data have
revealed intrinsic limitations that demonstrate why neither
of the two has already been fully adopted by the healthcare
sector.
On the one hand, the centralized approach provides undeniable advantages in terms of availability and flexibility,
although it introduces a single point of failure in the system
by accumulating all the trust on a single entity (i.e., the
data repository). Indeed, the security and confidentiality
of all the data rely on the ability of the central repository
to thwart both external (hackers) and internal (insiders)
attacks. Furthermore, as the number of health-data breaches
constantly increases [10], there is significant public pressure
on clinical sites to ensure that the privacy and security
of patients’ data can be properly protected, notably when
stored or processed by third parties. As a result, clinical
sites are worried about adopting the centralized approach
and outsourcing their data to a single central repository
(e.g., the cloud), especially when the data to be shared
is highly sensitive or identifying (e.g., genomic data). On
the other hand, the fully decentralized approach solves the
single-point-of-failure issue: clinical sites can individually
enforce local control on their own data by monitoring and
managing the different accesses. However, this decentralization imposes substantial costs on the clinical sites as
-----
they have to maintain an interoperable network, often with
very limited resources (both human and technical). For this
reason, the fully decentralized approach is also likely to
be unsustainable in the long run, especially for large scale
projects where multiple clinical sites are involved.
**�����������������������** **�������������** **�������������** **��������������**
**�������������** **�������������** **�������������** **�������������**
**������������������������** **��������������������������**
**����������������**
**�������������**
**�������������**
**�������������������**
**���������** **������������**
**���** **�������������**
**�����������������������������������������������**
Fig. 1: Comparison of approaches for sharing medical data.
(A) Centralized approach affected by the single-point-of-failure
problem. (B) Decentralized approach affected by high maintenance costs (both technical and human). (C) Hybrid and secure
approach enabled by MedCo, where clinical sites can securely
outsource their data to the storage and processing unit (SPU) of
their choice.
In this paper, to address the challenge of achieving
privacy-preserving, secure and scalable data sharing we
introduce MedCo. MedCo is the first operational system
that enables hundreds of clinical sites to share their clinical and genomic data through a hybrid or “somewhat”
decentralized approach that overcomes the limitations of
the approaches described above (see Figure 1). Instead of
concentrating the trust on a single central repository as
in the centralized approach, MedCo distributes the trust
among a set of different “storage and processing” units
to which clinical sites can securely outsource the storage
of their data. Together, the storage and processing units
form a secure, federated and interoperable network that
investigators can query for research purposes as if it were
a single unified database. MedCo enables each clinical site
to choose its preferred storage and processing unit in order
to offload the maintenance and availability costs that affect
the fully decentralized approach. Such a storage and processing unit can be hosted either by the clinical site itself,
by a governmental institution, or by a private/public cloud
provider with whom the clinical site establishes a datause agreement. For example, a clinical site with enough
resources can have its own storage and processing unit
hosted at its premises. Whereas, a clinical site with limited resources could use a cloud provider of its choice.
Potentially, each country could have a national storage and
processing unit, e.g., administered by the government or
a not-profit organization, to which all clinical sites within
the same country can outsource their data The different
national storage and processing units could then federate
to form an international, secure and distributed clinical
research network.
A critical advantage of MedCo, with respect to stateof-the-art systems for sharing medical data, is its ability to
provide strong security guarantees to clinical sites willing
to safely outsource the storage of their data to potentially
untrusted storage and processing units. Indeed, MedCo enables each site to encrypt its data with a shared key that
is collectively generated by all the storage and processing
units in the federation. As the encryption scheme used
by MedCo is additively homomorphic, investigators can
directly query and process the encrypted data stored at
different storage and processing units without the need
for decrypting them. This ensures end-to-end protection of
the data in the Anytrust adversary model. Only authorized
investigators can decrypt the result of a query/analysis and
none of the storage and processing units alone, even if
compromised, can decrypt the data stored at its premises.
Actually, in order to succeed and get access to the unencrypted data, an adversary would need to simultaneously
compromise all the storage and processing units in the
federation. Additionally, MedCo can also be configured to
minimize the risk of re-identification stemming from the
behavior of malicious or curious investigators that try to
abuse the querying system; this is achieved by providing
obfuscated results that provide formal and well-established
notions of privacy, e.g., differential privacy.
In order to ease its adoption in operational research
environments, we developed MedCo on top of existing and
well-established open-source technologies for clinical data
exploration, namely i2b2 [11] and SHRINE [12]. Currently,
i2b2 is used at more than 300 clinical sites worldwide. We
demonstrate the practicality of MedCo by testing it in a simulated federation of three clinical sites that outsource their
oncology data (both clinical and genomic) to three different
storage and processing units. We compare MedCo with a
standard deployment, based on i2b2 and SHRINE (that does
not provide any data protection guarantee) and we show
that MedCo’s performance overhead is practical.
In light of its low overhead, we believe that MedCo can
dramatically accelerate and automate IRB review processes
for sharing sensitive (and identifying) medical data with external researchers. Review processes can take several weeks,
if not months, to permit researchers to access the data,
and these processes are often denied because the necessary
privacy and security guarantees cannot be provided. As
such, MedCo paves the way to new and unexplored usecases where, for example, (i) researchers will be able to
securely query massive amounts of distributed clinical and
genetic data to obtain descriptive statistics indispensable
for generating new hypotheses in clinical research studies,
or (ii) clinicians will be able to find patients with similar
(possibly identifying) characteristics to those of the patient
under examination in order to take more informed decisions
in terms of diagnosis and treatment.
In summary, in this paper we make the following contributions:
- We introduce MedCo, the first operational system enabling the sharing of sensitive clinical and genomic information in a privacy-preserving secure and scalable way
**������������**
-----
- We developed MedCo to be fully compatible with state
of-the-art clinical research platforms such as i2b2 and
SHRINE, hence it can be seamlessly deployed by clinical
sites.
- We extensively tested MedCo in a simulated federation
of three sites, focusing on a clinical-oncology case with
tumor DNA data from The Cancer Genome Atlas, and we
demonstrated its practicality.
- We propose a new generic method to add dummy data
in order to mitigate frequency attacks that can target
the probabilistically encrypted data after they are transformed to deterministically encrypted data for the sake of
enabling equality-matching queries.
#### 2 RELATED WORK
Among the operational systems for sharing clinical or genomic information, SHRINE [12] (the networked version of
i2b2 [11]) and the GA4GH Beacon Network [13] are certainly
the most advanced and widespread. For example, SHRINE
is used in several PCORNet clinical data research networks.
However, as opposed to MedCo, they provide limited privacy guarantees (restricted to ad-hoc result obfuscation)
and no protection of data confidentiality besides standard
access control, thus significantly restraining the possibility
of outsourcing the storage and of processing of the data
to external parties in order to partially offload the costs
of maintaining an always-available interoperable network.
SHRINE provides an ad-hoc mechanism for obfuscating
query results and for locking-out investigators after a certain
number of queries, whereas MedCo features a privacybudget mechanism that achieves differential privacy. Conversely, the Beacon still suffers from risk of re-identification,
as none of the three practical strategies described in [14] has
been implemented yet.
To the best of our knowledge, there are two recent works
dealing with privacy-preserving queries in distributed medical databases; they represent the two main alternatives to
the encryption-based approach followed in this work: The
first one, PRINCESS [15], is based on trusted hardware:
The sites encrypt all their data under AES-GCM (Advanced
Encryption Standard - Galois Counter Mode) and send
them to an enclave that runs in a central server, featuring
an Intel SGX processor; this server decrypts and processes
the sensitive data thus, enabling the secure computation of
statistical models. Compared to our work, PRINCESS can
be more versatile in terms of allowed computations, but it
presents a single point of failure (the central server), and
it centralizes all trust in the enclave and in the attestation
protocol provided by Intel. Furthermore, the memory restrictions of the enclave limit the scalability of the scheme,
requiring compression and batching techniques to enable
processing of large genomic data, for which MedCo scales
much better.
The other recent approach, SMCQL [16], is based on
secure two-party computation; it introduces a framework
for private data network queries on a federated database
of mutually distrustful parties. SMCQL features a secure
query executor that implements different types of queries
(e.g., merge, join, distinct) on the distributed database by
relying on garbled circuits and Oblivious RAM (ORAM)
techniques. Whereas this work features truly decentralized
trust, it does not scale well to scenarios with more than two
sites that are typical in medical contexts with a high number
of collaborating hospitals.
#### 3 PRELIMINARIES
In this section, we briefly introduce the main cryptographic
concepts used throughout the paper.
**3.1** **Deterministic Encryption**
Deterministic encryption (DTE) [17] is a special type of
encryption that preserves the equality property of the plaintexts that, as opposed to probabilistic encryption, makes
ciphertexts indistinguishable and, a priori, unusable. Yet,
DTE also leaks this property; for a given plaintext and key,
DTE always produces the same ciphertext. More formally,
for A, B ⊆ Z with |A| ≤|B|, a function f : A → B is
equality-preserving if for all i, j A, f (i) = f (j) iff i = j. We
∈
say that an encryption scheme with plaintext and ciphertext
spaces D and R, respectively, is deterministic if EDTE(K, ·)
is an equality-preserving function from to for all K
D R ∈K
(where K is the key space).
DTE-based schemes have several advantages and are
mainly used in the context of encrypted database systems
(e.g., CryptDB [18]) as they enable relational databases to
perform equality searches on encrypted data in the same
way as they would operate on the plaintext data. As a
counterpart, they provide less security guarantees than
probabilistic encryption schemes, as they are vulnerable to
inference attacks due to the amount of information they
leak. Hence, their application has to be carefully assessed.
**3.2** **Homomorphic Encryption**
Homomorphic encryption (HE) is a special type of encryption that supports computation on encrypted data. Homomorphic encryption is probabilistic and provides semantic
security, meaning that no adversary without the secret key
can compute any function of the plaintext from the ciphertext. In 2009, Gentry [19] introduced for the first time a
special type of HE that enables arbitrary computations on
ciphertexts, called fully homomorphic encryption (FHE).
Despite its complete functionality, FHE is currently unpractical, as it introduces huge computational and storage
overheads that make it unusable for real-world applications.
For this reason, many variations of FHE have been proposed in the past few years, with the goal of improving
efficiency by sacrificing some flexibility. Such cryptosystems are called practical homomorphic cryptosystems, and
according to their functionality, they can be classified as
additively homomorphic if they satisfy only the addition of
ciphertexts, multiplicatively homomorphic if they satisfy only
multiplication, or somewhat homomorphic if they support (a
limited number of) additions and multiplications.
In this paper, we use the additively homomorphic cryptosystem ElGamal on Elliptic Curves, due to its low ciphertext expansion and fast homomorphic operations
-----
3.2.1 ElGamal On Elliptic Curves
The ElGamal cryptosystem on elliptic curves (ECElGamal) is an asymmetric, probabilistic and additivelyhomomorphic encryption scheme that achieves semantic
security, i.e., ciphertext indistinguishability. It enables additions and multiplications by constants in the ciphertext
domain. As every asymmetric cryptosystem, EC-ElGamal
features three algorithms:
- Key generation: Let E denote an elliptic curve over the
prime field GF(p) and G its base point. Then, the secret
key can be defined as an integer k ∈ GF(p), and the
public key can be derived as K = kG.
- Encryption: Let m be an integer and M = mG its
mapping to the corresponding point on the curve E.
Then, the encryption of M with the public key K is
denoted as EK(M ) = (C1, C2) = (rG, M +rK), where
r is a random nonce.
- Decryption: Given the ciphertext EK(M ) = (C1, C2)
and the secret key k, the decryption algorithm computes the original plaintext point as D(EK(M )) =
−kC1 + C2 = M . The original plaintext m is obtained
by inverting the mapping from the elliptic curve point
M .
Due to its additive homomorphism, EC-ElGamal enables
combining the encryptions of any two messages in order to
obtain an encrypted result that, when decrypted, equals the
sum of these two messages. More formally, let M1 and M2
be any two messages, and α and β be two scalars; then, we
have that αEK(M1) + βEK(M2) = EK(αM1 + βM2).
#### 4 MEDCO ECOSYSTEM
In this section, we introduce the ecosystem in which
MedCo operates. We begin by describing the system and
threat models. We then define the goals of MedCo with
respect to privacy/security and functionality.
�������������������
���������
��������������������
���������
�� ����
���� ���� ��
��
�� ��������������� **����������������**
��� �����������������������������
��
Fig. 2: MedCo’s system and threat models.
**4.1** **System Model**
We consider the system model depicted in Figure 2, where
several clinical sites (Si) want to collaborate in order to share
clinical and genomic data with investigators but do not
want to rely on any central third party or authority for stor
ing or managing their data. Moreover, because of the high
costs (both technical and human) for maintaining a fully
interoperable decentralized network and the increasing size
of the data, clinical sites want to securely outsource the storage of their data to a preferred storage and processing unit
(SPUj). Each site can have its own SPU, or multiple sites
can share the same SPU. All SPUs are organized together
in a peer-to-peer network and form a collective authority.
SPUs are responsible for (i) securely storing the data of the
clinical sites and (ii) securely processing a request of an
authorized investigator that wants to explore clinical sites’
data for generating and validating new research hypotheses
or for identifying cohorts of interest, by finding the patients
that match specific inclusion/exclusion clinical and genetic
criteria across the whole network.
**4.2** **Threat Model**
In this system model, we consider the following threats:
- Storage and processing units: We assume storage and processing units to be honest-but-curious (HBC) parties. Indeed, SPUs can be compromised by internal or external
adversaries that do not tamper with the data-sharing
protocol but can try to infer sensitive information about
the patients from the data stored at their premises and
from the data being processed during the protocol itself.
As a result, SPUs cannot be trusted by clinical sites and
they do not trust each other, either.
- Investigators: We assume investigators to be potentially
malicious-but-covert (MBC) adversaries. Indeed, an investigator can try to legitimately use the system in order
to infer sensitive information about the patients (without
being discovered) by performing consecutive queries and
exploiting the information leaked by the end-results. For
example, a malicious investigator with some background
information about a given individual can infer the presence of such individual into a sensitive cohort (e.g., patients who are HIV-positive) or even reconstruct a subset
of her medical record.
- Clinical sites: We assume clinical sites to be trusted parties.
Finally, we assume that investigators cannot collude with
SPUs, and that at least one SPU does not collude with the
others.
**4.3** **MedCo’s Goals**
To meet end-users expectations and be compliant with
regulations, MedCo has the following goals with respect to
functionality and privacy/security features.
4.3.1 Functionality Goals
The purpose of MedCo is to enable investigators to securely explore the clinical and genomic data stored at all
SPUs by the various clinical sites in the network. Therefore,
MedCo must provide the same functionalities as those provided by state-of-the-art distributed cohort explorers such
as SHRINE [12]:
- (F1) Cohort Exploration: An authorized investigator
should be able to obtain the number of patients per clinical
-----
site who satisfy a set of inclusion/exclusion clinical and
genetic criteria, optionally grouped by age, gender or ethnicity. More formally, MedCo must support SQL queries
such as
SELECT COUNT(patients)
FROM distributed_dataset
WHERE criteria_i AND/OR criteria_j
AND/OR ...
GROUP BY criteria_k;
- (F2) Cohort Selection: An authorized investigator should
be able to obtain the pseudonyms of the patients who
satisfy a set of inclusion/exclusion clinical and genetic
criteria at each clinical site. More formally, MedCo must
support SQL queries such as
SELECT patients
FROM distributed_dataset
WHERE criteria_i AND/OR criteria_j
AND/OR ...;
4.3.2 Security and Privacy Goals
MedCo must always provide the following privacy/security
features:
- (SP1) Trust Decentralization: There should be no single
point of failure in the system.
- (SP2) End-to-end Data Protection: The confidentiality of
the data stored at the SPUs must be protected at rest, in
transit and during computation. The data are encrypted
by the clinical site and the result of the query can be
decrypted only by the investigator issuing the query.
Depending on the access privileges of the investigator
querying the system, MedCo should be able to also provide
the following optional features (either one or both of them):
- (SP3) Unlinkability: The investigator must not be able to
trace a query response back to its original clinical site.
- (SP4) Result Obfuscation: The query result is obfuscated
in order to achieve formal privacy guarantees (e.g., differential privacy) and prevent re-identification.
#### 5 MEDCO CORE ARCHITECTURE & PROTOCOLS
In this section, we provide a detailed description of MedCo.
We begin with a brief overview of the system architecture
and core querying protocol. Then, we describe in detail
the different steps of the system initialization and the data
ingestion phases. Finally, we describe the steps of the secure
querying protocol that enables an investigator to efficiently
query the distributed encrypted data stored at the different
storage and processing units.
**5.1** **General Overview**
The main purpose of MedCo, whose architecture is depicted
in Figure 3, is to reassure clinical sites willing to share their
clinical and genomic data with investigators, by enabling
clinical sites to securely outsource the storage and processing of their data to a set of potentially untrusted storage and
processing units. In order to achieve the privacy and security goals mentioned in Section 4 3 MedCo enables SPUs
������������ ��������������
���� ������ ���������
������ ��� ��������
**����** �������
����� **�**
������� **�������**
�������������������� �� ����
**�** **�** **�**
�� �� �� �
���� **����������** ���� ��
**����**
��
������� �������
�������
**����������������**
��
�� ���������������� ���� ��������������������������������������������������������������������������
Fig. 3: MedCo core architecture and secure query protocol
comprising of: ETL process (steps A, B, C); Query generation
(step 1); Query re-encryption (step 2); Local query processing
(step 3); Local result obfuscation (step 4); Distributed results
shuffling (step 5); Distributed results re-encryption (steps 6):
Results decryption (step 7).
to collectively generate an encryption key for an additivelyhomomorphic encryption system[1], used by clinical sites to
encrypt their data before leaving the local trusted zone
of the site. Through a set of secure distributed protocols,
MedCo enables the SPUs (i) to switch the encryption of the
data from probabilistic encryption to deterministic encryption in order to securely process equality-matching queries,
and (ii) to re-encrypt the query result from an encryption
with the collective public key to an encryption under the
investigator’s public key, so that (only) the investigator can
eventually decrypt the result. And, depending on the access
privileges of the investigator issuing the query, MedCo can
securely shuffle and/or obfuscate the query results in order
to achieve unlinkability and/or differential privacy, respectively (see Section 4.3.2).
**5.2** **System Initialization**
During the initialization of MedCo, each storage and processing unit (SPUi) generates a pair of EC-ElGamal cryptographic keys (ki,Ki), where Ki = Gki, along with a secret
si. Then, all SPUs combine their EC-ElGamal public keys in
order to generate a single collective public key K = [�]i [K][i]
that will be used by the different clinical sites to encrypt the
data to be outsourced.
**5.3** **Data Extraction Transformation and Loading**
During the data-ingestion phase, i.e., extraction transformation and loading (ETL) phase, each clinical site extracts
patient-level data from its private EHR system or clinical research data warehouse, and transforms the data in
order to fit the “star-schema” data model [20] used by
1. For performance reasons, in this work we use EC-ElGamal, but any
other additively homomorphic scheme can be used as well
�
-----
MedCo. The star schema data model is based on the Entity
Attribute-Value (EAV) concept also used by widespread
clinical research systems such as i2b2 [11], where clinical
and genetic observations (or “facts”) about patients (e.g.,
diagnosis, medications, procedures, laboratory values and
genetic variants) are stored in a narrow table called “fact”
table. Observations are encoded by ontology concepts from
an extensible set of medical terminologies, e.g., the International Classification of Disease (ICD) or the US National
Drug Code (NDC). In this data model, four other “dimension” tables further describe the patients’ data and metadata. For example, the “patient dimension” table contains
pseudonymized demographic information of the patients,
and the “visit dimension” table stores information about the
visit, such as its date and time and the type of provider.
In such a data model, the information that clinical sites
want to protect from potential honest-but-curious adversaries at the storage and processing units is represented
by the mapping between the patients in the database and
the set of their clinical and genomic observations stored
in the “fact” table that are considered to be sensitive or
identifying. In order to protect such mapping, each site
separately performs the following three steps:
**A. Generation of Dummy Patients: Each site generates a**
set of dummy patients with plausible clinical observations
specifically chosen so that the distribution of observations
across patients in the “fact” table is as close as possible to the
uniform distribution. We explain the rationale behind this
step in detail in Section 6. To distinguish the real patients
from the dummies, each site also generates a binary flag to
be appended to the demographic information in the “patient
dimension” table. Such flag is set to 1 for real patients and
to 0 for dummy patients.
**B. Data Encryption: In order to break the link between the**
patients and their sensitive observations in the “fact” table,
each site encrypts with the collective public key K the set
of ontology concepts that encode these observations along
with the patients’ binary flags. As EC-ElGamal is a probabilistic encryption scheme, each clinical site obtains a set
of probabilistic ciphertexts that are totally indistinguishable
from each other.
**C. Data Loading and Re-Encryption: After encryption,**
each site uploads the encrypted data to the selected storage
and processing unit that immediately starts a Distributed
Deterministic Re-Encryption (DDR) protocol (the details of
this protocol are explained in Section 5.5) in which the
encrypted concepts are sent across the network of SPUs
so that their encryption is switched from probabilistic to
deterministic. This re-encryption is necessary for enabling
the secure processing of equality-matching queries (as those
defined in Section 4.3) that otherwise would be impossible with probabilistic ciphertexts. Due to the presence of
dummy patients, even if the deterministic nature of the
ciphertexts leaks the equality of the underlying plaintexts,
an honest-but-curious adversary is not able to perform a
frequency attack to distinguish ontology concepts based on
their frequency distribution. Dummy patients are indistinguishable from real patients, as long as the patients’ binary
flags are probabilistically encrypted
**5.4** **Secure Query Protocol**
We assume each investigator that uses MedCo has a pair
of EC-ElGamal cryptographic keys (kI, KI ) and, optionally,
is assigned an initial differential privacy budget �I during
the registration phase. The purpose of such a budget is to
limit the number of queries an investigator with low privileges can run on the system, hence �I -differential privacy
can be guaranteed. The proposed secure query protocol is
illustrated in Figure 3 and comprises the following steps:
**1. Query Generation: The secure query protocol starts with**
an authenticated and authorized investigator who wants to
obtain either the number of patients or the pseudonyms
of the patients who match a set of inclusion/exclusion
clinical and genetic criteria across the different clinical sites.
In clinical research, this procedure is called “cohort selection”. For this purpose, the investigator builds a query by
logically combining (i.e., through AND and OR operators)
a set of “sensitive” and “non-sensitive” concepts from a
common (i.e., shared across the different sites) ontology. The
“sensitive” concepts in the query are encrypted with the
collective public key K and the query is sent along with
the investigator’s public key KI to one of the storage and
processing units.
**2. Query Re-Encryption: The SPU that receives the query**
starts a Distributed Deterministic Re-Encryption (DDR) protocol (described in Section 5.5) in order to switch the encryption of the sensitive concepts in the query from probabilistic
to deterministic. Once the DDR protocol is over, the initial
SPU broadcasts the deterministic version of the query to the
other SPUs in the network.
**3. Local Query Processing: Each SPU locally processes the**
query by filtering the patients (both dummy and real) in
the “patient dimension” table whose observations in the
“fact” table (both the unencrypted and the deterministically
encrypted ones) match the concepts in the query. If the
query requests the list of matching patients’ pseudonyms,
each SPU returns the list of matching patients’ pseudonyms
along with the probabilistically encrypted binary flags.
If the query requests the number of matching patients,
each SPU homomorphically adds the matching-patients’
dummy flags and returns the encrypted result EK(Ri) =
EK([�]j∈φ [f][ j]i [) =][ �]j∈φ [E][K][(][f][ j]i [)][, where][ E][K][(][f][ j]i [)][ is the en-]
crypted flag of the j-th patient in site Si and φ is the
set of patients matching the query. In the homomorphic
summation, the binary flags of the dummy patients have
a null contribution (i.e., EK(0)), hence the encrypted final
result corresponds to the actual number of real matching
patients.
**4. Result Obfuscation: This step is optional and depends**
on (i) the type of query and (ii) the investigator’s privileges.
In order to guarantee differential privacy, each SPU can
obfuscate the encrypted patient counts computed during the
previous step by homomorphically adding noise sampled
from a Laplacian distribution. More specifically, let �q be
the privacy budget allocated for a given query q and μ be
the noise value drawn from a Laplacian distribution with
mean 0 and scale [Δ]�q[f] [, where the sensitivity][ Δ][f][ is equal to 1,]
due to Ri being a count. Then, the encrypted obfuscated
query result is obtained as EK( R[ˆ]i) = EK(Ri + μ) =
-----
EK(Ri)+EK(μ). We note that the query result is released to
the investigator only if the investigator’s differential privacy
budget is enough for such a query, i.e., if �I �q > 0.
−
**5. Result Shuffling: This step is also optional and depends,**
as the previous step, on (i) the type of query and (ii) the
investigator’s privileges. In order to break the link between
the encrypted (potentially obfuscated) query results generated at the different SPUs and the corresponding clinical
sites, the SPUs jointly run a Distributed Verifiable Shuffling
(DVS) protocol (described in Section 5.5) on the set of
encrypted patient counts. As a result, each SPU receives encrypted counts[2], that might have been generated by another
SPU.
**6. Result Re-Encryption: The query results securely com-**
puted by each SPU are encrypted with the collective key
K; to be decrypted by the investigator, each SPU runs
a Distributed Key Switching (DKS) protocol (described in
Section 5.5) that involves the other SPUs and switches the
encryption of the query results from an encryption with
K to an encryption with KI, the investigator’s public key.
After this, the newly encrypted query results are sent back
to the the SPU that initiated the protocol and then on to the
investigator.
**7. Result Decryption: As the query results are encrypted**
with KI, the investigator can use the corresponding secret key kI to decrypt them and obtain the corresponding
plaintext values. If the query results are the list of patients’ pseudonyms along with the patients’ binary flag,
the investigator can simply rule out the dummy patients
by discarding those who have the flag set to zero.
**5.5** **Secure Sub-Protocols**
The secure query protocol of MedCo is based on three secure
and distributed sub-protocols re-adapted from [21]. In this
section, we describe them in detail.
- Distributed Deterministic Re-Encryption (DDR) Proto**col. The DDR protocol enables a set of SPUs to determinis-**
tically re-encrypt data that are probabilistically encrypted
under the collective key generated by all SPUs, without
ever decrypting the data. The purpose of this protocol is to
enable equality-matching queries on probabilistically encrypted data that otherwise would not be possible. More
formally, let n be the number of SPUs in the network,
EK (M ) = (C1, C2) = (rG, M + rK) be the encryption
of a message M under the collective public key K. The
DDR protocol comprises two rounds through all SPUs. In
the first round, each SPUi sequentially uses its secret si
and adds siG to C2. After this first round, the resulting
ciphertext is ( C[˜]1,0, C[˜]2,0) = (rG, M + rK + [�]i[n]=1 [s][i][G][)][.]
In the second round, each SPU partially and sequentially
modifies this ciphertext. More specifically, when SPUi
receives the modified ciphertext ( C[˜]1,i−1, C[˜]2,i−1) from
SPUand Ci[˜]−21,i, it computes = si �C˜2,i− ( 1C −[˜]1,iC,˜C1[˜],i2−,i1)k, wherei�. At the end of theC[˜]1,i = siC[˜]1,i−1
second round, the deterministic re-encryption is obtained
2. The number of encrypted counts received by an SPU corresponds
to the number of sites that have outsourced the storage of their data to
that SPU
by keeping only the second component of the resulting
ciphertext DTs(M ) = C2,n = sM + [�]i[n]=1 [s][i][sG][, where]
s = [�]i[n]=1 [s][i][ is the collective secret corresponding to the]
product of each SPU’s secret.
- Distributed Verifiable Shuffling (DVS) Protocol. The
DVS protocol enables a set of SPUs to sequentially shuffle probabilistically encrypted data so that the outputs
cannot be linked back to the original ciphertexts. More
specifically, the DVS protocol uses the Neff shuffle [22].
It takes as input multiple sequences of EC-ElGamal
pairs (C1,i,j, C2,i,j) forming a a × b matrix, and outputs
a shuffled matrix of ( C[¯]1,i,j, C[¯]2,i,j) pairs such that for
all 1 ≤ i ≤ a and 1 ≤ j ≤ b, ( C[¯]1,i,j, C[¯]2,i,j) =
(C1,π(i),j + rπ[��](i),j[B, C][2][,π][(][i][)][,j][ +][ r]π[��](i),j[P] [)][, where][ r]i,j[��] [is a]
re-randomization factor, π is a permutation and P is a
public key.
- Distributed Key Switching (DKS) Protocol. The DKS
protocol enables a set of SPUs to convert a ciphertext
generated with the collective public key K into a ciphertext of the same data generated under any known
public key U, without ever decrypting them. The DKS
protocol never makes use of decryption. Let EK (M ) =
(C1, C2) = (rG, M + rK) be the encryption of a message
M with the collective public key K. The DKS protocol
starts with a modified ciphertext tuple ( C[˜]1,0, C[˜]2,0) =
(0, C2). Then, each SPU partially and sequentially modifies this element by generating a fresh random nonce vi
and computing ( C[˜]1,i, C[˜]2,i) where C[˜]1,i = C[˜]1,i−1 + viG
and C[˜]2,i = C[˜]2,i−1 − kiC1 + viU . The resulting ciphertext
corresponds to the message m encrypted under the public
key U, ( C[˜]1,n, C[˜]2,n) = (vG, M + vU ) from the original
ciphertext (C1, C2), where v = v1 + . . . + vn.
#### 6 DUMMY-ADDITION STRATEGIES
For cohort-exploration queries, the deterministic encryption
of the ontology concepts applied during the ETL phase
(see Section 5.3) avoids dictionary attacks by any subset of
colluding HBC SPUs due to the distribution of the secrets
si used in the DDR protocol. Nevertheless, a generationof-dummy-patients step is required prior to encryption in
order to avoid leaking to the SPUs (i) the ontology concepts
distribution and (ii) the query result. In this section, we
analyze the optimal dummy-generation strategy to achieve
this goal.
We assume, without loss of generality, that each patient
has a different set of observations; if there were equal
patients in the database, fake ontology concepts could be
added to make them different. The leakage to HBC SPUs
can be estimated by calculating (i) the adversary’s equivocation (i.e., conditional entropy) on the ontology concepts
of the “fact” table given their tagged versions, as an average measure, and (ii) the smallest anonymity set of the
ontology concepts, as a worst-case measure. The higher the
equivocation and the larger the anonymity set is, the lower
the leakage is. For this exposition, we will focus only on
the relation between patients and occurrences of sensitive
ontology concepts, leaving aside the temporal dimension.
This is a simplifying assumption, implying that (a) either
there are no causality relations between concepts or the time
-----
|Ontology code|Col2|a b c d e|Col4|Tagged|x y z r s|dummy flag|
|---|---|---|---|---|---|---|
|real patients|pid 1 pid 2 pid 3|1 1 1 1 0 0 1 1 1 1 1 0 1 1 1||pa pb pc pd pe|1 1 1 0 1 1 1 1 1 0 1 0 1 1 1 0 1 1 1 1 1 1 0 1 1|E(1) E(0) E(1) E(0) E(1)|
|dummy patients|pid 4 pid 5|1 1 0 1 1 1 1 1 0 1|||||
||||||||
#### M M [′] ←−� �−→ ←−� �−→
Fig. 4: Toy example. Ontology concepts mapping to real and added dummy patients with pseudo-identifiers pidi, and ontology
concepts a, b, c, d, e. pa, pb, pc, pd, pe are the randomly sorted version of the patient pseudo-identifiers, and x, y, z, r, s are the
shuffled and deterministically re-encrypted version of the ontology concepts. The binary flag is a probabilistic encryption of 1 for
real patients and 0 for dummies.
dimension is encrypted or not available in the database,
and that (b) the non-sensitive non-encrypted concepts are
independent of the encrypted ones; if this is not the case,
dependent concepts should be reclassified as sensitive and
be encrypted. We will follow the toy example shown in
Figure 4. This figure represents the (horizontally) folded
version of the (vertical) “fact” table, therefore coding each
patient as a row, each ontology concept as a column, and
each observed (resp. unobserved) concept in a patient as a
“1” (resp. “0”) in the corresponding cell.
More formally, let us define the matrix that associates
ontology concepts with patients as the tuple of a random
binary matrix M, where each row can be either a real
or a dummy patient, and each column represents one
ontology concept and two functions σp and σo, which
map the patient pseudo-identifiers (pidj in Fig. 4) to the
rows (pa, pb, pc, pd, pe in Fig. 4), and the observed ontology
concepts (a, b, c, d, e in Fig. 4) to the columns (x, y, z, r, s
in Fig. 4), respectively. These maps represent the shuffling applied to patients before they are assigned their
pseudo-identifiers, and the shuffling and deterministic reencryption applied to ontology concepts before they are
loaded into the SPU’s database. In order to focus on
the practical leakage of the deterministically encrypted
database, let us assume that the deterministic re-encryption
of the concepts and the probabilistic encryption of the patients’ binary flags do not leak anything about their inputs
(their trapdoors cannot be broken), even if they are based on
computational guarantees. Therefore, the adversary (each of
the SPUs) observes the realization of the row- and columnpermuted matrix: A ≡ [M[�] = M [�]], and her equivocation,
with respect to the original information given A, can be
expressed as
H(M, σo, σp|A) = H(M|σo, σp, A) (1)
+ H(σo|σp, A) + H(σp|A)
(a)
= H(σo|σp, A) + H(σp|A) (2)
(b)
≤ H(σo|A) + H(σp)
(c)
≤ H(σo) + H(σp).
Expression (1) can be divided in three terms: the first represents the entropy of M conditioned to the two permutations
and the observed contents of the cells, which is fully deterministic hence zero-entropy (step (a) in (2)); the second
term is the entropy of the ontology concepts permutation
conditioned to the observation of the matrix cells and the
patient permutation, and the third term is the entropy of
the patient permutation conditioned on the observed matrix
contents. We aim at maximizing these two terms.
The last term of the equivocation can be maximized by
making the dummy patients indistinguishable from the real
patients, i.e., drawn from the same distribution. Empirically,
this means that all the patients, real or dummy, have the
same type of distribution, and the contents of the rows
are independent of the position of the dummy patients in
the list. This also makes the two permutations independent
of each other even when conditioned on the contents of
M [�] (step (b) in (2)). In our toy example in Fig. 4, all the
real patients’ rows belong to the same type (weight 4); by
generating two new dummy patients with the same weight,
they become indistinguishable from real patients in our
simplified example.
In order to maximize the entropy of the ontology concepts mapping σo conditioned on A (step (c) in (2)), all the
permutations have to be equiprobable for the given M [�].
This is achieved by flattening the joint distribution of the
observed ontology concepts through the added dummies;
the geometric interpretation of this flattening is that any
column permutation can be cancelled out by a row permutation, such that it is not possible to univocally map any
ontology concept to any column in M [�]. In our toy example,
it can be seen that due to the two added dummies, any fixed
query yields the same number of patients independently of
the permutation applied to the query terms, which gives a
complete indistinguishability between all the deterministically encrypted ontology concepts even in light of the matrix
M [�]. It must be noted that the unobserved concepts do not
have to be added to the table, as the adversary does not
have a priori knowledge of which is the subset of observed
concepts, only its cardinality. Also, this strategy fully breaks
the correlation between ontology concepts; for example, if
the site added only one dummy patient with concepts a, b, e
to the real patients in Figure 4 the individual appearance
rate of the concepts would be flattened, but it would leak
that there is a correlation between the concepts c and d, that
could be identified in the encrypted matrix through an lpoptimization attack [23].
The last bound in (2) is the best that clinical sites can
do with the dummy-patient addition strategy knowing the
-----
matrix of real patients; it maximizes the uncertainty of
the attacker about the original ontology concepts, for any
real distribution of patients and ontology concepts. The
corresponding practical dummy-addition strategy can be
described as follows: Real rows are grouped according to
their weight (number of observations); if the whole set of
observed ontology concepts has n elements, for each group
of rows of weight k < n, dummy rows are added to
complete all the k-combinations of n elements, producing
� n �
rows (counting both real and dummies) per group.
k
In our toy example, (considering independent concepts) the
equivocation goes from 3.58 bits with no dummies to 10.23
bits with the two dummies, and the minimum anonymity
set raises from 2 to 5.
This strategy guarantees the maximum uncertainty for
the adversary for an arbitrary real distribution of concepts
across patients, but it generates a combinatorial number
of dummies, which is not feasible in general (unless the
number of observed concepts is very low). But if some
assumptions can be made about the concepts joint distribution, we can simplify the strategy. If dependencies are only
found within small groups of concepts, the groups being
mutually independent (this is the case for genomic information and dependencies found inside subsets of localized
variants), it is possible to constrain the needed number of
dummies by applying the same dummy-addition strategy
in a restricted block-wise fashion. In order to flatten only
the histogram of group weights, we group the concepts in
independent blocks of size n[�] n and apply the dummy�
generation permutation to the blocks (inter-block), but not to
the contents of each block, until the block distribution is flat,
therefore reducing the needed number of dummy rows. This
trade-off strategy creates an “anonymity set” of ontology
concepts of size n/n[�] in such a way that the adversary cannot distinguish between the set of concepts inside different
blocks. The drawback is that the equivocation is reduced,
as the resulting joint distribution of the ontology concepts
is only flat across blocks, but not inside each block. In the
worst case in terms of leakage (fully correlated concepts
within each block), the achievable adversary’s equivocation
becomes H(M, σo, σp|A) = H(σo|σp, A) + H(σp|A) ≤
H(σo,n/n� ) + H(σp), where σo,n/n� are the permutations of
the n/n[�] blocks of n[�] concepts each. This bound is achieved
when the blocks are mutually independent, hence the best
partitioning strategy consists in keeping correlated concepts
inside the same block. If fully independence between concepts can be assumed (n[�] = 1), it can be seen that flattening
the observations histogram leads to the same maximum
attacker equivocation as the complete permutation strategy
(Eq. (2)), but with a much lower number of added dummies.
In order to further reduce this number, it is possible to set
a minimum anonymity set size m for the concepts and add
dummies to water fill the observation histogram (block-wise
flat, instead of fully flat) until each concept has at least other
m 1 concepts featuring the same number of observations.
−
Finally, it must be noted that whenever a site’s database
is updated, dummies can be regenerated (and encryptions
re-randomized) when the ETL process (see Section 5.3) is
run again for the whole updated database. The DDR protocol uses a different fresh randomness so that the concepts
from the updated database cannot be linked back to the
concepts of the old one.
#### 7 PRIVACY & SECURITY ANALYSIS AND EXTEN
**SIONS (MEDCO+)**
The main privacy and security goals for MedCo are summarized in Section 4.3. In this section, we briefly discuss and
analyze the fulfillment of these targets for MedCo, and we
revisit possible extensions for more stringent requirements.
Security in MedCo is based on the cryptographic guarantees provided by the underlying decentralized sub-protocols
described in Section 5.5. All input sensitive data are either
deterministically (ontology concepts) or probabilistically
(patients’ binary flags) encrypted with collectively maintained keys, such that they cannot be decrypted without
the cooperation of all sites, thus guaranteeing confidentiality
and avoiding single points of failure (SP1 in Section 4.3). For
the full step-by-step security analysis of the distributed subprotocols, we refer the reader to [21]. Following this analysis, paired with the dummy strategy described in Section 6,
it can be seen that MedCo covers the unlinkability requirement (SP3 in Section 4.3) for the query results, thanks to the
DVS protocol; and it protects their confidentiality, as only
the authorized investigator can decrypt the query results
thanks to the DKS protocol (SP2 in Section 4.3). Conversely,
to avoid re-identification (or attribute disclosure) attacks
(SP4 in Section 4.3), MedCo also enables the application of
differentially private noise to the results and, due to the
proposed dummy strategy, it guarantees confidentiality of
the data also against all the SPUs that participate in the
system (SP2 in Section 4.3).
There are two extensions that can be applied to
MedCo in order to satisfy additional confidentiality and
integrity requirements: guaranteeing unlinkability among
investigators’ queries, and obtaining protection against (potentially) malicious SPUs.
**- Query confidentiality: In the basic MedCo system pre-**
sented in Section 5, HBC SPUs can link the ontology concepts used across different queries, as the deterministically
encrypted values of the same concepts are the same for
all the queries. In the case that query confidentiality is
also a requirement (e.g., investigators from pharmaceutical
companies), it is possible to address it by probabilistically
encrypting ontology concepts during the ETL phase and
by deterministically re-encrypting the obtained ciphertexts
with a fresh secret for each new query. Then, the effective
encryption key is different for each fresh run of the DDR
protocol, so it is not possible to link the query terms between
different runs of the shuffling-DDR. When this modified system (which we denote MedCo+) is paired with the proposed
dummy-addition strategy, the terms between queries are
indistinguishable and unlinkable, at the cost of transferring
and re-encrypting at runtime the encrypted database of each
site.
**- Malicious SPUs: MedCo’s threat model assumes HBC**
SPUs to be a credible and plausible assumption, based on
the damage to reputation that a SPU would suffer if it misbehaves in a collective data-sharing protocol. Nevertheless,
it is possible to cope with malicious SPUs by using proof
-----
generation protocols [21] that produce and publish zero
knowledge proofs for all the computations performed at
the SPUs, hence the proofs can be verified by any entity
in order to assess that no SPU deviated from the correct
behavior. This solution yields a hardened and resilient query
protocol, but the cost of producing all proofs results in
a typically unacceptable burden in common data sharing
applications, for which the basic proposed MedCo covers all
fundamental privacy and security requirements and yields a
very competitive performance, as shown in the next Section.
#### 8 IMPLEMENTATION AND EVALUATION
We implemented and tested MedCo on a clinical oncology
use-case by simulating a network of three clinical sites, each
one outsourcing the storage of their data to a different SPU.
**8.1** **Implementation**
To ease its adoption at clinical sites, we implemented
MedCo as three components that fully integrate within
the i2b2 [11] framework and its networking system
SHRINE [12]. i2b2 (Informatics for Integrating Biology and
the Bedside) (i2b2) is the state-of-the-art clinical platform
for enabling secondary use of electronic health records
(EHR) [11]. It is currently used at more than 300 medical
institutions, covering the data of more than 250 million
patients. Its back-end consists of a set of server-side software
modules implemented in Java, called “cells”, that are responsible for the business logic of the platform and are organized in a “hive”. The i2b2 data model is based on the “star
schema” [20]. Queries are built in a dedicated JavaScriptbased Web-client by logically combining ontology concepts
organized in a hierarchical tree-based structure. The three
components of MedCo are:
- A new i2b2 server cell, called “MedCo cell”, developed
in Java and Go. The MedCo cell is responsible for the
execution of the secure query protocol and communicates
with the other i2b2 cells through a REST API. We used the
UnLynx library [21] to implement the DDR, DVS and DKS
secure distributed sub-protocols.
- A new i2b2 Web-client plugin developed in JavaScript.
The plugin is responsible for managing the cryptographic
operations in the browser.
- A data importation tool, developed in Go, that is responsible for encrypting the sensitive ontology concepts and
generating the dummy patients.
These components are publicly available at [24]. We note
that MedCo is not limited to i2b2/SHRINE but can also
be integrated on top of other state-of-the-art platforms for
clinical and translational research, such as TranSMART [2],
in order to make them secure and distributed.
**8.2** **Oncology Use-Case**
The lack of privacy and security guarantees of existing tools
makes sharing sensitive oncological data outside the trusted
boundaries of clinical sites extremely difficult, if not impossible. For this reason, we tested MedCo on genomic and
clinical data from The Cancer Genome Atlas (TCGA) [25]
by performing typical queries for oncogenomics. We report
here two representative examples:
**Query A: Number of patients with skin cutaneous melanoma**
AND a mutation in BRAF gene affecting the protein at position
600. About half of melanoma patients harbor a mutation
in the BRAF gene at position V600E or V600K and can be
treated by the BRAF inhibitor vemurafenib [26]. The proportion of mutated BRAF melanoma is therefore an important
benchmark for a clinic or hospital.
**- Query B: Number of patients skin cutaneous melanoma AND**
a mutation in BRAF gene AND a mutation in (PTEN OR
CDKN2A OR MAP2K1 OR MAP2K2 genes). This query is
based on the fact that patients treated with vemurafenib
develop resistance through mutations that activate the MAP
kinase pathways [27]. When facing drug resistance, finding
another patient with a similar mutation profile could bring
invaluable information for clinical decisions.
We used genomic and clinical data of 8,000 cancer patients, 9 clinical attributes, and an average of 142 genetic
mutations per patient (more than 1 million observations in
total). We imported these data from the Mutation Annotation Format (MAF) into the i2b2 “star schema” data model.
Each mutation is represented as a code comprising the
concatenation of its chromosome, position, reference allele
and tumor allele. Clinical attributes are encoded with the
ICD-10 [28] and ICD-O [29] international terminologies.
**8.3** **Experimental Setup**
The initial testing environment comprises 3 servers interconnected by 10 Gbps links and featuring two Intel Xeon
E5-2680 v3 CPUs @2.5 GHz that support 24 threads on 12
cores, and 256 GB RAM. Each server represents an SPU and
hosts the i2b2/SHRINE Web client with the MedCo plugin,
the i2b2 hive including the SHRINE components, the new
MedCo cell, and the i2b2 database implemented in PostgreSQL. In order to test MedCo’s scalability, we increase the
number of servers up to 9 (see setup S3 below). To set up our
system and facilitate its deployment, we use Docker [30].
To evaluate MedCo’s performance, we consider five
different experimental setups, with each measurement averaged over 10 independent runs, and show MedCo’s computational and storage overhead with respect to an unprotected i2b2/SHRINE deployment:
**S1. ETL runtime for increasing dataset size: We analyze**
the amount of time needed to extract, transform and load
the data (pre-processing), which includes the formatting,
the initial probabilistic encryption, the deterministic reencryption of sensitive ontology concepts, and the loading
of the data in the i2b2 database.
**S2. Query runtime breakdown: We run queries A and B**
(see Section 8.2) on a federation of 3 SPUs, each storing the
full initial dataset (i.e., around 1 million observations on
8,000 patients at each SPU), and report the query-runtime
breakdowns for each step of the secure query protocol.
**S3. Query runtime for increasing dataset size: We run**
queries A and B (see Section 8.2) on a federation of 3 SPUs in
order to study MedCo’s scalability with respect to increasing
dataset sizes.
**S4. Query runtime overhead for increasing number of**
**SPUs: We run queries A and B (see Section 8 2) on a**
-----
federation with an increasing number of SPUs, each storing
the whole initial dataset.
**S5. Network traffic for varying query size: We study the**
amount of network traffic inter-SPU for queries with an
increasing number of ontology concepts.
**8.4** **Performance Results**
In the following, we report the performance results for
the aforementioned use-cases and experimental setups. We
show MedCo’s computational and storage overhead with
respect to an unprotected i2b2/SHRINE deployment.
As shown in Figure 5, the ETL phase (setup S1) is a
costly operation in MedCo. We can distinguish two separate
subphases: (i) the processing of the ontology (including the
parsing, the encryption and the distributed deterministic reencryption), which only depends linearly on the size of the
ontology and is usually constant, and (ii) the processing
of patients’ observations, which depends linearly on the
number of observations/patients but does not involve any
costly encryption operation hence it is much faster than
the ontology processing. We note that the ETL phase is
performed only once and can be significantly optimized
through parallel computing. If new data need to be added
after the first importation, there is no need to re-process the
ontology again.
Figure 6 provides query-runtime breakdowns for both
query A and query B (setup S2). The times for query-parsing
and encryption/decryption in the Web client, broadcasting
the query across the different SPUs, and result obfuscation
are all negligible, so we do not account for them. Unexpectedly, results show that the standard i2b2 query to the central
“fact” table is the most expensive operation in MedCo,
as it depends on the total number of observations in the
database. In this case, each SPU stores approximately 1 million observations (both genomic and clinical) per affiliated
clinical site (one site per SPU in our setting). This time is also
linear in the number of ontology concepts used in the query
Fig. 5: ETL time vs database size for experimental setup S1.
(a) Query A
(b) Query B
Fig. 6: Query-runtime breakdown for queries A and B in a
network with three sites and three SPUs for experimental setup
S2. The vertical black line signals the point where each node
has to wait for the others before it can proceed.
(96 for query A and 281 for query B) and it is inherent to the
standard i2b2 database management for SQL-queries to the
“fact” table. The times for fetching the encrypted patients’
binary flags from the “patient dimension” table and the
homomorphic aggregation (Step 3 in the query workflow)
depend linearly on the number of patients satisfying the
query criteria and can be extremely fast for rare ontology
concepts or rare combinations of concepts. For example, for
queries A and B, homomorphic aggregation takes around
30 and 8 milliseconds respectively, as only around 32 and
7 patients per site satisfy the query criteria. Differently, the
deterministic re-encryption time is linear in the number of
sensitive concepts in the query and number of SPUs in the
network, as each probabilistically encrypted concept has to
be sequentially modified by each SPU Such a process takes
-----
(a) Query A runtime vs database size. (b) Query B runtime vs database size.
(c) Queries A and B runtime vs number of SPUs. (d) Network traffic vs query size
Fig. 7: MedCo’s performance results for experimental setups S3-S5.
less time for query A than for query B, as they respectively
comprise 96 (95 mutations and 1 clinical attribute) and 281
(280 mutations and 1 clinical attribute) query attributes.
The remaining secure distributed operations introduced by
MedCo depend on the number of SPUs in the network, but
they are negligible, as they involve only one ciphertext, i.e.,
the encrypted query result.
Figure 7 shows the performance results for setups S3S5. The measurements are averaged out between SPUs. For
setup S3 (Subfig. 7a and Subfig. 7b), in order to study
MedCo’s ability to scale with increasing database sizes,
we randomly sample patients from the original dataset of
8k patients and create smaller datasets of 1k, 2k and 4k
patients per site. For setups S4 and S5 (Subfig. 7c and
Subfig. 7d), we use the initial dataset (8k patients). Results
show that MedCo is extremely efficient and performancewise comparable to the insecure i2b2/SHRINE deployment.
MedCo’s overhead only depends on the number of sensitive
concepts in the query, the number of matching patients satisfying the research criteria and marginally on the number of
SPUs in the network. As shown in Subfigure 7c, the number
of SPUs affects only the time needed by the distributed protocols to deterministically re-encrypt the sensitive ontology
concepts in the query and to re-encrypt the query end-result
under the investigator’s key.
In Subfigures 7a, 7b and 7c, we can also observe that
MedCo+ has a relatively higher runtime cost as a counterpart for achieving query unlinkability, because all the
observations in the “fact” table of each SPUs have to be
deterministically re-encrypted on the fly by the whole set of
SPUs for each new query. This is confirmed by Subfigure 7d
where the network traffic is significant and almost constant
for MedCo+, whereas for MedCo it is almost negligible and
it increases with the number of concepts in the query. We
note, however, that the privacy enhancements brought by
MedCo+ might be necessary only under specific circumstances (e.g., when an investigator from a pharmaceutical
company is using the system).
Finally, the storage overhead introduced by encryption
affects only the “concept dimension” table that stores the
-----
ontology, and it is in the order of 4x, as MedCo s de
terministic re-encryption converts each ontology concept,
represented by 64-bit integers, into a 32-bytes ciphertext.
Depending on the specific distribution of ontology codes
across patients, a varying number of dummy patients must
also be considered. In the tested oncology use-case, we
assume independent codes and follow the dummy-addition
strategy described in Section 6. As a result, we obtain an
increase factor of 3.6x.
#### 9 CONCLUSION
In this paper, we have presented MedCo, the first operational scalable system that enables secure sharing of
sensitive medical data, which so far was impossible due
to the low security guarantees of existing operational systems. MedCo relies on secure distributed protocols and a
new dummy-records addition strategy that enables different
privacy/security vs. efficiency trade-offs. With its generic
architecture, MedCo is easily deployable on top of existing
health information systems such as i2b2 or tranSMART.
Finally, results on a clinical oncology use-case have shown
practical query-response times and good scalability with
respect to the number of sites and amount of data. Therefore,
we firmly believe that MedCo represents a concrete solution
for fostering medical data sharing in a privacy-conscious
and regulatory-compliant way.
#### REFERENCES
[1] J. V. Selby, A. C. Beal, and L. Frank, “The patient-centered
outcomes research institute (pcori) national priorities for research
and initial research agenda,” Jama, vol. 307, no. 15, pp. 1583–1584,
2012.
[2] B. D. Athey, M. Braxenthaler, M. Haas, and Y. Guo, “tranSMART:
an open source and community-driven informatics and data
sharing platform for clinical and translational research,” AMIA
Summits on Translational Science Proceedings, vol. 2013, p. 6, 2013.
[3] Swiss Academies of Arts and Sciences, “Swiss Personalized
Health Network,” http://www.samw.ch/en/Projects/SPHN.
html, last Accessed: June 11, 2018.
[4] The Global Alliance for Genomics and Health, “A federated
ecosystem for sharing genomic, clinical data,” Science, vol. 352,
no. 6291, pp. 1278–1280, 2016.
[5] U.S. Department of Health & Human Services, “The health
insurance portability and accountability act (hipaa),” https://
www.hhs.gov/hipaa/index.html, last Accessed: June 11, 2018.
[6] EU Parlament, “The EU General Data Protection Regulation
(GDPR),” http://www.eugdpr.org/, last Accessed: June 11, 2018.
[7] “All of us research program,” https://allofus.nih.gov/, last Accessed: June 11, 2018.
[8] “The 100,000 genomes project protocol v3, genomics england.”
https://www.genomicsengland.co.uk/, last Accessed: June 11,
2018.
[9] T. G. A. for Genomics and Health, “Beacon network,” https://
beacon-network.org/, 2017, last Accessed: June 11, 2018.
[10] U.S. Department of Health and Human Services, “Breach
portal: Notice to the secretary of hhs breach of unsecured
protected health information,” https://ocrportal.hhs.gov/ocr/
breach/breach_report.jsf, last Accessed: June 11, 2018.
[11] S. N. Murphy, G. Weber, M. Mendis, V. Gainer, H. C. Chueh,
S. Churchill, and I. Kohane, “Serving the enterprise and beyond
with informatics for integrating biology and the bedside (i2b2),”
Journal of the American Medical Informatics Association, vol. 17,
no. 2, pp. 124–130, 2010.
[12] G. M. Weber, S. N. Murphy, A. J. McMurry, D. MacFadden,
D. J. Nigrin, S. Churchill, and I. S. Kohane, “The shared health
research information network (shrine): a prototype federated
query tool for clinical data repositories,” Journal of the American
Medical Informatics Association vol 16 no 5 pp 624–630 2009
[13] G. A. for Genomics and Health, The beacon project, https://
beacon-network.org/\#/, last Accessed: June 11, 2018.
[14] J. L. Raisaro, F. Tramèr, Z. Ji, D. Bu, Y. Zhao, K. Carey, D. Lloyd,
H. Sofia, D. Baker, P. Flicek, S. Shringarpure, C. Bustamante,
S. Wang, X. Jiang, L. Ohno-Machado, H. Tang, X. Wang, and J.-P.
Hubaux, “Addressing Beacon re-identification attacks: quantification and mitigation of privacy risks,” Journal of the American
Medical Informatics Association, no. 0, pp. 1–8, 2017.
[15] F. Chen, S. Wang, X. Jiang, S. Ding, Y. Lu, J. Kim, S. C. Sahinalp,
C. Shimizu, J. C. Burns, V. J. Wright et al., “Princess: Privacyprotecting rare disease international network collaboration via
encryption through software guard extensions,” Bioinformatics, p.
btw758, 2017.
[16] J. Bater, G. Elliott, C. Eggen, S. Goel, A. Kho, and J. Rogers,
“Smcql: Secure querying for federated databases,” Proc. VLDB
Endow., vol. 10, no. 6, pp. 673–684, Feb. 2017. [Online]. Available:
https://doi.org/10.14778/3055330.3055334
[17] M. Bellare, A. Boldyreva, and A. O Neill, “Deterministic and efficiently searchable encryption,” Advances in Cryptology-CRYPTO
2007, pp. 535–552, 2007.
[18] R. A. Popa, C. Redfield, N. Zeldovich, and H. Balakrishnan,
“Cryptdb: protecting confidentiality with encrypted query processing,” in Proceedings of the Twenty-Third ACM Symposium on
Operating Systems Principles. ACM, 2011, pp. 85–100.
[19] C. Gentry, “A fully homomorphic encryption scheme,” Ph.D.
dissertation, Stanford University, 2009.
[20] P. M. Nadkarni and C. Brandt, “Data extraction and ad hoc query
of an entity attribute value database,” Journal of the American
Medical Informatics Association, vol. 5, no. 6, pp. 511–527, 1998.
[21] D. Froelicher, P. Egger, J. S. Sousa, J. L. Raisaro, Z. Huang,
C. Mouchet, B. Ford, and J.-P. Hubaux, “UnLynx: A decentralized
system for privacy-conscious data sharing,” in Proceedings on
Privacy Enhancing Technologies, vol. 4, no. EPFL-CONF-229308,
2017, pp. 152–170.
[22] C. A. Neff, “Verifiable mixing (shuffling) of elgamal pairs,” VHTi
Technical Document, VoteHere, Inc, 2003.
[23] M. Naveed, S. Kamara, and C. V. Wright, “Inference
attacks on property-preserving encrypted databases,” in
Proceedings of the 22Nd ACM SIGSAC Conference on Computer
and Communications Security, ser. CCS ’15. New York,
NY, USA: ACM, 2015, pp. 644–655. [Online]. Available:
http://doi.acm.org/10.1145/2810103.2813651
[24] LCA1, EPFL, “Medco source code,” https://c4science.ch/w/
medco/, last Accessed: June 11, 2018.
[25] K. Tomczak, P. Czerwi´nska, and M. Wiznerowicz, “The cancer
genome atlas (tcga): an immeasurable source of knowledge,”
Contemporary oncology, vol. 19, no. 1A, p. A68, 2015.
[26] P. A. Ascierto, J. M. Kirkwood, J.-J. Grob, E. Simeone, A. M.
Grimaldi, M. Maio, G. Palmieri, A. Testori, F. M. Marincola, and
N. Mozzillo, “The role of braf v600 mutation in melanoma,”
Journal of translational medicine, vol. 10, no. 1, p. 85, 2012.
[27] H. Yang, D. Kircher, K. Kim, A. Grossmann, M. VanBrocklin,
S. Holmen, and J. Robinson, “Activated mek cooperates with
cdkn2a and pten loss to promote the development and maintenance of melanoma,” Oncogene, vol. 36, no. 27, pp. 3842–3851,
2017.
[28] W. H. Organization, The ICD-10 classification of mental and behavioural disorders: clinical descriptions and diagnostic guidelines.
World Health Organization, 1992, vol. 1.
[29] A. G. Fritz, International classification of diseases for oncology: ICDO. World Health Organization, 2000.
[30] D. Merkel, “Docker: lightweight linux containers for consistent
development and deployment,” Linux Journal, vol. 2014, no. 239,
p. 2, 2014.
-----
**Jean Louis Raisaro earned his PhD in Com**
puter and Communication Sciences in 2018 from
EPFL, Lausanne, Switzerland. Prior to that he
earned a MS and BS in Biomedical Informatics and Bioinformatics in 2012 and 2009 from
University of Pavia, Pavia, Italy. His main interests are the design and development of new
efficient privacy-enhancing technologies for the
protection of medical data with a special focus
on genetic data. He is an expert in applied cryptography, privacy and medical informatics.
**Juan Ramón Troncoso-Pastoriza Ph.D. Tele-**
com. Engineering (2012); elected member of
the IEEE Information Forensics and Security TC
and the IEEE Signal Processing Society Student
Services Committee for the period 2017-2019,
and associate editor of four journals on Information Security (EURASIP JIS, IET IFS, Elsevier
DSP and Elsevier JVCI). His research interests
include Secure Signal Processing, Applied Cryptography and Genomic Privacy, areas in which
he has published numerous papers in top conferences and journals and holds several international granted patents.
**Mickaël Misbach is pursuing a Master in Com-**
munication Systems at EPFL in Lausanne,
Switzerland, with a specialization in Information
Security. He is expected to graduate in September 2018. During the last years of his studies
he worked on medical data privacy, being the
main developer of the privacy-conscious cohort
explorer MedCo.
**João Sá Sousa is currently a Security / Privacy**
Software Engineer at EPFL under the direction
of professor Jean-Pierre Hubaux. He has a MS
and BS degree in Informatics Engineering at
the University of Coimbra and did a 3-month
internship at CMU-SV. His main interests include
Wireless Security, Genomic Privacy, Cryptography, Android Development, Web Development
and Business Management.
**Sylvain Pradervand received a Ph.D. degree**
in molecular biology from the University of Lausanne in 1998. After a postdoc studying transcriptomics in heart disease models at the University of California San Diego, he turned his interests to bioinformatics. He is currently leading
the bioinformatics team of the genomic technologies facility of the University of Lausanne and the
bioinformatics team of the clinical research support platform of the Lausanne University Hospital.
**Edoardo Missiaglia obtained his bachelor s de**
gree in biology (1994) from the University of
Padova, master’s degree in genetics (1998) from
the University of Bologna and PhD in pathological oncology (2003) from the University of
Verona. He worked at the ICRF (Cancer Research UK) (2001-2003) as research assistant
and at University of Verona (2003-05) and ICR
(2005-2010) as Post-Doc and bioinformatician.
He has been working as Project Manager at the
SIB (2010-2014). He became the scientific director of the molecular pathology laboratory of the Institute of Pathology at
CHUV in August 2014.
**Olivier Michielin is associate professor at the**
University of Lausanne. He obtained a diploma
of Physics in 1991 at the EPFL and an MD
from the University of Lausanne in 1997. He
pursued his PhD training under the supervision of Jean-Charles Cerottini (LICR) and Martin
Karplus (Harvard and Strasbourg Universities).
He was appointed Group Leader of the Swiss
Institute of Bioinformatics in 2002 and became
an Assistant Professor and Privat Docent at the
Medical Faculty of Lausanne in 2004 and 2005,
respectively. In parallel, he has trained as a medical oncologist and
obtained his board certification in 2007 at the Multidisciplinary Oncology
Center (CePO) of Lausanne where he is currently in charge of the
melanoma clinic.
**Bryan Ford leads the Decentralized/Distributed**
Systems (DEDIS) research group at the Swiss
Federal Institute of Technology in Lausanne
(EPFL). Ford focuses broadly on building secure decentralized systems, touching on topics
including private and anonymous communication, scalable decentralized systems, blockchain
technology, Internet architecture, and operating
systems. Ford earned his B.S. at the University
of Utah and his Ph.D. at MIT, then joined the faculty of Yale University where his work received
the Jay Lepreau Best Paper Award and grants from NSF, DARPA, and
ONR, including the NSF CAREER award.
**Jean-Pierre Hubaux is a full professor at EPFL.**
Through his research, he contributes to laying
the foundations and developing the tools for protecting privacy in tomorrow’s hyper-connected
world. He has pioneered the areas of privacy
and security in mobile/wireless networks and in
genomics. He is a Fellow of both IEEE (2008)
and ACM (2010).
-----
|
{
"disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1109/TCBB.2018.2854776?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/TCBB.2018.2854776, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "GREEN",
"url": "https://infoscience.epfl.ch/record/256348/files/08410926.pdf"
}
| 2,019
|
[
"JournalArticle"
] | true
| 2019-07-01T00:00:00
|
[] | 19,529
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Engineering",
"source": "s2-fos-model"
},
{
"category": "Physics",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/0273efebece84154e5d64f42df67e2139df992b6
|
[
"Computer Science"
] | 0.875587
|
A performance model for the communication in fast multipole methods on high-performance computing platforms
|
0273efebece84154e5d64f42df67e2139df992b6
|
The international journal of high performance computing applications
|
[
{
"authorId": "3083117",
"name": "H. Ibeid"
},
{
"authorId": "2274654",
"name": "Rio Yokota"
},
{
"authorId": "145907031",
"name": "D. Keyes"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"int j high perform comput appl",
"International Journal of High Performance Computing Applications",
"Int J High Perform Comput Appl"
],
"alternate_urls": [
"https://journals.sagepub.com/home/hpc"
],
"id": "8ce575db-79d9-4601-83df-e1e96f6b4e3b",
"issn": "1094-3420",
"name": "The international journal of high performance computing applications",
"type": "journal",
"url": "http://www.sagepub.com/journals/Journal201339/title"
}
| null |
|Col1|論文 / 著書情報 Article / Book Information|
|---|---|
|Title|A performance model for the communication in fast multipole methods on high-performance computing platforms|
|Author|Huda Ibeid, Rio Yokota, David Keyes|
|J ournal/Book name|International J ournal of High Performance Computing Applications, Vol. 30, No. 4, pp. 423--437|
|Issue date|2016, 3|
|DOI|http://dx.doi.org/10.1177/1094342016634819|
|Note|このファイルは著者（最終）版です。 This file is author (final) version.|
-----
## A Performance Model for the Communication in Fast Multipole Methods on HPC Platforms
#### Huda Ibeid, Rio Yokota, and David Keyes
Division of Computer, Electrical and Mathematical Sciences and Engineering
King Abdullah University of Science and Technology, Saudi Arabia
**Abstract**
Exascale systems are predicted to have approximately one billion cores, assuming Gigahertz cores. Limitations on affordable network topologies for distributed memory systems of such massive scale bring new
challenges to the currently dominant parallel programing model. Currently, there are many efforts to evaluate the hardware and software bottlenecks of exascale designs. It is therefore of interest to model application
performance and to understand what changes need to be made to ensure extrapolated scalability. The fast
multipole method (FMM) was originally developed for accelerating N -body problems in astrophysics and
molecular dynamics, but has recently been extended to a wider range of problems, including preconditioners
for sparse linear solvers [31]. Its high arithmetic intensity combined with its linear complexity and asynchronous communication patterns make it a promising algorithm for exascale systems. In this paper, we
discuss the challenges for FMM on current parallel computers and future exascale architectures, with a focus
on inter-node communication. We focus on the communication part only; the efficiency of the computational kernels are beyond the scope of the present study but see, e.g., [3]. We develop a performance model
that considers the communication patterns of the FMM, and observe a good match between our model
and the actual communication time on four HPC systems, when latency, bandwidth, network topology, and
multi-core penalties are all taken into account. To our knowledge, this is the first formal characterization of
inter-node communication in FMM that validates the model against actual measurements of communication
time. The ultimate communication model is predictive in an absolute sense; however, on complex systems,
this objective is often out of reach, or of a difficulty out of proportion to its benefit when there exists a
simpler model that is inexpensive and sufficient to guide coding decisions leading to improved scaling. The
current model provides such guidance.
### 1 Introduction
_N_ -body problems arise in many areas of physics (e.g.,
astrophysics, molecular dynamics, acoustics, electrostatics). In these problems, the system is described
by a set of N particles and the dynamics of the system arise from interactions that occur between every
pair of particles. This requires (N [2]) computational
_O_
complexity. For this reason, many efforts have been
directed at producing fast N -body algorithms. More
efficient algorithms of the particle interaction problem can be provided by a hierarchical approach using
tree structures. In this approach, the computational
domain is hierarchically subdivided, and the particles are clustered into a hierarchical tree structure.
An approximation of tunable accuracy is applied to
far-field interactions, whereas near-field interactions
are summed directly. When the far-field expansion
is calculated against the particles directly, this approach called a treecode [1]. When the far-field effect
is translated to local-expansions before summing their
effect, it is called a fast multipole method (FMM)
[15, 5]. These approaches bring the complexity down
to (N log N ) and (N ) for treecode and FMM, re_O_ _O_
spectively. FMM has been listed as one of the top
ten algorithms of the twentieth century [8] due to its
wide applicability and impact on scientific computing. It was originally developed for applications in
electrostatics and astrophysics, but continues to find
new areas of application such as aeroacoustics [29],
fluid dynamics [14], magnetostatics [28], and electrodynamics [33]. Because of its linear complexity, FMM
scales well with respect to the problem size, if implemented efficiently. For future computer systems the
conservation of flops is less important than the conservation of distant loads and stores to supply the
1
-----
arguments for the flops. FMM stands out among hierarchical O(N ) algorithms for its strong arithmetic
intensity.
Since the performance of a single-processor core has
plateaued, future supercomputing performance will
depend mainly on increases in system scale rather
than improvements in single-processor performance.
Processor counts are already in the millions for the
top system. Modeling application performance at
such scales is required to guide algorithmic choices
and tunings on existing architectures and evaluate
contemplated architectures. Since the performance
of the FMM has a large impact on a wide variety of
applications across a wide range of disciplines, it is
important to understand the challenges that FMM
implementations face on architectures with increased
parallelism, as well as to predict and locate bottlenecks that might cause performance degradation. On
future architectures where computation becomes relatively cheap compared to data movement, we anticipate that inter-node communication will become the
bottleneck. The priority of the present study is the
communication model of FMM.
To model the performance, we start with the baseline model, namely (α, β) model for communication,
where α is the latency and β is the inverse bandwidth.
Then, some penalties are added to the baseline model
based on machine constraints. These penalties include distance and reduced per-core bandwidth. Our
performance model is related to universal communication features and can be applied regardless of local FMM implementation choices, core-scale machine
characteristics that do not affect communication, and
arithmetic workload associated with other aspects of
the computation. Of course, the importance of communication as a bottleneck depends strongly on the
cost of other tasks, but it is important to be able to
evaluate communication costs as a component in an
overall cost model. The Byte-count parameters in our
model makes it adaptable to any of the various FMM
implementations, while the penalties in our model are
tunable to various architectures. We validate our performance model on four different architectures, Shaheen (BG/P), Mira (BG/Q), Titan (Cray XK7), and
Piz Dora (Cray XC40).
The focus of this paper is on characterizing the
FMM communication, not on introducing a new
model. For this purpose, we apply a performance
model developed originally and applied to multigrid
methods, which have a different communication pattern. A new application of an existing tool emphasizes
the versatility of the tool. Meanwhile, such detailed
analysis of the communication in FMM has not previously been reported, so there is particular relevance
to the FMM community, and to the HPC community that exploits, or will exploit at exascale, FMM
solvers.
The paper is organized as follows. Section 2 gives
an overview of related work. Section 3 summarizes
some performance challenges that face FMM on parallel machines. These challenges include massive parallelism and degradation due to inter-node communication. In Section 4, an exposition of the fast multipole method sufficiently detailed to expose communication properties is given. Section 5 describes our
performance model. Experiments done to validate the
performance models are provided in Section 6 and we
conclude in Section 8.
### 2 Related work
Performance modeling and characterization for understanding and predicting the performance of scientific applications on HPC platforms have been
targeted by many related projects. For example,
Clement and Quinn developed a performance prediction methodology through symbolic analysis of their
source code [6]. Mendes and Reed focused on predicting scalability of an application program executing on a given parallel system [24]. Mendes proposed
methodology to predict the performance scalability of
data parallel applications on multi-computers based
on information collected at compile time [23]. The approach of combining computation and communication
to obtain a general performance model is described by
Snavely et al. [27]. DeRose and Reed concentrate on
tool development for performance analysis [7]. Performance models for implicit CFD codes have been
considered [17]. The efficiency of the spectral transform method on parallel computers has been evaluated by Foster [10]. Kerbyson et al. provide an analytical model for the application SAGE [19]. Performance models for AMG were developed by Gahvari
_et al. [11], who have also analysed the performance_
of AMG over a dragonfly network in [12]. Traditional
evaluation of specific machines via benchmarking is
presented by Worley [30].
Scaling FMM to higher and higher processor counts
has been a popular topic [25, 18], while extensive
study of single-node performance optimization, tuning, and analysis of FMM has also been of interest
[4]. However, there has been little effort to model the
inter-node communication of FMMs. Lashuk et al.
derive the overall complexity of FMM on distributed
memory heterogeneous architectures [20], but do not
validate the model against the actual performance.
The present work is based on the communication
2
-----
model for AMG [11], and extends their theory to
FMM. To our knowledge, this is the first formal characterization of inter-node communication in FMM,
which validates the model against actual measurements of communication time.
### 3 Performance challenges
High performance computing systems have shown a
sustained exponential growth with performance improvement of approximately 10x every 3.6 years as
measured, for instance, by the Gordon Bell Prizes
or the Top500 benchmark over the past 2.5 decades.
This performance improvement comes at a cost in
code complexity and introduces many challenges.
Furthermore, the development of an exascale computing capability will cause significant and dramatic
changes in computing hardware architecture relative
to current petascale computers. In this section we
present some of the challenges faced by FMMs to
achieve good parallel performance on future exascale
systems.
#### 3.1 Trends in Computer Hardware
Computers consisting of nodes in the tens of thousands with cores per node in the hundreds have
emerged as the most widely used high-performance
computing platforms. These nodes communicate by
sending messages through a network, which leads to
lower scalability and less performance due to cores on
a single node contenting for access to the interconnect. We discuss multicore and manycore issues in
more detail when presenting our performance models
that take this into account.
#### 3.2 Communication
Two types of costs in terms of time and energy are
usually analyzed separately: computation (flops) and
communication (Bytes). Communication involves
moving data between levels of a memory hierarchy in
case of sequential algorithms and exchanging data between processors over a network in the case of parallel
algorithms. Therefore, without considering overlap,
the running time of an algorithm is the sum of three
terms: the number of flops times the time per flop,
the number of words moved divided by the bandwidth
(measured as words per unit time), and the number
of messages times the latency. The last two terms determine the time consumed by communication. The
time per flop is already an order of magnitude less
than reciprocal bandwidth and latency and the gaps
between computation and communication are growing exponentially with time. (See Table 2 under the
machine descriptions in Section 6 below.) Communication performance models can guide development of
algorithms to help reduce the communication.
### 4 Fast multipole method
_N_ -body methods are most commonly used to simulate
the interaction of particles in a potential field, which
has the form
Here, f (xi) represents a field value evaluated at a
point xi which is generated by the influence of sources
located at xj with weights qj. K(xi, xj) is the kernel
that governs the interactions between evaluation and
source particles. The direct approach to simulate the
_N_ -body problem is relatively simple; it evaluates all
pair-wise interactions among the particles. While this
method is exact to within machine precision, the solution is (N [2]) in its computational complexity, which
_O_
is prohibitively expensive for even modestly large data
sets. However, its simplicity and ease of implementation make it an appropriate choice when simulating
small particle sets (N < 1000) where high accuracy is
desired [26]. For a larger number of particles, many
faster algorithms have been invented, e.g., treecodes
[1] and, the fast multipole method (FMM) [15]. The
main idea behind these fast algorithms is to coarse
grain the effect of sufficiently far particles as permitted by rigorous analysis. The most common way to
achieve this approximation is to cluster the far particles into successively larger groups by constructing
a tree. The treecode clusters the far particles and
achieves (N log N ) complexity. The FMM further
_O_
clusters the near particles in addition to the far particles to achieve (N ) complexity.
_O_
In this section, we present an overview of fast algorithms that have been developed for the calculation
of N -body problems. First, the spatial hierarchy and
the fast approximate evaluation of these algorithms
are discussed. Then, a description of the communication introduced by the domain partitioning scheme
used in these algorithms is provided. The main focus
is on the data flow of the FMM algorithm for which
we develop the performance model.
#### 4.1 FMM Overview
This overview is intended to introduce some key ingredients of the FMM. The mathematics behind the
_f_ (xi) =
_N_
�
_qjK(xi, xj)_ (1)
_j=1_
3
-----
(a) 2-D view (b) Tree view
Figure 1: Hierarchical decomposition
specific FMM kernels is well documented elsewhere
and its detail conveniently decouples, given a simple
interface to the communication model. For details
of the mathematics we refer the reader to previous
publications on FMM [2, 5].
**4.1.1** **Basic Component**
Both treecodes [1] and the FMM [15] are based on two
key ideas: the tree representation for the spatial hierarchy, and the fast approximate evaluation. The spatial hierarchy means that the computational domain
is hierarchically decomposed into increasing levels of
refinement, and then the near and far subdomains can
be identified at each level. The three-dimensional spatial domain of the treecode and FMM is represented
by octrees, where the space is recursively subdivided
into eight cells until the finest level of refinement or
“leaf level. Figure 1 illustrates such a hierarchical
space decomposition for a two-dimensional domain
(a), associated to a quad-tree structure (b). The original FMM [16] is based on a series expansion of the
Laplace Green’s function (1/r) and therefore can be
applied to the evaluation of related potentials and/or
forces [13]. The approximation reduces the number
of operations in exchange for accuracy.
**4.1.2** **Flow of Calculation**
Figure 2, shows the flow of FMM where the effect of
the source particles, shown in red in the lower left
corner, are calculated on the target particles, shown
in blue in the lower right corner. The schematic is
a 2-D representation of what is actually a 3-D octree
structure. The calculation starts by transforming the
mass/charge of the source particles to a multipole
expansion (P2M). Then, the multipole expansion is
translated to the center of larger cells (M2M). Then,
the influence of multipoles on the particles is calculated in three steps. First, it translates the multipole
expansion to a local expansion (M2L). Next, the center of expansion is translated to smaller cells (L2L).
Finally, the effect of the local expansion in the far
field is translated onto the target particles (L2P). All
pairs interaction is used to calculate the effect of near
field on target particles (P2P).
#### 4.2 FMM Communication Scheme
Partitioning of the FMM global tree structure and
communication stencils is shown in Figure 3. The binary tree on the left side is a simplification of what
is actually an octree in a 3-D FMM. Likewise, the
schematics on the right are a 2-D representation of
what is actually a 3-D grid structure. Each leaf of the
global tree is a root of a local tree in a particular MPI
process, where the global tree has Lglobal levels, and
the local tree has Llocal levels. Each process stores
only the local tree, and communicates the halo region
at each level of the local and global tree as shown
in the red hatched region in the four illustrations on
the right. The blue, green, and black lines indicate
global cell boundaries, process boundaries, local cell
boundaries, respectively. The switch between local
and global trees produces a change in the communication pattern, as revealed in the heat map in Figure
4, where the switch is between levels 3 and 4.
### 5 Modeling Performance
Performance modeling is a key ingredient in high performance computing. It has a great importance in
the design, development and optimization of applications, architectures and communication systems. It
also plays a crucial role in understanding important
performance bottlenecks of complex systems. For this
reason, performance models are used to analyze, predict, and calibrate performance for systems of interest. The tree-based communication of FMM is in
4
-----
#### M2L
multipole to local
#### M2M
multipole to multipole
#### P2M M2L
particle to multipole
#### P2P
source particles particle to particle
#### L2L
local to local
#### L2P
local to particle
target particles
##### Lglobal
Llocal
Figure 2: Data-flow of FMM calculation. Data dependency is between red and blue points.
.
global cell boundaries
process boundaries
##### Global M2L
local cell boundaries Level : 0
Level : 1
Level : 2
Level : Lglobal-2 Global M2M
Level : Lglobal-1 Many process in one global cell
Level : Lglobal Many local cells in one process
Level : Lglobal+1
##### Local M2L
Level : Lglobal+Llocal-3
Level : Lglobal+Llocal-2
Level : Lglobal+Llocal-1 Local P2P
##### rank 0 rank 1
Figure 3: Splitting of the local and global tree in FMM.
|Col1|Level : Lglobal-2 Level : L-1|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|Col11|Global M2M|
|---|---|---|---|---|---|---|---|---|---|---|---|
||Level : Lglobal-1 Ma||||||||||ny process in one global cel L l P2P|
|Leve l|Level : Lglobal l : Lglobal+1 Level : Lglobal+Llocal- Level : Lglobal+L|3 local-2|Many local cells in one process Local M2 Loca|||||||||
|||||||||||||
|||||||||||||
|||||||||||||
|||||||||||||
|||||||||||||
||Level : Lglobal rank 0|+Llocal-1 rank 1||||||||||
|||||||||||||
|||||||||||||
|||||||||||||
|||||||||||||
|||||||||||||
creasingly important in HPC applications, both of
FMM itself and, for instance, of hierarchically lowrank (or “rank-structured”) matrices, which are under active development in theory and software. The
application of a model of demonstrated relevance to
one application to an entirely different application
makes a statement about the value and general applicability of the model. In this section we develop
a performance model to understand the performance
of the communication in FMM through a phase-byphase analysis based on four principal phases.
We start with a baseline model that is a combina
tion of the latency and inverse bandwidth. We subsequently refine this baseline model to reach a more
realistic model that is able to cover the relevant system architecture properties, with the exception that
overlapping communication with computation is not
considered in this work.
#### 5.1 FMM Communication Phases
As shown in Figure 3, our FMM uses a separate tree
structure for the local and global tree. In order to construct a performance model for the communication in
5
-----
(a) Level=7 (b) Level=6 (c) Level=5
(d) Level=4 (e) Level=3 (f) Level=2
Figure 4: Heat maps for level-by-level communication patterns for the M2L phase of an FMM with N=62,500
per process using 128 processes. Areas of black indicate zero messages between processes, the peak communication volume is represented in red. In this example, the switch between global and local trees is between
Level 3 and Level 4.
FMM, we estimate the amount of data that must be
sent at each level of the hierarchy. Table 1 shows the
number of cells that are sent, which correspond to
the illustrations in Figure 3. Lglobal is the depth of
the global tree, Llocal is the depth of the local tree.
We define N as the global number of particles, and P
as the number of processes (MPI ranks). The global
tree is constructed so that each MPI process is a leaf
node in the global tree. Therefore, the depth of the
global tree only depends on the number of processes
_P and not N_ . The depth of the global tree grows
with log8 P, whereas the depth of the local tree grows
with log8(N/P ). For the current calculations we are
assuming a nearly uniform particle distribution (as in
explicit solvent molecular dynamics) and therefore a
full octree structure.
**5.1.1** **Global M2L**
In Table 1 we show the number of cells to send per
level and the total amount of communication for all
levels. There are four types of communication in our
FMM, which correspond to the four stages shown
with the red hatching in Figure 3. The first is the
“Global M2L” communication, which sends 26 8
_×_
cells at each level, as shown at the top right of Figure
3. The green lines are the process boundaries and the
blue lines are the cell boundaries, which means one
FMM cell belongs to many processes in the global
tree. In order to avoid redundant communication, we
index each process that shares a global cell and perform a one-to-one communication between the processes with matching indices only. In order to further
reduce the communication, we select one process for a
group of eight cells to do the communication. There
6
-----
fore, the number of processes to communicate with
(pi) is always 26 and the number of cells to send is
always 8 for every process and for every level in the
global tree. In other words, for the “Global M2L”
communication the message size and number of sends
is constant regardless of N and P, and only the number of hops between the processes will increase depending on the network topology. On torus networks,
we map the MPI ranks to the torus and synchronize
the direction of the 26 one-to-one communications.
The communication per level is (1) and the number
_O_
of levels in the global tree is (log P ), so the total
_O_
communication complexity for this stage is (log P )
_O_
as shown in Table 1.
**5.1.2** **Global M2M**
The second type of communication is the “Global
M2M”, which sends 7 cells at each level, as shown in
Figure 3. We use a similar technique to the “Global
M2L” case to avoid redundant communication by
pairing the MPI ranks for the one-to-one communication when many processes share the same global cell.
The number of processes to communicate with is always seven and the number of cells to send is always
one for every process and for every level in the global
tree. Similar to the “Global M2L” case, only the number of hops during the one-to-one communication will
increase, and the rate depends on the network topology. The communication per level is (1) and the
_O_
number of levels is (log P ), so the total communi_O_
cation is (log P ) for the “Global M2M” stage.
_O_
**5.1.3** **Local M2L**
The third type of communication is the “Local M2L”,
which is shown in the red hatching in the second picture from the bottom on the right side of Figure 3.
The process boundaries shown in green are coarser
than the local cell boundaries shown in black, which
means that one process contains many cells, in contrast to the previous two communication types. In a
full octree structure, we know that all cells are nonempty so we simply need to send two layers of halo
cells for the M2L calculation at each level, as shown
in Figure 3. Therefore, the number of processes to
Table 1: Amount of communication in FMM
Cells to send / level Total comm.
Global M2L 26 8 (log P )
_×_ _O_
Global M2M 7 (log P )
_O_
Local M2L (2[i] + 4)[3] 8[i] ((N/P )[2][/][3])
_−_ _O_
Local P2P (2[i] + 2)[3] 8[i] ((N/P )[2][/][3])
_−_ _O_
communicate with is always the 26 neighbors, and
the number of cells to send depends on the level. At
level i of the local tree, there are 2[i] cells in each direction. Two layers of halo cells on each size will create
a volume of (2[i] +4)[3] cells, and subtracting the center
volume 8[i] will give (2[i] + 4)[3] 8[i] as shown in Table 1.
_−_
The leading term is (4[i]) since the 8[i] term cancels
_O_
out. Since the number of levels in the local tree grow
as log8(N/P ) the communication complexity for the
“Local M2L” is (4[log][8][(][N/P][ )]) = ((N/P )[2][/][3]). This
_O_ _O_
can also be understood as the surface to volume ratio of the bottom two illustrations in Figure 3. Since
_N/P is constant for weak scaling and decreases for_
strong scaling, this part does not affect the asymptotic weak/strong scalability of the FMM.
**5.1.4** **Local P2P**
The fourth type of communication in the FMM is the
“Local P2P”, which is shown in the bottom picture
on the right side of Figure 3. This communication
only happens at the bottom level of the local tree.
Similar analysis to the “Local M2L” stage shows that
(2[i] + 2)[3] 8[i] cells must be sent, as shown in Table 1.
_−_
In this case, i is exactly log8(N/P ) and we obtain
the same asymptotic amount of communication of
((N/P )[2][/][3]). Similar to the “Local M2L”, this part
_O_
does not affect the asymptotic weak/strong scalability of the FMM. However, the content of the data is
different from the previous three cases where the multipole expansion coefficients were being sent. In the
P2P communication the coordinates and the charges
of every particle that belongs to the cell must be sent.
Therefore, the asymptotic constant of (N/P )[2][/][3] is
_O_
typically much larger than that of the “Local M2L”,
and this could be the dominant part of the communication time depending on the number of particles per
leaf cell.
#### 5.2 Baseline Model ((α, β) model)
To model interprocess communication, we begin with
the basic (α, β) model, where α represents communication latency, where β is the send time per-Byte
(inverse bandwidth). Using the basic model, a message send cost can be represented as
_Tα−β = α + nβ_ (2)
where n is the number of Bytes in the message.
This basic model describes the communication over
an ideal architecture where the communication cost
does not depend on processor locations or network
traffic caused by many processors communicating at
the same time [9]. For a more realistic architecture,
7
-----
a more detailed model is needed. For this reason, we
add penalties to this basic model to take into account
machine-specific performance issues. In particular,
we consider communication distance, interconnection
switching delay, limited bandwidth, and the effect of
multiple cores on a single node contending for available resources.
#### 5.3 Distance Penalty ((α, β, γ) Model)
Following [11], we refine the assumption that distance
between processors in interconnected networks does
not have effect on communication time. To take into
account the effect of distance we refine the baseline
model according to the number of extra hops a message travels
_Tα−β−γ = α + nβ + (h −_ _hm)γ,_ (3)
where h is the number of hops a message travels, hm
is the smallest possible number of hops a message can
travel in the network, and γ is the delay per extra hop.
If there is no network contention and all messages
travel with minimum number of hops, this distance
penalty should have no effect.
#### 5.4 Bandwidth Penalty on β
The peak hardware bandwidth is rarely achieved
in message passing. Therefore, we multiply β by
_Bmax/B to incorporate the ratio between the peak_
hardware per-node bandwidth Bmax and the effective
bandwidth from the benchmark B.
_Tβ−P enalty = α + nβ [B][max]_ + (h − _hm)γ_ (4)
_B_
#### 5.5 Multicore Penalty on α or γ
Increasing the number of cores per node increases the
data traffic between nodes, and could potentially result in congestion. Furthermore, larger number of
cores per node introduces more noise caused by access to resources shared by multiple cores. To model
these effects, we multiply α and/or γ by the number
of active cores per node c. This model focuses on
the worst case behavior where a machine’s aggregate
bandwidth could be exceeded by all cores communicating simultaneously. The resulting models are
_Tα−P enalty = cα + nβ + (h −_ _hm)γ_ (5)
_Tγ−P enalty = α + nβ + c(h −_ _hm)γ_ (6)
### 6 Model Validation
#### 6.1 Machine Description
To validate our performance models we benchmark
our FMM code on four different architectures; Shaheen, Mira, Titan, and Piz Dora.
**Shaheen is 16 racks of an IBM BlueGene/P. Each**
rack contains 1024 PowerPC 450 CPUs with 4 cores
running at 850MHz with 32kB private L1 cache and
8MB shared L3 cache. Each compute node has 2GB
RAM with 13.6 GB/s memory bandwidth. The nodes
are connected by 3-D torus network with 5.1GB/s
injection bandwidth per node.
**Mira is 48 racks of an IBM BlueGene/Q. Each**
rack contains 1024 Power A2 CPUs with 16 + 1 cores
running at 1.6GHz with 16kB private L1 cache and
32MB shared L2 cache. Each compute node has 16GB
RAM with 42.6GB/s memory bandwidth. The nodes
are connected by a 5-D torus network with 20GB/s
injection bandwidth per node.
**Titan is a Cray XK7 system with 18, 688 com-**
pute nodes each equipped with an AMD Opteron 6274
CPU and NVIDIA Kepler K20X GPU. The CPU has
16 cores running at 2.2 GHz with 16 kB L1 cache, 2 4
_×_
MB L2 cache, and 8 2 MB L3 cache. The GPU has
_×_
15 64 cores running at 730 MHz with 64+48 kB L1
_×_
cache and 1.5 MB L2 cache. Each compute node has
32 GB of RAM with 51.2 memory bandwidth. The
nodes are connected by a 3-D torus with 20GB/s of
injection bandwidth per node. We do not use any of
the GPUs in the current study.
**Piz Dora is a Cray XC40 with 1256 compute**
nodes, each with two 12-core Intel Haswell CPUs
(Intel R Xeon R E5-2690 v3). Piz Dora has a total of
_⃝_ _⃝_
30144 cores (24 cores per node). Out of the total, 1192
nodes feature 64GB of RAM each, while the remaining 64 compute nodes have 128GB of RAM each (fat
nodes). The nodes are connected by a dragonfly network using the Aries interconnect where the routers
in each group are arranged as rows and columns of
a rectangle, with all-to-all links across each row and
column but not diagonally.
In order to obtain the machine parameters, the
```
b eff benchmark in the HPC Challenge suite [22] was
```
used to determine the parameters α and β. We report
the best-case latency and bandwidth measurements.
To find the parameter γ, we followed the same procedure as Gahvari et al. [11]. The machine parameters
for Shaheen, Mira, Titan, and Piz Dora are shown
in Table 2. Note that our definition of β is defined
as send time per Byte, whereas Gahvari et al. define
their β as send time per element (8 Bytes).
8
-----
Table 2: Machine parameters for latency α, inverse
bandwidth β, and distance penalty γ, on Shaheen,
Mira, Titan, and Piz Dora.
Shaheen Mira Titan Piz Dora
_α_ 4.12 µs 5.33 µs 1.67 µs 0.457 µs
_β_ 2.14 ns 1.32 ns 1.62 ns 0.4054 ns
_γ_ 29.9 ns 134 ns 284 ns 0.4838 µs
#### 6.2 Experimental Setup
We ran the FMM code for 10 steps and measured
the time spent on the communication for the “Global
M2L” and “Local M2L” phases. The results are then
divided by 10 to get the average time spent at each
level. The “Global M2M” phase was negligible and
the “Local P2P” phase only occurs at the bottom
level and is irrelevant to the scalability of the FMM,
so we do not consider these two phases in the current analysis. We used the Laplace kernel in three
dimensions with random distribution of particles in a
cube. We use periodic boundary conditions so that
there is no load imbalance at the edges of the domain.
The number of MPI processes was varied between
_P =_ 128, 1024, 8192, while the number of particles
_{_ _}_
per process was kept constant at N/P = 62, 500. On
all machines we used the maximum number of cores
on each node before increasing the number of nodes.
Timings were measured with gettimeofday() after a
```
MPI Barrier() call. We used the default rank map
```
ping to the nodes that the system provides.
Table 3 shows communication information and
statistics when running the FMM on 128, 1024, and
8192 processes. “Level” is the level within the tree
structure and goes from 0 to Lglobal +Llocal _−1, where_
_Llocal = 4 for N/P = 62, 500. Therefore, the bottom_
four levels in Table 3 (a), (b), and (c) belong to the
local tree. The depth of the global tree Lglobal is 4, 5,
and 6 for 128, 1024, and 8192 processes, respectively.
“Cells” is the total number of cells at that level of
the tree structure, which is simply 8[Level] for a full
octree. “Sends” is the number of processes to which
sends. As mentioned in Section 5.1 we have developed a communication scheme that limits the number
of sends to 26 regardless of the problem size, number
of processes, or the level. “Bytes” is the aggregate
data size that is sent by a given process at each level
of the tree. As shown in Table 1, the number of cells
for the “Global M2L” communication is 26 8. For
_×_
each cell we are sending 56 multipole expansion coefficients in single precision (4 Bytes). Therefore, the
total number of Bytes for the “Global M2L” phase is
26 8 56 4 = 46592. We can see from Table 1
_×_ _×_ _×_
that the amount of cells involved in the “Local M2L”
Table 3: Statistics of the M2L communication.
(a) 128 Processes
Level Cells Sends Bytes
0 1 0 0
1 8 0 0
2 64 26 46592
3 512 26 46592
4 4096 26 46592
5 32768 26 100352
6 262144 26 272384
7 2097152 26 874496
(b) 1024 Processes
Level Cells Sends Bytes
0 1 0 0
1 8 0 0
2 64 26 46592
3 512 26 46592
4 4096 26 46592
5 32768 26 46592
6 262144 26 100352
7 2097152 26 272384
8 16777216 26 874496
(c) 8192 Processes
Level Cells Sends Bytes
0 1 0 0
1 8 0 0
2 64 26 46592
3 512 26 46592
4 4096 26 46592
5 32768 26 46592
6 262144 26 46592
7 2097152 26 100352
8 16777216 26 272384
9 134217728 26 874496
communication can be calculated by (2[i] + 4)[3] 8[i],
_−_
where i is the level in the local tree (not the “Level”
shown in Table 3). For example, for level one in the
local tree, the amount of cells will be (2[1] + 4)[3] 8[1]
_−_
which is equivalent to 26 8. This is why the “Bytes”
_×_
is the same for the “Global M2L” and the first level
of the “Local M2L” in Table 3.
#### 6.3 Model Validation
We compare the actual communication time for the
M2L communication with our performance model on
Shaheen, Mira, Titan, and Piz Dora. We compare
against same combination of models as in the multigrid study [11]. The combinations are:
|Col1|Shaheen|Mira|Titan|Piz Dora|
|---|---|---|---|---|
|α|4.12 µs|5.33 µs|1.67 µs|0.457 µs|
|β|2.14 ns|1.32 ns|1.62 ns|0.4054 ns|
|γ|29.9 ns|134 ns|284 ns|0.4838 µs|
|Level|Cells|Sends|Bytes|
|---|---|---|---|
|0|1|0|0|
|1|8|0|0|
|2|64|26|46592|
|3|512|26|46592|
|4|4096|26|46592|
|5|32768|26|100352|
|6|262144|26|272384|
|7|2097152|26|874496|
|Level|Cells|Sends|Bytes|
|---|---|---|---|
|0|1|0|0|
|1|8|0|0|
|2|64|26|46592|
|3|512|26|46592|
|4|4096|26|46592|
|5|32768|26|46592|
|6|262144|26|100352|
|7|2097152|26|272384|
|8|16777216|26|874496|
|Level|Cells|Sends|Bytes|
|---|---|---|---|
|0|1|0|0|
|1|8|0|0|
|2|64|26|46592|
|3|512|26|46592|
|4|4096|26|46592|
|5|32768|26|46592|
|6|262144|26|46592|
|7|2097152|26|100352|
|8|16777216|26|272384|
|9|134217728|26|874496|
9
-----
1. Baseline model (α _β model)_
_−_
2. With distance penalty (α _β_ _γ model)_
_−_ _−_
3. With distance and bandwidth penalty (β
penalty)
4. With distance and bandwidth penalty, plus multicore penalty on latency (α, β penalty)
5. With distance and bandwidth penalty, plus multicore penalty on distance (β, γ penalty)
6. With distance and bandwidth penalty, plus multicore penalty on latency and distance (α, β, γ
penalty)
The results on Shaheen are shown in Figure 5. The
actual measured performance is shown as a black line,
where an error bar is drawn according to the standard
deviation in communication time among the different
MPI ranks. By comparing the Bytes in Table 3 with
the communication time in Figure 5, we see that the
deepest four levels that belong to the “Local M2L”
phase have a communication time that is proportional
to the data size being sent. The main discrepancy
in the models is caused by the β penalty, for which
the ratio between the theoretical injection bandwidth
and the b eff benchmark results is accounted for.
The actual communication time agrees well with the
models with α, β, and γ penalties.
For the shallow levels that belong to the “Global
M2L” phase, the communication time increases as the
level decreases/coarsens. Here, and in Figures 7, 8,
and 9 to follow, the “Global M2L” levels are 3 in
part (a), 3 and 4 in part (b), and 3, 4, and 5 in part
(c). The reason for the increase can be understood by
looking back at Figure 3, where the “Global M2L” is
communicating with farther processes at coarser levels of the tree. Since we are mapping the geometric
partitioning of the octree to the 3-D torus network of
Shaheen, the proximity in the octree directly translates to the proximity in the network. Therefore, even
though the data size is constant for all levels in the
“Global M2L” phase, the number of hops is larger,
which accounts for switching delays and also network
contention to some extent. This increases the communication time at coarser levels and the models that
incorporate γ are able to predict this behavior.
In Figure 6, the M2L communication time on Shaheen is plotted against the MPI rank to show the load
balance between the processes. Each color shows M2L
communication at a different level of the tree structure, and the numbers in the legend represent the levels. The communication time of each level is stacked
on top of each other so that the total hight of the area
(a) 128 processes
10 [-1]
actual
,-- Model
,---. Model
- Penalty
,, - Penalties
-, . Penalties
10 [-2],, -, . Penalties
10 [-3]
10 [-4]
2 3 4 5 6 7 8 9
Level
(b) 1024 processes
10 [-1]
actual
,-- Model
,---. Model
- Penalty
,, - Penalties
-, . Penalties
10 [-2],, -, . Penalties
10 [-3]
10 [-4]
2 3 4 5 6 7 8 9 10
Level
(c) 8192 processes
Figure 5: Performance model prediction and actual
time for M2L communication phase on Shaheen.
10
-----
|Col1|7 6 5 4 3|
|---|---|
|||
|||
|||
|||
MPIRANK
(a) 128 processes
0.012
0.01
8
0.008
7
6
0.006
5
4
0.004
3
0.002
0
200 400 600 800 1000
MPIRANK
(b) 1024 processes
0.012
0.01 9
8
0.008 7
6
0.006 5
4
0.004
3
0.002
0
2000 4000 6000 8000
MPIRANK
(c) 8192 processes
Figure 6: Load balance of M2L communication phase
on Shaheen.
plot represents the total M2L communication time
shown in Figure 5. The MPI ranks are sorted according to the total M2L communication time for better
visibility in the small differences between processes.
As can be seen from the figure, the load balance is
quite good. The imbalance seems to come from the
finest levels, which are 7, 8, and 9 for 128, 1024, and
8192 processes, respectively.
The M2L communication time on Mira is plotted
along with the six model predictions in Figure 7. Similarly to the runs on Shaheen, the main difference
in the model predictions is caused by the β penalty.
We also see a discrepancy between the model predictions with and without the α penalty for the “Global
M2L” phase (coarser levels). The multicore penalty is
very small on the Bluegene/Q. This lack of multicore
penalty has been observed in other applications where
the use of hybrid OpenMP+MPI approach did not
improve the performance over a flat MPI approach
[21]. Contrary to the runs on Shaheen, the communication time has a nearly flat profile for the “Global
M2L” phase. This is because the 5-D torus network
minimizes the number of hops and network contention
so the degradation at coarse levels of the tree is minimal. Far nodes in the octree are not so far in the
Bluegene/Q network topology.
Figure 8 shows the M2L communication time on
Titan along with the six model predictions. Similarly
to the previous two cases, the difference between the
model predictions is mainly due to the correction for
the inverse bandwidth. This difference in the theoretical injection bandwidth and measured effective
bandwidth seems to have the largest effect on all three
architectures. What is different from the previous two
cases is the large jump in the actual communication
time for the “Global M2L” phase. For example, for
the 8192 process run level 5 is taking about 10 times
more than level 6 even though the message size is
46, 592 Bytes for both cases. The γ term in the current performance models anticipates such behavior.
The error bars in the actual timings are quite large,
which indicates that there is a large load imbalance
compared to the previous two systems. The concaveconvex switch at level 5 in 8(b) is not well predicted
by the models, but the more refined models do pick it
up at level 6 in 8(c). Though a good match between
the measurements and simple models is not realized
for M2L at all granularities on Titan, performance
trends are generally well predicted.
The M2L communication time on Piz Dora is plotted along with the six model predictions in Figure 9.
In the case of 128 processes, the best fitting model
is the baseline model plus only the distance penalty.
Increasing the number of processes increases the pos
|Col1|8 7 6 5 4 3|
|---|---|
|||
|||
|||
|||
|||
|Col1|9 8 7 6 5 4 3|
|---|---|
|||
|||
|||
|||
|||
|||
11
-----
(a) 128 processes
10 [-1]
actual
,-- Model
,---. Model
- Penalty
,, - Penalties
-, . Penalties
10 [-2],, -, . Penalties
10 [-3]
10 [-4]
2 3 4 5 6 7 8 9
Level
(b) 1024 processes
10 [-1]
actual
,-- Model
,---. Model
- Penalty
,, - Penalties
-, . Penalties
10 [-2],, -, . Penalties
10 [-3]
10 [-4]
2 3 4 5 6 7 8 9 10
Level
(c) 8192 processes
Figure 7: Performance model prediction and actual
time for M2L communication phase on Mira.
(a) 128 processes
10 [-1] actual
,-- Model
,---. Model
- Penalty
,, - Penalties
-, . Penalties
10 [-2],, -, . Penalties
10 [-3]
10 [-4]
2 3 4 5 6 7 8 9
Level
(b) 1024 processes
10 [-1] actual
,-- Model
,---. Model
- Penalty
,, - Penalties
-, . Penalties
10 [-2],, -, . Penalties
10 [-3]
10 [-4]
2 3 4 5 6 7 8 9 10
Level
(c) 8192 processes
Figure 8: Performance model prediction and actual
time for M2L communication phase on Titan.
12
-----
10−1
10−2
10−3
10−4
sibility of contention and makes the model with all
penalties the best fitting model. Similar to the runs
on Titan, there is a large jump in the actual communication time for the “Global M2L” phase with
even worse load balancing suggested by the large error bars. The performance model is able to predict
the poor performance at the coarse levels.
### 7 Conclusion
2 3 4 5 6 7 8
Level
(a) 128 processes
10−1
10−2
10−3
10−4
2 3 4 5 6 7 8 9
Level
(b) 1024 processes
10−1
10−2
10−3
10−4
2 3 4 5 6 7 8 9 10
Level
(c) 8192 processes
Figure 9: Performance model prediction and actual
time for M2L communication phase on Piz Dora.
The goal of this work is to model the global communication of the FMM, to be able to anticipate challenges on future exascale machines. To improve model
fidelity, we consider penalties based on machine constraints including distance effects, reduced per core
bandwidth, and the number of cores per node. We
observe a good match between the (α, β, γ) model
with multicore penalties and the actual communication time. The discrepancy between the other models
means that all components of the model; latency (α),
bandwidth (β), hops (γ), and multicore penalty must
be taken into account when predicting the communication performance of FMM.
In our benchmark tests, we compare the performance models with measurements for the M2L communication, since this is the dominant part of the
FMM communication. Our observations are consistent with those of the studies by Gahvari et al.
[11], where the performance of an algebraic multigrid
method is analyzed using the same model. The measurements fall within the bounds of the performance
models, and match best with the model where latency,
bandwidth, hops, and multicore penalty are all taken
into account.
The present communication model is able to predict the performance on four HPC systems possessing
different characteristics. To our knowledge, this is the
first formal characterization of inter-node communication in FMM, which validates the model against actual measurements of communication time. Furthermore, the FMM implementation considered in this
paper has a provably best theoretical communication
complexity among FMM algorithms [32], so demonstrations for other implementations may be less relevant in practice.
Our current FMM code does not support asynchronous data transfer so we are not able to provide a
reference implementation for the performance model
that includes asynchronous data transfers.
The ultimate communication model is predictive
in an absolute sense; however, on complex systems,
this objective is often out of reach, or of a difficulty
out of proportion to its benefit when there exists a
13
-----
simpler model that is inexpensive and sufficient to
guide coding decisions leading to improved scaling.
The current model provides such guidance.
Looking into the future, we will most likely be
seeing more network topologies with larger diameter
(more hops). Large radix networks seem to be the
current trend, but with the exponential increase in
the node count the increase of the network diameter is unavoidable. Our communication model with
the distance penalty is able to capture the increase in
communication time at the coarse levels of the FMM
communication on Titan’s torus network. This should
allow predicting the communication bottlenecks on
future networks with larger diameter.
The performance model herein is applicable to
evolving heterogeneous systems, such as GPUs or
Xeon Phis. This is because the accelerators and coprocessors affect the per-node computation but not
the inter-node communication. Nor is the model affected by the on-node computational performance of
FMM, as long as the accelerators and coprocessors
are not using more than one MPI process, which is
the optimal way to use the current generation of such
hardware.
### Acknowledgements
We acknowledge system access and the generous assistance of the staffs at four facilities for the performance tests herein: the KAUST Supercomputing
Laboratory; the Argonne Leadership Computing Facility at Argonne National Laboratory, which is supported by the Office of Science of the U.S. Department of Energy under contract DE-AC02-06CH11357;
the Oak Ridge Leadership Computing Facility at Oak
Ridge National Laboratory, which is supported by the
Office of Science of the U.S. Department of Energy
under Contract No. DE-AC05-00OR22725; and the
Swiss National Supercomputing Centre (CSCS), under project ID g81.
### Author Biographies
_Huda Ibeid received her BSc degree in Computer En-_
gineering from the University of Jordan and is currently a PhD candidate in Computer Science at the
King Abdullah University of science and Technology
(KAUST). Her research interests include fast algorithms for particle-based simulations, fast algorithms
on parallel computers and GPUs, design of parallel numerical algorithms, parallel programming models and performance optimizations for heterogeneous
GPU-based systems.
_Rio Yokota obtained his PhD from Keio University,_
Japan, in 2009 and worked as a postdoctoral researcher with Prof. Lorena Barba at the University of
Bristol and then Boston University. He has worked
on the implementation of fast N -body algorithms on
special-purpose machines such as mdgrape-3, and
then on GPUs after CUDA was released, and on
vortex methods for fluids simulation. He joined the
King Abdullah University of Science and Technology
(KAUST) as a research scientist, where he continued
to work on fast multipole methods. He is now at the
Tokyo Institute of Technology as an Associate Professor.
_David Keyes is the director of the Extreme Com-_
puting Research Center at KAUST and an Adjoint
Professor of Applied Mathematics at Columbia University. Keyes graduated in Aerospace and Mechanical Sciences from Princeton University and earned a
doctorate in Applied Mathematics from Harvard University. He did postdoctoral work in the Computer
Science Department of Yale University. He works at
the algorithmic interface between parallel computing
and the numerical analysis of partial differential equations, across a spectrum of aerodynamic, geophysical,
and chemically reacting flows.
### References
[1] J. Barnes and P. Hut. O(N log N ) force-calculation
algorithm. Nature, 324:446–449, 1986.
[2] R. Beatson and L. Greengard. A short course on fast
multipole methods. In Wavelets, Multilevel Methods
_and Elliptic PDEs, pages 1–37. Oxford Science Pub-_
lications, 1997.
[3] A. Chandramowlishwaran, K. Madduri, and
R. Vuduc. Diagnosis, tuning, and redesign for
multicore performance: A case study of the fast
multipole method. In SC ’10 Proceedings of the
_2010 ACM/IEEE International Conference for High_
_Performance Computing, Networking, Storage and_
_Analysis, 2010._
[4] A. Chandramowlishwaran, S. Williams, L. Oliker,
I. Lashuk, G. Biros, and R. Vuduc. Optimizing and
tuning the fast multipole method for state-of-the-art
multicore architectures. In Proceeding of the Inter_national Parallel Distributed Processing Symposium_
_(IPDPS), pages 1–12, 2010._
[5] H. Cheng, L. Greengard, and V. Rokhlin. A fast
adaptive multipole algorithmin three dimensions.
_Journal of Computational Physics, 155(2):468–498,_
1999.
[6] M. J. Clement and M. J. Quinn. Symbolic performance prediction of scalable parallel programs. In
14
-----
_Proceedings of the International Parallel Processing_
_Symposium, pages 635–639, April 1995._
[7] L. DeRose and D. A. Reed. SvPablo: A multilanguage, architecture-independent performance
analysis system. In Proceeding of the International
_Conference on Parallel Processing, pages 311–318,_
Augest 1999.
[8] J. Dongarra and F. Sullivan. Guest Editors Introduction to The Top 10 Algorithms. _Computing in_
_Science and Engineering, 2:22–23, 2000._
[9] I. Foster. Designing and Building Parallel Programs.
Addison-Wesley, 1995.
[10] I. T. Foster and P. H. Worley. Parallel algorithms for
the spectral transform method. _SIAM Journal on_
_Scientific and Statistical Computing, 18(3):806–837,_
1997.
[11] H. Gahvari, A. H. Baker, M. Schulz, U. M. Yang,
K. E. Jordan, and W. Gropp. Modeling the performance of an algebraic multigrid cycle on HPC platforms. In ICS ’11 Proceedings of the International
_Conference on Supercomputing, pages 172–181, 2011._
[12] H. Gahvari, W. Gropp, K. E. Jordan, M. Schulz, and
U. M. Yang. Algebraic multigrid on a dragonfly network: First experience on a Cray XC30. In Pro_ceeding of the 5th International Workshop on Per-_
_formance Modeling, Benchmarking and Simulation_
_of High Performance Computer Systems (PMBS14),_
November 2014.
[13] N. L. Gorn and D. V. Berkov. Adaptation and performance of the fast multipole method for dipolar
systems. Journal of Magnetism and Magnetic Mate_rials, 272-276:698–700, 2004._
[14] L. Greengard, M. C. Kropinski, and A. Mayo. Integral equation methods for Stokes flow and isotropic
elasticity in the plane. _Journal of Computational_
_Physics, 125:403–414, 1996._
[15] L. Greengard and V. Rokhlin. A fast algorithm
for particle simulations. _Journal of Computational_
_Physics, 73(2):325–348, 1987._
[16] L. Greengard and Rokhlin V. On the efficient implementation of the fast multipole algorithm. Research
Report RR-602, Yale University, 1988.
[17] W. D. Gropp, D.K. Kaushik, D.E. Keyes, and B.F.
Smith. Toward realistic performance bounds for implicit CFD codes. In Proceedings of Parallel CFD’99,
pages 23–26, May 1999.
[18] P. Jetley, L. Wesolowski, F. Gioachin, L. V. Kale,
and T. R. Quinn. Scaling hierarchical N-body simulations on GPU clusters. In SC ’10 Proceedings of
_the 2010 ACM/IEEE International Conference for_
_High Performance Computing, Networking, Storage_
_and Analysis, pages 1–11, 2010._
[19] D. Kerbyson, H. Alme, A. Hoisie, F. Petrini,
A. Wasserman, and M. Gittings. Predictive performance and scalability modeling of a large-scale appli
cation. In Proceedings of the 2001 ACM/IEEE con_ference on Supercomputing, pages 1–12, 2001._
[20] I. Lashuk, A. Chandramowlishwaran, H. Langston,
T.-A. Nguyen, R. Sampath, A. Shringarpure,
R. Vuduc, L. Ying, D. Zorin, and G. Biros. A massively parallel adaptive fast multipole method on heterogeneous architectures. In Proceedings of the Con_ference on High Performance Computing Network-_
_ing, Storage and Analysis, pages 1–12, 2009._
[21] M. Lee, N. Malaya, and R. D. Moser. Petascale direct numerical simulation of turbulent channel flow
on up to 768k cores. In Proceedings of the Confer_ence on High Performance Computing Networking,_
_Storage and Analysis, Denver, CO, USA, Novermber_
16-22 2013.
[22] P. Luszczek and J. Dongarra. Introduction to the
HPC Challenge Benchmark Suite. Technical Report
ICL-UT-05-01, University of Tennessee, Knoxville,
March 2005.
[23] C. L. Mendes. Performance Scalability Prediction on
_Multicomputers._ PhD thesis, University of Illinois,
Urbana-Champaign, May 1997.
[24] C. L. Mendes and D. A. Reed. Integrated compilation and scalability analysis for parallel systems. In_ternational Conference on Parallel Architectures and_
_Compilation Techniques (PACT’98), pages 385– 392,_
October 1998.
[25] J. M. Perez-Jorda and W. Yang. On the scaling of
multipole methods for particle-paticle interactions.
_Chemical Physics Letters, 282:71–78, 1998._
[26] W. T. Rankin. Efficient Parallel Implementations of
_Multipole Based N-body Algorithm. PhD thesis, Duke_
University, 1999.
[27] A. Snavely, N. Wolter, and L. Carrington. Modeling application performance by convolving machine
signatures with application profiles. In Proceeding of
_the IEEE Workshop on Workload Characterization,_
pages 149–156, December 2001.
[28] B. Van de Wiele, F. Olyslager, and L. Dupre. Application of the fast multipole method for the evaluation of magneto-static fields in micromagnetic
computations. _Journal of Computational Physics,_
227:9913–9932, 2008.
[29] W. R. Wolf and S. K. Lele. Aeroacoustic integrals
accelerated by fast multipole method. AIAA Journal,
49(7):1466–1477, 2011.
[30] P. H. Worley. Performance evaluation of the IBM SP
and the Compaq AlphaServer SC. In Proceeding of
_the ACM International Conference of Supercomput-_
_ing 2000, pages 235–244, 2000._
[31] R. Yokota, J. Pestana, H. Ibeid, and D. E. Keyes.
Fast multipole preconditioners for sparse matrices
arising from elliptic equations. _arXiv:1308.3339v2,_
2014.
15
-----
[32] R. Yokota, G. Turkiyyah, and D. Keyes. Communication complexity of the fast multipole method and
its algebraic variants. Supercomputing Frontiers and
_Innovations, 1(1):63–84, 2014._
[33] J.-S. Zhao and W.-C. Chew. Three-dimensional multilevel fast multipole algorithm from static to electrodynamic. Microwave and Optical Technology Letters,
26(1):43–48, 2000.
16
-----
|
{
"disclaimer": "Notice: The following paper fields have been elided by the publisher: {'abstract', 'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1177/1094342016634819?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1177/1094342016634819, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "GREEN",
"url": "http://t2r2.star.titech.ac.jp/rrws/file/CTT100718789/ATD100000413/"
}
| 2,016
|
[
"JournalArticle"
] | true
| 2016-11-01T00:00:00
|
[] | 15,030
|
en
|
[
{
"category": "Medicine",
"source": "external"
},
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Medicine",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/0276e32e1440fee9e433f45192e8ccd7f31e3a56
|
[
"Medicine",
"Computer Science"
] | 0.892612
|
Platform for Efficient Switching between Multiple Devices in the Intensive Care Unit
|
0276e32e1440fee9e433f45192e8ccd7f31e3a56
|
Methods of Information in Medicine
|
[
{
"authorId": "7239058",
"name": "F. D. Backere"
},
{
"authorId": "48609947",
"name": "Thomas Vanhove"
},
{
"authorId": "2255655569",
"name": "E. Dejonghe"
},
{
"authorId": "2255655390",
"name": "M. Feys"
},
{
"authorId": "2255684665",
"name": "T. Herinckx"
},
{
"authorId": "2255618571",
"name": "J. Vankelecom"
},
{
"authorId": "145320688",
"name": "J. Decruyenaere"
},
{
"authorId": "2250414779",
"name": "F. D. Turck"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"Method Inf Med"
],
"alternate_urls": [
"http://www.schattauer.de/de/magazine/uebersicht/zeitschriften-a-z/methods.html"
],
"id": "95f5bdab-1f05-4090-899f-3869a15b5707",
"issn": "0026-1270",
"name": "Methods of Information in Medicine",
"type": "journal",
"url": "https://methods.schattauer.de/en/contents/methods-open.html"
}
| null |
# Platform for Efficient Switching between Multiple Devices in the
Intensive Care Unit
Femke De Backere[1], Thomas Vanhove[1], Emanuel Dejonghe[1], Matthias Feys[1], Tim Herinckx[1],
Jeroen Vankelecom[1], Johan Decruyenaere[2] and Filip De Turck[1]
1 Information Technology Department (INTEC), Ghent University - iMinds, Gaston Crommenlaan 8, bus 201, 9050 Ghent,
Belgium
2 Department of Intensive Care, Ghent University Hospital, De Pintelaan 185, B-9000 Gent, Belgium
# SUMMARY
_Objectives: Handheld computers, such as tablets and smartphones, are becoming more and_
more accessible in the clinical care setting and in Intensive Care Units (ICUs). By making the
most useful and appropriate data available on multiple devices and facilitate the switching
between those devices, staff members can efficiently integrate them in their workflow,
allowing for faster and more accurate decisions. This paper addresses the design of a platform
for the efficient switching between multiple devices in the ICU. The key functionalities of the
platform are the integration of the platform into the workflow of the medical staff and
providing tailored and dynamic information at the point of care.
_Methods: The platform is designed based on a 3-tier architecture with a focus on extensibility,_
scalability and an optimal user experience. After identification to a device using Near Field
Communication (NFC), the appropriate medical information will be shown on the selected
device. The visualization of the data is adapted to the type of the device. A web-centric
approach was used to enable extensibility and portability.
_Results: A prototype of the platform was thoroughly evaluated. The scalability, performance_
and user experience were evaluated. Performance tests show that the response time of the
system scales linearly with the amount of data. Measurements with up to 20 devices have
shown no performance loss due to the concurrent use of multiple devices.
_Conclusions: The platform provides a scalable and responsive solution to enable the efficient_
switching between multiple devices. Due to the web-centric approach new devices can easily
be integrated. The performance and scalability of the platform have been evaluated and it was
shown that the response time and scalability of the platform was within an acceptable range.
# MESH TERMS:
Decision Making, Computer-Assisted
Decision Support Systems, Clinical/organization & administration
Intensive Care Units
Information Systems
User-Computer Interface
# CORRESPONDENCE TO:
Femke De Backere
Department of Information Technology
Internet Based Communication Networks and Services (IBCN)
Ghent University - iMinds
Gaston Crommenlaan 8 (Bus 201), B-9050 Gent, Belgium
T: +32 9 33 14938
F: +32 9 33 14899
E: femke.debackere@intec.UGent.be
-----
# 1 INTRODUCTION
Handheld computers, such as tablets and smartphones, are becoming more and more popular,
even in the clinical care setting [1][2][3]. Moreover, with the increasing memory capabilities,
processing power and connectivity these devices can offer a portable platform for patient
management in the Intensive Care Unit (ICU) [4]. Furthermore, in a computerized ICU, a
computer is located next to every bed. Each department also has a unit PC and physicians
usually have a personal desktop and smartphone. Moreover, the number of devices on an ICU
is steadily increasing in recent years [5]. Therefore, there is a need for an efficient switching
mechanism between the different devices, used in the ICU, to ensure quality of care.
These devices have the capabilities and potential to be integrated within existing clinical
decision support systems (CDSS). CDSS are computer-driven technology solutions,
developed to provide support to physicians, nurses and patients using medical knowledge and
patient-specific information. Thus, these systems will not replace the medical staff, but will
merely give advice and guidance. This way, they are able to take all relevant data and
information into account. By filtering the information in an intelligent manner and presenting
it to the medical staff at the appropriate moment and in an intelligent way, these systems can
improve health care [6]. CDSS can be used in every aspect of the care process, from
preventive care and diagnosis to monitoring and follow up. Studies have already shown that
these systems improve quality, safety and effectiveness of medical decisions, resulting in
improved patient care, higher performance of the medical staff and more effective clinical
services [7]. Nevertheless, the uptake of CDSS is rather low and this is due to a number of
factors [8].
First, one of the main problems in the use of CDSS is the integration of applications into the
current workflow of the medical staff [9]. Kawamoto et al. [10]concluded that CDSS are more
successful when integrated into the work process of the medical staff. This also means the
integration with existing information systems of the hospital [11].
Second, the devices used in the ICU are not optimal embedded within CDSS. Sharing
information at the right time and place has a large influence on the use of these systems and
on the performance of the medical staff, moreover it is time-saving [12].
A third problem is representing all the relevant information of a specific patient. In the ICU,
up to 200,000 parameters are collected for each patient on a daily basis [13][14]. These
parameters are mainly originating from examinations and from monitoring data. Visualizing
this data in an optimal way and selecting only the most relevant information is a challenging
task [15].
Due to problems and the necessity to ensure continuity of healthcare services, improving
patient quality of life and rationalizing healthcare costs, new pervasive healthcare systems
[16] are being explored [17][18]. The research on clinical decision support has evolved over
50 years [19] resulting in new approaches such as pervasive and ubiquitous healthcare
[20][21].
As the medical staff in a clinical setting has very diverse tasks and the work is highly
fragmented [22][23]. On average, they do not spend more than 5 minutes on a specific task
-----
and in many cases only spend 1.5 minutes on a specific activity before switching to another
task. This means that personnel has to be highly adaptive and should be able to cope with an
ever changing environment and continually adjust their activities [24]. In an intensive care
setting, there is a wide variety of systems that are integrated in the workflow of the doctors
and the nurses. This means that during an activity they are taken into account the readings
and/or measurement from these systems into account to assess the situation of the patient or
they have to interact with different (software) systems to obtain the correct information about
a specific patient [25]. As the staff only has limited time to spend on certain tasks or activities,
accessing the correct infrastructure and tools can create a big overhead for the staff. The study
of Koch et al. [26] indicated that using integrated displays, where all important information in
contained in one screen, could be an advantage, if bidirectional communication between
different devices is implemented. Also, event recognition and treatment efficiency can be
improved when using a second display [27].
A better integration of the current infrastructure and handheld devices, such as tablets and
smartphones, can improve this situation. Therefore, there is a need for a platform, which is
capable of switching between the different devices, used in the ICU. However, there are still
obstacles concerning the efficient switching between devices and that should be taken into
account when developing such a platform. First, the user friendliness should be of paramount
importance. Introducing a new tool into such a complex setting as an ICU, should improve the
quality of care and the workflow of the medical staff. It should support the staff in their
current activities. Second, the speed of the switching mechanism is also important, as doctors
and nurses only spend 5 minutes on average on a task, they do not want to wait for a few
minutes while transferring the data from one device to another. Finally, the switching
mechanism should be carried out in such a manner, that it suggest when the situation is right
to switch to another device and automatically detect other devices in the vicinity. Keeping
user friendliness in mind, users should get the suggestion to switch and should not be
switched to another device automatically.
The purpose of this paper is twofold. On the one hand, the design and implementation of a
platform is presented, enabling doctors and other members from the medical staff to switch
between multiple devices. This platform is also capable of detecting which content is suitable
to be displayed on which device, e.g., text can be shown on all devices, while high-resolution
images are less suitable to be displayed on smartphones. On the other hand, the performance
of the platform is evaluated, to give valuable insights in the scalability, responsiveness and
user experience.
The remainder of the article is structured as follows. Section 2 details the objectives of our
platform, whereas Section 3 is devoted to the methodological approach. Section 4 deals with
the evaluation results. Finally, the main contributions are discussed and the main conclusions
of this research are highlighted in Sections 5 and 6.
# 2 OBJECTIVES
The aim of this research is to design a platform that allows for the efficient switching between
devices in an Intensive Care setting. This platform should offer following features:
- Integration in the workflow and at the point of care: to optimize the care and minimize
loss of time and costs, the visualization of the data by the platform should be
integrated into the workflow of the doctors and the medical staff. Moreover, the
-----
visualization should also be possible at the point of care, while examining the patients
and not only in the office of the physician.
- Tailored information: The information should be adapted dynamically to the
capabilities of the devices, e.g., high quality images, such as X-ray pictures, should not
be shown on a smartphone as the transfer of data would create an unacceptable delay
and it would not be easy to interpret this kind of images on a small screen.
- Dynamic information: When new information becomes available, the device should be
able to immediately visualize the new content.
- Displaying the appropriate medical information on the device: Based on the role and
the preferences of the end-user and the properties of the device, which are provided to
the platform by means of a database, the platform is capable of selecting and
visualizing the information in a user friendly manner. By enabling the medical staff to
enter their personal preferences, we make sure that it possible to deviate from the
settings made by the platform and we ensure a user-friendly experience. For example,
a cardiologist should see information concerning the heart instead of seeing kidney
data first.
Besides these functional requirements, the platform should also fulfill the following nonfunctional requirements:
- The platform should be generic and it should be possible to plug-in new devices at any
moment.
- As the number of devices in the ICU is increasing at a steady pace, the platform
should be scalable and able to cope with a large number of clients.
- The platform performance should be such that the loading times are in an acceptable
range.
The original research contribution of the paper is the design of a platform for the efficient
switching between devices in an Intensive Care setting, taking into account the above four
functional requirements and the three non-functional requirements. The design of the platform
is outlined in the paper and obtained performance results are presented, together with a
discussion section. The platform can also be used outside the intensive care setting, for
instance in ambulatory settings.
# 3 METHODS
The platform offers an environment, in which the efficient switching between devices is
facilitated. Section 3.1 details the general concept of the platform’s architecture, whereas
Section 3.2 describes the platform components and their interactions, by focusing on the
envisioned scenarios. Section 3.3 discusses the use of Near Field Communication as a
localization and switching standard between the devices. In Section 3.4, the components of
the platform and their interactions are described. Further implementation details, concerning
the three-tiered architecture are given in Sections 3.4.1, 3.4.2 and 3.4.3. Section 3.4.4 handles
the implementation details of the NFC communication. Finally, Section 3.5 details the
security and confidentiality techniques used within the platform.
-----
## 3.1 GENERAL
CONCEPT
Figure
1 illustrates the general concept of the platform. As can be seen in this figure, data from
a various range of sources is gathered in the Intensive Care Information System (ICIS). This
involves data from clinical observations, prescription information, monitoring parameters, lab
results as well as administrative data. Furthermore, information regarding the personal
preferences from doctors is stored in the Staff Preferences database and general knowledge is
kept in the Knowledge database, for example the capabilities of every type of device. The
information from these three databases is used for filtering and selecting the requested
information. Based on the capabilities of device and the preferences of the user, the
information can be filtered in an additional step, if necessary, and is sent to the device.
## 3.2 SCENARIO
From the general concept, as discussed in the previous section, the following scenario can be
envisioned.
1) The doctor is on his way for his round in the ICU ward and decides that he already
wants to check the last measurements of patient X. He takes his smartphone and gets a
concise overview, in the form of a table, of the patient’s status.
2) As the doctor arrives at the ICU ward, he wants to visualize the measurements onto
the bedside desktop PC. Therefore, he swipes his personal tag on the reader, attached
to the device.
3) Immediately, the patient’s data that the doctor was viewing on the smartphone is
shown on the bedside PC. As this screen has a larger size, the measurements are
shown as graphs, where possible.
4) Meanwhile, the nurse at the unit desktop PC is entering additional information about
patient X. Straightaway, all devices, currently visualizing data about patient X, will be
updated, ensuring an up to date view on patient X.
5) After a while, the doctor moves to the bed of patient Y and on the bedside PC he
visualizes this patient’s data. With this action, the smartphone application, still
visualizing the data of patient X, will refresh automatically and instantly show the data
of patient Y.
The visualization of this scenario is shown in Figure
2. Next to the floor plan, some examples
are given of which information is displayed on the screen of the involved devices.
The scenario, described in the previous paragraph, details the general concept of the proposed
platform. However, the platform makes it possible to switch between a wide range of different
devices that could be used in the ICU: smartphones, tablets, desktop computers at the nurse’s
station or in the doctor’s office, bedside pcs and smart TVs. In fact, all devices, which are
capable of visualizing web pages can be used with the platform. Different uses cases for
switching are:
- Switching between devices, from a device with a small screen to one with a bigger
screen, because the medical staff wants to have a more detailed overview of certain
variables. This can be done by means of a graph instead of a table or a listing of the
variables from the last hour.
- Switching from a screen that can be seen by visitors of patients, during visiting hours,
to a more personal device. This way, patient confidentiality can be taken into account.
Femke D
**Verwijd**
Femke D
**Verwijd**
-----
- Switching from a device, residing next to the patients bed, to a more personal device,
because the doctor is continuing his/her round.
## 3.3 NEAR
FIELD
COMMUNICATION
To implement the tags and readers, as mentioned in the previous section, we assume doctors
and the medical staff will use Near Field Communication (NFC)[28]. NFC is a new set of
standards that enables smartphones and other devices, with similar capabilities, to establish
radio communication. This connection is set up by bringing the devices in close proximity to
each other (usually only a few centimeters), or by touching each other. Not only
communication between two NFC enabled devices is possible, but also the communication
between an NFC reader and an unpowered NFC chip, which is often called a tag.
NFC is compatible with current existing Radio Frequency Identification (RFID) structures,
tags and smart cards [29][30]. There also is no technical barrier to use NFC, as the concept is
straightforward. The user just has to bring the two devices in their range to start
communication. As the communication range is short, it is easy to distinguish multiple
devices residing in each other’s neighborhood. This also means that there is little change that
there will be security issues, if no other device is in the vicinity, there will be no
communication [31].
3.4 PLATFORM
COMPONENTS
AND
INTERACTIONS
The platform is implemented by using Java EE 6 (Java Enterprise Edition 6), which defines a
standard for developing and implementing multi-tier applications, based on standardized
modular components. The Java EE framework offers a complete set of services to these
components and details concerning middleware activities are handled automatically, without
complex programming. A multi-tier, distributed application model is used by this platform.
Based on functionality, the application is split up into different components, which can be
installed on different machines, depending on the tier they belong to. Most Java EE enterprise
applications can be split into 3 tiers:
- Entities are contained in the _Persistence Tier. The Java Persistence API is used to_
implement entities and to persist them into a table in a relational database.
- Enterprise Java Beans (EJB) are defined in the _Business Tier. These beans are_
responsible for adding logic to the application. The Java EE framework ensures that
these EJBs offer scalability by means of resource pooling. There are two different
types of EJBs. Tasks of clients are performed by session beans. Based on the
requirements, session beans can be stateless, stateful or a singleton. Message-driven
beans are used when the application has to process asynchronous messages. In the
Business Tier, Web Services can be defined, which can call upon external services.
- Java Server Pages (JSP) and servlets are stored in the Web Tier. JSP and servlets can
be used to visualize dynamic web content and make it possible to enforce a separation
between the representation of data and the business logic.
The 3-tier architecture of the platform, as depicted in Figure
3, is based on a web centric
approach and addresses the functional needs as mentioned in Section 2. These choices ensure
that the platform is flexible and portable. Also, by choosing a web centric approach, all
Femke D
**Verwijd**
-----
devices with a web browser are able to plug into the system. The implementation code of the
Proof of Concept is made publically available on GitHub through the following URL:
https://github.ugent.be/fddbacke/DeviceSwitching.git.
3.4.1 PERSISTENCE
TIER
The Persistence Tier contains all the entities, representing the tables in the database. In fact,
entities are Plain Old Java Object (POJO), extended with annotations that can indicate for
example, an ID or multiplicities, such as a many-to-many relationship. The most important
entities in the platform are:
- _DeviceType: information about the specifications of a device type, e.g., the resolution_
of the screen.
- _Staff: knowledge about staff members’ preferences and limitations, concerning the_
data that a certain staff member can consult.
- _Variable: identifier to determine which type of data is stored, for example, body_
temperature, blood pressure, heartbeat per minute.
- _Patient: detailed information about the patient, such as name, episode number and_
unique national number.
`o` _PatientVariable: actual data about a specific_ _Variable, linked to the_ _Patient._
For example: Patient x has a body temperature of 37.2 degrees.
3.4.2 BUSINESS
TIER
In the Business Tier, the Facade design pattern is applied [32]. This implies that all
communication between the Web Tier and Business Tier will pass through a specific bean, the
_ManagementBean. This bean provides only high-level business methods in order to have a_
safe and simple interface. Furthermore, the methods in the ManagementBean can be split in
two types. First, there are methods that are related to the state of the application for a current
user. Second, there are also methods for retrieving and changing information related to the
patients. Since this defines a clear distinction, the _ManagementBean is connected to two_
different internal components.
The first component, the _StateManager, keeps track of the current run-time state of the_
application. This class is implemented as a Singleton. A Singleton session bean is instantiated
only once for each application and will exists for the whole lifecycle of that application. The
Singleton session bean was chosen since this bean has to keep the information about the
global state of the application. Information about the current number of devices and the users
logged in on these devices is not stored in the database, but is all stored inside the
_StateManager._
Run-time state information is not stored in the database as this would result in an extra delay.
The disadvantage of this decision is that information about which patient the user was
viewing and which devices were in use by the user, will be lost when the server has to reboot.
The second component, the DataBean, is used to communicate with the Persistence Tier. This
way, the application can take care of the data related to patients and staff members. Because
all global state-information is saved in the _StateManager, the_ _ManageBean can be_
implemented as a stateless session bean. This design choice makes the application more
scalable, since there can be more instances of this implementation available at the same time.
-----
3.4.3 WEB
TIER
The Web Tier consists of both code running on the device (the client) and code running on the
server.
The Asynchronous Javascript and XML (AJAX) design principle is used to create a fast and
responsive system [33]. However, Javascript Object Notation (JSON) was used instead of
XML to enable easier processing in the client [34]. A static HTML page (HyperText Markup
Language) is downloaded to the client and the content of this page is changed dynamically.
Communication with the server happens in the background, without any user intervention.
The server-side code consists of servlets running in the Java EE Web Container. These will
handle the requests from the client and forward them to the Business Tier. Another function
of the servlets is to convert the raw data from the Business Tier into a representation that the
client can process. This conversion depends on the type of device. For example, a graph will
be generated from the raw data for tablets and laptops, while smartphones will receive a
textual representation.
The data is presented to the user in the form of `Blocks. Each` `Block contains a specific`
_Variable, for example blood pressure or body temperature. When a user adds new information_
to a `Block, this information is sent to the Business Tier and all devices displaying this`
information will receive the updated content.
Clients also poll the server at fixed intervals for new content. To lower the load on the
Business Tier, a ChangeTracker object is introduced. The goal of this object is to keep track
of which devices need to be updated with new content. When a client checks whether there is
new content, the request will only be passed on to the Business Tier, if the _ChangeTracker_
indicates that new content is available. This reduces the load on the server.
3.4.4 NFC COMMUNICATION
IMPLEMENTATION
In order to establish a connection between the NFC infrastructure, used for the platform and
the platform itself, the IOTOPE[1] library is used. By developing a small client for the device,
equipped with an NFC reader, the platform can be notified, when a certain device is in the
vicinity of an NFC tag. Each NFC reader an decode tags. Whenever this situation occurs, the
client will perform a post to the NFC login server by transferring all the necessary data. This
is in fact standard IOTOPE functionality. This login server will send the request to the login
server of our platform and thus establish the connection.
The NFC communication within this platform is implemented in such a way that there is
support for two different setups. In the first setup, each user has a tag and a reader is
connected to each device. In the second setup, each device has a tag and a user can log in by
scanning the tag with his/her personal smartphone.
The implementation of this small client is also available on
https://github.ugent.be/fddbacke/DeviceSwitching.git.
## 3.5 SECURITY
AND
CONFIDENTIALITY
1 https://github.com/alexvanboxel/iotope-node/downloads
-----
In a clinical setting, security and privacy play an important role, as it is not the intention that
everybody can gain access to the information used within the platform. Therefore, some
security and privacy measurements are taking within the platform. The web application of the
platform, running on the handheld device, is facilitated by means of a web page. Thus, it is
possible to use existing, standard security methods for website. A secure Transport Layer
Security[2] (TLS) connection is used. For the web application, this means that the Hyper Text
Transfer Protocol (Secure) (HTTPS) is utilized. To prevent not-authorized users, accessing
the application, a login mechanism is implemented. This way, only personnel of the ICU can
gain access to the web application. Data can only be fetched from the databases of the ICU, if
the correct WiFi hotspot is used. No information is cached in the handheld device. When the
medical staff moves to another patient, the data is automatically forgotten. When a nurse or
doctor forgets to leave the handheld device at the ICU, the web application will notice this
when they leave the hospital, by means of GPS tracking, which is only possible outside
buildings, and the web application will be closed.
# 4 RESULTS
Measurements were performed to evaluate the performance of the platform. The results of
these tests give valuable information about the scalability, responsiveness and the user
experience.
## 4.1 EVALUATION
APPROACH
Time measurements were performed to benchmark the platform. Therefore, timestamps were
collected both at the client-side and server-side. This enables the calculation of the response
time under different circumstances and the determination of the most time consuming parts of
the application. Special care has been taken to assure that a possible mismatch between the
time on the client and the time on the server is excluded from the measurement results.
All evaluation tests were repeated 30 times. These results were then averaged to exclude
statistical fluctuations. Special code was included in the client to facilitate these
measurements. This code simulates user interaction and collects timestamps belonging to the
performed action. Reaction time of the NFC reader and the time to render the Document
Object Model (DOM) in the client browser are excluded from the measurements.
Response time will always be used as benchmark. This is defined as the time between the
selection of the patient and the time when the DOM with the patient’s information is updated
in the client.
## 4.2 EVALUATION
SETUP
All the tests were performed with the same server: a laptop with an Intel Core i5-2410M CPU
with 6 GB RAM and a SSD running Microsoft Windows 7. Glassfish Version 3.1.2 was used
as the Application Server. The client used in the tests, unless stated otherwise, is a laptop with
an Intel Core i5-3210M running Ubuntu 12.04. Firefox 17.0 was used as web browser. For the
communication between the devices a wireless 802.11 b/g router with 100Mbit/s Ethernet
connection was used. Unless stated otherwise, the devices were connected to the WiFi
network.
2 http://tools.ietf.org/html/rfc5246
-----
A Samsung Galaxy SII Smartphone, running Android 4.0 and a Samsung Galaxy Tab 10.1
tablet, running Android 3.1 were used as mobile devices. The tests were performed using their
standard web browsers.
## 4.3 EVALUATION
RESULTS
The first evaluation analyzes the response time in function of the number of blocks that are
sent to the client. This test was performed with a wired connection between server and client.
As can be seen in Figure
4, the relation between the response time and the number of blocks
shown in the client is linear. Presenting the data as graphs takes more time than presenting the
data in a tabular form. The response time increases also faster when graphs are generated.
This was expected since the number of blocks equals the number of graphs to be created.
Also, the time for generating three blocks, is almost equal for both representations. This is
because the first three blocks never contain any graphs.
To gain more insights in the response time, the different parts were analyzed. Three different
parts were identified:
- DOM-modification time at the client-side
- The communication delay between server and client. This delay consists of both the
network delay and the JSON-parsing at the client-side.
- The generation at the server-side, this can further be split in:
`o` Switch and select patient at the server-side
`o` Business logic without database interaction
`o` Interaction with the database
`o` Generating HTML
The result of this analysis with the data represented as graphs, or as tables can be found in
Figure
5 and Figure
6 respectively. The average and standard deviation are shown in Table
1. It
can be concluded that for the rendering of tables the most time-consuming parts are the DOM
modification at the client-side and the database interaction at the server-side. For the
rendering of graphs the HTML-generation at the server-side takes a considerable amount of
time. This is caused by the fact that the generation of the graphs is done as part of the
generation of the HTML code. At the server-side, the database interaction always takes a
considerable part of the total response time.
The average response times were measured when different devices were connected with each
device constantly sending requests to the server. With up to 20 devices connected, no
statistically significant differences in response time were measured, implying that the system
scales well.
Finally, the response time was measured on different types of devices to evaluate the user
experience. In normal operation, the system will show a different representation of the data
depending on the type of device. For these measurements, the same data was sent to each
device to enable comparison of the performance of the different devices. This data consists of
25 blocks, each containing 100 values and these are represented as a graph and as a table. The
results are shown in Table
2. On all devices the representation with graphs consistently takes
longer. The smartphone also performs better than the tablet for both the representation with
graphs and tables.
Femke D
**Verwijd**
Femke D
**Verwijd**
Femke D
**Verwijd**
Femke D
**Verwijd**
Femke D
**Verwijd**
-----
# 5 DISCUSSION
An accurate analysis of the evaluation results, obtained as described in the previous section,
indicates that the platform performs as desired. The response time for generating tables, as
shown in Figure
6 shows that most of the time goes to client DOM modification (57%) and
database interaction (31%). For generating graphs, as shown in Figure
5, 65% of the response
time is needed for HTML generation. Client DOM modification and database interaction take
respectively 12% and 16% of the response time.
Measurements with up to 20 devices have shown no performance loss due to the concurrent
use of multiple devices, proving that the application is scalable. The ICU of Ghent University[3]
is one of the largest ICUs in Belgium, which holds 56 beds. The department consists of 5
different units (cardiac, burn unit, surgical, pediatric and internal) and these are located in
different locations. Each unit in itself is again divided in several smaller units where each ICU
bed has a bedside PC. As the platform can be distributed in such a way that each small unit
has its own server, it can be ensured that the platform keeps running smoothly. Moreover, in
the experiments where 25 medical parameters were measured and 100 values per parameter
were stored, the system was able to consistently respond in less than 1 second when the data
is presented in tabular form.
The difference in response time between the different types of devices is less than a factor 3,
as can be observed in Table
2. This difference can be partly compensated by adapting the
representation of the data to the device type, which guarantees a consistent user experience on
all devices. Furthermore, the performance difference between the tablet and smartphone, used
in this experimental set-up, can most probably be explained by the newer Android version and
the faster CPU of the smartphone.
When the response time as a function of the number of the data points is measured, as shown
in Figure
4 and Table
1, a linear relationship is achieved as expected. The HTML generation
takes 50 times as much time than presenting the data in a tabular form. The increase in
response time as a function of the number of data points is smaller with graphs, since the
number of graphs that needs to be created is the same.
The main advantages of using the new NFC technology, instead of the more known RFID
technology are: (i) the capability of bi-directional communication, this way of
communication, instead of single mode communication as with RFID, allows for more
flexibility as the tags are able to communicate directly with each other, (ii) the ability to
emulate contactless smart cards, which advances the interoperability of NFC as there is no
need for an NFC tag or RFID card and information, stored on the NFC device, is used for
communication, however, making the system secure is a great challenge [35](iii) peer-to-peer
connections can be established, with this mode it is possible to exchange data at link-level
[36][37], and (iv) the speed of the connection establishment is negligible, NFC connections
are typically set up in less than 100 milliseconds as the connection between 2 devices is
created automatically[38].
The platform performance can be further improved by reducing the number of requests from
the client to the server: this can be realized by changing the polling architecture (from client
to server) to a push-based architecture (from server to client). In this way, when the server
discovers a change to the data of a viewer, it will push these changes to the client.
3 http://www.icu.be/eng/
Femke D
**Verwijd**
Femke D
**Verwijd**
Femke D
**Verwijd**
Femke D
**Verwijd**
Femke D
**Verwijd**
-----
The following technical considerations and challenges should be taken into account for the
platform. First, as the platform is used within a clinical setting, privacy and security are very
important. Therefore, we also plan to carefully consider the privacy and security requirements
in the extended version of the platforms. Second, Javascript was chosen to create a fast and
responsive system, and the execution of the code is done at client side, which limits the
necessary processing power in the back-end. However, this also means that the CPU of the
end user’s device is used, which can have an impact on the battery consumption of mobile
client. An adaptive approach to balance components between client server is currently being
considered. Another limitation of Javascript is that different layout engines will render the
code in a different manner, which may result in inconsistencies in terms of functionality and
interface. Proper front-end development tools and extensive automated software testing will
allow circumventing these incompatibility concerns. Third, by using AJAX, network latency
can impact the responsiveness of the platform. Lightweight alternatives are currently being
studied.
The proposed platform can be integrated within existing CDSS that are already deployed in
the intensive care unit, as previous research indicates that stand-alone CDSS are not enabled
to be executed on multiple computing platforms [39]. This can be done by generating the
entities, based on the relational databases of the ICU. As indicated in Figure
1, several
databases are integrated within our platform. If these databases are replaced with those used in
CDSS and the queries are adjusted accordingly, the platform can be fully operational in the
ICU, integrated with the other tools and systems. EHR (Electronic Health Record)
applications are considered to be an important part of CDSS, hence integrating the proposed
platform with existing EHR applications can be realized in a similar way: generating the
entities in the platform based on the database tables in the EHR application. This approach
was taken for the integration of the system with the EHR application in the Intensive Care
department of Ghent University hospital.
# 6 CONCLUSIONS
In this paper, a platform to access data through multiple devices is described. Based on the
functional and non-functional requirements, a 3-tier architecture was designed and
implemented. Due to the web-centric approach, the platform is portable, scalable and
extensible. Extensive timing measurements were performed to investigate the response time,
the scalability and user experience of the designed platform. The results of these evaluations
show that the response time of the platform scales linearly with the amount of data. The
response time for generating tables is always less than the response time for generating
graphs. The platform, presented in this paper, facilitates the use of multiple devices in an ICU
setting, which is integrated in the workflow of the doctors and the medical staff at the point of
care. Future research will focus on replacing the polling architecture into a push mechanism
and on the implementation of several caching strategies.
# CONFLICT OF INTEREST
The authors declare that they have no conflict of interest.
Femke D
**Verwijd**
-----
# REFERENCES
[1] Lapinsky, S.. Mobile computing in critical care. Journal of Critical Care 2007;22(1):41–44.
[2] Berger, E.. The iPad: Gadget or medical godsend? Annals of Emergency Medicine 2010;56(1):A21–
A22.
[3] Kubben, P.. Neurosurgical apps for iPhone, iPod touch, iPad and Android. Surgical Neurology
International 2010;1(1):89+.
[4] Lapinsky, P., Weshler, J., Mehta, S., Varkul, M., Hallett, D., Stewart, T.. Handheld computers in
Critical Care. Journal of Critical Care 2001;5:227+.
[5] Colpaert, K., Vanbelleghem, S., Danneels, C., Benoit, D., Steurbaut, K., Van Hoecke, S., et al. Has
information technology finally been adopted in Flemish Intensive Care Units? BMC Medical
Informatics and Decision Making 2010;10(62).
[6] Osheroff, J., Teich, J., Middleton, B., Steen, E., Wright, A., Detmer, D.. A roadmap for National
Action on Clinical Decision Support. JAMIA 2007;14(2):141–145.
[7] Garg, A., Adhikari, N., McDonald, H., Rosas-Arellano, P., Devereaux, P., Beyene, J., et al. Effects of
Computerized Clinical Decision Support Systems on Practitioner Performance and Patient Outcomes.
JAMA 2005;293(10):1223–1238.
[8] Moxey, A., Robertson, J., Newby, D., Hains, I., Williamson, M., Pearson, S.. Computerized Clinical
Decision Support for Prescribing: Provision does not guarantee Uptake. JAMIA 2010;17(1):25–33.
[9] Orwat, C., Graefe, A., Faulwasser, T.. Towards Pervasive Computing in Health Care - A Literature
Review. Social Science Research Network Working Paper Series 2008;8(26).
[10] Kawamoto, K., Lobach, D.. Clinical Decision Support provided within Physician Order Entry
Systems: A Systematic Review of Features Effective for changing Clinician Behavior. In: AMIA
Annl Sym Proc. 2003, p. 361–365.
[11] Decruyenaere, J., DeTurck, F., Vanhastel, S., Vandermeulen, F., Demeester, P., De Moor, G.. On the
design of a generic and scalable multilayer software architecture for data flow management in the
Intensive Care Unit. Methods of information in medicine 2003;42(1):79–88.
[12] Kawamoto, K., Houlihan, C., Balas, A., Lobach, D.. Improving Clinical Practice using Clinical
Decision Support Systems: A Systematic Review of Trials to identify Features Critical to Success.
BMJ (Clinical research ed) 2005;330(7494):765+.
[13] Van Hoecke, S., Decruyenaere, J., Danneels, C., Taveirne, K., Colpaert, K., Hoste, E., et al. Service
oriented subscription management of medical decision data in the Intensive Care Unit. Methods of
Information in Medicine 2008;47(4):364–380.
[14] Herasevich, V., Pickering, B., Dong, Y., Peters, S., Gajic, O.. Informatics Infrastructure for Syndrome
Surveillance, Decision Support, Reporting, and Modeling of Critical Illness. Mayo Clinic Proc
2010;85(3):247–254.
[15] Sittig, D., Wright, A., Osheroff, J., Middleton, B., Teich, J., Ash, J., et al. Grand Challenges in
Clinical Decision Support. J Biomed Inform 2008;41(2):387–392.
[16] Bardram, J.. Pervasive Healthcare as a Scientific Discipline. Methods of Information in Medicine
2008;47(3):178–185.
[17] Triantafyllidis, A., Koutkias, K., Chouvarda, I., Maglaveras, N.. An Open and Reconfigurable
Wireless Sensor Network for Pervasive Health Monitoring. Methods of Information in Medicine
2008;47(3):229–234.
[18] Blobel, B.. Architectural Approach to eHealth for Enabling Paradigm Changes in Health. Methods of
Information in Medicine 2010;49(2):123– 134.
[19] Mitchell, J., Gerdin, U., Lindberg, D., Lovis, C., Martin-Sanchez, F., Miller, R., et al. 50 years of
informatics research on decision support: What’s next. Methods of Information in Medicine
2011;50(6):525.
[20] Peek, N., Swift, S.. Intelligent data analysis for knowledge discovery, patient monitoring and quality
assessment. Methods of Information in Medicine 2012;51(4):318.
[21] Surján, G., et al. How to use health informatics to manage the information overflow created by itself?
Methods Inf Med 2013;52:97–98.
[22] Tentori, M., Favela, J.. Activity-aware computing for healthcare. IEEE Pervasive Computing
2008;7(2):51–57.
[23] Tentori, M., Hayes, G.R., Reddy, M.. Pervasive computing for hospital, chronic, and preventive care.
Foundations and Trends in Human- Computer Interaction 2012;5(1):1–95.
[24] Bardram, J.E., Bossen, C.. Mobility work: The spatial dimension of collaboration at a hospital.
Computer Supported Cooperative Work (CSCW) 2005;14(2):131–160.
[25] Bardram, E.. The trouble with login: on usability and computer security in ubiquitous computing.
Personal and Ubiquitous Computing 2005;9(6):357–367.
-----
[26] Koch, S.H., Weir, C., Westenskow, D., Gondan, M., Agutter, J., Haar, M., et al. Evaluation of the
effect of information integration in displays for icu nurses on situation awareness and task completion
time: A prospective randomized controlled study. International Journal of Medical Informatics
2013;82(8):665–675.
[27] Effken, J.A., Loeb, R.G., Kang, Y., Lin, Z.C.. Clinical information displays to improve ICU
outcomes. International journal of Medical Informatics 2008;77(11):765–777.
[28] Falke, O., Rukzio, E., Dietz, U., Holleis, P., Schmidt, A.. Mobile Services for Near Field
Communication. Tech. Rep. LMU-MI-2007-1; Vodafone Group Research and Development, Munich,
Embedded Interaction Research Group, University of Munic, Computing Department, Lancaster
University, UK, Fraunhofer IAIS, Sankt Augustin and b-it, University of Bonn; 2007.
[29] Ok,
K.,
Coskun,
V.,
Aydin,
M.N.,
Ozdenizci,
B..
Current
benefits
and
future
directions
of
NFC
services.
In:
Education
and
Management
Technology
(ICEMT),
2010
International
Conference
on.
IEEE;
2010,
p.
334–338.
[30] NFC-‐Forum,
Available:
http://www.nfc-‐forum.org.
[31] Csapodi,
M.,
Nagy,
A..
New
applications
for
nfc
devices.
In:
Mobile
and
Wireless
Communications
Summit, 2007. 16th IST. IEEE; 2007, p. 1–5.
[32] Gamma, E., Helm, R., Johnson, R., Vlissides, J.. Design Patterns: Elements of Reusable Object
Oriented Software. Addison-Wesley; 1994.
[33] Paulson, L.. Building Rich Web Applications with AJAX. Computer 2005;38(10):14–17.
[34] Wang, G.. Improving Data Transmission in Web Applications via the Translation between XML and
JSON. In: 2011 Third International Conference on Communications and Mobile Computing. IEEE;
2011, p. 182–185.
[35] Roland, M. "Software card emulation in NFC-enabled mobile phones: great advantage or security
nightmare." Fourth International Workshop on Security and Privacy in Spontaneous Interaction and
Mobile Phone Use. 2012.
[36] Madlmayr, G., Langer, J., Kantner, C. Scharinger, J.. “NFC devices: Security and privacy.”
In: Availability, Reliability and Security, 2008. ARES 08. Third International Conference on. IEEE,
2008. p. 642-647.
[37] Ok, K., Coskun, V., Aydin, M. N. Ozdenizci, B “Current benefits and future directions of NFC
services.” In:Education and Management Technology (ICEMT), 2010 International Conference on.
IEEE, 2010. p. 334-338.
[38] Want, R.. Near Field Communication. IEEE Pervasive Computing 2011;10(3):4–7.
[39] Farion, K., Michalowski, W., Wilk, S., OSullivan, D., Rubin, S., Weiss, D.. Clinical decision support
system for point of care use. Meth- ods Inf Med 2009;48:381–390.
-----
# FIGURES
**FIGURE
1
–
GENERAL
CONCEPT
OF
THE
PLATFORM.
DATA
FROM
A
WIDE
RANGE
OF
SOURCES
IS
GATHERED
IN
THE**
**INTENSIVE
CARE
INFORMATION
SYSTEM
(ICIS).
PERSONAL
PREFERENCES
OF
THE
STAFF
ARE
STORED
IN
THE
STAFF**
**PREFERENCES
DATABASE
AND
GENERAL
KNOWLEDGE
IS
KEPT
IN
THE
KNOWLEDGE
DATABASE.
INFORMATION
OF
THESE**
**DATA
SOURCES
IS
USED
TO
SELECT
AND
FILTER
THE
DATA
INTELLIGENTLY.
BASED
ON
THE
TYPE
OF
THE
DEVICE
AND
THE**
**PREFERENCES
OF
THE
USER
THE
INFORMATION
IS
VISUALIZED
ON
A
SPECIFIC
DEVICE.**
-----
**FIGURE
2
–
AN
ILLUSTRATIVE
SCENARIO
TO
SHOW
THE
PLATFORM:
ON
THE
LEFT
SIDE
OF
THE
FIGURE
A
PART
OF
THE
ICU**
**FLOOR
PLAN
IS
SHOWN,
ON
THE
RIGHT
SIDE
AN
EXAMPLE
OF
THE
VISUALISATION
ON
EACH
OF
THE
DEVICES,
USED
IN
THE**
**SCENARIO
IS
DISPLAYED.
1)
THE
DOCTOR
IS
ON
HIS
WAY
TO
THE
ICU
WARD
AND
ALREADY
GOES
THROUGH
THE
PATIENT’S**
**DATA
ON
HIS
SMARTPHONE,
BY
MEANS
OF
A
TABLE.
2)
WHEN
HE
ARRIVES
AT
THE
PATIENT’S
BEDSIDE
PC,
HE
USES
HIS**
**PERSONAL
TAG
TO
IDENTIFY
HIMSELF
TO
THE
COMPUTER.
3)
ON
THIS
SCREEN,
HE
SEES
A
MORE
DETAILED
OVERVIEW
OF**
**THE
DATA
SHAPED
AS
A
GRAPH.
4)
WHEN
THE
NURSE
ENTERS
NEW
DATA
OF
THE
PATIENT
INTO
THE
SYSTEM,
THE
NEW**
**INFORMATION
IS
IMMEDIATELY
VISUALIZED
ON
THE
SCREEN
OF
THE
PC
AND
SMARTPHONE
OF
THE
DOCTOR.
5)
WHEN**
**VISITING
A
NEXT
PATIENT,
THE
DOCTOR
CHANGES
PATIENT’S
ON
THE
APPLICATION
AND
ALL
SCREENS
ARE
AGAIN**
**UPDATED.**
**FIGURE
3
-‐
HIGH
LEVEL
OVERVIEW
OF
THE
PLATFORM'S
ARCHITECTURE.
THE
SERVLETS
COMMUNICATE
WITH
THE**
**BUSINESS
TIER
AND
CONVERT
RAW
DATA
INTO
A
SUITABLE
REPRESENTATION
FOR
THE
DEVICE.
THE
MANAGEMENTBEAN**
**IS
CONNECTED
TO
TWO
INTERNAL
COMPONENTS.
THE
STATEMANAGER
KEEPS
TRACK
OF
THE
CORRECT
RUN-‐TIME
STATE**
**AND
THE
DATABEAN
IS
USED
TO
COMMUNICATIE
WITH
THE
PERSISTENCE
TIER.**
-----
**FIGURE
4
–
RESPONSE
TIME
AS
A
FUNCTION
OF
THE
NUMBER
OF
BLOCKS
WITH
THE
DATA
REPRESENTED
AS
TABLES
OR
AS**
**GRAPHS.**
**FIGURE
5
-‐
ANALYSIS
OF
RESPONSE
TIME
WITH
DATA
REPRESENTED
IN
GRAPHS**
-----
**FIGURE
6
-‐
ANALYSIS
OF
RESPONSE
TIME
WITH
DATA
REPRESENTED
AS
TABLES**
-----
# TABLES
**TABLE
1
-‐
AVERAGE
AND
STANDARD
DEVIATION
IN
MS
FOR
THE
DIFFERENT
PARTS
OF
THE
RESPONSE
TIME**
**Graph**
**Average
[ms]
σ [ms]**
**Table**
**Average
[ms]
σ [ms]**
**Select
patient** 11.07 2.49 11.3 1.62
**Client
DOM
modification** 276.4 8.02 645.47 14.78
**Communication
delay** 122.93 17.97 84.2 12.66
**HTML
generation** 1435.7 174.88 26.03 14.30
**Business
logic
without**
**DB
communication**
10.8 10.08 10.9 12.47
**Database
interaction** 363.07 176.08 348.37 170.41
**TABLE
2
-‐
AVERAGE
AND
STANDARD
DEVIATION
IN
MS
CORRESPONDING
TO
THE
RESPONSE
TIME
FOR
THE
DIFFERENT**
**TYPES
OF
DEVICES**
**Graph**
**Average
[ms]
σ [ms]**
**Table**
**Average
[ms]
σ [ms]**
**Computer** 1126.27 184.40 2219.97 305.99
**Tablet** 2798.07 304.80 3652.93 537.28
**Smartphone** 2449.13 289.17 3218.07 348.94
-----
|
{
"disclaimer": "Notice: The following paper fields have been elided by the publisher: {'abstract', 'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.3414/ME13-02-0021?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3414/ME13-02-0021, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "other-oa",
"status": "GREEN",
"url": "https://biblio.ugent.be/publication/5968043/file/5968054.pdf"
}
| 2,014
|
[
"JournalArticle"
] | true
| 2014-06-06T00:00:00
|
[] | 12,508
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Biology",
"source": "s2-fos-model"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/0277681acf245005906e17d6996ed498421c5b68
|
[
"Computer Science"
] | 0.908015
|
Data-aware optimization of bioinformatics workflows in hybrid clouds
|
0277681acf245005906e17d6996ed498421c5b68
|
Journal of Big Data
|
[
{
"authorId": "3451478",
"name": "Athanassios M. Kintsakis"
},
{
"authorId": "71018930",
"name": "Fotis Psomopoulos"
},
{
"authorId": "143619722",
"name": "P. Mitkas"
}
] |
{
"alternate_issns": [
"2579-0048"
],
"alternate_names": [
"J Big Data",
"Journal on Big Data"
],
"alternate_urls": [
"http://www.springer.com/computer/database+management+&+information+retrieval/journal/40537",
"http://techscience.com/JBD/index.html",
"https://journalofbigdata.springeropen.com",
"https://journalofbigdata.springeropen.com/"
],
"id": "d60da343-ab92-4310-b3d7-2c0860287a9d",
"issn": "2196-1115",
"name": "Journal of Big Data",
"type": "journal",
"url": "http://www.journalofbigdata.com/"
}
|
Life Sciences have been established and widely accepted as a foremost Big Data discipline; as such they are a constant source of the most computationally challenging problems. In order to provide efficient solutions, the community is turning towards scalable approaches such as the utilization of cloud resources in addition to any existing local computational infrastructures. Although bioinformatics workflows are generally amenable to parallelization, the challenges involved are however not only computationally, but also data intensive. In this paper we propose a data management methodology for achieving parallelism in bioinformatics workflows, while simultaneously minimizing data-interdependent file transfers. We combine our methodology with a novel two-stage scheduling approach capable of performing load estimation and balancing across and within heterogeneous distributed computational resources. Beyond an exhaustive experimentation regime to validate the scalability and speed-up of our approach, we compare it against a state-of-the-art high performance computing framework and showcase its time and cost advantages.
|
## RESEARCH
## Open Access
# Data‑aware optimization of bioinformatics workflows in hybrid clouds
### Athanassios M. Kintsakis[*], Fotis E. Psomopoulos and Pericles A. Mitkas
*Correspondence:
akintsakis@issel.ee.auth.gr
Department of Electrical
and Computer
Engineering, Aristotle
University of Thessaloniki,
54124 Thessaloniki, Greece
**Abstract**
Life Sciences have been established and widely accepted as a foremost Big Data
discipline; as such they are a constant source of the most computationally challenging
problems. In order to provide efficient solutions, the community is turning towards
scalable approaches such as the utilization of cloud resources in addition to any
existing local computational infrastructures. Although bioinformatics workflows are
generally amenable to parallelization, the challenges involved are however not only
computationally, but also data intensive. In this paper we propose a data management
methodology for achieving parallelism in bioinformatics workflows, while simultaneously minimizing data-interdependent file transfers. We combine our methodology
with a novel two-stage scheduling approach capable of performing load estimation
and balancing across and within heterogeneous distributed computational resources.
Beyond an exhaustive experimentation regime to validate the scalability and speed-up
of our approach, we compare it against a state-of-the-art high performance computing framework and showcase its time and cost advantages.
**Keywords: Cloud computing, Component-based workflows, Bioinformatics, Big data**
management, Hybrid cloud, Comparative genomics
**Introduction**
There is no doubt that Life Sciences have been firmly established as a Big Data science
discipline, largely due to the high-throughput sequencers that are widely available and
extensively utilized in research. However, when it comes to tools for analyzing and interpreting big bio-data, the research community has always been one step behind the actual
acquisition and production methods. Although the amount of data currently available is
considered vast, the existing methods and extensively used techniques can only hint at
the knowledge that can be potentially extracted and consequently applied for addressing a plethora of key issues, ranging from personalized healthcare and drug design to
sustainable agriculture, food production and nutrition, and environmental protection.
Researchers in genomics, medicine and other life sciences are using big data to tackle
fundamental issues, but actual data management and processing requires more networking and computing power [14]. Big data is indeed one of today’s hottest concepts, but it
can be misleading. The name itself suggests mountains of data, but that’s just the start.
Overall, big data consists of three v’s: volume of data, velocity of processing the data, and
© The Author(s) 2016. This article is distributed under the terms of the Creative Commons Attribution 4.0 International License
[(http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium,](http://creativecommons.org/licenses/by/4.0/)
provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and
indicate if changes were made.
-----
variability of data sources. These are the key features of information that require particular tools and methodologies to efficiently address them.
The main issue with dealing with big data is the constantly increasing demands for
both computational resources as well as storage facilities. This in turn, has led to the
rise of large-scale high performance computing (HPC) models, such as cluster, grid and
cloud computing. Cloud computing can be defined as a potentially high performance
computing environment consisting of a number of virtual machines (VMs) with the ability to dynamically scale resources up and down according to the computational requirements. This computational paradigm has become a popular choice for researchers that
require a flexible, pay-as-you-go approach to acquiring computational resources that can
accompany their local computational infrastructure. The combination of public and privately owned clouds defines a hybrid cloud, i.e. an emerging form of a distributed computing environment.
From this perspective, optimizing the execution of data-intensive bioinformatics workflows in hybrid clouds is an interesting problem. Generally speaking, a workflow can be
described as the execution of a sequence of concurrent processing steps, or else computational processes, the order of which is determined by data interdependencies as well as
the target outcome. In a data-intensive workflow, data and metadata, either temporary
or persistent, are created and read at a high rate. Of course, a workflow can be both data
and computationally intensive and the two are often found together in bioinformatics
workflows. In such workflows, when scheduling tasks to distributed resources, the data
transfers between tasks are not a negligible factor and may comprise a significant portion of the total execution time and cost. A high level of data transfers can quickly overwhelm the storage and network throughput of cloud environments, which is usually on
the order of 10–20 MiB/s [6], while also saturating the bandwidth of local computational
infrastructures and leading to starvation of resources to other users and processes.
It is well known that a high level of parallelization can be achieved in a plethora of bioinformatics workflows by fragmenting the input of individual processes into chunks and
processing them independently, thus achieving parallelism in an embarrassingly parallel
way. This is the case in most evolutionary investigation, comparative genomics and NGS
data analysis workflows. This fact can be largely taken advantage of in order to achieve
parallelism by existing workflow management approaches emphasizing parallelization.
The disadvantage of this approach however is that it creates significant data interdependencies, which in turn lead to data transfers that can severely degrade performance
and increase overall costs.
In this work, we investigate the problem of optimizing the parallel execution of dataintensive bioinformatics workflows in hybrid cloud environments. Our motivation is
to achieve better time and cost efficiency than existing approaches by minimizing file
transfers in highly parallelizable data-intensive bioinformatics workflows. The main contributions of this paper are twofold; (a) We propose a novel data management paradigm
for achieving parallelism in bioinformatics workflows while simultaneously minimizing
data-interdependency file transfers, and (b) based on our data management paradigm,
we introduce a 2-stage scheduling approach balancing the trade-off between parallelization opportunities and minimizing file transfers when mapping the execution of bioinformatics workflows into a set of heterogeneous distributed computational resources
-----
comprising a hybrid cloud. Finally, in order to validate and showcase the time and cost
efficiency of our approach, we compare our performance with Swift, one of the most
widely used and state-of-the-art high performance workflow execution frameworks.
The rest of the paper is organized as follows: a review of the state-of-the-art on workflow management systems and frameworks in general and in the field of bioinformatics
in particular is presented in "Related work" section. "Methods" section outlines the general characteristics and operating principles of our approach. "Use case study" section
briefly presents the driving use case that involves the construction of phylogenetic profiles from protein homology data. "Results and discussion" section provides the results
obtained through rigorous experimentation, in order to evaluate the scalability and efficiency as well as the performance of our approach when compared against a high performance framework. Finally, concluding remarks and directions for future work are given
in "Conclusions and future work" section.
**Related work**
The aforementioned advantages of cloud computing have led to its widespread adoption
in the field of bioinformatics. Initial works were mostly addressed on tackling specific,
highly computationally intensive problems that outstretched the capabilities of local
infrastructures. As the analyses became more complex and incorporated an increasing
number of modules, several tools and frameworks appeared that aimed to streamline
computations and automate workflows. The field of bioinformatics has also sparked the
interest of many domain agnostic workflow management systems, some of the most prolific applications of which were bioinformatics workflows, thus leading to the development of pre-configured customized versions specifically for bioinformatics workflows
[34].
Notable works addressing well-known bottlenecks in computationally expensive pipelines, the most characteristic of which are Next Generation Sequencing (NGS) data analysis and whole genome assembling (WGA) include [18], Rainbow [9], CloudMap [29],
CloudBurst [40], SURPI [31] and RSD-Cloud [45]. These works, although highly successful, lack a general approach as they are problem specific and are often difficult to setup,
configure, maintain and most importantly integrate within a pipeline, when considering
the experience of a non-expert life sciences researcher.
Tools and frameworks aiming to streamline computations and automate standard
analysis bioinformatics workflows include Galaxy [17], Bioconductor [16], EMBOSS [39]
and Bioperl [43]. Notable examples of bioinformatics workflow execution in the cloud
include [11, 33] and an interesting review on bioinformatics workflow optimization in
the cloud can be found in [15]. In the past few years, there is a significant trend in integrating existing tools into unified platforms featuring an abundance of ready to use tools,
with particular emphasis on ease of deployment and efficient use of resources of the
cloud. A platform based approach is adopted by CloudMan [1], Mercury [38], CLoVR
[3], Cloud BioLinux [22] and others [24, 32, 42, 44]. Most of these works are addressing
the usability and user friendly aspect of executing bioinformatics workflows, while some
of them also support the use of distributed computational resources. However, they
largely ignore the underlying data characteristics of the workflow and do not perform
any data-aware optimizations.
-----
Existing domain agnostic workflow management systems including Taverna [48], Swift
[49], Condor DAGMan [23], Pegasus [13], Kepler [26] and KNIME [5] are capable of also
addressing bioinformatics workflows. A comprehensive review of the aspects of parallel workflow execution along with parallelization in scientific workflow managements
systems can be found in [8]. Taverna, KNIME and Kepler mainly focus on usability by
providing a graphical workflow building interface while offering limited to non-existent
support, in their basic distribution, for use of distributed computational resources. On
the other side, Swift, Condor DAGMan and Pegasus are mainly inclined over accomplishing parallelization on both local and distributed resources. Although largely successful in achieving parallelization, their scheduling policies are non data-aware and do
not address minimizing file transfers between sites.
Workflow management systems like Pegasus, Swift and Spark can utilize shared file
systems like Hadoop and Google Cloud Storage. The existence of a high performance
shared file system can be beneficial in a data intensive worfklow as data can be transferred directly between sites and not staged back and forth from the main site. However,
the advantages of a shared file system can be outmatched by a data-aware scheduling
policy which aims to minimize the necessity of file transfers to begin with. Furthermore,
the existence of a shared file system is often prohibitive in hybrid clouds comprising of
persistent local computational infrastructures and temporarily provisioned resources
in the cloud. Beyond the significant user effort and expertise required in setting up a
shared file system, one of the main technical reasons for this situation is that elevated
user operating system privileges are required for this operation, which are not usually
granted in local infrastructures.
A Hadoop MapReduce [12] approach is capable of using data locality for efficient
task scheduling. However, its advantages become apparent in a persistent environment
where the file system is used for long term storage purposes. In the case of temporarily
cloud provisioned virtual machines, the file system is not expected to exist either prior
or following the execution of the workflow and consequently all input data are loaded at
the beginning of the workflow. There is no guarantee that all the required data for a specific task will be placed in the same computational site and even if that were the case, no
prior load balancing mechanism exists for assigning all the data required for each task
to computational sites while taking into account the computational resources of the site
and the computational burden of the task. Additionally, a MapReduce approach requires
re-implementation of many existing bioinformatics tools which is not only impractical
but also unable to keep up to date with the vanilla and standardized versions.
Finally, it is important to note that none of the aforementioned related work clearly
addresses the problem of applying a data-aware optimization methodology when executing data-intensive bioinformatics workflows in hybrid cloud environments. It is exactly
this problem that we address in this work, by applying a data organization methodology
coupled with a novel scheduling approach.
**Methods**
In this section we introduce the operating principles and the underlying characteristics
of the data management and scheduling policy comprising our methodology.
-----
**Data management policy**
The fact that data parallelism can be achieved in bioinformatics workflows has largely
been taken advantage of in order to accelerate workflow execution. Data parallelism
involves fragmenting the input into chunks which are then processed independently. For
certain tasks of bioinformatics workflows, such as sequence alignment and mapping of
short reads which are also incidentally some of the most computationally expensive processes, this approach can allow for a very high degree of parallelism in multiprocessor
architectures and distributed computing environments. However, prior to proceeding
to the next step, data consistency requires that the output of the independently processed chunks be recombined. In a distributed computing environment, where the data
is located on multiple sites, this approach creates significant data interdependency issues
as data needs to be transferred from multiple sites in order to be recombined, allowing
the analysis to proceed to the next step. The same problem is not evident in a multiprocessor architecture, as the data exists within the same physical machine.
A sensible approach to satisfying data interdependencies with the purpose of minimizing, or even eliminating unnecessary file transfers would be to stage all fragments whose
output must be recombined on the same site. Following that, the next step, responsible
for processing the recombined output, can also be completed on the same site, and then
the next step, that will operate on the output of the previous, also on the same site, further advancing this course until it is no longer viable. It is becoming apparent that this
is a recursive process that takes into account the anticipated data dependencies of the
analysis. In this way, segments of the original workflow are partitioned into workflow
ensembles (workflows of similar structure but differing in their input data) that have no
data interdependencies and can then be executed independently in an approach reminiscent of a bag-of-tasks. Undoubtedly, not all steps included in a workflow can be managed this way, but a certain number can, often also being the most computationally and
data intensive.
Instead of fragmenting the input of data parallelizable tasks into chunks arbitrarily, we
propose fragmenting into chunks that can also sustain the data dependencies of a number of subsequent steps in the analysis. Future tasks operating on the same data can be
grouped back-to-back into forming a pipeline. To accomplish the aforementioned, we
model the data input space as comprising of Instances. An Instance (Inst) is a single data
entry, the simplest form data can exist independently. An example of an Inst would be
a single protein sequence in a .fasta file. Instances are then organized into organization
units (OU), which are sets of instances that satisfy the data dependencies of one or more
tasks. The definition of an OU is a set of Insts that can satisfy the data dependencies of a
number of consecutive tasks, thus allowing the formation of an OU pipeline.
However, before attempting to directly analyze the data involved, a key step is to preprocess the data instances in order to allow for a structured optimization of the downstream analysis process. A common occurrence in managing big data is the fact that
their internal organization is dependent on its specific source. Our data organization
model is applied through a preprocessing step that restructures the initial data organization into sets of Insts and OUs in a way reminiscent of a base transformation.
The process involves identifying Insts in the input data, and grouping them together
into OUs according to workflow data interdependencies. An identifier is constructed for
-----
each Inst that also includes the OU it belongs to. The identifier is permanently attached
to the respective data and therefore is preserved indefinitely. The initial integrity of the
input data is guaranteed to be preserved during workflow execution, thus ensuring the
accessibility to this information in later stages of the analysis and allowing for the recombination process. The identifier construction process is defined as follows.
**Definition 1 Each** _OU is a set that initially contains a variable number (denoted by_
_n, k, l, ...) of instances Instj where_ j = [1, n]. The internal order of instances within an OU
is preserved as the index assigned to each unique identifier Instj (i.e. the order 1 < i < n
of the instances) is reflected directly upon the constructed identifier. The total number of
_m_ _OUs themselves are grouped into a set of OUs and are each assigned unique identifi-_
ers OUi constructed in a semi-automated manner to better capture the semantic context
of the defined _OUs. Finally, the instance identifier,_ _InstID consists of the concatenated_
OUi and Instj parameters, as shown below:
OUs = {OU0, OU1, OU2, . . ., OUm} (1)
OU0 = {Inst0, ..., Instn}, OU1 = {Inst0, ..., Instk } ... (2)
OUn = {Inst0, ..., Instl}
InstID = F (OUi, Instj) = OUi_Instj (3)
At some point, some or all the pipelines may converge in what usually is a non parallelizable merging procedure. This usually happens at the end of the workflow, or in intermediate stages, before a new set of OU pipelines is formed and the analysis continues
onward.
**Scheduling policy**
It is obvious that this data organization approach although highly capable of minimizing
data transfers, it severely limits the opportunities for parallelization, as each OU pipeline
is processed in its entirety in a single site. In very small analyses where the number of
_OUs is less than the number of sites, obviously some sites will not be utilized, though_
this is a boundary case, unlikely to occur in real world analyses.
In a distributed computing environment, comprised of multiprocessor architecture
computational sites, ideally each _OU pipeline will be assigned to a single processor._
Given that today’s multiprocessor systems include a significant number of CPU cores,
the number of _OU pipelines must significantly exceed, by a factor of at least 10, the_
number of sites in order to achieve adequate utilization. Unfortunately, even that would
prove inadequate, as the computational load of OU pipelines may vary significantly, thus
requiring an even higher number of them in order to perform proper load balancing. It
is apparent that this strategy would be fruitful only in analyses where the computational
load significantly exceeds the processing capabilities of the available sites, spanning
execution times into days or weeks. In solely data-intensive workflows, with no computationally intensive component, under-utilization of multiprocessor systems may not
-----
become apparent as storage and network throughput are the limiting factors. Otherwise,
it will most likely severely impact performance.
Evidently, a mechanism for achieving parallelism in the execution of an OU pipeline in a
single site is required. Furthermore, in a heterogeneous environment of computational sites
of varying processing power and OU pipelines of largely unequal computational loads, load
balancing must be performed in order to map the OU pipelines into sites. To address these
issues we propose a novel 2-stage scheduling approach which combines an external scheduler at stage 1 mapping the OU pipelines into sites and an internal to each site scheduler at
stage 2 capable of achieving data and task parallelism when processing an OU pipeline.
**_External scheduler_**
The external scheduler is mainly concerned with performing load balancing of the OU
pipelines across the set of computational resources. As both the OU pipelines and the
computational sites are largely heterogeneous, the first step is performing an estimation
regarding both the OU pipeline loads and the processing power of the sites. The second
step, involves the utilization of the aforementioned estimations by the scheduling algorithm tasked with assigning the OU pipelines to the set of computational resources.
In order to perform an estimation of the load of an OU pipeline, a rough estimation
could be made based on the size of the OU input. A simple approach would be to use the
disk file size in MB but that would most likely be misleading. A more accurate estimation could be derived by counting the number of instances, this approach too however
is also inadequate as the complexity cannot be directly assessed in this way. In fact, the
computational load can only be estimated by taking into account the type of information presented by the file, which is specific to its file type. For example, given a .fasta file
containing protein sequences, the most accurate approach for estimating the complexity of a sequence alignment procedure would be to count the number of bases, rather
than count the number of instances. Fortunately, the number of distinct file types found
in the most common bioinformatics workflows is small, and therefore we have created
functions for each file type that can perform an estimation of the computational load
that corresponds to them. We already support formats of .fasta, .fastq and plain ASCII
(such as tab-delimited sequence similarity files) among others.
In order to better match the requirements of the data processing tasks to the available computational resources, the computational processing power of each site must also
be assessed. This is accomplished by running a generic benchmark on each site which is
actually a mini sample workflow that aims to estimate the performance of the site for similar workflows. The benchmarks we currently use are applicable on comparative genomics and pangenome analysis approaches, and measure the multithreaded performance
of the site, taking into account its number of CPU cores. We also use the generic tool
UnixBench [41] to benchmark the sites when no similar sample workflow is available.
The problem can now be modeled as one of scheduling independent tasks of unequal
load to processors of unequal computational power. As these tasks are independent,
they can be approached as a bag of tasks. Scheduling bag of tasks has been extensively
studied and many algorithms exist, derived from heuristic [46], list scheduling [20] or
metaheuristic optimization approaches [30]. In this work we utilize one of the highest
performing algorithms, the FPLT (fastest processor largest task) algorithm. According to
-----
FPLT, tasks are placed in descending order based on their computational load and each
task, starting from the largest task, is assigned to the fastest available processor. Whenever a processor completes a task, it is then added to the list of available processors, the
fastest of which is assigned the largest remaining task.
FPLT is a straightforward and lightweight algorithm, capable of outperforming other
solutions most of the time [20] when all tasks are available from the start, as is the case
here, without adding any computational burden. The disadvantage of FPLT is that when
the computational power of processors is largely unequal, a processor might be assigned
a task that severely exceeds its capabilities, thus delaying the makespan of the workflow.
This usually happens when some processors are significantly slower than the average
participating in the workflow.
The external scheduler initially performs an assessment of the type and load of the OU
pipelines. It then determines the capabilities of the available sites in processing the pipelines by retrieving older targeted benchmarks or completing new on the fly. The OU pipelines are then submitted to the sites according to FPLT and job failures are handled by
resubmission. The pseudocode of the external scheduler is presented in Algorithmic Box 1.
**_Internal scheduler_**
The internal scheduler is local to each site and is responsible for achieving data and
task parallelism when processing an _OU pipeline. Task parallelism involves executing_
independent tasks directly in parallel while data parallelism requires the identification
of tasks whose input can be fragmented in chunks and processed in parallel. The second requires that such tasks are marked as suitable for fragmentation at the workflow
description stage or maintaining a list of such tasks for automatic identification. Our
approach supports both.
-----
The internal scheduler automatically identifies the number of CPUs on the computational site and sets the number of simultaneous processing slots accordingly. It receives
commands from the master and assigns them to threads in order to execute them in
parallel. In case it receives a task where data parallelism is possible, it will fragment the
input into individual chunks, or else subsets, and then launch threads in order to process
them in parallel. A decision must be made on the number of fragments a task must be
split to, which involves a trade off between process initialization overhead and load balancing between threads.
Given the widely accepted assumption that the CPU cores of a given site have the same
computational capabilities, a simple solution would be to launch a number of threads
equal to the machine’s CPU count and divide the total number of input data, or else
the instances, across them. This solution is in turn predicated on the assumption that
the load assigned to a thread should directly correspond to the amount of data it has
to process and as such is prone to variations. In our case however, as all required data
exists within the same site, it is no longer desirable to distribute the data processing load
among the threads in advance, as the data can be accessed by any thread at any time
without any additional cost thus providing greater flexibility.
Therefore, when considering the situation within a single sitel, our approach can be
defined by the process of splitting the superset of all m _Insts of the OU pipeline into k_
subsets of fixed size n. The number of subsets is given when dividing m by n.
Superset{Inst0, ..., Instm} = Subset1{Inst0, ..., Instn} ∪ ... ∪ Subsetk {Inst0, ..., Instn} (4)
k = [m] (5)
n
Each given Subseti, is assigned to a thread responsible for completing the respective task.
Initially the subsets are placed into a list in random order. Each thread attempts to process the next available subset and this continues recursively until all available subsets are
exhausted. In order to synchronize this process and to ensure that no two threads process the same subset, a lock is established that monitors the list of subsets. Every time a
thread attempts to obtain the next available subset it must first acquire the lock. If the
lock is unavailable the thread is set to sleep in a waiting queue. If the lock is available, the
thread acquires the requested subset and increases an internal counter that points to the
next available subset. It then immediately releases the lock, an action that also wakes the
first thread that may be present in the queue. The pseudocode describing the operation
of the internal scheduler is presented in Algorithmic boxes 2 and 3.
-----
As the probability of two threads completing the execution of a subset at exactly the
same time is extremely low, the synchronization process has been proven experimentally to be very efficient, where most of the time there are no threads waiting on the
queue. The average waiting time along with the time of acquiring and releasing the lock
is usually minuscule. However, there is an important overhead that is associated with
the initialization of the process that will complete the task. An accurate estimation of
this overhead time is difficult to obtain as it is dependent on the actual processes being
launched and the overall status of the operating system at any given time. We estimate
this overhead to be around 300–1000 ms. A totalDelay parameter that indicates the estimated initialization delay involved in processing a given subset can be evaluated. This
parameter can be constructed by multiplying the number k of subsets with the overhead
parameter that reflects the average time wasted on synchronization and launching the
respective processes, and dividing the result by the number of threads, as follows:
overhead
totalDelay = k ∗ (6)
threadCount
It becomes apparent that minimizing the _totalDelay time is equal to minimizing the_
number of subsets k. The minimum value of k is equal to the number of threads in which
case the overhead penalty is suffered only once by each thread. However it is unwise
-----
to set _k equal to the number of threads as the risk of unequally distributing the data_
between the threads far outweighs the delay penalty.
We make the reasonable hypothesis that the execution times of chunks of fixed size
n = 1 resemble a Log Normal distribution, which is typically encountered in processing
times [4]. Our hypothesis was verified on an individual basis experimentally by running
a BLAST procedure as presented in Fig. 1. BLAST is the most computationally intensive
task of our use case study workflow presented in . Evidently, this does not apply to all
tasks but is a reasonable hypothesis and a common observation in processing times.
A Log Normal distribution appears approximately like a skewed to the right, positive
values only, normal distribution. This particular distribution presented in Fig. 1 allows
us to estimate that only 8.2 % of the processing times were twice as large as the average processing time. Moreover, less than 0.5 % of the processing times were larger than
five times the average processing time. It can easily be asserted that from a given set
size and below, it is highly unlikely for many of the slower processing times to appear
within it. However, it must be noted that this already low probability is further reduced
by the fact that this is a boundary situation, to be encountered by the end of the workflow where other threads have terminated. After experimentation we have established
that an empirical rule to practically eliminate the chance is to set n equal to 0.01 % of the
number m of instances.
**Fig. 1 An experimental run presenting the execution times of subsets with a size of one, when our specific**
membership function that involves BLAST alignment and phylogenetic profiling building (presented in "Use
case study" section ) is run. It is apparent that the execution times follow a Log Normal distribution which is
outlined by the red line
-----
The delayTime % defined by Eq. 7 is the total time wasted as a percentage of the actual
processing time.
totalDelay
delaytime % = m (7)
n [∗] [avgProcessingTime][ ∗] [threads][ ∗] [100]
Assuming that the average processing time, avgProcessingTime, of a single instance is at
least two and a half times greater than the overhead time and the number of threads is
at least eight, then by setting n at 0.01 % of m will lead to a delayTime % value equal to
0.05 % which is considered insignificant.
We conclude that a value of _n approximating 0.01 % of_ _m is a reasonable compro-_
mise. In practice, other limitations to the size of the subset n may exist, that are related
to the nature of the memberships functions involved and must be taken into account.
For example, in processes using hash tables extensively or having significant memory
requirements, a relatively high subset size would not be beneficial as there is risk for the
hash tables to be overloaded resulting in poor performance and high RAM usage.
It is evident that an accurate size n of the subsets cannot be easily calculated from a
general formula as it may have specific constraints due to the actual processes involved.
However, a general rule of thumb can be established of setting n around 0.01 % of m and
is expected to work reasonably well for the majority of cases. It is however, classified as a
parameter that can be optimized and thus its manipulation is encouraged on a use case
basis.
**Execution engine**
A number of requirements motivated us to implement a basic workflow execution
engine that was used in our experiments for validating our approach. These requirements are the deployment of containers on sites that include all the necessary software and tools, graphic workflow description, secure connections over SSH tunneling
and HTTPS and not requiring elevated user privileges for accessing sites. The execution environment is comprised of a number of computational sites having a UNIX based
operating system and a global, universally accessible cloud storage similar to Amazon S3,
referred to as object storage. The object storage is used to download input data, upload
final data and to share data between sites. It is not used for storing intermediate data
that temporarily exist within each site. We have implemented the proposed framework
using Java 8 and Shell scripting in Ubuntu Linux 14.04.
The overall architecture is loosely based on a master/slave model, where a master node
responsible for executing the external scheduler serves as the coordinator of actions
from the beginning to the completion of a given workflow. The master node is supplied
with basic information like the description of the workflow and input data, the object
storage and the computational sites. The workflow can be described as a directed acyclic
graph (DAG) in the GraphML [7] language by specifying graph nodes corresponding to
data and compute procedures and connecting them with edges as desired. To describe
the workflow in a GUI environment, the user can use any of the available and freely distributed graph design software tools that supports exporting to GraphML.
The only requirement for using a computational site is the existence of a standard user
account and accessibility over the SSH protocol. Each site is initialized by establishing a
-----
secure SSH connection through which a Docker [28] container equipped with the software dependencies required to execute the workflow is fetched and deployed. Workflow
execution on each site takes place within the container. The object storage access credentials are transferred to the containers and a local daemon is launched for receiving
subsequent commands from the master. The daemon is responsible for initiating the
internal scheduler and passing all received commands to it. Communication between the
master and the daemons running within the Docker container on each site is encrypted
and takes place over SSH tunneling. File transfers between sites and the object storage
are also encrypted and take place over the HTTPS protocol.
**Use case study**
The selected case study utilized in validating our approach is from the field of comparative genomics, and specifically the construction of the phylogenetic profiles of a set of
genomes. Phylogenetic profiling is a bioinformatics technique in which the joint presence or joint absence of two traits across large numbers of genomes is used to infer a
meaningful biological connection, such as involvement of two different proteins in the
same biological pathway [35, 37]. By definition, a phylogenetic profile of a genome is
an array where each line corresponds to a single sequence of a protein belonging to the
genome and contains the presence or absence of the particular entity across a number of
known genomes that participate in the study.
The first step in building phylogenetic profiles involves the sequence alignment of the
participating protein sequences of all genomes against themselves. It is performed by
the widely used NCBI BLAST tool [25] and the process is known as a BLAST all vs all
procedure. Each protein is compared to all target sequences and two values are derived,
the identity and the e-value. _Identity refers to the extent to which two (nucleotide or_
amino acid) sequences have the same residues at the same positions in an alignment,
and is often expressed as a percentage. _E-value (or expectation value or expect value)_
represents the number of different alignments with scores equivalent to or better than a
given threshold S, that are expected to occur in a database search by chance. The lower
the E-value, the more significant the score and the alignment.
Running this process is extremely computationally demanding, the complexity of
which is not straightforward to estimate [2], but can approach O(n[2]). For example, a simple sequence alignment between 0.5 million protein sequences, can take up to a week
on a single high-end personal computer. Even when employing high-performance infrastructures, such as a cluster, significant time as well as the expertise to both run and
maintain a cluster-enabled BLAST variant are required. Furthermore the output files
consume considerable disk space which for large analyses can easily exceed hundreds of
GBs.
Based on the sequence alignment data, each phylogenetic profile requires the comparison and identification of all homologues across the different number of genomes in
the study. The phylogenetic profiling procedure for each genome requires the sequence
alignment data of all its proteins against the proteins of all other genomes. Its complexity is linear to the number of sequence alignment matches generated by blast. Different
types of phylogenetic profiles exist, including binary, extended and best bi-directional all
3 of which are constructed in our workflow procedure.
-----
According to our data organization methodology, in this case proteins correspond to
_Insts and are grouped into OUs, which in this case are their respective genomes. Inde-_
pendent pipelines are formed for each _OU consisting firstly of the BLAST process_
involving the sequence alignment of the proteins of the _OU against all other proteins_
of all OUs and secondly of the three phylogenetic profile creation processes which utilizes the output of the first in order to create the binary, extended and best bi-directional
phylogenetic profile of the genome corresponding to the OU. These pipelines are then
scheduled according to the scheduling policy described in .
**Results and discussion**
A number of experiments have been performed in order to validate and evaluate our
framework. Therefore, this section is divided into (a) the validation experiments further
discussed in "Validation" subsection, where the methods outlined in "Methods" section are validated and (b) the comparison against Swift, a high performance framework,
further discussed in "Comparison against a high performance framework" subsection
where the advantages of our approach become apparent.
The computational resources used are presented in Table 1. Apart from the privately
owned resources of our institution, the cloud resources consist of a number of virtual
machines belonging to the European Grid Infrastructure (EGI) federated cloud and
operated by project Okeanos [21] of GRNET (Greek Research and Technology Network). Okeanos is based on the Synnefo (the meaning of the word is “cloud” in Greek)
open source cloud software which uses Google Ganeti and other third party open source
software. Okeanos, is the largest academic cloud in Greece, spanning more than 5400
active VMs and more than 500,000 spawned VMs.
As the resources utilize different processors of unequal performance, their performance was compared to the processors of the cloud resources which served as a baseline
reference. As such, the number of CPUs of each site was translated to a number of baseline CPUs, so a direct comparison can be performed. In this way, non integer numbers
appear in the number of baseline CPUs of each site.
This combination of local, privately owned computational resources with cloud-based
resources represents the typical use case we are addressing, individuals or research labs
that wish to extend their computational infrastructure by adopting resources of one or
multiple cloud vendors.
**Table 1 The pool of available computational resources along with their hardware type,**
**number of threads and number of baseline processors are presented**
**# Count** **CPU type** **# CPUs** **# Baseline CPUs**
1× 2 × Intel Xeon E5 2660 @ 2.2 GHZ 24 21.7
1× 2 × Intel Xeon E5 2660 @ 2.2 GHZ 12 16.7
1× Intel i7 6700 @ 4.0 GHZ 8 15
1× Intel i7 4790S @ 3.5 GHZ 8 11.3
10× AMD Opteron 6172 @ 2.1 GHZ 8 8
Total
14 – 132 144.7
All machines were equipped with more than 6 GB of RAM and were connected to the internet through a 100 MBps
connection
-----
The input data used in our experiments consists of an extended plant pangenome of 64
plant genomes including 39 cyanobacteria for which the complete proteome was available. The total size was 268 MB and includes 619,465 protein sequences nd 2.3 × 10[8]
base pairs. In order to accommodate our range of experiments, the data was divided into
sub-datasets.
It must be noted that, although the input data used may appear relatively small in file
size, it can be very demanding to process, requiring weeks on a single personal computer.
The particular challenge in this workflow is not the input size but the computational
requirements in conjunction with the size of the output as will become apparent in the
following sections. The dataset consist of files downloaded from the online and publicly
accessible databases of UniProt [10] and PLAZA[36] and can also be provided by our
repositories upon request. The source code of the proposed framework along with the
datasets utilized in this work can be found in our repository https://www.github.com/
akintsakis/odysseus.
**Validation**
In order to experimentally validate the optimal subset size value as outlined in "Internal
scheduler" section and the overall scalability performance of our approach, a number of
experiments were conducted utilizing the phylogenetic profiling use case workflow. All
execution times reported below involve only the workflow runtime and do not include
site initialization and code and initial data downloads as they require a nearly constant
time, irrespective of both problem size and number of sites and as such they would distort the results and not allow for accurately measuring scaling performance. For reporting purposes, the total time for site initialization is approximately 3–5 min.
**_Optimal subset value n_**
The phylogenetic profiling workflow was executed with an internal scheduler subset size
value _n of 0.0010, 0.0025, 0.0050, 0.0100, 0.0250, 0.0500 and 0.2500 % as a percentage_
of the total number of protein sequences in three distinct datasets comprising 189,378,
264,088 and 368,949 protein sequences. All sites presented in Table 1 except for the first
one, participated in this experiment. The site execution times for each subset size for all
three datasets are presented in boxplot form in Fig. 2. They verify the hypothesis presented in "Internal scheduler" subsection, we observe that the fastest execution time is
achieved when the subset size n is set close to our empirical estimation of 0.01 % of the
total dataset size. It is apparent that smaller or larger values of n lead to increased execution times. Generally, in both three datasets analyzed we observe the same behavior and
pattern of performance degradation when diverging from the optimal subset size.
Smaller values of _n lead to substantially longer processing times mainly due to the_
delay effect presented in Eq. 7. As _n increases, the effect gradually attenuates and is_
diminished for values larger than 0.0050 % of the dataset. Larger subset sizes impact performance negatively, with the largest size of 0.2500 % tested, yielding the slowest execution time overall. This can be attributed to the fact that for larger subset sizes, the load
may not be optimally balanced and some threads that were assigned disproportionately
higher load might prolong the overall total execution time while other threads are idle.
Additionally, large subset sizes can lead to reduced opportunities for parallelization,
-----
**Fig. 2 The execution times of all 13 participating sites in boxplot form are presented for the phylogenetic**
profiling workflow when executed with an internal scheduler subset size value n of 0.0010, 0.0025, 0.0050,
0.0100, 0.0250, 0.0500 and 0.2500 % as a percentage of the total number of protein sequences in three distinct datasets comprising 189,378, 264,088 and 368,949 protein sequences. The optimal value of n leading to
the fastest execution times is 0.01 % of the input dataset
especially on smaller OUs that are broken into fewer chunks than the available threads
on site, thus leaving some threads idle.
The average memory usage of all execution sites for each subset size for all three datasets is presented in Fig. 3. It is apparent that both the subset size and the size of the
dataset increase memory consumption. Between smaller subset sizes, differences in
memory usage are insignificant and inconsistent, thus difficult to measure. As we reach
the larger subsets, the differences become more apparent. Due to the current workflow
not being memory intensive, increases in memory usage are only minor. However, in a
-----
**Fig. 3 Average memory utilization of all 13 participating sites of the the phylogenetic profiling workflow**
when executed with an internal scheduler subset size value n of 0.0010, 0.0025, 0.0050, 0.0100, 0.0250, 0.0500
and 0.2500 % as a percentage of the total number of protein sequences in three distinct datasets comprising
189,378, 264,088 and 368,949 protein sequences. Both subset size and dataset size seem to increase memory
consumption, though the differences are minimal due to the workflow not being memory demanding
memory demanding workflow these differences could be substantial. Although the size
of the dataset to be analyzed cannot be tuned, the subset size can and it should be taken
into account in order to remain within the set memory limits. A subset size n value of
0.0100 % is again a satisfactory choice when it comes to keeping memory requirements
on the low end.
Although we have validated that an adequate and cost-effective approach is to set the
value of n at 0.0100 % of the total size of the dataset, we must state that optimal selection of n is also largely influenced by the type of workflow and thus its manipulation is
encouraged on a use case basis.
**_Execution time reduction scaling_**
In this experiment the performance scalability of our approach was evaluated. For the
needs of this experiment, a subset of the original dataset was formed, consisting of only
the 39 cyanobacteria. The purpose was to evaluate the speed-up and efficiency and compare it to the ideal case, in which a linear increase in the total available processing power
would lead to an equal reduction in processing time. Speed-up S(P), and efficiency E(P),
-----
are fundamental metrics for measuring the performance of parallel applications and are
defined in literature as follows:
S(p) = [T] [(][1][)] (8)
T (p)
where T(1) is the execution time with one processor and T(p) is the execution time with
p processors.
T (1)
E(p) = (9)
p ∗ T (p)
The above equations found in literature assume that p processors of equal computational
power are used. As in our case we use resources of uneven computational performance,
we translate their processing power into baseline processor units. Consequently, p can
take continuous and not discrete values, corresponding to the increase in computational
power as measured in baseline processors.
All sites presented in Table 1 participated in this experiment. The sites were sorted in
ascending order according to their multithreaded performance and the workflow was
executed a number of times equal to the number of sites, by increasing the number of
participating sites one at a time.
The execution times of all sites are presented in boxplot form for all workflow runs in
Fig. 4. The X axis represents the total computational power score of the sites participating in the workflow and the Y axis, in logarithmic scale, represents the site execution
time in seconds. The dashed magenta line is the ideal workflow execution time (corresponding to linear speed-up) and it intersects the mean values of all boxplots. As we can
**Fig. 4 The workflow execution times of various configurations of participating sites are presented in boxplot**
form when the cyanobacteria phylogenetic profiling validation workflow is executed. Initially, only one site
participates in the workflow and the execution time is the longest as seen on the left of the figure. As new
sites are gradually added the execution times are decreased. The dashed magenta line represents the ideal
reduction in time based on the increase in total computational resources
-----
see variations in site execution time for each workflow run are consistent with no large
deviations present. There are outliers in some workflow runs towards the lower side,
where one site would terminate before others as there are no more OU pipelines to process. Despite being an outlier value however, they are not laid too far away in absolute
quantities. Execution times fall consistently as new sites are added and computational
resources are increased.
A closer inspection of the results can be found in Table 2 where the execution times
and the average and makespan speed-up and efficiency are analyzed. As expected, the
average speed-up is almost identical to the ideal case, where the speed-up is equal to
_p and efficiency approaches the optimal value of 1. This was to be expected, as our_
approach does not introduce any overhead and keeps file transfers to a minimum, almost
like as if all the processing took place on a single site. The minuscule variations observed
can be attributed to random variations in the processing power of our sites and/or our
benchmarking of the sites and to random events controlled by the OS. It can be observed
that when using a high number of CPUs the efficiency tends to marginally drop to 0.97.
This is attributed to the fact the data intensive part of the workflow is limited by disk
throughput and cannot be accelerated by increasing the CPU count. Although the data
intensive part is approximately 3–5 % of the total workflow execution time when excluding potential file transfers, using such a high number of CPUs for this workflow begins to
approach the boundaries of Amdahl’s law [19].
On average, the makespan efficiency is 0.954 % for all runs. It can be presumed that
the makespan speed-up and efficiency tend to reach lower values when a higher number
of sites are involved. This is to be expected as some sites terminate faster than others
when the pool of OU pipelines is exhausted and as such their resources are no longer
utilized. This effect becomes apparent mostly when using a very high number of CPUs
for the given workflow that results in a workflow completion time of less than 30 min.
Although it is apparent in this experiment, we are confident it will not be an issue in
**Table 2 Speed-up and efficiency as scalability metrics of our approach**
**#CPUs** **Execution times in seconds** **Average** **Makespan**
**Average** **Min** **Max** **Std** **S(p)** **E(p)** **S(p)** **E(p)**
8.0 35,601 35,601 35,601 0.0 8.00 1.00 8.00 1.00
16.0 17,834 17,629 18,037 288.3 15.9 0.99 15.8 0.98
24.0 11,866 11,845 11,880 18.1 24.0 1.00 24.0 0.99
32.0 8930 8,628 9,039 201.4 31.9 0.99 31.5 0.98
40.0 7127 6,768 7,240 201.5 39.9 0.99 39.3 0.98
48.0 5929 5,713 6,159 234.7 48.0 1.00 46.2 0.96
56.0 5095 4,834 5,285 206.8 55.9 0.99 53.9 0.96
64.0 4467 4,097 4,546 150.5 63.8 0.99 62.6 0.97
72.0 3968 3,767 4,207 178.0 71.7 0.99 67.7 0.94
80.0 3571 3,122 3,688 164.5 79.7 0.99 77.1 0.96
91.3 3138 2,970 3,378 182.7 90.7 0.99 84.3 0.92
106.2 2741 2,561 2,888 85.0 103.9 0.98 98.5 0.93
123.0 2378 2,080 2,604 191.8 119.8 0.97 109.3 0.89
144.7 2029 1,937 2,183 69.4 140.3 0.97 130.4 0.90
-----
real world cases as using 14 sites for this workflow, can be considered as an overkill and
therefore slightly inefficient.
In general, the average speed-up and efficiency is the metric of interest when evaluating the system’s cost efficiency and energy savings as our approach automatically shuts
down and releases the resources of sites that have completed their work. The makespan
speedup corresponds to the actual completion time of the workflow when all sites have
terminated and the resulting data is available. Our approach attempts to optimize the
makespan speed-up but with no compromise in the average speed-up, i.e. the system’s
cost efficiency. We can conclude from this experiment that the average speed-up is close
to ideal and the makespan speed-up is inferior to the ideal case by about 5 % on average
and can approach 10 % when using a high number of resources when compared to the
computational burden of the workflow.
**Comparison against a high performance framework**
To establish the advantages of our approach against existing approaches, we chose to
execute our use case phylogenetic profiling workflow in Swift and perform a comparison. Swift [49] is an implicitly parallel programming language that allows the writing of
scripts that distribute program execution across distributed computing resources [47],
including clusters, clouds, grids, and supercomputers. Swift is one of the highest performing frameworks for executing bioinformatics workflows in a distributed computing
environment. The reason we chose Swift is that it is a well established framework that
emphasizes parallelization performance and in use in a wide range of applications also
including bioinformatics.
Swift has also been integrated [27] into the popular bioinformatics platform Galaxy,
in order to allow for utilization of distributed resources. Although perfectly capable of
achieving parallelization, Swift is unable to capture the underlying data characteristics
of the bioinformatics workflows addressed in this work, thus leading to unnecessary file
transfers that increase execution times and costs and may sometimes even become overwhelming to the point of causing job failures.
The testing environment included all sites presented in Table 1 except for the first one,
as we were unable to set the system environment variables required by Swift, due to not
having elevated privileges access to it. In the absence of a pre-installed shared file system, the Swift filesystem was specified as local, where all data were staged from the site
where Swift was executing from. This is the default Swift option that is compatible with
all execution environments and does not require a preset shared file system. The maximum number of jobs on each site was set equal to the site’s number of CPUs.
Three datasets were chosen as input to the phylogenetic profiling workflow, these are
the total of 64 plant genomes and its subsets of 58 and 52 genomes. The datasets were
chosen with the purpose of approximately doubling the execution time of each workflow run when compared to the previous one. Uptime, system load and network traffic
among others were monitored on each site. In order to perform a cost analysis, we utilized parameters from the Google Cloud Compute Engine pricing model, according to
which, the cost per hour to operate the computational resources is 0.232$ per hour per
8 baseline CPUs and the cost of network traffic is 0.12$ per GB as per Google’s internet
egress worldwide cheapest zone policy.
-----
The makespan execution time, total network traffic and costs of our approach against
Swift when executing the phylogenetic profiling workflow for the three distinct datasets
are presented in Table 3. The values presented are average values of 3 execution runs.
As can be seen, for workflow runs 1 and 2, Swift is approximately 20 % slower in makespan and 16 % slower in the case of workflow run 3. This is attributed mostly to the time
lost waiting for the file transfers to take place in the case of Swift. It must be noted that
we were unable to successfully execute workflow 3 until termination with Swift, due to
network errors near the end of the workflow that we attribute to the very large number
of required file transfers. Had the workflow reached termination, we expect Swift to be
about 17–18 % slower. As the particular use case workflow is primarily computationally
intensive, an increase in the input size of the workflow increases the computational burden faster than the data intensive part, thus the performance gap is slightly smaller in the
case of workflow 3.
The total network traffic includes all inbound and outbound network traffic of all sites.
It is apparent that it is significantly higher in Swift thus justifying the increased total
execution time accounting to file transfers. Regarding the cost of provisioning the VMs,
it was calculated by multiplying the uptime of each site with the per processors baseline
cost of operation. The external scheduler of our approach will release available resources
when the pool of OU pipelines is exhausted, thus leading to cost savings that can range
from 10 to 25 % when compared to keeping all resources active until the makespan time.
Oppositely, this feature is not supported by Swift and as such in this case all sites are
active until makespan time, leading to increased costs. The cost savings of our approach
regarding provisioning of VMs were higher than 40 % in all three workflow.
The cost of network transfers is difficult to interpret as it is dependent on the locations and the providers of the computational resources. The cost presented here is a
worst case estimate that would take place when all network traffic between sites were
**Table 3 Execution time, network traffic and cost comparison of our approach against Swift**
**Workflow 1 1.16 ∗** **10[8] Bases** **Workflow 2 1.67 ∗** **10[8] Bases** **Workflow 3 2.3 ∗** **10[8] Bases**
Makespan
Ours 9302 s 18278 s 33752 s
Swift 11194 s 21933 s 39229 s
Diff +20.3 % +19.9 % +16.2 %
Network total traffic
Ours 0.183 GB 0.338 MB 0.403 GB
Swift 57.525 GB 88.944 GB 140.982 GB
Cost of Provisioning VMs
Ours 8.86 $ 17.6 $ 32.63 $
Swift 13.05 $ 25.56 $ 45.73 $
Diff +47.2 % +45.2 % +40.0 %
Cost of network transfers
Ours 0.02 $ 0.04 $ 0.05 $
Swift 6.90 $ 10.67 $ 16.91 $
Total cost
Ours 8.88 $ 17.64 $ 32.68 $
Swift 19.95 $ 36.23 $ 62.64 $
Diff +124.6 % +105.3 % +91.6 %
-----
charged at the nominal rate. That is not always true, for example if all sites were located
within the same cloud facility of one vendor there would be no cost at all for file transfers. However, they would still slow down the workflow leading to increased uptime
costs, unless the sites were connected via a high speed link like InfiniBand often found in
supercomputer configuration environments. In a hybrid cloud environment, which this
work addresses, as computational sites will belong to different cloud vendors and private
infrastructures, the file transfer cost can be significant and may even approach the worst
case scenario. In total, our approach is significantly more cost effective than Swift, which
can be anywhere from 40 to 47 % to more than 120 % more expensive, depending on the
pricing of network file transfers.
To further analyze the behavior of our framework against Swift, in Fig. 5 we present
the system load and network activity of all sites when executing the phylogenetic profiling workflow with the 64 genome input dataset for both our approach and Swift. The
Swift system load and network activity are denoted by the blue line and red line respectively, while the system load and network activity of our approach are denoted by the
green and magenta lines respectively. Figure 6 plots each line separately for site 0, allowing for increased clarity. A system load value of 1 means that the site is fully utilized,
while values higher than 1 means that the site is overloaded. A network activity value of
**Fig. 5 System load and network activity of all sites when executing the phylogenetic profiling workflow with**
the 64 genome input dataset for both our approach and Swift
-----
**Fig. 6 System load and network activity of site 0 when executing the phylogenetic profiling workflow with**
the 64 genome input dataset for both our approach and Swift
1 corresponds to utilization of 100 MBps. The network activity reported is both incoming and outgoing, so the maximum value it can reach is 2, which means 100 MBps of
incoming and outgoing traffic simultaneously, though this is difficult to achieve due to
network switch limitations.
Regarding our approach, the network traffic magenta line is barely visible, marking
only a few peaks, that coincide with drops in system load as denoted by the green line.
This is to be expected as network traffic takes place while downloading the input data of
the next OU pipeline and simultaneously uploading the output of the just processed OU
pipeline, during which the cpu is mostly inactive. It is apparent that the number of sections between the load drops are equal to the number of OU pipelines, 64 in this case.
Other than that, the system load is consistently at a value of 1.
In the Swift execution case, load values are slightly higher than 1 in all sites except site
1 which has 12 instead of 8 CPUs. This can be attributed to the slightly increased computational burden of submitting the jobs remotely and transferring inputs and outputs to
the main site. The internal scheduler of our approach operating on each site can be more
efficient. Network traffic is constant and on the low end for the duration of the workflow,
as data is transferred to and from the main site. However, near the end of the workflow,
system load drops and network traffic increases dramatically, especially on site 0 which
is the main site from which Swift operates and stages all file transfers to and from the
-----
other sites. As the computationally intensive part of most OU pipelines comes to an end,
the data intensive part then requires a high number of file transfers that overloads the
network and creates a bottleneck. This effect significantly slows down the makespan and
is mostly responsible for the increased execution times of Swift and the costly file transfers. In large workflows where the data to be moved is hundreds of GBs, it can even lead
to instability due to network errors.
**Conclusions and future work**
In this work, we presented a versatile framework for optimizing the parallel execution
of data-intensive bioinformatics workflows in hybrid cloud environments. The advantage of our approach is that it achieves surpassing time and cost efficiency than existing solutions through minimization of file transfers between sites. It accomplishes that
through the combination of a data management methodology that organizes the workflow into pipelines with minimal data interdependencies along with a scheduling policy
for mapping their execution into a set of heterogeneous distributed resources comprising a hybrid cloud.
Furthermore, we compared our methodology with Swift, a state of the art high performance framework and achieved superior cost and time efficiency in our use case
workflow. By minimizing file transfers, the total workflow execution time is reduced thus
leading to directly decreasing costs based on uptime of computational resources. Costs
can also decrease indirectly, as file transfers can be costly especially in hybrid clouds
where resources are not located within the facility of a single cloud vendor.
We are confident that our methodology can be applied to a wide range of bioinformatics workflows sharing similar characteristics with our use case study. We are currently
working on expanding our use case basis by implementing workflows in the fields of
metagenomics, comparative genomics, and haplotype analysis according to our methodology. Additionally, we are improving our load estimation functions so as to more accurately capture the computational load of a given pipeline through an evaluation of the
initial input.
In the era of Big data, cost-efficient high performance computing is proving to be
the only viable option for most scientific disciplines [14]. Bioinformatics is one of the
most representative fields in this area, as the data explosion has overwhelmed current
hardware capabilities. The rate at which new data is produced is expected to increase
significantly faster compared to the advances, and the cost, in hardware computational
capabilities. Data-aware optimization can be the a powerful weapon in our arsenal when
it comes to utilizing the flood of data to advance science and to provide new insights.
**Authors’ contributions**
AMK and FEP conceived and designed the study and drafted the manuscript. AMK implemented the platform as a
software solution. PAM participated in the project design and revision of the manuscript. AMK and FEP analyzed and
interpreted the results and coordinated the study. FEP edited the final version of the manuscript. All authors read and
approved the final manuscript.
**Acknowledgements**
This work used the European Grid Infrastructure (EGI) through the National Grid Infrastructure NGI_GRNET - HellasGRID.
We also thank Dr. Anagnostis Argiriou (INAB-CERTH) for access to their computational infrastructure.
**Competing interests**
The authors declare that they have no competing interests.
-----
Received: 17 August 2016 Accepted: 11 October 2016
**References**
1. Afgan E, Baker D, Coraor N, Chapman B, Nekrutenko A, Taylor J. Galaxy cloudman: delivering cloud compute clusters.
BMC Bioinform. 2010;11(Suppl 12):S4.
2. Altschul SF, Gish W, Miller W, Myers EW, Lipman DJ. Basic local alignment search tool. J Mol Biol. 1990;215(3):403–10.
3. Angiuoli SV, Matalka M, Gussman A, Galens K, Vangala M, Riley DR, Arze C, White JR, White O, Fricke WF. Clovr: a
virtual machine for automated and portable sequence analysis from the desktop using cloud computing. BMC
Bioinform. 2011;12(1):356.
4. Baker KR, Trietsch D. Principles of sequencing and scheduling. Hoboken: Wiley; 2013.
5. Berthold MR, Cebron N, Dill F, Gabriel TR, Kötter T, Meinl T, Ohl P, Sieb C, Thiel K, Wiswedel B. Knime: the konstanz
information miner. In: Data analysis, machine learning and applications. Berlin: Springer; 2008. p. 319–26
6. Bocchi E, Mellia M, Sarni S. Cloud storage service benchmarking: methodologies and experimentations. In: Cloud
networking (CloudNet), 2014 IEEE 3rd international conference on, IEEE; 2014. p. 395–400
7. Brandes U, Eiglsperger M, Herman I, Himsolt M, Marshall MS. Graphml progress report structural layer proposal. In:
Graph drawing. Berlin: Springer; 2001. p. 501–12
8. [Bux M, Leser U. Parallelization in scientific workflow management systems. 2013. arXiv preprint arXiv:1303.7195](http://arxiv.org/abs/1303.7195)
9. Chong Z, Ruan J, Wu CI. Rainbow: an integrated tool for efficient clustering and assembling rad-seq reads. Bioinformatics. 2012;28(21):2732–7.
10. Consortium U, et al. The universal protein resource (uniprot). Nucleic Acids Res. 2008;36(suppl 1):D190–5.
11. De Oliveira D, Ocaña KA, Ogasawara E, Dias J, Gonçalves J, Baião F, Mattoso M. Performance evaluation of parallel
strategies in public clouds: a study with phylogenomic workflows. Future Gener Comput Syst. 2013;29(7):1816–25.
12. Dean J, Ghemawat S. Mapreduce: simplified data processing on large clusters. Commun ACM. 2008;51(1):107–13.
13. Deelman E, Singh G, Su MH, Blythe J, Gil Y, Kesselman C, Mehta G, Vahi K, Berriman GB, Good J, et al. Pegasus: a
framework for mapping complex scientific workflows onto distributed systems. Sci Progr. 2005;13(3):219–37.
14. Duarte AM, Psomopoulos FE, Blanchet C, Bonvin AM, Corpas M, Franc A, Jimenez RC, de Lucas JM, Nyrönen T, Sipos
G, et al. Future opportunities and trends for e-infrastructures and life sciences: going beyond the grid to enable life
science data analysis. Front Genet. 2015:6.
15. Emeakaroha VC, Maurer M, Stern P, Łabaj PP, Brandic I, Kreil DP. Managing and optimizing bioinformatics workflows
for data analysis in clouds. J Grid Comput. 2013;11(3):407–28.
16. Gentleman RC, Carey VJ, Bates DM, Bolstad B, Dettling M, Dudoit S, Ellis B, Gautier L, Ge Y, Gentry J, et al. Bioconductor: open software development for computational biology and bioinformatics. Genome Biol. 2004;5(10):R80.
17. Goecks J, Nekrutenko A, Taylor J, et al. Galaxy: a comprehensive approach for supporting accessible, reproducible,
and transparent computational research in the life sciences. Genome Biol. 2010;11(8):R86.
18. Gurtowski J, Schatz MC, Langmead B. Genotyping in the cloud with crossbow. Curr Prot Bioinform. 2012:15–3.
19. Hill MD, Marty MR. Amdahl’s law in the multicore era. Computer. 2008;7:33–8.
20. Iosup A, Sonmez O, Anoep S, Epema D. The performance of bags-of-tasks in large-scale distributed systems. In: Proceedings of the 17th international symposium on high performance distributed computing. New York: ACM; 2008.
p. 97–108
21. Koukis V, Venetsanopoulos C, Koziris N. [∼] Okeanos: Building a cloud, cluster by cluster. IEEE Internet Comput.
2013;3:67–71.
22. Krampis K, Booth T, Chapman B, Tiwari B, Bicak M, Field D, Nelson KE. Cloud biolinux: pre-configured and ondemand bioinformatics computing for the genomics community. BMC Bioinform. 2012;13(1):42.
23. Litzkow MJ, Livny M, Mutka MW. Condor-a hunter of idle workstations. In: Distributed computing systems, 8th
international conference on, IEEE; 1988. p. 104–11.
24. Liu B, Madduri RK, Sotomayor B, Chard K, Lacinski L, Dave UJ, Li J, Liu C, Foster IT. Cloud-based bioinformatics workflow platform for large-scale next-generation sequencing analyses. J Biomed Inform. 2014;49:119–33.
25. Lobo I. Basic local alignment search tool (blast). Nature Educ. 2008;1(1):215.
26. Ludäscher B, Altintas I, Berkley C, Higgins D, Jaeger E, Jones MB, Lee EA, Tao J, Zhao Y. Scientific workflow management and the kepler system. Concurr Comput Pract Exp. 2006;18(10):1039–65.
27. Maheshwari K, Rodriguez A, Kelly D, Madduri R, Wozniak J, Wilde M, Foster I. Enabling multi-task computation on
galaxy-based gateways using swift. In: Cluster computing (CLUSTER), 2013 IEEE international conference on, IEEE;
2013. p. 1–3.
28. Merkel D. Docker: lightweight linux containers for consistent development and deployment. Linux J.
2014;2014(239):2.
29. Minevich G, Park DS, Blankenberg D, Poole RJ, Hobert O. Cloudmap: a cloud-based pipeline for analysis of mutant
genome sequences. Genetics. 2012;192(4):1249–69.
30. Moschakis IA, Karatza HD. Multi-criteria scheduling of bag-of-tasks applications on heterogeneous interlinked
clouds with simulated annealing. J Syst Soft. 2015;101:1–14.
31. Naccache SN, Federman S, Veeraraghavan N, Zaharia M, Lee D, Samayoa E, Bouquet J, Greninger AL, Luk KC, Enge
B, et al. A cloud-compatible bioinformatics pipeline for ultrarapid pathogen identification from next-generation
sequencing of clinical samples. Genome Res. 2014;24(7):1180–92.
32. Nagasaki H, Mochizuki T, Kodama Y, Saruhashi S, Morizaki S, Sugawara H, Ohyanagi H, Kurata N, Okubo K, Takagi T,
et al. Ddbj read annotation pipeline: a cloud computing-based pipeline for high-throughput analysis of next-generation sequencing data. DNA Res. 2013;dst017.
33. Ocaña KA, De Oliveira D, Dias J, Ogasawara E, Mattoso M. Designing a parallel cloud based comparative genomics
workflow to improve phylogenetic analyses. Future Gener Comput Syst. 2013;29(8):2205–19.
-----
34. Oinn T, Addis M, Ferris J, Marvin D, Senger M, Greenwood M, Carver T, Glover K, Pocock MR, Wipat A, et al. Taverna: a
tool for the composition and enactment of bioinformatics workflows. Bioinformatics. 2004;20(17):3045–54.
35. Pellegrini M, Marcotte EM, Thompson MJ, Eisenberg D, Yeates TO. Assigning protein functions by comparative
genome analysis: protein phylogenetic profiles. Proc Natl Acad Sci. 1999;96(8):4285–8.
36. Proost S, Van Bel M, Sterck L, Billiau K, Van Parys T, Van de Peer Y, Vandepoele K. Plaza: a comparative genomics
resource to study gene and genome evolution in plants. Plant Cell. 2009;21(12):3718–31.
37. Psomopoulos FE, Mitkas PA, Ouzounis CA, Promponas VJ, et al. Detection of genomic idiosyncrasies using fuzzy
phylogenetic profiles. PLoS One. 2013;8(1):e52854.
38. Reid JG, Carroll A, Veeraraghavan N, Dahdouli M, Sundquist A, English A, Bainbridge M, White S, Salerno W, Buhay C,
et al. Launching genomics into the cloud: deployment of mercury, a next generation sequence analysis pipeline.
BMC Bioinform. 2014;15(1):30.
39. Rice P, Longden I, Bleasby A, et al. Emboss: the European molecular biology open software suite. Trends Genet.
2000;16(6):276–7.
40. Schatz MC. Cloudburst: highly sensitive read mapping with mapreduce. Bioinformatics. 2009;25(11):1363–9.
41. Smith B, Grehan R, Yager T, Niemi D. Byte-unixbench: a unix benchmark suite. 2011.
42. Sreedharan VT, Schultheiss SJ, Jean G, Kahles A, Bohnert R, Drewe P, Mudrakarta P, Görnitz N, Zeller G, Rätsch G.
Oqtans: the rna-seq workbench in the cloud for complete and reproducible quantitative transcriptome analysis.
Bioinformatics. 2014:btt731.
43. Stajich JE, Block D, Boulez K, Brenner SE, Chervitz SA, Dagdigian C, Fuellen G, Gilbert JG, Korf I, Lapp H, et al. The
bioperl toolkit: Perl modules for the life sciences. Genome Res. 2002;12(10):1611–8.
44. Tang W, Wilkening J, Desai N, Gerlach W, Wilke A, Meyer F. A scalable data analysis platform for metagenomics. In: Big
data, 2013 IEEE international conference on, IEEE; 2013. p. 21–6.
45. Wall DP, Kudtarkar P, Fusaro VA, Pivovarov R, Patil P, Tonellato PJ. Cloud computing for comparative genomics. BMC
Bioinform. 2010;11(1):259.
46. Weng C, Lu X. Heuristic scheduling for bag-of-tasks applications in combination with qos in the computational grid.
Future Gener Comput Syst. 2005;21(2):271–80.
47. Wilde M, Hategan M, Wozniak JM, Clifford B, Katz DS, Foster I. Swift: a language for distributed parallel scripting.
Parallel Comput. 2011;37(9):633–52.
48. Wolstencroft K, Haines R, Fellows D, Williams A, Withers D, Owen S, Soiland-Reyes S, Dunlop I, Nenadic A, Fisher P,
et al. The taverna workflow suite: designing and executing workflows of web services on the desktop, web or in the
cloud. Nucleic Acids Res. 2013:gkt328.
49. Zhao Y, Hategan M, Clifford B, Foster I, Von Laszewski G, Nefedova V, Raicu I, Stef-Praun T, Wilde M. Swift: fast, reliable,
loosely coupled parallel computation. In: Services, 2007 IEEE Congress on, IEEE; 2007. p. 199–206.
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.1186/s40537-016-0055-2?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1186/s40537-016-0055-2, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "GOLD",
"url": "https://journalofbigdata.springeropen.com/counter/pdf/10.1186/s40537-016-0055-2"
}
| 2,016
|
[
"JournalArticle"
] | true
| 2016-10-21T00:00:00
|
[
{
"paperId": "1f18c130264b82562d1f5b93cdb7463f33d23028",
"title": "Principles of Sequencing and Scheduling"
},
{
"paperId": "67b4c7080ffc94ccc495a37aa6dcbe8a34302b05",
"title": "Future opportunities and trends for e-infrastructures and life sciences: going beyond the grid to enable life science data analysis"
},
{
"paperId": "ab15e4a5e569e000675883d0b2374c212f8b4420",
"title": "Multi-criteria scheduling of Bag-of-Tasks applications on heterogeneous interlinked clouds with simulated annealing"
},
{
"paperId": "bd0cf8ceaa409b1bde3b4223b2e897da8c467f95",
"title": "Cloud storage service benchmarking: Methodologies and experimentations"
},
{
"paperId": "d4d6bef7e3df07197a9882a89f140a394e3c6a5c",
"title": "A cloud-compatible bioinformatics pipeline for ultrarapid pathogen identification from next-generation sequencing of clinical samples"
},
{
"paperId": "e0a622bd6e897e8597d82b379486e36133e0838d",
"title": "Cloud-based bioinformatics workflow platform for large-scale next-generation sequencing analyses"
},
{
"paperId": "875d90d4f66b07f90687b27ab304e04a3f666fc2",
"title": "Docker: lightweight Linux containers for consistent development and deployment"
},
{
"paperId": "c4d05bf0062d648800465d13b65bb0d81b5f7b7b",
"title": "Launching genomics into the cloud: deployment of Mercury, a next generation sequence analysis pipeline"
},
{
"paperId": "d03293cdd2079ab23b881728f81aecb8d0a3c40b",
"title": "Oqtans: the RNA-seq workbench in the cloud for complete and reproducible quantitative transcriptome analysis"
},
{
"paperId": "7f302d785d779b5ebf34b76bc4dfb9150e0d2cd2",
"title": "A scalable data analysis platform for metagenomics"
},
{
"paperId": "98f769c1584c7fac8b89bd929a3856b6ecac7314",
"title": "Designing a parallel cloud based comparative genomics workflow to improve phylogenetic analyses"
},
{
"paperId": "37a878ddf4893b768580a7cc3d50a0adac2043e4",
"title": "Managing and Optimizing Bioinformatics Workflows for Data Analysis in Clouds"
},
{
"paperId": "735bbb3782fade94e053b46a179b82e53e06f063",
"title": "Enabling multi-task computation on Galaxy-based gateways using swift"
},
{
"paperId": "8ba0dd6ce4fa74ccce43034cdec7de5a3871ad2c",
"title": "Performance evaluation of parallel strategies in public clouds: A study with phylogenomic workflows"
},
{
"paperId": "07b62aa82b3cc693873e7cb4f0fa6501ad25ea3c",
"title": "DDBJ Read Annotation Pipeline: A Cloud Computing-Based Pipeline for High-Throughput Analysis of Next-Generation Sequencing Data"
},
{
"paperId": "a0b789d1f3afe9cf6fadfaa8f120479181f0c4e5",
"title": "The Taverna workflow suite: designing and executing workflows of Web Services on the desktop, web or in the cloud"
},
{
"paperId": "99c40f08918f70aab65ef62fe0f2531f9f33ca93",
"title": "~okeanos: Building a Cloud, Cluster by Cluster"
},
{
"paperId": "f5c0924d91e8e8b9a6ea6cae8f73e604b39b407b",
"title": "Parallelization in Scientific Workflow Management Systems"
},
{
"paperId": "b5eddfe356dc226e146df0c63dac55846b83610f",
"title": "Detection of Genomic Idiosyncrasies Using Fuzzy Phylogenetic Profiles"
},
{
"paperId": "5a4923537633e6e3621203f3807d736d534a5fea",
"title": "CloudMap: A Cloud-Based Pipeline for Analysis of Mutant Genome Sequences"
},
{
"paperId": "2d5376233101455b91772aeea1e2d84320588dcb",
"title": "Rainbow: an integrated tool for efficient clustering and assembling RAD-seq reads"
},
{
"paperId": "0e231e08c29455a8cc1bb3a82f10a5b4bf308c39",
"title": "Genotyping in the Cloud with Crossbow"
},
{
"paperId": "8f262d37490e1393123d2b3459f73dd06f27159a",
"title": "Cloud BioLinux: pre-configured and on-demand bioinformatics computing for the genomics community"
},
{
"paperId": "9ca2a17fa3669323dbcb367e6c70ff9acb969b6d",
"title": "Swift: A language for distributed parallel scripting"
},
{
"paperId": "7a199a0c579160a93157f3d6975ded8aeb295fd9",
"title": "CloVR: A virtual machine for automated and portable sequence analysis from the desktop using cloud computing"
},
{
"paperId": "83cc2ae159eea5f4174a7974a631b7aa88d00a7c",
"title": "Galaxy CloudMan: delivering cloud compute clusters"
},
{
"paperId": "ebd342793ee80999119cf163f5774d649fc0c928",
"title": "Galaxy: a comprehensive approach for supporting accessible, reproducible, and transparent computational research in the life sciences"
},
{
"paperId": "252b1b24dae7d8ba553fe4eb9c659aabe53ccdba",
"title": "Cloud computing for comparative genomics"
},
{
"paperId": "63f4c96346eed2fc43dfdc34c6e4f366bca127fe",
"title": "PLAZA: A Comparative Genomics Resource to Study Gene and Genome Evolution in Plants[W]"
},
{
"paperId": "540a803d3d99f3b7814f37652c056718c701b5c5",
"title": "CloudBurst: highly sensitive read mapping with MapReduce"
},
{
"paperId": "3a02beb0ffb573c31a6be82565f62ba7abe98bc7",
"title": "Amdahl's Law in the Multicore Era"
},
{
"paperId": "ec46be6dd0fa93893a804396cfa62b4d6371339e",
"title": "The performance of bags-of-tasks in large-scale distributed systems"
},
{
"paperId": "7fd2213b834a5e745deecc0ee375e5be05fd1b2f",
"title": "Swift: Fast, Reliable, Loosely Coupled Parallel Computation"
},
{
"paperId": "0b4b0d518f6c1f814b6e91ac110495f2264d1d7e",
"title": "Taverna: lessons in creating a workflow environment for the life sciences"
},
{
"paperId": "2022160699559f695d4007e6328cc01261d3beed",
"title": "Pegasus: A framework for mapping complex scientific workflows onto distributed systems"
},
{
"paperId": "6f10e2b58bf58286f3f961cbda878370309d7c29",
"title": "Heuristic scheduling for bag-of-tasks applications in combination with QoS in the computational grid"
},
{
"paperId": "7f1b4ef3351041755e40cab188748af613949f4e",
"title": "Basic Local Alignment Search Tool (BLAST)"
},
{
"paperId": "3d9fbcf35f53bd84c75fd99daa6b2c69397b0a01",
"title": "The Universal Protein Resource (UniProt)"
},
{
"paperId": "36789799e464aa7465125ff8e778939843a0e89b",
"title": "Taverna: a tool for the composition and enactment of bioinformatics workflows"
},
{
"paperId": "fd495d6cf7c3169bc58550fdf32be6e16e2800f8",
"title": "Bioconductor: open software development for computational biology and bioinformatics"
},
{
"paperId": "33780e4aba639a97f9fb7f7e773853f74dd494b7",
"title": "The Bioperl toolkit: Perl modules for the life sciences."
},
{
"paperId": "ddf06cf0d375fb9404fe30c5f1d7858d74080e9c",
"title": "EMBOSS: the European Molecular Biology Open Software Suite."
},
{
"paperId": "adea14ce6f45a439d3480bf2f4c030b77ea89102",
"title": "Assigning protein functions by comparative genome analysis: protein phylogenetic profiles."
},
{
"paperId": "a03dc4d8aa1c2ebd5c3a3c1d87d531f52d63c120",
"title": "Basic local alignment search tool."
},
{
"paperId": "ea6b2281bab9dd7efbc2ad6b95492a5263861cc9",
"title": "Condor-a hunter of idle workstations"
},
{
"paperId": "5c632485f98035bd6c8c263af8193cc9ad1358da",
"title": "The use of microprocessors as automobile on-board controlers"
},
{
"paperId": null,
"title": "Byte-unixbench: a unix benchmark suite"
},
{
"paperId": "627be67feb084f1266cfc36e5aed3c3e7e6ce5f0",
"title": "MapReduce: simplified data processing on large clusters"
},
{
"paperId": "f199791a6381fd8eacff37a00c8f7ab195bbdb64",
"title": "KNIME: The Konstanz Information Miner"
},
{
"paperId": "06429adb228b2241f2854819b6c9963120da6a48",
"title": "GraphML Progress Report ? Structural Layer Proposal"
}
] | 17,230
|
en
|
[
{
"category": "Engineering",
"source": "external"
},
{
"category": "Engineering",
"source": "s2-fos-model"
},
{
"category": "Environmental Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/027bd79f0d6c728825aaac3fa41f6178e8b30145
|
[
"Engineering"
] | 0.860423
|
Decentralized Blended Acquisition
|
027bd79f0d6c728825aaac3fa41f6178e8b30145
|
[
{
"authorId": "145271273",
"name": "G. Berkhout"
}
] |
{
"alternate_issns": null,
"alternate_names": null,
"alternate_urls": null,
"id": null,
"issn": null,
"name": null,
"type": null,
"url": null
}
| null |
### Decentralized Blended Acquisition
#### Guus Berkhout, Delft University of Technology
**SUMMARY**
The concept of blending and deblending is reviewed, making
use of traditional and dispersed source arrays. The network
concept of distributed blended acquisition is introduced. A
million-trace robot system is proposed, illustrating that decentralization may bring about a revolution in the way we acquire
seismic data in the future.
**INTRODUCTION**
In traditional seismic surveys, interference between shot records
is minimized by choosing the temporal interval and/or the lateral distance between consecutive shots sufficiently large. However, in the concept of simultaneous shooting shot records do
overlap, allowing denser source sampling in a favorable economic way. Denser source sampling takes care of the desired
property that each subsurface gridpoint is illuminated from a
larger number of angles and, therefore, will improve the image
quality in terms of signal-to-noise ratio and spatial resolution.
In the seismic literature, already an abundance of references
on simultaneous shooting can be found. Examples of recent
publications are Beasley (2008), Berkhout (2008), Howe et al.
(2008), Pecholcs et al. (2010), Berkhout et al. (2012), Beasley
et al. (2012), Abma et al. (2012), Krupovnickas et al. (2012).
In blended acquisition, being a special version of simultaneous shooting, the ‘simultaneous’ source wavefield is incoherent (see Figure 1).
objective of blended acquisition is to maximize the emission
of full-bandwidth, non-aliased, far-field signal energy within a
pre-specified acquisition time.
In traditional seismic surveys a single coherent source (array)
is used for each shot record. This localized source unit must
transmit the full temporal frequency band for a wide range of
emission angles. Today’s seismic vibrators and airgun arrays
are designed such that they have a large bandwidth, ranging
over many octaves. In practice, however, such source designs
are a compromise from a systems engineering point of view.
I propose that the individual source units in a blended array
(1) are not chosen to be equal and (2) do not need to satisfy
the wide-band requirements. Instead, they may be dedicated
narrowband designs with superior emission properties around
their central frequency. The ultimate criterion is that the combined incoherent source wavefield has the required temporal
and angular spectral properties at each gridpoint in the subsurface. In addition, I propose that the traditional centralized
concept in seismic acquisition is replaced by a decentralized
network alternative.
**THEORETICAL CONSIDERATIONS**
Seismic data can be arranged in data matrix P. In the frequency domain P represents a frequency slice of the total data
volume and one element Pij is one frequency component of
the trace measured at detector position i generated by source
_j. In my notation P(zd, zs) means that the source and detector_
positions are situated at depth levels zs and zd respectively. If
we choose for the moment zs = zd = z0 (typical for land data),
then the model of data matrix P can be written as (Berkhout,
1982):
**P(z0, z0) = D(z0)X(z0, z0)S[+](z0),** (1)
where matrix X is the Earth’s transfer operator that includes
the interaction with the surface. In source matrix S[+](z0) each
column represents a (directional) source. In detector matrix D
each row represents a receiver (array). The response of each
source column (S[⃗]j[+][) is given by the corresponding column of]
the data matrix (P[⃗]j).
Using expression 1, the result of one blended experiment can
be formulated by (Berkhout, 2008):
**P(z0, z0)[⃗]Γj** (z0) = D(z0)X(z0, z0)S[+](z0)[⃗]Γj (z0). (2a)
Figure 1: Subdivision of simultaneous shooting methods,
based on the degree of incoherency.
Such an incoherent wavefield is physically generated by firing
a multitude of sources, each source with its own code (such
as temporal delay, nonlinear phase function, pseudo-random
time series), together forming a blended source array. Unlike
a traditional source array, a blended source array may cover a
large spatial area, meaning that one blended source array illuminates subsurface gridpoints from many different angles. The
Column vector _[⃗]Γj(z0) contains the blending information. This_
is illustrated in Figure 2: elements Γkj (z0) are complex-valued
scalars, describing time delays or a more complex code, while
the involved sources are indicated by the positions (k) of the
scalars in column vector _[⃗]Γj_ (z0). Note that equation 2a is based
on the linearity of seismic data in wavefields. This can be eas
-----
�
_Sk[�]�kj_
#### Decentralized Blended Acquisition
**DESCRIPTION OF DEBLENDING ALGORITHM**
_z0_
### S[�]�j
blending code
### � � � Sk[�]�kj
_k_
_[x][�]s_
### S[�]�j
|Col1|Col2|
|---|---|
one unit of a
blended source array
includes classical
field array
_k_
### S� �[�] j
_k_
blended source array
(downward radiating)
Figure 2: One blended source array consists of a multitude of
source units, each unit having its own code.
ily seen if we rewrite this equation as follows:
� _⃗Pk(z0, z0)Γkj_ (z0) = D(z0)X(z0, z0) � _⃗Sk[+][(][z][0][)Γ][kj][(][z][0][)][,]_
_k_ _k_
(2b)
showing that the weighted sources of the blended source array generate a weighted set of shot records, the latter being
referred to as a blended shot record. Equation 2b can be made
specific for marine data by showing explicitly the ghost effect.
If we allow the individual elements (k) of a blended source
array to be at different depth levels (zk), then we may write:
� _⃗Pk(z0, zk)Γkj(zk) = D(z0)X(z0, z0)_ � _⃗Sk[+][(][z][0][, z][k][)Γ][kj][(][z][k][)]_
_k_ _k_
(3a)
where, assuming a surface reflectivity of -1,
In deblending, blended measurements are given and unblended
data need be computed (inversion process). In this closed-loop
process, numerically simulated measurements - output of forward modeling according to equations 2a and 2b - are compared with the real measurements. By minimizing the difference between the two datasets the unblended samples (parameters) can be estimated. To explain this inversion process, let us
minimize the following unconstrained least-squares criterion
(zd and zs are omitted for notational convenience):
2 2
Δ⃗Pj ′ = _⃗Pj ′_ _._ (5a)
��� ��� ��� _[−]_ **[P][⃗][Γ][j]** ���
Bear in mind that in minimization equation 5a
where _P[⃗]k[(][i][)]_ in 6 is approaching _P[⃗]k in 7 asymptotically._
In the first iteration (i = 1) ΔP[′] = P[′], meaning that the inversion process starts with pseudo-deblending. It is interesting
to realize that Λ may be a scaled unity matrix or a diagonal matrix or a bandmatrix, depending on the properties of blending
matrix Γ. During the presentation properties of the algorithm
will be illustrated with examples. The computational diagram
is shown in Figure 3.
�
**P[⃗]Γj =** _⃗PkΓkj_ (5b)
_k_
represents the modeling output and vector _P[⃗]k equals the de-_
blended shot record for shot k. The iterative solution of minimization problem 5a is given by:
_⃗P_ [(]k[i][)] = _P[⃗]_ [(]k[i][−][1)] + �ΔP[′][�][i][−][1]Λ[⃗]Γ[H]k _[,]_ (6)
where diagonal matrix Λ contains the weights.
The validity of iterative, weighted, least-squares solution 6 can
be quickly verified by substituting the expression of ΔP[′] in
equation 6, leading to the well-known analytic equation:
� �
**P[′]Λ[⃗]Γ[H]k** [=][ P] ΓΛ[⃗]Γ[H]k _,_ (7)
_⃗Sk[+][(][z][0][, z][k][) =][ W][∗][(][z][0][, z][k][)][⃗S]k[+][(][z][k][)][ −]_ **[W][(][z][0][, z][k][)][⃗S]k[−][(][z][k][)][.]**
(3b)
In equation 3b matrix W(z0, zk) describes the propagation between source depth zk and surface level z0 and superscript *
denotes the complex conjugate. Note that the incident wavefield in gridpoint i at depth level zm, being generated by blended
source array j at the surface z0, is given by:
P[+]ij [(][z][m][, z][0][) =][ ⃗W][ †]i [(][z][m][, z][0][)][S][+][(][z][0][)][⃗][Γ][j] [(][z][0][)][.] (4)
Here, _W[⃗]_ _i[†]_ [describes wavefield propagation from all source ar-]
ray points at surface level z0 to subsurface gridpoint i at depth
level zm.
From the foregoing it follows that blended acquisition has two
important advantages: (1) the number of source points per km[2]
is increased and (2) the survey time per km[2] may be decreased.
Both aspects refer to data quality: more signal energy per unit
area and unit time is transmitted into the subsurface (less spatial aliasing and larger signal to background noise ratio). The
second aspect also refers to economics. Particularly in special
situations, think of areas where access is restricted to a limited period of time, blending may be the only solution that is
practically feasible.
**P�** subtraction adaptive (�P�)i�1 estimation parameter _[P]�j_ (i�1) parameter selection **P**
|Col1|adaptive subtraction|
|---|---|
([P][�]j�)[i]�1
i+1 � (i�1)
_[P]j_
|P(i1) j parameter selection G P(i1) j|parameter selection|Col3|
|---|---|---|
�
Figure 3: Computational diagram of deblending in terms of inversion, showing the four principal algorithmic modules (estimation, selection, modeling and subtraction) in each iteration.
-----
#### Decentralized Blended Acquisition
**DISPERSED SOURCE ARRAYS**
For the design of blended source arrays, the individual sources
at surface locations k (S[⃗]k[+][Γ][kj] [), see equation 4, need to be opti-]
mized by considering the properties of the composite incident
wavefield at subsurface locations i (Pij[+][). It means that the in-]
dividual sources of a blended array may consist of narrowband
sources with different central frequencies (‘components’), as
long as the sum of all arriving components (‘composite result’)
satisfies the full bandwidth requirements.
According to the Nyquist criterion, the ideal source spacing
should be smaller than half the smallest wavelength a source
transmits. In case of different source types, e.g., low-, mid- and
high-frequency sources, it means that each type has its own optimum spacing. Note that this is largest for the low-frequency
sources and smallest for the high-frequency sources! I call this
type of blended source configuration: Dispersed Source Array
(DSA).
It is important to realize that a DSA acts like a modern audio surround system: the different loudspeaker units are decentralized, taking care of the different sub-bands within the
total audio frequency range. This subdivision leads to entirely
different loudspeaker designs for the low, mid and high frequencies (see Figure 4). The audio-seismic comparison highlights the fundamental difference of the DSA concept with systems such as Polychromatic Acquisition (CREWES consortium) and SeisMovie (Meunier et al., 2001), where broadband
source units operate in a multi-monochromatic manner.
detector side, showing excellent results (Soubaras, 2010). Combining the two is the way to go.
**DECENTRALIZED BLENDED ACQUISITION**
Based on the blending method and the DSA concept, it is proposed to make another fundamental improvement in seismic
data acquisition. This improvement is achieved by changing
the system architecture. I propose to focus future acquisition
developments on the major opportunities that are offered by
the decentralized network architecture. By moving from a sin_gle complex, centralized system to a network of simple, decen-_
tralized subsystems, more information is collected with less
complexity.
Decentralization is the major change we have seen in many
technological solutions during the last decade; particularly think
of information, communication and computation systems in
the IC-sector. Central systems have been transformed to networks, increasing the capability and efficiency beyond expectation. Figure 5 visualizes two system architectures. Figure
a. centralized network (N=5) b. decentralized network (N[2]=25)
Figure 5: Two types of system architectures. Until today, seismic acquisition occurs with a centralized architecture (a).
1. ONE BROADBAND 2. DIFFERENT NARROWBAND 3. DIFFERENT DISTRIBUTED
SOURCE SOURCES NARROWBAND SOURCES
Figure 4: Application of the DSA concept in broadband high
performance audio systems. Note the significantly different
designs for the different frequency bands.
Inhomogeneous blending with DSAs has a number of attractive potential advantages: (1) the dedicated narrowband units
of a blended array represent technically simple, no-compromise
source units, (2) destructive interference within a source array
is avoided, allowing angle-independent source wavelets, (3)
each source type has its own spatial sampling interval, allowing multi-scale acquisition grids, (4) each source type has its
own depth level, allowing ghost matching in the field (marine),
(5) deblending DSA data is relatively simple: the first step
(source decoding + bandpass filtering) is already very effective, (6) DSAs are more flexible to comply with the emerging
strict regulation on sea life protection (marine).
It is interesting to mention here that the advantages of multilevel depth sources were already demonstrated in a EAGE workshop on marine seismic in Cyprus (Cambois and Osnes, 2009).
Recently, the variable depth option was also proposed at the
5a shows schematically a conventional broadcast architecture,
allowing N one-way connections from the central source subsystem to the N receiver subsystems. Hence, with this architecture the information received increases linear with N . Figure 5b shows a decentralized network architecture, where every element functions both as a source and receiver subsystem.
Now there exist N [2] connections in the network, meaning that
the information received increases quadratically with N, see
Figure 6.
decentralized N[2 ]
offsets &
centralized N
azimuths
1 _N_
Figure 6: The difference in information content between a centralized and decentralized system.
-----
#### Decentralized Blended Acquisition
If we look at the current seismic acquisition systems, then
we may conclude that the industry makes use of the so-called
broadcast architecture: one seismic source (array) sends its energy - via the Earth - to the N seismic detectors. In the past
decades we have seen that the number of detectors have been
continuously increased to as much as 100.000 and further increases are in progress. This has increased the complexity of
the acquisition system tremendously. Actually, current seismic
systems are great technological achievements.
I propose to the industry to abandon the centralized acquisition
concept: the linear relationship is not an attractive proposition.
Instead, it is proposed to concentrate on the exciting opportunities that are offered by the network architecture. For example,
if we use an acquisition network with a swarm of 100 simple
source-detector subsystems, where each subsystem consists of
a DSA robot dragging one short 100-detector cable, then the
total number of traces per blended shot record equals one million traces (100x100[2] )! Figure 7 gives an artist impression of
such a network.
Figure 7: Artist impression of a distributed seismic acquisition network. Each robot consists of an optimized narrowband
source and a small detector array, e.g., with 100 receivers only.
A swarm of one hundred of these robots configure a one million trace system.
concentrate future developments on the network architecture
concept, showing a quadratic behavior in seismic information
(N [2]).
By moving from a single complex, centralized system to a
network of simple, decentralized subsystems, robotization becomes an attractive proposition: a one million channel system
can be realized by a small number of simple source-detector
robots.
**FINAL REMARK**
Berkhout and Blacquiere (2012) conclude that the signal to
background-noise ratio of a field-blended survey must be higher
than of a comparable traditional survey. This is because the
power of the signal (total signal energy divided by the effective survey time) increases in blended acquisition, not only because the number of sources increases, but also due to the fact
that the survey time may decrease. On the other hand, the
power of the background noise is independent of whatever we
do in the blending process. Hence, a shorter recording time not
only favors economics, it also favors quality, particularly in
areas with a high background noise level. This conclusion emphasizes the enormous potential of blended acquisition for the
industry. As a consequence, I expect that unblended seismic
acquisition will become a technology of the past.
**ACKNOWLEGDMENT**
I would like to acknowledge the sponsors of the Delphi consortium at Delft University of Technology for the stimulating
discussions on robotized blended acquisition and I also want
to thank them for their financial support.
**CONCLUSIONS**
With a multitude of dedicated narrow-band source units, being
referred to as Dispersed Source Arrays, the blended incident
wavefield at a particular subsurface gridpoint contains broadband, multi-angle, multi-azimuth information. The theoretical
spatial sampling requirements can be fulfilled by allowing lowfrequency sources to be distributed more sparsely than highfrequency sources (‘multi-scale shooting grids’). In the marine
case source depths can be optimized (‘ghost matching’).
It is also proposed to rethink the centralized acquisition concept. Instead, I propose to concentrate future developments on
the network architecture concept, where information collection is linear in the number of detectors (N ). A plea is made to
-----
http://dx.doi.org/10.1190/segam2013-0845.1
**EDITED REFERENCES**
Note: This reference list is a copy-edited version of the reference list submitted by the author. Reference lists for the 2013
SEG Technical Program Expanded Abstracts have been copy edited so that references provided with the online metadata for
each paper will achieve a high degree of linking to cited sources that appear on the Web.
**REFERENCES**
### Abma, R., Zhang, Q., Arogunmati, A., and Beaudoin, G., 2012, An overview of BP’s marine
## independent simultaneous source field trials: 82nd Annual International Meeting, SEG, Expanded Abstracts, http://dx.doi.org/10.1190/segam2012-1404.1.
### Beasley, C. J., B. Dragoset, and A. Salama, 2012, A 3D simultaneous source field test processed using alternating projections: A new active separation method: Geophysical Prospecting, 60, 591–601, http://dx.doi.org/10.1111/j.1365-2478.2011.01038.x.
Beasley, C. J., 2008, A new look at marine simultaneous sources: The Leading Edge, 27, 914–917,
http://dx.doi.org/10.1190/1.2954033.
### Berkhout, J., and G. Blacquière, 2012, Utilizing dispersed source arrays in blended acquisition: 82nd
## Annual International Meeting, SEG, Expanded Abstracts, http://dx.doi.org/10.1190/segam2012-0302.1.
### Berkhout, A., D. Verschuur, and G. Blacquière, 2012, Illumination properties and imaging promises of blended, multiple-scattering seismic data: A tutorial: Geophysical Prospecting, 60, 713–732, http://dx.doi.org/10.1111/j.1365-2478.2012.01081.x.
Berkhout, A. J., 1982, Seismic migration, imaging of acoustic energy by wave field extrapolation. Part A: Theoretical aspects: Elsevier.
Berkhout, A. J., 2008, Changing the mindset in seismic data acquisition: The Leading Edge, 27, 924–938,
http://dx.doi.org/10.1190/1.2954035.
### Doulgeris, P., K. Bube, G. Hampson, and G. Blacquière, 2012, Convergence analysis of a coherency- constrained inversion for the separation of blended data: Geophysical Prospecting, 60, 769–781, http://dx.doi.org/10.1111/j.1365-2478.2012.01088.x.
Howe, D., M. Foster, T. Allen, B. Taylor, and I. Jack, 2008, Independent simultaneous sweeping — A method to increase the productivity of land seismic crews: 78th Annual International Meeting,
## SEG, Expanded Abstracts, 2826–2830, http://dx.doi.org/10.1190/1.3063932.
### Krupovnickas, T., K. Matson, C. Corcoran, and R. Pascual, 2012, Marine simultaneous source OBS survey suitability for 4D analysis : 82nd Annual International Meeting, SEG, Expanded
## Abstracts, http://dx.doi.org/10.1190/segam2012-0815.1.
### Meunier, J., F. Huguet, and P. Meynier, 2001, Reservoir monitoring using permanent sources and vertical receiver antennae: The Céré-la-Ronde case study: The Leading Edge, 20, 622–629, http://dx.doi.org/10.1190/1.1439008.
Pecholcs, P. I., S. K. Lafon, T. Al-Ghamdi, H. Al-Shammery, P. G. Kelamis, S. X. Huo, O. Winter, J.-B. Kerboul, and T. Klein, 2010, Over 40,000 vibrator points per day with real-time quality control: Opportunities and challenges: 80th Annual International Meeting, SEG, Expanded Abstracts, 111–115, http://dx.doi.org/10.1190/1.3513041.
-----
|
{
"disclaimer": "Notice: The following paper fields have been elided by the publisher: {'abstract', 'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1190/SEGAM2013-0845.1?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1190/SEGAM2013-0845.1, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "GREEN",
"url": "https://repository.tudelft.nl/file/File_86897a5b-1732-40b7-a9d6-171b42f50f7f"
}
| 2,013
|
[
"Review"
] | true
| 2013-01-10T00:00:00
|
[] | 5,618
|
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/0287942a57191b829f0a68658bb43646c7d5160d
|
[
"Computer Science"
] | 0.834799
|
Enforcing Security and Assurance Properties in Cloud Environment
|
0287942a57191b829f0a68658bb43646c7d5160d
|
International Conference on Utility and Cloud Computing
|
[
{
"authorId": "31559735",
"name": "Aline Bousquet"
},
{
"authorId": "1885414",
"name": "Jérémy Briffaut"
},
{
"authorId": "1803718",
"name": "E. Caron"
},
{
"authorId": "145634347",
"name": "E. M. Domínguez"
},
{
"authorId": "2064925777",
"name": "Javier Franco"
},
{
"authorId": "2018068",
"name": "Arnaud Lefray"
},
{
"authorId": "143875640",
"name": "Ó. López"
},
{
"authorId": "2066752310",
"name": "Saioa Ros"
},
{
"authorId": "1403816047",
"name": "Jonathan Rouzaud-Cornabas"
},
{
"authorId": "2208354",
"name": "C. Toinard"
},
{
"authorId": "3185042",
"name": "Mikel Uriarte"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"IEEE/ACM Int Conf Util Cloud Comput",
"Utility and Cloud Computing",
"Int Conf Util Cloud Comput",
"Util Cloud Comput",
"UCC",
"IEEE/ACM International Conference Utility and Cloud Computing"
],
"alternate_urls": null,
"id": "d03a5bfe-db75-4dc9-95f3-ae92b081f42c",
"issn": null,
"name": "International Conference on Utility and Cloud Computing",
"type": "conference",
"url": null
}
|
Before deploying their infrastructure (resources, data, communications, ) on a Cloud computing platform, companies want to be sure that it will be properly secured. At deployment time, the company provides a security policy describing its security requirements through a set of properties. Once its infrastructure deployed, the company want to be assured that this policy is applied and enforced. But describing and enforcing security properties and getting strong evidences of it is a complex task. To address this issue, in [1], we have proposed a language that can be used to express both security and assurance properties on distributed resources. Then, we have shown how these global properties can be cut into a set of properties to be enforced locally. In this paper, we show how these local properties can be used to automatically configure security mechanisms. Our language is context-based which allows it to be easily adapted to any resource naming systems e.g., Linux and Android (with SELinux) or PostgreSQL. Moreover, by abstracting low-level functionalities (e.g., deny write to a file) through capabilities, our language remains independent from the security mechanisms. These capabilities can then be combined into security and assurance properties in order to provide high-level functionalities, such as confidentiality or integrity. Furthermore, we propose a global architecture that receives these properties and automatically configures the security and assurance mechanisms accordingly. Finally, we express the security and assurance policies of an industrial environment for a commercialized product and show how its security is enforced.
|
## Enforcing Security and Assurance Properties in Cloud Environment
### Aline Bousquet, Jérémy Briffaut, Eddy Caron, Eva María Dominguez, Javier
Franco, Arnaud Lefray, Oscar López, Saioa Ros, Jonathan Rouzaud-Cornabas,
Christian Toinard, et al.
To cite this version:
#### Aline Bousquet, Jérémy Briffaut, Eddy Caron, Eva María Dominguez, Javier Franco, et al.. Enforcing Security and Assurance Properties in Cloud Environment. 8th IEEE/ACM International Conference on Utility and Cloud Computing (UCC 2015), University of Cyprus, Dec 2015, Limassol, Cyprus. hal-01240557
### HAL Id: hal-01240557
https://inria.hal.science/hal-01240557
#### Submitted on 9 Dec 2015
#### HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
#### L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
#### Distributed under a Creative Commons Attribution NonCommercial NoDerivatives| 4 0
-----
# Enforcing Security and Assurance Properties in Cloud Environment
#### Aline Bousquet[∗], J´er´emy Briffaut[∗], Eddy Caron[†], Eva Mar´ıa Dominguez[¶], Javier Franco[‡], Arnaud Lefray[∗][†], Oscar L´opez[§], Saioa Ros[§], Jonathan Rouzaud-Cornabas[∥], Christian Toinard[∗] and Mikel Uriarte[§]
_∗INSA Centre Val de Loire, Univ. Orl´eans, LIFO EA 4022, Bourges, France_
†University of Lyon - LIP, CNRS - ENS Lyon - Inria - UCB Lyon, France
‡Industry and Advanced Manufacturing department, Vicomtech-IK4, Spain
§Research and Development department, Nextel S.A., Spain
- ransport - Technology and Development, IKUSI, Spain
_∥Universiy of Lyon, CNRS, Inria, INSA-Lyon, LIRIS, UMR5205, F-69621, France_
Email: aline.bousquet, jeremy.briffaut, christian.toinard @insa-cvl.fr, eddy.caron, arnaud.lefray @ens-lyon.fr,
_{_ _}_ _{_ _}_
jonathan.rouzaud-cornabas@inria.fr, jfranco@vicomtech.org, sros, olopez, muriarte @nextel.es, eva.dominguez@ikusi.com
_{_ _}_
**_Abstract—Before deploying their infrastructure (resources,_**
**data, communications, ...) on a Cloud computing platform,**
**companies want to be sure that it will be properly secured.**
**At deployment time, the company provides a security policy**
**describing its security requirements through a set of properties.**
**Once its infrastructure deployed, the company want to be assured**
**that this policy is applied and enforced. But describing and**
**enforcing security properties and getting strong evidences of it is**
**a complex task.**
**To address this issue, in [1], we have proposed a language that**
**can be used to express both security and assurance properties**
**on distributed resources. Then, we have shown how these global**
**properties can be cut into a set of properties to be enforced**
**locally. In this paper, we show how these local properties can**
**be used to automatically configure security mechanisms. Our**
**language is context-based which allows it to be easily adapted**
**to any resource naming systems e.g., Linux and Android (with**
**SELinux) or PostgreSQL. Moreover, by abstracting low-level**
**functionalities (e.g., deny write to a file) through capabilities, our**
**language remains independent from the security mechanisms.**
**These capabilities can then be combined into security and assur-**
**ance properties in order to provide high-level functionalities, such**
**as confidentiality or integrity. Furthermore, we propose a global**
**architecture that receives these properties and automatically**
**configures the security and assurance mechanisms accordingly.**
**Finally, we express the security and assurance policies of an**
**industrial environment for a commercialized product and show**
**how its security is enforced.**
**_Keywords—Security, Cloud, Assurance, Enforcement, Use-case_**
I. INTRODUCTION
In security, three main concepts commonly known as
the CIA-triad (not to be confused with the US agency) has
been widely used for decades: Confidentiality, Integrity and
Availability. Both the Departement Of Defense guidelines
(TCSEC/Orange Book) [2] edited in 1985 and the more recent Common Criteria (ISO/IEC 15408) international standard
define security as an integration of availability, confidentiality
and integrity.
In a survey [3] on Cloud adoption practices, the Cloud
Security Alliance (CSA) indicates that 73% of the participating
industries are concerned about the security of their data. Thus,
while many companies are transitioning to Cloud computing,
they are also worried about the security risk. But Cloud
platforms lack of reliable security [4]. Furthermore, Halpern
et al. [5] state that security policies described in a natural
language have quite ambiguous semantics. To answer these
problems, we need to provide a way (a language) to let the
Cloud tenants (e.g., companies) express their security requirements i.e., through a security policy. This security policy must
them be enforced on the Cloud platform and assurance reports
(i.e., proofs) of this enforcement must be given to the tenants.
A single security mechanism cannot protect a heterogeneous and multi-layer system such as a Cloud [6]. Consequently, it is a set of uncoupled (and already existing)
mechanisms that will be used to enforce the security. However,
even if the mechanisms are uncoupled, it is mandatory to
carefully take into account their capabilities (i.e., what they are
able to enforce) and configure all of them at once to provide
the wanted security. But, each of these mechanisms also comes
with its own configuration language.
In [1], we have defined a specification language for global
security properties (i.e., properties that involve distributed resources). We have shown how these global properties can be
automatically cut into a set of local properties. These local
properties can be used to automatically configure security
mechanisms. Moreover, our common independent language
abstracting low-level capabilities can be used to provide proofs
of security enforcement (i.e., assurance).
As said previously, the tenants also require to receive
a proof that their security is indeed enforced during the
whole life cycle of the infrastructure. Accordingly, using our
language, the tenants can expressed their security assurance
requirements.
Once the security and assurance policies has described a set
of properties to enforce, an architecture is required to automatically configure distributed and heterogeneous mechanisms.
Furthermore, this architecture must also send back assurance
reports to the tenants.
To show the usability and capacities of our solution, we
-----
describe how our language has been used to define the security
policy of a complete industrial application. Then, we show how
our architecture is used to automatically enforce the policy and
generate assurance reports of it.
This paper is organized as follows. In Section II, we present
a set of existing security mechanisms that could be used to
provide security in Clouds, and the related work around Cloud
security and assurance. Section III describes the language and
the architecture we use for the security policy enforcement
and assurance. Section IV details an industrial use-case and
the whole process to secure it, and Section V concludes this
paper.
II. RELATED WORK
The solution proposed in this paper aims to both enforce
and assure security properties, such as confidentiality or integrity. Hence, this section first describes the works related
to the definition of security properties and their enforcement.
Then, we quickly present the security mechanisms we will use
in Section IV. Finally, we present some existing solutions for
assurance.
_A. Security Policy and Enforcement_
Because of the ever increasing adoption of Cloud Computing platforms, many researches have been done to improve
their security. As stated in [6], a security policy language
is required to allow the tenants to express their security
requirements. Indeed, this makes sense that the tenants define
their security as they are the ones that know the best their
infrastructure and its security requirements.
Some of works related to security policy languages are
specific to a programming language and require the modification of the application sources. Ponder2 [7] is a distributed
object management system. The Ponder2 language can express
security and management policies for distributed systems. It
is declarative and object-oriented and can be used to declare
different types of policies. Consequently, it can only be used on
Java application augmented with the Ponder2 solution. Same
for the A4Cloud project described in [8] and its associated
language, A-PPL. Furthermore, this solution focuses on privacy and accountability, but does not address other classes of
security, such as isolation.
Works such as [9], [10] also strength the need of combing
multiple security mechanisms to provide an end-to-end and
cross-layer security. VESPA [10] is one of such architecture
for protecting cloud infrastructures using a policy-based management approach. However, this work is oriented toward the
use of automatic computing to create self-protection loops.
Consequently, they lack a language allowing the tenants to
express their security. Nonetheless, a combination with our
works could be a a way for future work. In [11], authors
present MEERKATS, a mission-oriented Cloud architecture
dedicated to security. It is composed of several components
that aim to address several types of attacks and seek to provide
high flexibility in the use of the protection mechanisms.
Nevertheless, MEERKATS lacks a simple way of expressing the
security requirements of an infrastructure. In [12], the authors
present a policy-based security framework. Their ASPF policy
consists of an attribute map (that links system elements to
their attributes) and a set of rules (indicating which actions
are allowed). While the ASPF framework can enforce security
policy, only low-level security properties can be expressed,
which makes the definition of the security policy complex.
Finally, several works such as [13] have been done around
the use of XACML to define security policies. But they focus
on a specific type of mechanisms, access control. Moreover,
XACML is a complex language [14] that requires the verification of the policy conformance regarding its syntax and its
semantic. Furthermore, XACML does not express high level
security requirements such as integrity but rather it expresses
the policy directly using low-level capabilities. Accordingly,
the size of the security policy is larger and thus the risk of
making mistakes increases as these policies are written by
human versus generated ones. Nevertheless, it could be possible to automatically generate such policy from our security
policy language. Thus, such work could be used as a security
mechanisms by our solution. [15] presents a privacy-aware
access control system since privacy is an important concern
for most users. However, the PRIME architecture is based on
XACML and therefore presents the same limitations.
Each of these solutions either focuses on one kind of
protection (mainly access control) or uses low-level security
policy (tedious to define). To the best of our knowledge, there
is no current research on the configuration of existing security
mechanisms through a common abstract language.
_B. Security Mechanisms_
Many different security mechanisms exist, providing a wide
range of features. We present some of them that we use in our
solution. SELinux [16] is a Linux Security Module (LSM) providing MAC (Mandatory Access Control). PAM [17] provides
authentication management support for Linux. Iptables [18]
is a standard Linux firewall. We use the tunnel functionality
provided by OpenSSH [19] to secure the communications
between the machines.
Even if not used in this paper, cryptographic solution (such
the one presented in [20] for the security of medical records or
even homomorphic encryption [21]) could be added to the list
of the security mechanisms that our solution takes into account.
Indeed, the solution we propose is mechanism-agnostic.
_C. Assurance_
Operational Security Assurance [22] provides the ground
for confidence that deployed security mechanisms are running
as expected. Some researches have been done to evaluate
security assurance. For instance, Common Criteria [23] evaluates security functionality and assurance by means of tests
conducted by users. However, this process is static and time
consuming. Consequently, it cannot be directly applied for
a continuous evaluation of security assurance. Furthermore,
Common Criteria focuses on the implementation phase of the
product rather than on the operation phase, when the product
is used.
Assurance Profile [24] is a formalized document that
defines a common set of security assurance measurement
requirements for a service infrastructure and facilitates a future
evaluation against these needs. This is the approach selected
for the assurance framework development.
-----
XCCDF (eXtensible Configuration Checklist Description
Format) is a standard that can perform assurance checks. It
belongs to SCAP [25], a set of specifications from NIST
to standardize the format and the naming of information
reporting concerning specific security configurations. XCCDF
provides security checklists and benchmarks to support an
automated compliance testing over a set of target systems.
OpenSCAP [26] is an auditing tool implementing SCAP and
XCCDF.
III. ARCHITECTURE AND LANGUAGE
As we have seen, many security mechanisms are efficient
but are focused toward a specific issue and/or type of protection. It is important to understand that we do not propose any
new or more secure mechanism, but we rather consider the
existing ones and automatically configure/coordinate them in
order to enforce high-level security properties. In this section,
we first present our functional architecture. Then, we present
our language and how it is used to enforce a security policy.
Eventually, we show how it is possible to automatically assess
the correctness of the enforcement.
_A. Functional Architecture_
As depicted in Fig. 1, our solution consists in a 3-steps
cycle. First, the tenant specifies its security policy using the
language detailed in Section III-B. Then, from this highlevel policy specification, the policy is enforced by firstly
selecting security mechanisms and secondly configuring them.
At the end, the policy specification and the list of selected
mechanisms are used to generate the assurance profile. The
assurance part will verify whether or not the security properties
(from the policy) are duly enforced. These assurance checks
are sent as feedbacks to the specification step to notify the
tenant if the enforcement is correct/incorrect but also if the
available mechanisms are sufficient (or not) to enforce the
policy.
Fig. 1: Functional Architecture
_B. Policy Specification_
During the policy specification step, a knowledgeable tenant (i.e., a security expert) expresses a set of security properties
to enforce e.g., confidentiality, integrity.
In [1], we have defined the Cloud Security Property Language (CSPL) that allows to specify security properties. In
particular, we have shown how to automatically transform
a property on a set of distributed objects (that we refer as
a global property) into a set of properties on local objects
(referred as local properties). In this paper, we are focusing
on the enforcement of local properties with a given security
mechanism and verifying the correct enforcement of this
property. Therefore, in the following we consider only local
properties.
CSPL is a context-based language. A context is a set of
attributes where each attribute characterizes an entity or a set
of entities. At the highest level, entities can be classified into
2 categories: subject (i.e., the active resources such as users
and processes) and object (i.e., the passive resources such as
files). For instance, the context configApp = (File="Configuration")
:(Domain="App") identifies the configuration files (attribute File)
of an application (attribute Domain). Therefore, it is possible to
use the same set of contexts on different systems. A specific
mapping file is required for each system to associate the
resources (e.g., files, users, sockets, processes) name (e.g., the
full path of a file, the user id, IP addresses, processes name)
and the context which they belong to. For example, to associate
the application’s configuration to the corresponding files, the
following line is added to the mapping file (“o” for “object”):
o /opt/dbhook/dbhook.conf configApp
Using contexts to identify resources (or set of resources),
CSPL allows to define security properties and by relying
on contexts to address entities, the expression of security
properties is independent from the resources naming of the
target system. These properties are independent from any
security mechanism; in fact, multiple mechanisms can realize
the same security property. Our proposition is to select a
security mechanism able to enforce a given property from a
pool of available mechanisms.
For instance, the property P1 expresses that the context
SCInt integrity has to be guaranteed with the exception to
SCAuth contexts that are allowed to go against this property
_i.e., no one is allowed to modify it except the resources with_
the context SCAuth.
P1:= Integrity (Context SCInt, Context SCAuth);
Then, P1 must be instantiated. For example, to specify
that the integrity of the application’s configuration files with
the context configApp should be protected and only be
the user with the context adminRoot = (Username="appAdmin"):(Role=
"StandardUser|appAdmin") is allowed to go against, the following
property is instantiated: Integrity(configApp, adminRoot).
From the tenant’s point of view, the security properties
are abstract i.e., the tenant only considers the semantics of
the properties, and not the underlying security mechanisms.
However, the properties need to be precisely defined in order to
be enforced. Thus, we introduce the concept of capabilities. A
capability is an elementary function provided by one or several
security mechanisms. For instance, C1:= deny_all_write_accesses(
Context) is a capability that can be provided by access control
mechanisms (e.g., Unix’s DAC permissions or SELinux) but
also by other security mechanisms. It can be used to enforce
an integrity property. Consequently, the integrity property P1
can be defined as follows:
P1:=Integrity (Context SCInt, Context SCauth) {
deny_all_write_accesses(SCInt);
allow_write_access(SCInt, SCauth);
}
The context SCInt represents an (set of) object(s) to secure,
while the context SCauth is the identity of a (set of) subject(s)
-----
that can write to i.e., modify, SCInt. Two capabilities are
involved in this property. The capability deny_all_write_accesses
denies all write accesses to the context while the capability
allow_write_access allows the context SCauth to counter the
previous capability.
In a Cloud environment, a security property can address
multiple machines (e.g., Virtual Machines). In our language,
a context mapped to an IP address refers to a machine. For
example, a mapping file can include the following lines (“c”
for “computer”):
c 192.168.30.8 hostClient
c 172.22.11.181 hostReverseProxy
The user can use particular attributes to include networkspecific metadata such as port e.g., tunServer:=hostReverseProxy
:(Port="5900"). These kinds of contexts can be used in typed
security properties i.e., properties accepting only specific types
of contexts (here IP:Port). For example, Confidentiality_Tunnel
(hostClient, tunServer) creates a tunnel between two machines.
In this paper and especially in our use-case (see Section IV), we will use the following security properties. Each of
the parameter (i.e., contexts) of the properties can be a single
context or a set of contexts.
**Isolation(Context SC1): Isolate a context SC1 from other**
_•_
resources and vice-versa (e.g., it isolates an application and its resources from the rest of the system)
**Confidentiality(Context SC1, Context SC2):** Deny every
_•_
contexts from reading a context SC1 with the exception
of the context SC2
**Integrity(Context SC1, Context SC2): Deny every contexts**
_•_
from modifying SC1 with the exception of the context
SC2
**Confidentiality_Tunnel(IP:Port SC1, IP:Port SC2): Create**
_•_
a secure tunnel between 2 contexts SC1 and SC2 of
types IP:Port.
**Access(Context SC1, Context SC2):** Allow connections
_•_
from SC2 to SC1
**Authentication(Context SC1, Context SC2, Context SC3):**
_•_
Upon successful connection from SC1 on SC2, modify
the context of SC1 to SC3
Finally, we enrich the set of security properties with the
following assurance property :
**Assurance(int secs): Run assurance checks at frequency**
_•_
secs (i.e., every secs seconds) for every security
property.
_C. Policy enforcement_
The second step of our solution, the policy enforcement,
must first automatically select a set of suitable security mechanisms enabling the enforcement of each security property.
And then, it must automatically configure them accordingly.
The component in charge of the policy enforcement step
is call the Secure Element Extended (SE[E] ). In the mean
time, the assurance property triggers the generation of an
assurance profile based on the security properties and the
selected mechanisms.
Let’s first present the selection and enforcement of the
security properties. Figure 2 presents the architecture of the
_SE[E]_ and how it can enforce the security policy. The SE[E]
takes as input the security properties and the contexts/resources
mapping. The SE[E] also considers the security mechanisms
that are available on the system and their capabilities. Then,
it selects the right security mechanisms and enforces the
properties by configuration.
PolicyLoad1 Property2Projection Engine 4 Get suitable SEE MappingEngine ContextGet PluginsSELinuxPlugin MechanismsSecuritySELinux
For each Select 3 plugins Capabilities SSH TunnelPlugin SSH
capability Plugin SelectorPlugin 5Send suitablesplugins DirectoryShared variablePublish iptablesPlugin Configure SSM9 iptables
7 6 pluginSelect Directory variableGet Plugin PAM PAM
Plugins Manager 8 Apply each capability OscapPlugin Oscap
using selected plugin
Fig. 2: Architecture of the SE[E] .
First, the SE[E] loads the security policy i.e., the contexts,
the properties, and the contexts/resources mapping (step 1
on Figure 2). Then, the SE[E] iterates over security properties
of the policy. For each capability of each property (step 2),
the SE[E] selects a plug-in (i.e., a security mechanism) that
can apply it (steps 3 to 6). This is done by querying the
Capabilities Directory that contains the association between
the capabilities and the security mechanisms (steps 4 and 5).
Once the Plugin Selector has the list of matching mechanisms,
it selects one of them (best-effort algorithm, step 6). When a
capability/plug-in mapping has been found for the property, it
is sent to the Plugins Manager that controls the plug-ins (step
7). Then, the Plugins Manager contacts each plug-in that needs
to perform some actions (step 8) and the plug-ins configure
their associated security mechanism (step 9).
The use of plug-ins offers a modular model: new mechanisms can be easily added by developing the associated plugin. Plug-ins implement a simple interface to communicate with
the SE[E], but the way they interact with their mechanisms is
up to the plug-in’s developer.
For instance, if the SE[E] receives the integrity property
**Integrity(configApp, adminRoot) defined in the previous section,**
it can enforce it using several security mechanisms. If this
property is enforced using SELinux, then the SE[E] generates a
SELinux module that forbids any write operation to the files
labeled configApp that does not come from a user labeled
adminRoot.
The SE[E] also provides secure communication capabilities,
especially for the case of properties between multiple systems.
Thus, the two sides (i.e., the selected mechanisms applying to
the two contexts) of the communication must use compatible
mechanisms to enforce a property. For instance, let us consider
the property Confidentiality_Tunnel(hostClient, tunServer). The
server allows the connection of the user through the defined
port, and the client sets up the tunnel. The coordination is done
by the SE[E] ’s communication capabilities.
Now, let’s present the generation of assurance files for the
|Shared Directory|variable Get|
|---|---|
|Plugin Selector|plugins 5Send suitables|Capabilities Directory|
|---|---|---|
|Col1|Col2|Col3|SEE|Col5|Col6|Col7|
|---|---|---|---|---|---|---|
|1 Load|Projection Engine||SEE Mapping Get Engine Context 4Get suitable plugins Capabilities 5Send suitablesDirectory plugins Publish Shared variable 6pS le ule gc int Directory vaG riae bt le 8Apply each capability|Plugins Plugin SELinux Plugin SSH Tunnel Plugin iptables Plugin PAM Plugin Oscap|9 Configure|Security Mechanisms|
|Policy|Property 2 3 For each Select capabilityPlugin Plugin Selector 7 Plugins Manager|||||SELinux|
||For each capability|||||SSH|
||||||||
|||||||iptables|
||7||||SSM||
||||||||
|||||||PAM|
||||||||
|||||||Oscap|
||Plugins Manager||||||
||||using selected plugin||||
-----
assurance framework presented in Section III-D. To validate
the enforcement of the security policy, multiple files are generated and given as input to the Assurance step, namely the XCCDF and system-specific scripts. System-specific scripts are
generated using a process similar to the property enforcement:
each property definition includes an assurance specification,
using capabilities. For instance, the assurance of the Integrity
property is defined as follows:
P1:=Integrity (Context SCInt, Context SCauth) {
**assurance {**
boolean c = true;
for (SCUserTmp IN get_all_users()) {
if (SCUserTmp.Id == SCauth.Id) {
c &= check_write (SCInt, SCauth);
} else {
c &= (NOT check_write (SCInt, SCauth)); }
} return c; }}
As a result, the system-specific script will contain the
implementation of the check_write assurance capability for the
context SCInt and the authorized context SCAuth. This generated
script is called a Based Measure (BM) as it is the lowest level
of assurance measure. Therefore, the XCCDF is simply a list
of Based Measures. The XCCF file and the related scripts are
given as input to the assurance step.
_D. Assurance_
In order to be able to evaluate continuously the security
assurance for a service, it is necessary to implement a process
composed of several steps: modeling, measuring, aggregation,
evaluation and presentation of the security assurance reports.
This process is supported by a set of software components
that compose our assurance framework. Fig. 3 presents our
assurance framework architecture.
Fig. 3: Assurance Framework Architecture.
_1) System Measurement Collection: As stated before, the_
assurance step receives an XCCDF file and the related systemspecific scripts called Based Measures. The SE[E] is responsible for launching the measurement collection process realized
by the Assurance Collector Engine (ACE). This engine
includes a BM Agent which executes several system-specific
scripts. Script results are associated with some metadata
including extra information to unequivocally identify their
origins and contexts. Note that the ACE, based on OpenSCAP,
is the only assurance module deployed in each virtual machine.
Next, the Measurement Aggregator (MA) receives these
measurements from each node, validates and classifies them
according to their metadata, before storing them in the Assur**ance DB.**
_2) Assurance Results Presentation: We have seen how to_
execute low-level assurance checks and collect their results. In
the following, we present how to add semantics to the collected
results i.e., interpret them, and how to present all assurance
checks to the tenant in a modular and concise manner. Our
assurance model defining the entities/files relations from lowlevel measures to high-level views is presented in Fig. 4.
Fig. 4: Assurance Model.
First, we have not determined yet if the collected values
mean a correct or faulty enforcement i.e., we need to interpret
them. Hence, we call Derived Measure (DM) the interpretation of a Based Measure.
Depending on the number of security properties and the
size of the system (i.e., number of objects), it is possible
to have a significantly large set of assurance checks (or
Derived Measures) which can be an impediment to the tenant’s
verification task. Our solution is to hierarchically aggregate
these measurements.
Therefore, a set of Derived Measures is aggregated in
an Operational Measurement Requirement (OMR) via an
aggregation function. In particular, if all Derived Measures
have successfully passed their checks, then the OMR is marked
as successful. In other words, an OMR is a set of system
assurance checks. The Operational Profile (OP) contains both
the definition of OMRs (i.e., the list of Derived Measures) and
the definition of Derived Measures.
Our next level of abstraction is to allow the tenant to
specify several Security Assurance Views (SAVs) where an
assurance view is an aggregation of Operational Measurement
Requirements. The definition of Security Assurance Views is
done in the Assurance Profile (AP) file.
In Fig. 3, the Assurance Modeling Tool takes in input
the Assurance Profile and the Operational Profile to maintain
a Security Assurance Model. Depending on the layer of the
-----
assurance model (e.g., SAV, OMR, or DM), the Assurance
**Assessment Engine is responsible for deciding if the collected**
assurance values meet the expectations and for computing the
aggregation results.
Finally, the Assurance Visualization Tool provides a
Graphical User Interface for the user to monitor the assurance
status. For a monitored service, the user will be able to select
the different SAVs available in the Security Assurance Model.
This tool presents an assurance report adapted to the tenant’s
concerns.
To sum up, the modeling and configuration of the assurance
framework rely on the definition and refinement of 3 XML
files : 1) the AP for defining the high level measurement
requirements and the assurance views adapted to the tenant’s
concerns, 2) the OP for establishing the links with the real
environment, and 3) the XCCDF for specifying how to execute the assurance measurements. Examples of these files are
provided in Section IV.
IV. IKUSI’S USECASE
_A. Description of Usecase_
In this section, we present an industrial use-case based on
Ikusi application : Airport Management. It aims to provide a
centralized-operational management for airports management
services. It involves the coordination of a group of processes,
where both human and IT system interactions are required.
It is a classical 3-tier web architecture i.e., a HTTP frontend
(tier-1), an application server (tier-2) and a database (tier-3).
This architecture is deployed on top of an IaaS Cloud and is
provided to end-users through a SaaS model. Moreover, one
instance of the application server is launched for each client
_i.e., for each airport._
Services provided by the architecture include the management of an operational data repository for each airport
operator and passenger, the real time management of flight
status updates, and the dynamic allocation and optimization of
assigned resources according to data from air flight companies
and airport operators.
It is based on message exchange modules, on resource
allocation and on billing management airport services to provide airlines with an operational platform based on Cloud
computing technology. It also incorporates enhanced security
solutions based on a network of secure element developed in
the SEED4C project.
The use-case is presented in Fig. 5. Four different kinds of
machines or VMs are involved. First, the machine ctseed1
is the client machine. It is the device that is used within
the airport to access the airport’s services. Secondly, the
reverseproxy VM (i.e., tier-1) is a proxy used by the enduser to access the airport’s services. The musik VMs (both
musik1 and musik2) belong to an airport (MAD[1] or EAS[2])
and are accessed by the end-user machine through the proxy
(i.e., an instance of the application server, tier-2). The corresponding VM is selected based on the location of the end-user.
Apart from their airport domain, these VMs are identical, so
1MAD: Madrid Airport code
2EAS: San Sebastian Airport code
we only consider one of them in this use-case (the security
policy would be duplicated). Each of these VMs runs a
Musik application that accesses the database (running in the
seed4c_mysql machine) i.e., tier-3.
_B. Security policy_
Based on the use case description, a security policy is
defined through the graphical tool Sam4C (see Fig 5).
The next listing presents an excerpt from the security
policies for the different VMs of the use-case:
1 // Policy for the Database VM
2 Isolation(DomainAODB);
3
4 Integrity(BinaryAODB);
5 Integrity(ConfigAODB, AdminRoot);
6 Integrity(KeyAODB, AdminRoot);
7 Integrity(LogAODB, ServiceAODB);
8
9 Confidentiality(FileAODB, ServiceAODB);
10 Confidentiality(KeyAODB, AdminRoot);
11 Confidentiality(ConfigAODB, AdminRoot);
12 Confidentiality(ConfigAODB, ServiceDB);
13 Confidentiality(LogAODB, AdminRoot);
14 Confidentiality(LogAODB,AdminOperator);
15
16 Authentication(HostReverseProxy, ServiceSSH, "SystemUser|
CloudProvider|AdminRoot|AdminOperator|User");
17
18 Access (MysqlPort|MysqlProxyPort|SSHPort|NTPPort, AnyIP);
19
20 Assurance(Freq);
21
22 // Policy for the ReverseProxy VM
23 Integrity(BinaryModuleWeb);
24 Integrity(BinaryWeb);
25 Integrity(ConfigWeb);
26
27 Confidentiality(ConfigWeb,AdminRoot);
28 Confidentiality_Tunnel(tunClient, tunServer);
29
30 Access (SSHPort|NTPPort, AnyIP);
31
32 Authentication(anyone, ServiceSSH, "SystemUser|CloudProvider
|AdminRoot|AdminOperator|User" );
The first security property (line 2) of this listing sandboxes
the whole application. Lines 4 to 7 forbid anyone to edit the
application’s binary, but allow several write accesses to its
files (configuration, keys, and logs). Lines 9 to 14 forbid read
access to the application resources except for the application
itself. Line 16 specifies the context evolution upon an SSH
connection: a role is given to the authenticating user depending
on his login data. Line 18 opens several ports for all incoming
IP addresses. Line 20 defines the assurance tests to perform.
The second part of the listing describes the policy for
the reverse proxy. Lines 23 to 25 guarantee the integrity of
the Web application. Line 27 requests the confidentiality of
the configuration files. Line 28 specifies that the network
communication between the proxy and the client should be
kept confidential. Line 30 opens some ports. Finally, line 32
manages the contexts evolution upon SSH connections.
The contexts used in this policy are associated to system
resources. An extract from the association file is displayed in
the next listing:
1 o /opt/dbhook(/.*)? FileAODB
2 o /opt/dbhook/dbhook.conf ConfigAODB
3 o /opt/dbhook/keys(/.*)? KeyAODB
4 o /opt/dbhook/log(/.*)? LogAODB
5 o /opt/dbhook/proxydaemon.sh BinaryAODB
-----
Fig. 5: Usecase Description
6 o /etc/rc\.d/init\.d/dbhook BinaryAODB
7 o /opt/oscap/ssm/results/SSM-results-$date.xml SSMResultFile
8 o /opt/oscap/ssm/SSM-xccdf.xml SSMXccdfFile
9
10 p /usr/bin/mysqld_safe ServiceDB
11 p /usr/libexec/mysqld ServiceDB
12 p /usr/bin/mysql-proxy ServiceAODB
13 p /usr/sbin/sshd ServiceSSH
14
15 u cloudprovider CloudProvider
16 u tenant-admin AdminRoot
17 u tenant-operator AdminOperator
18 u user User
19
20 c 172.22.11.181 HostReverseProxy
21 c 172.22.11.178 HostServerBBDD
22 c 212.81.220.68 HostClient
Lines 1 to 8 of the mapping file associate the contexts to
files. Lines 10 to 13 are for the processes, lines 15 to 18 for
the users, and lines 20 to 22 for the computers (IP addresses).
_C. Security Enforcement_
The security policy is enforced by several security mechanisms. The SE[E] detects what are the available mechanisms
and selects those that can enforce the properties.
In this usecase, four mechanisms collaborate to enforce the
whole policy.
_1) SELinux:_ First security mechanism available is
SELinux. It enforces properties from three groups: isolation,
confidentiality, and integrity. To enforce them, the plug-in
generates a SELinux module.
Upon receiving an isolation property for a domain, the
plug-in creates a SELinux module to isolate all elements of
this domain from the rest of the system. Then, the plug-in
will allow some interactions corresponding to confidentiality
and integrity properties. To enforce the properties Isolation(
DomainAODB), Integrity(ConfigAODB, AdminRoot), and Confidentiality(
ConfigAODB, "ServiceDB|AdminRoot") from policy, the following module is generated (see next listing). Lines 2-5 define the domain
and SELinux contexts, while lines 7-8 give authorization rules.
Lines 12-13 associate SELinux contexts to resources.
1 $ cat Aodb.te
2 policy_module(Aodb,1.0.0)
3 see_create_service_domain(Aodb)
4 see_create_files_type(Aodb_conf_t)
5 see_create_files_type(Aodb_file_t)
6
7 see_files_type_read_write(Aodb_t,Aodb_conf_t)
8 see_files_type_read(idAodbAdmin_t,Aodb_conf_t)
9 [...]
10
11 $ cat Aodb.fc
12 /opt/dbhook/dbhook.conf gen_context(system_u:object_r:
Aodb_conf_t,s0)
13 /usr/bin/mysql-proxy gen_context(system_u:object_r:
Aodb_exec_t,s0)
14 [...]
_2) PAM:_ The PAM plug-in enforces authentication
properties. Indeed, such property specifies how contexts can evolve to have correct properties applied.
Moreover, it controls the authentication rights and allows or denies a user authentication. Upon encountering the property **Authentication(anyone, ServiceSSH, "SystemUser|**
CloudProvider|AdminRoot|AdminOperator|User"), PAM plug-in adds a
rule to PAM configuration in order to detect a successful login:
session required pam_exec.so /etc/see/scripts/notifyLogin
When a successful authentication occurs, PAM executes
the script notifyLogin (see next listing), which informs the
_SE[E]_ (through Ncat) of a connection and sends data, such as
the user name, the remote host, or the date.
1 $ cat notifyLogin
2 #!/bin/sh
3 [ "$PAM_TYPE" = "open_session" ] || exit 0
4 {echo "User: $PAM_USER"
5 echo "Ruser: $PAM_RUSER"
6 echo "Rhost: $PAM_RHOST"
7 echo "Service: $PAM_SERVICE"
8 echo "TTY: $PAM_TTY"
9 echo "Date: ‘date‘"
10 echo "Server: ‘uname -a‘"
11 echo "PID: $$"
-----
12 echo "PPID: $PPID"
13 } | ncat -U --send-only /var/run/seePam
_3) iptables: The iptables plug-in enforces the network_
access properties. For instance, the iptables plug-in can allow
network communications on a specific port or from a given IP
address.
The Access properties in the use-case’s policy are used to
open some ports. For instance, the access property Access (
SSHPort, AnyIP) is enforced using the following iptables rule:
iptables -I INPUT -m state --state NEW -p tcp --dport 22 -j
ACCEPT
This plug-in can be requested to apply additional capabilities by other plug-ins. For instance, the SSH tunneling plug-in
can dynamically request a specific port to be opened by the
iptables plug-in.
_4) SSH Tunneling: The SSH Tunneling plug-in enforces_
the creation of confidential tunnels between machines inside
and outside the Cloud. Apart from the infrastructure in the
Cloud, the Airport Management use-case includes physical
machines located in the airport. As part of the use-case, we
need to monitor what is displayed on the machines from the
Cloud application. The remote port forwarding process comes
with a solution to this issue allowing flows redirection.
The remote port forwarding process ensures the confidentiality of the communication, because SSH is an encrypted
protocol. Furthermore, thank to the public key cryptography,
both sides of the communication channel are authenticated.
The communication between the machines is essential
since the SSH server machine has to allocate its own local
ports. They will be assigned to a SSH client in order to allow
it to do the port forwarding. The enforcement is made of 3
steps: 1) the SSH server machine allocates a local port for the
client to set up the tunnel, 2) the SSH client gets the allocated
port, and 3) the SSH client creates the remote tunnel.
Fig. 6 shows this process for a remote port forwarding
tunnel creation using the port 5900 as the port on client
machine.
Fig. 6: Tunnel creation process example
To enforce the tunnel, the command ssh -R 7900:0.0.0.0:5900
ctseed1@reverse-seed4c is executed, where 7900 is the allocated
port for the SSH server, 5900 is the objective port for the SSH
client, reverse-seed4c is the server’s hostname, and ctseed1 is the
user on the SSH server machine used by the client machine.
_D. Assurance_
The Assurance Model used in the airport management usecase is based on the security policy and focuses on monitoring
the effectiveness of the security mechanisms. The model
checks that deployed security mechanisms (e.g., SELinux,
Iptables, and OpenSSH tunnel) are running as expected. It also
checks that the security properties are fulfilled, in terms of data
integrity, data confidentiality, and data availability.
For instance, the enforcement of the property Integrity
(ConfigAODB, AdminRoot) (line 5 of the security policy in Section IV-B) can be checked using the script from listing 1. This
script is generated by the SE[E] during the enforcement step
(Section IV-C), depending on the properties of the security
policy.
1 $ cat BM_fileInt-1.1.sh
2 #!/bin/bash
3 RET=$XCCDF_RESULT_PASS
4 check_write(){su -c "test -w "$1"" $2; return $?;}
5 FILES=[...] # list of files in integrity property
6 USERS=[...] # list of all users
7 OK_USERS=[...] # list of authorized users
8
9 for file in "${FILES[@]}" ; do
10 for user in "${USERS[@]}" ; do
11 check_write $file $user
12 WRITE_OK=$?
13
14 if [[ " ${OK_USERS[@]} " =˜ " $user " ]] ; then
15 if [[ $WRITE_OK -ne "0" ]] ; then
16 RET=$XCCDF_RESULT_FAIL
17 echo "Unexpected access denial: $user->$file"
18 fi
19 else
20 if [[ $WRITE_OK -eq "0" ]] ; then
21 RET=$XCCDF_RESULT_FAIL
22 echo "Unauthorized access: $user->$file"
23 fi
24 fi
25 done
26 done
27 exit $RET
Listing 1: Script checking the integrity of a file
The script BM_fileInt-1.1.sh checks the integrity of
a file by testing which users are allowed to write it. Line
4 defines a function to check if a file can be written by a
specific user. Lines 5 to 7 get the files and users involved in
the property (not detailed here due to lack of space). Then,
the script loops over the files (line 9) and the users (line 10)
and tries to open the files for writing (line 11). If the property
and the test result do not match (lines 15 and 20), the return
value is set to XCCDF_RESULT_FAIL (lines 16 and 21), so that
the script will exit with a failure. Otherwise, the script exits
with the return value XCCDF_RESULT_PASS, indicating that the
integrity property has been properly enforced.
As presented before, the assurance framework is steered
by 3 files, namely the Assurance Profile (AP), the Operational
Profile (OP) and the XCCDF.
The excerpt of the Assurance Profile presented in Listing 2 defines one Security Assurance View (SAV) with two
Operational Measurement Requirements (OMRs), OMR_1 and
OMR_3 (lines 10-11), needed for the evaluation of data integrity
(lines 7-13).
1 [...]
2 <SecurityAssuranceView id="SAV_1">
3 <Statement>Security Functions effectiveness</Statement>
-----
4 <SAVObject id="1_Data_Int">
5 <Description>Data Integrity</Description>
6 <MetricsAggregFunction>#%t</MetricsAggregFunction>
7 <Metric id="SF_Int_Active">
8 <Description>Availability of security functions affecting
data integrity</Description>
9 <ReqAggregFunction>#t==##</ReqAggregFunction>
10 <ConcernedMeasurementReq>OMR_1</ConcernedMeasurementReq>
11 <ConcernedMeasurementReq>OMR_3</ConcernedMeasurementReq>
12 [...]
13 </Metric>
14 [...]
15 </SAVObject>
16 [...]
17 </SecurityAssuranceView>
18 [...]
Listing 2: AP file for the Airport Management use case.
The XCCDF file in Listing 3 defines the last step of the
measurement chain. It specifies the assurance checks (with
their related scripts, for example BM_fileInt-1.1.sh, line
14) that have to be executed to collect the base measures
(here, BM-fileInt-1.1, lines 9 to 16) needed to evaluate
upper levels of the assurance model.
1 [...]
2 <Profile id="properties_IO">
3 <description>Properties Assurance</description>
4 <select idref="BM-fileInt-1.1" selected="true" />
5 <select idref="BM-fileConf-1.1" selected="true" />
6 <select idref="BM-netConf-1.1" selected="true" />
7 </Profile>
8 <Group id="properties_group">
9 <Rule id="BM-fileInt-1.1" selected="true">
10 <title>File Integrity</title>
11 <description>Check that file integrity is enforced</
**description>**
12 <check system="http://open-scap.org/page/SCE">
13 <check-import import-name="stdout" />
14 <check-content-ref href="BM_fileInt-1.1.sh"/>
15 </check>
16 </Rule>
17 [...]
18 </Group>
19 [...]
Listing 3: XCCDF file for the Airport Management use case.
In order for the Assurance Profile and the XCCDF file
to inter-operate, the Operational Profile (Listing 4) links
the Operational Measurement Requirements OMR_3 of the
Assurance Profile (lines 13 to 18) with the Based Measures
BM-fileInt-1.1 of the XCCDF file (lines 3 to 9). It also
specifies the machine from which to collect this data (line 7).
1 [ . . . ]
2 <DerivedMeasures>
3 _<DerivedMeasure_ **id** =”DM− f i l e I n t −1.1− musik1 ”>
4 _<Description>Check_ t h a t f i l e i n t e g r i t y i s e f f e c t i v e</
**Description>**
5 _<InterpretFunction>” pass ” . equals ($0)</ InterpretFunction>_
6 _<ConcernedBaseMeasure>BM−_ f i l e I n t −1.1</ ConcernedBaseMeasure
_>_
7 _<ConcernedDevice>Musik1</ ConcernedDevice>_
8 _<P e r i o d i c i t y>180000</ P e r i o d i c i t y>_
9 _</ DerivedMeasure>_
10 [ . . . ]
11 </ DerivedMeasures>
12 <MeasurementRequirements>
13 _<MeasurementRequirement id_ =”OMR 3”>
14 _<MRAggregFunction># t ==##</ MRAggregFunction>_
15 _<DerivedMeasure>DM−_ f i l e I n t −1.1− musik1</ DerivedMeasure>
16 _<DerivedMeasure>DM−_ f i l e I n t −1.1− musik2</ DerivedMeasure>
17 _<DerivedMeasure>DM−_ f i l e I n t −1.1−db</ DerivedMeasure>
18 _</ MeasurementRequirement>_
19 [ . . . ]
20 </ MeasurementRequirements>
21 [ . . . ]
Listing 4: OP file for the Airport Management use case.
The Assurance Collector Engine executes the script
BM_fileInt-1.1.sh (listing 1) in order to check to enforcement of integrity properties in the security policy (here, the
property on line 5 of the policy).
Both the Assurance Profile and the Operational Profile are
imported in the Assurance Modeling Tool and derived into the
Airport Management Assurance Model, displayed in Fig. 7
by the Assurance Visualization Tool. The model shows the
Security Assurance Views defined in the Assurance Profile,
in this case the Security Functions effectiveness view, with its
corresponding measurement requirements fed by the assurance
checks. The left framework of the model allows the navigation
by the model structure and shows the assurance compliance
in a colour basis. The right framework shows the details
on the selected model component. In this case it shows the
base measures corresponding to SELinux (MAC) mechanisms
status, but the results obtained from the integrity property
verification can also be displayed.
Fig. 7: Airport Management Assurance Model Evaluation and Visualization.
_E. Results_
Table I presents some statistics concerning the security
policy for this usecase.
Number of
Templates 8
Properties 47
For the client node 1
For the proxy VM 7
For the application VM 12
For the database VM 27
SSMs collaborating to enforce the security properties 5
(SELinux, iptables, PAM, SSH, Oscap)
Assurance scripts for the properties 8
Assurance scripts for the SSMs 4
TABLE I: Use-case Policy Statistics
As we can see, the policy for this use-case uses only 8
different properties templates, since our high-level properties
cover a wide range of security needs. The policy itself uses
|Number of|Col2|
|---|---|
|Templates|8|
|Properties For the client node For the proxy VM For the application VM For the database VM|47 1 7 12 27|
|SSMs collaborating to enforce the security properties (SELinux, iptables, PAM, SSH, Oscap)|5|
|Assurance scripts for the properties Assurance scripts for the SSMs|8 4|
-----
50 contexts and 47 properties for the protection of the whole
use-case, which is a very low number considering all the security functionalities covered. Moreover, this policy is entirely
generated from a GUI, so the Cloud tenant does not have to
write these contexts and properties himself. Besides, this policy
manages both the enforcement and the assurance, so that Cloud
tenant has information about the status of the enforcement,
through a graphical dashboard.
V. CONCLUSION
In this paper, we have presented a solution to specify,
enforce and assure security properties in a Cloud environment.
Our solution handles the enforcement by re-using existing
security and assurance mechanisms, such as SELinux, iptables,
PAM, SSH, or Oscap. Our solution is composed of several
elements: 1) a language that can express the security and assurance properties independently from the system, the resources
naming, and the available mechanisms, 2) an enforcement
engine, the SE[E], that receives the properties and enforces
them by configuring existing mechanisms, and 3) an assurance
framework that models, measures, aggregates, evaluates and
presents the security assurance results. Our solution has shown
its efficiency on a complete industrial use-case for airport
system management: 1) the policy expressing the security
requirements of the use-case has been defined, 2) the policy
has been enforced using several mechanisms that collaborate to
offer an end-to-end protection (across the different machines),
and 3) the assurance framework has confirmed the proper
enforcement of the security policy.
In our future works, we will define generic policy templates
that could be used to secure the system base, in addition to
the policy on the tenant’s software architecture. This added
protection would improve the overall security of the system.
Besides, we plan to extend the language so that the results
generated by the assurance framework are sent back to the
enforcement engine: this would allow the enforcement engine
to update the configuration of the security mechanisms to adapt
the protection in case something is not working as expected.
**_Acknowledgments_**
This work was done thanks to the financial support of the
Celtic+ project Seed4C (Eureka Celtic+ CPP2011/2-6).
REFERENCES
[1] L. Bobelin, A. Bousquet, J. Briffaut, J.-F. Couturier, C. Toinard,
E. Caron, A. Lefray, and J. Rouzaud-Cornabas, “An advanced securityaware cloud architecture,” in High Performance Computing & Simu_lation (HPCS), 2014 International Conference on._ IEEE, 2014, pp.
572–579.
[2] United States Department of Defense, “Trusted Computer System
Evaluation Criteria (Orange Book),” Tech. Rep., 1985.
[3] Cloud Security Alliance, “Cloud Adoption Practices and Priorities
Survey Report,” 2015, https://cloudsecurityalliance.org/download/cloudadoption-practices-priorities-survey-report/.
[4] T. Jaeger and J. Schiffman, “Outlook: Cloudy with a Chance of
Security Challenges and Improvements,” IEEE Security & Privacy
_Magazine, vol. 8, no. 1, pp. 77–80, Jan. 2010. [Online]. Available: http:_
//ieeexplore.ieee.org/lpdocs/epic03/wrapper.htm?arnumber=5403158
[5] J. Y. Halpern and V. Weissman, “Using first-order logic to reason
about policies,” ACM Trans. Inf. Syst. Secur., vol. 11, no. 4, pp.
21:1–21:41, Jul. 2008. [Online]. Available: http://doi.acm.org/10.1145/
1380564.1380569
[6] R. Sandhu, R. Boppana, R. Krishnan, J. Reich, T. Wolff, and J. Zachry,
“Towards a Discipline of Mission-Aware Cloud Computing,” in Pro_ceedings of the 2010 ACM workshop on Cloud computing security_
_workshop, 2010, pp. 13–18._
[7] K. Twidle, N. Dulay, E. Lupu, and M. Sloman, “Ponder2: A policy
system for autonomous pervasive environments,” in Autonomic and
_Autonomous Systems, 2009. ICAS’09. Fifth International Conference_
_on._ IEEE, 2009, pp. 330–335.
[8] S. Pearson, V. Tountopoulos, D. Catteddu, M. S¨udholt, R. Molva,
C. Reich, S. Fischer-H¨ubner, C. Millard, V. Lotz, M. G. Jaatun et al.,
“Accountability for cloud and other future internet services.” in Cloud_Com, 2012, pp. 629–632._
[9] J. Chen, Y. Wang, and X. Wang, “On-demand security architecture for
cloud computing,” Computer, vol. 45, no. 7, pp. 73–78, 2012.
[10] A. Wailly, M. Lacoste, and H. Debar, “Vespa: multi-layered selfprotection for cloud resources,” in Proceedings of the 9th international
_conference on Autonomic computing._ ACM, 2012, pp. 155–160.
[11] A. D. Keromytis, R. Geambasu, S. Sethumadhavan, S. J. Stolfo, J. Yang,
A. Benameur, M. Dacier, M. Elder, D. Kienzle, and A. Stavrou, “The
meerkats cloud security architecture,” in Distributed Computing Systems
_Workshops (ICDCSW), 2012 32nd International Conference on. IEEE,_
2012, pp. 446–450.
[12] R. He, M. Lacoste, and J. Leneutre, “A policy management framework
for self-protection of pervasive systems,” in Autonomic and Autonomous
_Systems (ICAS), 2010 Sixth International Conference on._ IEEE, 2010,
pp. 104–109.
[13] C. Ngo, P. Membrey, Y. Demchenko, and C. de Laat, “Policy and
context management in dynamically provisioned access control service
for virtualized cloud infrastructures,” in Availability, Reliability and
_Security (ARES), 2012 Seventh International Conference on, Aug 2012,_
pp. 343–349.
[14] V. C. Hu, E. Martin, J. Hwang, and T. Xie, “Conformance checking
of access control policies specified in xacml,” in Computer Software
_and Applications Conference, 2007. COMPSAC 2007. 31st Annual_
_International, vol. 2._ IEEE, 2007, pp. 275–280.
[15] C. A. Ardagna, S. De Capitani di Vimercati, S. Paraboschi, E. Pedrini,
and P. Samarati, “An xacml-based privacy-centered access control
system,” in Proceedings of the first ACM workshop on Information
_security governance._ ACM, 2009, pp. 49–58.
[16] R. Haines, The SELinux Notebook - Third Edition, 2012.
[17] V. Samar, “Unified login with pluggable authentication modules (pam),”
in Proceedings of the 3rd ACM conference on Computer and commu_nications security._ ACM, 1996, pp. 1–10.
[18] O. Andreasson, “Iptables tutorial 1.2. 2,” 2001.
[19] T. Ylonen and C. Lonvick, “The secure shell (ssh) connection protocol,”
2006.
[20] M. Li, S. Yu, Y. Zheng, K. Ren, and W. Lou, “Scalable and secure
sharing of personal health records in cloud computing using attributebased encryption,” Parallel and Distributed Systems, IEEE Transactions
_on, vol. 24, no. 1, pp. 131–143, 2013._
[21] A. L´opez-Alt, E. Tromer, and V. Vaikuntanathan, “On-the-fly
multiparty computation on the cloud via multikey fully homomorphic
encryption,” in _Proceedings_ _of_ _the_ _Forty-fourth_ _Annual_ _ACM_
_Symposium_ _on_ _Theory_ _of_ _Computing,_ ser. STOC ’12. New
York, NY, USA: ACM, 2012, pp. 1219–1234. [Online]. Available:
http://doi.acm.org/10.1145/2213977.2214086
[22] S. Haddad, S. Dubus, A. Hecker, T. Kanstr´en, B. Marquet, and
R. Savola, “Operational security assurance evaluation in open infrastructures,” in Risk and Security of Internet and Systems (CRiSIS), 2011
_6th International Conference on._ IEEE, 2011, pp. 1–6.
[23] C. Criteria, “Common Criteria for Information Technology Evaluation
v3.1 (ISO/IEC 15408),” 2012.
[24] ETSI, “ETSI TR 187 023: Security Assurance Profile for Secured Telecom Operations Statement of needs for security assurance measurement
in operational telecom infrastructures,” 2012, www.etsi.org/deliver/etsi
tr/187000 187099/187023/01.01.01 60/tr 187023v010101p.pdf.
[25] SCAP, “SCAP: Security Content Automation Protocol,” 2014, http://
scap.nist.gov/.
[26] OpenSCAP, “OpenSCAP Website,” 2014, http://www.open-scap.org.
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.1109/UCC.2015.45?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/UCC.2015.45, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBYNCND",
"status": "GREEN",
"url": "https://hal.inria.fr/hal-01240557/file/UCC_2015.pdf"
}
| 2,015
|
[
"JournalArticle",
"Conference"
] | true
| 2015-12-01T00:00:00
|
[
{
"paperId": "58a56d344c96de6d2ae898d0ae7b4f8fc4b6ade8",
"title": "Scalable and Secure Sharing of Personal Health Records in Cloud Computing Using Attribute-Based Encryption"
},
{
"paperId": "cff2f444a32c7318b182ea8fcc64799b43c693c3",
"title": "SCALABLE AND SECURE SHARING OF PERSONAL HEALTH RECORDS IN CLOUD COMPUTING"
},
{
"paperId": "da8d7822c57b6fcaefc35bce2f6ac845f0641305",
"title": "Development of protection profile and security target for Indonesia electronic ID card's (KTP-el) card reader based on common criteria V3.1:2012/SNI ISO/IEC 15408:2014"
},
{
"paperId": "1a798f7f0df116c2ac43769764594f24df3d97b4",
"title": "An advanced security-aware Cloud architecture"
},
{
"paperId": "b578fe9ed0f71a0ea4d6acc4fabbe6cab91447fe",
"title": "Accountability for cloud and other future Internet services"
},
{
"paperId": "b8f22f6fbf4f91607f842c37c0367826067830af",
"title": "VESPA: multi-layered self-protection for cloud resources"
},
{
"paperId": "00ffdd19d85c066a1b9359992eb31168b1e12e1f",
"title": "Policy and Context Management in Dynamically Provisioned Access Control Service for Virtualized Cloud Infrastructures"
},
{
"paperId": "115193759bcf8ffcb2cf8d8a29cabbf54228fbc3",
"title": "On-Demand Security Architecture for Cloud Computing"
},
{
"paperId": "4a16120f3da0159922d3e066b8e61f1e4dfb1760",
"title": "The MEERKATS Cloud Security Architecture"
},
{
"paperId": "c3f095c11102196896ed10ed27b9adb2e5bca68f",
"title": "On-the-fly multiparty computation on the cloud via multikey fully homomorphic encryption"
},
{
"paperId": "0eda5846a8ab5ac5ec175331c98554de389b3449",
"title": "Towards a discipline of mission-aware cloud computing"
},
{
"paperId": "18b1791690af2873c800c9be6ef1a51470cc15ab",
"title": "A Policy Management Framework for Self-Protection of Pervasive Systems"
},
{
"paperId": "23c9b3ca9b7f1ca3226ba3cbfbc26e7cf7935880",
"title": "An XACML-based privacy-centered access control system"
},
{
"paperId": "d49b3b10d0da631e8e0f39870bc001a8e76ee864",
"title": "Proceedings of the first ACM workshop on Information security governance"
},
{
"paperId": "c1e0cf409bebefa4997c84fe2a3b401f52c06138",
"title": "Ponder2: A Policy System for Autonomous Pervasive Environments"
},
{
"paperId": "d17f69b0061d54f5c8b114faabc969b770591497",
"title": "Conformance Checking of Access Control Policies Specified in XACML"
},
{
"paperId": "3c154b28ba5110443894e2c26edb183a1504a802",
"title": "Using first-order logic to reason about policies"
},
{
"paperId": null,
"title": "“Cloud Adoption Practices and Priorities Survey Report,”"
},
{
"paperId": null,
"title": "OpenSCAP Website"
},
{
"paperId": null,
"title": "SCAP: Security Content Automation Protocol"
},
{
"paperId": null,
"title": "Common Criteria for Information Technology Evaluation v3.1 (ISO/IEC 15408)"
},
{
"paperId": null,
"title": "The SELinux Notebook -Third Edition"
},
{
"paperId": "0958d851a6b1ff1b4b5b862e3758d2ac10bf2779",
"title": "Outlook: Cloudy with a Chance of Security Challenges and Improvements"
},
{
"paperId": "e67fb2446f914ef589835c540bcd60c5542a5c0e",
"title": "The Secure Shell (SSH) Connection Protocol"
},
{
"paperId": "6d701f92a3b7cf34fe4cc770da9782a564aa9307",
"title": "Trusted Computer System Evaluation Criteria ( Orange Book ) December"
},
{
"paperId": null,
"title": "Iptables tutorial 1.2. 2"
},
{
"paperId": "6adacd3f01ef16bd70b179132cdf8fa5ea6e6531",
"title": "Unified Login with Pluggable Authentication Modules ( PAM )"
},
{
"paperId": "2e602d3dd1d9e950fcb73da5f999a812e019d3c2",
"title": "2011 6th International Conference on Risks and Security of Internet and Systems (crisis) Operational Security Assurance Evaluation in Open Infrastructures"
},
{
"paperId": null,
"title": "// Policy for the Database VM 2 Isolation( DomainAODB )"
},
{
"paperId": null,
"title": "Policy for the ReverseProxy VM 23 Integrity( BinaryModuleWeb ); 24 Integrity( BinaryWeb ); 25 Integrity( ConfigWeb )"
},
{
"paperId": null,
"title": "Access ( MysqlPort | MysqlProxyPort | SSHPort | NTPPort , AnyIP )"
},
{
"paperId": null,
"title": "Assurance( Freq )"
},
{
"paperId": null,
"title": "32 Authentication(anyone , ServiceSSH"
},
{
"paperId": null,
"title": "ETSI TR 187 023: Security Assurance Profile for Secured Telecom Operations Statement of needs for security assurance measurement in operational telecom infrastructures"
}
] | 14,989
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/028821c2f74bce87b58d99cf63a204fe5cce94e9
|
[
"Computer Science"
] | 0.879975
|
Optimal content placement for peer-to-peer video-on-demand systems
|
028821c2f74bce87b58d99cf63a204fe5cce94e9
|
2011 Proceedings IEEE INFOCOM
|
[
{
"authorId": "143858231",
"name": "Bo Tan"
},
{
"authorId": "1768703",
"name": "L. Massoulié"
}
] |
{
"alternate_issns": null,
"alternate_names": null,
"alternate_urls": null,
"id": null,
"issn": null,
"name": null,
"type": null,
"url": null
}
| null |
# Optimal Content Placement for Peer-to-Peer Video-on-Demand Systems[1]
#### Bo (Rambo) Tan
Department of Electrical and Computer Engineering
University of Illinois at Urbana-Champaign
Urbana, IL 61801, USA
Email: botan2@illinois.edu
#### Laurent Massouli´e
Technicolor Paris Research Lab
Issy-les-Moulineaux Cedex 92648, France
Email: laurent.massoulie@technicolor.com
**(a) Distributed Server Network** **(b) Pure Peer-to-Peer Network**
**_Abstract—In this paper, we address the problem of content_**
**placement in peer-to-peer systems, with the objective of maxi-**
**mizing the utilization of peers’ uplink bandwidth resources. We**
**consider system performance under a many-user asymptotic. We**
**distinguish two scenarios, namely “Distributed Server Networks”**
**(DSN) for which requests are exogenous to the system, and “Pure**
**P2P Networks” (PP2PN) for which requests emanate from the**
**peers themselves. For both scenarios, we consider a loss network**
**model of performance, and determine asymptotically optimal**
**content placement strategies in the case of a limited content**
**catalogue. We then turn to an alternative “large catalogue”**
**scaling where the catalogue size scales with the peer population.**
**Under this scaling, we establish that storage space per peer**
**must necessarily grow unboundedly if bandwidth utilization is**
**to be maximized. Relating the system performance to properties**
**of a specific random graph model, we then identify a content**
**placement strategy and a request acceptance policy which jointly**
**maximize bandwidth utilization, provided storage space per peer**
**grows unboundedly, although arbitrarily slowly, with system size.**
I. INTRODUCTION
The amount of multimedia traffic accessed via the Internet,
already of the order of exabytes (10[18]) per month, is expected
to grow steadily in the coming years. A peer-to-peer (P2P)
architecture, whereby peers contribute resources to support
service of such traffic, holds the promise to support such
growth more cheaply than by scaling up the size of data
centers. More precisely, a large-scale P2P system based on
resources of individual users can absorb part of the load that
would otherwise need to be served by data centers.
In the present work we address specifically the Video-onDemand (VoD) application, for which the critical resources
at the peers are storage space and uplink bandwidth. Our
objective is to ensure that the largest fraction of traffic is
supported by the P2P system. More precisely, we look for
content placement strategies that enable content downloaders
to maximally use the peers’ uplink bandwidth, and hence maximally offload the servers in the data centers. Such strategies
must adjust to the distinct popularity of video contents, as a
more popular content should be replicated more frequently.
We consider the following mode of operation: Video requests are first submitted to the P2P system; if they are
1Part of the results developed in this paper have made the object of a “brief
announcement” in [12] and further shown in more detail in [13].
Fig. 1: Two architectures of P2P VoD systems
accepted, uplink bandwidth is used to serve them at the
video streaming rate (potentially via parallel substreams from
different peers). They are rejected if their acceptance would
require disruption of an ongoing request service. Rejected
requests are then handled by the data center. Alternative modes
of operation could be envisioned (e.g., enqueueing of requests,
service at rates distinct from the streaming rate, joint service
by peers and data center,...). However the proposed model is
appealing for the following reasons. It ensures zero waiting
time for requests, which is desirable for VoD application;
analysis is facilitated, since the system can be modeled as
a loss network [7], for which powerful theoretical results are
available; and finally, as our results show, simple placement
strategies ensure optimal operation in the present model.
In the P2P system we are considering, there are two kinds
of peers: boxes and pure users. Their difference is that boxes
do contribute resources (storage space and uplink bandwidth)
to the system, while pure users do not. This paper focuses on
the following two architectures (illustrated in Figure 1):
- Distributed Server Network (DSN): Requests to download contents come only from pure users, and can be
regarded as external requests.
- Pure P2P Network (PP2PN): There are no pure users
in the system, and boxes do generate content requests,
which can be regarded as “internal”.
The rest of the paper is organized as follows: We review
related work in Section II and introduce our system model in
Section III. For the Distributed Server Network scenario, the
-----
so-called “proportional-to-product” content placement strategy
is introduced and shown to be optimal in a large system limit
in Section IV, where extensive simulation results are also provided. For the Pure P2P Network scenario, a distinct placement
strategy is introduced and proved optimal in Section V. These
results apply for a catalogue of contents of limited size. An
alternative model in which catalogue size grows with the user
population is introduced in Section VI, where it is shown
that the “proportional-to-product” placement strategy remains
optimal in the DSN scenario in this large catalogue setting,
for a suitably modified request management technique.
II. RELATED WORK
The number and location of replicas of distinct content
objects in a P2P system have a strong impact on such system’s
performance. Indeed, together with the strategy for handling
incoming requests, they determine whether such requests
must either be delayed, or served from an alternative, more
expensive source such as a remote data center. Requests which
cannot start service at once can either be enqueued (we then
speak of a waiting model) or redirected (we then speak of a
loss model).
Previous investigations of content placement for P2P VoD
systems were conducted by Suh et al. [11]. The problem tackled in [11] differs from our current perspective, in particular no
optimization of placement with respect to content popularity
was attempted in this work. Performance analysis of both
queueing and loss models are considered in [11]. Valancius
et al. [17] considered content placement dependent on content
popularity, based on a heuristic linear program, and validated
this heuristic’s performance in a loss model via simulations.
Tewari and Kleinrock [14], [15] advocated to tune the
number of replicas in proportion to the request rate of the
corresponding content, based on a simple queueing formula,
for a waiting model, and also from the standpoint of the load
on network links. They further established via simulations that
Least Recently Used (LRU) storage management policies at
peers emulated rather well their proposed allocation.
Wu et al. [18] considered a loss model, and a specific timeslotted mode of operation whereby requests are submitted
to randomly selected peers, who accommodate a randomly
selected request. They showed that in this setup the optimal
cache update strategy can be expressed as a dynamic program.
Through experiments, they established that simple mechanisms such as LRU or Least Frequently Used (LFU) perform
close to the optimal strategy they had previously characterized.
Kangasharju et al. [6] addressed file replication in an environment where peers are intermittently available, with the aim
of maximizing the probability of a requested file being present
at an available peer. This differs from our present focus in that
the bandwidth limitation of peers is not taken into account,
while the emphasis is on their intermittent presence. They
established optimality of content replication in proportion to
the logarithm of its popularity, and identified simple heuristics
approaching this.
Boufkhad et al. [3] considered P2P VoD from yet another
viewpoint, looking at the number of contents that can be
simultaneously served by a collection of peers.
Content placement problem has also been addressed towards
other different optimization objectives. For example, Almeida
et al. [1] aim at minimizing total delivery cost in the network,
and Zhou et al. [19] target jointly maximizing the average
encoding bit rate and average number of content replicas
as well as minimizing the communication load imbalance of
video servers.
Cache dimensioning problem is considered in [9], where
Laoutaris et al. optimized the storage capacity allocation
for content distribution networks under a limited total cache
storage budget, so as to reduce average fetch distance for
the request contents with consideration of load balancing and
workload constraints on a given node. Our paper takes a
different perspective, focusing on many-user asymptotics so
the results show that the finite storage capacity per node is
never a bottleneck (even in the “large catalogue model”, it
also scales to infinity more slowly than the system size).
There are obvious similarities between our present objective
and the above works. However, none of these identifies explicit
content placement strategies at the level of the individual peers,
which lead to minimal fraction of redirected (lost) requests in
a setup with dynamic arrivals of requests.
Finally, there is a rich literature on loss networks (see in
particular Kelly [7]); however our present concern of optimizing placement to minimize the amount of rejected traffic in a
corresponding loss network appears new.
III. MODEL DESCRIPTION
We now introduce our mathematical model and related
notations. Denote the set of all boxes as . Let = B and
B |B|
index the boxes from 1 to B. Box b has a local cache Jb that
can store up to M contents, all boxes having the same storage
space M . We further assume that each box can simultaneously
serve U concurrent requests, where U is an integer, i.e., each
box has an uplink bandwidth equal to U times the video
streaming rate. In particular we assume identical streaming
rates for all contents.
The set of available contents is defined as . Let = C
C |C|
and index contents from 1 to C. Thus a given box b will be
able to serve requests for content c for all c ∈Jb.
In a Pure P2P Network, when box b has a request for
a certain content c, which is coincidentally already in its
cache, a “local service” is provided and no download service
is needed, hence the service to this request consumes no
bandwidth resource. The effect of local service on deriving
an optimal content placement strategy will be discussed in
detail in Section V.
In a Distributed Server Network, however, local service will
never occur since all the requests are external with respect to
-----
the system resources[2].
For a new request that needs a download service, an
attempt is made to serve this request by some box holding
content c, while ensuring that previously accepted requests
can themselves be assigned to adequate boxes, given the cache
content and bandwidth resources of all boxes. This potentially
involves “repacking” of requests, i.e., reallocation of all the
bandwidth resources in the system (“box-serving-request”
mapping) to accommodate this new download demand pattern.
If such repacking can be found, then the request is accepted;
otherwise, it is rejected from the P2P system.
It will be useful in the sequel to characterize the concurrent
numbers of requests that are amenable to such repacking. Let
n = {nc}c∈C be the vector of numbers nc of requests per
content c. Clearly, a matching of these requests to server boxes
is feasible if and only if there exist nonnegative integers zcb
(number of concurrent downloads of content c from box b)
such that
� zcb = nc, ∀ c ∈C;
b:c∈Jb
� zcb ≤ U, ∀ b ∈B. (1)
c:c∈Jb
A more compact characterization of feasibility follows by an
application of Hall’s theorem [2] (detailed in Appendix B),
giving that n is feasible if and only if:
∀S ⊆C, � nc ≤ U |{b ∈B : S ∩Jb ̸= ∅}| . (2)
c∈S
We now introduce statistical assumptions on request arrivals
and durations. New requests for content c occur at the instants
of a Poisson process with rate νc. We assume that the video
streaming rate is normalized to 1, and is the same for all
contents. We further assume that all videos have the same
duration, again normalized at 1. Under these assumptions, the
amount of work per time unit brought into the system by
content c equals νc.
With the above assumptions at hand, assuming fixed cache
contents, the vector n of requests under service is a particular
instance of a general stochastic process known as a loss
network model. Loss networks were introduced to represent
ongoing calls in telephone networks, and exhibit rich structure.
In particular, the corresponding stochastic process is reversible,
and admits a closed-form stationary distribution. For the
Distributed Server Network model, the stationary distribution
reads:
π(n) = [1] � νc[n][c] (3)
Z c∈C nc! [I][{][n][ is feasible][}][.]
In words, the numbers of requests nc are independent Poisson
random variables with parameter νc, conditioned on feasibility
of the whole vector n.
2In fact the external users issuing requests could keep local copies of
previously accessed content, and hence experience “local service” upon reaccessing the same content. But we do not need consider this as this happens
outside the perimeter of our system.
Our objective is then to determine content placement strategies so that in the corresponding loss network model, the
fraction of rejected requests is minimal. The difficulty in doing
this analysis resides in the fact that the normalizing constant Z
is cumbersome to evaluate. Nevertheless, simplifications occur
under large system asymptotics, which we will exploit in the
next sections.
We conclude this section by the following remark. For simplicity we assumed in the above description that a particular
content is either fully replicated at a peer, or not present at
all, and that a request is served from only one peer. It should
however be noted that we can equally assume that contents
are split into sub-units, which can be placed onto distinct
peers, and downloaded from such distinct peers in parallel
sub-streams in order to satisfy a request. This extension is
detailed in Appendix F.
IV. OPTIMAL CONTENT PLACEMENT IN DISTRIBUTED
SERVER NETWORKS
We first describe a simple adaptive cache update strategy
driven by demand, and show why it converges to a “predetermined” content placement called “proportional-to-product”
strategy. We then establish the optimality of this “proportionalto-product” placement in a large system asymptotic regime.
_A. The Proportional-to-Product Placement Strategy_
A simple method to adaptively update the caches at boxes
driven by demand is described as follows:
**Demand-Driven Cache Update**
Whenever a new request comes, with probability ǫB (ǫ is
chosen such that ǫB 1), the server picks a box b uniformly at
≤
random, and attempts to push content c into this box’s cache. If
c is already in there, do nothing; otherwise, remove a content
selected uniformly at random from the cache.
Since external demands for content c are according to a
Poisson process with rate νc, we find that under the above
simple strategy, content c is pushed at rate ǫνc into a particular
box which is not caching content c. Recall that each box
stores M distinct contents, and let j denote a candidate “cache
state”, which is a size M subset of the full content set . For
C
convenience, let denote the collection of all such j.
J
With the above strategy, the caches at each box evolve
independently according to a continuous-time Markov process.
The rate at which cache state j is changed to j[′], where
j[′] = j + c d for some contents d j, c / j, which
{ } \ { } ∈ ∈
we denote by q(j, j[′]), is easily seen to be q(j, j[′]) = ǫνc/M .
Indeed, content d is evicted with probability 1/M, while
content c is introduced at rate ǫνc.
It is easy to verify that the distribution p( ) given by
p(j) = [1]
Z
� νc, j ∈J, (4)
c∈j
-----
for some suitable normalizing constant Z, verifies the follwing
equation:
p(j)q(j, j[′]) = p(j[′])q(j[′], j), j, j[′] . (5)
∈J
The latter relations, known as the local balance equations,
readily imply that p( ) is a stationary distribution for the above
Markov process; since the process is irreducible, this is the
unique stationary distribution.
Thus, we can conclude that under this cache update
strategy, the random cache state at any box eventually follows
this stationary distribution. This is what we refer to as the
**“proportional-to-product” placement strategy, and it is the**
one we advocate in the Distributed Server Network scenario.
_Remark 1: The customized parameter ǫ should not be too_
large, otherwise the burden on the server will be increased due
to use of “push”. Neither should it be too small, otherwise the
Markov chain will converge too slowly to the steady state.
⋄
Under the cache update strategy, the distribution of cache
contents needs time to converge to the steady state. However,
if we have a priori information about content popularity, we
can use a sampling strategy as an alternative way to directly
generate proportional-to-product content placement in one go.
One method works as follows:
**Sampling-Based Preallocation**
Select successively M contents at random in an i.i.d. fashion, according to the probability distribution {νˆc}, where
νˆc = νc/ [�]c[′]∈C [ν][c][′][ is the normalized popularity. If there are]
duplicate selections of some content, re-run the procedure.
It is readily seen that this yields a sample with the desired
distribution.
An alternative sampling strategy which can be faster than
the one described above when very popular items are present
is given in the Appendix C.
_B. A Loss Network Under Many-User Asymptotics_
We now consider the asymptotic regime called “many user–
**fixed catalogue” scaling: The number of boxes B goes to**
infinity. The system load, defined as
�c∈C [ν][c]
ρ ≜, (6)
BU
is assumed to remain fixed, which is achieved in the present
section by assuming that the content collection is kept fixed,
C
while the individual rates {νc} scale linearly with B. We also
assume that the normalized content popularities {νˆc} remain
fixed as B increases. It thus holds that νc = ˆνcρBU for all
c . Note that although boxes are pure resources rather than
∈C
users, scaling of {νc} with B to infinity actually indicates a
“many-user” scenario.
To analyze the performance of our proposed proportionalto-product strategy, we require that the cache contents are sampled at random according to this strategy and are subsequently
kept fixed. This can either reflect the situation where we use
the previously introduced sampling strategy, or alternatively
the situation where the cache update strategy has already made
the distribution of cache states converge to the steady state, and
occurs at a slower time scale than that at which new requests
arise and complete.
Note that, as B grows large, the right-hand side in the
feasibility constraint (2) verifies, by the strong law of large
numbers,
|{b ∈B : S ∩Jb ̸= ∅}| ∼ B � mj. (7)
j:j∩S̸=∅
Here, {mj} corresponds to a particular content placement
strategy, under which each box holds a size M content set
j with probability mj, and this happens independently over
boxes. Specifically, mj = Z1 �c∈j [ν][ˆ][c][ (where][ Z][ is a nor-]
malizing constant) corresponds to our proportional-to-product
placement strategy.
We now establish a sequence of loss networks indexed by
a large parameter B. For the B[th] loss network, requests for
content c (regarded as “calls of type c”) arrive at rate
∈C
νc(B) = (ρU ˆνc) · B, each “virtual link” S ⊆C has a capacity
WS(B) ≜ (U � mj) · B, (8)
j:j∩S̸=∅
and c represents that virtual link is part of the “route”
∈S S
which serves call of type c.[3] This particular setup has been
identified as the “large capacity network scaling” in Kelly [7].
There, it is shown that the loss probabilities in the limiting
regime where B can be characterized via the analysis
→∞
of an associated variational problem.
We now describe the corresponding results in [7]
relevant to our present purpose. For the B[th] loss
network, consider the problem of finding the mode of
the stationary distribution (3), which corresponds to
maximizing c∈C[(][n](cB) [log][ ν]c(B) − log n(cB) [!)][ over feasible][ n][(][B][)] [.]
[�] (B) (B) (B) (B)
Then, approximate log nc [!][ by][ n]c [log][ n]c − nc according
to Stirling’s formula and replace the integer vector n[(][B][)]
by a real-valued vector x[(][B][)] . This leads to the following
optimization problem:
**[OPT 1]**
maxx[(][B][)] �(x(cB) [log][ ν]c(B) − x(cB) [log][ x](cB) + x(cB) [)] (9)
c∈C
s.t. ∀S ⊆C, � x(cB) ≤ WS(B) (10)
c∈S
over x[(][B][)] 0.
≥
3Note that this construction in fact admits a form of fixed routing which is
equivalently transformed from a dynamic routing model where each particular
box is regarded as a link and calls of type c can use any single-link route
corresponding to a box holding content c. This equivalent transform is based
on the assumption that repacking is allowed (cf. Section 3.3. in [7]). We have
already found this equivalent transform by converting feasibility condition (1)
to (2) in Section III.
-----
The corresponding Lagrangian is given by:
L(x[(][B][)], y[(][B][)] ) = �(xc(B) [log][ ν]c(B) − xc(B) [log][ x]c(B) + x(cB) [)]
c∈C
+ � yS[(][B][)] [(][W]S(B) − � xc(B) [)][,]
S⊆C c∈S
where {yS(B)[}]S⊆C [are Lagrangian multipliers. The KKT con-]
ditions for this convex optimization problem comprise the
original constraints and the following ones:
4), or if the catalogue size C scales with the box population
size B, a case not covered by the classical literature on loss
networks, and to which we turn in Section VI-B.
_Proof: First, we consider ρ_ 1. Letting
≥
− � y¯S[(][B][)]
S:c∈S
= 1/ρ, c, (14)
∀ ∈C
�
exp
�
y¯S[(][B][)][(][W]S(B) − � x¯c(B) [) = 0][,][ ¯][y]S[(][B][)] ≥ 0, ∀S ⊆C,
c∈S
∂L(¯x[(][B][)](B,) ¯y[(][B][)] ) = log νc(B) − log ¯xc(B) − � y¯S[(][B][)] = 0, ∀ c ∈C
∂xc
S:c∈S
(11)
where (¯x[(][B][)], ¯y[(][B][)] ) is a solution to the optimization problem.
From equation (11), we further get
x¯(cB) = νc(B) exp(− � y¯S[(][B][)] [)][,][ ∀] [c][ ∈C][.] (12)
S:c∈S
we have
∀c ∈C, � y¯S[(][B][)] = log ρ. (15)
S:c∈S
Putting equation (15) into (12) leads to
∀c ∈C, ¯x(cB) = νc(B) [/ρ.]
Thus, inequality (10) in OPT 1 becomes
Then the result that we will need from Kelly [7] is the following: for the B[th] loss network, the steady state probability
of accepting request for c, denoted by Ac(B) [, verifies]
∀S ⊆C, � νc(B) ≤ ρ � mjBU. (16)
c∈S j:j∩S̸=∅
Since νc(B) = ρBU · ˆνc and [�]c∈C [ν][ˆ][c][ = 1][, inequality (16)]
further becomes, upon explicitly writing out the normalization
constant Z:
� νˆc ≤ � νˆc · �
c∈G c∈C G: G∩S̸=∅
G⊆C
|G|=M
�
+ O �B[−] [1]2 �, c, (13)
∀ ∈C
∀S ⊆C, � νˆc · �
c∈S G: G⊆C
|G|=M
� νˆc. (17)
c∈G
A(cB) = exp
�
− � y¯S[(][B][)]
S:c∈S
where ¯yS(B) are the Lagrangian multipliers of the previous
optimization problem.
_C. Optimality of Proportional-to-Product Content Placement_
Note that the global acceptance probability, denoted by
Asys, which also reads Asys = [�]c∈C [ν][ˆ][c][A][c][, cannot exceed]
min(1, 1/ρ). Indeed, it is clearly no larger than 1. It cannot
exceed 1/ρ either, otherwise the system would treat more
requests than its available resources.
We now prove that the proportional-to-product content
placement not only achieves the optimal global acceptance
probability Asys = min(1, 1/ρ), but also achieves fair
individual acceptance probabilities, i.e., Ac = Asys for all c.
More precisely, we have the following theorem:
_Theorem 1: By using mj =_ c∈j [ν][ˆ][c][/Z][ for all][ j][ ⊆C]
[�]
s.t. j = M, where Z is the normalizing constant, we have
lim |B→∞| A(cB) = min{1, 1/ρ}, ∀c ∈C, for fixed ρ and C. ⋄
Before giving the proof, we comment on the result. One
point to note is that because of (7), the above optimal
acceptance rate is achieved with probability one under
any random sampling which follows the proportional-toproduct scheme. Secondly, the optimality of the asymptotic
acceptance probability does not depend on M, as long as
M 1. Thus for this particular scaling regime, storage space
≥
is not a bottleneck. As we shall see in the next two sections,
increasing M **does improve performance if either local**
services occur, as in the Pure P2P Network scenario (Section
Two types of product terms (mapped to subsets ) appear
K ⊆C
on both sides:
I. [�]c∈K [ν][ˆ][c][:][ |K|][ =][ M][ + 1][,][ K ∩] [S][ ̸][=][ ∅][.]
II. ([�]c∈K [ν][ˆ][c][)][ ·][ ˆ][ν][c][′][:][ c][′][ ∈K ∩] [S,][ |K|][ =][ M] [.]
To show whether inequality (17) hold, we only have to prove
that given any, for each product term (related to a )
S ⊆C K
which appears in one inequality corresponding to a certain,
S
its multiplicity on the left hand side is no more than that on
the right hand side.
1. For a product term of Type I:
- On the LHS: Since [�]c∈K [ν][ˆ][c][ =][ �]c∈G [ν][ˆ][c][ ·][ ˆ][ν][c][′][ for]
some and c[′], where is a size M
G ⊆C ∈S ∩K G
content set, c[′], and = + c[′] . It is easy to
̸∈G K G { }
see that we have different choice of c[′] in
|S ∩K|
a, so the multiplicity of this product term on the
K
LHS equals .
|S ∩K|
- On the RHS: When 2, for any c[′],
|S ∩K| ≥ ∈K
c[′] is a size M content set of which the intersect
K\{ }
with is not empty, hence the multiplicity equals
S
(= M +1). When = 1, the exception to
|K| |S ∩K|
the above case is that if c[′], then c[′] is
∈S ∩K K\{ }
a size M content set which has no intersect with
S
and is actually impossible to appear in the second
summation term (over all size M content sets s.t.
G
= ) in inequality (17). Thus, the multiplicity
G ∩S ̸ ∅
equals 1 (= M ).
|K| −
From above, we can see that the multiplicity of the
product term on the LHS is always no more than that
on the RHS.
-----
2. For a product term of Type II:
is actually already a size M content set s.t. =
K G G ∩C ̸
. Therefore, it is easy to see that on both sides, the
∅
multiplicities of this product term are both 1.
Now we can conclude that inequality (17) holds for all,
S ⊆C
and continue to check the complementary slackness. Given
ρ 1, one simple solution to equation (15) reads:
≥
∀S ⊆C, ¯yS[(][B][)] = log ρ · I{S=C}. (18)
Besides, inequality (17) is tight for = (we even do not
S C
need to check this when ρ = 1). Therefore, complementary
slackness is always satisfied with solution (18).
So far we have proved that the KKT condition holds when
ρ 1. When ρ < 1, we modify (14) by letting
≥
− � y¯S[(][B][)]
S:c∈S
= 1, c, (19)
∀ ∈C
�
Fig. 2: System loss rates under different traffic loads
10 contents and serve at most U = 4 concurrent requests.
The duration of downloading each content is exponentially
distributed with mean equal to 1 time unit. The parameter ǫ
in the cache update algorithm is set as 1/B such that upon a
request, one box will definitely be chosen for cache update.
For every algorithm, we take the average over 10 independent repetitive experiments, each of which is observed for 10
time units. According to the sample path, the initial 1/5 of the
whole period is regarded as a “warm-up” period and hence
ignored in the calculation of final statistics.[4]
Some implementation details are not captured by our theoretical model, but should be considered in simulations. Upon
a request arrival, the most idle box (i.e., with the largest
number of free connections) among all the boxes which hold
the requested content is chosen to provide the service, for the
purpose of load balancing. If none of them is idle, we use a
heuristic repacking algorithm which iteratively reallocates the
ongoing services among boxes, in order to handle as many
requests as possible while still respects load balancing. One
important parameter which trades off the repacking complexity
and the performance is the maximum number of iterations
t[max]r, which is set as “undefined” by default (i.e., the iterations
will continue until the algorithm terminates; theoretically there
are at most C iterations). Other details regarding the repacking
algorithm can be found in Appendix D. We will see an
interesting observation about t[max]r later.
Figure 2 evaluates system loss rates under different traffic
loads ρ. Our two algorithms SAMP and CU, which target the proportional-to-product placement, both match the
theoretically optimum very well.[5] On the other hand, the
UNIF algorithm, which does not utilize any information about
content popularity, incurs a large loss even if the system is
underloaded (ρ < 1). The gain of proportional-to-product
placement over UNIF becomes less significant as the traffic
4We can get enough samples during each observation period of 10 time
units (for example, when ρ = 1, B = 4000 and U = 4, the average arrivals
would be 160000). It has also been checked that after the warm-up period,
the distribution of cache states well approximates the proportional-to-product
placement and is kept quite stably for the remaining observation period.
5In fact, around ρ = 1, they perform a little worse than the optimum. The
reason is that ρ = 1 is the “critical traffic load” (a separation point between
zero-loss and nonzero-loss ranges), under which the simulation results are
easier to incur deviation from the theoretical value.
exp
�
and hence there is an additional factor 1/ρ > 1 on the RHS
of inequality (17). Since the old version of inequalities (17) is
proved to hold, the new version automatically holds, but none
of them is tight now. However, from (19) we have ¯yS(B) =
0,, which means complementary slackness is always
∀S ⊆C
satisfied (similar to ρ = 1).
Therefore, according to equation (13), it can be concluded
that by using mj = [�]c∈j [ν][ˆ][c][/Z][ for all][ j][, we can achieve]
A(cB) = min{1, 1/ρ} + O �B[−] [1]2 �, ∀c ∈C,
so limB→∞ A(cB) = min{1, 1/ρ}.
_D. Simulation Results_
In this subsection, we use extensive simulations to evaluate
the performances of the two implementable schemes proposed
in Subsection IV-A which follow the “proportional-to-product”
placement strategy, namely the sampling-based preallocation
scheme and the demand-driven cache update (labeled as
**“SAMP” and “CU”, respectively).**
We compare the results with the theoretical optimum (i.e.,
loss rate for each content equals (1 1/ρ)[+]; the curves
−
are labeled as “Optimal”) and a uniform placement strategy
(labeled as “UNIF”) defined as the following: first, permute
all the contents uniformly at random, resulting in a content
sequence {ci}, for 1 ≤ i ≤ C; then, push the M contents
indexed by subsequence {c(j mod C)[}]bM+1≤j≤(b+1)M [into]
the cache of box b, for 1 b B. UNIF is also used to
≤ ≤
generate the initial content placement for CU so that the loss
rate can be reduced during the warm-up period.
If not further specified, the default parameter setting is as
follows: The popularity of contents {νˆc} follows a zipf-like
distribution (see e.g. [4]), i.e.,
(c0 + c)[−][α]
νˆc = (20)
�c[′]∈C[(][c][0][ +][ c][′][)][−][α][,]
with a decaying factor α > 0 and the shift c0 ≥ 0. We use
α = 0.8 and c0 = 0. The content catalogue size C = 500 and
the number of boxes B = 4000. Each box can store M =
-----
80%
SAMP
70% CU
60% UNIF
Optimal
50%
40%
30%
20%
10%
0%
0 0.4 0.8 1.2 1.6 2
α
|SAMP CU UNIF Optimal|Col2|Col3|Col4|Col5|
|---|---|---|---|---|
||||||
Fig. 3: System loss rates with different α (ρ = 1)
Fig. 5: System loss rates with different number of boxes
Fig. 4: Effect of repacking on the system loss rate
load grows, which can be easily expected.
In Figure 3, when the decaying factor α in the zipflike distribution increases, the distribution of placed contents
generated by UNIF has a higher discrepancy from the real
content popularity distribution, so UNIF performs worse. On
the other hand, the two proportional-to-product strategies are
insensitive to the change of content popularity, as we expected.
Figure 4 shows the effect of repacking on the system loss
rate. In sub-figure (a), we find that under SAMP, repacking is
not necessary. In sub-figure (b) which shows the performances
of CU, when ρ is low, one iteration of repacking is sufficient
to make the performance close enough to the optimum; when
ρ is high, repacking also becomes unnecessary. The main takeaway message from this figure is that we can execute a repacking procedure of very small complexity without sacrificing
much performance. The reason is that when the server picks
a box to serve a request, it already respects the rule of load
balancing.
We then explain why CU still needs one iteration of
repacking to improve the performance when ρ is low. Note
that during the cache update, it is possible that the box is
currently uploading the “to-be-kicked-out” content to some
users. If repacking is enabled, those ongoing services can be
repacked to other boxes (see details in Appendix D), but if
t[max]r = 0 (no repacking), they will be terminated and counted
as losses. When ρ is high, however, boxes are more likely to
be busy, which leads to the failure of repacking, so repacking
Fig. 6: Loss rate of requests for each content (ρ = 1)
makes no difference.
Recall that the proportional-to-product placement is only
optimal when the number of boxes B . Figures 5 and
→∞
6 then show the impact of a finite B. In Figure 5, as B
decreases, the system loss rate of every algorithms increases
(compared to the two proportional-to-product strategies, UNIF
is less sensitive to B). In Figure 6, non-homogeneity in the
individual loss rates of requests for each content also reflects
a deviation from the theoretical result (when B, the
→∞
loss rates of the requests for all the contents are proved to be
identical). As expected, increasing the number of boxes (from
4000 to 8000) makes the system closer to the limiting scenario
and the individual loss rates more homogeneous. Another
observation is that as the popularity of a content decreases (in
the figure, the contents are indexed in the descending order of
their popularity), the individual loss rate increases. However,
according to Figure 2, those less popular contents do not affect
the system loss rate much even if they incur high loss, since
their weights {νˆc} are also lower.
In fact, if we choose a smaller content catalogue size C or
a larger cache size M, simulations show the negative impact
of a finite B will be reduced (the figures are omitted here).
This tells us that if C scales with B rather than being fixed,
the proof of optimality under the loss network framework in
Subsection IV-B is no longer valid and M must be a bottleneck
against the performance of the optimal algorithm. We will
solve this problem by introducing a certain type of “large
catalogue model” later in Section VI.
-----
V. OPTIMAL CONTENT PLACEMENT IN PURE
PEER-TO-PEER NETWORKS
In the Pure P2P Network scenario, when box b has a request
for content c which is currently in its own cache, a “local
service” will be provided and no download bandwidth in the
network will be consumed. To simplify our analysis, each
request for a specific content is assumed to originate from
a box chosen uniformly at random (this in particular assumes
identical tastes of all users).
This means that the effective arrival rate of the requests for
content c which generates traffic load actually equals ˜νc ≜
νc(1 − m˜ c), where ˜mc is defined as the fraction of boxes who
have cached content c. Let ρc ≜ ρνˆc denote the traffic load
generated by requests for content c, and λc denote the fraction
of the system bandwidth resources used to serve requests for
content c. Obviously, [�]c∈C [λ][c][ ≤] [1][. The traffic load absorbed]
by the P2P system either via local services or via service from
another box is then upper-bounded by
ρ˜ = � ρc ˜mc + [ρc(1 − m˜ c)] ∧ λc, (21)
c∈C
where “ ” denotes the minimum operator.
∧
We will use this simple upper bound to identify an optimal placement strategy in the present Pure P2P Network
scenario. To this end, we shall establish that our candidate
placement strategy asymptotically achieves this performance
bound, namely absorbs a portion ˜ρ in the limit where B tends
to infinity.
To find the optimal strategy, we introduce a variable
xc ≜ [ρc(1 − m˜ c)] ∧ λc for all c. Note further that the fraction
λc is necessarily bounded from above by ˜mc, as only those
boxes holding c can devote their bandwidth to serving c. It
is then easy to see that the quantity ˜ρ in (21) is no larger
than the optimal value of the following linear programming
problem:
**[OPT 2]**
- For c = c[∗] + 1, ˜mc = λc = xc = 1 − [�]c[c]=[∗] M [m][˜] [c][.]
- For c[∗] + 2 ≤ c ≤ C, ˜mc = λc = xc = 0. ⋄
The proof consists in checking that the KKT conditions
are met for the above candidate solution. Details are given in
Appendix E.
The above optimal solution suggests the following placement strategy:
**“Hot-Warm-Cold” Content Placement Strategy**
Divide the contents into three different classes according to
their popularity ranking (in descending order):
- Hot: The M 1 most popular contents. At each box,
−
M 1 cache slots are reserved for them to make sure
−
that requests for these contents are always met via local
service.
- Warm: The contents with indices from M to c[∗] + 1 (or
c[∗] if c=M [m][˜] [c][ = 1][). For these contents, a fraction][ ˜][m][c]
[�][c][∗]
of all the boxes will store content c in their remaining one
cache slots, where the value of ˜mc is given in Theorem 2.
All requests for these contents (except c[∗] + 1 if it is
classified as “warm”) can be served, at the expense of all
bandwidth resources.
- Cold: The other less popular contents are not cached at
all.
max
m˜,λ,x
�(ρc ˜mc + xc)
c∈C
s.t. ∀ c ∈C, 0 ≤ m˜ c ≤ 1, 0 ≤ λc ≤ m˜ c;
∀ c ∈C, 0 ≤ xc ≤ λc, xc ≤ ρc(1 − m˜ c);
� m˜ c = M, � λc 1.
≤
c∈C c∈C
The following theorem gives the structure of an optimal
solution to OPT 2, and as a result suggests an optimal
placement strategy.
_Theorem 2: Assume that {νˆc} are ranked in descending_
order. The following solution solves OPT 2:
_Remark 2: The requests for the c[∗]_ most popular contents
(“hot” contents and “warm” contents except content c[∗] + 1)
incur zero loss, while the requests for the C c[∗] 1 least
− −
popular contents incur 100% loss. There is a partial loss in
the requests for content c[∗] + 1 if [�]c[c]=[∗] M [m][˜] [c][ <][ 1][.]
Note that the placement for “warm” contents looks like the
“water-filling” solution in the problem of allocating transmission powers onto different OFDM channels to maximize the
overall achievable channel capacity in the context of wireless
communications [16].
⋄
Under this placement strategy, the maximum upper bound
on the absorbed traffic load reads
c[∗]
�
c=M
ρc
1 + ρc
.
�
ρ˜ =
c[∗]
� ρc + (ρc∗+1 + 1)
c=1
�
1
−
- For 1 ≤ c ≤ M − 1, ˜mc = 1, λc = xc = 0.
- For M ≤ c ≤ c[∗], ˜mc = λc = xc = ρc/(1 + ρc), where
c[∗] satisfies that
We then have the following corollary:
_Corollary 1: Considering the large system limit B_,
→∞
with fixed catalogue and associated normalized popularities
{ ˆνc} as considered in Subsection IV-B, the proposed “hotwarm-cold” placement strategy achieves an asymptotic fraction of absorbed load equal to the above upper bound ˜ρ, and
is hence optimal in this sense.
⋄
_Proof: With the proposed placement strategy, hot (respec-_
tively, cold) contents never trigger accepted requests, since all
incoming requests are handled by local service (respectively,
rejected). For warm contents, because each box holds only one
c[∗]
�
c=M
ρc 1, but
≤
1 + ρc
c[∗]+1
�
c=M
ρc
- 1.
1 + ρc
-----
warm content, it can only handle requests for that particular
warm content. As a result, the processes of ongoing requests
for distinct warm contents evolve independently of one another. For a given warm content c, the corresponding number
of ongoing requests behaves as a simple one-dimensional loss
network with arrival rate νc(1 − m˜ c) and service capacity
m˜ cBU . For c = M, . . ., c[∗], one has ˜mc = ρc/(1 + ρc) where
ρc = νc/(BU ), so both the arrival rate and the capacity of
the corresponding loss network equal ˜mcBU . The asymptotic
acceptance probability as B then converges to 1 and
→∞
the accepted load due to both local service and services
from other boxes converges to ρc. For content c[∗] + 1 (if
m˜ c∗+1 > 0), the corresponding loss network has arrival rate
νc∗+1(1−m˜ c∗+1) and service capacity ˜mc∗+1BU . Then, in the
limit B, the accepted load (due to both local services
→∞
and services from other boxes) reads ρc∗+1 ˜mc∗+1 + ˜mc∗+1
(which is actually smaller than ρc[∗]+1). Summing the accepted
loads of all contents yields the result.
VI. LARGE CATALOGUE MODEL
Keeping the many-user asymptotic, we now consider an
alternative model of content catalogue, which we term the
“large catalogue” scenario. The set of contents is divided
C
into a fixed number of “content classes”, indexed by i .
∈I
In class i, all the contents have the same popularity (arrival
rate) νi. The number of contents within class i is assumed
to scale in proportion to the number of boxes B, i.e., class i
contains αiB contents for some fixed scaling factor αi. We
further define α ≜ [�]i [α][i][. With the above assumptions, the]
system traffic load ρ in equation (6) reads
ρ = [1]
U
� αiνi. (22)
i∈I
bottleneck? Is the proportional-to-product placement strategy
still optimal under the large-catalogue scaling?
_A. Necessity of Unbounded Storage_
We first establish that bounded storage will strictly
constrain utilization of bandwidth resources. To this end we
need the following lemma:
_Lemma 1: Consider the system under large catalogue scal-_
ing, with fixed weights αi and cache size M per box. Define
M [′] ≜ 2M/α . Then
⌈ ⌉
(i) More than half of the contents are replicated at most M [′]
times, and
(ii) For each of these contents, the loss probability is at least
E(inf i νi, M [′]U ) > 0, where E(·, ·) is the Erlang function [7]
defined as:
−1
� C ν[n] �
E(ν, C) ≜ [ν][C] � .
C! n!
n=1
⋄
_Proof: We first prove part (i). Note that the total number_
of content replicas in the system equals BM . Thus, denoting
by f the fraction of contents replicated at least M [′] + 1 times,
it follows that fαB(M [′] + 1) BM, which in turn yields
≤
M M
f
≤
α ( 2M/α + 1) 2M + α [<][ 1]2 [,]
⌈ ⌉ [≤]
which implies statement (i).
To prove part (ii), we establish the following general property for a loss network (equivalent to our original system) with
call types j ∈J, corresponding arrival rates νj, and capacity
(maximal number of competing calls) Cl on link ℓ for all
ℓ . We use ℓ j to indicate that the route for calls of type
∈L ∈
j comprises link ℓ. Denoting the loss probability of calls of
type j in such a loss network as pj, we then want to prove
pj ≥ E(νj, Cj[′] [)][,] (23)
where Cj[′] [≜] [min][ℓ][∈][j][ C][ℓ][, i.e., the capacity of the bottleneck]
link on the route for calls of type j.
Note that the RHS of the above inequality is actually the
loss probability of a loss network with only calls of type j
and capacity Cj[′] [. Fixing index][ j][, we define this loss network]
as an auxiliary system and consider the following coupling
construction which allows us to deduce inequality (23): Let Xk
be the number of active calls of type k in the original system
for all k, and let Xj[′] [denote the number of active calls of type]
j in the auxiliary system. Initially, Xj(0) = Xj[′][(0)][. The non-]
zero transition rates for the joint process ({Xk}k∈K, Xj[′] [)][ are]
given by
k ̸= j : Xk → Xk + 1 at rate νk � I{[�]k∋ℓ [X][k][<C][ℓ][}][,]
ℓ∈j
k ̸= j : Xk → Xk − 1 at rate Xk,
(Xj, Xj[′] [)][ →] [(][X][j][ + 1][, X]j[′] [+ 1)] at rate νj[both],
(Xj, Xj[′] [)][ →] [(][X][j][ + 1][, X]j[′][)] at rate νj[ori],
(Xj, Xj[′] [)][ →] [(][X][j][, X]j[′] [+ 1)] at rate νj[aux],
(Xj, Xj[′] [)][ →] [(][X][j][ −] [1][, X]j[′] [−] [1)] at rate Xj, +
(Xj, Xj[′] [)][ →] [(][X][j][, X]j[′] [−] [1)] at rate �Xj[′] [−] [X][j]�,
The primary motivation for this model is mathematical convenience: by limiting the number of popularity values we limit
the “dimensionality” of the request distribution, even though
we now allow for a growing number of contents. It can also be
justified as an approximation, that would result from batching
into a single class all contents with a comparable popularity.
Such classes can also capture the movie type (e.g. thriller,
comedy) and age (assuming popularity decreases with content
age).
We use ˆυi to denote the normalized popularity of content
class i ∈I and it reads i∈I [υ][ˆ][i][ = 1][. It is reasonable to regard]
[�]
each ˆυi as fixed. ˆνi ≜ υˆi/(αiB) represents the normalized
popularity of a specific content in class i, which decreases as
the number of contents in this class αiB increases, since users
now have more choices within each class. In practice, an online
video provider company which uses the Distributed Server
Network architecture adds both boxes and available movies of
each type to attract more user traffic, under a constraint of a
maximum tolerable traffic load ρ.
Returning to the Distributed Server Network model of
Section IV, we consider the following questions: What amount
of storage is required to ensure that memory space is not a
-----
where
νj[both] ≜ νjI{Xj[′] [<C]j[′] [}][ ·] � I{[�]k∋ℓ [X][k][<C][ℓ][}][,]
ℓ∈j
νj[ori] ≜ νjI{Xj[′] [=][C]j[′] [}][ ·] � I{[�]
ℓ∈j
k∋ℓ [X][k][<C][ℓ][}][,]
νj[aux] ≜ νjI{Xj[′] [<C]j[′] [}][ · I][{∃][ℓ][∈][j][ s.t.][ �]
k∈ℓ [X][k][=][C][ℓ][}][.]
It follows from Theorem 8.4 in [5] that {Xk} is indeed a loss
network process with the original dynamics, and that Xj[′] [is]
a one-dimensional loss network with capacity Cj[′] [and arrival]
rate νj. From the construction, we can see that all transitions
preserve the inequality Xj(t) ≤ Xj[′] [(][t][)][ for all][ t][ ≥] [0][, due to the]
following reason: Once Xj increases by 1, Xj[′] [either increases]
by 1 or equals the capacity limit Cj[′] [, and for the latter case, the]
corresponding transition rate νj[ori] implies that Xj ≤ Cj[′] [=][ X]j[′][.]
Similarly, once Xj[′] [decreases by 1, either][ X][j][ also decreases]
by 1, or in the case that Xj does not decrease, it must be that
the transition rate Xj[′] [−] [X][j][ is strictly positive. In any case, the]
above inequality is preserved.
We further let Aj (t), A[′]j[(][t][)][ denote the number of type][ j]
external calls, Lj(t), L[′]j[(][t][)][ the number of type][ j][ call rejec-]
tions, and Dj(t), Dj[′] [(][t][)][ the number of type][ j][ call completions,]
respectively in the original and auxiliary systems, during time
interval [0, t]. It follows from our construction that whenever
the service for a call of type j completes in the original
system, the service for a call of type j also completes in the
auxiliary system, hence Dj(t) ≤ Dj[′] [(][t][)][ for all][ t][ ≥] [0][. Since]
Xj(t) = Aj(t)−Dj(t)−Lj(t), Xj[′] [(][t][) =][ A][′]j[(][t][)][−][D]j[′] [(][t][)][−][L][′]j[(][t][)]
and Aj(t) = A[′]j[(][t][)][, we have][ L][j][(][t][)][ ≥] [L][′]j[(][t][)][. Upon dividing]
this inequality by A(t) and letting t tend to infinity, one
retrieves the announced inequality (23) by the ergodic theorem.
Back to the context of our P2P system, for those contents
which are replicated at most M [′] times (i.e., the contents
considered in part (i)), the rejection rate of content c of type
j reads pj ≥ E(inf i νi, Cj[′] [)][ ≥] [E][(inf] [i][ ν][i][, M][ ′][U] [)][.]
The above lemma readily implies the following corollary:
_Corollary 2: Under the assumptions in Lemma 1, The over-_
all rejection probability is at least [1]
2 [E][(min][i][ ν][i][, M][ ′][U] [)][. Indeed,]
for bounded M, M [′] is also bounded, and E(mini νi, M [′]U )
is bounded away from 0.
⋄
Thus, even when the system load ρ is strictly less than 1,
with bounded M there is a non-vanishing fraction of rejected
requests, hence a suboptimal use of bandwidth.
_B. Efficiency of Proportional-to-Product Placement_
We consider the following “Modified Proportional-to**Product Placement”: Each of the M storage slots at a given**
box b contains a randomly chosen content. The probability of
selecting one particular content c is νi/(ρBU ) if it belongs to
class i. In addition, we assume that the selections for all such
MB storage slots are done independently of one another.
_Remark 3: This content placement strategy can be viewed_
as a “balls-and-bins” experiment. All the MB cache slots in
the system are regarded as balls, and all the |C| (= i [α][i][B][)]
[�]
contents are regarded as bins. We throw each of the MB
balls at random among all the bins. Bin c (corresponding
|C|
to content c which belongs to class i) will be chosen with
probability νi/(ρBU ). Alternatively, the resulting allocation
can be viewed as a bipartite random graph connecting boxes
to contents.
⋄
Note that this strategy differs from the “proportional-toproduct” placement strategy proposed in Section IV, in that
it allows for multiple copies of the same content at the same
box. However, by the birthday paradox, we can prove the
following lemma which shows that up to a negligible fraction
of boxes, the above content placement does coincide with the
proportional-to-product strategy.
_Lemma 2: By using the above content placement strategy,_
at a certain box, if M ≪ �(mini αi)B,
Pr(all the M cached contents are different) 1. (24)
≈
⋄
_Proof: In the birthday paradox, if there are m people_
and n equally possible birthdays, the probability that all the
m people have different birthdays is close to 1 whenever
m n. Here in our problem, at a certain box, the M
≪ [√]
cache slots are regarded as “people” and the contents are
|C|
regarded as “birthdays.” Although the probability of picking
one content is non-uniform, the probability of picking one
content within a specific class is uniform. One can think of
picking a content for a cache slot as a two-step process: With
probability αiνi/ [�]j [α][j][ν][j][, a content in class][ i][ is chosen. Then]
conditioned on class i, a specific content is chosen uniformly
at random among all the αiB contents in class i.
Contents from different classes are obviously different.
When M ≪ [√]αiB, even if all the M cached contents are
from class i, the probability that they are different is close to
1. Thus, M ≪ [√]mini αiB is sufficient for (24) to hold.
To prove that under this particular placement, inefficiency
in bandwidth utilization vanishes as M, we shall in
→∞
fact consider a slight modification of the “request repacking”
strategy considered so far for determining which contents to
accept:
**Counter-Based Acceptance Rule**
A parameter L > 0 is fixed. Each box b maintains at all
times a counter Zb of associated requests. For any content
c, the following procedure is used by the server whenever a
request arrives: A random set of L distinct boxes, each of
which holds a replica of content c, is selected. An attempt is
made to associate the newly arrived request with all L boxes,
but the request will be rejected if its acceptance would lead
any of the corresponding box counters to exceed LU .
-----
_Remark 4: Note that in this acceptance rule, associating a_
request to a set of L boxes does not mean that the requested
content will be downloaded from all these L boxes. In fact,
as before, the download stream will only come from one of
the L boxes, but here we do not specify which one is to be
picked.
It is readily seen that the above rule defines a loss network.
Moreover, it is a stricter acceptance rule than the previously
considered one. Indeed, it can be verified that when all ongoing
requests have an associated set of L boxes, whose counters
are no larger than LU, there exist nonnegative integers Zcb
such that b:c∈Jb [Z][cb][ =][ Ln][c][,][ ∀] [c][ ∈C][ and][ �]c:c∈Jb [Z][cb][ ≤]
[�]
LU, b, then feasibility condition (2) holds a fortiori.
∀ ∈B ⋄
We introduce an additional assumption, needed for technical
reasons.
_Assumption 1: A content which is too poorly replicated is_
never served. Specifically, a content must be replicated at
**least M** [3][/][4] **times to be eligible for service.**
⋄
Our main result in this context is the following theorem:
_Theorem 3: Consider fixed M_, αi, νi, and corresponding
load ρ < 1. Then for suitable choice of parameter L, with
high probability (with respect to placement) as B, the
→∞
loss network with the above “modified proportional-to-product
placement” and “counter-based acceptance rule” admits a
content rejection probability φ(M ) for some function φ(M )
decreasing to zero as M .
→∞ ⋄
The interpretation of this theorem is as follows: The fraction of lost service opportunities, for an underloaded system
(ρ < 1), vanishes as M increases. Thus, while Corollary 2
showed that M is necessary for optimal performance,
→∞
this theorem shows that it is also sufficient: there is no need
for a minimal speed (e.g. M log B) to ensure that the loss
≥
rate becomes negligible.
The proof is given in Appendix A.
VII. CONCLUSION
In peer-to-peer video-on-demand systems, the information
of content popularity can be utilized to design optimal content
placement strategies, which minimizes the fraction of rejected
requests in the system, or equivalently, maximizes the utilization of peers’ uplink bandwidth resources. We focused
on P2P systems where the number of users is large. For
the limited content catalogue size scenario, we proved the
optimality of a proportional-to-product placement in the Distributed Server Network architecture, and proved optimality
of “Hot-Warm-Cold” placement in the Pure P2P Network
architecture. For the large content catalogue scenario, we also
established that proportional-to-product placement leads to
optimal performance in the Distributed Server Network. Many
interesting questions remain. To name only two, more general
popularity distributions (e.g. Zipf) for the large catalogue
scenario could be investigated; the efficiency of adaptive cache
update rules such as the one discussed in Section IV-A, or
classical alternatives such as LRU, in conjunction with a loss
network operation, also deserves more detailed analysis.
REFERENCES
[1] J. M. Almeida, D. L. Eager, M. K. Vernon, and S. J. Wright. Minimizing
delivery cost in scalable streaming content distribution systems. IEEE
_Transactions on Multimedia, 6(2):356–365, Apr. 2004._
[2] B. Bollob´as. Modern Graph Theory. Springer, New York, 1998.
[3] Y. Boufkhad, F. Mathieu, F. de Montgolfier, D. Perino, and L. Viennot.
Achievable catalog size in peer-to-peer video-on-demand systems. In
_Proc. of the Seventh International Workshop on Peer-to-Peer Systems_
_(IPTPS), 2008._
[4] L. Breslau, P. Cao, L. Fan, G. Philips, and S. Shenker. Web caching
and zipf-like distributions: Evidence and implications. In Proc. of IEEE
_INFOCOM, Mar. 1999._
[5] M. Draief and L. Massoulie. Epidemics and rumours in complex
networks. In the London Mathematical Society Lecture Note Series.
Cambridge University Press, 2010.
[6] J. Kangasharju, K. W. Ross, and D. A. Turner. Optimizing file availability in peer-to-peer content distribution. In Proc. of IEEE INFOCOM,
2007.
[7] F. Kelly. Loss networks. The Annals of Applied Probability, 1(3):319–
378, 1991.
[8] A. Klenke and L. Mattner. Stochastic ordering of classical discrete
distributions. Advances in Applied probability, 42(2):392–410, 2010.
[9] N. Laoutaris, V. Zissimopoulos, and I. Stavrakakis. On the optimization of storage capacity allocation for content distribution. Computer
_Networks, 47:409–428, 2003._
[10] M. Mitzenmacher and E. Upfal. Probability and computing: randomized
_algorithms and probabilistic analysis._ Cambridge University Press,
2005.
[11] K. Suh, C. Diot, J. Kurose, L. Massoulie, C. Neumann, D. Towsley,
and M. Varvello. Push-to-peer video-on-demand system: Design and
evaluation. _IEEE Journal on Selected Areas in Communications,_
25(9):1706–1716, 2007.
[12] B. R. Tan and L. Massoulie. Brief announcement: Adaptive content
placement for peer-to-peer video-on-demand systems. In Proc. of 29th
_Annual ACM SIGACT-SIGOPS Symposium on Principles of Distributed_
_Computing (PODC), Jul. 2010._
[13] B. R. Tan and L. Massoulie. Optimal content placement for peer-to-peer
video-on-demand systems. In Proc. of IEEE INFOCOM, Apr. 2011.
[14] S. Tewari and L. Kleinrock. On fairness, optimal download performance
and proportional replication in peer-to-peer networks. In Proc. of IFIP
_Networking, 2005._
[15] S. Tewari and L. Kleinrock. Proportional replication in peer-to-peer
networks. In Proc. of IEEE INFOCOM, 2006.
[16] D. Tse and P. Viswanath. Fundamentals of Wireless Communication.
Cambridge University Press, 2005.
[17] V. Valancius, N. Laoutaris, L. Massoulie, C. Diot, and P. Rodriguez.
Greening the internet with nano data centers. In Proc. of the 5th interna_tional conference on Emerging networking experiments and technologies_
_(CoNEXT), pages 37–48, 2009._
[18] J. Wu and B. Li. Keep cache replacement simple in peer-assisted VoD
systems. In Proc. of IEEE INFOCOM Mini-Conference, 2009.
[19] X. Zhou and C.-Z. Xu. Efficient algorithms of video replication and
placement on a cluster of streaming servers. Journal of Network and
_Computer Applications, 30(2):515–540, Apr. 2007._
APPENDIX
_A. Proof of Theorem 3_
The proof has five sequential stages:
_1) The chance for a content to be “good”_
Let Nc denote the number of replicas of content c of class
i. Then, Nc admits a binomial distribution with parameters
(MB, ρBUνi [)][. We call content][ c][ a “good” content if][ |][N][c][ −]
E[Nc]| < M [2][/][3], i.e.,
Nc − νρUiM < M 2/3. (25)
���� ����
-----
As Nc = [�]i[MB]=1 [Z][i][, where][ Z][i][ ∼] [Ber][(][p][)][ (][p][ ≜] ρBUνi [) are i.i.d.,]
according to the Chernoff bound,
�
Pr Nc ≥ M [2][/][3] + [ν]ρU[i][M]
�
e[−][MB][·][I][(][a][)], (26)
≤
where a ≜ �M [2][/][3] + [ν]ρU[i][M] � /MB and I(x) ≜ supθ{xθ −
ln(E[e[θZ][i] ]) is the Cram´er transform of the Bernoulli random
}
variable Zi. Instead of directly deriving the RHS of inequality
(26), which can be done but needs a lot of calculations (see
Appendix G), we upper bound it by using a much simpler
approach here: For the same deviation, a classical upper bound
on the Chernoff bound of a binomial random variable is
provided by the Chernoff bound of a Poisson random variable
which has the same mean (see e.g. [5]). Therefore, the RHS
of inequality (26) can be upper bounded by
� � ρU ��
exp 1 +,
− [ν][i][M]
ρU [·][ ˆ][I] νiM [1][/][3]
where I[ˆ](x) is the Cram´er transform of a unit mean Poisson
random variable, i.e., I[ˆ](x) = x log x x + 1. By Taylor’s
−
expansion of I[ˆ](x) at x = 1, the exponent in the last expression
is equivalent to
� ρU
νiM [1][/][3]
2
�
+ o �M [−][2][/][3][��]
− [ν][i][M]
ρU
[·]
� 1
2
= M [1][/][3] + o �M [1][/][3][�] = Θ �M [1][/][3][�] .
− [ρU] −
2νi
On the other hand, when M is large, −M [2][/][3] + [ν]ρU[i][M] ≥ 0
holds, hence we have
�
Pr Nc ≤−M [2][/][3] + [ν]ρU[i][M]
�
e[−][MB][·][I][ˆ][(ˆ][a][)], (27)
≤
memory slot at a particular box (we index a slot by j for
1 j MB), and f (ξ) corresponds to the number of good
≤ ≤
contents in class i based on the placement ξ, i.e., Xi = f (ξ).
It is easy to see that in our case c = 1, hence we have
Pr(|Xi − E[Xi]| ≥ t) ≤ 2e[−][2][t][2][/][(][MB][)], ∀t > 0.
Taking t = (MB)[2][/][3] in the above inequality further yields
�
Pr |Xi − E[Xi]| ≥ (MB)[2][/][3][�] ≤ 2e[−][2(][MB][)][1][/][3].
Thus, we have
� �
Pr Xi ≥ 1 − 2e[−][Θ(][M] [1][/][3][)][�] - αiB − (MB)[2][/][3][�]
(a)
�
≥ Pr Xi ≥ E[Xi] − (MB)[2][/][3][�]
�
≥ Pr |Xi − E[Xi]| < (MB)[2][/][3][�]
1 2e[−][2(][MB][)][1][/][3], (29)
≥ −
where (a) holds since
E[Xi] = Pr(content c is good) · αiB
�
≥ 1 − 2e[−][Θ(][M] [1][/][3][)][�] - αiB.
Note that in order for the lower bound on Xi shown in the
above probability to be Θ(B), M o(B[1][/][2]) is a sufficient
∼
condition.
_3) The chance for a box to be “good”_
We call a replica “good” if it is a replica of a good content,
and use Ci to denote the number of good replicas of class i.
We also call a box “good” if the number of good replicas of
class i held by this box lies within
αiνiM O(M [2][/][3]).
±
ρU
As we did for “good contents,” we will also use the Chernoff
bound to prove that a box is good with high probability.
Let Ei represent an event that the number Xi of good
contents within class i satisfies
�
Xi ≥ 1 − 2e[−][Θ(][M] [1][/][3][)][�] αiB − (MB)[2][/][3], (30)
which has a probability of at least 1 2e[−][Ω((][MB][)][1][/][3][)], accord−
ing to inequality (29) when M o(B[1][/][2]). Conditional on
∼
Ei, according to the lower bound in inequality (25) (i.e., the
definition of “good contents”) and inequality (30), we have
� νiM ���
Ci ≥ 1 − 2e[−][Θ(][M] [1][/][3][)][�] αiB
ρU
[−] [M][ 2][/][3]
�
(MB)[2][/][3]
−
= Pr
�MB
� Zˆi ≥ MB · ˆa
i=1
�
where (−Z[ˆ]i) ∼ Ber(p), ˆa ≜ M [−][1][/][3]/B −p ∈ [−1, 0] when B
is large, and it is easy to check that I[ˆ](ˆa) = I( aˆ). Similarly
−
as above by upper bounding e[−][MB][·][I][(][−][a][ˆ][)], we can find that the
exponent of the upper bound is also Θ �M [1][/][3][�]. Therefore,
−
Pr(content c is good) 1 2e[−][Θ(][M] [1][/][3][)]. (28)
≥ −
_2) The number of “good contents” in each class_
Denoting by Xi the number of good contents in class i, we
want to use a corollary of Azuma-Hoeffding inequality (see
e.g. Section 12.5.1 in [10] or Corollary 6.4 in [5]) to upper
bound the chance of its deviation from its mean. This corollary
applies to a function f of independent variables ξ1, . . ., ξn, and
states that if the function changes by an amount no more than
some constant c when only one component ξi has its value
changed, then for all t > 0,
Pr( f (ξ) E[f (ξ)] t) 2e[−][2][t][2][/][(][nc][2][)].
| − | ≥ ≤
Back to our problem, each independent variable ξj correspond to the choice of a content to be placed in a particular
On the other hand, from the upper bound in inequality (25)
and the fact Xi ≤ αiB, we obtain that
Ci ≤ MB · [α]ρU[i][ν][i] �1 + O(M [−][1][/][3])� . (32)
= MB [α][i][ν][i]
ρU
�1 O(M [−][1][/][3] + M [2][/][3]B[−][1][/][3])� .
−
(31)
-----
Conditional on Ei, to constitute a box, sample without replacement from the determined content replicas. Denote the
number of good replicas of class i stored in a particular box
(say, box b) by ζi, which actually represents the number of
good replicas in the M samples sampled without replacement
from all the MB replicas, among which Ci are good ones
(conditional on Ei). This means that, conditional on Ei, ζi
follows a hypergeometric distribution H(MB, Ci, M ). It can
be found that (see e.g. Theorem 1 in [8]) conditional on Ei,
Hi ≤st ζi ≤st Gi. Here, “≤st” represents stochastic ordering,
and
from inequality (34), we further have
Pr ζi − αiρUνiM ≥ O(M 2/3)�
����� ����
≤ 2e[−][Θ(][M] [1][/][3][)] - Pr(Ei) + (1 − Pr (Ei))
= 1 − (1 − 2e[−][Θ(][M] [1][/][3][)]) Pr (Ei)
1 (1 2e[−][Θ(][M] [1][/][3][)])(1 2e[−][Ω((][MB][)][1][/][3][)])
≤ − − −
= 2e[−][Θ(][M] [1][/][3][)] 2e[−][Ω((][MB][)][1][/][3][)]. (35)
−
�
Gi ∼ Bin M, [α][i][ν][i]
ρU
�
Hi ∼ Bin M, [α][i][ν][i]
ρU
�1 + O(M [−][1][/][3])�[�],
�1 O(M [−][1][/][3] + M [2][/][3]B[−][1][/][3])�[�],
−
where the second parameters of the distributions of Gi and
Hi are determined according to inequalities (32) and (31)
respectively.
We will see why we need these two “binomial bounds” on
ζi. By definition,
Pr(box b is not good)
= Pr ��
i∈I
αiνiM
ζi − ρU
�����
O(M 2/3)�[�]
≥
����
≤ � Pr ζi − αiρUνiM
i∈I �����
where for all i,
∈I
O(M 2/3)�, (33)
≥
����
Putting inequality (35) back to inequality (33) immediately
results in
Pr(box b is good) 1 2 e[−][Θ(][M] [1][/][3][)]. (36)
≥ − |I|
_4) The number of “good boxes”_
We use a similar approach as in Stage 2 to bound the
number of good boxes, say Y, which can be represented as a
function g(ξ) where ξ = (ξ1, ξ2, · · ·, ξMB) is the same content
placement vector defined in Stage 2. Still, g(ξ) changes by an
amount no more than 1 when only one component ξi has its
value changed, then for all t > 0, Pr( Y E[Y ] t)
| − | ≥ ≤
2e[−][2][t][2][/][(][MB][)], and taking t = (MB)[2][/][3] further yields
�
Pr Y E[Y ] (MB)[2][/][3][�] 2e[−][2(][MB][)][1][/][3].
| − | ≥ ≤
Similarly as we obtain inequality (29), we finally come to
� �
Pr Y B 1 2 e[−][Θ(][M] [1][/][3][)][��] 1 2e[−][2(][MB][)][1][/][3].
≥ − |I| ≥ −
(37)
_5) The performance of a loss network_
Finally, consider the performance of the loss network
defined by the “Counter-Based Acceptance Rule.” We
introduce an auxiliary system to establish an upper bound
on the rejection rate. In the auxiliary system, upon arrival
of a request for content c, L different requests are mapped
to L distinct boxes holding a replica of c, but here they are
accepted or rejected individually rather than jointly. Letting
Zb (respectively, Zb[′][) denote the number of requests associated]
to box b in the original (respectively, auxiliary) system, one
readily sees that Zb ≤ Zb[′] [at all times and all boxes and for]
each box b, the process Zb[′] [evolves as a one-dimensional loss]
network. We now want to upper bound the overall arrival rate
of requests to a good box:
_(a) Non-good contents_
Assume that upon a request arrival, we indeed pick L
content replicas, rather than L distinct boxes holding the
requested content (as specified in the acceptance rule). This
entails that, if two replicas of this content are present at one
box, then this box can be picked twice. However, since a
vanishing fraction of boxes will have more than one replicas
of the same content when M ≪ �(mini αi)B (as proved
in Lemma 2), we can strengthen the definition of a “good”
box to ensure that, on top of the previous properties, a good
box should hold M distinct replicas. It is easy to see that the
Pr ζi − αiρUνiM ≥ O(M 2/3)�
����� ����
= Pr ζi − αiρUνiM ≥ O(M 2/3), Ei�
����� ����
+ Pr ζi − αiρUνiM ≥ O(M 2/3), Eic�
����� ����
≤ Pr ζi − αiρUνiM ≥ O(M 2/3) Ei� - Pr (Ei)
����� ���� ����
+ Pr (Ei[c][)][ .] (34)
By definition of stochastic ordering,
Pr ζi − αiρUνiM ≥ O(M 2/3) Ei
����� ���� ����
� �
≤ Pr Gi ≥ [α][i]ρU[ν][i][M] + O(M [2][/][3])
�
� �
+ Pr Hi ≤ [α][i]ρU[ν][i][M] − O(M [2][/][3])
(a)
2e[−][Θ(][M] [1][/][3][)],
≤
where (a) can be obtained using a similar Chernoff bounding
approach as for Nc in Stage 1 of this proof. Thus, continuing
-----
fraction of good boxes will still be of the same order as with
the original weaker definition.
With these modified definitions, consider one non-good
content c of class i cached at a good box. Its unique replica
will be picked with probability L/Nc when the sampling of
L replicas among the Nc existing ones is performed. Thus,
since we ignore requests for all content c with Nc ≤ M [3][/][4]
(according to Assumption 1), the request rate will be at most
νiLM [−][3][/][4].
Besides, there are at most O(M [2][/][3]) non-good content
replicas held by one good box. The reason is as follows: By
definition, a good box holds at least
� � αiνiM O(M [2][/][3])� = M O(M [2][/][3]) (38)
− −
ρU
i∈I
good content replicas among all classes, so the remaining
slots, being occupied by non-good content replicas, are at most
O(M [2][/][3]). Therefore, the overall arrival rate of requests for
non-good contents to a good box is upper bounded by
νnon-good = O(M [2][/][3] - LM [−][3][/][4]) = O(LM [−][1][/][12]). (39)
_(b) Good contents_
The rate generated by a good content c of class i is νiL/Nc.
Now, by definition of a good content, one has:
Nc ≥ [ν]ρU[i][M] [(1][ −] [O][(][M][ −][1][/][3][))][.]
network, say E(λ, C), as a certain conditional probability of
S Poi(λ), i.e.,
∼
E(λ, C) = Pr(S = C S C) = [Pr(][S][ =][ C][)]
| ≤
Pr(S C) [.]
≤
Using the Chernoff bound, we have Pr(S C) e[−][λI][(][C/λ][)],
≥ ≤
where I(x) = x log x x + 1, hence
−
Pr(S C) e[−][λI][(][C/λ][)]
≥
E(λ, C)
≤
1 − Pr(S ≥ C) [≤] 1 − e[−][λI][(][C/λ][)][ .]
This entails that the rate of requests for this content is upper
bounded by
ρLU
M [(1 +][ O][(][M][ −][1][/][3][))][.]
By definition of a “good box,” there are at most αiνiM/ρU +
O(M [2][/][3]) good content replicas of class i cached in this good
box. Therefore, the overall arrival rate of requests for good
contents to a good box is upper bounded by
νgood = �
i∈I
� ρLU �
M [(1 +][ O][(][M][ −][1][/][3][))]
� αiνiM + O(M [2][/][3])�
×
ρU
= (ρLU )(1 + O(M [−][1][/][3])). (40)
To conclude, for any good box b, the process Zb[′] [evolves]
as a one-dimensional loss network with arrival rate no larger
than
ν = νnon-good + νgood = ρLU + O(LM [−][1][/][12]),
by combining the two results in (39) and (40).
Next, we are going to upper bound the loss probability
of Zb[′][. Since][ ν][ is an upper bound on the arrival rate, the]
probability that Zb[′] [=][ LU][ is upper bounded by][ E][(][ρLU][ +]
O(LM [−][1][/][12]), LU ). One can actually further upper bound this
Erlang function by e[−][Θ(][L][)]. To see this, let us first rewrite
the loss probability (Erlang function) of a general 1-D loss
Back to the Erlang function in our problem, I(C/λ) = I((ρ +
O(M [−][1][/][12]))[−][1]), hence,
Pr(Zb[′] [=][ LU] [)][ ≤] [E][(][ρLU][ +][ O][(][LM][ −][1][/][12][)][, LU] [)][ ≤] [e][−][Θ(][L][)][,]
(41)
where the second inequality holds under the assumption that
ρ < 1 (otherwise, the exponent will become 0 or +Θ(L)).
The number of good replicas in good boxes is, due to
inequality (37) and equation (38), at least MB(1 O(M [−][1][/][3])),
−
with a high probability (at least 1 2e[−][2(][MB][)][1][/][3]). On the other
−
hand, the total number of replicas of good contents is at most
MB, which is the total number of replicas (or available cache
slots).
Now pick some small ǫ (0, 1/3) and let X[˜] denote the
∈
number of good contents which have at least M [2][/][3+][ǫ] replicas
outside good boxes. Then necessarily, with a probability of at
least 1 2e[−][2(][MB][)][1][/][3],
−
XM˜ [2][/][3+][ǫ] MB MB(1 O(M [−][1][/][3])) = O(BM [2][/][3]),
≤ − −
i.e., X[˜] O(BM [−][ǫ]). According to inequality (29), the total
≤
number of good contents is Θ(B) (specifically, very close to
= αB) with a probability of at least 1 2 e[−][2(][MB][)][1][/][3],
|C| − |I|
hence we can conclude that, with high probability, for a
fraction of at least 1 O(M [−][ǫ]) of good contents, each of them
−
has at least a fraction 1 O(M [−][1][/][3+][ǫ]) of its replicas stored
−
in good boxes (since a good content has νi
ρU [M][ ±][ O][(][M][ 2][/][3][)]
replicas in total by definition). We further use [˜] to represent
C
the set of such contents.
Recall that Ac was defined in Subsection IV-B as the steadystate probability of accepting a request for content c in the
original system. For all c,
∈ C[˜]
Ac ≥ Pr(all the L sampled replicas are in good boxes)
× Pr(Zb < LU, ∀b s.t. box b is sampled)
(a) L
�1 O(M [−][1][/][3+][ǫ])�
≥ −
× Pr(Zb[′] [< LU,][ ∀][b s.t.][ box][ b][ is sampled][)][.]
(b) L
�1 O(M [−][1][/][3+][ǫ])� �1 Le[−][Θ(][L][)][�] .
≥ − - −
(42)
Here, (b) is obtained according to inequality (41). The argument why (a) holds is as follows: We have Nc ≈ νiM/(ρU )
replicas (assuming that content c is of class i), among which
Nc[′] [=][ N][c][(1][ −] [O][(][M][ −][1][/][3+][ǫ][))][ are in good boxes. Then, the]
-----
probability that L samples fall in the good boxes can be written
explicitly as
Nc[′][(][N][ ′]c [−] [1)][ · · ·][ (][N][ ′]c [−] [L][ + 1)]
Nc(Nc − 1) · · · (Nc − L + 1) [,]
which can be approximated as the first part on the RHS we
write above, under the assumption that L M . The second
≪
part is due to the fact that Zb[′] [≤] [Z][b][ for all box][ b][.]
It should be recalled that within this stage of proof, finally
coming to inequality (42) actually needs everything to be
conditional on the following events:
- The number of good boxes is Θ(B);
- The number of good contents is Θ(B);
- A box caches M distinct replicas,
and as B, M →∞ and M ≪ �(mini αi)B, all of them
p
have high probabilities. Additionally, [˜] as B, M .
C →C →∞
Therefore, further letting L but keeping L M [1][/][3][−][ǫ],
→∞ ≪
we will find that the RHS of inequality (42) is approximated
as
1 O(LM [−][1][/][3+][ǫ]) Le[−][Θ(][L][)] 1,
− − ≈
and then conclude that the requests for almost all the contents
will have near-zero loss.
_B. Proof of Equivalence between Feasibility Conditions (1)_
_and (2)_
_1) Sufficiency of Condition (2): We use Hall’s theorem to_
prove the sufficiency.
**[Hall’s theorem] Suppose J = {J1, J2, · · · } is a collection of**
sets (not necessarily countable). A SDR (“System of Distinct
Representatives”) for J is defined as X = {x1, x2, · · · },
where xi ∈ Ji. Then, there exists a SDR (not necessarily
unique) iff. meets the following condition:
J
, � A . (43)
∀T ⊆J |T | ≤| |
A∈T
⋄
In our P2P VoD system, denote the content set as =
C
{c1, c2, · · ·, cN }. Given the ongoing download services of
each content {ni}i[N]=1[, we get a “distinguishable content set”]
C¯ = {c[(1)]1 [, c]1[(2)][,][ · · ·][, c]1[(][n][1][)]; c[(1)]2 [, c]2[(2)][,][ · · ·][, c]1[(][n][2][)]; · · · ;
c[(1)]N [, c]N[(2)][,][ · · ·][, c]N[(][n][N] [)]},
where c[(]i[k][)] represents the k-th download service of content i
for 1 ≤ k ≤ ni, and has its “potential connection set”
Ji[(][k][)] = {lb[(][j][)] : 1 ≤ j ≤ U, ci ∈ b, b ∈B},
i.e., the set of all the connections of those boxes which have
content ci. A collection of the “potential connection sets” for
all {c[(]i[k][)]} is then
J = {J1[(1)][, J]1[(2)][,][ · · ·][, J]1[(][n][1][)]; · · · ; JN[(1)][, J]N[(2)][,][ · · ·][, J]N[(][n][N] [)]},
and a SDR for is
S
X = {x[(1)]1 [, x]1[(2)][,][ · · ·][, x]1[(][n][1][)]; · · · ; x[(1)]N [, x]N[(2)][,][ · · ·][, x][(]N[n][N] [)]},
s.t. x[(]i[k][)] ∈ Ji[(][k][)], which means each c[(]i[k][)] is affiliated with a
distinct connection (i.e., a feasible solution in our model).
Now we want to prove the existence of such a SDR, i.e.,
to prove equation (43). For, there is a one-to-one
∀T ⊆J
mapping between and a [¯] . Further, this [¯] can be
T S ⊆ C[¯] S
mapped to a where
S ⊆C
S = {ci : ∃1 ≤ k ≤ ni, s.t. c[(]i[k][)] ∈ S}[¯],
i.e., is the set of all contents considered in [¯] without
S S
considering multiple services of each content. Then,,
∀T ⊆J
RHS = | � Ji[(][k][)]| = � U
Ji[(][k][)]∈T b:∃ci∈S s.t. ci∈b
and
Therefore, if
= U |{b ∈B : S ∩Jb ̸= ∅}|
LHS = |T | = |S| ≤[¯] � ni.
ci∈S
∀S ⊆C, � ni ≤= U |{b ∈B : S ∩Jb ̸= ∅}|
ci∈S
holds, then equation (43) holds. The sufficiency is proved.
_2) Necessity of Condition (2): For any_,
S ⊆C
� nc = � � Zcb = � � Zcb
c∈S c∈S b:c∈Jb b: ∃c∈S c∈S∩Jb
s.t. c∈Jb
(a)
≤ � U = U |{b ∈B : S ∩Jb ̸= ∅}|,
b: ∃c∈S s.t. c∈Jb
where the inequality (a) is due to the second constraint in
condition (1). Hence, the necessity is proved.
_C. Approximation to Proportional-to-Product Placement Us-_
_ing Bernoulli Sampling_
An alternative sampling strategy to get the proportional-toproduct placement is as follows:
To push contents to box b (1 b B), the server will
≤ ≤
1. Generate C independent Bernoulli random variables
Xc ∼ Ber(pc) for all c ∈C, where pc = βνˆc/(1 + βνˆc),
νˆc is the normalized version of νc, and β is a customized
constant parameter.
2. If [�]c∈C [X][c][ =][ M][ (which means a valid cluster of size]
M is generated), push content c to box b if Xc = 1;
Otherwise, go back to Step 1.
We now analyze why this scheme works: after generating a
valid size-M subset, the probability that this subset is a certain
-----
subset Gj equals
Pr(Xc = 1, ∀c ∈Gj ; Xc = 0, ∀c ̸∈Gj| � Xc = M )
c∈C
�c∈Gj [p][c][ ·][ �]c̸∈Gj [(1][ −] [p][c][)]
=
Pr([�]c∈C [X][c][ =][ M] [)]
= � pc � �c∈C [p][c]
c∈Gj 1 − pc Pr([�]c∈C [X][c][ =][ M] [)]
�
= � νˆc/Z,
c∈Gj
where Z = Pr([�]c∈C [X][c][ =][ M] [)][/][(][β][M][ �]c∈C [p][c][)][, which actu-]
ally equals the normalizing factor for [�]c∈Gj [ν][ˆ][c][.]
We then consider the computational complexity of this
approximation algorithm. Assuming that {νˆc} is sorted in the
descending order, we have
C
� (1 − pc)
c=M+1
Pr(� Xc = M ) ≥
c∈C
M
� pc ·
c=1
= �Cc�=1Mc[(1 +]=1 [β][ β][ν][ˆ][c][ν][ˆ][c][)] ≜ P [∗].
So the computational complexity is upper bounded by
O(BC/P [∗]). Note that the constant parameter β can be
adjusted to get a higher Pr([�]c∈C [X][c][ =][ M] [)][ in order to reduce]
computational complexity. To achieve this, we can just choose
a β which maximizes its lower bound P [∗], so
∂ log P [∗]
= [M]
∂β β
[−]
C
�
c=1
νˆc = 0. (44)
1 + βνˆc
_D. Detailed Implementation in the Simulations_
_1) A Heuristic Repacking Algorithm: We first describe the_
concept of “repacking.” When the cache size M = 1, all the
bandwidth resources at a certain box belongs to the content
the box caches. When M 2, however, this is not the case:
≥
all the contents cached in one box are actually competitors for
the bandwidth resources at that box. Let’s consider a simple
example in which B = 2, M = 2 and U = 1: Box 1 which
caches content 1 and 2 is serving a download of content 2,
while box 2 which caches content 2 and 3 is idle. When a
request for content 1 comes, the only potential candidate to
serve it is box 1, but since the only connection is already
occupied by a download of content 2, the request for content
1 has to be rejected. However, if this ongoing download can be
“forwarded” to the idle box 2, the new request can be satisfied
without breaking the old one. We call this type of forwarding
“repacking.”
In the the feasibility condition (1) and its equivalent form
(2), we actually allow perfect repacking to identify a feasible
{nc}. In a real system, perfect repacking needs to enumerate
all the possible serving patterns and choose the best one based
on some criterion, which is usually computationally infeasible.
We then propose a heuristic repacking algorithm which is not
so complex but can achieve similar functionality and improve
performances, although imperfect.
Several variables need to be defined before we describe the
algorithm:
- nc: the system-wide ongoing downloads of content c,
which does not count the downloads from the server.
- Bc[k][: The set of boxes which have content][ c][ (“potential]
candidate boxes”) and k free connections, for 0 k U .
≤ ≤
- Dc: number of boxes which has content c. Dc =
�Uk=0 [|B]c[k][|][.]
- ub: a U -dimensional vector, of which the i-th component
represents the content box b is using its i-th connection
to upload (a value 0 represents a free connection).
- co: the “orphan content” which is affiliated with a new
request or an ongoing download but has not been assigned
with any box.
- Co: the set of contents which has once been chosen as
orphan contents.
- tR: the number of repacking already done.
Note that when choosing a box to serve a request, load balancing is already considered, which to some extent reduces the
chance of necessary repacking in later operations. However,
repacking is still needed for an incoming request for content
c as soon as ∪k>0Bc[k] [=][ ∅][.]
**Repacking Algorithm**
After getting a request for content c while ∪k>0Bc[k] [=][ ∅][, the]
server
1. Initialize co := c, Co := {c}, and tR := 0.
2. Let C[¯] = {c[′] : nc′/Dc′ > nco/Dco and c[′] ̸∈Co}, i.e.,
a set of contents which haven’t become orphans during
The server can use any numerical methods (e.g., Newton’s
method) to seek a root of equation (44). In fact, this lower
bound P [∗] on Pr([�]c∈C [X][c][ =][ M] [)][ is not tight, since it is just]
the largest item in the sum expression. When the popularity
is close to uniformness (e.g., in a zipf-like distribution, α
is small), this largest item is no longer dominant, so the
lower bound P [∗] is quite untight, which means we actually
overestimate the computation complexity by only evaluating
its upper bound. However, this will not affect the real gain we
obtain after choosing the optimal β according to equation (44).
Recall that we also proposed a simple sampling strategy in
Section IV-A. It is easy to see that when some contents are
much more popular than the others (e.g., zipf-like α is large),
the probability that duplicates appear in one size-M sample
is high, hence largely increases the number of resampling.
Thus, it would be faster if we choose the Bernoulli sampling.
However, when the popularity is quite uniform, the simple
sampling works very well. An extreme case is that under the
uniform popularity distribution,
M−1
�
i=1
�
1
− [i]
C
Pr a valid size-M subset =
{ }
�MC � - M !
=
C[M]
�
,
which shows that when C is large, you can get a valid sample
almost every time.
-----
this repacking process and of which the utilization factor
(may be larger than 1) is larger than that of the current
orphan content co. If C[¯]o = ∅, regard co as a loss and
TERMINATE.
3. Choose c[∗] = arg maxc′∈C¯{nc′/Dc′}. Uniformly pick
one (box, connection) pair from
{(b, i) : b ∈Bc[0][, c][∗] [is the][ i][-th component of][ u][b][}][.]
4. Use the chosen box b and its i-th connection to continue
uploading the remaining part of content co. At the
same time, c[∗] which was served using that connection
becomes a new orphan, i.e., co := c[∗]. Update ub and
{nc}. Set tR := tR + 1.
5. If ∪k>0Bc[k]o
[̸][=][ ∅][, i.e., there exists a free connection to]
serve the new co, then use the load-balancing-based box
selection rule to select a box to continue uploading the
remaining part of co. The repacking process is perfect
(no remaining orphan) and TERMINATE. Otherwise,
- If tR = t[max]R, a customized algorithm parameter
(0 ≤ t[max]R ≤ C), regard co as a loss and TERMINATE.
- Otherwise, set Co := Co + {co}, and go to Step 2.
_2) A Practical Issue in Cache Update: When a box b is_
chosen for cache update (and it does not hold the content
c corresponding to the request), it might still be uploading
content c[′] which is to be replaced. This fact is not captured by
the Markov chain model. In practice, those ongoing services
must be terminated. Since we have introduced the repacking
scheme, they become “orphans” ready for repacking. We
implement the procedure as follows:
1. Rank these orphans by their remaining service time in
the ascending order, i.e., the original download which is
sooner to be completed is given higher priority.
2. Do repacking one by one until one orphan fails to be
repacked. Note that here the repacking algorithm starts
from Step 5, since there may already be some boxes
with both content c and free connections.
_E. Proof of Theorem 2_
The Lagrangian of OPT 2 is
L( ˜m, λ, x; u, v, y, z, w, η, γ)
The KKT condition includes the feasible set defined in OPT
2 and the following:
∂L
∂xc = 1 − yc − zc = 0, ∀c;
∂L
∂m˜ c = ρc − uc + vc − ρczc − γ = 0, ∀c;
∂L
∂λc = −vc + yc − η + wc = 0, ∀c;
uc( ˜mc − 1) = 0, uc ≥ 0, ∀c;
vc(λc − m˜ c) = 0, vc ≥ 0, ∀c;
yc(xc − λc) = 0, yc ≥ 0, ∀c;
zc(xc − ρc + ρc ˜mc) = 0, zc ≥ 0, ∀c;
wcλc = 0, wc ≥ 0, ∀c.
We then put the solution stated in the theorem into KKT
condition to check whether the condition is satisfied. The
analysis is as follows:
- For 1 ≤ c ≤ M − 1, since ˜mc = 1 and λc = xc = 0, we
obtain that vc = 0, yc + zc = 1, ρc(1 − zc) = uc + γ,
and yc = η − wc. Letting wc = 0, we further have:
uc = ρcη−γ, yc = η, zc = 1−η. To keep uc, yc, zc ≥ 0,
we must have η ∈ [0, 1] and γ ≤ ρcη, for 1 ≤ c ≤ M −1.
Thus, since {ρc} are also ranked in the descending order,
we have
γ ≤ ρM −1η. (45)
- For M ≤ c ≤ c[∗], since ˜mc = λc = xc = ρc/(1 + ρc),
we obtain that uc = wc = 0, yc + zc = 1, ρc(1 − zc) =
γ − vc, yc = η + vc. We further have:
vc = [γ][ −] [ρ][c][η]
ρc + 1 [, y][c][ =][ η]ρc[ +] + 1[ γ] [, z][c][ = 1][ −] ρ[η]c[ +] + 1[ γ] [.]
To keep vc, yc, zc ≥ 0, we must have ρcη ≤ γ ≤ ρc +
1 η, for M c c[∗]. Thus,
− ≤ ≤
ρM η ≤ γ ≤ ρc∗ + 1 − η. (46)
- For c = c[∗] + 1, when mc = 0, it degenerates to the next
case. When ˜mc > 0, since ˜mc = λc = xc < ρc(1 − m˜ c),
we obtain that uc = wc = zc = 0, yc = 1, ρc + vc =
γ, η + vc = 1. We further have
γ = ρc∗+1 + 1 − η. (47)
- For c[∗] +2 ≤ c ≤ C, since ˜mc = λc = xc = 0, we obtain
that uc = zc = 0, yc = 1, vc = γ−ρc, wc = η+vc−1 =
η + γ − ρc − 1. To keep vc, wc ≥ 0, and due to the fact
that η ∈ [0, 1], we must have γ ≥ ρc, for c[∗] +2 ≤ c ≤ C.
Thus,
γ ≥ ρc[∗]+2. (48)
For inequalities (45), (46), (48) and equation (47) to hold
simultaneously, we can choose a η which satisfies
ρc∗+1 + 1
ρM −1 + 1 [≤] [η][ ≤] [ρ][c]ρ[∗]M[+1] + 1[ + 1] [,]
= �
c∈C
�
ρc ˜mc + xc − uc( ˜mc − 1) − vc(λc − m˜ c)
�
−yc(xc − λc) − zc(xc − ρc + ρc ˜mc) + wcλc
γ
−
.
η
−
�� λc − 1
c∈C
�
�� m˜ c − M
c∈C
�
which also satisfies η [0, 1]. Therefore, the theorem is
∈
proved.
-----
m˜ cIt should be mentioned that when∗+1 = 0, the case “c = c[∗] + 1” can be combined with[�]c[c]=[∗] M [m][˜] [c][ = 1][, i.e.,]
the next case “c[∗] + 2 c C”, hence equation (47) does not
≤ ≤
exist while inequality (48) is changed to γ ≥ ρc[∗]+1. Then, we
can just choose a η which satisfies
0 η
≤ ≤ [ρ][c][∗][+1][ + 1]
ρM + 1 [.]
_F. Storage of Segments and Parallel Substreaming_
then within one stream, some substreams may complete earlier
than the others. Therefore, the above equality needs to be
added as a constraint (and used to come up with the following
result), i.e., the bandwidth for the K substreams should be
reserved until the whole streaming is completed.
Then, in the proof of the optimality of “proportional-toproduct” placement for DSN, every expression keeps the same,
except that the feasibility constraint (10) is changed to
� x(cB) ≤ � mjBUK, (51)
c:θ∈c j:j∩S̸=∅
We have mentioned before that compared to the “storage
of complete contents and downloads by single streaming”
setting, a more widely used mechanism in practice is that
each box stores one specific segment of a video content
and a download (streaming) comprises parallel substreaming
from different boxes. To model this mechanism, we have the
following simplifying assumptions: Each content is divided
into K segments with equal length which are independently
stored. Each box can store up to M segments (actually it
does not matter if we keep the original storage space of each
box, i.e., M complete contents, which now can hold MK
segments, since the storage space is a customized parameter)
and these M segments do not necessarily belong to M distinct
contents. The bandwidth of each box is kept as U, so now
each box can accommodate UK parallel substreaming, each
with download rate 1/K (the average service duration is still
kept as 1 because each segment is 1/K of the original content
length). The definition of “traffic load” ρ is then the same as in
equation (6). A request for a content will be divided into subrequests submitted to the boxes holding those corresponding
segments of this content, generating K parallel substreaming
flows in total (one box can serve more than one substreaming
service for this request if it caches more than one distinct
segments of this content).
Let θ represent a segment and θ c indicate that θ is a
∈
segment of content c. Recall that we use nc to denote the
number of concurrent downloads (now called “streams”) of
content c in the network. We further use nθ to denote the
number of substreams corresponding to segment θ.
Now the original feasibility constraint (1) becomes
� zθb = nθ, ∀ θ ∈ Θ;
b: θ∈Jb
� zθb ≤ UK, ∀ b ∈B, (49)
θ: θ∈Jb
where Θ represents the whole set of segments and zθb denotes
the the number of concurrent substreams downloading segment
θ from box b. It is easy to see that the equivalent version which
can be proved by Hall’s theorem becomes:
∀S ⊆ Θ, � nθ ≤ KU |{b ∈B : S ∩Jb ̸= ∅}|, (50)
θ∈S
where with a little abuse of notation, is used to denote a
S
subset of Θ, instead of as before.
C
Since we have assumed that video duration and video
streaming rate are all the same, one naturally has nθ = nc for
all θ c. If we let randomness exist in the service duration,
∈
Θ, �
∀S ⊆
θ∈Θ
and the “proportional-to-product” placement {mj} is now
with respect to each segment, i.e., mj = [�]θ∈j [ν][ˆ][θ][/Z][ for]
all j Θ s.t. j = M, where Z is the normalizing
⊆ | |
constant and ˆνθ = ˆνc if θ ∈ c. With an observation that
�θ∈Θ [ν][ˆ][θ][ =][ K][ �]c∈C [ν][ˆ][c][ =][ K][, we can still come to an]
inequality same with inequality (17), except that c and are
C
replaced by θ and Θ respectively. All the succeeding steps are
exactly the same in the proof of optimality.
_G. Another Approach to Bound the Chance of “Good Con-_
_tents” in Proving Theorem 3_
At the first stage of proving Theorem 3, we mentioned that
we can also directly derive the Chernoff bound on the RHS of
inequality (26) to get the result. The derivation is given below:
Recall that I(x) = supθ{xθ − ln(E[e[θZ][i]])} is the Cram´er
transform of the Bernoulli random variable Zi. It is easy to
check that
I(x) =
� � x � � 1−x �
a ln p + (1 − x) ln 1−p if x ∈ [0, 1]
+ else
∞
Also recall that a ≜ �M [2][/][3] + [ν]ρU[i][M] � /MB = M [−][1][/][3]/B + p,
where p ≜ ρBUνi [. Since we are considering a large][ B][,][ a][ ∈] [[0][,][ 1]]
holds. Thus, denoting ¯p = 1 p for brevity, the exponent of
−
RHS of inequality (26) reads
MB I(a)
−
� 1 �
= (pMB + M [2][/][3]) ln 1 +
−
pM [1][/][3]B
� 1 �
(¯pMB M [2][/][3]) ln 1
− − - −
pM¯ [1][/][3]B
pMB
= +
− [pMB][ +][ M][ 2][/][3]
pM [1][/][3]B 2(pM [1][/][3]B)[2]
pMB¯
+ [pMB][¯] [ −] [M][ 2][/][3] +
pM¯ [1][/][3]B 2(¯pM [1][/][3]B)[2][ +][ o][(][M][ 1][/][3][)]
=
− [M][ 1][/][3]
2B
=
− [M][ 1][/][3]
2
� ρU 1
+
νi B(1 − ρU[ν][i] [)]
� 1
p [+ 1]p¯
�
+ o(M [1][/][3])
�
+ o(M [1][/][3])
= Θ �M [1][/][3][�] . (52)
−
With similar steps as above, we can show the exponent
exponent of the RHS of inequality (27) is also Θ �M [1][/][3][�].
−
Therefore, inequality (28) is proved.
-----
|
{
"disclaimer": "Notice: The following paper fields have been elided by the publisher: {'abstract', 'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1109/INFCOM.2011.5935250?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/INFCOM.2011.5935250, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "GREEN",
"url": "https://arxiv.org/pdf/1004.4709"
}
| 2,010
|
[
"JournalArticle"
] | true
| 2010-04-27T00:00:00
|
[] | 29,301
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Medicine",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Engineering",
"source": "s2-fos-model"
},
{
"category": "Environmental Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/028a57e04bdebfd6e0de0b6b95c1070f6a0b1fd8
|
[
"Computer Science",
"Medicine"
] | 0.882433
|
Neural Network Ensembles for Sensor-Based Human Activity Recognition Within Smart Environments
|
028a57e04bdebfd6e0de0b6b95c1070f6a0b1fd8
|
Italian National Conference on Sensors
|
[
{
"authorId": "79805155",
"name": "Naomi Irvine"
},
{
"authorId": "51071379",
"name": "C. Nugent"
},
{
"authorId": "2108431596",
"name": "Shuai Zhang"
},
{
"authorId": "2155654884",
"name": "Hui Wang"
},
{
"authorId": "38218996",
"name": "Wing W. Y. Ng"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"SENSORS",
"IEEE Sens",
"Ital National Conf Sens",
"IEEE Sensors",
"Sensors"
],
"alternate_urls": [
"http://nbn-resolving.de/urn/resolver.pl?urn=urn:nbn:ch:bel-142001",
"http://www.mdpi.com/journal/sensors",
"https://www.mdpi.com/journal/sensors"
],
"id": "3dbf084c-ef47-4b74-9919-047b40704538",
"issn": "1424-8220",
"name": "Italian National Conference on Sensors",
"type": "conference",
"url": "http://www.e-helvetica.nb.admin.ch/directAccess?callnumber=bel-142001"
}
|
In this paper, we focus on data-driven approaches to human activity recognition (HAR). Data-driven approaches rely on good quality data during training, however, a shortage of high quality, large-scale, and accurately annotated HAR datasets exists for recognizing activities of daily living (ADLs) within smart environments. The contributions of this paper involve improving the quality of an openly available HAR dataset for the purpose of data-driven HAR and proposing a new ensemble of neural networks as a data-driven HAR classifier. Specifically, we propose a homogeneous ensemble neural network approach for the purpose of recognizing activities of daily living within a smart home setting. Four base models were generated and integrated using a support function fusion method which involved computing an output decision score for each base classifier. The contribution of this work also involved exploring several approaches to resolving conflicts between the base models. Experimental results demonstrated that distributing data at a class level greatly reduces the number of conflicts that occur between the base models, leading to an increased performance prior to the application of conflict resolution techniques. Overall, the best HAR performance of 80.39% was achieved through distributing data at a class level in conjunction with a conflict resolution approach, which involved calculating the difference between the highest and second highest predictions per conflicting model and awarding the final decision to the model with the highest differential value.
|
# sensors
_Article_
## Neural Network Ensembles for Sensor-Based Human Activity Recognition Within Smart Environments
**Naomi Irvine** **[1,]*, Chris Nugent** **[1]** **, Shuai Zhang** **[1], Hui Wang** **[1]** **and Wing W. Y. NG** **[2]**
1 School of Computing, Ulster University, Co. Antrim, Northern Ireland BT37 0QB, UK;
cd.nugent@ulster.ac.uk (C.N.); s.zhang@ulster.ac.uk (S.Z.); h.wang@ulster.ac.uk (H.W.)
2 School of Computer Science and Engineering, South China University of Technology,
Guangzhou 510640, China; wingng@ieee.org
***** Correspondence: irvine-n2@ulster.ac.uk
[����������](http://www.mdpi.com/1424-8220/20/1/216?type=check_update&version=1)
Received: 13 November 2019; Accepted: 21 December 2019; Published: 30 December 2019 **�������**
**Abstract: In this paper, we focus on data-driven approaches to human activity recognition (HAR).**
Data-driven approaches rely on good quality data during training, however, a shortage of high quality,
large-scale, and accurately annotated HAR datasets exists for recognizing activities of daily living
(ADLs) within smart environments. The contributions of this paper involve improving the quality of
an openly available HAR dataset for the purpose of data-driven HAR and proposing a new ensemble
of neural networks as a data-driven HAR classifier. Specifically, we propose a homogeneous ensemble
neural network approach for the purpose of recognizing activities of daily living within a smart home
setting. Four base models were generated and integrated using a support function fusion method
which involved computing an output decision score for each base classifier. The contribution of this
work also involved exploring several approaches to resolving conflicts between the base models.
Experimental results demonstrated that distributing data at a class level greatly reduces the number
of conflicts that occur between the base models, leading to an increased performance prior to the
application of conflict resolution techniques. Overall, the best HAR performance of 80.39% was
achieved through distributing data at a class level in conjunction with a conflict resolution approach,
which involved calculating the difference between the highest and second highest predictions per
conflicting model and awarding the final decision to the model with the highest differential value.
**Keywords: human activity recognition; neural networks; ensemble neural networks; model conflict**
resolution; smart environments
**1. Introduction**
Human Activity Recognition (HAR) is a challenging and dynamic research field that has been
attracting significant interest in recent years [1], as human activities are intricate and highly diverse.
Particularly, sensor-based approaches to HAR have become prevalent in pervasive computing,
largely due to advancements with sensing technologies and wireless sensor networks. HAR is
a fundamental component in an extensive range of application areas, including connected health,
pervasive computing, surveillance systems, human computer interaction (HCI), and ambient assisted
living (AAL) in smart home settings. Other notable interest domains include human/object detection
and recognition based on object analysis and processing, for example, tracking and detection [2,3],
computer engineering [4], physical sciences [5], health-related issues [6], natural sciences, and industrial
academic areas [7]. Notably, the progression of AAL technologies is becoming vital, due to the
continuously increasing cost of healthcare provision, the aging population, and the need to support
“aging in place”. In this domain, several dedicated smart home projects have been aimed at AAL
for the elderly and disabled, for example CASAS [8], Gator Tech [9], MavHome [10], DOMUS [11],
-----
_Sensors 2020, 20, 216_ 2 of 26
and Aware Home [12]. These environments all employ a large number of sensors that capture activity
data via a range of sensor modalities. They possess the common aim of supporting smart home
inhabitants in carrying out activities of daily living (ADLs) and providing them with non-intrusive,
AAL environments to promote their independence and quality of life. ADL monitoring in smart
environments is an important aspect to consider for assessing the health status of inhabitants, therefore
the automatic detection of these activities is a significant motivation for conducting HAR research [13].
Various sensors are available for the purpose of image object capturing and processing, including
binary sensors, digital cameras, and depth data in image analysis fields [14,15].
Sensor-based approaches to HAR can be deemed generally within two categories: data-driven or
knowledge-driven. Data-driven approaches make use of datasets to learn activity models through
applying machine learning and data mining techniques [16], whereas knowledge-driven approaches
build activity models through exploiting rich prior knowledge in the domain of interest [17]. This work
focuses on data-driven approaches to HAR and addresses the current challenges of their application to
openly available datasets. Within the context of this work, the availability of openly available datasets
prompted focus on data-driven approaches, whilst an awareness of the difficulties in accessing domain
knowledge averted attention away from knowledge-driven approaches. Nevertheless, data quality
is a substantial consideration as data-driven approaches depend on good training data, however,
in the realms of HAR, a shortage of high quality, large-scale, and accurately annotated HAR datasets
exists for recognizing ADLs within smart environments [18]. This work avails of low-quality data
and emphasizes that good practice concerning data preparation can help improve HAR performance.
In relation to this, it has been observed that many machine learning algorithms rely on large amounts
of data during the training phase to achieve the desired generalization capabilities [18].
Ensemble learners have been explored widely, due to their ability to improve machine learning
performance [19], with the main motivation being the desire to improve generalization capabilities [20].
By combining a set of imperfect models, the acknowledged limitations of individual learners can
be more efficiently managed, in that the errors recognized in each component can be minimized as
an ensemble, through the implementation of effective combination approaches [20].
In this paper, contributions include improving the quality of an openly available HAR dataset
for the purpose of data-driven HAR, since it has been observed that data quality is a substantial
consideration for data-driven approaches to HAR, as well as proposing a new ensemble of neural
networks as a data-driven HAR classifier. Furthermore, various approaches to resolving conflicts
that occur between base models in ensemble classifiers are investigated, and the effects of various
data distributions that form the complement class per model are analyzed, as each model in the
ensemble contains unique classes. It has been observed that the various data distributions to generate
the complement class per model greatly impact the number of conflicts arising between the base
models, thus demonstrating that the effective generation of these classes is an important consideration.
The importance of adhering to good data preparation practices is also highlighted, as restructuring
and balancing the data has supported and notably improved HAR performance.
The remainder of the paper is structured as follows. Section 2 provides an overview of HAR and
describes ensemble approaches to activity classification. Following this, Section 3 describes the dataset
used in this study and issues identified with the data. Section 4 provides the methods and materials
implemented. Results are then presented and discussed in Section 5, followed by conclusions and
future work in Section 6.
**2. Related Works**
This Section presents relevant background information and related works. Section 2.1 provides
information relating to HAR within smart home settings, Section 2.2 describes neural networks with
regards to their recent use for HAR tasks and Section 2.3 describes ensemble learners with particular
consideration to ensemble generation and integration techniques.
-----
_Sensors 2020, 20, 216_ 3 of 26
_2.1. Human Activity Recognition (HAR)_
HAR is concerned with the ability to recognize and interpret human activities automatically
through the deployment of sensors and the processing of the data they generate [21]. Various approaches
to recognizing activities within smart environments have been explored, including the extensive use
of wearable devices [22,23] and video-based approaches [24], which is largely due to the increased
accessibility of these technologies. Nevertheless, these approaches have associated limitations to
consider, including concerns with ethics, comfort, privacy invasion, and obtrusiveness. For example, it
has been reported that many elderly inhabitants in AAL scenarios are often reluctant and unwilling
to continuously adopt the use of body-worn sensors, in addition to expressing reluctance to the
installation of video-based monitoring [25]. Consequently, in an attempt to address the identified
concerns and prevent user acceptance issues, binary sensors deployed in the surrounding environment
are becoming increasingly promising for long-term activity monitoring in the ubiquitous computing
domain, as these devices eliminate the privacy concerns identified with other approaches to HAR,
whilst also being non-invasive to smart home inhabitants [16].
Binary sensors have been used in a recent HAR study conducted by [26] to recognize nine ADLs,
such as cleaning, cooking and sleeping, performed by four smart home inhabitants. The sensors
deployed included motion detectors integrated within, or attached to, smart appliances. These also
incorporated ON/OFF states for cleaning appliances, e.g., a vacuum, ceiling lights, cooking heaters,
TV and PC, as well as OPEN/CLOSE states for kitchen appliances such as the fridge. The chosen classifier
was a Random Forest model which achieved 68% accuracy, however, the researchers suggested this
figure could be increased by applying more effective methods. In addition to this, in [27], binary sensors
were deployed within a home monitoring environment to recognize four basic activity classes, namely
relaxing, preparing a meal, eating, and transitioning from bed to toilet. A Deep Convolutional
Neural Network (DCNN) was proposed for activity classification, where the binary sensor data
generated by four door sensors and 31 passive infrared (PIR) motion sensors were converted into
representative activity images. The images generated were used to train the DCNN model which
obtained an accuracy of 99.36% in recognizing the four ADLs observed in the study. Although this
approach performed significantly well, a greater number of activity classes could have been explored.
Another study conducted by [28] explored the potential of ADL recognition using neural networks
within a smart home setting. Experiments involved the design and implementation of recurrent
(RNN) and convolutional (CNN) neural networks to recognize activities, e.g., cooking, bathing and
sleeping. Data acquired through the deployment of binary sensors consisting of pressure sensors,
reed switches, float sensors and PIR motion sensors, was used to train the various neural network
classifiers, with results showing that the RNN and CNN models significantly outperformed other
common classifiers during comparisons achieving 89.8% and 88.2% accuracies, respectively.
HAR requires a feature extraction stage where a set of features are chosen as inputs to a classification
model in order to represent the activities being detected. Various state-of-the-art features have been
determined for HAR, however, these vary depending on the sensors used to capture activity data.
For example, in the realms of wearable technologies that produce accelerometry data, extracting the
maximum, minimum, and range features are beneficial in differentiating between activities that
comprise movements of varying ranges [29]. Additionally, calculating the signal magnitude area
(SMA) of an accelerometry signal has proven advantageous in differentiating between static and
dynamic activities [30]. Alternatively, considering the vision-based HAR domain, visual objects can be
represented, for example, using local descriptors [31] or calculating the centroids from the contour of
depth silhouettes [32]. In this domain, features are commonly extracted with a template-based approach,
for example, through human silhouette representations, or a model-based approach, i.e., where the
body is defined by a skeleton-based outline with joint points used as feature representation [33].
-----
_Sensors 2020, 20, 216_ 4 of 26
_2.2. Neural Networks_
Neural Networks (NNs) are discriminative models that have been attracting attention recently
and are becoming a popular classifier for activity recognition tasks [34]. The Multilayer Perceptron
(MLP) is a notable type of feed-forward NN often used for activity recognition tasks [35–38]. They are
capable of modelling complex, non-linear relationships and provide an alternative approach to
pattern recognition, which is valuable for application in the HAR domain [35]. NNs require
high computational capacities which had restricted their use previously, however, due to recent
advancements in technology, more complex architectures are being explored with potential to offer
better performance and support [39]. In [37], various approaches to recognizing 11 common ADLs
were explored, including the use of a single hidden layer NN, a deep NN architecture, and a fuzzy
rule-based approach. The shallow NN performed best with an accuracy of 97.72%, followed by the
deep NN approach with 96.59%, with the researchers stating the potential of deep NNs had not
been shown during the study, whilst also stating that this could be due to insufficient amounts of
training data. IN addition to this, in a study conducted by [40], an efficiency investigation was carried
out which compared HAR performance using shallow to deep NN approaches. The shallow NN
outperformed the convolutional neural network (CNN) on the evaluated HAR datasets, with the
shallow NN achieving 99.2% on the WARD data and 96.7% on the UCI_DB data [41], in comparison to
97.7% and 94.2% with the CNN model, respectively. Conclusions of this study stated that the optimal
choice for HAR tasks is the use of shallow NNs with two or three layers, rather than the implementation
of more complex architectures, particularly if the dataset contains a small number of training samples.
_2.3. Ensemble Learners_
A technique often used to improve classification performance is to combine multiple models
together, i.e., to create an ensemble method, rather than relying on the performance of a single
model [29]. Ensemble learning involves two key considerations: ensemble generation and ensemble
integration [42]. The generation phase includes generating the base models and determining the
size of the ensemble. If the models created are achieved using a consistent induction algorithm,
it is known as a homogeneous approach, whereas a heterogeneous approach involves creating
models using various different algorithms [43]. In [44], a heterogeneous ensemble approach was
implemented to recognize various activities within the CASAS smart home testbeds. The ensemble
included four base classifiers, which included a Hidden Markov Model (HMM), a NN, a Support
Vector Machine (SVM), and Conditional Random Fields (CRF). The results were promising and
revealed performance improvements over the use of a single classification model. Further to
this, [45] implemented an ensemble classification approach to activity recognition using several
heterogeneous base classifiers. The five common base classifiers included an SVM, Decision Tree (DT),
kNN, NN, and Naïve Bayes. Results demonstrated that the ensemble approach, combined through
majority voting, performed extremely well in classifying twelve activities. As for homogeneous
approaches, [46] proposed an ensemble of random forest learners with the aim of generating
a more accurate, stable classifier to recognize activities from the PAMAP physical activity dataset.
Activity recognition performance was very high, and the generalization capability of the produced
classifier had improved significantly. In [47], multiple HMM base models were combined using
a decision templates method to recognize activities collected by a smartphone-embedded triaxial
accelerometer. Their approach addressed the interclass similarity and intraclass variability HAR
challenges, with results showing the ensemble generated performed significantly well with data
representing six activity classes and collected by 30 participants. In addition to this, [48–51] proposed
homogeneous ensemble approaches for HAR. An observation has been made that less research effort
exists on heterogeneous ensembles due to more difficulties arising in controlling interactions between
the various learning processes [43]. More recently, researchers have been exploring ensemble learners
on the basis of deep learning approaches. For example, [52–54] proposed ensemble deep learning
techniques for HAR, which revealed positive results and robustness. Nevertheless, NNs, and more
-----
_Sensors 2020, 20, 216_ 5 of 26
specifically, deep learning techniques, require a large number of training samples to enhance their
performance [55].
2.3.1. Ensemble Generation
During ensemble generation, data partitioning is a commonly considered approach aimed
towards diversifying the input data of the base models, so that the subspaces of inputs become
complementary [56]. Boosting and Bagging are two common data partitioning ensemble methods
used to combine multiple classification models that have been trained on different subsets of the
training data [29]. Boosting involves the combination of multiple base classifiers to generate a strong
committee classifier that may provide significantly enhanced performance in comparison to the base
classifiers, achieved through reweighting the misclassified data samples and therefore boosting their
performance [29]. SMOTEBoost and RUSBoost are adaptations of the known AdaBoost approach,
where random undersampling or SMOTE is applied to the base classifiers training data, along with the
reweighting phase in accordance with the AdaBoost algorithm, as demonstrated in a study conducted
by [57]. Both SMOTEBoost and RUSBoost inject a great degree of arbitrariness through generating
or removing instances, resulting in improved robustness to noise [57]. Bagging, on the other hand,
averages the outputs produced by each base model, where each model is trained on different training
sets consisting of data generated through sampling with replacement [42]. Examples of well-known
bagging-based approaches include OverBagging, UnderBagging, and SMOTEBagging. Particularly,
SMOTEBagging has been recommended for handling multi-class imbalanced data problems where the
instances within each bag are significantly diverse [58]. In a recent study [59], two bagging-based hybrid
methods were proposed to deal with imbalanced datasets, namely, ADASYNBagging and RSYNBagging.
The ADASYNBagging approach uses the bagging algorithm in conjunction with the ADASYN-based
oversampling method, whereas the RSYNBagging approach uses the ADASYN-based oversampling
method as well as random undersampling alongside the bagging algorithm. The performances of the
proposed hybrid approaches were compared against UnderBagging and SMOTEBagging techniques
and evaluated on twelve datasets, with promising experimental results obtained. The benefits of the
proposed hybrid approaches were demonstrated, as they outperformed the benchmark methods on
eight of the twelve datasets evaluated. Another approach considered during ensemble generation is to
manipulate the inputs of the base classifiers at a feature level, for example, training the base models on
various different subsets of features [56].
2.3.2. Ensemble Integration
The ensemble integration phase determines how the predictions produced by the base
models should be integrated together to increase performance by obtaining a single outcome [42].
Multiple fusion strategies exist and can be considered at a class label level, a trainable level, or a support
level, according to [55]. The class label fusion technique involves each of the base classifiers voting for
a certain class, then the final output is decided upon through either a majority voting or weighted
majority voting strategy. Majority voting decides on the final output prediction based on the class
that has been chosen most often or unanimously by the base classifiers, whereas weighted majority
voting assigns weights to each model, often based on their performances, where the classifier with the
highest output after weight assignments wins the overall prediction [55]. In a study conducted by [60],
majority voting was implemented to decide upon the final outputs of an ensemble approach based on
AdaBoost. Three weak learners were used, namely a Decision Tree, Logistic Regression, and Linear
Discriminant Analysis (LDA). In addition to AdaBoost, Bagging and Stacking methods were also
explored, with the best performance produced by the Bagging approach. Another study [46] used
weighted majority voting with an ensemble of Random Forest classifiers. Each classifier was assigned
different weights per activity, with the final outcome attained through combining the classification
outcomes from each base model via the weighted votes.
-----
_Sensors 2020, 20, 216_ 6 of 26
Fusion techniques at a trainable level consider the chosen fusion weights during the learning
process and implement optimization strategies to increase classification performance whilst also
reducing computation cost [55]. These include weighted summations of hypotheses, where higher
weights are assigned to those with lower error rates and the Dempster–Shafer theory to handle
uncertainty in the decision-making process. In [61] the outputs of various SVM classifiers, trained on
different input feature subsets, were subsequently combined using the Dempster-Shafer fusion rule.
The four-step process included creating decision templates for all training instances, calculating the
proximity between decision templates and classifier outputs, computing the belief degrees for each
output class, and finally, applying the Dempster rule to combine the degrees of belief derived from
each base classifier.
Finally, support function fusion involves computing an output decision score for each base
classifier, which is derived from the estimated likelihood of a class [56]. This estimation can be
computed as an a posteriori probability attained through probabilistic models, using fuzzy membership
functions, or through combining NN outputs according to their performance. In [62], five classifiers
were combined using an average of probabilities fusion method to recognize six activities. This method
used the average of the probability distributions for each base classifier to make a final decision,
achieving the best HAR performance in comparison to a majority voting approach that was also
implemented during this study. Another study [63], implemented a support function fusion where
a Naïve Bayesian fusion method was compared to a majority voting approach to fuse several HMM
base models. The Naïve Bayesian approach involves calculating the posterior probability of the HMM
outputs, which achieved the best activity recognition performance during the study.
After reviewing the literature, we focus on ensemble learning for HAR in this work, due to their
perceived benefits, rather than relying on the performance of a single model. Particularly, an ensemble
of NNs are explored, although, due to a lack of high quality data in ADL datasets and a low quantity
of data, it was decided to employ lighter weight models rather than exploring deeper architectures.
The literature has shown that shallow NNs have previously achieved similar performance to deep NN
architectures for HAR tasks, with provided recommendations to use shallow architectures particularly
in cases where a small number of training samples are available. As outlined, there has been less
effort made with heterogeneous ensembles in the research community due to difficulties existing
in controlling interactions between the various learning processes, consequently, this work focuses
on a homogeneous approach to generating an ensemble of NNs. As stated in [64], one of the crucial
problems to consider with ensemble learning is the combination rule employed to determine a final
class decision amongst the base models. In this work, a support function fusion method is used to
integrate base models, and various approaches to effectively resolve conflicts that occur between the
base models are investigated to determine a final output decision.
In summary, this section contained detailed relevant background information and related
works. The recent potential of NNs for HAR and pattern recognition problems was presented,
which demonstrated that shallow architectures are preferred in scenarios where the dataset contains
a small number of training samples. Ensemble approaches to activity recognition were also discussed
because of their recent performance in the HAR domain. As mentioned, this work focuses on
data-driven approaches to HAR, thus the importance of data quality in relation to those is considered
in Section 3, along with a description of the dataset used to conduct experiments.
**3. Dataset for Data-Driven HAR**
An overview of the HAR data is presented in this section with an emphasis on the quality of data
acquired. The UCAmI Cup challenge is also described, as the dataset used in this study was derived
from this competition. Section 3.1 outlines details of the original dataset, Section 3.1.1 highlights
the problems identified, and Section 3.2 details the restructured dataset created as a result of the
encountered problems and to demonstrate more realistic capabilities of binary datasets for HAR
in smart environments.
-----
_Sensors 2020, 20, 216_ 7 of 26
Data collection is becoming a critical concern among the countless challenges in machine learning,
largely due to limited amounts of training data being available to researchers in their respective
fields and the quality of the data being collected [65]. In the realms of machine learning, it is known
that the majority of effort and time is consumed through preparing the data, which involves data
collection, cleansing, interpretation, and feature engineering [65]. Data quality is an imperative
consideration in applications of data-driven approaches to HAR, as the performance of models are
largely dependent on the quality of training data. Noise can be introduced during data collection by
the participants and/or sensors which adversely affects the performance of data-driven techniques [66].
Common issues include missing or erroneous values and mislabeled data [67]. Data cleansing is known
as the process of removing inconsistencies or errors, such as outliers and/or noise from a collection of
data [68]. According to [66], addressing the presence of outliers and noise is vital as their existence
can substantially influence experimental results produced by data-driven approaches. Nevertheless,
an unclear border is often present between normal and abnormal data, where a considerably large
“gray area” may exist [69]. In supervised learning, noise can transpire at an attribute or class level.
In [70], an effort was made to evaluate the impact of noise on classification performance on 17
datasets, generated within various domains. Each dataset was manually introduced to various levels
of noise to investigate how it affected model performance. Their findings demonstrated that as noise
levels increase, performance decreases, and that attribute noise is generally less harmful than class
noise. Furthermore, [71] compared and evaluated how well several classifiers performed with noisy,
poor quality data. Conclusions stated that robustness to noisy data and classification performance
varied significantly amongst the algorithms observed, with the Random Forest and kNN models
proving most resilient to noise.
The data used in this study was generated for the 1st UCAmI Cup challenge, where participants
were invited to use their tools and techniques to analyze a HAR dataset with the aim of achieving the
highest accuracy on the unseen test set. In [72], the challenge organizers describe the UCAmI Cup
dataset comprehensively. Knowledge-driven rule-based approaches outperformed the data-driven
approaches to the activity recognition problem, with many of the participants reporting issues and
limitations found within the data [73–76]. The approach implemented by [73] involved a domain
knowledge-based solution inspired by a Finite State Machine, achieving 81.3% accuracy. In [74], a hybrid
model was proposed using a hidden Markov chain and logic model. The researchers combined their
knowledge-driven and probabilistic models using a weighted averaging method, however, they
reported that they had expected a better result than 45.0% accuracy on the test set. In addition
to this, [75] used a Naïve Bayes approach with emphasis on location-aware, event-driven activity
recognition. The applied method interpreted events as soon as they became available in real-time,
omitting the need of an explicit segmentation phase, and generated activity estimations using an activity
prediction model. Reported results show mean accuracies of around 68%, with the researchers stating
that given the high number of activity classes, the outcome achieved was reasonable. Another approach
implemented in [76] used various common machine learning algorithms, including a Decision Tree,
Nearest Neighbour, Support Vector Machine, and three ensemble approaches including Random
Forest, Boosting, and Bagging. The researchers reported a training set accuracy of 92.1%, however,
their approach achieved 60.1% on the provided test data which demonstrated poor generalization.
Their suggested cause for the low outcome was the high imbalance of classes in the training set,
and they stated that the training algorithm required more labelled training data to perform better.
_3.1. UCAmI Cup Dataset_
The HAR dataset was collected over 10 days by researchers in the UJAmI Smart Lab [72].
The UJAmI Smart Lab is divided into five regions: an entrance, a workplace, a living room, a bedroom
with an integrated bathroom, and a kitchen, which measures approximately 25 square meters combined,
as presented in Figure 1. The dataset was captured by a single male inhabitant completing morning,
afternoon, and evening routines, representing 246 occurrences of 24 activity classes, as presented
-----
_Sensors 2020, 20, 216_ 8 of 26
in Table 1. The training set consisted of 7 days of labelled data, with the remaining 3 days of data being
_Sensors 2019, 19, x FOR PEER REVIEW_ 8 of 26
provided as an unlabeled test set.
**Figure 1. Location of Binary Sensors in the UJAmI Smart Lab [72].**
**Figure 1. Location of Binary Sensors in the UJAmI Smart Lab [72].**
**Table 1. Activity Classes in the UCAmI Cup Dataset [72], where M, A, and E indicate the Morning,**
**Table 1. Activity Classes in the UCAmI Cup Dataset [72], where M, A, and E indicate the Morning,**
Afternoon, and Evening routines, respectively.
Afternoon, and Evening routines, respectively.
**ID** **Name** **Instances** **Routine** **ID** **Name** **Instances** **Routine**
**ID** **Name** **Instances** **Routine** **ID** **Name** **Instances** **Routine**
Take Leave
Act01 MedicationTake 52 A, E Act13 Leave Smart Smart Lab 33 M, A
Act01 52 A, E Act13 33 M, A
Medication Prepare Visitor toLab
Act02 63 M Act14 7 M, A
BreakfastPrepare Visitor to Smart Lab
Act02 Prepare 63 M Act14 Put waste 7 M, A
Act03 Breakfast 118 A Act15 Smart Lab 75 A, E
lunch in the bin
Prepare Put waste in
Act04Act03 Prepare 76118 EA Act15 Act16 Wash 2275 A, E M
Dinnerlunch the bin hands
Prepare Brush
Act05Act04 Breakfast 7876 ME Act16 Act17 Wash hands teeth 13222 M, A, EM
Dinner
Use the
Act06Act05 Breakfast Lunch 10178 AM Act17 Act18 Brush teeth 44132 M, A, EM, A, E
toilet
Act06 Lunch 101 A Act18 Use the toilet Wash 44 M, A, E
Act07 Dinner 86 E Act19 13 A, E
Act07 Dinner 86 E Act19 Wash dishes dishes 13 A, E
Put washing Put
Act08Act08 Eat a snack Eat a snack 1212 AA Act20 Act20 washing 2020 M, AM, A
in machine
in machine
Work at the Work at
Act09Act09 Watch TV Watch TV 7070 A, EA, E Act21 Act21 2020 MM
the tabletable
Act10 Enter Smart Enter 21 A, E Act22 Dressing 86 M, A, E
Act10 Smart Lab 21 A, E Act22 Dressing 86 M, A, E
Lab
Play
Act11 Play a 28 M, E Act23 Go to bed 30 E
Act11 a videogame 28 M, E Act23 Go to bed 30 E
videogame Relax on
Act12 85 M, A, E Act24 Wake up 32 M
Relax on the sofa
Act12 85 M, A, E Act24 Wake up 32 M
the sofa
A set of 30 binary sensors consisting of magnetic contact switches, PIR motion detectors,
A set of 30 binary sensors consisting of magnetic contact switches, PIR motion detectors, and
and pressure sensors were deployed in the UJAmI Smart Lab to capture human interactions within the
pressure sensors were deployed in the UJAmI Smart Lab to capture human interactions within the
environment, as presented in Figure 1. The two changeable states of the magnetic contact switches were
environment, as presented in Figure 1. The two changeable states of the magnetic contact switches
open/close, which were attached to, or integrated within, doors and objects, such as the medication
were open/close, which were attached to, or integrated within, doors and objects, such as the
box. The motion detectors generated and recorded movement/no movement states to identify whether
medication box. The motion detectors generated and recorded movement/no movement states to
identify whether an inhabitant had moved in or out of the 7-meter sensing range. Finally, the pressure
e o de loyed e e ated eithe e u e/ o e u e tate a d a e e t i the bed a d the ofa to
-----
_Sensors 2020, 20, 216_ 9 of 26
an inhabitant had moved in or out of the 7-meter sensing range. Finally, the pressure sensors deployed
generated either pressure/no pressure states and was present in the bed and the sofa to detect any
interactions. A comprehensive description of each binary sensor is presented in Table 2.
**Table 2. Description of binary sensors [72].**
**ID** **Object** **Type** **States**
Movement/No
SM1 Kitchen area Motion
movement
Movement/No
SM3 Bathroom area Motion
movement
Movement/No
SM4 Bedroom area Motion
movement
Movement/No
SM5 Sofa area Motion
movement
M01 Door Contact Open/Close
TV0 TV Contact Open/Close
D01 Refrigerator Contact Open/Close
D02 Microwave Contact Open/Close
D03 Wardrobe Contact Open/Close
D04 Cups cupboard Contact Open/Close
D05 Dishwasher Contact Open/Close
D07 WC Contact Open/Close
D08 Closet Contact Open/Close
D09 Washing machine Contact Open/Close
D10 Pantry Contact Open/Close
C01 Medication box Contact Open/Close
C02 Fruit platter Contact Open/Close
C03 Cutlery Contact Open/Close
C04 Pots Contact Open/Close
C05 Water bottle Contact Open/Close
C07 XBOX Remote Contact Present/Not present
C08 Trash Contact Open/Close
C09 Tap Contact Open/Close
C10 Tank Contact Open/Close
C12 Laundry basket Contact Present/Not present
C13 Pyjamas drawer Contact Open/Close
C14 Bed Pressure Pressure/No pressure
C15 Kitchen faucet Contact Open/Close
H01 Kettle Contact Open/Close
S09 Sofa Pressure Pressure/No pressure
3.1.1. Data Challenges
A number of issues were identified with the original binary dataset that hindered the performance
of recognizing ADLs in a smart environment setting. These included:
- Number of classes. The number of classes in the original dataset were very high given the
low number of instances per activity and low amount of data overall. As discussed previously,
data-driven approaches rely on large amounts of good quality data. Furthermore, certain classes
were too closely related to one another to recognize with binary data alone. For example,
the following activities relied on one door sensor: entering the smart lab, leaving the smart lab,
and having a visitor to the smart lab. Binary sensors are limited in inferring activities in that
they provide information at an abstract level [77], therefore Act08 eating a snack was difficult
to distinguish compared to Act03 prepare breakfast, Act04 prepare lunch, and Act05 prepare
dinner, as these activities all used similar sensors. Thus in order to capture activities at a finer
level, the presentation and interpretation of binary data often requires further knowledge of the
environment [78]. This issue was discussed by a UCAmI Cup participant in [75], where conclusions
-----
_Sensors 2020, 20, 216_ 10 of 26
had stated that their achieved activity recognition performance was reasonable given the large
number of activity classes present in the dataset.
- Imbalanced dataset. The distribution of instances per class in the original dataset were highly
diverse, which may have caused minority classes to be overlooked by the classification model.
For example, Act19 wash dishes was represented by 13 instances of data, whereas other activities
such as Act17 brush teeth had more than 100 instances. Furthermore, the distribution of instances
per class in the provided training and test sets were highly varied. For example, Act09 was
very under-represented in the training set, yet the test set included a large number of Act09
instances. Noteworthy, Act09 also produced very similar sensor characteristics to Act12, which was
problematic in the initial experiments, as the training set included large amounts of Act12 data.
This issue was discussed in [74], where researchers stated that their approach also found difficulty
in classifying Act12, due to the poor representation of this activity in the training set, and suggested
that the data should be better distributed to improve HAR performance.
- Quantity of data. As previously stated, data-driven approaches require lots of data during the
training phase to learn activity models and to ensure these models can generalize well to new
data. NN require lots of data to learn complex activity models [79], though the original dataset
was relatively small. Thus, more labelled training data could have improved initial experiments.
In [76], UCAmI Cup participants suggested the cause for their low HAR performance was the
high imbalance of classes in the training set and stated that the training algorithm required more
labelled training data to perform better.
- Missing sensors. Act21 work at table had no binary sensor located near the table to distinguish this
activity, as presented in Figure 1. This issue caused confusion as the sensor firing for Act21 in the
labelled training set was seen to be a motion sensor located in the bedroom, which is irrelevant to
Act21 and therefore seen as erroneous. In addition to missing sensors, there were also missing
values from sensors that were expected to fire during certain activities. As previously stated,
some researchers participating in the UCAmI Cup challenge reported that they found missing
values or mislabeling of some activities within the training set. In [73] this issue was discussed,
where participants stated that during one instance of Act10 enter the smart lab, the only binary
sensor that is expected to fire (M01), does not change states.
- Interclass similarity. This is a common HAR challenge that occurs when certain activities generate
similar sensor characteristics, though they are physically different [80]. Table 3 shows the activities
that produced similar sensor characteristics, resulting in difficulties arising in discriminating
between these activities during classification.
**Table 3. Activities producing similar sensor characteristics within the UCAmI Cup data.**
**Activity Group** **Activity Name** **Common Sensors**
Enter Smart Lab, Leave Smart Lab,
Act10, Act13, Act14 M01 Door
and Visitor to Smart Lab
Act23, Act24 Go to Bed and Wake Up C14 Bed
S09 Pressure Sofa
Act09, Act12 Watch TV and Relax on Sofa
SM5 Sofa Motion
Prepare Breakfast, Prepare Lunch,
Act02, Act03, Act04, Act08
Prepare Dinner, Prepare Snack
SM1 Kitchen Motion
D10 Pantry
C03 Cutlery
As a result of the various problems identified with the dataset, it was decided to restructure the
data to reveal the potential of using binary sensors alone within smart environments.
-----
_Sensors 2020, 20, 216_ 11 of 26
_Sensors 2019, 19, x FOR PEER REVIEW_ 11 of 26
_3.2. Restructured Dataset_
First, the provided training and test sets were combined to better represent activity classes
within the training data. Figure 2 shows the distribution of the combined 10 days of 24 activity classes First, the provided training and test sets were combined to better represent activity classes within
the training data. Figurefor all the available data in the UCAmI Cup. As can be viewed in Figure 2, certain classes were very 2 shows the distribution of the combined 10 days of 24 activity classes for
all the available data in the UCAmI Cup. As can be viewed in Figureunder-represented, with a third of all activity classes containing less than 30 instances. These classes 2, certain classes were very
under-represented, with a third of all activity classes containing less than 30 instances. These classeswere removed, as they would be under-represented in the training phase and therefore would not
were removed, as they would be under-represented in the training phase and therefore would notgeneralize well to unseen data. Consequently, 8.82% of instances were removed, which comprised
generalize well to unseen data. Consequently, 8.82% of instances were removed, which comprised thethe following classes: Act08, Act11, Act16, and Act19-Act21. An opportunity to combine certain
following classes: Act08, Act11, Act16, and Act19-Act21. An opportunity to combine certain similarsimilar activity classes was also identified so that the data could be used effectively. For example
activity classes was also identified so that the data could be used eAct10, Act13, and Act14 were combined to produce ActN1 door, as they all make use of a single door ffectively. For example Act10,
Act13, and Act14 were combined to produce ActN1 door, as they all make use of a single door sensor,sensor, and Act09 and Act12 were combined to produce ActN2 watch TV on sofa, as they mainly
and Act09 and Act12 were combined to produce ActN2 watch TV on sofa, as they mainly consisted ofconsisted of the inhabitant sitting on the sofa. Furthermore, Act02 and Act05, Act03 and Act06, and
the inhabitant sitting on the sofa. Furthermore, Act02 and Act05, Act03 and Act06, and finally Act04finally Act04 and Act07 were combined to produce ActN3 breakfast, ActN4 lunch, and ActN5 dinner,
and Act07 were combined to produce ActN3 breakfast, ActN4 lunch, and ActN5 dinner, respectively,respectively, as these sets of activities were similar. Table 4 presents the restructured dataset.
as these sets of activities were similar. Table 4 presents the restructured dataset.
**Figure 2. Distribution of the 24 UCAmI Cup activity classes with threshold shown.**
**Figure 2. Distribution of the 24 UCAmI Cup activity classes with threshold shown.**
**Table 4. Activity classes in the restructured dataset.**
**Table 4. Activity classes in the restructured dataset.**
**ID** **Name** **Instances** **Routine** **ID** **Name** **Instances** **Routine**
**ID** **Name** **Instances** **Routine** **ID** **Name** **Instances** **Routine**
Take
Act01 MedicationTake 52 A, E Act24 Wake up 32 M
Act01 52 A, E Act24 Wake up 32 M
Medication Put waste
Act15 75 A, E ActN1 Door 61 M, A, E
Put waste in the bin
Act15 Brush 75 A, E ActN1 Watch TVDoor 61 M, A, E
Act17 in the bin 132 M, A, E ActN2 155 M, A, E
teeth on sofa
Watch TV on
Act17 Act18 Brush teeth Use the 132 44 M, A, E M, A, E ActN2 ActN3 Breakfast 141155 M, A, E M
toilet sofa
Act22 Use the Dressing 86 M, A, E ActN4 Lunch 219 A
Act18 Act23 Go to bed 44 30 M, A, E E ActN3 ActN5 Breakfast Dinner 162141 EM
toilet
Act22 Dressing 86 M, A, E ActN4 Lunch 219 A
**4. Proposed HAR Classification ModelAct23** Go to bed 30 E ActN5 Dinner 162 E
The materials and methods implemented are described within this section. The data pre-processing
**4. Proposed HAR Classification Model**
phase is explained, including data segmentation and feature extraction, which are two fundamental
aspects of the activity recognition process [The materials and methods implemented are described within this section. The data pre-80]. Following this, the ensemble approach and conflict
resolution techniques are presented.processing phase is explained, including data segmentation and feature extraction, which are two
fundamental aspects of the activity recognition process [80]. Following this, the ensemble approach
and conflict resolution techniques are presented
-----
_Sensors 2020, 20, 216_ 12 of 26
_Sensors 2019, 19, x FOR PEER REVIEW_ 12 of 26
_4.1. Data Pre-Processing 4.1. Data Pre-Processing_
Since the data restructuring process involved combining the provided train and test sets to Since the data restructuring process involved combining the provided train and test sets to produce
produce a set of data that better represents activity classes in the training data, it was subsequently a set of data that better represents activity classes in the training data, it was subsequently required to
required to extract a new test set. Thus, 15% of the data was randomly selected and removed to extract a new test set. Thus, 15% of the data was randomly selected and removed to generate an unseen
generate an unseen test set. The raw data files containing data streams produced by binary sensors test set. The raw data files containing data streams produced by binary sensors include a timestamp,
include a timestamp, the sensor ID, the sensor state, and the inhabitant name, as presented in Figure 3. the sensor ID, the sensor state, and the inhabitant name, as presented in Figure 3.
**Figure 3. Excerpt from a raw binary data file.**
**Figure 3. Excerpt from a raw binary data file.**
The raw data was segmented into 30-second non-overlapping time windows to identify the The raw data was segmented into 30-second non-overlapping time windows to identify the
segments of data that are likely to contain information regarding activities. Time-based windowing segments of data that are likely to contain information regarding activities. Time-based windowing
involves dividing the entire dataset equally into time segments that include a fixed quantity of data involves dividing the entire dataset equally into time segments that include a fixed quantity of
per window [29]. It is a common approach for segmenting data streams collected through data per window [29]. It is a common approach for segmenting data streams collected through
environmental sensors, however, no clear consensus exists for choosing the optimal window size for environmental sensors, however, no clear consensus exists for choosing the optimal window size
ADL recognition [81], therefore a 30 second window size was chosen, as this was the regulation for ADL recognition [81], therefore a 30 second window size was chosen, as this was the regulation
adhered to in the UCAmI Cup challenge. A total of 31 features were included, which consisted of one adhered to in the UCAmI Cup challenge. A total of 31 features were included, which consisted of one
feature per binary sensor and an additional time routine feature representing whether the activity feature per binary sensor and an additional time routine feature representing whether the activity
had occurred in the morning, afternoon, or evening, to help distinguish between the similar activities had occurred in the morning, afternoon, or evening, to help distinguish between the similar activities
previously outlined. For example, as Act23 go to bed and Act24 wake up use the same pressure sensor previously outlined. For example, as Act23 go to bed and Act24 wake up use the same pressure sensor
located in the bed, the inclusion of a time routine feature can help distinguish these activities due to located in the bed, the inclusion of a time routine feature can help distinguish these activities due to
the human nature of habitually waking up in the morning and going to bed in the evening. the human nature of habitually waking up in the morning and going to bed in the evening.
_4.2. Ensemble Approach_
_4.2. Ensemble Approach_
Ensemble methods for classification have been explored recently, due to their potential to improve
Ensemble methods for classification have been explored recently, due to their potential to
robustness, performance and generalization capabilities in comparison to single model approaches [40].
improve robustness, performance and generalization capabilities in comparison to single model
Our approach consists of four MLPs as base classifiers to generate a homogeneous ensemble method.
approaches [40]. Our approach consists of four MLPs as base classifiers to generate a homogeneous
A model is created per time routine: Morning, Afternoon, and Evening as some activities uniquely
ensemble method. A model is created per time routine: Morning, Afternoon, and Evening as some
occur within specific routines. Additionally, a Mixed model is created to consider activities that occur
activities uniquely occur within specific routines. Additionally, a Mixed model is created to consider
arbitrarily throughout the day. Figure 4 presents the four base classifiers where n indicates the number
activities that occur arbitrarily throughout the day. Figure 4 presents the four base classifiers where
of classes per model. M, A, and E represent the Morning, Afternoon, and Evening models, respectively,
n indicates the number of classes per model. M, A, and E represent the Morning, Afternoon, and
and finally MI represents the Mixed model.
Evening models, respectively, and finally MI represents the Mixed model.
-----
_Sensors 2020, 20, 216_ 13 of 26
_Sensors 2019, 19, x FOR PEER REVIEW_ 13 of 26
**Figure 4.Figure 4. Four base classifiers presented per time routine, where n indicates the number of classes per Four base classifiers presented per time routine, where n indicates the number of classes per**
model. M, A, and E represent the Morning, Afternoon, and Evening models, respectively, and finally model. M, A, and E represent the Morning, Afternoon, and Evening models, respectively, and finally
MI represents the Mixed model. MI represents the Mixed model.
Definitions: Definitions:
Input:
Input: _R_
_X =_ �→x 1, →x 2, . . ., →x M� ∈ _BN×d,_
𝑋= [𝑥⃗�, 𝑥⃗�, …, 𝑥⃗�][�] ∈ 𝐵[�×�],
where N is the number of instances, d is the number of features, d=31.
where N is the number of instances, d is the number of features, d=31.
Output:
Output:
→ � �
𝑥⃗� _x= [𝑥 i =_ ��, 𝑥x[1]1��[,][ x], …, 𝑥2[2][,][ . . .]��[,]] where [ x][d]i where𝑥�� _x∈[0,1][d]i_ [∈] [[][0, 1]. []][.]
� �R
Y = y1, y2, . . ., yN ∈ [1, . . ., 12].
Y = [y�, y�, …, y�][�] ∈[1, …, 12].
Base Models:
Base Models:
Models M1, M2, M3, and M4 represent the Morning, Afternoon, Evening, and Mixed base models,
Models M1, M2, M3, and M4 represent the Morning, Afternoon, Evening, and Mixed base models,
respectively, in the proposed ensemble approach.
respectively, in the proposed ensemble approach.
Given the instance Given the instance𝑥⃗� base model output →x i base model output𝑀� is given by M _j is given by_
𝑓�� = 𝑓fi[j] [=]�[ f](𝜑[ j][�]�ϕ(𝑥[j]�())xi, )�,
where index where indexj = [1, …, 4]; j = [1, . . ., 4];𝜑[�] ϕ(𝑥[j]�()x is the input to the activation function of base model i) is the input to the activation function of base model𝑀 M� and j and f𝑓[j][�] is is the
the output of each base model 𝑀�
output of each base model M _j_
� � � � �
For simplicity, the output can be represented as For simplicity, the output can be represented as𝑓� = �𝑝�, …, 𝑝 fi[j] [=]���,p where 1[j] [,][ . . .][,][ p]m[j]𝑚j� represents the number, where mj represents the
number of outputs from base model M _j._
of outputs from base model Predicted class _k[ˆ]_ _[j]_ 𝑀�.
_i_
Predicted class maximum p values𝑘[�]�� ∈[1, …, 12] pi[j][,1] [∈]= max[[][1,][ . . .] from base model �p[, 12]1[j] [,][ . . .][]][ from base model][,][ p]m[j] _j�._ 𝑀� is the class represented by the output with [ M] _[j][ is the class represented by the output with]_
maximum The second largest value in the output vector is notated asp values 𝑝��,� = max [𝑝��, …, 𝑝�� �]. _pi[j][,2][.][ p][ values will be used for later]_
conflict resolution in Algorithms 2–5.
The second largest value in the output vector is notated as Base Model Compositions: 𝑝��,�. p values will be used for later conflict
Universal set C represents the set of all classes of activities; C[j] represents activity classes represented
resolution in Algorithms 2–5.
by the time domain of each base model M _j_
Base Model Compositions:
Universal set C represents the set of all classes of activities; 𝐶[�] represents activity classes represented
by the time domain of each base model 𝑀�
-----
_Sensors 2020, 20, 216_ 14 of 26
_C�[j]_ is the complement class for base model M _j and it combines the activity classes not in the C[j]_
denoted below
�
_k ∈_ _C : k �_ _C[j][�]_
Example: Morning Base Model M1 contains activities from classes
_C[1]_ = [Act24, ActN3]
_C�[1]_ = [{ActN4, Act23, ActN5, Act01, Act15, Act17, Act18, Act22, ActN1, ActN2}]
There are mj = 3 number of classes, where all but one class, the complement, are in C[1].
The morning model contains two main activity classes, namely Act24 wake up and ActN3
breakfast, as these activities occur in a typical morning routine. ActN4 lunch, is the only main class
within the afternoon model as lunch usually occurs in the afternoon. The evening model contains
two main classes, namely Act23 go to bed and ActN5 dinner, as these activities habitually occur in an
evening routine. Finally, the mixed model contains seven main activity classes that do not regularly
occur within a specific time routine. For example, Act15 put waste in the bin and Act22 dressing are
activities commonly performed at any time during the day. The activity class outputs per model are
presented in Table 5.
**Table 5. Activity class outputs per model.**
**#output** **Model ID** **Name** **Activity Classes**
m1 = 3 _M1_ Morning
m2 = 2 _M2_ Afternoon
m3 = 3 _M3_ Evening
m4 = 8 _M4_ Mixed
_C[1]_ = [Act24, ActN3] ← 2 classes
_C�[1]_ = [ActN4, Act23, ActN5, Act01,
Act15, Act17, Act18, Act22, ActN1,
ActN2] ← 1 class
_C[2]_ = [ActN4] ← 1 class
_C�[2]_ = [Act24, ActN3, Act23, ActN5,
Act01, Act15, Act17, Act18, Act22,
ActN1, ActN2] ← 1 class
_C[3]_ = [Act23, ActN5] ← 2 classes
_C�[3]_ = [Act24, ActN3, ActN4, Act01,
Act15, Act17, Act18, Act22, ActN1,
ActN2] ← 1 class
_C[4]_ = [Act01, Act15, Act17, Act18,
Act22, ActN1 ActN2] ← 7 classes
_C�[4]_ = [Act24, ActN3, ActN4, Act23,
ActN5] ← 1 class
A framework for the implemented homogenous ensemble approach is presented in Figure 5,
where the conflict resolution approaches are compared. Each base model is presented with an input
feature vector consisting of data produced by 30 binary sensors and an additional time routine feature,
resulting in a total of 31 input features. Each of the base models produce output predictions derived
from the estimated likelihood of each class, which are subsequently combined through the support
function fusion [56] during the ensemble integration phase.
-----
feature, resulting in a total of 31 input features. Each of the base models produce output predictions
derived from the estimated likelihood of each class, which are subsequently combined through the Sensors 2020, 20, 216 15 of 26
support function fusion [56] during the ensemble integration phase.
**Figure 5. Framework for the homogeneous ensemble approach. M1, M2, and M3 represent the Morning,**
**Figure 5.** Framework for the homogeneous ensemble approach. _M1,_ _M2, and_ _M3 represent the_
Afternoon, and Evening models, respectively, and M4 represents the Mixed model.
Morning, Afternoon, and Evening models, respectively, and M4 represents the Mixed model.
Due to each model having no overlapping classes, each needs to be trained with a complement
class, which consists of representative activity samples from each of the main classes contained within
the remaining models. The aim of this is that each model will be able to identify whether or not new
activity instances belong to that model, thus when a model receives an unseen input of an activity
-----
_Sensors 2020, 20, 216_ 16 of 26
class existing within its complement, it should recognize that the activity does not exist as a main
class in the model and should, therefore, eliminate itself from the decision process. For example, if the
morning model is presented with an activity instance contained in the _C[�][1]_ class, e.g., ActN4, as presented
in Table 5, it should recognize that ActN4 belongs to the complement class and should therefore exclude
itself from the decision making process. To analyze the effects on model conflicts of various data
distributions that construct the complement classes per model, we explore two approaches towards
generating these classes. Section 4.2.1 explains the generation of the complement class data at a model
level, where activity instances are distributed evenly between the remaining models, and Section 4.2.2
explains the generation of the complement class data at a class level, where activity instances are
distributed evenly between the remaining classes.
4.2.1. Complement Class Generation at a Model Level
Distributing instances at a model level involves balancing the complement class data equally
between the remaining models. The first step in the process is to calculate how many instances this class
should contain, in total. Per model, this is calculated as the average number of main class instances.
This total is then divided by the number of remaining models to achieve an equal distribution of
instances per model. Following this, the class distributions are calculated by dividing the number of
instances per model by the number of main classes within each model. Table 6 presents the distribution
of instances at a model level.
**Table 6. Model level-distribution of instances for complement class compositions.**
**Model**
**Complement** **Distribution**
**(No. of Instances)**
**Class Distribution**
**(No. of Instances)**
Act17 (03)
Act18 (04)
Act22 (04)
ActN1 (04)
ActN2 (04)
Act15 (09)
Act17 (09)
Act01 (09)
Act22 (09)
ActN1 (09)
ActN2 (09)
Act15 (04)
Act17 (04)
Act22 (04)
ActN1 (04)
ActN2 (04)
Act23 (12)
ActN5 (12)
ActN4 (24)
Act23 (12)
ActN5 (12)
Act01 (03)
Act15 (03)
Act24 (31)
ActN3 (31)
Act23 (31)
ActN5 (31)
Act18 (08)
Act24 (13)
ActN3 (14)
ActN4 (27)
Act18 (03)
Act01 (04)
Act24 (12)
ActN3 (12)
ActN4 (24)
complement class _C[�][1]_ of M1
complement class _C[�][2]_ of M2
complement class _C[�][3]_ of M3
complement class _C[�][4]_ of M4
Afternoon (24)
Evening (24)
Mixed (25)
Morning (62)
Evening (62)
Mixed (62)
Morning (27)
Afternoon (27)
Mixed (27)
Morning (24)
Afternoon (24)
Evening (25)
4.2.2. Complement Class Generation at a Class Level
Distributing instances at a class level involves balancing the complement class data equally
between the remaining classes within the models. As with the previous approach, the first step
involves calculating the average number of main class instances per model to attain the total instances
for each complement class. Following this, the previously calculated total is divided by the number
of remaining classes across the remaining models to achieve an equal distribution of instances per
class. Finally, all instances per class were multiplied by 2 to better represent each class. For example,
to generate the M1 complement class, the average number of main class instances was calculated
-----
_Sensors 2020, 20, 216_ 17 of 26
first, resulting in 74. Subsequently, to achieve an equal distribution of instances per class within the
complement, 74 was divided by the 10 remaining classes, resulting in 7.4 instances required per class.
Finally, to better represent each class during training, this number was multiplied by 2, resulting in 14.8
(15) instances per class. Table 7 presents the distribution of instances at a class level.
**Table 7. Class level-distribution of instances for complement class compositions.**
**Model**
**Complement** **Distribution**
**(No. of Instances)**
**Class Distribution**
**(No. of Instances)**
Act17 (15)
Act18 (15)
Act22 (15)
ActN1 (15)
ActN2 (15)
Act15 (34)
Act17 (34)
Act01 (34)
Act22 (34)
ActN1 (34)
ActN2 (34)
Act15 (16)
Act17 (16)
Act22 (16)
ActN1 (16)
ActN2 (16)
Act23 (29)
ActN5 (29)
ActN4 (15)
Act23 (15)
ActN5 (15)
Act01 (15)
Act15 (15)
Act24 (34)
ActN3 (34)
Act23 (34)
ActN5 (34)
Act18 (34)
Act24 (16)
ActN3 (16)
ActN4 (16)
Act18 (16)
Act01 (16)
Act24 (29)
ActN3 (29)
ActN4 (29)
complement class _C[�][1]_ of M1
complement class _C[�][2]_ of M2
complement class _C[�][3]_ of M3
complement class _C[�][4]_ of M4
_4.3. Model Conflict Resolution_
Afternoon (15)
Evening (30)
Mixed (105)
Morning (68)
Evening (68)
Mixed (238)
Morning (32)
Afternoon (16)
Mixed (112)
Morning (58)
Afternoon (29)
Evening (58)
As mentioned, support function fusion [56] is explored through combining the output predictions
produced by each MLP base model during the ensemble integration phase. The combined predictions
are subsequently analyzed to determine whether a single model has chosen the final output, i.e., all
models except one had chosen the complement class. If this is not the case, and more than one model
has chosen a main class output, a conflict has occurred between these models during the decision
making process, as seen in Algorithm 1. We investigate several approaches to the model conflict
resolution to determine the final output class per instance.
**Algorithm 1. Process of finding conflicts between models**
1: For Each instance →x i ∈ _B1×d_
2: _if ∃j�kˆij_ [∈] _[C]_ _[j][�]_ Λ∃jj �kˆijj [∈] _[C]_ _[jj][ Λ][ j][ �]_ _[jj]�_
Then use conflict resolution approaches in Algorithms 2/3/4/5 as there are at least 2
3:
conflicting cases
The first method of resolving conflicts, presented in Algorithm 2, is simply to award the final
decision to the model with the highest output prediction. This approach has previously been established
as a soft-level combiner [82], as it makes use of the output predictions given by the classifiers as the
posterior probabilities of each output class. A limitation of this method, however, is that it provides
limited confidence of the output prediction. For example, consider the two largest output values of one
base model are 0.56 and 0.54, respectively. If the final class decision is awarded according to the highest
output value in this case, there is less confidence in the quality of classification, which implies a less
secure output prediction. To overcome this, another technique, presented in Algorithm 3, is proposed
-----
_Sensors 2020, 20, 216_ 18 of 26
to calculate the difference between the highest and second highest predictions per conflicting model,
where subsequently the final decision is given to the model with the highest differential value, i.e., this
is the model with the strongest class prediction. Following this, the impact of a weighting technique
is investigated in Algorithm 4 on the basis of the number of classes per model, as each base model
contains a different number of unique classes. This approach considers the output predictions from
each conflicting base classifier and the number of classes the base models are trained on, i.e., the output
predictions from each base model are multiplied by the number of classes within those base models.
For example, if a conflict occurs between model M2 and model M4, which contain two and eight classes,
respectively, the two class problem may be less complex than the eight class problem, and therefore
a lower weighting is specified for M2. Finally, we explore the potential of another weighted method
in Algorithm 5, which builds upon the previous approach. Weightings are implemented on the basis of
the number of classes, as well as the training performance per model, i.e., the output predictions from
each conflicting base classifier are multiplied by the number of classes in that model and the training
performance achieved. According to [83], a base classifier that outperforms other base classifiers
in an ensemble approach should be given a higher confidence when deciding upon the final output
prediction, as the training performance measure is indicative of the classifiers’ effectiveness in predicting
the correct output class. The training performance measure in Algorithm 5 is the classification accuracy
obtained by each conflicting model.
Repeated notations:
The largest value in the output vector is notated as p[j][,1] .
_i_
The second largest value in the output vector is notated as p[j][,2]
_i_ [.]
**Algorithm 2. Conflict resolution approach 1**
**Input:** →x i, base models Mr, Ms
**Output: class yi**
1: _i f p[r][,1]_ - p[s][,1]
_i_ _i_
2: Then yi = _k[ˆ]i[r]_
3: Else yi = _k[ˆ]i[s]_
**Algorithm 3. Conflict resolution approach 2**
**Input:** →x i, base models Mr, Ms
**Output: class yi**
� � � �
1: _i f_ _p[r]i_ [,1] − _p[r]i_ [,2] - _p[s]i_ [,1] − _p[s]i_ [,2]
2: Then yi = _k[ˆ]i[r]_
3: Else yi = _k[ˆ]i[s]_
**Algorithm 4. Conflict resolution approach 3**
**Input:** →x i, base models Mr, Ms
**Output: class yi**
1: _i f p[r]i_ [,1] × mr > p[s]i [,1] × ms
2: Then yi = _k[ˆ]i[r]_
3: Else yi = _k[ˆ]i[s]_
-----
_Sensors 2020, 20, 216_ 19 of 26
**Algorithm 5. Conflict resolution approach 4**
**Input:** →x i, base models Mr, Ms
**Output: class yi**
1: _Acc[r]_
_train_ [represents training performance for base model][ M][r]
2: _Acc[s]_
_train_ [represents training performance for base model][ M][s]
3: _if p[r]i_ [,1] × mr × Acc[r]train [>][ p]i[s][,1] × ms × Acc[s]train
4: Then yi = _k[ˆ]i[r]_
5: Else yi = _k[ˆ]i[s]_
**5. Results and Discussion**
The results show that the class level distribution technique, described in Section 4.2.2,
greatly reduces the number of conflicts that occur between the various base models, in comparison to
the model level distribution technique, as shown in Table 8. This is due to improved representations of
activities within the complement classes per model during the training phase of the base classifiers.
For example, with the class level distribution technique activity instances were distributed evenly
between classes, therefore evenly representing each activity within the complement class. Contrarily,
the model level distribution technique involved balancing the complement class data equally between
the remaining models, which meant the class distributions within these models were imbalanced.
For example, with the model level distribution technique, the _C[�][1]_ complement class contained 24
instances of ActN4 and only 03 instances of Act17, whereas with the class level distribution technique,
the _C[�][1]_ complement class contained 15 instances each of ActN4 and Act17. Consequently, with the
implementation of the latter distribution technique, the base classifiers are stronger at deciding when
an unseen instance belongs to their complement class, eliminating themselves from the decision-making
process and therefore reducing the number of conflicts that occur.
**Table 8. Number of conflicts.**
**No. of Conflicts Per Fold**
1 2 3 4 5 6 7 8 9 10 Avg.
Complement Class –
76 57 69 52 49 35 60 45 62 56 56.1
Model Level Approach
Complement Class –
21 37 11 13 13 42 29 39 11 17 23.3
Class Level Approach
Classification performance from each of the two data distribution techniques were analyzed
before and after conflict resolution approaches were applied, as presented in Figure 6. Considering the
complement class generation at a model level, the preliminary performance accuracy of 60.28% is much
less than that of the complement class generation at a class level, which achieves a preliminary accuracy
of 72.12%. This is due to less model conflicts occurring in the latter approach, which shows the base
models were stronger during the decision-making process. As for the final accuracies produced after
conflict resolution techniques had been applied, the class level approach outperformed the model
level approach in all four cases. Finally, overall, the best HAR performance of 80.39% was achieved
using complement data generated at a class level in conjunction with the conflict resolution approach
presented in Algorithm 3, i.e., resolving conflicts through calculating the difference between the highest
and second highest predictions per conflicting model, where the final decision is given to the model
with the highest differential value.
-----
_Sensors 2020, 20, 216_ 20 of 26
_Sensors 2019, 19, x FOR PEER REVIEW_ 20 of 26
### HAR Performance per Conflict Resolution Approach
90
77.6 79.27 80.2280.39 76.2 79.05 79.8380.17
80 72.12
70
60.28
60
50
40
30
20
10
0
Algorithm 2 Algorithm 3 Algorithm 4 Algorithm 5
Prelim % Final % Final % Final % Final %
Complement Class - Model Level Approach Complement Class - Class Level Approach
**Figure 6. Human Activity Recognition (HAR) performance per conflict resolution approach.**
**Figure 6. Human Activity Recognition (HAR) performance per conflict resolution approach.**
Table 9 presents an analysis of incorrectly classified instances with regards to the first data Table 9 presents an analysis of incorrectly classified instances with regards to the first data
distribution approach where complement class data was generated at a model level, as discussed distribution approach where complement class data was generated at a model level, as discussed
previously in Section 4.2.1, whereas Table 10 presents an analysis of incorrectly classified instances previously in Section 4.2.1, whereas Table 10 presents an analysis of incorrectly classified instances with
with regards to the second data distribution approach, where complement class data was generated regards to the second data distribution approach, where complement class data was generated at a class
at a class level, as discussed previously in Section 4.2.2. The “incorrect” instances reported describe level, as discussed previously in Section 4.2.2. The “incorrect” instances reported describe those that
those that were incorrectly classified by the target model, for example, there may not have been any were incorrectly classified by the target model, for example, there may not have been any conflicting
conflicting models, yet the incorrect class was chosen by the base classifier. The number of incorrectly models, yet the incorrect class was chosen by the base classifier. The number of incorrectly classified
classified instances are important to consider when analyzing the effectiveness of each conflict instances are important to consider when analyzing the effectiveness of each conflict resolution
resolution approach, as these cases would permanently be incorrect, regardless of the application of approach, as these cases would permanently be incorrect, regardless of the application of conflict
conflict resolution techniques. resolution techniques.
The “right but incorrect” cases are those that were correctly classified by the target base model,
although they were not chosen during the final decision-making process after applying the conflict Table 9. Ensemble approach 1—analysis of incorrect instances.
resolution approaches. These cases are considered when evaluating the most effective approach of
**Fold**
the four explored, as they could have resulted in a correct classification, given the application of an
1 2 3 4 5 6 7 8 9 10 Avg.
effective conflict resolution technique.
Algorithm Incorrect 22 22 21 29 29 20 30 22 20 22 23.7
2 Right but
17 18 21 12 17 16 9 14 20 20 16.4
IncorrectTable 9. Ensemble approach 1—analysis of incorrect instances.
Algorithm Incorrect 23 22 21 29 29 22 29 22 20 24 24.1
3 Right but **Fold**
10 14 10 9 12 12 9 12 14 11 11.3
Incorrect 1 2 3 4 5 6 7 8 9 10 Avg.
Algorithm Algorithm IncorrectIncorrect 22 22 23 22 21 21 2929 2929 2220 30 29 22 22 20 20 22 22 23.7 23.9
2 4 Right but Incorrect Right but 17 18 21 12 17 16 9 14 20 20 16.4
31 22 13 23 11 15 23 18 10 21 18.7
Algorithm IncorrectIncorrect 23 22 21 29 29 22 29 22 20 24 24.1
Algorithm3 Right but Incorrect Incorrect 22 10 22 14 21 10 299 2912 2212 9 29 12 22 14 20 11 22 11.3 23.8
Algorithm 5 Right butIncorrect 22 23 21 29 29 22 29 22 20 22 23.9
14 10 13 7 13 15 9 17 14 11 12.3
4 Right but Incorrect Incorrect 31 22 13 23 11 15 23 18 10 21 18.7
Algorithm Incorrect 22 22 21 29 29 22 29 22 20 22 23.8
5 Right but Incorrect 14 10 13 7 13 15 9 17 14 11 12.3
**Table 10. Ensemble approach 2—analysis of incorrect instances.**
**Fold**
1 2 3 4 5 6 7 8 9 10 Avg.
Algorithm Incorrect 33 26 35 33 25 32 27 28 40 26 30.5
2 Right but Incorrect 6 9 2 6 4 11 8 10 2 8 6.6
Algorithm Incorrect 33 26 35 33 25 31 27 28 40 26 30.4
3 Right but Incorrect 5 7 3 2 6 7 6 5 0 6 4.7
-----
_Sensors 2020, 20, 216_ 21 of 26
**Table 10. Ensemble approach 2—analysis of incorrect instances.**
**Fold**
1 2 3 4 5 6 7 8 9 10 Avg.
Algorithm Incorrect 33 26 35 33 25 32 27 28 40 26 30.5
2 Right but
6 9 2 6 4 11 8 10 2 8 6.6
Incorrect
Algorithm Incorrect 33 26 35 33 25 31 27 28 40 26 30.4
3 Right but
5 7 3 2 6 7 6 5 0 6 4.7
Incorrect
Algorithm Incorrect 33 26 35 33 25 31 27 28 40 25 30.3
4 Right but
8 21 4 3 6 6 11 7 1 5 7.2
Incorrect
Algorithm Incorrect 33 26 34 33 25 31 27 28 40 25 30.2
5 Right but
8 8 5 2 6 5 6 7 1 5 5.3
Incorrect
The “right but incorrect” cases are those that were correctly classified by the target base model,
although they were not chosen during the final decision-making process after applying the conflict
resolution approaches. These cases are considered when evaluating the most effective approach of the
four explored, as they could have resulted in a correct classification, given the application of an effective
conflict resolution technique.
The conflict resolution approach presented in Algorithm 3 was the most effective when applied to
both data distributions, as there were the lowest number of “right but incorrect” instances (on average
11.3 and 4.7, respectively), closely followed by the approach in Algorithm 5. The lower the number
of “right but incorrect” cases helps to determine which conflict resolution approach is most effective
in deciding upon which base model should be awarded the final class decision. For example, consider
the conflict resolution technique in Algorithm 3 with ensemble approach 2, as presented in Table 10.
There were 23.3 conflicts occurring on average (refer to Table 8). Upon analysis of the incorrectly
classified instances, 30.4, on average, were incorrectly classified, whereas 4.7, on average, could have
been correctly classified, though an incorrect base model won the final decision after applying conflict
resolution. Finally, this means that as a result of applying Algorithm 3, an average of 18.6 conflicting
cases were correctly resolved, improving the final HAR performance.
As shown in Figure 6, the best HAR performance of 80.39% was achieved using complement data
generated at a class level in conjunction with the conflict resolution approach presented in Algorithm 3.
Given the non-parametric nature of the neural networks, two non-parametric benchmark classifiers
were chosen to evaluate the proposed ensemble approach, namely, Support Vector Machine (SVM) and
Nearest Neighbour (kNN) classifiers. The multiclass SVM classifier was an error-correcting output
codes (ECOC) model required for multiclass learning, consisting of multiple binary learners. Figure 7
presents the performance of our ensemble approach in comparison to the chosen non-parametric
benchmark classifiers. The kNN model achieved an accuracy of 70.95%, whereas the SVM model
achieved 76.54%, thus demonstrating that the proposed ensemble approach outperformed both
benchmark classifiers.
-----
p y
whereas the SVM model achieved 76.54%, thus demonstrating that the proposed ensemble approach Sensors 2020, 20, 216 22 of 26
outperformed both benchmark classifiers.
#### Benchmark Comparison of HAR Performance
82
80.39
80
78
76.54
76
74
72 70.95
70
68
66
Ensemble NN kNN SVM
Classification Models
**Figure 7. Figure 7.HAR performance of the proposed ensemble Neural Network (NN) approach compared to HAR performance of the proposed ensemble Neural Network (NN) approach compared to**
Nearest Neighbour (kNN) and Support Vector Machine (SVM) classifiers. Nearest Neighbour (kNN) and Support Vector Machine (SVM) classifiers.
**6. Conclusions**
**6. Conclusions**
In this work, we focused on data-driven approaches to HAR and addressed the current challenges
In this work, we focused on data-driven approaches to HAR and addressed the current
of their application to openly available datasets. We proposed an ensemble approach to recognize
challenges of their application to openly available datasets. We proposed an ensemble approach to
ADLs within a smart environment setting, with particular emphasis on exploring various approaches
recognize ADLs within a smart environment setting, with particular emphasis on exploring various
to resolving conflicts that occur between base models in ensemble classifiers and analyzing the effects
approaches to resolving conflicts that occur between base models in ensemble classifiers and
of various data distributions that generate the complement class per base model. It was observed
analyzing the effects of various data distributions that generate the complement class per base model.
that distributing data at a class level greatly reduces the number of conflicts that occur between
the base models, leading to an increased preliminary performance before the application of conflict
resolution techniques. It was also found that the best method of resolving conflicts, in comparison to
other approaches explored, is to award the final decision to the model with the highest differential
value between the highest and second highest predictions per conflicting model. We evaluated our
proposed HAR classification model, the ensemble NN method, by comparing the achieved HAR
performance with two non-parametric benchmark classifiers. The ensemble NN method outperformed
both benchmark models, demonstrating the effectiveness of the proposed ensemble approach.
This work is limited in that feature selection techniques were not applied to determine an optimal
subset of input features. According to [84], feature selection is an increasingly significant consideration
in machine learning, with the primary aim of its application being to reduce the dimensionality
in large, multi-dimensional datasets. Thus, future work would involve the application of feature
selection techniques to determine the optimal subset of features required for the classification problem.
Additionally, this work is limited in that the proposed approach was evaluated on one HAR dataset,
therefore future work would involve evaluating the methods on another dataset so that results are not
subjective to only the current dataset.
**Author Contributions: Conceptualization, N.I., C.N., S.Z., H.W., and W.W.Y.N.; methodology, N.I., C.N.,**
S.Z., and W.W.Y.N.; software, N.I.; validation, N.I., C.N., S.Z., and W.W.Y.N.; formal analysis, N.I., C.N.,
S.Z., and W.W.Y.N.; investigation, N.I.; resources, N.I., C.N., S.Z., H.W., and W.W.Y.N.; data curation, N.I.;
writing-original draft preparation, N.I.; writing-review and editing, N.I., C.N., S.Z., and W.W.Y.N.; visualization,
N.I., C.N., S.Z., and W.W.Y.N.; supervision, C.N., S.Z., H.W., and W.W.Y.N.; project administration, N.I., C.N.,
S.Z., H.W., and W.W.Y.N.; funding acquisition. All authors have read and agreed to the published version of
the manuscript.
**Funding: This research was supported through a Northern Ireland Department for the Economy (DfE) PhD**
scholarship. The APC was funded through the DfE PhD scholarship.
-----
_Sensors 2020, 20, 216_ 23 of 26
**Conflicts of Interest: The authors declare no conflict of interest.**
**References**
1. Ranasinghe, S.; al Machot, F.; Mayr, H.C. A review on applications of activity recognition systems with
[regard to performance and evaluation. Int. J. Distrib. Sens. Netw. 2016, 12, 1550147716665520. [CrossRef]](http://dx.doi.org/10.1177/1550147716665520)
2. Zhao, W.; Yan, L.; Zhang, Y. Geometric-constrained multi-view image matching method based on semi-global
[optimization. Geo-Spat. Inf. Sci. 2018, 21, 115–126. [CrossRef]](http://dx.doi.org/10.1080/10095020.2018.1441754)
3. Abdallah, Z.S.; Gaber, M.M.; Srinivasan, B.; Krishnaswamy, S. Adaptive mobile activity recognition system
[with evolving data streams. Neurocomputing 2015, 150, 304–317. [CrossRef]](http://dx.doi.org/10.1016/j.neucom.2014.09.074)
4. Bakli, M.S.; Sakr, M.A.; Soliman, T.H.A. A spatiotemporal algebra in Hadoop for moving objects.
_[Geo-Spat. Inf. Sci. 2018, 21, 102–114. [CrossRef]](http://dx.doi.org/10.1080/10095020.2017.1413798)_
5. Awad, M.M. Forest mapping: A comparison between hyperspectral and multispectral images and
[technologies. J. For. Res. 2018, 29, 1395–1405. [CrossRef]](http://dx.doi.org/10.1007/s11676-017-0528-y)
6. Jalal, A.; Quaid, M.A.K.; Hasan, A.S. Wearable sensor-based human behavior understanding and recognition
in daily life for smart environments. In Proceedings of the 2018 International Conference on Frontiers of
Information Technology (FIT), Islamabad, Pakistan, 17–19 December 2018; pp. 105–110.
7. Lee, Y.; Choi, T.J.; Ahn, C.W. Multi-objective evolutionary approach to select security solutions. CAAI Trans.
_[Intell. Technol. 2017, 2, 64–67. [CrossRef]](http://dx.doi.org/10.1049/trit.2017.0002)_
8. Cook, D.J.; Crandall, A.S.; Thomas, B.L.; Krishnan, N.C. CASAS: A Smart Home in a Box. Computing Practices
**[2013, 46, 62–69. [CrossRef] [PubMed]](http://dx.doi.org/10.1109/MC.2012.328)**
9. Helal, S.; Chen, C. The Gator Tech Smart House: Enabling Technologies and Lessons Learned. In Proceedings
of the 3rd International Convention on Rehabilitation Engineering & Assistive Technology, Singapore,
22–26 April 2009.
10. Cook, D.J.; Youngblood, M.; Heierman, E.O.; Gopalratnam, K.; Rao, S.; Litvin, A.; Khawaja, F. MavHome:
An Agent-Based Smart Home. In Proceedings of the First IEEE International Conference on Pervasive
Computing and Communications, 2003 (PerCom 2003), Fort Worth, TX, USA, 26 March 2003; pp. 521–524.
11. [The DOMUS Laboratory. Available online: http://domuslab.fr (accessed on 8 November 2019).](http://domuslab.fr)
12. [The Aware Home. Available online: http://awarehome.imtc.gatech.edu (accessed on 8 November 2019).](http://awarehome.imtc.gatech.edu)
13. Krishnan, N.C.; Cook, D.J. Activity Recognition on Streaming Sensor Data. Pervasive Mob. Comput. 2014, 10,
[138–154. [CrossRef] [PubMed]](http://dx.doi.org/10.1016/j.pmcj.2012.07.003)
14. Kamal, S.; Jalal, A.; Kim, D. Depth images-based human detection, tracking and activity recognition using
[spatiotemporal features and modified HMM. J. Electr. Eng. Technol. 2016, 11, 1857–1862. [CrossRef]](http://dx.doi.org/10.5370/JEET.2016.11.6.1857)
15. Buys, K.; Cagniart, C.; Baksheev, A.; de Laet, T.; de Schutter, J.; Pantofaru, C. An adaptable system for
RGB-D based human body detection and pose estimation. J. Vis. Commun. Image Represent. 2014, 25, 39–52.
[[CrossRef]](http://dx.doi.org/10.1016/j.jvcir.2013.03.011)
16. Chen, L.; Hoey, J.; Nugent, C.D.; Cook, D.J.; Yu, Z.; Member, S. Sensor-Based Activity Recognition. IEEE Trans.
_[Syst. Man Cybern. Part C 2012, 42, 790–808. [CrossRef]](http://dx.doi.org/10.1109/TSMCC.2012.2198883)_
17. Azkune, G.; Almeida, A.; López-De-Ipiña, D.; Chen, L. Extending Knowledge-Driven Activity Models
[through Data-Driven Learning Techniques. Expert Syst. Appl. 2015, 42, 3115–3128. [CrossRef]](http://dx.doi.org/10.1016/j.eswa.2014.11.063)
18. Cleland, I.; Donnelly, M.P.; Nugent, C.D.; Hallberg, J.; Espinilla, M. Collection of a Diverse, Naturalistic
and Annotated Dataset for Wearable Activity Recognition. In Proceedings of the 2018 IEEE International
Conference on Pervasive Computing and Communications Workshops (PerCom Workshops), Athens, Greece,
19–23 March 2018.
19. Akhand, M.A.H.; Murase, K. Neural Networks Ensembles: Existing Methods and New Techniques; LAP LAMBERT
Academic Publishing: Khulna, Bangladesh, 2010.
20. Sharkey, A.J.C. Combining Artificial Neural Nets; Springer: London, UK, 1999.
21. Aggarwal, J.K.; Xia, L.; Ann, O.C.; Theng, L.B. Human activity recognition: A review. In Proceedings of
the 2014 IEEE International Conference on Control System, Computing and Engineering (ICCSCE 2014),
Batu Ferringhi, Malaysia, 28–30 November 2014; pp. 389–393.
22. Hegde, N.; Bries, M.; Swibas, T.; Melanson, E.; Sazonov, E. Automatic Recognition of Activities of Daily
Living utilizing Insole-Based and Wrist-Worn Wearable Sensors. IEEE J. Biomed. Health Inform. 2017, 22,
[979–988. [CrossRef] [PubMed]](http://dx.doi.org/10.1109/JBHI.2017.2734803)
-----
_Sensors 2020, 20, 216_ 24 of 26
23. Liu, J.; Sohn, J.; Kim, S. Classification of Daily Activities for the Elderly Using Wearable Sensors. J. Healthc.
_[Eng. 2017, 2017, 8934816. [CrossRef]](http://dx.doi.org/10.1155/2017/8934816)_
24. Pirsiavash, H.; Ramanan, D. Detecting activities of Daily Living in First-Person Camera Views. In Proceedings
of the 2012 IEEE Conference on Computer Vision and Pattern Recognition, Providence, RI, USA,
16–21 June 2012; pp. 2847–2854.
25. Roy, N.; Misra, A.; Cook, D. Ambient and Smartphone Sensor Assisted ADL Recognition in Multi-Inhabitant
[Smart Environments. J. Ambient Intell. Humaniz. Comput. 2016, 7, 1–19. [CrossRef]](http://dx.doi.org/10.1007/s12652-015-0294-7)
26. Moriya, K.; Nakagawa, E.; Fujimoto, M.; Suwa, H.; Arakawa, Y.; Kimura, A.; Miki, S.; Yasumoto, K.
Daily Living Activity Recognition with Echonet Lite Appliances and Motion Sensors. In Proceedings of the
2017 IEEE International Conference on Pervasive Computing and Communications Workshops (PerCom
Workshops), Kona, HI, USA, 13–17 March 2017.
27. Gochoo, M.; Tan, T.; Huang, S. DCNN-Based Elderly Activity Recognition Using Binary Sensors.
In Proceedings of the 2017 International Conference on Electrical and Computing Technologies and
Applications (ICECTA), Ras Al Khaimah, United Arab Emirates, 21–23 November 2017.
28. Singh, D.; Merdivan, E.; Hanke, S. Convolutional and Recurrent Neural Networks for Activity Recognition
in Smart Environment. In Towards Integrative Machine Learning and Knowledge Extraction; Springer: Cham,
Switzerland, 2017; pp. 194–209.
29. Cook, D.J.; Krishnan, N.C. Activity Learning: Discovering, Recognizing and Predicting Human Behavior from
_Sensor Data, 1st ed.; Wiley: Hoboken, NJ, USA, 2015._
30. Mannini, A.; Rosenberger, M.; Haskell, W.L.; Sabatini, A.M.; Intille, S.S. Activity Recognition in Youth Using
[Single Accelerometer Placed at Wrist or Ankle. Med. Sci. Sports Exerc. 2017, 49, 801–812. [CrossRef]](http://dx.doi.org/10.1249/MSS.0000000000001144)
31. Huang, Q.; Yang, J.; Qiao, Y. Person re-identification across multi-camera system based on local descriptors.
In Proceedings of the 2012 Sixth International Conference on Distributed Smart Cameras (ICDSC), Hong Kong,
China, 30 Octorber–2 November 2012; pp. 1–6.
32. Farooq, A.; Jalal, A.; Kamal, S. Dense RGB-D map-based human tracking and activity recognition using skin
joints features and self-organizing map. KSII Trans. Internet Inf. Syst. 2015, 9, 1856–1869.
33. Kamal, S.; Jalal, A. A Hybrid Feature Extraction Approach for Human Detection, Tracking and Activity
[Recognition Using Depth Sensors. Arab. J. Sci. Eng. 2016, 41, 1043–1051. [CrossRef]](http://dx.doi.org/10.1007/s13369-015-1955-8)
34. Böttcher, S.; Scholl, P.M.; van Laerhoven, K. Detecting Transitions in Manual Tasks from Wearables:
[An Unsupervised Labeling Approach. Informatics 2018, 5, 16. [CrossRef]](http://dx.doi.org/10.3390/informatics5020016)
35. Trost, S.G.; Wong, W.-K.; Pfeiffer, K.A.; Zheng, Y. Artificial Neural Networks to Predict Activity Type and
[Energy Expenditure in Youth. Med. Sci. Sport Exerc. 2012, 44, 1801–1809. [CrossRef]](http://dx.doi.org/10.1249/MSS.0b013e318258ac11)
36. Lara, O.D.; Labrador, M.A. A Survey on Human Activity Recognition using Wearable Sensors. IEEE Commun.
_[Surv. Tutor. 2013, 15, 1192–1209. [CrossRef]](http://dx.doi.org/10.1109/SURV.2012.110112.00192)_
37. Synnott, J.; Nugent, C.; Zhang, S.; Calzada, A.; Cleland, I.; Espinilla, M.; Quero, J.M.; Lundstrom, J.
Environment Simulation for the Promotion of the Open Data Initiative. In Proceedings of the 2016 IEEE
International Conference on Smart Computing (SMARTCOMP), St. Louis, MO, USA, 18–20 May 2016.
38. Oniga, S.; József, S. Optimal Recognition Method of Human Activities Using Artificial Neural Networks.
_[Meas. Sci. Rev. 2015, 15, 323–327. [CrossRef]](http://dx.doi.org/10.1515/msr-2015-0044)_
39. [Greengard, S. GPUs Reshape Computing. Commun. ACM 2016, 59, 14–16. [CrossRef]](http://dx.doi.org/10.1145/2967979)
40. Suto, J.; Oniga, S. Efficiency investigation from shallow to deep neural network techniques in human activity
[recognition. Cogn. Syst. Res. 2019, 54, 37–49. [CrossRef]](http://dx.doi.org/10.1016/j.cogsys.2018.11.009)
41. Anguita, D.; Ghio, A.; Oneto, L.; Parra, X.; Reyes-Ortiz, J.L. A Public Domain Dataset for Human Activity
Recognition Using Smartphones. Eur. Symp. Artif. Neural Netw. 2013, 437–442, 9782874190827.
42. Rooney, N.; Patterson, D.; Nugent, C. Ensemble Learning for Regression. Encyclopedia Data Warehous. Mining
_Inf. Sci. Ref. N. Y. US 2010, 2, 777–782._
43. Mendes-Moreira, J.; Soares, C.; Jorge, A.M.; de Sousa, J.F. Ensemble Approaches for Regression: A Survey.
_[ACM Comput. Surv. 2012, 45, 10. [CrossRef]](http://dx.doi.org/10.1145/2379776.2379786)_
44. Fatima, I.; Fahim, M.; Lee, Y.-K.; Lee, S. Classifier ensemble optimization for human activity recognition
in smart homes. In Proceedings of the 7th International Conference on Ubiquitous Information Management
and Communication, Kota Kinabalu, Malaysia, 17–19 January 2013; pp. 1–7.
45. Ni, Q.; Zhang, L.; Li, L. A Heterogeneous Ensemble Approach for Activity Recognition with Integration of
[Change Point-Based Data Segmentation. Appl. Sci. 2018, 8, 1695. [CrossRef]](http://dx.doi.org/10.3390/app8091695)
-----
_Sensors 2020, 20, 216_ 25 of 26
46. Feng, Z.; Mo, L.; Li, M. A Random Forest-based ensemble method for activity recognition. In Proceedings of
the 2015 37th Annual International Conference of the IEEE Engineering in Medicine and Biology Society
(EMBC), Milan, Italy, 25–29 August 2015; pp. 5074–5077.
47. Kim, Y.J.; Kang, B.N.; Kim, D. Hidden Markov Model Ensemble for Activity Recognition Using Tri-Axis
Accelerometer. In Proceedings of the 2015 IEEE International Conference on Systems, Man, and Cybernetics,
Kowloon, China, 9–12 October 2016; pp. 3036–3041.
48. Sagha, H.; Bayati, H.; Millán, J.D.R.; Chavarriaga, R. On-line anomaly detection and resilience in classifier
[ensembles. Pattern Recognit. Lett. 2013, 34, 1916–1927. [CrossRef]](http://dx.doi.org/10.1016/j.patrec.2013.02.014)
49. A Genetic Algorithm-based Classifier Ensemble Optimization for Activity Recognition in Smart Homes.
_[KSII Trans. Internet Inf. Syst. 2013, 7, 2853–2873. [CrossRef]](http://dx.doi.org/10.3837/tiis.2013.11.018)_
50. JMin, K.; Cho, S.B. Activity recognition based on wearable sensors using selection/fusion hybrid ensemble.
In Proceedings of the 2011 IEEE International Conference on Systems, Man, and Cybernetics, Anchorage,
AK, USA, 9–12 October 2011; pp. 1319–1324.
51. Diep, N.N.; Pham, C.; Phuong, T.M. Motion Primitive Forests for Human Activity Recognition Using
Wearable Sensors. In Pacific Rim International Conference on Artificial Intelligence; Springer: Cham, Switzerland,
2016; pp. 340–353.
52. Ijjina, E.P.; Mohan, C.K. Hybrid deep neural network model for human action recognition. Appl. Soft Comput.
_[J. 2016, 46, 936–952. [CrossRef]](http://dx.doi.org/10.1016/j.asoc.2015.08.025)_
53. Hwang, I.; Park, H.M.; Chang, J.H. Ensemble of deep neural networks using acoustic environment
classification for statistical model-based voice activity detection. Comput. Speech Lang. 2016, 38, 1–12.
[[CrossRef]](http://dx.doi.org/10.1016/j.csl.2015.11.003)
54. Guan, Y.; Ploetz, T. Ensembles of Deep LSTM Learners for Activity Recognition using Wearables. Proc. ACM
_[Interact. Mob. Wearable Ubiquitous Technol. 2017, 1, 11. [CrossRef]](http://dx.doi.org/10.1145/3090076)_
55. Nweke, H.F.; Teh, Y.W.; Mujtaba, G.; Al-garadi, M.A. Data fusion and multiple classifier systems for human
[activity detection and health monitoring. Inf. Fusion 2019, 46, 147–170. [CrossRef]](http://dx.doi.org/10.1016/j.inffus.2018.06.002)
56. Woz’niak, M.; Woz’niak, W.; Graña, M.; Corchado, E. A survey of multiple classifier systems as hybrid
[systems. Inf. Fusion 2014, 16, 3–17. [CrossRef]](http://dx.doi.org/10.1016/j.inffus.2013.04.006)
57. Díez-Pastor, J.F.; Rodríguez, J.J.; García-Osorio, C.; Kuncheva, L.I. Random Balance: Ensembles of variable
[priors classifiers for imbalanced data. Knowl.-Based Syst. 2015, 85, 96–111. [CrossRef]](http://dx.doi.org/10.1016/j.knosys.2015.04.022)
58. Feng, W.; Huang, W.; Ren, J. Class imbalance ensemble learning based on the margin theory. Appl. Sci. 2018,
_[8, 815. [CrossRef]](http://dx.doi.org/10.3390/app8050815)_
59. Ahmed, S.; Mahbub, A.; Rayhan, F.; Jani, R.; Shatabda, S.; Farid, D.M. Hybrid Methods for Class Imbalance
Learning Employing Bagging with Sampling Techniques. In Proceedings of the 2017 2nd International
Conference on Computational Systems and Information Technology for Sustainable Solution (CSITSS),
Bangalore, India, 21–23 December 2017; pp. 1–5.
60. Farooq, M.; Sazonov, E. Detection of chewing from piezoelectric film sensor signals using ensemble classifiers.
In Proceedings of the 2016 38th Annual International Conference of the IEEE Engineering in Medicine and
Biology Society (EMBC), Orlando, FL, USA, 16–20 August 2016; pp. 4929–4932.
61. Mohammadi, E.; Wu, Q.M.J.; Saif, M. Human activity recognition using an ensemble of support vector
machines. In Proceedings of the 2016 International Conference on High Performance Computing & Simulation
(HPCS), Innsbruck, Austria, 18–22 July 2016; pp. 549–554.
62. Bayat, A.; Pomplun, M.; Tran, D.A. A study on human activity recognition using accelerometer data from
[smartphones. Procedia Comput. Sci. 2014, 34, 450–457. [CrossRef]](http://dx.doi.org/10.1016/j.procs.2014.07.009)
63. Zappi, P.; Stiefmeier, T.; Farella, E.; Roggen, D.; Benini, L.; Tröster, G. Activity recognition from on-body sensors
by classifier fusion: Sensor scalability and robustness. In Proceedings of the 2007 3rd international conference
on intelligent sensors, sensor networks and information, Melbourne, QLD, Australia, 3–6 December 2007;
pp. 281–286.
64. Krawczyk, B.; Wo´zniak, M. Untrained weighted classifier combination with embedded ensemble pruning.
_[Neurocomputing 2016, 196, 14–22. [CrossRef]](http://dx.doi.org/10.1016/j.neucom.2016.02.040)_
65. Roh, Y.; Heo, G.; Whang, S.E. A Survey on Data Collection for Machine Learning: A Big Data—AI Integration
[Perspective. IEEE Trans. Knowl. Data Eng. 2019, 1, 1–19. [CrossRef]](http://dx.doi.org/10.1109/TKDE.2019.2946162)
66. Gorunescu, F. Data Mining: Concepts, Models and Techniques, 1st ed.; Springer: Berlin, Germany, 2011.
-----
_Sensors 2020, 20, 216_ 26 of 26
67. Kantardzic, M. Data Mining: Concepts, Models, Methods and Algorithms, 2nd ed.; Wiley-IEEE Press: Hoboken,
NJ, USA, 2002.
68. Maimon, O.; Rokach, L. Data Mining and Knowledge Discovery Handbook, 1st ed.; Springer: Boston, MA, USA,
2005.
69. Han, J.; Kamber, M.; Pei, J. Data Mining: Concepts and Techniques, 3rd ed.; Elsevier Inc.: Waltham, MA, USA,
2012.
70. Zhu, X.; Wu, X. Class Noise vs. Attribute Noise: A Quantitative Study of Their Impacts. Artif. Intell. Rev.
**[2004, 22, 177–210. [CrossRef]](http://dx.doi.org/10.1007/s10462-004-0751-8)**
71. Folleco, A.A.; Khoshgoftaar, T.M.; van Hulse, J.; Napolitano, A. Identifying Learners Robust to Low Quality
Data. Informatica 2009, 33, 245–259.
72. Espinilla, M.; Medina, J.; Nugent, C. UCAmI Cup. Analyzing the UJA Human Activity Recognition Dataset
[of Activities of Daily Living. MDPI Proc. UCAmI 2018, 2, 1267. [CrossRef]](http://dx.doi.org/10.3390/proceedings2191267)
73. Karvonen, N.; Kleyko, D. A Domain Knowledge-Based Solution for Human Activity Recognition: The UJA
[Dataset Analysis. MDPI Proc. UCAmI 2018, 2, 1261. [CrossRef]](http://dx.doi.org/10.3390/proceedings2191261)
74. Lago, P.; Inoue, S. A Hybrid Model Using Hidden Markov Chain and Logic Model for Daily Living Activity
[Recognition. MDPI Proc. UCAmI 2018, 2, 1266. [CrossRef]](http://dx.doi.org/10.3390/proceedings2191266)
75. Seco, F.; Jiménez, A.R. Event-Driven Real-Time Location-Aware Activity Recognition in AAL Scenarios.
_MDPI Proc. UCAmI 2018, 2, 1240._
76. Cerón, J.D.; López, D.M.; Eskofier, B.M. Human Activity Recognition Using Binary Sensors, BLE Beacons,
an Intelligent Floor and Acceleration Data: A Machine Learning Approach. MDPI Proc. UCAmI 2018, 2,
[1265. [CrossRef]](http://dx.doi.org/10.3390/proceedings2191265)
77. Ding, D.; Cooper, R.A.; Pasquina, P.F.; Fici-Pasquina, L. Sensor technology for smart homes. Maturitas 2011,
_[69, 131–136. [CrossRef]](http://dx.doi.org/10.1016/j.maturitas.2011.03.016)_
78. Amiribesheli, M.; Benmansour, A.; Bouchachia, A. A review of smart homes in healthcare. J. Ambient Intell.
_[Humaniz. Comput. 2015, 6, 495–517. [CrossRef]](http://dx.doi.org/10.1007/s12652-015-0270-2)_
79. Jain, N.; Srivastava, V. DATA MINING TECHNIQUES: A SURVEY PAPER. Int. J. Res. Eng. Technol. 2013, 2,
116–119.
80. Bulling, A.; Blanke, U.; Schiele, B. A Tutorial on Human Activity Recognition using Body-Worn Inertial
[Sensors. ACM Comput. Surv. 2014, 46, 33. [CrossRef]](http://dx.doi.org/10.1145/2499621)
81. Banos, O.; Galvez, J.-M.; Damas, M.; Pomares, H.; Rojas, I. Window Size Impact in Human Activity
[Recognition. Sensors 2014, 14, 6474–6499. [CrossRef] [PubMed]](http://dx.doi.org/10.3390/s140406474)
82. Mohandes, M.; Deriche, M.; Aliyu, S.O. Classifiers Combination Techniques: A Comprehensive Review.
_[IEEE Access 2018, 6, 19626–19639. [CrossRef]](http://dx.doi.org/10.1109/ACCESS.2018.2813079)_
83. Yijing, L.; Haixiang, G.; Xiao, L.; Yanan, L.; Jinling, L. Adapted ensemble classification algorithm based on
multiple classifier system and feature selection for classifying multi-class imbalanced data. Knowl.-Based Syst.
**[2016, 94, 88–104. [CrossRef]](http://dx.doi.org/10.1016/j.knosys.2015.11.013)**
84. Suto, J.; Oniga, S.; Sitar, P.P. Comparison of wrapper and filter feature selection algorithms on human activity
recognition. In Proceedings of the 2016 6th International Conference on Computers Communications and
Control (ICCCC), Oradea, Romania, 10–14 May 2016; pp. 124–129.
© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
[(CC BY) license (http://creativecommons.org/licenses/by/4.0/).](http://creativecommons.org/licenses/by/4.0/.)
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://pmc.ncbi.nlm.nih.gov/articles/PMC6982871, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "GOLD",
"url": "https://www.mdpi.com/1424-8220/20/1/216/pdf"
}
| 2,019
|
[
"JournalArticle"
] | true
| 2019-12-30T00:00:00
|
[
{
"paperId": "8772b78e1ff9a3b6db46492ae58af3108a0d5b47",
"title": "Efficiency investigation from shallow to deep neural network techniques in human activity recognition"
},
{
"paperId": "800cd71393a568128e3797b832f3005a664e662f",
"title": "Data fusion and multiple classifier systems for human activity detection and health monitoring: Review and open research directions"
},
{
"paperId": "db049db27febe25c8af0007ef98ea26cba24bd66",
"title": "Wearable Sensor-Based Human Behavior Understanding and Recognition in Daily Life for Smart Environments"
},
{
"paperId": "3a83d8595e6727269c876fcebd23ee9ddd524b76",
"title": "A Survey on Data Collection for Machine Learning: A Big Data - AI Integration Perspective"
},
{
"paperId": "badee62939c4c87f8ae4bb93012418f634195a99",
"title": "Event-Driven Real-Time Location-Aware Activity Recognition in AAL Scenarios"
},
{
"paperId": "aa42bd924240622f176eff4fe0ce4990c2c55163",
"title": "UCAmI Cup. Analyzing the UJA Human Activity Recognition Dataset of Activities of Daily Living"
},
{
"paperId": "09b563b54a53efe8b564621332df4a4b8f22190f",
"title": "A Hybrid Model Using Hidden Markov Chain and Logic Model for Daily Living Activity Recognition"
},
{
"paperId": "558bce500a47dd94d46c9494b52423d7d4c8b835",
"title": "A Domain Knowledge-Based Solution for Human Activity Recognition: The UJA Dataset Analysis"
},
{
"paperId": "81d3ea8a18a23207a000a9ef9eb31d02ee6f8a4a",
"title": "Human Activity Recognition Using Binary Sensors, BLE Beacons, an Intelligent Floor and Acceleration Data: A Machine Learning Approach"
},
{
"paperId": "745364feecfc3429484a51f9d121b8c8c4225f18",
"title": "A Heterogeneous Ensemble Approach for Activity Recognition with Integration of Change Point-Based Data Segmentation"
},
{
"paperId": "8b085c3215389c4107850eac223c3e3ad253901d",
"title": "Automatic Recognition of Activities of Daily Living Utilizing Insole-Based and Wrist-Worn Wearable Sensors"
},
{
"paperId": "e29d22c32f5c89a5c425c916a1157009aad52521",
"title": "Class imbalance ensemble learning based on the margin theory"
},
{
"paperId": "dcb71a8f3f7e3caa905ab9568973043f3a385aee",
"title": "Detecting Transitions in Manual Tasks from Wearables: An Unsupervised Labeling Approach"
},
{
"paperId": "c6fd6422039178a0abaec8156f64e9b61a54ecfe",
"title": "Collection of a Diverse, Realistic and Annotated Dataset for Wearable Activity Recognition"
},
{
"paperId": "9b120751fd878fbf19040a43d04281cc1c17d0f4",
"title": "Geometric-constrained multi-view image matching method based on semi-global optimization"
},
{
"paperId": "a91545462f2f425161f93f41389e9a1036fbda51",
"title": "A spatiotemporal algebra in Hadoop for moving objects"
},
{
"paperId": "ede39f75804f815d9b7cd2b98d751ebd388f6c18",
"title": "Hybrid Methods for Class Imbalance Learning Employing Bagging with Sampling Techniques"
},
{
"paperId": "f45854a32044148793280fe08590746096b625d3",
"title": "Classification of Daily Activities for the Elderly Using Wearable Sensors"
},
{
"paperId": "9ecca1856f7858ecc58de632352c30061a074e88",
"title": "Forest mapping: a comparison between hyperspectral and multispectral images and technologies"
},
{
"paperId": "45296e2b4e2bf98eb60953091d0906619a42fd72",
"title": "DCNN-based elderly activity recognition using binary sensors"
},
{
"paperId": "5a954a4817848c5b0a921b199a53131fb9ab07cd",
"title": "Multi-objective evolutionary approach to select security solutions"
},
{
"paperId": "04f79497dfb0ee31f0376e465749bf176a8d0ccf",
"title": "Activity Recognition in Youth Using Single Accelerometer Placed at Wrist or Ankle"
},
{
"paperId": "4b8446ea306d7d3f8218a8850158b0ae084a988e",
"title": "Ensembles of Deep LSTM Learners for Activity Recognition using Wearables"
},
{
"paperId": "b65f5acd18f62cbbd0d43f53b195e7321e32685e",
"title": "Daily living activity recognition with ECHONET Lite appliances and motion sensors"
},
{
"paperId": "9ef98f87a6b1bd7e6c006f056f321f9d291d4c45",
"title": "Depth Images-based Human Detection, Tracking and Activity Recognition Using Spatiotemporal Features and Modified HMM"
},
{
"paperId": "012037e6bfef86f140f56eb97e453fcb1ff5cca1",
"title": "Hybrid deep neural network model for human action recognition"
},
{
"paperId": "339410e5d7e41b3477472f7e27dc46ce950939cb",
"title": "GPUs reshape computing"
},
{
"paperId": "86100cac4647f6cb190af5373b8463b177df0e7c",
"title": "Motion Primitive Forests for Human Activity Recognition Using Wearable Sensors"
},
{
"paperId": "de21627c01e8324aaddd6a3a303be7e8fdcb43da",
"title": "Detection of chewing from piezoelectric film sensor signals using ensemble classifiers"
},
{
"paperId": "f559c8f003f2f336e6cca50fffaf724c4bfc7ea7",
"title": "A review on applications of activity recognition systems with regard to performance and evaluation"
},
{
"paperId": "3bd9e2b35052f43ac25d9c27a30e388d9a4a56f3",
"title": "Untrained weighted classifier combination with embedded ensemble pruning"
},
{
"paperId": "fd59460de812c8b250a5756e54c5152b77dd74df",
"title": "Ensemble of deep neural networks using acoustic environment classification for statistical model-based voice activity detection"
},
{
"paperId": "e0ab926cd48a47a8c7b16e27583421141f71f6df",
"title": "Human activity recognition using an ensemble of support vector machines"
},
{
"paperId": "327e601151bb52868e2e353a5f163aa849dd59c8",
"title": "Environment Simulation for the Promotion of the Open Data Initiative"
},
{
"paperId": "5cb4199d013f7ee92bae3cfe65a2170ea5b067ec",
"title": "Comparison of wrapper and filter feature selection algorithms on human activity recognition"
},
{
"paperId": "24d633be7c2068a12e78a78865d4a5b19555cb1d",
"title": "Ambient and smartphone sensor assisted ADL recognition in multi-inhabitant smart environments"
},
{
"paperId": "76a368170db29ef70d34794a91eb18b0bc9cc130",
"title": "Optimal Recognition Method of Human Activities Using Artificial Neural Networks"
},
{
"paperId": "126a413f410146470217456b3becbbd114623c60",
"title": "A Hybrid Feature Extraction Approach for Human Detection, Tracking and Activity Recognition Using Depth Sensors"
},
{
"paperId": "c6cd47443f9e8d523387594935d56827c50e9e26",
"title": "A Random Forest-based ensemble method for activity recognition"
},
{
"paperId": "fd5ec6492b4b25db921ae665f730bbfce27d869c",
"title": "Hidden Markov Model Ensemble for Activity Recognition Using Tri-Axis Accelerometer"
},
{
"paperId": "9d8429a413cd1c4ec8cf1024b5fbec79de91ec57",
"title": "Random Balance: Ensembles of variable priors classifiers for imbalanced data"
},
{
"paperId": "5984388a90b1a675b837703c67880c9a7e28b8bd",
"title": "Dense RGB-D Map-Based Human Tracking and Activity Recognition using Skin Joints Features and Self-Organizing Map"
},
{
"paperId": "5591b632e78bd12791872f32a1a013ab7ded4195",
"title": "Extending knowledge-driven activity models through data-driven learning techniques"
},
{
"paperId": "3fce056a3bed699a27f902116468fce921975d6c",
"title": "A review of smart homes in healthcare"
},
{
"paperId": "99e601284696121e47f160cfa8464faaef609a28",
"title": "Adaptive mobile activity recognition system with evolving data streams"
},
{
"paperId": "cb264f6ab37ac664389344d88bc2b8451005485d",
"title": "Activity Learning: Discovering, Recognizing, and Predicting Human Behavior from Sensor Data"
},
{
"paperId": "e3a48e6b95cd0511a6c2c31ce98a8c86e61f25b5",
"title": "Human activity recognition: A review"
},
{
"paperId": "615a72aadb94a994f9d49548c0dffbeed1631ba2",
"title": "Window Size Impact in Human Activity Recognition"
},
{
"paperId": "8acdf4a1e88b443b917d80e97850af7bbb1e3534",
"title": "A survey of multiple classifier systems as hybrid systems"
},
{
"paperId": "fc79c74063d5183cfbc0994be7b2dcab26bed937",
"title": "Activity recognition on streaming sensor data"
},
{
"paperId": "aea9e985c1d20bd591ca4bceae41d662f06fa174",
"title": "A Genetic Algorithm-based Classifier Ensemble Optimization for Activity Recognition in Smart Homes"
},
{
"paperId": "fefa1bf3fca6c1ce08ecbd63e4604815c11038a0",
"title": "DATA MINING TECHNIQUES: A SURVEY PAPER"
},
{
"paperId": "27c1d65b998cd3aeee052c364c704ff6018bc24d",
"title": "On-line anomaly detection and resilience in classifier ensembles"
},
{
"paperId": "0cf5e93a4fd936672d9763ee577224fcbfa06980",
"title": "CASAS: A Smart Home in a Box"
},
{
"paperId": "8d3041129b500b90521c7d768996fc2de11b0e47",
"title": "A Survey on Human Activity Recognition using Wearable Sensors"
},
{
"paperId": "9cf59101b32f7550312f676de7588df3abf6f2a9",
"title": "Classifier ensemble optimization for human activity recognition in smart homes"
},
{
"paperId": "9e605d72ab39f7a37a13e912556ee034c0d06133",
"title": "Sensor-Based Activity Recognition"
},
{
"paperId": "e30a47911cfb98934dc640eb669b9351827bd37b",
"title": "Ensemble approaches for regression: A survey"
},
{
"paperId": "a19a5a604e07c525f9a8969c05f1c9ed5b497053",
"title": "Person re-identification across multi-camera system based on local descriptors"
},
{
"paperId": "fb5c28ec125db520e7f3ef9a0e4be7831d9316f6",
"title": "Artificial neural networks to predict activity type and energy expenditure in youth."
},
{
"paperId": "9e81caf9dd31b893ebbee3970c312619b7eac7bf",
"title": "Detecting activities of daily living in first-person camera views"
},
{
"paperId": "f4e0ab55f1571a31263a511a92aec8691fc60408",
"title": "Activity recognition based on wearable sensors using selection/fusion hybrid ensemble"
},
{
"paperId": "3adf3a23dd16e6489973d23f6877af615102233e",
"title": "Data Mining - Concepts, Models and Techniques"
},
{
"paperId": "230ef581d3361ac90da495983002dfb50cfcd660",
"title": "Sensor technology for smart homes."
},
{
"paperId": "36e9538f1998b5922ebbcc72695251bbfc324ee2",
"title": "Neural Networks Ensembles: Existing Methods and New Techniques"
},
{
"paperId": "a826892385d54211968a44df6ee8dcb6bc96e963",
"title": "The Gator Tech Smart House: enabling technologies and lessons learned"
},
{
"paperId": "c2bee99b8834e17af7778ef51ccd84cb161deda9",
"title": "Identifying learners robust to low quality data"
},
{
"paperId": "1b4ef3e5a783a18de6439f345af20831dcdcf29e",
"title": "Activity recognition from on-body sensors by classifier fusion: sensor scalability and robustness"
},
{
"paperId": "dfca81b6743e88edb352e601c95af1502dea3077",
"title": "Data Mining: Concepts, Models, Methods, and Algorithms"
},
{
"paperId": "c63a8640e5d426b1c8b0ca2ea45c20c265b3f2ad",
"title": "Class Noise vs. Attribute Noise: A Quantitative Study"
},
{
"paperId": "4b61ae6fc24e533e73b41cd502f94d74e2cd4436",
"title": "MavHome: an agent-based smart home"
},
{
"paperId": "6b0b93396792958c5e58da1933b46a4f565e82c6",
"title": "Data Mining And Knowledge Discovery Handbook"
},
{
"paperId": "756d8af94094d276a2fe6c8e297a3373e3c0b5f9",
"title": "Classifiers Combination Techniques: A Comprehensive Review"
},
{
"paperId": "c6b4e80e9aa37f9f1d1da145e58ae0b57dbaeef7",
"title": "Adapted ensemble classification algorithm based on multiple classifier system and feature selection for classifying multi-class imbalanced data"
},
{
"paperId": "cb206d1aed84388d3a4b83470957819bb6f1cf75",
"title": "Adapted ensemble classification algorithm based on multiple classifier system and feature selection for classifying multi-class imbalanced data"
},
{
"paperId": "13c21287d85fb93d22d041be777e9adadd18abef",
"title": "Convolutional and Recurrent Neural Networks for Activity Recognition in Smart Environment"
},
{
"paperId": "9baa45693e37d6f14a9a0efae7c5f552808b36cb",
"title": "The 11th International Conference on Mobile Systems and Pervasive Computing (MobiSPC-2014) A Study on Human Activity Recognition Using Accelerometer Data from Smartphones"
},
{
"paperId": "b9eb00ee1656f40ae3bbfd8631bda30c1dd9206d",
"title": "A tutorial on human activity recognition using body-worn inertial sensors"
},
{
"paperId": "9f97bec3ba3071b41609f8f4590b8635116d7e9f",
"title": "An adaptable system for RGB-D based human body detection and pose estimation"
},
{
"paperId": "83de43bc849ad3d9579ccf540e6fe566ef90a58e",
"title": "A Public Domain Dataset for Human Activity Recognition using Smartphones"
},
{
"paperId": "de8a9421577b8d22953b84d0574984eab0f33182",
"title": "Ensemble Learning for Regression"
},
{
"paperId": "982b955c900b04e9da64e3b39422690c13d6b94f",
"title": "Data Mining - Concepts and Techniques"
},
{
"paperId": "d5355193db69c1ec9e03cf09e6c5045612d0b273",
"title": "M: Cluster Analysis"
},
{
"paperId": "20202f93d49e782395a6c3bd829678e3fd352924",
"title": "Combining Artificial Neural Nets"
},
{
"paperId": null,
"title": "This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license"
},
{
"paperId": null,
"title": "The DOMUS Laboratory The Aware Home"
}
] | 25,913
|
en
|
[
{
"category": "Engineering",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/028b33b36dcefb6ae6139e06cefa758df632ffeb
|
[
"Engineering"
] | 0.923838
|
Is Security Realistic in Cloud Computing?
|
028b33b36dcefb6ae6139e06cefa758df632ffeb
|
Journal of International Technology and Information Management
|
[
{
"authorId": "153309689",
"name": "S. Srinivasan"
},
{
"authorId": "2119102915",
"name": "Jesse H. Jones"
}
] |
{
"alternate_issns": [
"1941-6679"
],
"alternate_names": [
"J Int Technol Inf Manag"
],
"alternate_urls": null,
"id": "cb34de5f-6970-41cc-b2fe-c801d63d384d",
"issn": "1543-5962",
"name": "Journal of International Technology and Information Management",
"type": "journal",
"url": null
}
|
INTRODUCTION Cloud computing today is benefiting from the technological advancements in communication, storage and computing. The basic idea in cloud computing is to take advantage of economies of scale if IT services could be provided on demand with a decentralized infrastructure. This idea is a natural evolution from the IT time-share model of the 1960s and 1970s. Today, technology has advanced significantly and many more organizations have computing demands that are elastic in nature. Organizations large and small require reliable computing resources in order to succeed in business. Large businesses deal with complex systems where as Small and Medium sized Enterprises (SMEs) need access to affordable computing resources. Based on these aspects we can summarize some of the rationale for today's cloud computing needs as follows: * acquiring and managing the IT resources requires specialized skills, * maintaining a reliable IT infrastructure is expensive, * rapid technology advancements make it difficult to keep current the IT expertise, * internet has opened up many opportunities for individuals as well as small businesses, * number of entities requiring computing resources has grown exponentially, * SMEs' demand for computing resources varies significantly over time, * providing data security is a complex undertaking. In the above paragraph we have identified some of the major reasons as to why cloud computing would be advantageous to use. When a significant part of the business depends on a type of service that the business does not fully control, the question arises as to how the business can meet its obligations to its customers. As highlighted above, IT services are essential to the success of the business but it would be cost prohibitive for the business to manage an IT center with the required expertise and fluctuating demand on resources for processing and storage. Thus, a business using cloud computing must understand the security challenges that it would be responsible for and how cloud computing could help in this regard. We address the security challenges by first noting the differences in the types of cloud computing that a business might be using. In order to address the security challenges associated with cloud computing, we need to understand first the meaning of cloud computing. The primary reason for this is that the term 'cloud computing' is used as a catch-all for a wide ranging array of services. After a careful analysis of numerous sources in the literature we have arrived at the following working definition of 'cloud computing' based primarily on the National Institute of Standards and Technology definition: Cloud computing consists of both the infrastructure and services that facilitate reliable on-demand access to resources that can be allocated and released quickly by the user without provider intervention using the pay-as-you-go model (NIST, 2011). It is worth noting in this context that Mell and Grance further amplified on this general definition in their NIST report that is now widely accepted as one of the important definitions of cloud computing (Mell, 2011). Today's cloud computing has three basic types: Software as a Service (SaaS), Platform as a Service (PaaS) and Infrastructure as a Service (IaaS). In the simplest of terms 'cloud computing' has come to embody SaaS. Similar to the IT time-share model mentioned earlier, SaaS provides both the server hardware and software to an organization without any of the complications of managing an IT system. The simplest example of SaaS service would be email for an organization. The cloud provider benefits from the economies of scale in managing a large infrastructure because of their strength in that area and is able to provide the necessary computing resources to the user, majority of who are SMEs, at an affordable cost. SaaS leaves the full control of the computing system with the provider. …
|
# Journal of International Technology and Information Management Journal of International Technology and Information Management
[Volume 22](https://scholarworks.lib.csusb.edu/jitim/vol22) [Issue 4](https://scholarworks.lib.csusb.edu/jitim/vol22/iss4) [Article 3](https://scholarworks.lib.csusb.edu/jitim/vol22/iss4/3)
11-4-2013
# Is Security Realistic in Cloud Computing? Is Security Realistic in Cloud Computing?
S. Srinivasan
Texas Southern University
[Follow this and additional works at: https://scholarworks.lib.csusb.edu/jitim](https://scholarworks.lib.csusb.edu/jitim?utm_source=scholarworks.lib.csusb.edu%2Fjitim%2Fvol22%2Fiss4%2F3&utm_medium=PDF&utm_campaign=PDFCoverPages)
Recommended Citation Recommended Citation
Srinivasan, S. (2013) "Is Security Realistic in Cloud Computing?," Journal of International Technology and
Information Management: Vol. 22: Iss. 4, Article 3.
[DOI: https://doi.org/10.58729/1941-6679.1020](https://doi.org/10.58729/1941-6679.1020)
[Available at: https://scholarworks.lib.csusb.edu/jitim/vol22/iss4/3](https://scholarworks.lib.csusb.edu/jitim/vol22/iss4/3?utm_source=scholarworks.lib.csusb.edu%2Fjitim%2Fvol22%2Fiss4%2F3&utm_medium=PDF&utm_campaign=PDFCoverPages)
This Article is brought to you for free and open access by CSUSB ScholarWorks. It has been accepted for inclusion
in Journal of International Technology and Information Management by an authorized editor of CSUSB
[ScholarWorks. For more information, please contact scholarworks@csusb.edu.](mailto:scholarworks@csusb.edu)
-----
# Is Security Realistic In Cloud Computing?
**S. Srinivasan**
**Jesse H. Jones School of Business**
**Texas Southern University**
**USA**
**ABSTRACT**
_Cloud computing is rapidly emerging as an attractive IT option for businesses. As a concept_
_cloud computing is well received because of the benefits it offers but many users are not clear_
_about the scope of security in cloud computing. Many surveys point out that security in the cloud_
_remains the top concern for many businesses in their decision making consideration in spite of_
_the cost advantages it offers. In order to identify the security concerns we analyzed over 50_
_research articles and industry white papers published over the past five years. In this paper we_
_focus on the question “Is security realistic in cloud computing?” In presenting the justification_
_that it is possible to expect adequate security features in the cloud we address several related_
_issues. In this context we first briefly describe the three types of cloud services – SaaS, PaaS and_
_IaaS. Then we focus on the security aspects that businesses must pay attention to in order to_
_succeed. Next, we consider the importance of trust in the service providers and how they could_
_build customer trust in their services. This discussion leads to service reliability in the cloud and_
_what the cloud providers have learned from cloud outages in order to build trust. Also, we_
_highlight how the security features offered in the cloud support compliance requirements. We_
_conclude the paper with some relevant information on the legal aspects related to cloud_
_computing._
**INTRODUCTION**
Cloud computing today is benefiting from the technological advancements in communication,
storage and computing. The basic idea in cloud computing is to take advantage of economies of
scale if IT services could be provided on demand with a decentralized infrastructure. This idea is
a natural evolution from the IT time-share model of the 1960s and 1970s. Today, technology has
advanced significantly and many more organizations have computing demands that are elastic in
nature. Organizations large and small require reliable computing resources in order to succeed in
business. Large businesses deal with complex systems where as Small and Medium sized
Enterprises (SMEs) need access to affordable computing resources. Based on these aspects we
can summarize some of the rationale for today’s cloud computing needs as follows:
- acquiring and managing the IT resources requires specialized skills,
- maintaining a reliable IT infrastructure is expensive,
- rapid technology advancements make it difficult to keep current the IT expertise,
- internet has opened up many opportunities for individuals as well as small businesses,
- number of entities requiring computing resources has grown exponentially,
- SMEs’ demand for computing resources varies significantly over time,
- providing data security is a complex undertaking.
-----
In the above paragraph we have identified some of the major reasons as to why cloud computing
would be advantageous to use. When a significant part of the business depends on a type of
service that the business does not fully control, the question arises as to how the business can
meet its obligations to its customers. As highlighted above, IT services are essential to the
success of the business but it would be cost prohibitive for the business to manage an IT center
with the required expertise and fluctuating demand on resources for processing and storage.
Thus, a business using cloud computing must understand the security challenges that it would be
responsible for and how cloud computing could help in this regard. We address the security
challenges by first noting the differences in the types of cloud computing that a business might
be using.
In order to address the security challenges associated with cloud computing, we need to
understand first the meaning of cloud computing. The primary reason for this is that the term
‘cloud computing’ is used as a catch-all for a wide ranging array of services. After a careful
analysis of numerous sources in the literature we have arrived at the following working
definition of ‘cloud computing’ based primarily on the National Institute of Standards and
Technology definition: _Cloud computing consists of both the infrastructure and services that_
_facilitate reliable on-demand access to resources that can be allocated and released quickly by_
_the user without provider intervention using the pay-as-you-go model (NIST, 2011). It is worth_
noting in this context that Mell and Grance further amplified on this general definition in their
NIST report that is now widely accepted as one of the important definitions of cloud computing
(Mell, 2011).
Today’s cloud computing has three basic types: Software as a Service (SaaS), Platform as a
Service (PaaS) and Infrastructure as a Service (IaaS). In the simplest of terms ‘cloud computing’
has come to embody SaaS. Similar to the IT time-share model mentioned earlier, SaaS provides
both the server hardware and software to an organization without any of the complications of
managing an IT system. The simplest example of SaaS service would be email for an
organization. The cloud provider benefits from the economies of scale in managing a large
infrastructure because of their strength in that area and is able to provide the necessary
computing resources to the user, majority of who are SMEs, at an affordable cost. SaaS leaves
the full control of the computing system with the provider. Some of the major commercial SaaS
providers are Amazon, Google, Microsoft and SalesForce.
PaaS provides the customer a platform, such as the Windows operating system with the
necessary server capacity to run the applications for the customer. PaaS is used mainly by
developers who need to test their applications under a variety of conditions. The PaaS cloud
service provider manages the system for its upkeep and provisioning of tools such as .NET and
Java whereas the customer is responsible for the selection of applications that run on the
platform of their choice using the available tools. Thus, the customer is responsible for the
security challenges associated with the applications that they run. For example, a customer
running a SQL Server database on the platform should be aware of the vulnerabilities of the
database system. Hence, the customer should have the expertise to manage such applications on
the platform used under this pay-as-you-go model. The benefit to the customer is that if their
hardware needs change or if they require a Linux/UNIX platform for some other applications,
-----
then provisioning them takes only a few days as opposed to few weeks to make the new system
operational. Major PaaS cloud service providers are Google App Engine and Windows Azure.
IaaS provides the customer the same features as PaaS but the customer is fully responsible for
the control of the leased infrastructure. IaaS may be viewed as the computing system of the
customer that is not owned by them. Unlike PaaS, IaaS requires the organization to have the
necessary people with extensive computing expertise. The IaaS customer would be responsible
for all security aspects of the system that they use except physical security, which would be
handled by the cloud provider. Amazon and IBM are examples of IaaS providers. Combining the
information presented so far about these three types of cloud services with additional cloud
service providers, we have Table 1 that provides a quick snapshot of the available resources.
**Table 1: Summary of cloud service providers.**
**Provider** **Type of** **Product name**
**service**
SaaS AWS
Amazon PaaS Elastic Beanstalk
IaaS EC2, S3
SaaS Gmail, GoogleDocs
Google
PaaS App Engine
Microsoft PaaS Azure
SaaS Sales Cloud
Salesforce.com
PaaS Force.com
Rackspace PaaS Rackspace Cloud
IaaS Rackspace Cloud
SaaS CloudBurst
IBM IaaS Blue Cloud
EMC IaaS Atmos
Apple SaaS iCloud
AT & T SaaS Synaptic Hosting
VMware IaaS vCloud Director
It is worth noting that these three types of services are gaining ground. According to the
Ponemon Institute/CA Technologies 2011 study, among cloud service providers, SaaS accounts
for 55 percent, PaaS accounts for 11 percent and IaaS accounts for 34 percent. Besides these
three service types available, a potential user must also consider the four different cloud
deployment models for meeting their computing needs. The four cloud deployment models are
public cloud, private cloud, hybrid cloud, and community cloud. The most common cloud
deployment model is the public cloud. In the public cloud the customer shares the resources with
other customers. On the other hand, in a private cloud the resource are dedicated to the
organization and has greater security because the computing resources are not shared with other
customers. Private cloud is affordable only for large organizations. A natural evolution from
public cloud and private cloud service models is the hybrid cloud which uses both proprietary
computing resources and/or private cloud resources that the organization manages directly and
the public cloud for some of the computing requirements, especially the ones with varying
|Provider|Type of service|Product name|
|---|---|---|
|Amazon|SaaS|AWS|
||PaaS|Elastic Beanstalk|
||IaaS|EC2, S3|
|Google|SaaS|Gmail, GoogleDocs|
||PaaS|App Engine|
|Microsoft|PaaS|Azure|
|Salesforce.com|SaaS|Sales Cloud|
||PaaS|Force.com|
|Rackspace|PaaS|Rackspace Cloud|
||IaaS|Rackspace Cloud|
|IBM|SaaS|CloudBurst|
||IaaS|Blue Cloud|
|EMC|IaaS|Atmos|
|Apple|SaaS|iCloud|
|AT & T|SaaS|Synaptic Hosting|
|VMware|IaaS|vCloud Director|
-----
demands on resources (Bhattacharjee, 2009). Two of the major hybrid cloud providers currently
are VMware and HP. Another important statistic to note is that 65 percent of the cloud service
customers use public cloud service while 18 percent each use private cloud and hybrid cloud
services. These three types of cloud services aim to meet the customer requirements at different
levels of engagement in managing the computing hardware and software. This has a direct
correlation to the size of the organization in choosing the type of cloud service. For this reason
we can broadly classify the cloud computing users as belonging to either the public cloud or the
private cloud. Small and medium sized businesses typically use the public cloud and large
organizations use the private cloud. All the cloud service providers mentioned earlier provide
both public and private cloud services. In the private cloud, a large organization which has a data
center to manage, is able to use large amounts of storage and computing power dedicated to just
their organization only. The private cloud facilitates the large organization to handle demand
elasticity similar to the public cloud provider.
The community cloud is used by organizations with a common focus such as health care,
automotive and financial services. The community cloud represents a vertical market in which
the organizations stand to benefit by having a dedicated server that addresses the specialized
needs of that sector. For example, in the media industry companies are looking for ways to
simplify content production at low-cost. This requires collaboration among a large group of
people. A community cloud facilitates the location of necessary computing resources for content
production and editing. By using a community cloud dedicated to the media industry this need is
met. Windows Azure platform is used as a public cloud for this community cloud architecture.
Having provided a brief overview of the three basic cloud types and the four deployment models,
let us next review the security aspects in the cloud as discussed in several research articles and
industry white papers. One of the main reasons for the cloud to provide cost efficiencies is its
ability to leverage the economies of scale in their hardware and their ability to offer Virtual
Machines (VMs) on a single hardware for multiple clients. Moreover, cloud providers enable
visibility to the customer on the location of their VM in the cloud. How this feature is exploited
by attackers to launch side-channel attacks on the cloud is the major contribution of Ristenpart,
Tromer, Schacham and Savage (2009). In their oft cited paper “Hey, You, Get off of my cloud,”
these UC San Diego and MIT researchers highlight the security concerns of many businesses.
They point out the data leakage aspect in a public cloud (Ristenpart, 2009). In a multi-tenant
environment on a physical infrastructure, which is very common in a public cloud, such attacks
are capable of extracting encryption keys. Thus, one of the heavily relied upon defense to secure
data storage in the cloud becomes vulnerable. Armbrust et al., discuss in their paper the top 10
obstacles to cloud adoption. These UC Berkeley researchers show the current status of the cloud
service and how the technology needs to improve further to address customer security concerns.
This paper points out how, in spite of advancements in interoperability among different
platforms, the storage APIs tend to be proprietary. This basically locks in a cloud customer from
switching to another cloud service provider easily (Armbrust, 2010). Providing very high
reliability of service in the cloud requires extensive infrastructure deployment with plenty of
redundancy built-in. Major service providers like Amazon, Google, Microsoft and Salesforce
have the ability to assure very high availability of their services. All these services have
experienced some well publicized outages which cause concern for businesses in their desire to
switch to the cloud.
-----
The significance of cloud security is the focus of one of the four parts of the book Cloud
Computing by Antonopoulos and Gillam. In this edited book the authors have included several
chapters on cloud security (Antonopoulos, 2010). In particular, the work of Durbano, Rustvold,
Saylor and Studarus focus on the significance of standards in enabling cloud security. Their work
points out the gaps in ISO 27002 security controls (Durbano, 2010). Chen, Paxson and Katz
answer the question of ‘What is new about cloud computing security?’ Their analysis shows that
many of the cloud security issues are not really new except that they hinge upon multi-tenancy
trust considerations and auditability of service providers’ ability to back up their claims with data
on security aspects (Chen, 2010).
One of the challenges for any new technology is the availability of global standards. Cloud
computing is evolving rapidly but there are not many commonly accepted standards yet. ISO
27001, NIST and Cloud Security Alliance are all working toward providing guidelines for the
cloud industry. One of the Cloud Security Alliance guidelines involves the Top 9 Cloud
Computing Threats in 2013. Some of these threats relate to data breaches in the cloud, data loss
due to data leakage, insecure APIs and abuse of cloud services (Cloud Security Alliance, 2013).
We already pointed out one such abuse from the work of Ristenpart et al involving side channel
attacks. Next we look at the literature review article of Yang and Tate in which they classify 205
articles that appeared in cloud computing (Yang, 2012). They started this line of research in 2009
when they reviewed 54 articles. Since then the field has grown significantly and they included
several of the articles that we are examining in this brief review. Similar to Yang and Tate’s
work, Idziorek and Tannian surveyed all research articles in the area of public cloud computing
and focused on cloud computing security. This article points out several reasons on the
impediments still facing cloud computing adoption (Idziorek, 2012). Likewise, Modi et al
surveyed the issues affecting cloud computing adoption and their vulnerabilities. This paper
identifies some solutions to strengthen security and privacy in the cloud (Modi, 2013). Related to
this work is the technical book by Trivedi and Pasley on Cloud Computing Security. As
developers of cloud security solutions with a major technology company these authors identify
several security solutions based on cloud architecture, design and the way the customers deploy
their cloud based solutions (Trivedi, 2012). Continuing this line of research on cloud computing
security, Zissis and Lekkas propose the creation of a trusted third party focused on cloud
security. The authors point out that this arrangement would create a security mesh for all cloud
users that will lead to a trusted environment (Zissis, 2012).
Many businesses use cloud computing for data storage. This feature provides the business a cost
effective solution to store as much data as necessary and at the same time provide related data
backup, recovery and business continuity benefits. However, it also introduces the risk of not
having full control over the data storage as it is physically outside the control of the business.
This has led to several risks for businesses. To address this concern Wang et al propose a flexible
distributed method. In their approach they propose a method that achieves storage correctness
and supports dynamic operations such as data update and delete (Wang, 2009). John Viega from
a major security service firm analyzed the security aspects of the three major cloud services –
SaaS, PaaS and IaaS. His analysis shows that in the case of SaaS the main concern for the
customer relates to the service providers’ ability to protect the infrastructure from attack and
ensure non-leakage of data in the multi-tenant environment. In PaaS, even though the developers
who subscribe to this service will be able to develop their own security solutions, they are still
-----
dependent on the service providers’ way of protecting the service below their application level
for intrusion prevention. For IaaS, the major concern is the way the virtual machines are
configured. A related concern with IaaS service is the reliability of the service provider (Viega,
2009). Mark Ryan has a special focus in his paper on privacy concerns related to the cloud
because his paper addresses an area of interest for many academic researchers. The goal of
Ryan’s paper is on the privacy aspects related to the two major conference management systems
in use – EDAS and EasyChair. The paper highlights the many benefits of the conference
management systems on the cloud and also highlights some concerns such as the leakage of
reviewer information, cumulative success records of many researchers related to their
submissions for a variety of conferences over a long period of time and aggregated reviewing
profile of the reviewers. These data could be accidentally or maliciously disclosed by systems
administrators on these cloud systems where they are privy to large volumes of data. Even
though this is a very small segment of the cloud service industry, this paper’s focus is on the
potential privacy concerns for data stored on the cloud (Ryan, 2011).
The next set of papers that we examine relates to cloud computing risks and how they are
addressed. Gartner Research identifies seven cloud computing risks that are quite common.
These are presented in the context of a potential cloud customer evaluating a cloud service. Some
of these concerns relate to how the service provider handles privileged access to system
resources, their regulatory compliance activities related to physical security of the system and
third party audit such as SAS 70 Type II audit report, where they store the data and how they
segregate belonging to different customers so that they do not co-mingle (Brodkin, 2008). In
summarizing the cloud security concerns of many European partners, Daniele Catteddu in his
lengthy report points out that the two major benefits of cloud services, namely the economies of
scale and the operational flexibility are ‘both a friend and foe.’ The main thrust of this report is
that the cloud customer needs an assurance that the service providers are following sound
security practices to mitigate the risks faced by the customer and the provider (Catteddu, 2010).
Similar to the above report, the World Privacy Forum developed a report on the privacy
implications of data stored in the cloud. This report especially focuses on the many legal aspects
of compliance based on laws such as HIPAA (Health Insurance Portability and Accountability
Act), GLBA (Gramm-Leach-Bliley Act), ECPA (Electronic Communications Privacy Act), and
Fair Credit Reporting Act. The report notes that the information stored by an individual or a
business with a cloud service provider may have less protection than when the same information
is held by the information creator. Moreover, governments find it easy to obtain lot of
information from a centralized source such as a cloud provider (Gellman, 2009). The main
contribution of this report is in raising the awareness of the cloud customers relative to privacy
issues in the cloud.
One of the indicators of a mature model is the availability of enough case law to understand how
courts interpret the technological aspects. Using this metric cloud computing is not yet mature
enough to have a solid body of case law. To understand the ambiguities of how to interpret the
implementation models in the cloud we cite two instances. In the first case, Cartoon Network
sued CSC Holdings (parent company of Cablevision) for copyright infringement in 2009. In this
case, Cablevision provided customers the ability to store the recordings of their choice on the
cloud and access the same using their authentication credentials. Cartoon Network contended
that Cablevision was violating their copyright by sharing their content with others using the
-----
cloud storage. The court ruled that Cablevision was simply providing their authenticated
customers a storage facility in the cloud and not illegally sharing any copyrighted material. In the
second case, Arista Records sued Usenet.com in 2009 for violating their copyright on their
musical content by enabling unauthorized redistribution of their copyrighted content through the
bulletin board system in the cloud managed by Usenet. The model used by Usenet enabled the
cloud users to share their stored content with others unlike the Cablevision case. Since the cloud
was used in this case for data sharing the court ruled that it was a copyright violation by Usenet
(Wittow, 2011).
The analysis so far was focused on the US experience with the use of cloud computing.
However, cloud computing is a global phenomenon. In United Kingdom the government’s cloud
computing initiative is known as G-cloud. In the brief article focused on G-cloud and NIST
definitions of cloud computing, author Craig-Wood develops a comprehensive picture of all
aspects of cloud (Craig-Wood, 2010). We have developed Figure 2 based on this view. This
figure summarizes effectively our prior discussion on the three cloud types, four cloud
deployment models and some of the major advantages of cloud services. Just as cloud computing
is used in UK, the Australian experience relative to the legal requirements of the cloud provider
is described in their white paper by Vincent and Crooks. The details presented in this white paper
relate to privacy laws, location of data in the cloud, how foreign governments might get access to
this data, security beaches and service availability (Vincent, 2013). These are all important
security considerations for a cloud service consumer to consider prior to making a commitment
to use the cloud.
-----
## Figure 2: Cloud Types-Deployment Models-Features.
Cloud Featureeees
On-demand Self-Service Pay-as-You-Go Pooled Broad Network Public CloudPrivate CloudHybrid CloudCommunity CloudSaaPaaIaaService Deployment Models
We review next the German experience with respect to cloud computing based on adherence to
privacy laws. This topic is discussed by Doelitzscher, Reich and Sulistio in their Cloud Security
Project with particular emphasis on Small and Medium-sized Enterprises (SMEs). They
introduce a six layer security model that involves risk analysis and encryption (Doelitzscher,
2010). We conclude this security review of existing cloud computing literature with a brief
outline of Bhensook and Senivongse’s assessment of security requirements compliance by
service providers. This paper makes an extensive analysis of Cloud Security Alliance’s
recommendations by using the Goal Question Metric (GQM) for security requirements
compliance. The weighted scoring model that they develop is then tested using Amazon Web
Services’ (AWS) compliance. The results show that in most cases AWS is compliant with
various metrics being measured (Bhensook, 2012).
|Col1|Col2|Col3|Col4|
|---|---|---|---|
|On-demand Self-Service Pay-as-You-Go Pooled Broad Network Public CloudPrivate CloudHybrid CloudCommunity CloudSaaPaaIaaService Deployment Models||||
|||||
On-demand Self-Service Pay-as-You-Go Pooled Broad Network Public CloudPrivate CloudHybrid CloudCommunity CloudSaaPaaIaaService Deployment Models
-----
## NEW PARADIGM
Cloud computing is a significant shift in the way IT services are managed. Organizations large
and small have managed IT services over the years with varying levels of investments. Today,
with advancements in communication technology, many new options have opened up for
existing businesses and new entrepreneurs want to use more of the capabilities of IT. These two
aspects have spawned the significant growth of cloud computing, which gives the customer the
ability to benefit from the pay-as-you-go model. Cloud computing has enabled the service
providers to benefit from the economies of scale.
This change in service rendering is necessitated by the fact that today’s workforce is increasingly
mobile and consequently the need for access to remote resources is greater. Moreover, demand
fluctuations for IT services are a reality. Businesses cannot afford to provision IT services to
meet peak demand, which occur infrequently. Cloud computing provides an ideal solution to
meet these needs without incurring significant cost in services provisioning.
Investments necessary to have a reliable IT service kept many prospective entrepreneurs from
creating online ventures. On the web, businesses large and small look alike. Cloud computing is
providing entrepreneurs the opportunity to try their ideas out, with IT services no longer holding
them back as a barrier to entry. The major beneficiaries of cloud computing are small and
medium sized businesses as this new concept provides them an opportunity to try out high-end
services with no up-front cost, allowing them to use the pay-as-you-go model.
Large enterprises also stand to benefit from cloud computing, although of a different nature.
Large enterprises manage data centers and the IT paradigm shift referred to earlier mean more in
the context of accessing data from the data centers. In this context private clouds have been
introduced where the benefits of storage management and elasticity in demand for computing
services are the key drivers. Moreover, the cloud technology also offers high level of reliability
and availability of systems without significant capital layout. Often, the benefits of cloud
computing are realized by taking a hybrid approach. The hybrid approach gives the large
organizations the ability to manage their IT centers and at the same time expand their computing
capacity without large capital investment by utilizing the cloud resources. This is especially
useful to meet seasonal peak demands with hybrid clouds. Organizations with seasonal high
demands that could benefit from hybrid clouds are in the entertainment industry around holidays,
sports networks with on-demand service and tax service providers.
In assessing cloud computing’s appeal we should consider the usage levels of organizational
servers. Server utilization level gives a good metric to see if the investment cost in hardware is
worth it. The U.S. federal government started looking at the server utilization in its data centers
several years ago and found that the utilization level was low. According to a 2010 report by the
Computer Sciences Corporation (CSC), a global technology services company, that among all
data centers in use the server utilization rate is between 6 percent and 20 percent. The data from
this report is shown in Figure 2.1. Even Google’s server utilization rate is around 40 percent.
One reason for the low utilization is the lack of virtualization and the need to use dedicated
servers for multiple operating systems as well as separation of sensitive applications. Cloud
-----
computing is a natural fit to address the low utilization aspect because of high levels of
virtualization. With multiple users sharing the computing resources, cloud computing has a very
high level of server utilization (Hayes, 2008).
**Figure 2.1: Server Usage Statistics.**
Cloud computing architecture enables businesses to meet demand elasticity in computing
resources. Business organizations have great difficulty in dealing with demand elasticity for cost
considerations. A useful model to compare here is how networks manage elasticity in bandwidth
demand. For cost reasons network bandwidth provisioning uses the Committed Information Rate
(CIR) model. Likewise, cloud computing provides a similar feature in meeting demand elasticity
in both storage and computing power. Without the ability to meet demand elasticity, businesses
may end up with an underprovisioned service. In that case customers would abandon such
services. Jeff Bezos, CEO of Amazon, highlights the success of extreme demand for computing
power within a very short period of time. In this case a nascent web services company, Animoto,
grew so rapidly that its server needs grew from 50 to 3,500 servers over a three day period.
Amazon was able to accommodate such a high demand easily. This is a good illustration of high
demand elasticity.
In the traditional model, the end user had control over the creation, maintenance and deletion of a
document. In the cloud environment, the end user is spared the trouble of maintaining the
computing system and reaps the benefits of the application software. This is a positive aspect of
cloud storage. However, it is not entirely clear to the end user that when a document is deleted it
is going to be inaccessible from the storage system. There have been instances where the
-----
document lingered on in the storage system of the cloud provider. These types of issues are
unique to cloud computing and thus are a departure from the standard expectation of a computer
system. Thus, we note that a shift in approach is needed in order to have control over the online
information.
Many large organizations are considering cloud based services as a cost-effective way to plan for
disaster recovery. In this case the organization has its own computing resources that it controls
and plans to use the cloud services for disaster recovery purposes. In the traditional model for
disaster recovery the organization would use a warm or cold site as its backup facility, which is a
recurring expenditure for the company. The cloud model eliminates this recurring expenditure,
instead the organization pays for the services it uses when needed. The main cloud service type
being considered for disaster recovery is IaaS. Data backup is another service area for which
cloud computing is gaining ground. In the traditional model companies perform an incremental
backup daily and full backup weekly. The backed up data gets stored off site and handled by
companies such as Iron Mountain. In the cloud computing model a large organization would use
the cloud services for backup and recovery, which by its very design is at an offsite far away
from the company location. The organization could architect the backup process in such a way
that it is an automated full backup continuously. The promise of these two services in the cloud
has brought Microsoft and Iron Mountain together to offer data backup and recovery. Using the
Windows Azure platform Microsoft performs data backup and Iron Mountain manages the stored
data for the customer based on their expertise in this field. The customer pays for this service
based on the amount of storage used and the retention period for backup data. This service has
the added benefit of offsite storage built-in that is essential for disaster recovery and backup
because the cloud provider is remotely located relative to the customer. An essential component
of efficient data backup is data de-duplication, also known as ‘intelligent compression.’ The deduplication method allows for storing only one copy of the data and providing a pointer from all
future occurrences of the same data. Data de-duplication can be performed at the file or block
level. The latter is more efficient than the former. In typical email backups many users may have
the same file as an attachment and so the same file is backed up multiple times. Using the deduplication approach only one copy is saved and all other references point to the same copy. This
is a typical file level de-duplication. Most often de-duplication is more efficient at the block
level. In this approach each block of data is hashed using MD5 or SHA-1 and the index is stored.
Future hashes of blocks producing the same index are treated as duplicates and not stored. There
are sophisticated methods available to detect hash collision, which is rare (Armbrust, 2010).
## SECURITY CHALLENGES
Managing security is a major challenge to both large and small organizations. Transitioning to
cloud computing increases the challenges faced by these companies many-folds because of the
several unknowns. One important aspect of security is physical security. An organization that
owns the computing resources knows how to provide physical security. An organization that uses
cloud computing does not know where the physical resources are located and so providing that
form of security is transferred to the service provider. This raises the important question of
liability in the event of a security compromise. Many organizations do not take this aspect into
consideration in their security planning. Ultimately it would be the organization that would be
-----
held accountable by its customers for any security loss due to failure in physical security. Major
providers such as Amazon, Google, SalesForce and Microsoft provide the necessary physical
security guarantee.
Security and trust are closely related. An important contributor to trust is transparency. One way
to achieve customer trust in cloud environments is to share the security practices as they relate to
physical security, backup and recovery, compliance, incident handling and logs of security
attacks. With the help of non-disclosure agreements cloud providers can share details of logs on
incident handling and security attacks. Information security’s core requirements are
confidentiality, integrity and availability. We add the fourth piece – access control – to this
security scenario. In this section we will explore the various security challenges facing a
customer using cloud computing.
First and foremost, the cloud customer must understand the levels of security the provider is
guaranteeing. This includes system up time, system upkeep with respect to software updates and
patches and sharing a variety of logs that will enable the cloud customer to meet certain
compliance requirements such as HIPAA and SOX. Since the cloud provider manages the
servers on which the customer applications run, access control to the applications is an important
part of security as well. Access control in cloud environments differ based on the type of service
– SaaS, PaaS, IaaS.
Today’s cloud environment is dominated by SaaS and in this case the cloud provider should
provide the necessary authentication mechanisms to grant access. The leading SaaS service is
CRM, estimated to be about US$ 3.8 Billion worldwide in 2011. A Goldman-Sachs survey in
2010 points out that small and medium sized businesses would consider cloud service 70 percent
of the time and would prefer a cloud solution, if available, 58 percent of the time. This high level
of confidence in SaaS cloud services shows the maturity of this particular market.
The success of cloud computing as a viable alternative to individual businesses managing their
own IT centers is commendable. From an operational perspective this is an efficient model.
However, from a security perspective there are several major concerns to overcome. As
mentioned above, physical security of systems is one concern. A more serious concern is in
access control – physical access to hardware, privileged access to operating environment, access
to application software. In Figure 3.1 we combine the access control feature with the type of
protection that a service provider could guarantee the customer for security of their data in this
design. The goal of this proposed architecture for cloud relative to the cloud customer is to
provide a way for them to see how their data and interaction with the cloud are both secure. The
Authentication layer inside the cloud provides a way for the cloud customer to control who gets
access to their virtual machine and data. Knowing who the privileged users are from the service
provider perspective will enable the customer to monitor the access log, which they should be
able to get in an automated manner from the service provider. The firewall inside the cloud gives
the cloud customer the level of assurance that they are accustomed to in their own systems. The
customer would be able to have the firewall configured by the service provider the way they
want, thus giving an added level of protection to their data. By using encrypted storage in the
cloud and holding the encryption key themselves, the customer will have added security
protection in case of data leakage because of storage comingling.
-----
**Figure 3.1: Cloud Computing Security Infrastructure.**
Customer
Authentication Layer
Firewall
Server Encrypted
Storage
From a security perspective cloud computing should be viewed as a potential single point of
failure. Cloud providers have high redundancy in their architecture, yet relying solely on one
cloud provider may raise reliability issues. Moreover, when many customers are still ranking
cloud security as their main concern in their switch to cloud computing, the providers have to
offer many trust building measures. One such is system reliability. This has suffered some
setback because of some well publicized instances of cloud outage. Some of these major outages
are the April 2011 Amazon AWS failure (Amazon, 2011), October 2009 DDoS attack on
Bitbucket that utilized Amazon EC2 and the November 2007 Rackspace failure due to power
supply disruption. Since reliability could still be an issue for some customers, one possible
solution to using cloud services is to split the types of cloud services among multiple service
providers. A common approach to this could be to use one cloud provider for operations and
another provider for storage solutions.
Besides the access control concerns with cloud storage and the violations cited in this regard, it is
a major decision for a customer to have a data backup outside of the cloud service provider. The
reason for this is that when a cloud service provider has to comply with law enforcement demand
for removal of a storage device for investigative purposes, many other customers might also be
affected. This is not a hypothetical scenario. In March 2009 a data center in Dallas owned by a
cloud service provider was raided by law enforcement and as such many other customers were
also affected because their data was unavailable for access for an extended period of time.
Server
Authentication Layer
Customer
-----
Security, interoperability and portability are considered the major barriers for cloud computing
adoption. In many instances the cloud service provider does not think it is their responsibility to
protect customer data. In the 2011 Ponemon Institute/CA Technologies cloud security study a
surprising finding is reported (Ponemon, 2011). It points out that a majority of the cloud service
providers do not think that cloud security is their responsibility and even more disturbing is that
these providers do not think that their services protect and secure customers’ sensitive data.
Cloud service customers should be aware of how secure their sensitive data are and how to
protect the same.
Cloud computing model requires virtualization in a big way. In virtual environments data
comingling is a reality. We identified in Figure 3.1 how a customer could work with the service
provider in establishing a level of security to prevent comingling of data on a central server.
Since cloud service providers operate in a multi-tenant environment in their public cloud, an
unintended consequence of comingling data from multiple customers on the same physical
device is data leakage. The author recommends the architecture specified in Figure 3.2 as a
preferred solution to cloud provider’s infrastructure vis-à-vis the customer need for access logs.
The figure shows the need for virtualization of servers and isolation of the virtual servers. Also,
the customer should be aware of the privileged users at the cloud provider who will have access
to their virtual servers as well as the ability to get data from user access logs for their respective
virtual servers in an automated manner. Ristenpart et al have highlighted the data leakage
problems associated with comingling data on the same physical device (Ristenpart, 2009). Data
leakage could lead to privacy violations. When cloud providers do not see that it is their
responsibility to protect customer data, as highlighted by the Ponemon Institute study, it puts an
extra burden on cloud customers to use encryption for all their data. We expect more and more
customers are going to be sensitive to the potential of data leakage from the cloud, which is not a
problem if they had their own computing resources. The security does not come from encrypted
storage alone. The customer must hold the encryption key in their system at their place of
business and not place it on the cloud, even if it is in their virtual server. Ristenpart et al have
shown how an attacker could place their virtual server on the same physical server as the
customer and perform side-channel attacks to recover the encryption key.
**Figure 3.2: Cloud Provider Infrastructure.**
Provider Privileged Users
Customers VS = Virtual Server
VS1 VS2 VS3 …… AL = Access Log
VSn
AL1 AL2 AL3 ……
ALn
|VS1 VSn|Col2|VS2|Col4|VS3|Col6|……|Col8|Col9|
|---|---|---|---|---|---|---|---|---|
|AL1|Col2|AL2|Col4|AL3|Col6|……|Col8|Col9|
|---|---|---|---|---|---|---|---|---|
VS1 VS2 VS3 ……
VSn
AL1 AL2 AL3 ……
Customers
-----
In order to provide data security, both customers and cloud service providers must realize
that security has to become information-centric, i.e., security moves with the data. This way any
operational decision by the cloud service provider will protect the customer data. Another
important aspect of handling the cloud security challenge is to trust the processes and
infrastructure. To this end,
1. Businesses should know from cloud provider how the following are handled:
– Physical security
– Number of privileged users with access to hardware
– Log data for system access, up time, incident handling, attacks on the system
– Data backup and recovery
– Compliance with regulations
2. Security and trust are interrelated
3. Important contributor to trust is transparency
The significance of compliance such as SAS 70 Type II Audit builds trust for the customer.
Providing data security involves not only having adequate measures but also compliance with
relevant laws. In order to verify compliance an organization would require log data related to
access control and incident handling. The cloud customer should be able to get such data from
the provider. This is not standard practice with the cloud providers. System up time is another
issue that should be looked into by the cloud customer. The system up time could be verified
only with log data. In Table 3.1 we highlight the compliance aspects from Amazon and Google,
two of the major cloud service providers.
Table 3.1. Summary of Security Features by Amazon and Google
|Provider|Security Features|Compliance Support|
|---|---|---|
|Amazon| ISO 27001 certified SAS 70 Type II Audit certified PCI DSS Level 1 certified FISMA certified Physically secure data centers distributed around the world User control for data encryption Data backup Customer-aware storage region Does not allow third-party cloud providers Processes in place to prevent unauthorized insider access to customer data SLA guarantee of 99.95% up time Supports multifactor authentication|SOX compliant GLBA compliant HIPAA compliant|
-----
|Col1| Supports host-based firewall|Col3|
|---|---|---|
|Google|SAS 70 Type II Audit certified Multi-layered physical security Hardware customization, maintenance Secure storage handling Privacy protection for customer data Data centers located around the world Redundancy built into storage management Ease of movement for customer data into and out of Google Apps Provides additional levels of access control designed by customers|Google does not explicitly certify any of their products as meeting standard compliance requirements such as HIPAA, SOX, GLBA, FISMA, etc. Instead, it provides features that customers can enable in order to be compliant on their own using Google Apps.|
Trust aspects include knowing the reliability of cloud service. To this end the cloud service
provider should be able to provide the customer their infrastructure capability with respect to
system uptime. It is an integral part of knowing the provider’s ability to assure security. Table
3.2 gives the amount of downtime allowed when cloud providers claim a certain level of uptime,
usually denoted by a set of 9s. For example, four 9s means 99.99 percent up time. Highly
available computing is expensive. Addition of one 9 to the guaranteed uptime nearly doubles the
cost. Given these limitations the cloud provider uptime claims should be backed by data.
**Table 3.2 : System downtime chart based on up time claim by cloud provider.**
System Availability Maximum downtime Per
year
99.999% 00:05:15
99.99% 00:52:35
99.9% 08:45:56
99% 87:39:29
Source: Stratus.com
Standards play an important role in security. In the case of Cloud Computing, the standards are
still evolving. The groups that are making contributions to cloud standards are the Distributed
Management Task Force, Object Management Group, National Institute of Standards and
Technology, European Telecommunications Standards Institute and the Storage Networking
Industry Association. These standards help address the issues raised earlier so that customers can
be familiar with the level of security that they are getting from the cloud provider.
## FUTURE OF CLOUD COMPUTING
Cloud computing has many benefits to offer. Foremost among them is in meeting the elasticity of
demand and the ability to use the ‘pay-as-you-go’ model. Cloud computing is often referred to as
“converting capital expenses to operating expenses.” Cloud computing provides risk transference
for the customer in that the customer need not worry about the cost involved in overprovisioning
or under provisioning resources. It is important to note that a business that consistently under
provisions resources will lose customers permanently.
|System Availability|Maximum downtime Per year|
|---|---|
|99.999%|00:05:15|
|99.99%|00:52:35|
|99.9%|08:45:56|
|99%|87:39:29|
-----
Cloud computing has become the preferred storage solution to deal with Big Data. We are seeing
an exponential growth in data due to several social media applications. A welcome contribution
to cloud computing has been added by Google through its MapReduce architecture. MapReduce
provides a simple way to process large volumes of data quickly. The open source
implementation of MapReduce method is Hadoop by Apache. Many organizations use Hadoop
Distributed File System (HDFS) to process large volumes of data in a reliable way. Without the
availability of cloud services many businesses will not be able to afford the high cost of
infrastructure needed to use Hadoop. Handling security in the cloud is a complex process and
many researchers are working on incremental aspects of providing security to a variety of aspects
in cloud processing. Hamlen et al (Hamlen, 2010) discuss one such security solution for the
cloud. In particular, they discuss ways to efficiently store data in remote locations and query
encrypted data. This approach to storing data in the cloud after encrypting it provides a level of
security for sensitive data. Some of the cost advantages that the cloud provides will be lost if we
were to add the cost of encryption. If we store data in the cloud in encrypted form then we need
to develop processes to query encrypted data. Hamlen et al use HDFS for virtualization and
apply Hadoop processes for secure storage.
Another area where cloud computing could play a major role in the future is in backup storage.
Many businesses do not consider backup as highly critical and with cloud services providing a
viable low cost alternative, it is expected to catch on in the future. Companies like Carbonite and
CrashPlan have simplified the process of backup. Recovery goes hand-in-hand with backup and
customers are responsible for the policies regarding recovery testing periodically. Given the
cloud infrastructure capability this should be a simple solution as the customer would be able to
have access to the necessary hardware and software available to test the recovery aspects on a
pay-as-you-go model.
As cloud computing emerges as an affordable service for many businesses it is inevitable that
there will be several legal issues that arise as to ownership of data, liability for protection of data
and safeguarding of privacy. There is not a large body of case law in this regard. We cite two
recent court cases where challenges were mounted because of the use of cloud service. In the
first case, Cartoon Network challenged that CableVision’s Remote Storage DVR service violated
the Copyright laws. The CableVision service offered the ability for the customer to record shows
and store them in the cloud in their personal storage area for later playback. Since the access to
this storage was based on customer’s access control mechanism of userID/Password, the court
ruled that CableVision was not violating the Copyright laws (Cartoon Network, 2008).
According to a study by Harvard University’s Lerner and others, the impact of this ruling was
significant as it resulted in investments of over a billion dollars in this area of cloud service.
However, there were conflicting rulings in Europe (unfavorable in France, favorable in
Germany) on similar aspects of storage in the cloud by individual users (Borek, 2012).
Cloud service technology has the ability to allow each user of their service to have their own
storage space as described in the CableVision case above. The cloud technology is also capable
of storing a single copy of an artifact and make it available only to authorized users. MP3Tunes,
a new cloud-based music service, created a locker in which a customer could place a song and
access it later on demand with appropriate authentication. As part of the cloud technology’s
-----
capability, MP3Tunes held only one copy of a music in the locker even when new users attempt
to store the same song. Capitol Records, one of the largest music companies, sued MP3Tunes as
offering a music distribution without proper license of copyrighted material. The Manhattan
District Court in U.S. ruled that MP3Tunes did not violate the owner’s copyright for the song
(Capital Records, 2011) as it was leveraging one aspect of the technology known as
deduplication. The users accessing the song all had stored that song in the cloud. These two cases
point to the challenges faced by the cloud service industry as the service matures.
## SUMMARY
Cloud computing is efficient and cost effective. It takes the burden out of many businesses to
maintain a technical system and yet reap the benefits of its availability. In our analysis so far we
were able to identify the many security challenges posed by the cloud service. The solutions
identified in various places in the above discussion show that we can answer in the affirmative
the question raised by the title of this paper concerning cloud security. Cloud computing has
opened up the opportunity to many entrepreneurs to focus on their core business and count on
cloud services to provide the necessary computing power. Many businesses experience demand
elasticity and cloud computing is a natural fit in providing cost effective service in this area.
Security aspects did not get much attention for a few years now and many customers were
focusing on availability of service. As more and more businesses start using this service the
question of security is coming to the forefront. We have addressed some of the security concerns
with cloud computing. One area that is bound to receive greater attention is in compliance with
government laws in each jurisdiction in which the cloud operates and the global industry
requirements such as the one by the Payment Card Industry as many businesses depend on cloud
computing.
**REFERENCES**
[Amazon. (2011). AWS Security Center. http://aws.amazon.com/security/ Accessed 11/15/13](http://aws.amazon.com/security/)
Antonopoulos, N., & Gillam, L. (2010). Cloud Computing: Principles, Systems and
Applications. Springer: London, UK.
Armbrust, M., Fox, A., Grith, R., Joseph, A. D., Katz, R., Konwinski, A., Lee, G., Patterson, D.,
Rabkin, A., Stoica, I. & Zaharia, M. (2010). A view of cloud computing.
_Communications of ACM, 53(4), 50-58._
Bhattacharjee, R. 2009. An Analysis of the Cloud Computing Platforms, _MIT MS Thesis,_
Cambridge, MA.
Bhensook, N., & Senivongse, T. (2012). An Assessment of Security Requirements Compliance
of Cloud Providers, _Proceedings of the IEEE 4[th] International Conference on Cloud_
_Computing Technology and Science, 520-525._
Borek, C., Christensen, L., Hess, P., Lerner, J. & Rafert, G. (2012). Lost in the Clouds: The
Impact of Copyright Scope on Investment in Cloud Computing Ventures,
[http://www.intertic.org/Conference/Lerner.pdf Accessed 11/22/13](http://www.intertic.org/Conference/Lerner.pdf)
-----
Brodkin, J. (2008). Gartner: Seven Cloud Computing Risks,
[http://www.idi.ntnu.no/emner/tdt60/papers/Cloud_Computing_Security_Risk.pdf](http://www.idi.ntnu.no/emner/tdt60/papers/Cloud_Computing_Security_Risk.pdf)
Accessed 11/15/13.
Capital Records vs MP3Tunes. (2011). Manhattan U.S. District Court, 1:07-cv-09931.
Cartoon Network vs CSC Holdings. (2008). US 2[nd] Circuit Court Ruling, 536 F .3d 121.
Catteddu, D. (2010). Cloud Computing: Benefits, Risks and Recommendations for Information
Security, _Web Applications Security, edited by Serrao, C., Diaz, V. A. & Cerullo, F.,_
Springer: Berlin, Germany.
Chen, Y., Paxson, V., & Katz, R. H. (2010). What is new about cloud computing security. UC
_Berkeley Technical Report EECS-2010-5._
[Cloud Security Alliance. (2013). http://www.cloudsecurityalliance.org Accessed 11/15/13](http://www.cloudsecurityalliance.org/)
Craig-Wood, K. (2010). Definition of Cloud Computing incorporating NIST and G-Cloud views,
[http://www.katescomment.com/definition-of-cloud-computing-nist-g-cloud/](http://www.katescomment.com/definition-of-cloud-computing-nist-g-cloud/)
Doelitzscher, F., Reich, C., & Sulistio, A. (2010). _Proceedings of the 10[th] IEEE International_
_Conference on Computer and Information Technology, 930-935._
[Durbano, J. P., Rustvold, D., Saylor, G., & Studarus, J.](http://link.springer.com/search?facet-author=%22Derek+Rustvold%22) (2010). Securing the cloud, _Cloud_
_Computing, edited by Antonoupolos, N. & Gillam, L., 289-302, Springer: London, UK._
Gellman, R. (2009). Privacy in the clouds: risks to privacy and confidentiality from cloud
computing. World Privacy Forum.
Hamlen, K., Kantarcioglu, M., Khan, L., & Bhavani, T. (2010). Security Issues for Cloud
Computing, International Journal of Information Security and Privacy, 4(2), 39-51.
Hayes, B. (2008). Cloud Computing, Communications of ACM, 51(7), 9–11.
Idziorek, J., & Tannian, M. (2012). Security analysis of public cloud computing. International
_Journal of Communication Networks and Distributed Systems, 9(1&2), 4-20._
Mell, P., & Grance, T. (2011). The NIST Definition of Cloud Computing, _NIST, Gaithersburg,_
MD.
Modi, C., Patel, D., Borisaniya, B., Patel, A., & Rajarajan, M. (2013). A survey on security
issues and solutions at different layers of cloud computing. Journal of Supercomputing,
63(2), 563-592.
-----
NIST SP 800-145. (2011). The NIST Definition of cloud computing.
[http://csrc.nist.gov/publications/nistpubs/800-145/SP800-145.pdf](http://csrc.nist.gov/publications/nistpubs/800-145/SP800-145.pdf) Accessed 11/15/13
Ponemon. (2011). Security of cloud computing providers study, Ponemon Institute, May.
Ristenpart, T., Tromer, E., Shacham, H., & Savage, S. (2009). Hey, You, Get Off of My Cloud:
Exploring Information Leakage in Third-Party Compute Clouds, _Proceedings of 16[th]_
_ACM Conference on Computer and Communications Security, Chicago, IL._
Ryan, M. D. (2011). Cloud Computing Privacy Concerns on our Doorstep, Communications of
_ACM, 54(1), 36-38._
Trivedi, K., & Pasley, K. 2012. Cloud Computing Security, Cisco Press: Indianapolis, IN
Viega, J. (2009). Cloud computing and the common man. IEEE Computer, 42(8), 106-108.
Vincent, M., & Crooks, K. (2013). Cloud Computing: What legal commitments can you expect
from your provider? _White_ _Paper,_ SheltonIP, Sydney, Australia, 1-21.
[http://i.haymarket.net.au/Assets/cloudcover2013.pdf Accessed 11/15/13](http://i.haymarket.net.au/Assets/cloudcover2013.pdf)
Wang, C., Wang, Q., Ren, K., & Lou, W. (2009, July)). Ensuring data storage security in cloud
computing, 17[th] International. Workshop on Quality of Service, 1-9.
Wittow, M. H. (2011). Cloud Computing: Recent Cases and Anticipating New Types of Claims,
_The Computer and Internet Lawyer, 28(1), 1-8._
Yang, H., & Tate, M. (2012). A descriptive literature review and classification of cloud
computing research, Communications of AIS, 31(2), 35-60.
Zissis, D., & Lekkas, D. (2012). Addressing cloud computing security issues. Future Generation
_Computer Systems, 28(3), 583-592._
-----
|
{
"disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.58729/1941-6679.1020?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.58729/1941-6679.1020, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "BRONZE",
"url": "https://scholarworks.lib.csusb.edu/cgi/viewcontent.cgi?article=1020&context=jitim"
}
| 2,013
|
[] | true
| 2013-04-01T00:00:00
|
[] | 13,652
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Engineering",
"source": "external"
},
{
"category": "Mathematics",
"source": "external"
},
{
"category": "Medicine",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Engineering",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/028f829fc1d47a64dd4ca5db564f8b50deba9ff5
|
[
"Computer Science",
"Engineering",
"Mathematics",
"Medicine"
] | 0.862817
|
An Optimality Summary: Secret Key Agreement with Physical Unclonable Functions
|
028f829fc1d47a64dd4ca5db564f8b50deba9ff5
|
IACR Cryptology ePrint Archive
|
[
{
"authorId": "10417153",
"name": "O. Günlü"
},
{
"authorId": "2307454",
"name": "Rafael F. Schaefer"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"IACR Cryptol eprint Arch"
],
"alternate_urls": null,
"id": "166fd2b5-a928-4a98-a449-3b90935cc101",
"issn": null,
"name": "IACR Cryptology ePrint Archive",
"type": "journal",
"url": "http://eprint.iacr.org/"
}
|
We address security and privacy problems for digital devices and biometrics from an information-theoretic optimality perspective to conduct authentication, message encryption/decryption, identification or secure and private computations by using a secret key. A physical unclonable function (PUF) provides local security to digital devices and this review gives the most relevant summary for information theorists, coding theorists, and signal processing community members who are interested in optimal PUF constructions. Low-complexity signal processing methods are applied to simplify information-theoretic analyses. The best trade-offs between the privacy-leakage, secret-key, and storage rates are discussed. Proposed optimal constructions that jointly design the vector quantizer and error-correction code parameters are listed. These constructions include modern and algebraic codes such as polar codes and convolutional codes, both of which can achieve small block-error probabilities at short block lengths, corresponding to a small number of PUF circuits. Open problems in the PUF literature from signal processing, information theory, coding theory, and hardware complexity perspectives and their combinations are listed to stimulate further advancements in the research on local privacy and security.
|
# entropy
_Review_
## An Optimality Summary: Secret Key Agreement with Physical Unclonable Functions
**Onur Günlü *** **and Rafael F. Schaefer**
Information Theory and Applications Chair, Technische Universität Berlin, 10623 Berlin, Germany;
rafael.schaefer@tu-berlin.de
*** Correspondence: guenlue@tu-berlin.de; Tel.: +49-30-314-26632**
**Abstract: We address security and privacy problems for digital devices and biometrics from an**
information-theoretic optimality perspective to conduct authentication, message encryption/decryption,
identification or secure and private computations by using a secret key. A physical unclonable function (PUF) provides local security to digital devices and this review gives the most relevant summary
for information theorists, coding theorists, and signal processing community members who are
interested in optimal PUF constructions. Low-complexity signal processing methods are applied
to simplify information-theoretic analyses. The best trade-offs between the privacy-leakage, secretkey, and storage rates are discussed. Proposed optimal constructions that jointly design the vector
quantizer and error-correction code parameters are listed. These constructions include modern and
algebraic codes such as polar codes and convolutional codes, both of which can achieve small blockerror probabilities at short block lengths, corresponding to a small number of PUF circuits. Open
problems in the PUF literature from signal processing, information theory, coding theory, and hardware complexity perspectives and their combinations are listed to stimulate further advancements in
the research on local privacy and security.
[����������](https://www.mdpi.com/1099-4300/23/1/16?type=check_update&version=3)
**�������**
**Citation: Günlü, O.; Schaefer, R.F. An**
Optimality Summary: Key
Agreement with Physical Unclonable
Functions. Entropy 2021, 23, 16.
[https://doi.org/10.3390/e23010016](https://doi.org/10.3390/e23010016)
Received: 4 November 2020
Accepted: 16 December 2020
Published: 24 December 2020
**Publisher’s Note: MDPI stays neu-**
tral with regard to jurisdictional clai
ms in published maps and institutio
nal affiliations.
**Copyright: © 2020 by the authors. Li-**
censee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and con
ditions of the Creative Commons At
[tribution (CC BY) license (https://](https://creativecommons.org/licenses/by/4.0/)
[creativecommons.org/licenses/by/](https://creativecommons.org/licenses/by/4.0/)
[4.0/).](https://creativecommons.org/licenses/by/4.0/)
**Keywords: physical unclonable functions (PUFs); private authentication; secret key generation;**
information theoretic privacy; code constructions for security
**1. Motivations**
Fundamental advances in cryptography were made in secret during the 20th century.
One exception was Claude E. Shannon’s paper “Communication Theory of Secrecy Systems” [1]. Until 1967, the literature on security was not extensive, but a book [2] with a
historical review of cryptography changed this trend [3]. Since then, the amount of sensitive
data to be protected against attackers has increased significantly. Continuous improvements
in security are needed and every improvement creates new possibilities for attacks [4].
Recent hardware-intrinsic security systems, biometric secrecy systems, 5th generation
of cellular mobile communication networks (5G) and beyond, as well as the internet of
things (IoT) networks, have numerous noticeable characteristics that differentiate them
from existing mechanisms. These include large numbers of low-complexity terminals with
light or no infrastructure, stringent constraints on latency, and primary applications of
inference, data gathering, and control. Such characteristics make it difficult to achieve
a sufficient level of secrecy and privacy. Traditional cryptographic protocols, requiring
certificate management or key distribution, might not be able to handle various applications
supported by such technologies and might not be able to assure the privacy of personal
information in the data collected. Similarly, low complexity terminals might not have the
necessary processing power to handle such protocols, or latency constraints might not
permit the processing time required for cryptographic operations. Similarly, traditional
methods that store a secret key in a secure nonvolatile memory (NVM) can be illustrated
to be not secure because of possible invasive attacks to the hardware. Thus, secrecy and
-----
_Entropy 2021, 23, 16_ 2 of 23
privacy for information systems are issues that need to be rethought in the context of recent
networks, digital circuits, and database storage.
Information-theoretic security is an emerging approach to provide secrecy and privacy,
for example, for wireless communication systems and networks by exploiting the unique
characteristics of the wireless communication channel. Information-theoretic security methods such as physical layer security (PLS) use signal processing, advanced coding, and
communication techniques to secure wireless communications at the physical layer. There
are two key advantages of PLS. Firstly, it enables the use of resources available at the
physical layer such as multiple measurements, channel training mechanisms, power, and
rate control, which cannot be utilized by the upper layers of the protocol stack. Secondly,
it is based on an information-theoretic foundation for secrecy and privacy that does not
make assumptions on the computational capabilities of adversaries, unlike cryptographic
primitives. By considering the security and privacy requirements of recent digital systems
and the potential benefits from information-theoretic security and privacy methods, it can
be seen that information-theoretic methods can complement or even replace conventional
cryptographic protocols for wireless networks, databases, and user authentication and
identification. Since information-theoretic methods do not generally require pre-shared secret keys, they might considerably simplify the key management in complicated networks.
Thus, these methods might be able to fulfill the stringent hardware area constrains of
digital devices and delay constraints in 5G/6G applications, or to avoid unnecessary computations, increasing the battery life of low power devices. Information-theoretic methods
offer “built-in” secrecy and privacy, generally independent of the network infrastructure,
providing better scalability with respect to an increase in the network or data size.
A promising local solution to information-theoretic security and privacy problems
is a physical unclonable function (PUF) [5]. PUFs generate “fingerprints” for physical
devices by using their intrinsic and unclonable properties. For instance, consider ring
oscillators (ROs) with a logic circuit of multiple inverters serially connected with a feedback
of the output of the last inverter into the input of the first inverter, as depicted in Figure 1.
RO outputs are oscillation frequencies 1/x, where _x is the oscillation period, that are_
� �
unique and uncontrollable since the difference between different RO outputs is caused
by submicron random manufacturing variations that cannot be controlled. One can use
RO outputs as a source of randomness, called a PUF circuit, to extract secret keys that are
unique to the digital device that embodies these ROs. The complete method that puts out a
unique secret key by using RO outputs is called an RO PUF. Similarly, binary static random
access memory (SRAM) outputs are utilized as a source of randomness to implement
SRAM PUFs in almost all digital devices because most digital devices have embedded
SRAMs used for data storage. The logic circuit of an SRAM is depicted in Figure 2 and the
logically stable states of an SRAM cell are (Q, Q) = (1, 0) and (0, 1). During the power-up,
the state is undefined if the manufacturer did not fix it. The undefined power-up state
of an SRAM cell converges to one of the stable states due to random and uncontrollable
mismatch of the inverter parameters, fixed when the SRAM cell is manufactured [6]. There
is also random noise in the cell that affects the cell at every power-up. Since the physical
mismatch of the cross-coupled inverters is due to manufacturing variations, an SRAM cell
output during power-up is a PUF output that is a response with one challenge, where the
challenge is the address of the SRAM cell [6].
#### 1 x[^]
### ENABLE
**Figure 1. Ring oscillator (RO) logic circuit.**
-----
_Entropy 2021, 23, 16_ 3 of 23
###### Q Q
**Figure 2. Static random access memory (SRAM) logic circuit.**
PUFs resemble biometric features of human beings. In this review, we will list stateof-the-art methods that bridge the gap between the practical secrecy systems that use PUFs
and the information-theoretic security limits by
- Modeling real PUF outputs to solve security problems with valid assumptions;
- Analyzing methods that make information-theoretic analysis tractable, for example,
by transforming PUF symbols so that the transform-domain outputs are almost independent and identically distributed (i.i.d.), and that result in smaller hardware area
than benchmark designs in the literature;
- Stating the information-theoretic limits for realistic PUF output models and providing
optimal and practical (i.e., low-complexity and finite-length) code constructions that
achieve these limits;
- Illustrating best-in-class nested codes for realistic PUF output models.
In short, we start with real PUF outputs to obtain mathematically-tractable models of
their behavior and then list optimal code constructions for these models. Since we discuss
methods developed from the fundamentals of signal processing and information theory,
any further improvements in this topic are likely to follow the listed steps in this review.
_Organization and Main Insights_
In Section 2, we provide a definition of a PUF, list its existing and potential applications, and analyze the most promising PUF types. The PUF output models and design
challenges faced when manufacturing reliable, low-complexity, and secure PUFs are listed
in Section 3. The main security challenge in designing PUFs, i.e., output correlations,
is tackled in Section 4 mainly by using a transform coding method, which can provably
protect PUFs against various machine learning attacks. The reliability and secrecy performance (e.g., the number of authenticated users) metrics used for PUF designs are defined
and jointly optimized in Section 5. PUF security and complexity performance evaluations
for the defined transform coding method are given in Section 6. Performance results for
error-correction codes used in combination with previous code constructions that are used
for key extraction with PUFs, are shown in Section 7 in order to illustrate that previous
key extraction methods are strictly suboptimal. We next define the information theoretic
metrics and the ultimate key-leakage-storage rate regions for the key agreement with PUFs
problem, as well as comparing available code constructions for the key agreement problem
in Section 8. Optimal code constructions for the key extraction with PUFs are implemented
in Section 9 by using nested polar codes, which are used in 5G networks in the control
channel, to illustrate significant gains from using optimal code constructions. In Section 10,
we provide a list of open PUF problems that might be interesting for information theorists,
coding theorists, and signal processing researchers in addition to the PUF community.
**2. PUF Basics**
We give a brief review of the literature on PUFs and discuss the problems with previous
PUF designs that can be tackled by using signal processing and coding-theoretic methods.
-----
_Entropy 2021, 23, 16_ 4 of 23
A PUF is defined as an unclonable function embodied in a device. In the literature,
there are alternative expansions of the term PUF such as “physically unclonable function”,
suggesting that it is a function that is only physically-unclonable. Such PUFs may provide
a weaker security guarantee since they allow their functions to be digitally-cloned. For any
practical application of a PUF, we need the property of unclonability both physically and
digitally. We therefore consider a function as a PUF only when the function is a physical
function, i.e. it is in a device, and it is not possible to clone it physically and digitally.
Physical identifiers such as PUFs are heuristically defined to be complex challengeresponse mappings that depend on the random variations in a physical object. Secret
sequences are derived from this complex mapping, which can be used as a secret key. One
important feature of PUFs is that the secret sequence generated is not required to be stored
and it can be regenerated on demand. This property makes PUFs cheaper (no requirement
for a memory for secret storage) and safer (the secret sequence is regenerated only on
demand) alternatives to other secret generation and storage techniques such as storing the
secret in an NVM [5].
There is an immense number of PUF types, which makes it practically impossible to
give a single definition of PUFs that covers all types. We provide the following definition
of PUFs that includes all PUF types of interest for this review.
**Definition 1 ([5]). We define a PUF as a challenge-response mapping embodied by a device such**
_that it is fast and easy for the device to put out the PUF response and hard for an attacker, who does_
_not have access to the PUF circuits, to determine the PUF output to a randomly chosen input, given_
_that a set of challenge-response (or input-output) pairs is accessible to him._
The terms used in Definition 1, i.e., fast, easy, and hard, are relative terms that should
be quantified for each PUF application separately. There are physical functions, called
physical one-way functions (POWFs), in the literature that are closely related to PUFs. Such
functions are obtained by applying the cryptographic method of “one-way functions”,
which refers to easy to evaluate and (on average) difficult to invert functions [7], to physical
systems. As the first example of POWFs, the pattern of the speckle obtained from waves
that propagate through a disordered medium is a one-way function of both the physical
randomness in the medium and the angle of the beam used to generate the optical waves [8].
Similar to POWFs, biometric identifiers such as the iris, retina, and fingerprints
are closely related to PUFs. Most of the assumptions made for biometric identifiers are
satisfied also by PUFs, so we can apply almost all of the results in the literature for
biometric identifiers to PUFs. However, it is common practice to assume that PUFs can
resist invasive (physical) attacks, which are considered to be the most powerful attacks
used to obtain information about a secret in a system, unlike biometric identifiers that are
constantly available for attacks. The reason for this assumption is that invasive attacks
permanently destroy the fragile PUF outputs [5]. This assumption will be the basis for the
PUF system models used throughout this review. We; therefore, assume that the attacker
does not observe a sequence that is correlated with the PUF outputs, unlike biometric
identifiers, since physical attacks applied to obtain such a sequence permanently change
the PUF outputs.
_2.1. Applications of PUFs_
A PUF can be seen as a source of random sequences hidden from an attacker who does
not have access to the PUF outputs. Therefore, any application that takes a secret sequence
as input can theoretically use PUFs. We list some scenarios where PUFs fit well practically:
- Security of information in wireless networks with an eavesdropper, i.e., a passive
attacker, is a PLS problem. Consider Wyner’s wiretap channel model introduced
in [9]. This model is the most common PLS model, which is a channel coding problem
unlike the secret key agreement problem we consider below that is a source coding
problem. A randomized encoder helps the transmitter in keeping the message secret
-----
_Entropy 2021, 23, 16_ 5 of 23
by confusing the eavesdropper. Therefore, at the WTC transmitter, PUFs can be used
as the local randomness source when a message should be sent securely through the
wiretap channel.
- Consider a 5G/6G mobile device that uses a set of SRAM outputs, which are available
in mobile devices, as PUF circuits to extract secret keys so that the messages to be
sent are encrypted with these secret keys before sending the data over the wireless
channel. Thus, the receiver (e.g., a base station) that previously obtained the secret
keys (sent by mobile devices, e.g., via public key cryptography) can decrypt the
data, while an eavesdropper who only overhears the data broadcast over the wireless
channel cannot easily learn the message sent.
- The controller area network (CAN) bus standard used in modern vehicles is illustrated
in [10] to be susceptible to denial-of-service attacks, which shows that safety-critical
inputs of the internal vehicle network such as brakes and throttle can be controlled by
an attacker. One countermeasure is to encrypt the transmitted CAN frames by using
block ciphers with secret keys generated from PUF outputs used as inputs.
- IoT devices such as wearable or e-health devices may carry sensitive data and use a
PUF to store secret keys in such a way that only a device to which the secret keys are
accessible can command the IoT devices. One common example of such applications
is when PUFs are used to authenticate wireless body sensor network devices [11].
- Cloud storage requires security to protect users’ sensitive data. However, securing the
cloud is expensive and the users do not necessarily trust the cloud service providers.
A PUF in a universal serial bus (USB) token, i.e., Saturnus, has been trademarked
to encrypt user data before uploading the data to the cloud, decrypted locally by
reconstructing the same secret from the same PUF.
- System developers want to mutually authenticate a field programmable gate array
(FPGA) chip and the intellectual property (IP) components in the chip, and IP developers want to protect the IP. In [12], a protocol is described to achieve these goals with
a small hardware area that uses one symmetric cipher and one PUF.
Other applications of PUFs include providing non-repudiation (i.e., undeniable transmission or reception of data), proof of execution on a specific processor, and remote integrated circuit (IC) enabling. Every application of PUFs has different assumptions about the
PUF properties, computational complexity, and the specific system models. Therefore, there
are different constraints and system parameters for each application. We focus mainly on the
application where a secret key is generated from a PUF for user, or device, authentication
with privacy and secrecy guarantees, and low complexity.
_2.2. Main PUF Types_
We review four PUF types, i.e., silicon, arbiter, RO, and SRAM PUFs. We consider
mainly the last two PUF types for algorithm and code designs due to their common use
in practice and because signal processing techniques can tackle the problems arising in
designing these PUFs. For a review of other PUF types that are mostly considered in the
hardware design and computer science literatures, and various classifications of PUFs, see,
for example, [4,13,14]. The four PUF types considered below can be shown to satisfy the
assumption that invasive attacks permanently change PUF outputs, since digital circuit
outputs used as the source of randomness in these PUF types change permanently under
invasive attacks due to their dependence on nano-scale alterations in the hardware.
_2.3. Silicon and Arbiter PUFs_
Common complementary metal-oxide-semiconductor (CMOS) manufacturing processes are used to build silicon PUFs, where the response of the PUF depends on the circuit
delays, which vary across integrated circuits (ICs) [5]. Due to high sensitivity of the circuit
delays to environmental changes (e.g., ambient temperature and power supply voltage),
arbiter PUFs are proposed in [15], for which an arbiter (i.e., a simple transparent data latch)
is added to the silicon PUFs so that the delay comparison result is a single bit. The differ
-----
_Entropy 2021, 23, 16_ 6 of 23
ence of the path delays is mapped to, for example, the bit 0 if the first path is faster, and
the bit 1 otherwise. The difference between the delays can be small, causing meta-stable
outputs. Since the output of the mapper is generally pre-assigned to the bit 0, the signals
that are incoming are required to satisfy a setup time (tsetup), required by the latch to
change the output to the bit 1, resulting in a bias in the arbiter PUF outputs. Symmetrically
implementable latches (e.g., set-reset latches) should be used to overcome this problem,
which is difficult because FPGA routing does not allow the user to enforce symmetry in the
hardware implementation. We discuss below that PUFs without symmetry requirements,
for example, RO PUFs, provide better results.
_2.4. RO PUFs_
The RO logic circuit is depicted in Figure 1, where an odd number of inverters are
connected serially with feedback. The first logic gate in Figure 1 is a NAND gate, giving the
same logic output as an inverter gate when the ENABLE signal is 1 (ON), to enable/disable
the RO circuit. The manufacturing-dependent and uncontrollable component in an RO
is the total propagation delay of an input signal to flow through the RO, determining the
oscillation frequency [1]
_xˆ_ [of an RO that is used as the source of randomness. A self-sustained]
oscillation is possible when the ring that oscillates at the oscillation frequency [1]
_xˆ_ [of the RO]
provides a phase shift of 2π with a voltage gain of 1.
Consider an RO with m ≥ 3 inverters. Each inverter should provide a phase shift of _m[π]_
with an additional phase shift of π due to the feedback. Therefore, the signal should flow
through the RO twice to provide the necessary phase shift [16]. Suppose a propagation
1
delay of τd for each inverter, so the oscillation frequency of an RO is [1] . We remark
_xˆ_ [=] 2mτd
that since RO outputs are generally measured by using 32-bit counters, it is realistic to
assume that a measured RO output [1]
_xˆ_ [is a realization of a continuous distribution that can]
be modeled by using the histogram of a family of RO outputs with the same circuit design,
as assumed below.
The propagation delay τd is affected by nonlinearities in the digital circuit. Furthermore, there are deterministic and additional random noise sources [16]. Such effects should
be eliminated to have a reliable RO output. Rather than improving the standard RO designs,
which would impose the condition that manufacturers should change their RO designs, the
first proposal to fix the reliability problem was to make hard bit decisions by comparing
RO pairs [17], as illustrated in Figure 3.
Oscillator 1
ENABLE
Bit
><
Response
Counter
Oscillator N
Challenge
**Figure 3. The first and most common RO physical unclonable function (PUF) design [17].**
In Figure 3, the multiplexers are challenged by a bit sequence of length at most
_⌈log2 N⌉_ so that an RO pair out of N ROs is selected. The counters count the number of
times a rising edge is observed for each RO during a fixed time. A logic bit decision is
made by comparing the counter values, which can be bijectively mapped to the oscillation
-----
_Entropy 2021, 23, 16_ 7 of 23
frequencies. For instance, when the upper RO has a greater counter value, then the bit 0 is
generated; otherwise, the bit 1. Given that ROs are identically laid out in the hardware,
the differences in the oscillation frequencies are determined mainly by uncontrollable
manufacturing variations. Furthermore, it is not necessary to have a symmetric layout
when hard-macro hardware designs are used for different ROs, unlike arbiter PUFs.
The key extraction method illustrated in Figure 3 gives an output of ([N]2 [)][ bits, which]
are correlated due to overlapping RO comparisons. This causes a security threat and makes
the RO PUF vulnerable to various attacks, including machine learning attacks. Thus, nonoverlapping pairs of ROs are used in [17] to extract each bit. However, there are systematic
variations in the neighboring ROs due to the surrounding logic, which also should be
eliminated to extract sequences with full entropy. Furthermore, ambient temperature and
supply voltage variations are the most important effects that reduce the reliability of RO
PUF outputs. A scheme called 1-out-of-k masking is proposed as a countermeasure to these
effects, which compares the RO pairs that have the maximum difference between their
oscillation frequencies for a wide range of temperatures and voltages to extract bits [17].
The bits extracted by such a comparison are more reliable than the bits extracted by using
previous methods. The main disadvantages of this scheme are that it is inefficient due to
unused RO pairs, and only a single bit is extracted from the (semi-) continuous RO outputs.
We review transform-coding based RO PUF methods below that significantly improve on
these methods without changing the standard RO hardware designs.
_2.5. SRAM PUFs_
There are multiple memory-based PUFs such as SRAM, Flip-flop, DRAM, and Butterfly PUFs. Their common feature is to possess a small number of challenge-response pairs
with respect to their sizes. As the most promising memory-based PUF type that is already
used in the industry, we consider SRAM PUFs that use the uncontrollable settling state of
bi-stable circuits [18]. In the standard SRAM design, there are four transistors used to form
the logic of two cross-coupled inverters, as depicted in Figure 2, and two other transistors
to access the inverters. The power-up state, i.e., (Q, Q) = (1, 0) or (0, 1), of an SRAM cell
provides one secret bit. Concatenating many such bits allows to generate a secret key from
SRAM PUFs on demand. We provide an open problem about SRAM PUFs in Section 10.
**3. Correlated, Biased, and Noisy PUF Outputs**
PUF circuit outputs are biased (nonuniform), correlated (dependent), and noisy
(erroneous). We review a transform-coding algorithm that extracts an almost i.i.d. uniform
bit sequence from each PUF, so a helper-data generation algorithm can correct the bit errors
in the sequence generated from noisy PUF outputs. Using this transform-coding algorithm,
we also obtain memoryless PUF measurement-channel models, so standard informationtheoretic tools, which cannot be easily applied to correlated sequences, can be used.
**Remark 1. The bias in the PUF circuit outputs is considered in the PUF literature to be a big**
_threat against the security of the key generated from PUFs since the bias allows to apply, for example,_
_machine learning attacks. However, it is illustrated in [19] (Figure 6) that the output bias does_
_not change the information-theoretic rate regions significantly, illustrating that there exist code_
_constructions that do not require PUF outputs to be uniformly distributed._
We consider two scenarios, where a secret key is either generated from PUF outputs
(i.e., generated secret [GS] model) or they are bound to PUF outputs (chosen secret [CS]
model). An example of GS methods is code-offset fuzzy extractors (COFE) [20], and
an example of the CS methods is the fuzzy-commitment scheme (FCS) [21]. We first
analyze a method that significantly improves privacy, reliability, hardware cost and secrecy
performance, by transforming the PUF outputs into a frequency domain, which are later
used in the FCS. We remark that the information-theoretic analysis of the CS model follows
directly from the analysis of the GS model [22], so one can use either model for comparisons.
-----
_Entropy 2021, 23, 16_ 8 of 23
PUF output correlations might cause information leakage about the PUF outputs
(i.e., privacy leakage) and about the secret key (i.e., secrecy leakage) [22,23]. Furthermore,
channel codes are required to satisfy the constraint on the reliability due to output noise.
The transform coding method proposed in [24] adjusts the PUF output noise to satisfy the
reliability constraint in addition to reducing the PUF output correlations.
_3.1. PUF Output Model_
Consider a (semi-)continuous output physical function such as an RO output as a
source with real valued outputs ˆx. Since in a two-dimensional (2D) array the maximum
distance between RO hardware logic circuits is less than in a one-dimensional array, decreasing the variations in the RO outputs caused by surrounding hardware logic circuits [25],
we consider a 2D RO array of size l = r _c that can be represented as a vector random_
_×_
variable _X[l]. Each device embodies a single 2D RO array that has the same circuit design_
[�]
and we have _X[�][l]_ _∼_ _f �Xl_, where f �Xl is a probability density function. Mutually independent
and additive Gaussian noise denoted as _Z[l]_ disturbs the RO outputs, i.e., we have noisy RO
[�]
outputs _Y[l]_ = _X[l]_ +Z[l]. Since _X[l]_ and _Y[l]_ are dependent, using these outputs a secret key can
[�] [�] [�] [�] [�]
be agreed [26,27].
**Remark 2. PUF outputs are noisy, as discussed above in this section. However, the first PUF**
_outputs are used by, for example, a manufacturer to generate or embed a secret key, which is called_
_the enrollment procedure. Since a manufacturer can measure multiple noisy outputs of the same_
_RO to estimate the noiseless RO output, we can consider that the PUF outputs measured during_
_enrollment are noiseless. However, during the reconstruction step, for example, an IoT device_
_observes a noisy RO output, which can be the case because the IoT device cannot measure the_
_RO outputs multiple times due to delay and complexity constraints. Therefore, we consider a_
_key-agreement model where the first measurement sequence (during enrollment) is noiseless and the_
_second measurement sequence (during reconstruction) is noisy; see also Section 8. Extensions to_
_key agreement models with two noisy sequences, where the noise components can be correlated, are_
_discussed in [23,28,29]._
We extract i.i.d. symbols from _X[l]_ and _Y[l]_ such that information theoretic tools used
[�] [�]
in [30] for the FCS can be applied. An algorithm is proposed in [24] to obtain almost
i.i.d. uniformly-distributed and binary vectors X[n] and Y[n] from _X[l]_ and _Y[l], respectively._
[�] [�]
For such X[n] and Y[n], we can define a binary error vector as E[n] = X[n] _Y[n], where_ is
_⊕_ _⊕_
the modulo-2 sum. We then obtain the random sequence E[n] Bernoulli[n](p), so the
_∼_
channel PY|X ∼ BSC(p) is a binary symmetric channel (BSC) with crossover probability p.
We discuss a transform-coding method below, which further provides reliability guarantees
for each bit generated.
The FCS can reconstruct a secret key from dependent random variables with zero
secrecy leakage [21]. For the FCS, depicted in Figure 4, an encoder Enc( ) maps a secret
_·_
key S, which is uniformly distributed in the set 1, 2, . . .,, into a codeword C[n]
_∈S_ _{_ _|S|}_
with binary symbols that are later added to the PUF-output sequence X[n] in modulo-2
during enrollment. The output is called helper data W, sent to a database via a noiseless,
public and authenticated communication link. The sum of W and Y[n] in modulo-2 is
_R[n]_ = W _Y[n]_ = C[n] _E[n], mapped to a secret key estimate_ _S[ˆ] during reconstruction by the_
_⊕_ _⊕_
decoder Dec( ).
_·_
We next give information-theoretic rate regions for the FCS; see [31] for informationtheoretic notation and basics.
-----
_Entropy 2021, 23, 16_ 9 of 23
_S_
_Sˆ_
Enrollment Reconstruction
**Figure 4. The fuzzy commitment scheme (FCS).**
**Definition 2. The FCS can achieve a secret-key vs. privacy-leakage rate pair (Rs,Rℓ) with zero**
_secrecy leakage (i.e., perfect secrecy) if, given any δ > 0, there is some n ≥_ 1, and an encoder and
_decoder pair for which we have Rs =_ [log][ |S|] _and_
_n_
Pr[S[ˆ] ̸= S] = PB ≤ _δ_ (reliability) (1)
_H(S) ≥_ _n(Rs −_ _δ)_ (key uniformity) (2)
_I(S; W)=_ 0 (perfect secrecy) (3)
_I(X[n]; W) ≤_ _n(Rℓ_ + δ) (privacy) (4)
_where (3) suggests that S and W are independent and (4) suggests that the rate of dependency_
_between X[n]_ _and W is bounded. The achievable secret-key vs. privacy-leakage rate, or key-leakage,_
_region RFCS for the FCS is the union of all achievable pairs._
**Theorem 1 ([30]). The key-leakage region RFCS for the FCS with perfect secrecy, uniformly-**
_distributed X and Y, and a channel PY|X ∼_ _BSC(p) is_
_RFCS = {(Rs, Rℓ)_ : 0 ≤ _Rs ≤_ 1 − _Hb(p),_ _Rℓ_ _≥_ 1 − _Rs}_ (5)
_where Hb(p) = −p log p −_ (1 − _p) log(1 −_ _p) is defined as the binary entropy function._
The region R of all achievable (secret-key, privacy-leakage) rate pairs for the CS model
with a negligible secrecy-leakage rate is [22]
�
= �
_R_
_PU|X_
�
(Rs, Rℓ): 0 ≤ _Rs ≤_ _I(Y; U),_ _Rℓ_ _≥_ _I(X; U) −_ _I(Y; U)_
(6)
such that U _X_ _Y forms a Markov chain and it suffices to have_ + 1. The aux_−_ _−_ _|U|≤|X |_
iliary random variable U represents a distorted version of X through a channel PU|X.
The FCS is optimal only at the point (Rs[∗][,][ R][∗]ℓ [)=(][1][−][H][b][(][p][)][,][ H][b][(][p][))][ [][30][], corresponding to]
the maximum secret-key rate.
**4. Transformation Steps**
Transform coding methods decrease RO output correlations for ROs that are in the
same 2D array by using, for example, a linear transformation. We discuss a transformcoding algorithm proposed in [32] as an extension of [24] to provide reliability guarantees
to each generated bit. Joint optimization of the error-correction code and quantizer in
order to maximize the reliability and secrecy are the main steps. The output of these
post-processing steps is a bit sequence X[n] (or its noisy version Y[n]) utilized in the FCS.
-----
_Entropy 2021, 23, 16_ 10 of 23
It suffices to discuss only the enrollment steps, depicted in Figure 5, since the same steps
are used also for reconstruction.
_X[l]_ are correlated RO outputs, where the cause of correlations is, for example, the sur�
rounding logic in the hardware. A transform Tr×c(·) with size r×c transforms RO outputs
to decrease output correlations. We model each output T in the transform domain, i.e.,
_transform coefficient, calculated by transforming the RO outputs given in the dataset [33] by_
using the Bayesian information criterion (BIC) [34] and the corrected Akaike’s information
criterion (AICc) [35], suggesting a Gaussian distribution as a good fit for the discrete Haar
transform (DHT), discrete Walsh–Hadamard transform (DWHT), DCT, and Karhunen–
Loève transform (KLT).
##### X[^][ l]
##### Post-Processing
#### Transform Hist.
##### Quant. with X [n]
#### Trxc Equali.
###### Gray Map
**Figure 5. Transformation steps [24].**
In Figure 5, the histogram equalization changes the probability density of the i-th
coefficient Ti into a standard normal distribution so that quantizers are the same for all
transform coefficients, decreasing the storage. Obtained coefficients _Ti are independent_
[�]
when the transform coefficients Ti are jointly Gaussian and the transform Tr×c(·) decorrelates the RO outputs perfectly. For such a case, scalar quantizers do not introduce any
performance loss. Bit extraction methods and scalar quantizers are given below for the FCS
with the independence assumption, which can be combined with a correlation-thresholding
approach in practice.
**5. Joint Quantizer and Error-Correction Code Design**
The steps in Figure 5 are applied to obtain a uniform binary sequence X[n]. We utilize
a quantizer ∆(·) that assigns quantization-interval values of k = 1, 2, · · ·, 2[K][i], where Ki
represents the number of bits obtained from the i-th coefficient. We have
∆(t[ˆ]i) = k if _bk−1 <_ _t[ˆ]i ≤_ _bk_ (7)
� _k_ �
where we have bk = Φ[−][1], and Φ[−][1](·) is the standard Gaussian distribution’s
2[K][i]
quantile function. A length-Ki bit sequence represents the output k. Since the noise
has zero mean, we use a Gray mapping to determine the sequences assigned to each k,
so neighboring sequences differ only in one bit.
_Quantizers with Given Maximum Number of Errors_
We discuss a conservative approach that suppose either bits assigned to a quantized
transform coefficient all flip or they are all correct. Let the correctness probability Pc of a
coefficient be the probability that all bits assigned to a transform coefficient are correct,
used to choose the number of bits extracted from a coefficient in such a way that one can
design a channel encoder with a bounded minimum distance decoder (BMDD) to satisfy
|Col1|Col2|
|---|---|
|RO Array rxc|Col2|Col3|Col4|
|---|---|---|---|
|l||||
|||||
|Transform T rxc||||
-----
_Entropy 2021, 23, 16_ 11 of 23
the reliability constraint PB ≤ 10[−][9], a common value for the block-error probability of
PUFs that use CMOS circuits [17].
Let Q(·) be the Q-function, f �T the probability density of the standard Gaussian distribution, and σn[2]ˆ [the noise variance. The correctness probability can be calculated as]
_f �T(t[ˆ])dt[ˆ]_ (8)
� � _bk+1−tˆ_
_−Q_
_σnˆ_
2[K]−1
_Pc(K)=_ ∑
_k=0_
_bk+1_
�
_bk_
� � _bk−tˆ_
_Q_
_σnˆ_
�[�]
where K is the length of the bit sequence assigned to a quantizer with quantization boundaries bk from (7) for an equalized Gaussian transform coefficient _T. In (8), we calculate the_
[�]
probability that the additive noise will not change the quantization interval assigned to the
transform coefficient, i.e., all bits associated with the transform coefficient stay the same
after adding noise.
Assume that all errors in up to Cmax coefficients can be corrected by a channel decoder,
that the correctness probability Pc,i(K) of the i-th coefficient _Ti is greater than or equal to_
[�]
_Pc(Cmax), and that errors occur independently. We first find the minimum correctness_
probability that satisfies PB ≤ 10[−][9], denoted as Pc(Cmax), by solving
_l_
#### ∑
_c=Cmax+1_
�l
_c_
�
(1−Pc(Cmax))[c]Pc(Cmax)[l][−][c] _≤_ 10[−][9] (9)
which allows to find the maximum bit-sequence length Ki for the i-th transform coefficient
such that Pc,i(K) ≥ _Pc(Cmax). The first transform coefficient, i.e., DC coefficient,_ _T[�]1 can_
in general be estimated by an attacker, which is the first reason why it is not used for
key extraction. As the second reason, temperature and voltage changes affect RO outputs
highly linearly, which affects the DC coefficient the most [36]. Thus, we fix K1 = 0, so the
total number of extracted bits can be calculated as
_l_
_n(Cmax)=_ ∑ _Ki._ (10)
_i=2_
We first sort Ki values in descending order such that Ki[′] _[≥]_ _[K]i[′]+1_ [for all][ i] [=] [1, 2,][ . . .][,][ l][ −] [2. Thus,]
up to
_Cmax_
_e(Cmax) =_ ∑ _Ki[′]_ (11)
_i=1_
bit errors must be corrected for the worst case scenario. Using a BMDD, a block code with
minimum distance dmin ≥ 2e(Cmax)+1 can satisfy this requirement [37].
The advanced encryption standard (AES) requires a seed of, e.g., a secret key with
length 128 bits. If the FCS is applied to PUFs to extract such a secret key for the AES, the block
code designed should have a code length ≤ _n(Cmax) bits, code dimension ≥128 bits, and_
minimum distance dmin ≥ 2e(Cmax) + 1, given a Cmax. Such an optimization problem is
generally hard to solve but, using an exhaustive search over different Cmax values and
over different algebraic codes, one can show the existence of a channel code that satisfies
all constraints. Considering codes with low-complexity implementations is preferred for,
e.g., IoT applications. We remark that the correctness probability might be significantly
greater than Pc(Cmax), that the probability that less than Ki bits are actually in error when
the i-th coefficient is erroneous is high, and that the bit errors do not necessarily happen in
the coefficients from which the maximum-length bit sequences are obtained. Therefore,
we next illustrate that even though e(Cmax) errors cannot be corrected, the constraint
_PB ≤_ 10[−][9] is satisfied.
-----
_Entropy 2021, 23, 16_ 12 of 23
**6. PUF Performance Evaluations**
Represent RO outputs _X[l]_ as a vector random variable with the autocovariance matrix
[�]
**C** 8 and 16 16 RO arrays, whose autocovariance matrix is estimated by
**X�X� and consider 8×** _×_
using the RO outputs in [33]. Using the dataset, we next compare the performance of the
DWHT, DCT, KLT, and DHT in terms of their security, decorrelation efficiency, uniqueness,
and complexity.
_6.1. Decorrelation Efficiency_
Consider the autocovariance matrix CTT of the transform coefficients so that the
decorrelation efficiency ηc, used as a decorrelation performance metric, of a fixed transform
is [38]
_l_
∑ _|CTT(a, b)|1{a_ _̸=_ _b}_
_b=1_
(12)
_l_
_b∑=1_ _|CX�X�_ (a, b)|1{a _̸=_ _b}_
_ηc = 1 −_
_l_
∑
_a=1_
_l_
∑
_a=1_
where 1 is the indicator function. The KLT has a decorrelation efficiency of 1,
_{·}_
i.e., optimal [38]. Average ηc values of remaining transforms are given in Table 1 and
they have good (i.e., high) and similar decorrelation efficiency performance. The DHT
and DCT have the highest efficiency for 8 8 RO arrays; while, for 16 16 RO arrays,
_×_ _×_
the DWHT is the best transform. Table 1 suggests that an array size increase improves ηc.
**Table 1. Average decorrelation-efficiency results for RO outputs.**
**DWHT** **DCT** **DHT**
_ηc for 16 × 16_ 0.9988 0.9987 0.9986
_ηc for 8 × 8_ 0.9977 0.9978 0.9978
_6.2. Complexity of Transforms_
Computational complexity of r _c = 8_ 8 and 16 16 RO arrays are considered,
_×_ _×_ _×_
which are powers of 2 so that there are fast algorithms to implement the DWHT, DCT, and
DHT. The KLT has a computational complexity of O(n[3]) for r = c = n; while, the DWHT
and DCT have O(n[2] log2 n), and the DHT has O(n[2]) [39]. Efficient implementations of
the DWHT that do not require multiplications exist [32], which can be applied also to the
transforms proposed in [40]. The DWHT is therefore a good candidate for implementing
RO PUFs for IoT applications. For instance, a hardware implementation of 2D DWHT in an
evaluation board of Xilinx ZC706 with a Zynq-7000 XC7Z045 system-on-chip is illustrated
in [32] to require approximately 11% smaller hardware area and 64% less processing time
than the benchmark RO PUF hardware implementation in [41].
_6.3. Security and Uniqueness_
The extracted bit sequence is required to be uniformly distributed to use the rate region
_RFCS in (5). The randomness measure called uniqueness is the average fractional Hamming_
distance between bit sequences generated from different RO PUFs. All transforms have
similar uniqueness results with a mean Hamming distance of 0.500 and Hamming distance
variance is 7 × 10[−][4]. These results are close to optimal uniqueness results, expected because
of equipartitioned quantization intervals and high decorrelation efficiencies, that are better
than previous uniqueness results with mean values of 0.462 [17] and 0.473 [33].
The national institute of standards and technology (NIST) has randomness tests to
check if an extracted binary sequence can be differentiated from a uniformly-random
binary sequence [42]. The bit sequences with the DWHT pass most of the applicable tests,
considered to be an acceptable result [42]. The KLT performs the best because of its optimal
decorrelation performance.
-----
_Entropy 2021, 23, 16_ 13 of 23
**7. Error-Correction Codes for PUFs with Transform Coding**
Suppose that bit sequences extracted by using the transform-coding method are i.i.d.
and uniformly distributed, so perfect secrecy is satisfied. We assume that signal processing
steps mentioned above perform well, so we can conduct standard information- and codingtheoretic analysis. We provide a list of codes designed for the transform-coding algorithm
by using the reliability metric considered above.
Select a channel code for the quantizer designed above for a fixed maximum number
of errors for a secret key of size 128 bits. The correctness probabilities for the coefficients
with the smallest and highest probabilities are depicted in Figure 6. Transform coefficients
that represent the low-frequency coefficients are the most reliable, which are at the upperleft corner of the 2D transform-coefficient array with indices such as 1, 17, 2, 18, 3, 19. These
coefficients thus have the highest signal-to-noise ratios (SNRs). Conversely, the least
reliable coefficients are observed to be coefficients that represent intermediate frequencies,
indicating that one can define a metric called SNR-packing efficiency, defined similarly as
the energy-packing efficiency, and show that it follows a more complicated scan order than
the classic zig-zag scan order used for the energy-packing efficiency.
1
0.8
0.6
0.4
0.2
0
1 2 3 4 5 6 7 8 9 10
|Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|
|---|---|---|---|---|---|---|---|---|
||||||||||
|Co|||||||||
||eff. 1||||||||
|Co Co|eff. 2 eff. 17||||||||
|Co Co Co|eff. 31 eff. 128 eff. 150||||||||
_K_
**Figure 6. Transform coefficients’ correctness probabilities.**
Fix Cmax, defined above, and calculate Pc(Cmax) via (9), n(Cmax) via (10), and e(Cmax)
via (11). If Cmax ≤ 10, Pc(Cmax) is large and Pc,i(K = 1) _≤_ _Pc(Cmax) for all i = 2, . . ., l. In ad-_
dition, if 11 _≤_ _Cmax ≤_ 15, then n(Cmax) ≤ 128 bits. Furthermore, if Cmax increases, Pc(Cmax)
decreases, so the maximum of the number Kmax(Cmax)= _K1[′]_ [(][C][max][)][ of bits extracted among]
all used coefficients increases, increasing the hardware complexity. Thus, consider only the
cases where Cmax ≤ 20. Table 2 shows Pc(Cmax), n(Cmax), and e(Cmax) for a range of Cmax
values used for channel-code selection.
**Table 2. Code-design constraints.**
**Cmax** **20** **19** **18** **17** **16**
_Pc_ 0.9844 0.9860 0.9875 0.9889 0.9902
_Kmax_ 3 3 3 3 3
_n_ 259 255 250 224 144
_e_ 25 23 21 20 18
Consider Reed–Solomon (RS) and binary (extended) Bose–Chaudhuri–Hocquenghem
(BCH) codes, whose minimum-distance dmin is high. There is no BCH or RS code with
parameters satisfying any of the (n(Cmax), e(Cmax)) pairs in Table 2 such that its dimension
-----
_Entropy 2021, 23, 16_ 14 of 23
is 128 bits. However, the analysis leading to Table 2 is conservative. Thus, we next
_≥_
find a BCH code whose parameters are as close as possible to an (n(Cmax), e(Cmax)) pair
in Table 2. Consider the binary BCH code that can correct all error patterns with up to
_eBCH = 18 errors with the block length of 255 and code dimension of 131 bits._
First, extract exactly one bit from each transform coefficient, i.e., Ki = 1 for all
_i_ = 2, 3, . . ., l, so n = l 1 = 255 bits are extracted, resulting in mutually-independent
_−_
bit errors Ei. Thus, all error patterns with up to e = 20 bit errors should be corrected by
the chosen code rather than e(20)= 25 bit errors. However, this value is still greater than
_eBCH = 18._
The block error probability PB for the BCH code C(255, 131, 37) with a BMDD is equal
to the probability of encountering more than 18 errors, i.e., we have
255
_PB =_ ∑
_j=19_
�
#### ∑ (1 − Pc,i) [•] ∏ Pc,i
_A∈Fj_ _i[∏]∈A_ _i∈A[c]_
�
(13)
where Pc,i is the correctness probability of the i-th coefficient _Ti as in (8) for i_ = 2, 3, . . ., 256,
[�]
_A[c]_ denotes the complement of the set A, and Fj is the set of all size-j subsets of the set
_{2, 3, . . ., 256}. Pc,i values are different and they represent probabilities of independent_
events because we assume that the transform coefficients are independent. We apply the
discrete Fourier transform characteristic function method [43] to evaluate the block-error
probability with the result PB ≈ 1.26 × 10[−][11] _< 10[−][9]. The block-error probability_
(i.e., reliability) constraint is therefore satisfied by the BCH code (255, 131, 37), although
_C_
the conservative analysis suggested otherwise. This code achieves a (secret-key, privacyleakage) rate pair of (Rs, Rℓ) = ( [131]255 [, 1][−] 255[131] [)][ ≈] [(][0.514, 0.486][)][ bits/source-bit, which is]
significantly better than previous results. We next consider the region of all achievable
rate pairs for the CS model and the FCS for a BSC PY|X with crossover probability pb =
1 − _l−11_ [∑]i[l]=2 _[P][c][,][i][(][K][i][ =][ 1][)]_ _[≈]_ [0.0097, i.e., probability of being in error averaged over all used]
coefficients with the above defined quantizer. The (secret-key, privacy-leakage) rate pair of
the BCH code, regions of all rate pairs achievable by the FCS and CS model, the maximum
secret-key rate point, and a finite-length bound [44] for the block length of n = 255 bits and
_PB =_ 10[−][9] are depicted in Figure 7 for comparisons.
1
0.8
0.6
0.4
0.2
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
|Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|Col11|Col12|
|---|---|---|---|---|---|---|---|---|---|---|---|
|||||||||||||
|||||||||||||
|||Propos|ed Code|||||||||
|||Fuzzy CS Mo|Commit del|ment||||||||
|||(R∗, R∗ l s Finite-l|) ength B|ound||||||||
_Rl_
**Figure 7.** The operation point of the considered Bose–Chaudhuri–Hocquenghem (BCH) code
_C(255, 131, 37), the maximum secret-key rate point (R[∗]ℓ_ [,][ R][s][∗][ )][, regions of achievable rate pairs][ according]
to (5) and (6), and a finite-length bound for BSC(0.0097), n = 255 bits, and PB = 10[−][9].
Denote the maximum secret-key rate as Rs[∗] _[≈]_ [0.922 bits/source-bit and the corre-]
sponding minimum privacy-leakage rate as R[∗]
_ℓ_ _[≈]_ [0.079 bits/source-bit. The gap between]
(R[∗]ℓ [,][ R]s[∗][)][ at which the FCS is optimal and the rate tuple achieved by the BCH code can be]
explained by the short block length and small block-error probability. However, the finitelength bound given in [44] (Theorem 52) suggests that the FCS can achieve the rate tuple
-----
_Entropy 2021, 23, 16_ 15 of 23
(Rs, Rℓ) = (0.691, 0.309) bits/source-bit, shown in Figure 7. Better channel code designs
and decoders (possibly with higher hardware implementation complexity) can improve the
performance, but they might not be feasible for IoT applications. Figure 7 shows that there
are other code constructions (that are not standard error-correcting codes) that can achieve
smaller privacy-leakage and storage rates for a fixed secret-key rate, illustrated below.
**8. Code Constructions for PUFs**
Consider the two-terminal key agreement problem, where the identifier outputs
during enrollment are noiseless. We mention two optimal linear code constructions from [45]
that are based on distributed lossy source coding (or Wyner–Ziv [WZ] coding) [46]. The random linear code construction achieves the GS and CS models’ key-leakage-storage regions
and the nested polar code construction jointly designs vector quantization (during enrollment) and error correction (during reconstruction) codes. Designed nested polar codes
improve on existing code designs in terms of privacy-leakage and storage rates, and one
code achieves a rate tuple that existing methods cannot achieve.
Several practical code constructions for key agreement with identifiers have been
proposed in the literature. For instance, the COFE and the FCS both require a standard errorcorrection code to satisfy the constraints of, respectively, the key generation (GS model)
and key embedding (CS model) problems, as discussed above. Similarly, a polar code
construction is proposed for the GS model in [47]. These constructions are sub-optimal in
terms of storage and privacy-leakage rates.
A Golay code is used as a vector quantizer (VQ) in [22] in combination with distributed
lossless source codes (or Slepian–Wolf [SW] codes) [48] to increase the ratio of key vs.
storage rates (or key vs. leakage rates). Thus, we next consider VQ by using WZ coding
to decrease storage rates. The WZ-coding construction turns out to be optimal, which is
not coincidental. For instance, the bounds on the storage rate of the GS model and on the
WZ rate (storage rate) have the same mutual information terms optimized over the same
conditional probability distribution. This similarity suggests an equivalence that is closely
related to the concept of formula duality. In fact, the optimal random code construction,
encoding, and decoding operations are identical for both problems. One therefore can
call the GS model and WZ problem functionally equivalent. Such a strong connection
suggests that there might exist constructive methods that are optimal for both problems for
all channels, which is closely related to the operational duality concept.
Consider the GS model, where a secret key is generated from a physical or biometric source, depicted in Figure 8(a). The encoder Enc( ) observes during enrollment the
_·_
noiseless i.i.d. sequence X[n] _∼_ _PX to generate public helper data W and a secret key S,_
i.e., (S, W) = Enc(X[n]). The decoder Dec( ) observes during reconstruction the helper data
_·_
_W and a noisy measurement Y[n]_ of X[n] through a memoryless channel PY|X to estimate
the secret key, i.e., _S = Dec(Y[n], W). Similarly, the CS model is shown in Figure 8(b), where_
[�]
a secret key S independent of (X[n], Y[n]) is chosen and embedded into the helper data,
i.e., W = Enc(X[n], S). The alphabets,,, and are finite sets, which can be achieved
_X_ _Y_ _S_ _W_
if, for example, the transform-coding algorithm discussed above is applied.
-----
_Entropy 2021, 23, 16_ 16 of 23
_S_ _S�_
(a) (b) (a) (b)
(a)
(S, W) = Enc(X[n]) _W_
(b) _S� = Dec(Y[n], W)_
_W_ = Enc(X[n], S)
_PX_ _X[n]_ _PY|X_ _Y[n]_
Enrollment Reconstruction
**Figure 8. The (a) generated-secret (GS) and (b) chosen-secret (CS) models.**
**Definition 3. For GS and CS models, a key-leakage-storage tuple (Rs, Rℓ, Rw) is achievable if,**
_given any δ > 0, there is an encoder, a decoder, and some n_ _≥_ 1 such that Rs = [log][ |S|] _and_
_n_
Pr[S[ˆ] ̸= S] = PB ≤ _δ_ (reliability) (14)
_I(W; S) ≤_ _nδ_ (weak secrecy) (15)
_I(X[n]; W) ≤_ _n(Rℓ_ + δ) (privacy) (16)
_H(S) ≥_ _n(Rs −_ _δ)_ (uni f ormity) (17)
log |W| ≤ _n(Rw + δ)_ (storage) (18)
_are satisfied. The key-leakage-storage regions Rgs for the GS model and Rcs for the CS model are_
_the closures of the sets of achievable tuples for these models._
**Theorem 2 ([22]). The key-leakage-storage regions Rgs and Rcs for the GS and CS models,**
_respectively, are_
�
_Rcs =_
_PU|X_
�
(Rs, Rℓ, Rw):
�
_Rgs =_
_PU|X_
�
(Rs, Rℓ, Rw):
0 ≤ _Rs ≤_ _I(Y; U),_
_Rℓ_ _≥_ _I(X; U) −_ _I(Y; U),_
�
_Rw ≥_ _I(X; U) −_ _I(Y; U)_ _,_
_and_
0 ≤ _Rs ≤_ _I(Y; U),_
_Rℓ_ _≥_ _I(X; U) −_ _I(Y; U),_
�
_Rw ≥_ _I(X; U)_
_where U −_ _X −_ _Y form a Markov chain. Rgs and Rcs are convex sets and |U|≤|X | + 1 suffices_
_for both rate regions._
**Remark 3. Improvement of the weak secrecy to strong secrecy, where (15) is replaced with**
_I(W; S) ≤_ _δ, is possible by using multiple identifier output blocks as described in [49], e.g., by using_
_multiple PUFs in the same device._
Assume, as above, that X[n] _∼_ Bernoulli[n]( 2[1] [)][ and the channel][ P][Y][|][X] _[∼]_ [BSC][(][p][A][)][ for]
_pA ∈_ [0, 0.5]. Define the star-operation as q ∗ _pA = q(1 −_ _pA) + (1 −_ _q)pA. The key-leakage-_
storage region of this GS model is
�
_Rgs,bin =_
_q∈[0,0.5]_
�
(Rs, Rℓ, Rw):
0 ≤ _Rs ≤_ 1 − _Hb(q ∗_ _pA),_
_Rℓ_ _≥_ _Hb(q ∗_ _pA) −_ _Hb(q),_
�
_Rw ≥_ _Hb(q ∗_ _pA) −_ _Hb(q)_ . (19)
-----
_Entropy 2021, 23, 16_ 17 of 23
_Comparisons Between Code Constructions for PUFs_
We consider three best code constructions proposed for the GS and CS models, which
are COFE and the polar code construction in [47] for the GS model, and FCS for the CS
model, in order to compare them with the WZ-coding constructions. The FCS and COFE
achieve only a single point on the key-leakage rate region boundary, i.e., Rs[∗] [=][ I][(][X][;][ Y][)][ and]
_R[∗]_
_ℓ_ [=][ H][(][X][|][Y][)][.]
Adding a VQ step, one can improve these two methods. During enrollment rather
than X[n], its quantized version Xq[n] [can be used for this purpose, which can be asymptotically]
represented as summing the original helper data and another independent random variable
_J[n]_ Bernoulli[n](q), i.e., W = X[n] _C[n]_ _J[n]_ is the (new) helper data. Modified FCS and
_∼_ _⊕_ _⊕_
COFE can achieve the key-leakage region when a union of all achieved rate tuples is taken
over all q [0, 0.5]. Nevertheless, the helper data of the modified FCS and COFE have
_∈_
length n bits, i.e., the storage rate is 1 bit/source-bit, which is suboptimal.
The storage rate of 1 bit/source-bit is decreased by using the polar code construction
proposed in [47]. Nevertheless, this construction cannot achieve the key-leakage-storage
region. In addition, in [47] there is an assumption that a “private” key that is shared
between the encoder and decoder is available, which is not realistic because there is a
need for hardware protection against invasive attacks to have such a private key. If such a
hardware protection is feasible, there is no need to utilize an on-demand key reconstruction
and storage method like a PUF. The previous methods cannot, therefore, achieve the keyleakage-storage region for a BSC, unlike the distributed lossy source coding constructions
proposed in [45]. To compare such WZ-coding constructions, we use the ratio of key vs.
storage rates as the metric, which determines the design procedures to control the storage
and privacy leakage.
**9. Optimal Nested Polar Code Constructions**
The first channel codes with asymptotic information-theoretic optimality and low
decoding complexity are polar codes [50], whose finite length performance is good when a
list decoder is utilized. Nesting two codes is simple with polar codes due to their simple
matrix representation; therefore, one can use them for distributed lossy source coding [51].
The channel polarization phenomenon, i.e., converting a channel into polarized binary
channels by using a polar transform, is the core of polar codes. The polar transform takes a
sequence U[n] with unfrozen and frozen bits as input and converts it into a codeword that
has also length n. The decoder then observes a noisy codeword in addition to the fixed
frozen bits of U[n] in order to estimate the bit sequence U[n]. A polar code with block length
_n, and frozen bit sequence G[|F|]_ at indices are denoted as (n,, G[|F|]). We next utilize
_F_ _C_ _F_
nested polar codes that are proposed for WZ coding in [51].
_9.1. The GS Model Polar Code Construction_
Consider two nested polar codes C(n, F, V) and C1(n, F1, V) such that F = F1 ∪Fw
and V = [W, V], where W is of length m2 and V is of length m1. Suppose m1 and m2 satisfy
_m1_
(20)
_n_ [=][ H][b][(][q][)][ −] _[δ]_
_m1 + m2_
= Hb(q ∗ _pA) + δ_ (21)
_n_
for a δ > 0 and some distortion q ∈ [0, 0.5]. Two polar codes C(n, F, V) and C1(n, F1, V)
are nested since the set of indices F1 refer to frozen channels with values V, which are
common to both polar codes, and the code C has further frozen channels with values W at
indices Fw.
Since the rate of C1 is greater than the capacity of the lossy source coding problem for
an average distortion q, it functions as a VQ with distortion q. Furthermore, since the rate
of C is less than the channel capacity of the BSC(q ∗ _pA), it functions as an error-correcting_
code. We want to calculate the values W during enrollment, stored as the public helper
-----
_Entropy 2021, 23, 16_ 18 of 23
data, such that (V, W, Y[n]) can be used during reconstruction to estimate the key S with
length n − _m1 −_ _m2, which is depicted in Figure 9. We assign the all-zero vector to V, so to_
not increase storage, which does not affect the average distortion E[q] between Xq[n] [and][ X][n]
defined below; see [51] (Lemma 10) for a proof.
_S_
_Sˆ_
Helper Data
and Key _W_
Key
Extraction
Extraction
_U[n]_ _Uˆ_ _[n]_
Polar Polar Polar
_V_ _W_ _V_
Decoder C1 Transform Decoder C
_Xq[n]_
BSC(q ∗ _pA)_ _Y[n]_
_PX_ _X[n]_ BSC(pA)
##### Enrollment Reconstruction
**Figure 9. Second WZ-coding construction for the GS model.**
During enrollment, the PUF outputs X[n] _∼_ Bernoulli[n]( [1]2 [)][ are observed by a polar]
decoder of C1 and considered as noisy measurements of a sequence Xq[n] [measured through]
a BSC(q), i.e., X[n] is quantized into Xq[n] [by a polar decoder of][ C]1[. The polar decoder puts out]
the sequence U[n] and the bit values W at its indices Fw are publicly stored as the helper
data. Furthermore, the bit values at indices j ∈{1, 2, . . ., n} \ F are assigned as the secret
key S. We remark that the polar transform of U[n] is the sequence Xq[n] [that is the quantized]
(or distorted) version of X[n]. Consider the error sequence Eq[n] [=][ X][n][ ⊕] _[X]q[n][, which also models]_
the distortion between Xq[n] [and][ X][n][. The error sequence is shown in [][51][] (Lemma 11) to]
resemble a sequence that is distributed according to Bernoulli[n](q) when n tends to ∞.
During reconstruction, a polar decoder of then observes Y[n], a noisy version of X[n]
_C_
measured through a BSC(pA). The frozen bits V = [V, W] of C are available to the polar
decoder in order to estimate U[n], from which the secret key estimate _S can be obtained by_
[�]
finding the bit values at indices j ∈{1, 2, . . ., n} \ F .
Next, a design procedure to implement practical nested polar codes that satisfy these
properties is summarised.
Nested polar codes C ⊆C1 must be constructed jointly such that the sets of indices
_F and F1 result in codes that satisfy the security and reliability constraints simultane-_
ously. Suppose the block length n, key length n − _m1 −_ _m2, target block-error probability_
_PB = Pr[S ̸=_ _S[�]], and BSC crossover probability pA are given, which depends on the PUF_
application considered. Then we have the following design procedure [45]:
- Design a polar code with rate _[n][−][m][1][−][m][2]_, corresponding to fixing its indices that
_C_ _F_
_n_
determine the frozen bits. This step is a conventional error-correcting code design task.
-----
_Entropy 2021, 23, 16_ 19 of 23
- Find the maximum BSC crossover probability pc for which the code C achieves the
target block-error probability PB, which can be achieved by evaluating the performance of for a BSC over a crossover probability range. Using the inverse of the
_C_
star-operation pc = E[q] ∗ _pA, the target distortion averaged over a large number of_
realizations of X[n] that should be achieved by C1 is E[q] = _[p][c][ −]_ _[p][A]_ . This step can be
1 − 2pA
applied via Monte-Carlo simulations.
- Find an index set F1, representing the frozen set of C1, such that F1 ⊂F and the
target distortion E[q] is achieved with a minimal amount of helper data. This step
can be applied by starting with F1′ [=][ F][ and then computing the resulting average]
distortion E[q[′]] obtained from Monte-Carlo simulations. If E[q[′]] is greater than E[q],
we remove elements from F1′ [according to polarized bit channel reliabilities. This step]
is repeated until the resulting average distortion E[q[′]] is less than the target (or desired)
distortion E[q].
An additional degree of freedom is provided by varying the distortion level in the
design procedure above, making the design procedure suitable for numerous applications.
Using this degree of freedom, PUFs with different BSC crossover probabilities pA can be
supported by using the same nested polar codes with different distortion levels. Similarly,
different PUF applications with different target block-error probabilities PB can also be
supported by using the same nested codes with different distortion levels.
_9.2. Designed GS Model Nested Polar Codes_
We design nested polar codes to generate a secret key S of length log |S| = _n−m1−m2 =_
128 bits, used in the AES. Furthermore, the common target block-error probability for PUFs
used in an FPGA is PB = 10[−][6] and the common BSC PY|X crossover probability for SRAM
and RO PUFs is pA = 0.15 [6,36]. We consider these PUF applications and parameters to
design nested polar codes that improve on previously proposed codes.
_Code 1: Suppose a block length of n = 1024 bits and a fixed list size of 8 for polar_
successive cancellation list (SCL) decoders are used for nested codes. First, the code
_C_
with rate 128/1024 is designed to determine pc, which is defined in the design procedure
steps above, obtained by using the SCL decoder. We obtain the crossover probability value
_pc = 0.1819, corresponding to a target distortion of E[q] = 0.0456. This target distortion is_
obtained with a minimal helper data W length of m2 = 650 bits.
_Code 2: Suppose a block length of n = 2048 bits. Applying the design procedure steps_
given above, we obtain for Code 2 the value pc = 0.2682, resulting in a target distortion of
_E[q] = 0.1689. This target distortion is obtained with a minimal helper data W length of_
_m2 = 611 bits._
For these nested polar code designs, the error probability PB is considered as the
average error probability over a large number of input realizations, corresponding to
a large number of PUF circuits that have the same circuit design. This result can be
improved by satisfying the target error probability for each input realization, which can be
implemented by using the maximum distortion rather than E[q] in the design procedure
discussed above. A block-error probability that is 10[−][6] can be guaranteed for 99.99%
_≤_
of all realizations of input X[n] by including an additional 32 bits for the helper data W
for Code 1 and an additional 33 bits for Code 2. The numbers of additional bits included
are small because the distortion q has a small variance for the block lengths considered.
For code comparisons below, we depict the sizes of helper data needed to guarantee the
target block-error probability of PB = 10[−][6] for 99.99% of all PUF realizations.
_9.3. Comparisons of Codes_
The boundary points of Rgs,bin for pA = 0.15 are projected onto the storage-key
(Rw, Rs) plane and depicted in Figure 10. The point (Rs[∗][,][ R]w[∗] [)][, defined in Section][ 3.1][, is also]
depicted. Furthermore, we use the random coding union bound from [44] (Theorem 16)
to obtain the rate pairs that can be achieved by using the FCS or COFE. These points are
-----
_Entropy 2021, 23, 16_ 20 of 23
shown in Figure 10 in addition to the rate tuples achieved by the previous SW-coding based
_Entropypolar code design from [ 2020, 1, 0_ 47], and Codes 1 and 2 discussed above. 20 of 24
0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2
|Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|Col11|
|---|---|---|---|---|---|---|---|---|---|---|
||||||||||||
||||||||||||
|||||||Rgs,bin Boundary (R w∗, R s∗ ) Code 1, n=1024|||||
|||||||Code 2, n=2048 Prev. Polar Code [47], n|=1024||||
|||||||Best Code in [6], n=14 FCS/COFE achievable,|08 n=102|4|||
|||||||FCS/COFE achievable,|n=204|8|||
||||||||||||
||||||||||||
||||||||||||
||||||||||||
_Rw_
**Figure 10.Figure 10. Storage-key rates for the GS model with Storage-key rates for the GS model with pA p =A 0.15. The = 0.15. The (Rw[∗]**, ( RsR[∗] )w[∗] point is the best, Rs[∗] ) point is the best
possible point achieved by SW-coding constructions, which lies on the dashed line representing
possible point achieved by SW-coding constructions, which lies on the dashed line representing
_Rw + Rs = H(X). The block error probability satisfies PB ≤_ 10[−][6] and the key length is 128 bits for all
_Rw +code points. Rs = H(X). The block error probability satisfies PB ≤_ 10[−][6] and the key length is 128 bits for
all code points.
coding method in the binary case. The previous SW-coding based polar code construction improves the
The COFE and FCS result in a storage rate of 1 bit/source-bit, which is strictly subop
rate tuples achieved by the COFE and FCS in terms of the ratio of key vs. storage rates. Code 1 achieves
the key-leakage-storage tuple oftimal. The previous SW-coding based polar code construction in [ (0.125, 0.666, 0.666) bits/source-bit and Code 2 of47] achieves a rate tuple (0.063, 0.315, 0.315)
bits/source-bit, which significantly improve on all previous code constructions without any privatesuch that Rs + Rw = 1 bit/source-bit, as expected because it is an SW-coding construction
key assumption. Thus, Codes 1 and 2 results also suggest that for these parameters increasing thethat corresponds to a syndrome coding method in the binary case. The previous SW-coding
block length increases thebased polar code construction improves the rate tuples achieved by the COFE and FCS in Rs/Rw ratio, which is 0.188 for Code 1 and 0.199 for Code 2. Furthermore,
the privacy-leakage and storage rate tuple achieved by Code 2 cannot be achieved by using previousterms of the ratio of key vs. storage rates. Code 1 achieves the key-leakage-storage tuple
constructions without applying the time sharing method, because Code 2 achieves the privacy-leakageof (0.125, 0.666, 0.666) bits/source-bit and Code 2 of (0.063, 0.315, 0.315) bits/source-bit,
(and storage) rate of 0.315 bits/source-bit that is less than the minimal privacy-leakage (and storage)which significantly improve on all previous code constructions without any private key
ratesassumption. Thus, Codes 1 and 2 results also suggest that for these parameters increasing R[∗]ℓ [=][ R]w[∗] [=][ H]b[(][p]A[)][ ≈] [0.610 bits/source-bit that can be achieved by using previous code]
constructions.
the block length increases the Rs/Rw ratio, which is 0.188 for Code 1 and 0.199 for Code 2.
To find an upper bound on the the ratio of key vs. storage rates for the maximum secret-key
Furthermore, the privacy-leakage and storage rate tuple achieved by Code 2 cannot be
rate point, we apply the sphere packing bound from [52, Eq. (5.8.19)] for the channel pA = 0.15 and
achieved by using previous constructions without applying the time sharing method,
code parameters n = 1024, and PB = 10[−][6]. The sphere packing bound shows that the rate of C, as
because Code 2 achieves the privacy-leakage (and storage) rate of 0.315 bits/source-bit that
depicted in Figure 9, must satisfy R 0.273 bits/source-bit. Suppose the key rate is fixed to its
_C ≤_
maximum valuehave the ratio ofis less than the minimal privacy-leakage (and storage) ratesbits/source-bit that can be achieved by using previous code constructions. R Rss/ =Rw R ≤C and the storage rate is fixed to its minimum value0.375. Similarly, for n = 2048 we obtain the ratio of R[∗]ℓ [=][ R] R Rw[∗]ws/ =[=]R[ H]w 1 ≤ −b[(]0.437. The[p]RAC, so we[)][ ≈] [0.610]
two finite-length results that are valid for WZ-coding constructions with nested codes indicate thatTo find an upper bound on the the ratio of key vs. storage rates for the maximum
ratio of key vs. storage rates achieved by Codes 1 and 2 can be further increased. Using differentsecret-key rate point, we apply the sphere packing bound from [52] (Equation (5.8.19)) for
nested polar codes that improve the minimum-distance properties, as in [the channel pA = 0.15 and code parameters n = 1024, and PB =53 10], or using nested algebraic[−][6]. The sphere packing
codes for which design methods are available in the literature, as in [bound shows that the rate of C, as depicted in Figure 9, must satisfy54], one can reduce the gaps RC ≤ 0.273 bits/sourceto the finite-length bounds calculated for nested code constructions. We remark again that suchbit. Suppose the key rate is fixed to its maximum value Rs = RC and the storage rate is
optimality-seeking approaches, for example, based on information-theoretic security, provide the rightfixed to its minimum value Rw = 1 − _RC_, so we have the ratio of Rs/Rw ≤ 0.375. Similarly,
insights into the best solutions for the digital era’s security and privacy problems.for n = 2048 we obtain the ratio of Rs/Rw ≤ 0.437. The two finite-length results that are
valid for WZ-coding constructions with nested codes indicate that ratio of key vs. storage
rates achieved by Codes 1 and 2 can be further increased. Using different nested polar
codes that improve the minimum-distance properties, as in [53], or using nested algebraic
codes for which design methods are available in the literature, as in [54], one can reduce the
gaps to the finite-length bounds calculated for nested code constructions. We remark again
that such optimality-seeking approaches, for example, based on information-theoretic
-----
_Entropy 2021, 23, 16_ 21 of 23
security, provide the right insights into the best solutions for the digital era’s security and
privacy problems.
**10. Discussions and Open Problems**
- We want to use low-complexity scalar quantizers after transformation without extra
secrecy leakage; however, the decorrelation efficiency metric does not fully represent
the dependency between transform coefficients. What is the right metric to use
for choosing the transform used in combination with scalar quantizers? Is mutual
information between transform coefficients an appropriate metric for this purpose?
The choice of the transform should also depend on a reliability metric such as SNRpacking efficiency so that the transform, quantizers, and the error-correction codes
can be designed jointly. What is the right reliability metric for this purpose?
- It is shown in [36] that the ambient temperature and supply voltage affect the RO
outputs deterministically rather than adding extra random noise, which was assumed
in the RO PUF literature. What are the right output models for common PUF types,
i.e., what are the deterministic and random components, and how are they related?
- SRAM PUFs are already used in products. In the literature there is no extensive analysis of the output correlations between different SRAMs in the same device possibly
because SRAM outputs are binary and it is difficult to model the correlation between
binary symbols. However, SRAM outputs are modeled in [6] as binary-quantized
sums of independent Gaussian random variables. Is it possible to determine or approximate the correlations between the Gaussian random variables of different SRAMs?
If yes, this might be useful for an attacker to obtain information about the secret
sequence generated from the SRAM PUF output, which causes extra secrecy leakage.
- The transform-coding approach discussed above provides reliability guarantees for
RO arrays with random outputs, which considers an average over all ROs manufactured. The worst case scenario is when the transform coefficient value is on the
quantization boundary, for which the secret-key capacity is 0 bit. If one replaces
the average reliability metric used above by a lower bound on the reliability of each
RO, i.e., a worst-case scenario metric, how would this change the rate of the errorcorrection code used? For a fixed code, what should be the optimal bound on the
reliability of each RO to maximize the yield, i.e., the percentage of ROs among all
manufactured ROs for which the worst-case reliability guarantee is satisfied?
- Are the WZ problem and the GS model operationally equivalent?
- Linear block-code constructions discussed above are for uniformly-distributed PUF
outputs. Can one construct other (random) linear block codes that are asymptotically
optimal for nonuniform PUF outputs? Is it necessary to use an extension of the COFE
for this purpose?
- Consider the nested polar code design procedure given above. Construction of a
code for n ≤ 512 is not possible with the procedure discussed above because q ∗ _pA_
increases with increasing q for q [0, 0.5]. Is it possible to construct a nested polar
_∈_
code for n = 512 by improving the decoder and the code design procedure?
**Author Contributions: O.G. conceived the study, designed, and conducted the experiments; O.G.**
and R.F.S. contributed to the writing of the paper and analyzed the data, and a combined effort
of O.G. and R.F.S. improved the algorithms discussed. All authors have read and agreed to the
published version of the manuscript.
**Funding: O.G. and R.F.S. are supported by the German Federal Ministry of Education and Research**
(BMBF) within the national initiative for “Post Shannon Communication (NewCom)” under the
Grant 16KIS1004. We acknowledge support by the German Research Foundation (DFG) and the
Open Access Publication (OAP) Fund of TU Berlin.
**Institutional Review Board Statement: Not applicable.**
**Informed Consent Statement: Not applicable.**
-----
_Entropy 2021, 23, 16_ 22 of 23
**Data Availability Statement: Not applicable.**
**Conflicts of Interest: The funders had no role in the design of the study; in the collection, analyses,**
or interpretation of data; in the writing of the manuscript, or in the decision to publish the results.
The authors declare no conflict of interest.
**References**
1. [Shannon, C.E. Communication theory of secrecy systems. Bell Syst. Tech. J. 1949, 28, 656–715. [CrossRef]](http://dx.doi.org/10.1002/j.1538-7305.1949.tb00928.x)
2. Kahn, D. The Codebreakers: The Story of Secret Writing; Macmillan Publishers: New York, NY, USA, 1967.
3. Schneier, B. Applied Cryptography: Protocols, Algorithms, and Source Code in C, 2nd ed.; John Wiley & Sons: Hoboken, NJ, USA, 1996.
4. Böhm, C.; Hofer, M. Physical Unclonable Functions in Theory and Practice; Springer: New York, NY, USA, 2012.
5. Gassend, B.; Clarke, D.; Dijk, M.V.; Devadas, S. Silicon physical random functions. In Proceedings of the 9th ACM Conference on
Computer and Communications Security, Washington, DC, USA, 18–22 November 2002; pp. 148–160.
6. Maes, R.; Tuyls, P.; Verbauwhede, I. A Soft Decision Helper Data Algorithm for SRAM PUFs. In Proceedings of the 2009 IEEE
International Symposium on Information Theory, Seoul, Korea, 28 June–3 July 2009; pp. 2101–2105.
7. Goldreich, O. Modern Cryptography, Probabilistic Proofs and Pseudorandomness; Springer: Berlin Heidelberg, Germany, 1998;
Volume 17.
8. Pappu, R. Physical One-Way Functions. Ph.D. Thesis, Massachusetts Institute of Technology, Cambridge, MA, USA, 2001.
9. [Wyner, A.D. The wire-tap channel. Bell Syst. Tech. J. 1975, 54, 1355–1387. [CrossRef]](http://dx.doi.org/10.1002/j.1538-7305.1975.tb02040.x)
10. Palanca, A.; Evenchick, E.; Maggi, F.; Zanero, S. A stealth, selective, link-layer denial-of-service attack against automotive
networks. In Detection of Intrusions and Malware, and Vulnerability Assessment; Springer: Cham, Switzerland, 2017; pp. 185–206.
11. Lee, Y.S.; Lee, H.J.; Alasaarela, E. Mutual authentication in wireless body sensor networks (WBSN) based on Physical Unclonable
Function (PUF). In Proceedings of the 2013 9th International Wireless Communications and Mobile Computing Conference
(IWCMC), Sardinia, Italy, 1–5 July 2013; pp. 1314–1318.
12. Simpson, E.; Schaumont, P. Offline hardware/software authentication for reconfigurable platforms. In International Workshop on
_Cryptographic Hardware and Embedded Systems; Springer: Berlin/Heidelberg, Germany, 2006; pp. 311–323._
13. Herder, C.; Yu, M.; Koushanfar, F.; Devadas, S. Physical Unclonable Functions and Applications: A Tutorial. Proc. IEEE 2014,
_[102, 1126–1141. [CrossRef]](http://dx.doi.org/10.1109/JPROC.2014.2320516)_
14. Huth, C.; Guillaume, R.; Strohm, T.; Duplys, P.; Samuel, I.A.; Güneysu, T. Information reconciliation schemes in physical-layer
[security: A survey. Comput. Netw. 2016, 109, 84–104. [CrossRef]](http://dx.doi.org/10.1016/j.comnet.2016.06.014)
15. Lim, D.; Lee, J.W.; Gassend, B.; Suh, G.E.; Dijk, M.V.; Devadas, S. Extracting Secret Keys From Integrated Circuits. IEEE Trans.
_Very Large Scale Integr. Syst. 2005, 13, 1200–1205._
16. Mandal, M.K.; Sarkar, B.C. Ring oscillators: Characteristics and Applications. Indian J. Pure Appl. Phys. 2010, 48, 136–145.
17. Suh, G.E.; Devadas, S. Physical Unclonable Functions for Device Authentication and Secret Key Generation. In Proceedings of
the 2007 44th ACM/IEEE Design Automation Conference, San Diego, CA, USA, 4–8 June 2007; pp. 9–14.
18. Guajardo, J.; Kumar, S.S.; Schrijen, G.J.; Tuyls, P. FPGA Intrinsic PUFs and Their Use for IP Protection. In International Workshop on
_Cryptographic Hardware and Embedded Systems; Springer: Berlin/Heidelberg, Germany, 2007; pp. 63–80._
19. Günlü, O.; Kramer, G.; Skorski, M. Privacy and secrecy with multiple measurements of physical and biometric identifiers. In Proceedings of the 2015 IEEE Conference on Communications and Network Security (CNS), Florence, Italy, 28–30 September 2015;
pp. 89–94.
20. Dodis, Y.; Ostrovsky, R.; Reyzin, L.; Smith, A. Fuzzy extractors: How to generate strong keys from biometrics and other noisy
[data. SIAM J. Comput. 2008, 38, 97–139. [CrossRef]](http://dx.doi.org/10.1137/060651380)
21. Juels, A.; Wattenberg, M. A fuzzy commitment scheme. In Proceedings of the 6th ACM Conference on Computer and Communications Security, Singapore, 1–4 November 1999; pp. 28–36.
22. Ignatenko, T.; Willems, F.M.J. Biometric systems: Privacy and secrecy aspects. IEEE Trans. Inf. Forensics Secur. 2009, 4, 956–973.
[[CrossRef]](http://dx.doi.org/10.1109/TIFS.2009.2033228)
23. Günlü, O.; Kramer, G. Privacy, Secrecy, and Storage With Multiple Noisy Measurements of Identifiers. IEEE Trans. Inf. Forensics
_[Secur. 2018, 13, 2872–2883. [CrossRef]](http://dx.doi.org/10.1109/TIFS.2018.2834303)_
24. Günlü, O.; I¸scan, O. DCT Based Ring Oscillator Physical Unclonable Functions. In Proceedings of the 2014 IEEE International[˙]
Conference on Acoustics, Speech and Signal Processing (ICASSP), Florence, Italy, 4–9 May 2014; pp. 8198–8201.
25. Maiti, A.; Schaumont, P. Improved ring oscillator PUF: An FPGA-friendly secure primitive. J. Cryptol. 2011, 24, 375–397.
[[CrossRef]](http://dx.doi.org/10.1007/s00145-010-9088-4)
26. Ahlswede, R.; Csiszár, I. Common Randomness in Information Theory and Cryptography—Part I: Secret Sharing. IEEE Trans. Inf.
_[Theory 1993, 39, 1121–1132. [CrossRef]](http://dx.doi.org/10.1109/18.243431)_
27. Maurer, U.M. Secret Key Agreement by Public Discussion from Common Information. IEEE Trans. Inf. Theory 1993, 39, 733–742.
[[CrossRef]](http://dx.doi.org/10.1109/18.256484)
28. Günlü, O.; Schaefer, R.F.; Poor, H.V. Biometric and physical identifiers with correlated noise for controllable private authentication.
_arXiv 2020, arXiv:2001.00847._
29. Günlü, O.; Schaefer, R.F.; Kramer, G. Private authentication with physical identifiers through broadcast channel measurements.
In Proceedings of the 2019 IEEE Information Theory Workshop (ITW), Visby, Sweden, 25–28 August 2019; pp. 1–5.
-----
_Entropy 2021, 23, 16_ 23 of 23
30. Ignatenko, T.; Willems, F.M. Information leakage in fuzzy commitment schemes. IEEE Trans. Inf. Forensics Secur. 2010, 5,
[2337–2348. [CrossRef]](http://dx.doi.org/10.1109/TIFS.2010.2046984)
31. Cover, T.M.; Thomas, J.A. Elements of Information Theory; John Wiley & Sons: Hoboken, NJ, USA, 2012.
32. Günlü, O.; Kernetzky, T.; I¸scan, O.; Sidorenko, V.; Kramer, G.; Schaefer, R.F. Secure and Reliable Key Agreement with Physical[˙]
[Unclonable Functions. Entropy 2018, 20, 340. [CrossRef]](http://dx.doi.org/10.3390/e20050340)
33. Maiti, A.; Casarona, J.; McHale, L.; Schaumont, P. A Large Scale Characterization of RO-PUF. In Proceedings of the 2010 IEEE
International Symposium on Hardware-Oriented Security and Trust (HOST), Anaheim, CA, USA, 13–14 June 2010; pp. 94–99.
34. [Schwarz, G. Estimating the dimension of a model. Ann. Stat. 1978, 6, 461–464. [CrossRef]](http://dx.doi.org/10.1214/aos/1176344136)
35. Sugiura, N. Further analysis of the data by Akaike’s information criterion and the finite corrections. Commun. Stat. Theory Methods
**[1978, 7, 13–26. [CrossRef]](http://dx.doi.org/10.1080/03610927808827599)**
36. Günlü, O.; I¸scan, O.; Kramer, G. Reliable secret key generation from physical unclonable functions under varying environmental[˙]
conditions. In Proceedings of the 2015 IEEE International Workshop on Information Forensics and Security (WIFS), Rome, Italy,
16–19 November 2015; pp. 1–6.
37. Lin, S.; Costello, D.J. Error Control Coding; Prentice-Hall: Englewood Cliffs, NJ, USA, 2004.
38. Ohm, J.R. Multimedia Signal Coding and Transmission; Springer: Berlin/Heidelberg, Germany, 2015.
39. Wang, R. Introduction to Orthogonal Transforms: With Applications in Data Processing and Analysis; Cambridge University Press:
Cambridge, UK, 2012.
40. Günlü, O.; Schaefer, R.F. Low-Complexity and Reliable Transforms for Physical Unclonable Functions. In Proceedings of the
2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Barcelona, Spain, 4–8 May 2020;
pp. 2807–2811.
41. Maes, R.; Herrewege, A.V.; Verbauwhede, I. PUFKY: A fully functional PUF-based cryptographic key generator. In Cryptographic
_Hardware Embedded Systems; Springer: Berlin/Heidelberg, Germany, 2012; pp. 302–319._
42. Rukhin, A.; Soto, J.; Nechvatal, J.; Smid, M.; Barker, E. A Statistical Test Suite for Random and Pseudorandom Number Generators for
_Cryptographic Applications; Technical Report; The National Institute of Standards and Technology: Gaithersburg, MD, USA, 2001._
43. Hong, Y. On Computing the Distribution Function for the Sum of Independent and Nonidentical Random Indicators; Technical Report;
Department of Statistics, Virginia Tech.: Blacksburg, VA, USA, 2011.
44. Polyanskiy, Y.; Poor, H.V.; Verdú, S. Channel Coding Rate in the Finite Blocklength Regime. IEEE Trans. Inf. Theory 2010,
_[56, 2307–2359. [CrossRef]](http://dx.doi.org/10.1109/TIT.2010.2043769)_
45. Günlü, O.; I¸scan, O.; Sidorenko, V.; Kramer, G. Code Constructions for Physical Unclonable Functions and Biometric Secrecy[˙]
[Systems. IEEE Trans. Inf. Forensics Secur. 2019, 14, 2848–2858. [CrossRef]](http://dx.doi.org/10.1109/TIFS.2019.2911155)
46. Wyner, A.D.; Ziv, J. The rate-distortion function for source coding with side information at the decoder. IEEE Trans. Inf. Theory
**[1976, 22, 1–10. [CrossRef]](http://dx.doi.org/10.1109/TIT.1976.1055508)**
47. Chen, B.; Ignatenko, T.; Willems, F.M.; Maes, R.; van der Sluis, E.; Selimis, G. A Robust SRAM-PUF Key Generation Scheme
Based on Polar Codes. In Proceedings of the GLOBECOM 2017—2017 IEEE Global Communications Conference, Singapore,
4–8 December 2017; pp. 1–6.
48. [Slepian, D.; Wolf, J. Noiseless coding of correlated information sources. IEEE Trans. Inf. Theory 1973, 19, 471–480. [CrossRef]](http://dx.doi.org/10.1109/TIT.1973.1055037)
49. Maurer, U.; Wolf, S. Information-theoretic key agreement: From weak to strong secrecy for free. In Proceedings of the International
Conference on the Theory and Applications of Cryptographic Techniques, Bruges, Belgium, 14–18 May 2000; pp. 351–368.
50. Arikan, E. Channel Polarization: A Method for Constructing Capacity-Achieving Codes for Symmetric Binary-Input Memoryless
[Channels. IEEE Trans. Inf. Theory 2009, 55, 3051–3073. [CrossRef]](http://dx.doi.org/10.1109/TIT.2009.2021379)
51. Korada, S.B.; Urbanke, R.L. Polar Codes are Optimal for Lossy Source Coding. IEEE Trans. Inf. Theory 2010, 56, 1751–1768.
[[CrossRef]](http://dx.doi.org/10.1109/TIT.2010.2040961)
52. Gallager, R.G. Low-Density Parity-Check Codes; M.I.T. Press: Cambridge, MA, USA, 1963.
53. Günlü, O.; Trifonov, P.; Kim, M.; Schaefer, R.F.; Sidorenko, V. Randomized Nested Polar Subcode Constructions for Privacy,
Secrecy, and Storage. arXiv 2020, arXiv:2004.12091.
54. Jerkovits, T.; Günlü, O.; Sidorenko, V.; Kramer, G. Nested Tailbiting Convolutional Codes for Secrecy, Privacy, and Storage. arXiv
**2020, arXiv:2004.13095.**
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/2012.08924, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "GOLD",
"url": "https://www.mdpi.com/1099-4300/23/1/16/pdf?version=1611228187"
}
| 2,020
|
[
"Review",
"JournalArticle"
] | true
| 2020-11-05T00:00:00
|
[
{
"paperId": "eb5d870bd20abde2f7a39e87bf4be306ced452cc",
"title": "Nested Tailbiting Convolutional Codes for Secrecy, Privacy, and Storage"
},
{
"paperId": "6f136d593b88ac58351ee7f92a52b06a3c37b0ae",
"title": "Randomized Nested Polar Subcode Constructions for Privacy, Secrecy, and Storage"
},
{
"paperId": "230498fca7fcd3ce688f45e61e7f705453d6a374",
"title": "Low-Complexity and Reliable Transforms for Physical Unclonable Functions"
},
{
"paperId": "9431185026ea605543f9db8160db28e6b609f7c9",
"title": "Differential privacy for eye tracking with temporal correlations"
},
{
"paperId": "8ccc6a9057e90f0f64c8fa0627dd24ed211b0c60",
"title": "Biometric and Physical Identifiers with Correlated Noise for Controllable Private Authentication"
},
{
"paperId": "3a0331c3071c0f76eeed27845533612068a9b5f8",
"title": "Private Authentication with Physical Identifiers Through Broadcast Channel Measurements"
},
{
"paperId": "be80f1e19aaf3c1dcc9c070a7bb728dc165be903",
"title": "Secure and Reliable Key Agreement with Physical Unclonable Functions †"
},
{
"paperId": "75847bb8a5932607b8b360b28bf748d7bead8bb8",
"title": "Code Constructions for Physical Unclonable Functions and Biometric Secrecy Systems"
},
{
"paperId": "38415e0dc8401a79882c28bf6c23d32eec3231b0",
"title": "A Stealth, Selective, Link-Layer Denial-of-Service Attack Against Automotive Networks"
},
{
"paperId": "219b07d9eba79624c8cf5f782ce96e4812197c5e",
"title": "A Robust SRAM-PUF Key Generation Scheme Based on Polar Codes"
},
{
"paperId": "e665bd5eb45d04e2c3e3f255926f14c4cba96a86",
"title": "Information reconciliation schemes in physical-layer security: A survey"
},
{
"paperId": "6e8637f4fb7a7ea9f2ef733e26fe04b9903c6c1c",
"title": "Privacy, Secrecy, and Storage With Multiple Noisy Measurements of Identifiers"
},
{
"paperId": "8436e46720eb5df67b20900e78479083402bbe1c",
"title": "Privacy and secrecy with multiple measurements of physical and biometric identifiers"
},
{
"paperId": "0f8717c1c19827dd54cb196b9aee54d876005a2c",
"title": "Reliable secret key generation from physical unclonable functions under varying environmental conditions"
},
{
"paperId": "a1f389c111d8bbeae315b05eba2487bbbf16ceb2",
"title": "Multimedia Signal Coding and Transmission"
},
{
"paperId": "a1da194a3f03d7d7216b03ee5ff408d6e070f4b3",
"title": "Physical Unclonable Functions and Applications: A Tutorial"
},
{
"paperId": "4e58acf47d72b4adca677e81b2ee1fe8bb50e3f0",
"title": "DCT based ring oscillator Physical Unclonable Functions"
},
{
"paperId": "3ac8bc8affff2e004898d2528ba16d6488557774",
"title": "Mutual authentication in wireless body sensor networks (WBSN) based on Physical Unclonable Function (PUF)"
},
{
"paperId": "c5dca6736779ab2e7a94735533c4a9c138568914",
"title": "Physical Unclonable Functions in Theory and Practice"
},
{
"paperId": "03c857ecdb86724133b2a473ce5d99f34a4e83c1",
"title": "PUFKY: A Fully Functional PUF-Based Cryptographic Key Generator"
},
{
"paperId": "058e4b42459c2703a2c915036242d12b939c22b9",
"title": "Introduction to Orthogonal Transforms: With Applications in Data Processing and Analysis"
},
{
"paperId": "df5cac0129326f07ff1a64cb1a1646acd500d174",
"title": "Improved Ring Oscillator PUF: An FPGA-friendly Secure Primitive"
},
{
"paperId": "03c4cc2526ce04c7f0532e6d3ec5556c03957c51",
"title": "A large scale characterization of RO-PUF"
},
{
"paperId": "14db5217b7da79c4365df56ae91a75344cd0fa2c",
"title": "Information Leakage in Fuzzy Commitment Schemes"
},
{
"paperId": "9d4e83db13ea001165d8968a3e869e8716c78377",
"title": "Channel Coding Rate in the Finite Blocklength Regime"
},
{
"paperId": "c1f27dd01a563a82646e295824ab449511c69034",
"title": "Ring oscillators: Characteristics and applications"
},
{
"paperId": "e49181d1f312f8a0a9689005c0a3321116e9a1ec",
"title": "Biometric Systems: Privacy and Secrecy Aspects"
},
{
"paperId": "d66b3ea2f3538675e037f863a1e56968f3ed5ca9",
"title": "A soft decision helper data algorithm for SRAM PUFs"
},
{
"paperId": "3418666b85e8d0c58c364900c7beedada91fbbe4",
"title": "Polar codes are optimal for lossy source coding"
},
{
"paperId": "62f9b3c57d73092f2125e9be5568dd1aa0a06d18",
"title": "Channel Polarization: A Method for Constructing Capacity-Achieving Codes for Symmetric Binary-Input Memoryless Channels"
},
{
"paperId": "bacf21343fa323ca220cc70e1ec06e782ecbe2e2",
"title": "FPGA Intrinsic PUFs and Their Use for IP Protection"
},
{
"paperId": "852839104a55cbedffda3a6b8650c108ae2da4dd",
"title": "Physical Unclonable Functions for Device Authentication and Secret Key Generation"
},
{
"paperId": "b934b5a983562d396ff6a2e722045ddc0c5ec80b",
"title": "Offline Hardware/Software Authentication for Reconfigurable Platforms"
},
{
"paperId": "fe195a79c748c2f883e4eebef4402dc6c612399b",
"title": "Extracting secret keys from integrated circuits"
},
{
"paperId": "7dbdb4209626fd92d2436a058663206216036e68",
"title": "Elements of Information Theory"
},
{
"paperId": "f9000e8e1e1851a3b95dcc3c3b02cd3b2788ab11",
"title": "Error Control Coding"
},
{
"paperId": "5f00233479dd7d08475f708c911c5b3999a78d8f",
"title": "Fuzzy Extractors: How to Generate Strong Keys from Biometrics and Other Noisy Data"
},
{
"paperId": "5e310160e95d9f7d92369725b27da453d3e5365b",
"title": "Physical One-Way Functions"
},
{
"paperId": "d3f002dd0d6f56bc3bdf9493f734f1abe5a999d0",
"title": "A Statistical Test Suite for Random and Pseudorandom Number Generators for Cryptographic Applications | NIST"
},
{
"paperId": "961fc0f6c68303cde94a28fdb24e6eaf3ef040b4",
"title": "A Statistical Test Suite for Random and Pseudorandom Number Generators for Cryptographic Applications"
},
{
"paperId": "209ec57396225554522738b10ad05cbf2d274fe7",
"title": "Information-Theoretic Key Agreement: From Weak to Strong Secrecy for Free"
},
{
"paperId": "3cc379e3eb3db1126f5cf6f0055be911159ffe5d",
"title": "A fuzzy commitment scheme"
},
{
"paperId": "6e7ae7f8d422ae8b7ef25ddadfb39d1abf46a85f",
"title": "Modern Cryptography, Probabilistic Proofs and Pseudorandomness"
},
{
"paperId": "8b55961e7be0ac5deb5f0c304ab0c1de1a827621",
"title": "Applied Cryptography: Protocols, Algorithms, and Source Code in C"
},
{
"paperId": "5553970601d94c090eb2d3431b653106bf271c4c",
"title": "Common randomness in information theory and cryptography - I: Secret sharing"
},
{
"paperId": "68f8fafd79445110b91c0f24480fbd8da160c31c",
"title": "Secret key agreement by public discussion from common information"
},
{
"paperId": "37e44d1de8003d8394d158ec6afd1ff0e87e595b",
"title": "Estimating the Dimension of a Model"
},
{
"paperId": "c02bae0c9e68c4417d9d6226790c9a1366991494",
"title": "The wire-tap channel"
},
{
"paperId": "71c969f0a7204259cdd394283ab02d0fc34e5fc2",
"title": "Noiseless coding of correlated information sources"
},
{
"paperId": "86ba7660ed50b2be9ccd801070680152c5646b55",
"title": "The codebreakers : the story of secret writing"
},
{
"paperId": "e073a7c5a6418d96fc16d8337a6056a457e75c1e",
"title": "Communication theory of secrecy systems"
},
{
"paperId": "b1304465c267596576a151c91ac948d225a28c6b",
"title": "Key Agreement with Physical Unclonable Functions and Biometric Identifiers"
},
{
"paperId": null,
"title": "On computing the distribution function for the sum of independent and nonidentical random indicators"
},
{
"paperId": "c80460c741362038b3e366313414b2edcaa015a4",
"title": "Physical one-way functions"
},
{
"paperId": "d389f72307f2748c7a5169ebc15872818b139c16",
"title": "A STATISTICAL TEST SUITE FOR RANDOM AND PSEUDORANDOM NUMBER GENERATORS FOR CRYPTOGRAPHIC APPLICATIONS"
},
{
"paperId": "ce133305355a373843656145dfcd6a4563d75520",
"title": "Applied cryptography - protocols, algorithms, and source code in C, 2nd Edition"
},
{
"paperId": "0b3995e33daf8fea7a772585ff7dac925af7fae9",
"title": "Further analysts of the data by akaike' s information criterion and the finite corrections"
},
{
"paperId": "d2607df2b01d1123cae2716b876776510c41f836",
"title": "The rate-distortion function for source coding with side information at the decoder"
},
{
"paperId": "206f827fad201506c315d40c1469b41a45141893",
"title": "Low-density parity-check codes"
},
{
"paperId": "d8548b8fce3861732afdcc83213a8a3cbf27b83c",
"title": "Silicon Physical Random Functions (cid:3)"
}
] | 24,444
|
en
|
[
{
"category": "Medicine",
"source": "external"
},
{
"category": "Medicine",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/029062eca1fb9a279c20694127fcf1c8313966ff
|
[
"Medicine"
] | 0.809361
|
Exome/Genome-Wide Testing in Newborn Screening: A Proportionate Path Forward
|
029062eca1fb9a279c20694127fcf1c8313966ff
|
Frontiers in Genetics
|
[
{
"authorId": "4855327",
"name": "V. Rahimzadeh"
},
{
"authorId": "2121443177",
"name": "J. Friedman"
},
{
"authorId": "51893879",
"name": "G. de Wert"
},
{
"authorId": "3148841",
"name": "B. Knoppers"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"Front Genet"
],
"alternate_urls": [
"http://journal.frontiersin.org/journal/genetics",
"https://www.frontiersin.org/journals/genetics"
],
"id": "9ec189b3-db0f-41d8-9c82-21ef5fa9b87e",
"issn": "1664-8021",
"name": "Frontiers in Genetics",
"type": "journal",
"url": "http://www.frontiersin.org/genetics/"
}
|
Population-based newborn screening (NBS) is among the most effective public health programs ever launched, improving health outcomes for newborns who screen positive worldwide through early detection and clinical intervention for genetic disorders discovered in the earliest hours of life. Key to the success of newborn screening programs has been near universal accessibility and participation. Interest has been building to expand newborn screening programs to also include many rare genetic diseases that can now be identified by exome or genome sequencing (ES/GS). Significant declines in sequencing costs as well as improvements to sequencing technologies have enabled researchers to elucidate novel gene-disease associations that motivate possible expansion of newborn screening programs. In this paper we consider recommendations from professional genetic societies in Europe and North America in light of scientific advances in ES/GS and our current understanding of the limitations of ES/GS approaches in the NBS context. We invoke the principle of proportionality—that benefits clearly outweigh associated risks—and the human right to benefit from science to argue that rigorous evidence is still needed for ES/GS that demonstrates clinical utility, accurate genomic variant interpretation, cost effectiveness and universal accessibility of testing and necessary follow-up care and treatment. Confirmatory or second-tier testing using ES/GS may be appropriate as an adjunct to conventional newborn screening in some circumstances. Such cases could serve as important testbeds from which to gather data on relevant programmatic barriers and facilitators to wider ES/GS implementation.
|
Edited by:
Laura V. Milko,
University of North Carolina at Chapel
Hill, United States
Reviewed by:
Milan Macek,
Charles University, Czechia
Jonathan Berg,
University of North Carolina at Chapel
Hill, United States
*Correspondence:
Vasiliki Rahimzadeh
[vrahim@stanford.edu](mailto:vrahim@stanford.edu)
Specialty section:
This article was submitted to
Human and Medical Genomics,
a section of the journal
Frontiers in Genetics
Received: 29 January 2022
Accepted: 27 May 2022
Published: 04 July 2022
Citation:
Rahimzadeh V, Friedman JM,
de Wert G and Knoppers BM (2022)
Exome/Genome-Wide Testing in
Newborn Screening: A Proportionate
Path Forward.
Front. Genet. 13:865400.
[doi: 10.3389/fgene.2022.865400](https://doi.org/10.3389/fgene.2022.865400)
p y
[doi: 10.3389/fgene.2022.865400](https://doi.org/10.3389/fgene.2022.865400)
# Exome/Genome-Wide Testing in Newborn Screening: A Proportionate Path Forward
Vasiliki Rahimzadeh [1]*, Jan M. Friedman [2], Guido de Wert [3] and Bartha M. Knoppers [4]
1Stanford Center for Biomedical Ethics, Stanford University, Stanford, CA, United States, 2Department of Medical Genetics,
University of British Columbia, Vancouver, BC, Canada, [3]Department of Health, Ethics and Society, Maastricht University,
Maastricht, Netherlands, [4]Centre of Genomics and Policy, McGill University, Montreal, QC, Canada
### Population-based newborn screening (NBS) is among the most effective public health programs ever launched, improving health outcomes for newborns who screen positive worldwide through early detection and clinical intervention for genetic disorders discovered in the earliest hours of life. Key to the success of newborn screening programs has been near universal accessibility and participation. Interest has been building to expand newborn screening programs to also include many rare genetic diseases that can now be identified by exome or genome sequencing (ES/GS). Significant declines in sequencing costs as well as improvements to sequencing technologies have enabled researchers to elucidate novel gene-disease associations that motivate possible expansion of newborn screening programs. In this paper we consider recommendations from professional genetic societies in Europe and North America in light of scientific advances in ES/GS and our current understanding of the limitations of ES/GS approaches in the NBS context. We invoke the principle of proportionality—that benefits clearly outweigh associated risks—and the human right to benefit from science to argue that rigorous evidence is still needed for ES/GS that demonstrates clinical utility, accurate genomic variant interpretation, cost effectiveness and universal accessibility of testing and necessary follow-up care and treatment. Confirmatory or second-tier testing using ES/GS may be appropriate as an adjunct to conventional newborn screening in some circumstances. Such cases could serve as important testbeds from which to gather data on relevant programmatic barriers and facilitators to wider ES/GS implementation.
Keywords: exome sequencing, genome sequecing, newborn screening, population health genomics, access, public
health ethics
## INTRODUCTION
Population-based newborn screening (NBS) is among the most effective public health programs ever
launched (Tonniges, 2000; Sahai and Marsden, 2009; Berry, 2015). Updated national estimates in the
United States suggest nearly 12,900 newborns screened positive for childhood onset disorders that
previously led to severe morbidity or mortality and were listed on the Recommended Universal
Screening Panel (RUSP) (5) between 2015 and 2017 (Sontag et al., 2020). Key to the success of NBS
programs has been their affordability and near universal access and participation. Pre-symptomatic
treatment of newborns who screen positive for some of these conditions is much more cost-effective
-----
and less burdensome on healthcare systems than treating the
conditions once they become symptomatic (Carroll and Downs,
2006). Preventing the development of symptomatic disease is a
particularly important consideration with respect to genetic
diseases that can be detected by ES/GS analysis because most
do not have specific treatments that can prevent disease onset or
progression.
Since early validation studies of mass screening tests for
metabolic disorders in the 1960s (McCandless and Wright,
2020), NBS methods as well as their formal adoption and
oversight have evolved considerably. Interest has been building
to expand NBS programs to also include more rare genetic
diseases that can be identified using ES/GS approaches (Holm
et al., 2018; Genomics England and the UK National Screening
Committee, 2021; Gold et al., 2022; Lu et al., 2022).
Improvements to genome sequencing technologies that enable
researchers to elucidate novel gene-disease associations and to
diagnose conditions undiscoverable using traditional biochemical
or other biomarker testing, and the wide availability and declining
costs of genomic testing are among the reasons ES/GS might be
advantageous as a first-tier clinical test for diagnosing genetic
diseases.
At the outset, it is important to distinguish NBS meant to
identify pre-symptomatic infants rare but potentially devastating
conditions e.g., phenylketonuria (PKU), severe combined
immunodeficiency disease of congenital heart defects, from
screening for risk stratification meant to guide lifestyle
modification or surveillance protocols routinely offered to
adults. Current universal NBS protocols fall into the first
category; ES/GS of newborn infants for most genetic diseases
would fall into the second category. This is true whether one
considers all known genetic diseases or only a subset in which
non-specific interventions may be able to reduce the risk or age of
symptomatic onset.
Using ES/GS as a tool in NBS may also inappropriately
conflate the recognition of a disease-associated genetic variant
with diagnosis of the disease. Diagnosing a genetic disease
requires a physician to interpret an ES/GS result in the
context of an individual’s complete clinical picture–the
medical history, family history, physical exam, and other
laboratory and imaging studies–in light of what is known
about the range of clinical manifestations, inheritance pattern,
penetrance, and variability of the disease. Complete clinical
assessment is the only confirmatory “test” available for most
genetic diseases. If universal NBS relied on sequencing the entire
genome, exome or specific regions of the exome, then complete
clinical assessment for the genetic disease indicated would be
necessary to confirm the molecular “diagnosis” in every case.
Population-based NBS of any kind should only be offered as part
of a comprehensive public health program that includes clinical
follow-up, therapeutic interventions, quality assurance,
governance and oversight, and public and professional
education (Friedman et al., 2017) in addition to the
confirmatory complete clinical assessment and genetic
counselling (if the condition found is a genetic disease). If ES/
GS is being considered as a replacement for current NBS,
evidence that the ES/GS methods are superior to the existing
methods is necessary. Adoption of sequencing-based NBS
without consideration of the unique ethical, legal and social
issues it raises (Eichinger et al., 2021; Woerner et al., 2021)
risks widening disparities in availability and access to standard
NBS, particularly in under-resourced settings.
In this paper, we review recommendations from professional
bodies regarding integration of genomic sequencing methods in
public NBS programs in Europe (Howard et al., 2015) and North
America, where the authors are based. We limit our discussion of
relevant ethical, legal and social issues associated with universal
ES/GS as a population screening tool for newborns,
acknowledging, as others do (Johnston et al., 2018), that
different professional obligations and standards exist in clinical
screening, diagnostic, and direct-to-consumer contexts. Our
analysis focuses on applications of universal genomic
sequencing of the genome, exome, or a portion of the exome
that includes a large number of disease-associated genes. We refer
to as “ES/GS,” rather than on targeted sequencing of one or a few
genes for confirmatory testing of conditions identified by
conventional NBS (Bhattacharjee et al., 2015).
Indeed, there are compelling advantages for supporting
genomic sequencing method applied in the NBS context.
Genomic sequencing has been shown to detect previously fatal
diseases in affected newborns, as well as provide information to
patients and families about genetic predisposition risks for later
onset diseases (Holm et al., 2018) and inform preventative clinical
action. Scholars have also argued that biological family may
receive ancillary benefits from recognition of disease-associated
variants in an infant by enabling prenatal diagnosis or specialized
care for future pregnancies, earlier diagnosis or prevention of
disease in relatives, or the empowerment provided by better
knowledge (Ceyhan-Birsoy et al., 2019; Biesecker et al., 2021).
However, the “gap between what sequencing results can reveal
and the kinds of information most people need to improve their
health, combined with widely publicized hopes for the
revolutionary power of genomics, creates the very real risk
that patients, research participants, health care professionals,
policy-makers, and others may have unrealistic expectations of
what sequencing can achieve and little appreciation for its
downsides” (Johnston et al., 2018).
Public opinion research suggests that family preferences vary
considerably regarding whether and how to return genomic
sequencing results (Lipstein et al., 2010; Fernandez et al., 2014;
Botkin et al., 2015; Joseph et al., 2016; Pereira et al., 2021), to say
nothing of current shortages of genetic counsellors and genetic
specialist physicians needed or enhancements to genomic literacy
and education for health professionals and the general public
should ES/GS become routine in NBS (Lewis et al., 2016). Key
policy questions also remain unresolved. These include: What
rights and protections apply for genomic and related health data
involving newborns when they become adults? How will public
health agencies ensure that appropriate infrastructures for
sequencing, variant interpretation, diagnostic confirmation,
treatment or non-medical interventions, genetic counselling,
clinical follow-up, and program governance and quality
assurance are in place and accessible to all infants, even those
in under-resourced settings? And whether requirements for
-----
explicit informed consent to ES/GS-based NBS would need to
obtained from the parents and, if so, should it include permission
for others (researchers, family members, police, etc.) to access
stored newborn sequencing data in the future.
We assess these questions by evaluating the proposed benefits
and foreseeable risks of implementing ES/GS in NBS. In our
analysis, we apply the principle of proportionality to our
discussion—that benefits of sequencing should clearly
outweigh associated risks—and consider the human right to
benefit from science -especially that of the asymptomatic, atrisk newborn to be found. We conclude that routine universal ES/
GS implementation is not justified at the present time, even if the
analysis is restricted to a subset of disease-associated genes.
Stronger evidence is needed to establish the clinical utility of
ES/GS, accurate genomic variant interpretation, and cost
effectiveness for newborn screening, as well as policies
ensuring universal access and equitable resourcing for not only
the testing but also for comprehensive diagnostic confirmation,
treatment, genetic counselling, and clinical follow-up of affected
patients. Moreover, this evidence should demonstrate the
population health benefits of universal ES/GS-wide screening
of newborns and not simply that anticipated harms of
incorporating ES/GS are minimal. Prioritizing expanded access
over expanded testing is likely to lead to more equitable
distribution of the public health benefits of newborn screening
programs.
## PRINCIPLE OF PROPORTIONALITY
The principle of proportionality suggests an intervention may be
ethically permissible if its anticipated benefits on balance justify
exposure to associated harms and hence a helpful framework with
which to assess ES/GS-based screening (Sénécal et al., 2018). The
principle is rooted in moral and legal theory of punishment. 17th
Century constitutional law theorists, for example, invoked the
principle to judge the statutory fairness between restrictions
imposed to implement a corrective measure and the severity
of the act(s) the measure purports to mitigate (Walen, 2021). In
research, the proportionality principle underpins decisions
institutional/ethics review boards make regarding the relative
risks and benefits of a study to prospective participants and is
subsequently codified in national human subjects research
regulations (OHRP, 2017; Canadian Institutes of Health
Research, 2018) and international biomedical research norms
(Council for International Organizations of Medical Sciences
(CIOMS) in collaboration with the World Health
Organization, 2016; WMA, 2022). It has also been has more
recently been applied to guide privacy protections when sharing
genomic and related health data (Wright et al., 2016).
And last, but not least, some more recent versions of the
normative framework for screening add the principle of
proportionality as a central, over-arching, screening criterion:
“The overall benefits of screening should outweigh the harm”
(Andermann et al., 2008; Health Council of the Netherlands,
2008). The appeal of the proportionality principle to the NBS
debate is astutely summarized by Kalkman and Dondorp in their
position against screening newborns for non-treatable
conditions: “the dividing line in the debate is . . . whether such
screening should be regarded as catering to a parental “right to
know,” or as a public health service that should be subject to
standards of evidence and proportionality” (Kalkman and
Dondorp, 2022).
## The Benefits of Accurate and Timely Diagnosis
New precision methods to detect disease-causing genetic variants
have greatly improved (Dondorp and de Wert, 2013). ES/GS
could identify infants with rare genetic diseases not currently
recognized using standard NBS. In theory, newborns who screen
positive by ES/GS have the potential to benefit from: early
diagnosis; disease onset prevention using available approaches;
opportunities for genetic counselling for their families; eligibility
for participation in clinical trials or other research studies; and
avoiding long and difficult diagnostic odysseys.
ES/GS should not, in our view, replace standard methods for
any disease screening unless the former has been shown to have
better sensitivity and specificity for the disease. For conditions
that are not included in current NBS programs, development and
uniform adoption of an approach will be needed to select the
conditions for which ES/GS are expected to provide tangible
benefit to the newborn. An exome- or genome-wide analysis that
generates more harms than benefits or for which the harms and
benefits have not been established is ethically unjustifiable–a
more targeted analysis is to be preferred; see for example
(Milko et al., 2019). But agreement on a uniform approach for
selecting conditions detectable only using ES/GS is proving
elusive for NBS programs worldwide (Jansen et al., 2017).
Assuming agreement on the approach were achieved, the
question would become whether every disease gene that we
look for using ES/GS must meet the same criteria required to
add conditions to the RUSP.
The benefit-harm calculus is further complicated by the type
of disorder being screened. One significant challenge facing
public health decision-makers and clinicians alike is
determining when to add conditions to the RUSP that are
identifiable only through ES/GS methods. For diseases for
which standard screening is superior, ES/GS may be
considered as an add-on to current first-tier screening
programs. Findings from a comparison study for example
showed that traditional NBS using tandem mass spectrometry
had greater sensitivity and specificity than ES for the diseases that
are currently being screened, but ES was useful for confirmatory
(Adhikari et al., 2020).
## Screening for Late-Onset Conditions
Debates abound in the literature regarding the ethics of testing
children for conditions likely to present later in life or which may
be clinically relevant for parents or other biological family
members in the immediate term. The presumption of clinical
benefit to the parents and family members, however, has been
challenged (Buchbinder and Timmermans, 2011; Ross and
Clayton, 2019). Screening parents themselves using ES/GS for
-----
previously unrecognized conditions would not only be more
clinically effective but, most importantly, avoids
instrumentalizing the child for parental benefit. We
furthermore object to predictive testing for later-onset
disorders taking account both the harm principle and the
principle of respect for the child’s future right to informational
self-determination, a specification of the child’s proposed right to
an open future (Davis, 1997). Professional guidelines are
consistent with these principles, advocating that publicly
funded, universal NBS should be limited to diseases that can
be diagnosed in the newborn period and which can be effectively
treated or prevented during childhood (de Wert et al., 2021;
Miller et al., 2021). As others have argued, “Providing additional
genomic information beyond the most actionable conditions,
while potentially of interest to many parents, may increase the
complexity of informed consent and thereby serve to distract
from the primary health benefits” (Roman et al., 2020).
Broadening the scope of NBS beyond its primary aim of
detecting rare disorders in asymptomatic children has the
potential to adversely impact the universal delivery of NBS, to
say nothing of the impacts on public trust and widespread
support for NBS.
## Testing Capability and Challenges in Genomic Variant Interpretation
Standard clinical analyses of ES/GS data do not reliably identify
some kinds of disease-causing genetic variants, including short
tandem repeat expansions, mobile element insertions, and
complex or small structural variants. Knowing that ES/GSbased NBS has been done may preclude or delay appropriate
genetic testing for symptomatic genetic disease in an older child
or adult.
Interpretation of NBS results requires extensive knowledge of
benign, as well as disease-causing variants for every gene tested.
The sensitivity and specificity of ES/GS for most rare genetic
diseases are unknown and likely to remain so because sample
sizes are small and studies difficult to power sufficiently. In
addition, the penetrance and phenotypic spectrum associated
with pathogenic variants for most genetic disease loci are
unknown. Thus, it is difficult or impossible to know if an
asymptomatic baby with a “molecular diagnosis” of a rare
genetic disease will ever develop the disease or, in the event
the child does develop the disease, when it will occur or how
severe it will be. Moreover, genetic disease diagnosis is Bayesian.
That is, the probability of finding a pathogenic variant is small in
a healthy newborn with no family history of the genetic disease.
Since there is no primary indication for NBS, the a priori risk that
an infant will develop any particular genetic disease is extremely
small. This makes “positive” results more likely to be false
positives and less likely to be true positives, even if the
analytical validity of the test is very high.
Our inability at the present time to interpret the pathogenicity
of most genomic variants is perhaps the strongest reason against
adopting ES/GS in population-based NBS, despite improvements
to clinical annotation of variants (Amendola et al., 2020) and
broader accessibility to relevant databases at the point of care
(Rehm et al., 2015). The problems of interpretation also
exacerbate the effects of false positives/negatives on families
and the healthcare system that are likely to result if variants of
hundreds or thousands of potential disease genes are analyzed
(Adhikari et al., 2020).
The confidence of variant classification and clinical
interpretation of genetic results will determine their predictive
value. In line with the ethical principle of proportionality,
proponents of ES/GS-based NBS will need to specify
thresholds for what genes and/or variants should be disclosed
in a screening context based on better understanding of
anticipated benefits and harms associated with those decisions.
The general issue remains that ES/GS is currently used as a
diagnostic test, i.e., to confirm a clinical diagnosis of suspected
genetic disease. However, in NBS, ES/GS would be used as a
screening test to identify children who are at high risk of a genetic
disease implied by the “molecular diagnosis.” If ES/GS were
indeed used as a screening test, confirmatory testing to
manage the inevitable false positives must be available. The
distinction between the ES/GS result, regardless of its ACMG
classification, and the actual diagnosis of a disease in the child
would have to be explicit, generally accepted, and universally
understood to avoid stigmatization, discrimination, insurance
coverage, among other social issues.
Interpretation of ES/GS variants requires comparisons to allele
frequencies in both diagnosed and healthy populations and has
direct implications for justice and health equity. This is because
ES/GS interpretation is dependent on genetic ancestry. Variant
interpretation upon which positive predictive values for ES/GS
are measured has been established almost exclusively from
individuals of European descent (Popejoy and Fullerton, 2016;
Peterson et al., 2019). Given such underrepresentation of diverse
ancestries, clinical interpretation of ES/GS results could be less
reliable for newborns of non-European ancestry. Without
adequate representation in datasets from individuals with
diverse genetic ancestry, some newborns will benefit more
from ES/GS than others. Clinical variant interpretation using
resources such as ClinVar (Wain et al., 2018) and gnomAD
(Gudmundsson et al., 2021) is therefore growing in importance,
given they provide clinical assertions about genomic variants and
associations with disease across genetically diverse populations.
In general, problems of underrepresentation have prompted the
development of new tools to monitor trends and identify gaps in
genomic databases (Wang et al., 2022). Indeed, the global catalog
of clinically actionable variants is expected to grow as reference
data sets become larger, better curated and strive to be more
representative of world populations.
## Re-Analysis and Obligations to Update Variant Interpretation
It is anticipated that routine re-analysis of “negative” screens
might increase the diagnostic rate by 3%–5% per year and identify
variants of concern in children who later present with clinical
features suggestive of a genetic disease (Wenger et al., 2017;
Costain et al., 2018). To capture these clinical benefits, NBS
programs would need to systematically update screens and store
-----
ES/GS datasets in the health record to ensure results reflect up-todate classification of genomic variants and take into account
attendant costs and privacy risks. The treating physician may no
longer be following the family and follow up with a new provider
may be difficult and expensive. If a variant of uncertain
significance were reclassified but not reported to the family
based on clinical course, would NBS programs be subject to
legal action if a child later manifests the disease (Clayton et al.,
2021)? The expenditures and risks of storing all children’s
genomic data long-term to enable such systematic re-analysis
may also exceed those of re-sequencing only those children for
whom it is clinically indicated in the future (Veenstra et al., 2021).
## Stigma, Psychological Impacts and Medicalization
Recent studies investigating the psychosocial impacts of
expanding ES/GS in the newborn context have yielded
different results. In a randomized trial of NBS with and
without ES, researchers found both clinicians and parents
valued information gleaned from standard of care NBS more
than from exome sequencing but for different reasons (Pereira
et al., 2019). Parents expressed knowing in advance how to
prepare for a child with special needs was a benefit to
sequencing, but worried about the psychosocial distresses
brought on by variants of unknown significance and potential
for discrimination among other things (Pereira et al., 2019). The
potential for social stigma and medicalization of children with a
molecular diagnosis who are pre-symptomatic (or destined never
to exhibit the disease because it is non-penetrant) is also a
concern. This scenario would be particularly concerning if
enhanced surveillance or prophylactic treatments impinge on
the child’s quality of life or expose them to interventions with
adverse effects.
## Genomic Data Privacy and Protection
Key policy questions persist with respect to what rights and
protections should apply to genomic and related health data
collected at birth when newborns reach adulthood. The moral
justification for mandatory NBS rests on the premise that finding
the asymptomatic, at risk child is within the child’s best interests
(United Nations Convention on the Rights of the Child, 1989).
Child welfare considerations and the “the opportunity to
intervene and dramatically alter a child’s life course and
expectancy has been regarded as sufficient to preempt any
claims of parental autonomy” (Goldenberg and Sharp, 2012).
It is unlikely, however, that the huge volumes of data generated
from ES/GS followed by untargeted whole exome/genome
analysis will meet the criteria needed to justify overruling
parental decision-making authority.
Yet samples taken from dried blood spots collected and stored
using Guthrie cards are rich data sources needed to advance
population health research. While most samples are de-identified
or pseudonymized according to applicable laws/regulations when
used for research, the generation of ES/GS data as part of NBS
introduces novel ethical, legal and social challenges for data
protection, agency and consent for the future adult (Khoury
et al., 2003; Lewis, 2014). Genomic data are highly identifying
and may implicate not only the individual tested but also their
biological relatives. Concerns regarding loss of privacy and
misuse of genomic data have emerged as key themes in the
empirical literature on expansion of sequencing in NBS, and were
found to be especially acute among participants of color (Joseph
et al., 2016; Tsosie et al., 2021). It is unclear if the benefits of
storing children’s genomic data in a centralized research data
repository outweighs the privacy and security risks, particularly if
children are not given the opportunity to consent themselves.
Re-consenting minors when they become adults to the
continued use of their data collected at birth is supported in
theory but logistically challenging to implement in practice
(Knoppers et al., 2016; Rothwell et al., 2017; Nordfalk and
Ekstrøm, 2019). Legislation passed in the United States in
2014, for example, requires that researchers seek broad
consent for the use of the child’s dried blood spots for
research beyond NBS (Newborn Screening Saves Lives
Reauthorization Act, 2014). However this law preceded
revisions to the United States Common Rule which now
exempts research using de-identified data, thus removing a
layer of specific consent (Lewis and Goldenberg, 2015;
Rothwell et al., 2017). Empirical studies involving parents of
both healthy and affected newborns suggest NBS programs
should err on the side of greater transparency in terms of when,
how and for what purposes their child’s samples and data will
be used (Downie et al., 2021). Policy makers would need to
determine whether, or how permissions for future use of ES/
GS data for research will be incorporated into screening, and it
remains unknown what effect this will have on public
willingness to sustain state sponsored NBS programs that
adopt ES/GS.
## ES/GS and the Wilson and Jungner Criteria
Disagreement regarding which disorders are screened for has
largely (though not entirely) been avoided in some
jurisdictions through standardization (Advisory Committee
on Heritable Disorders in Newborns and Children, 2018) and
concerted efforts are ongoing to harmonize screening lists
internationally (Vittozzi et al., 2010; Franková et al., 2021).
Wilson and Jungner anticipated such discrepancies and in
1968, developed criteria that outlined practical principles for
screening services (Box 1) (Wilson and Jungner, 1968). While
there have been recent calls to update the criteria to better
align with technological advances in testing methods (King
et al., 2021) and apply more nuanced decision analysis
approaches (Prosser et al., 2012), the Wilson and Jungner
criteria remain the generally accepted guidelines.
The threat to NBS participation should be a top concern if
conditions are added to mandatory screening that challenge the
Wilson-Jungner criteria or do not reflect how healthcare is
accessed or paid for in a particular jurisdiction. Universal ES/
GS with untargeted analysis in the NBS context poses several
direct challenges to these criteria.
First, while there are many accepted treatments for conditions
commonly screened for, most rare genetic diseases that are
detectable by ES/GS do not have proven therapies.
-----
Second, establishing a clinical diagnosis in an asymptomatic
infant with a “molecular diagnosis” of a rare variant is resourceintensive, requiring specialized clinical assessment and variant
interpretation, additional testing, and counseling services
(Appelbaum et al., 2020). Newborn screening by any method
should be accessible to every infant (Friedman et al., 2017; de
Wert et al., 2021). To meet this universality target, healthcare
centers must be equipped with appropriate sequencing
infrastructure. Both human and material resources will
therefore be needed in addition to those already allocated for
existing NBS programs. At present, ES is available as a diagnostic
tool primarily from certain clinical laboratories and through
direct-to-consumer genetic testing services. A comparison of
community report cards published by the National
Organization for Rare Disorders (National Organization for
Rare Disorders Newborn screening State report card, 2021)
demonstrates that many NBS programs already face various
resource limitations and vast differences exist in screening
availability by U.S. states (Roman et al., 2020).
Disparities in NBS access and quality could be seen to violate
the parens patriae doctrine which upholds that it is the duty of the
State and its courts to protect the interests of persons in situations
of vulnerability, for example children. NBS programs organized
by the State are an extension of this duty (Knoppers, 1992), and
the reasons many jurisdictions adopt an implied consent to NBS.
GS/ES-based NBS may well be different; if explicit consent is
required, extant research suggests families are more likely to
refuse consent, thus inadvertently denying their child the benefits
of current NBS(Bombard et al., 2014; Joseph et al., 2016;
Friedman et al., 2017; Genetti et al., 2019).
Moreover, the right of everyone to benefit from science and
its applications is protected under Article 27 of the United
Nations Declaration of Human Rights. While not a legally
binding agreement, 193 countries have ratified at least one of
the nine core international treaties which codify the
Declaration’s commitments to basic rights and freedoms.
Article 24 of the Convention on the Rights of the Child
further obligates signatories to implement interventions that
reduce infant and child mortality, to provide effective health
care, and to combat childhood disease, among other legally
binding responsibilities. Taken together, international
conventions have been powerful tools for motivating the
development and sustainability of public health programs
(Reinbold, 2019) including NBS. Applying a human rights
frame to the current debate favors expanding access to
established NBS methods that have shown to be clinically
effective, and which enable more children to directly benefit
from proven methods. Ensuring universal access to high
quality NBS irrespective of birthplace, gender and income,
however, continues to be a global challenge (Krotoski et al.,
2009; Borrajo, 2021).
Third, most genetic conditions diagnosed through ES/GS in early
childhood have unknown natural histories or are unrecognizable
during early childhood because the diseases are so rare and have only
been described in a small number of patients.
Fourth, ES/GS is widely misunderstood among patients and
clinicians alike, challenging overall public acceptance as a testing
method. Issues of particular concern include data privacy, family
decision-making when faced with an uncertain result and
possible insurance discrimination (Pereira et al., 2019; Wojcik
et al., 2021).
Fifth, recent analyses of global NBS coverage indicate that cost
remains a barrier to even standard NBS access in low- and
middle-income countries (Therrell et al., 2015, 2020; Howson
et al., 2018; Therrell and Padilla, 2018). Since ES/GS cannot
replace all current NBS by other methods, sequencing computing
and storage costs for genomic data would be needed in addition to
current laboratory costs to mitigate real privacy and security risks.
Studies further show that clinical demand for medical geneticists
and genetic counsellors far exceeds available services (O’Daniel,
2010; Boothe et al., 2021). Ultimately, however, NBS alone cannot
reasonably be expected to universally improve health outcomes
without addressing systemic health disparities, underlying social
determinants of health (Melzer, 2022) and barriers to healthcare
access (Goldstein et al., 2020) experienced predominantly by
marginalized racial/ethnic groups (Sohn and Timmermans,
2019).
## CONCLUSION
Owing to the public health importance of universal access to
NBS, applying ES/GS as screening tools in the newborn context
is unsubstantiated as yet clinically and pragmatically. Ongoing
translational research and technological advances will emerge
in the coming years which are sure to improve our
-----
understanding of the opportunities and limitations of ES/GS in
detecting and preventing early disease. Considering this
evolving evidence, policy makers ought to be persuaded by
a burden a proof that clearly demonstrates superior public
health benefits of ES/GS beyond those achievable through
traditional NBS methods. Attempts to concentrate efforts
only on justifying the minimalness of any anticipated harms
associated with ES/GS in NBS risks sidelining the real ethical,
legal and social issues which have thus far tempered the
promises of precision medicine in general.
Our position thus exposes a central tension in the debate
between providing universal access to traditional NBS and
respecting parents’ decision-making about much more
extensive screening that they may perceive to be in the child’s
best interests but that many adults may not opt for themselves. All
screening programs expose individuals to potential harms that
must be balanced against the benefits anticipated. This is not
unique to genome-wide sequencing-based screening programs
and is true even if only a selected “slice” of genes represented in
the exome data were analyzed. The reality that some infants will
screen positive and never experience symptoms does not justify
excluding possible ES/GS for NBS. Rather the balance of benefits
and harms must be quantified and considered in any policy
decision regarding screening programs to ensure aggregate
benefits outweigh foreseeable aggregate harms. Indeed, NBS
programs must expand to provide all newborns access to
screening that is of proven value, meet established criteria for
## REFERENCES
Adhikari, A. N., Gallagher, R. C., Wang, Y., Currier, R. J., Amatuni, G.,
Bassaganyas, L., et al. (2020). The Role of Exome Sequencing in Newborn
[Screening for Inborn Errors of Metabolism. Nat. Med. 26, 1392–1397. doi:10.](https://doi.org/10.1038/s41591-020-0966-5)
[1038/s41591-020-0966-5](https://doi.org/10.1038/s41591-020-0966-5)
Advisory Committee on Heritable Disorders in Newborns and Children (2018).
Recommended Uniform Screening Panel.
Amendola, L. M., Muenzen, K., Biesecker, L. G., Bowling, K. M., Cooper, G. M.,
Dorschner, M. O., et al. (2020). Variant Classification Concordance Using the
ACMG-AMP Variant Interpretation Guidelines across Nine Genomic
[Implementation Research Studies. Am. J. Hum. Genet. 107, 932–941. doi:10.](https://doi.org/10.1016/j.ajhg.2020.09.011)
[1016/j.ajhg.2020.09.011](https://doi.org/10.1016/j.ajhg.2020.09.011)
Andermann, A., Blancquaert, I., Beauchamp, S., and Déry, V. (2008). Revisting
Wilson and Jungner in the Genomic Age: a Review of Screening Criteria over
[the Past 40 Years. Bull. World Health Organ 86, 317–319. doi:10.2471/blt.07.](https://doi.org/10.2471/blt.07.050112)
[050112](https://doi.org/10.2471/blt.07.050112)
Appelbaum, P. S., Parens, E., Berger, S. M., Chung, W. K., and Burke, W. (2020). Is
There a Duty to Reinterpret Genetic Data? the Ethical Dimensions. Genet. Med.
[22, 633–639. doi:10.1038/s41436-019-0679-7](https://doi.org/10.1038/s41436-019-0679-7)
[Berry, S. A. (2015). Newborn Screening. Clin. Perinatology 42, 441–453. doi:10.](https://doi.org/10.1016/j.clp.2015.03.002)
[1016/j.clp.2015.03.002](https://doi.org/10.1016/j.clp.2015.03.002)
Bhattacharjee, A., Sokolsky, T., Wyman, S. K., Reese, M. G., Puffenberger, E.,
Strauss, K., et al. (2015). Development of DNA Confirmatory and High-Risk
Diagnostic Testing for Newborns Using Targeted Next-Generation DNA
[Sequencing. Genet. Med. 17, 337–347. doi:10.1038/gim.2014.117](https://doi.org/10.1038/gim.2014.117)
Biesecker, L. G., Green, E. D., Manolio, T., Solomon, B. D., and Curtis, D. (2021).
[Should All Babies Have Their Genome Sequenced at Birth? BMJ, n2679. doi:10.](https://doi.org/10.1136/bmj.n2679)
[1136/bmj.n2679](https://doi.org/10.1136/bmj.n2679)
Bombard, Y., Miller, F. A., Hayeems, R. Z., Barg, C., Cressman, C., Carroll, J. C., et al.
(2014). Public Views on Participating in Newborn Screening Using Genome
[Sequencing. Eur. J. Hum. Genet. 22, 1248–1254. doi:10.1038/ejhg.2014.22](https://doi.org/10.1038/ejhg.2014.22)
proportionality (e.g., Wilson-Jungner) and shown to yield greater
and more equitably distributed public health gains.
## DATA AVAILABILITY STATEMENT
The original contributions presented in the study are included in
the article/Supplementary Material, further inquiries can be
directed to the corresponding author.
## AUTHOR CONTRIBUTIONS
All authors conceived of and contributed to the ideas represented
in this paper. Author VR drafted the initial and revised
manuscripts following peer review. Authors JF, GdW, and BK
contributed to both editorial and substantive revisions to earlier
drafts of the manuscript and during peer review. All authors
approved the final version of the manuscript.
## FUNDING
VR received funding for this work from the NIH Division of Loan
Repayment as well as the Stanford Training Program in ELSI
Research (T32HG008953). BK is supported by the Canada
Research Chair in Law and Medicine.
Boothe, E., Greenberg, S., Delaney, C. L., and Cohen, S. A. (2021). Genetic
Counseling Service Delivery Models: A Study of Genetic Counselors’
Interests, Needs, and Barriers to Implementation. Jrnl Gene Coun 30,
[283–292. doi:10.1002/jgc4.1319](https://doi.org/10.1002/jgc4.1319)
Borrajo, G. J. C. (2021). Newborn Screening in Latin America: A Brief Overview of
[the State of the Art. Am. J. Med. Genet. C Semin. Med. Genet. 187, 322. doi:10.](https://doi.org/10.1002/ajmg.c.31899)
[1002/ajmg.c.31899](https://doi.org/10.1002/ajmg.c.31899)
Botkin, J. R., Belmont, J. W., Berg, J. S., Berkman, B. E., Bombard, Y., Holm, I. A.,
et al. (2015). Points to Consider: Ethical, Legal, and Psychosocial Implications
of Genetic Testing in Children and Adolescents. Am. J. Hum. Genet. 97, 6–21.
[doi:10.1016/j.ajhg.2015.05.022](https://doi.org/10.1016/j.ajhg.2015.05.022)
Buchbinder, M., and Timmermans, S. (2011). Newborn Screening and Maternal
[Diagnosis: Rethinking Family Benefit. Soc. Sci. Med. 73, 1014–1018. doi:10.](https://doi.org/10.1016/j.socscimed.2011.06.062)
[1016/j.socscimed.2011.06.062](https://doi.org/10.1016/j.socscimed.2011.06.062)
Canadian Institutes of Health Research (2018). Canadian Institutes of Health
Research, Natural Sciences and Engineering Research Council of Canada, and
Social Sciences and Humanities Research Council, Tri-Council Policy Statement:
Ethical Conduct for Research Involving Humans.
Carroll, A. E., and Downs, S. M. (2006). Comprehensive Cost-Utility Analysis of
[Newborn Screening Strategies. Pediatrics 117, S287–S295. doi:10.1542/peds.](https://doi.org/10.1542/peds.2005-2633H)
[2005-2633H](https://doi.org/10.1542/peds.2005-2633H)
Ceyhan-Birsoy, O., Murry, J. B., Machini, K., Lebo, M. S., Yu, T. W., Fayer, S., et al.
(2019). Interpretation of Genomic Sequencing Results in Healthy and Ill
Newborns: Results from the BabySeq Project. Am. J. Hum. Genet. 104,
[76–93. doi:10.1016/j.ajhg.2018.11.016](https://doi.org/10.1016/j.ajhg.2018.11.016)
Clayton, E. W., Appelbaum, P. S., Chung, W. K., Marchant, G. E., Roberts, J. L., and
Evans, B. J. (2021). Does the Law Require Reinterpretation and Return of
[Revised Genomic Results? Genet. Med. 23, 833–836. doi:10.1038/s41436-020-](https://doi.org/10.1038/s41436-020-01065-x)
[01065-x](https://doi.org/10.1038/s41436-020-01065-x)
Costain, G., Jobling, R., Walker, S., Reuter, M. S., Snell, M., Bowdin, S., et al. (2018).
Periodic Reanalysis of Whole-Genome Sequencing Data Enhances the
Diagnostic Advantage over Standard Clinical Genetic Testing. Eur. J. Hum.
[Genet. 26, 740–744. doi:10.1038/s41431-018-0114-6](https://doi.org/10.1038/s41431-018-0114-6)
-----
Council for International Organizations of Medical Sciences (CIOMS) in
collaboration with the World Health Organization (2016). International
Ethical Guidelines for Health-Related Research Involving Humans. Geneva,
Switzerland: World Health Organization.
Davis, D. S. (1997). Genetic Dilemmas and the Child’s Right to an Open Future.
[Hastings Cent. Rep. 27, 7. doi:10.2307/3527620](https://doi.org/10.2307/3527620)
de Wert, G., Dondorp, W., Dondorp, W., Clarke, A., Dequeker, E. M. C., Cordier,
C., et al. (2021). Opportunistic Genomic Screening. Recommendations of the
[European Society of Human Genetics. Eur. J. Hum. Genet. 29, 365–377. doi:10.](https://doi.org/10.1038/s41431-020-00758-w)
[1038/s41431-020-00758-w](https://doi.org/10.1038/s41431-020-00758-w)
Dondorp, W. J., and de Wert, G. M. W. R. (2013). The ’thousand-Dollar Genome’:
[an Ethical Exploration. Eur. J. Hum. Genet. 21, S6–S26. doi:10.1038/ejhg.](https://doi.org/10.1038/ejhg.2013.73)
[2013.73](https://doi.org/10.1038/ejhg.2013.73)
Downie, L., Halliday, J., Lewis, S., and Amor, D. J. (2021). Principles of Genomic
[Newborn Screening Programs. JAMA Netw. Open 4, e2114336. doi:10.1001/](https://doi.org/10.1001/jamanetworkopen.2021.14336)
[jamanetworkopen.2021.14336](https://doi.org/10.1001/jamanetworkopen.2021.14336)
Eichinger, J., Elger, B. S., Koné, I., Filges, I., Shaw, D., Zimmermann, B., et al.
(2021). The Full Spectrum of Ethical Issues in Pediatric Genome-wide
[Sequencing: a Systematic Qualitative Review. BMC Pediatr. 21, 387. doi:10.](https://doi.org/10.1186/s12887-021-02830-w)
[1186/s12887-021-02830-w](https://doi.org/10.1186/s12887-021-02830-w)
Fernandez, C. V., Bouffet, E., Malkin, D., Jabado, N., O’Connell, C., Avard, D., et al.
(2014). Attitudes of Parents toward the Return of Targeted and Incidental
[Genomic Research Findings in Children. Genet. Med. 16, 633–640. doi:10.1038/](https://doi.org/10.1038/gim.2013.201)
[gim.2013.201](https://doi.org/10.1038/gim.2013.201)
Franková, V., Driscoll, R. O., Jansen, M. E., Loeber, J. G., Kožich, V., Bonham, J.,
et al. (2021). Regulatory Landscape of Providing Information on Newborn
[Screening to Parents across Europe. Eur. J. Hum. Genet. 29, 67–78. doi:10.1038/](https://doi.org/10.1038/s41431-020-00716-6)
[s41431-020-00716-6](https://doi.org/10.1038/s41431-020-00716-6)
Friedman, J. M., Cornel, M. C., Goldenberg, A. J., Lister, K. J., Sénécal, K., Vears, D.
F., et al. (2017). Genomic Newborn Screening: Public Health Policy
[Considerations and Recommendations. BMC Med. Genomics 10, 9. doi:10.](https://doi.org/10.1186/s12920-017-0247-4)
[1186/s12920-017-0247-4](https://doi.org/10.1186/s12920-017-0247-4)
Genetti, C. A., Schwartz, T. S., Robinson, J. O., VanNoy, G. E., Petersen, D., Pereira,
S., et al. (2019). Parental Interest in Genomic Sequencing of Newborns:
Enrollment Experience from the BabySeq Project. Genet. Med. 21, 622–630.
[doi:10.1038/s41436-018-0105-6](https://doi.org/10.1038/s41436-018-0105-6)
Genomics England and the UK National Screening Committee (2021).
Implications of Whole Genome Sequencing for Newborn Screening: A Public
Dialogue.
Gold, N. B., Harrison, S. M., Rowe, J. H., Gold, J., Furutani, E., Biffi, A., et al. (2022).
Low Frequency of Treatable Pediatric Disease Alleles in gnomAD: An
Opportunity for Future Genomic Screening of Newborns. Hum. Genet.
[Genomics Adv. 3, 100059. doi:10.1016/j.xhgg.2021.100059](https://doi.org/10.1016/j.xhgg.2021.100059)
Goldenberg, A. J., and Sharp, R. R. (2012). The Ethical Hazards and Programmatic
[Challenges of Genomic Newborn Screening. JAMA 307, 461. doi:10.1001/jama.](https://doi.org/10.1001/jama.2012.68)
[2012.68](https://doi.org/10.1001/jama.2012.68)
Goldstein, N. D., Palumbo, A. J., Bellamy, S. L., Purtle, J., and Locke, R. (2020). State
and Local Government Expenditures and Infant Mortality in the United States.
[Pediatrics 146, e20201134. doi:10.1542/peds.2020-1134](https://doi.org/10.1542/peds.2020-1134)
Gudmundsson, S., Singer-Berk, M., Watts, N. A., Phu, W., Goodrich, J. K.,
Solomonson, M., et al. (2021). Variant Interpretation Using Population
[Databases: Lessons from gnomAD. Hum. Mutat. doi:10.1002/humu.24309](https://doi.org/10.1002/humu.24309)
Health Council of the Netherlands (2008). Screening: Between Hope and Hype.
Holm, I. A., Agrawal, P. B., Agrawal, P. B., Ceyhan-Birsoy, O., Christensen, K. D.,
Fayer, S., et al. (2018). The BabySeq Project: Implementing Genomic
[Sequencing in Newborns. BMC Pediatr. 18, 225. doi:10.1186/s12887-018-](https://doi.org/10.1186/s12887-018-1200-1)
[1200-1](https://doi.org/10.1186/s12887-018-1200-1)
Howard, H. C., Knoppers, B. M., Knoppers, B. M., Cornel, M. C., Wright Clayton,
E., Sénécal, K., et al. (2015). Whole-genome Sequencing in Newborn Screening?
A Statement on the Continued Importance of Targeted Approaches in
Newborn Screening Programmes. Eur. J. Hum. Genet. 23, 1593–1600.
[doi:10.1038/ejhg.2014.289](https://doi.org/10.1038/ejhg.2014.289)
Howson, C. P., Cedergren, B., Giugliani, R., Huhtinen, P., Padilla, C. D., Palubiak,
C. S., et al. (2018). Universal Newborn Screening: A Roadmap for Action. Mol.
[Genet. Metabolism 124, 177–183. doi:10.1016/j.ymgme.2018.04.009](https://doi.org/10.1016/j.ymgme.2018.04.009)
Jansen, M. E., Metternick-Jones, S. C., and Lister, K. J. (2017). International
Differences in the Evaluation of Conditions for Newborn Bloodspot
Screening: a Review of Scientific Literature and Policy Documents. Eur.
[J. Hum. Genet. 25, 10–16. doi:10.1038/ejhg.2016.126](https://doi.org/10.1038/ejhg.2016.126)
Johnston, J., Lantos, J. D., Goldenberg, A., Chen, F., Parens, E., Koenig, B. A., et al.
(2018). Sequencing Newborns:A Call for Nuanced Use of Genomic
[Technologies. Hastings Cent. Rep. 48, S2–S6. doi:10.1002/hast.874](https://doi.org/10.1002/hast.874)
Joseph, G., Chen, F., Harris-Wai, J., Puck, J. M., Young, C., and Koenig, B. A. (2016).
Parental Views on Expanded Newborn Screening Using Whole-Genome
[Sequencing. PEDIATRICS 137, S36–S46. doi:10.1542/peds.2015-3731H](https://doi.org/10.1542/peds.2015-3731H)
Kalkman, S., and Dondorp, W. (2022). The Case for Screening in Early Life for
’non-Treatable’ Disorders: Ethics, Evidence and Proportionality. A Report from
[the Health Council of the Netherlands. Eur. J. Hum. Genet. doi:10.1038/s41431-](https://doi.org/10.1038/s41431-022-01055-4)
[022-01055-4](https://doi.org/10.1038/s41431-022-01055-4)
Khoury, M. J., McCabe, L. L., and McCabe, E. R. B. (2003). Population Screening in
[the Age of Genomic Medicine. N. Engl. J. Med. 348, 50–58. doi:10.1056/](https://doi.org/10.1056/NEJMra013182)
[NEJMra013182](https://doi.org/10.1056/NEJMra013182)
King, J. R., Notarangelo, L. D., and Hammarström, L. (2021). An Appraisal of the
Wilson & Jungner Criteria in the Context of Genomic-Based Newborn
Screening for Inborn Errors of Immunity. J. Allergy Clin. Immunol. 147,
[428–438. doi:10.1016/j.jaci.2020.12.633](https://doi.org/10.1016/j.jaci.2020.12.633)
Knoppers, B. M., Sénécal, K., Boisjoli, J., Borry, P., Cornel, M. C., Fernandez, C. V.,
et al. (2016). Recontacting Pediatric Research Participants for Consent when
They Reach the Age of Majority. IRB 38, 1–9.
Knoppers, B. M. (1992). Canadian Child Health Law: Health Rights and Risks of
Children. Toronto, Ont., Canada: Thompson Educational Publishing, Inc.
Krotoski, D., Namaste, S., Raouf, R. K., El Nekhely, I., Hindi-Alexander, M.,
Engelson, G., et al. (2009). Conference Report: Second Conference of the
Middle East and North Africa Newborn Screening Initiative: Partnerships
for Sustainable Newborn Screening Infrastructure and Research
[Opportunities. Genet. Med. 11, 663–668. doi:10.1097/GIM.0b013e3181ab2277](https://doi.org/10.1097/GIM.0b013e3181ab2277)
Lewis, M. A., Paquin, R. S., Roche, M. I., Furberg, R. D., Rini, C., Berg, J. S., et al.
(2016). Supporting Parental Decisions about Genomic Sequencing for
Newborn Screening: The NC NEXUS Decision Aid. Pediatrics 137, S16–S23.
[doi:10.1542/peds.2015-3731E](https://doi.org/10.1542/peds.2015-3731E)
Lewis, M. H., and Goldenberg, A. J. (2015). Return of Results from Research Using
Newborn Screening Dried Blood Samples. J. Law. Med. Ethics 43, 559–568.
[doi:10.1111/jlme.12299](https://doi.org/10.1111/jlme.12299)
Lewis, M. H. (2014). Newborn Screening Controversy. JAMA Pediatr. 168, 199.
[doi:10.1001/jamapediatrics.2013.4980](https://doi.org/10.1001/jamapediatrics.2013.4980)
Lipstein, E. A., Nabi, E., Perrin, J. M., Luff, D., Browning, M. F., and Kuhlthau, K. A.
(2010). Parents’ Decision-Making in Newborn Screening: Opinions, Choices,
[and Information Needs. Pediatrics 126, 696–704. doi:10.1542/peds.2010-0217](https://doi.org/10.1542/peds.2010-0217)
Lu, C. Y., McMahon, P. M., and Wu, A. C. (2022). Modeling Genomic Screening in
[Newborns. JAMA Pediatr. 176, 344. doi:10.1001/jamapediatrics.2021.5798](https://doi.org/10.1001/jamapediatrics.2021.5798)
McCandless, S. E., and Wright, E. J. (2020). Mandatory Newborn Screening in the
United States: History, Current Status, and Existential Challenges. Birth Defects
[Res. 112, 350–366. doi:10.1002/bdr2.1653](https://doi.org/10.1002/bdr2.1653)
Melzer, S. M. (2022). Addressing Social Determinants of Health in Pediatric Health
Systems: Balancing Mission and Financial Sustainability. Curr. Opin. Pediatr.
[34, 8–13. doi:10.1097/MOP.0000000000001083](https://doi.org/10.1097/MOP.0000000000001083)
Milko, L. V., O’Daniel, J. M., DeCristo, D. M., Crowley, S. B., Foreman, A. K. M.,
Wallace, K. E., et al. (2019). An Age-Based Framework for Evaluating Genome[Scale Sequencing Results in Newborn Screening. J. Pediatr. 209, 68–76. doi:10.](https://doi.org/10.1016/j.jpeds.2018.12.027)
[1016/j.jpeds.2018.12.027](https://doi.org/10.1016/j.jpeds.2018.12.027)
Miller, D. T., Lee, K., Gordon, A. S., Amendola, L. M., Adelman, K., Bale, S. J., et al.
(2021). Recommendations for Reporting of Secondary Findings in Clinical
Exome and Genome Sequencing, 2021 Update: a Policy Statement of the
American College of Medical Genetics and Genomics (ACMG). Genet. Med.
[23, 1391–1398. doi:10.1038/s41436-021-01171-4](https://doi.org/10.1038/s41436-021-01171-4)
National Organization for Rare Disorders Newborn screening State report card
(2021). National Organization for Rare Disorders Newborn Screening State
[Report Card. Available at: https://rarediseases.org/policy-issues/newborn-](https://rarediseases.org/policy-issues/newborn-screening/)
[screening/.](https://rarediseases.org/policy-issues/newborn-screening/)
Newborn Screening Saves Lives Reauthorization Act (2014). Newborn Screening
Saves Lives Reauthorization Act of 2014, 42 U.S.C. §§ 300b-8 to 300b-17.
Nordfalk, F., and Ekstrøm, C. T. (2019). Newborn Dried Blood Spot Samples in
Denmark: the Hidden Figures of Secondary Use and Research Participation.
[Eur. J. Hum. Genet. 27, 203–210. doi:10.1038/s41431-018-0276-2](https://doi.org/10.1038/s41431-018-0276-2)
-----
O’Daniel, J. M. (2010). The Prospect of Genome-Guided Preventive Medicine: A
Need and Opportunity for Genetic Counselors. J. Genet. Couns. 19, 315–327.
[doi:10.1007/s10897-010-9302-4](https://doi.org/10.1007/s10897-010-9302-4)
OHRP (2017). United States Common Rule. Rockville, MD: OHRP.
Pereira, S., Robinson, J. O., Gutierrez, A. M., Petersen, D. K., Hsu, R. L., Lee, C. H.,
et al. (2019). Perceived Benefits, Risks, and Utility of Newborn Genomic
[Sequencing in the BabySeq Project. Pediatrics 143, S6–S13. doi:10.1542/peds.](https://doi.org/10.1542/peds.2018-1099C)
[2018-1099C](https://doi.org/10.1542/peds.2018-1099C)
Pereira, S., Smith, H. S., Frankel, L. A., Christensen, K. D., Islam, R., Robinson, J. O.,
et al. (2021). Psychosocial Effect of Newborn Genomic Sequencing on Families
[in the BabySeq Project. JAMA Pediatr. 175, 1132–1141. doi:10.1001/](https://doi.org/10.1001/jamapediatrics.2021.2829)
[jamapediatrics.2021.2829](https://doi.org/10.1001/jamapediatrics.2021.2829)
Peterson, R. E., Kuchenbaecker, K., Walters, R. K., Chen, C.-Y., Popejoy, A. B.,
Periyasamy, S., et al. (2019). Genome-wide Association Studies in Ancestrally
Diverse Populations: Opportunities, Methods, Pitfalls, and Recommendations.
[Cell. 179, 589–603. doi:10.1016/j.cell.2019.08.051](https://doi.org/10.1016/j.cell.2019.08.051)
Popejoy, A. B., and Fullerton, S. M. (2016). Genomics Is Failing on Diversity.
[Nature 538, 161–164. doi:10.1038/538161a](https://doi.org/10.1038/538161a)
Prosser, L. A., Grosse, S. D., Kemper, A. R., Tarini, B. A., and Perrin, J. M. (2012).
Decision Analysis, Economic Evaluation, and Newborn Screening: Challenges
[and Opportunities. Genet. Med. 14, 703–712. doi:10.1038/gim.2012.24](https://doi.org/10.1038/gim.2012.24)
Rehm, H. L., Berg, J. S., Brooks, L. D., Bustamante, C. D., Evans, J. P., Landrum, M.
J., et al. (2015). ClinGen - the Clinical Genome Resource. N. Engl. J. Med. 372,
[2235–2242. doi:10.1056/NEJMsr1406261](https://doi.org/10.1056/NEJMsr1406261)
Reinbold, G. W. (2019). Effects of the Convention on the Rights of the Child on
Child Mortality and Vaccination Rates: a Synthetic Control Analysis. BMC Int.
[Health Hum. Rights 19, 24. doi:10.1186/s12914-019-0211-9](https://doi.org/10.1186/s12914-019-0211-9)
Roman, T. S., Crowley, S. B., Roche, M. I., Foreman, A. K. M., O’Daniel, J. M.,
Seifert, B. A., et al. (2020). Genomic Sequencing for Newborn Screening: Results
[of the NC NEXUS Project. Am. J. Hum. Genet. 107, 596–611. doi:10.1016/j.ajhg.](https://doi.org/10.1016/j.ajhg.2020.08.001)
[2020.08.001](https://doi.org/10.1016/j.ajhg.2020.08.001)
Ross, L. F., and Clayton, E. W. (2019). Ethical Issues in Newborn Sequencing
[Research: The Case Study of BabySeq. Pediatrics 144, e20191031. doi:10.1542/](https://doi.org/10.1542/peds.2019-1031)
[peds.2019-1031](https://doi.org/10.1542/peds.2019-1031)
Rothwell, E., Goldenberg, A., Johnson, E., Riches, N., Tarini, B., and Botkin, J. R.
(2017). An Assessment of a Shortened Consent Form for the Storage and
Research Use of Residual Newborn Screening Blood Spots. J. Empir. Res. Hum.
[Res. Ethics 12, 335–342. doi:10.1177/1556264617736199](https://doi.org/10.1177/1556264617736199)
Sahai, I., and Marsden, D. (2009). Newborn Screening. Crit. Rev. Clin. Laboratory
[Sci. 46, 55–82. doi:10.1080/10408360802485305](https://doi.org/10.1080/10408360802485305)
Sénécal, K., Unim, B., Unim, B., and Maria Knoppers, B. (2018). Newborn
Screening Programs: Next Generation Ethical and Social Issues. Obm Genet.
[2, 1. doi:10.21926/obm.genet.1803027](https://doi.org/10.21926/obm.genet.1803027)
Sohn, H., and Timmermans, S. (2019). Inequities in Newborn Screening: Race and
[the Role of Medicaidq. SSM - Popul. Health 9, 100496. doi:10.1016/j.ssmph.](https://doi.org/10.1016/j.ssmph.2019.100496)
[2019.100496](https://doi.org/10.1016/j.ssmph.2019.100496)
Sontag, M. K., Yusuf, C., Grosse, S. D., Edelman, S., Miller, J. I., McKasson, S., et al.
(2020). Infants with Congenital Disorders Identified through Newborn
Screening - United States, 2015-2017. MMWR Morb. Mortal. Wkly. Rep. 69,
[1265–1268. doi:10.15585/mmwr.mm6936a6](https://doi.org/10.15585/mmwr.mm6936a6)
Therrell, B. L., Lloyd-Puryear, M. A., Lloyd-Puryear, M. A., Ohene-Frempong, K.,
Ware, R. E., Padilla, C. D., et al. (2020). Empowering Newborn Screening
Programs in African Countries through Establishment of an International
[Collaborative Effort. J. Community Genet. 11, 253–268. doi:10.1007/s12687-](https://doi.org/10.1007/s12687-020-00463-7)
[020-00463-7](https://doi.org/10.1007/s12687-020-00463-7)
Therrell, B. L., Padilla, C. D., Loeber, J. G., Kneisser, I., Saadallah, A., Borrajo, G.
J. C., et al. (2015). Current Status of Newborn Screening Worldwide: 2015.
[Seminars Perinatology 39, 171–187. doi:10.1053/j.semperi.2015.03.002](https://doi.org/10.1053/j.semperi.2015.03.002)
Therrell, B. L., and Padilla, C. D. (2018). Newborn Screening in the Developing
Countries. Curr. Opin. Pediatr. 30, 734–739. [doi:10.1097/MOP.](https://doi.org/10.1097/MOP.0000000000000683)
[0000000000000683](https://doi.org/10.1097/MOP.0000000000000683)
Tonniges, T. F. (2000). Serving the Family from Birth to the Medical Home.
Newborn Screening: a Blueprint for the Future - a Call for a National Agenda on
State Newborn Screening Programs. Pediatrics 106, 389–422.
Tsosie, K. S., Yracheta, J. M., Kolopenuk, J. A., and Geary, J. (2021). We Have
“Gifted” Enough: Indigenous Genomic Data Sovereignty in Precision Medicine.
[Am. J. Bioeth. 21, 72–75. doi:10.1080/15265161.2021.1891347](https://doi.org/10.1080/15265161.2021.1891347)
United Nations Convention on the Rights of the Child (1989). United Nations
[Convention on the Rights of the Child. Available at: https://www.ohchr.org/en/](https://www.ohchr.org/en/professionalinterest/pages/crc.aspx)
[professionalinterest/pages/crc.aspx (Accessed January 24, 2022).](https://www.ohchr.org/en/professionalinterest/pages/crc.aspx)
Veenstra, D. L., Rowe, J. W., Pagán, J. A., Shelton Brown, H., Schneider, J. E.,
Gupta, A., et al. (2021). Reimbursement for Genetic Variant Reinterpretation:
[Five Questions Payers Should Ask. Am. J. Manag. Care 27, e336–e338. doi:10.](https://doi.org/10.37765/ajmc.2021.88763)
[37765/ajmc.2021.88763](https://doi.org/10.37765/ajmc.2021.88763)
Vittozzi, L., Hoffmann, G. F., Cornel, M., and Loeber, G. (2010). Evaluation of Population
Newborn Screening Practices for Rare Disorders in Member States of the European
[Union. Orphanet J. Rare Dis. 5, P26. doi:10.1186/1750-1172-5-S1-P26](https://doi.org/10.1186/1750-1172-5-S1-P26)
Wain, K. E., Palen, E., Savatt, J. M., Shuman, D., Finucane, B., Seeley, A., et al.
(2018). The Value of Genomic Variant ClinVar Submissions from Clinical
Providers: Beyond the Addition of Novel Variants. Hum. Mutat. 39, 1660–1667.
[doi:10.1002/humu.23607](https://doi.org/10.1002/humu.23607)
Walen, A. (2021). “Retributive Justice,” in The Stanford Encyclopedia of Philosophy.
Editor E. N. Zalta (Stanford, CA: Metaphysics Research Lab, Stanford
University). Available at: [https://plato.stanford.edu/archives/sum2021/](https://plato.stanford.edu/archives/sum2021/entries/justice-retributive/)
[entries/justice-retributive/(Accessed April 13, 2022).](https://plato.stanford.edu/archives/sum2021/entries/justice-retributive/)
Wang, T., Antonacci-Fulton, L., Howe, K., Lawson, H. A., Lucas, J. K.,
Phillippy, A. M., et al. (2022). The Human Pangenome Project: a Global
[Resource to Map Genomic Diversity. Nature 604, 437–446. doi:10.1038/](https://doi.org/10.1038/s41586-022-04601-8)
[s41586-022-04601-8](https://doi.org/10.1038/s41586-022-04601-8)
Wenger, A. M., Guturu, H., Bernstein, J. A., and Bejerano, G. (2017). Systematic
Reanalysis of Clinical Exome Data Yields Additional Diagnoses: Implications
[for Providers. Genet. Med. 19, 209–214. doi:10.1038/gim.2016.88](https://doi.org/10.1038/gim.2016.88)
Wilson, J. M. G., and Jungner, G. (1968). Principles and Practice of Screening for
Disease. Geneva: World Health Organization.
WMA (2022). WMA - the World Medical Association-WMA Declaration of
Helsinki – Ethical Principles for Medical Research Involving Human Subjects.
[Available at: https://www.wma.net/policies-post/wma-declaration-of-helsinki-](https://www.wma.net/policies-post/wma-declaration-of-helsinki-ethical-principles-for-medical-research-involving-human-subjects/)
[ethical-principles-for-medical-research-involving-human-subjects/(Accessed](https://www.wma.net/policies-post/wma-declaration-of-helsinki-ethical-principles-for-medical-research-involving-human-subjects/)
February 11, 2022).
Woerner, A. C., Gallagher, R. C., Vockley, J., and Adhikari, A. N. (2021). The Use of
Whole Genome and Exome Sequencing for Newborn Screening: Challenges
[and Opportunities for Population Health. Front. Pediatr. 9, 663752. doi:10.](https://doi.org/10.3389/fped.2021.663752)
[3389/fped.2021.663752](https://doi.org/10.3389/fped.2021.663752)
Wojcik, M. H., Zhang, T., Ceyhan-Birsoy, O., Genetti, C. A., Lebo, M. S., Yu, T. W.,
et al. (2021). Discordant Results between Conventional Newborn Screening and
Genomic Sequencing in the BabySeq Project. Genet. Med. 23, 1372–1375.
[doi:10.1038/s41436-021-01146-5](https://doi.org/10.1038/s41436-021-01146-5)
Wright, C. F., Hurles, M. E., and Firth, H. V. (2016). Principle of
Proportionality in Genomic Data Sharing. Nat. Rev. Genet. 17, 1–2.
[doi:10.1038/nrg.2015.5](https://doi.org/10.1038/nrg.2015.5)
Conflict of Interest: The authors declare that the research was conducted in the
absence of any commercial or financial relationships that could be construed as a
potential conflict of interest.
Publisher’s Note: All claims expressed in this article are solely those of the authors
and do not necessarily represent those of their affiliated organizations, or those of
the publisher, the editors and the reviewers. Any product that may be evaluated in
this article, or claim that may be made by its manufacturer, is not guaranteed or
endorsed by the publisher.
Copyright © 2022 Rahimzadeh, Friedman, de Wert and Knoppers. This is an open[access article distributed under the terms of the Creative Commons Attribution](https://creativecommons.org/licenses/by/4.0/)
[License (CC BY). The use, distribution or reproduction in other forums is permitted,](https://creativecommons.org/licenses/by/4.0/)
provided the original author(s) and the copyright owner(s) are credited and that the
original publication in this journal is cited, in accordance with accepted academic
practice. No use, distribution or reproduction is permitted which does not comply
with these terms.
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://pmc.ncbi.nlm.nih.gov/articles/PMC9289115, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "GOLD",
"url": "https://www.frontiersin.org/articles/10.3389/fgene.2022.865400/pdf"
}
| 2,022
|
[
"JournalArticle"
] | true
| 2022-07-04T00:00:00
|
[
{
"paperId": "772d3c795067541b768fdf08ca01051cc73af257",
"title": "The Human Pangenome Project: a global resource to map genomic diversity"
},
{
"paperId": "e2d1dfab7d7dd990b8e059fd312ce516c2f40dff",
"title": "The case for screening in early life for ‘non-treatable’ disorders: ethics, evidence and proportionality. A report from the Health Council of the Netherlands"
},
{
"paperId": "d123d7c5f8e189e4d047fa2ee205d97ce16413cd",
"title": "Modeling Genomic Screening in Newborns."
},
{
"paperId": "6526ccfdde1c76c184f55ed095ab6180d7ab76d3",
"title": "Addressing social determinants of health in pediatric health systems: balancing mission and financial sustainability"
},
{
"paperId": "7f4ae21e4a02f14bc6eda2d91d00b044f2f164d6",
"title": "Should all babies have their genome sequenced at birth?"
},
{
"paperId": "87c472f54dba27c1385d25e8bf1e2b96e3a7f316",
"title": "Reimbursement for genetic variant reinterpretation: five questions payers should ask."
},
{
"paperId": "f29fc1a2c8edf1e44bc9c0338e8c636f05b7635d",
"title": "The full spectrum of ethical issues in pediatric genome-wide sequencing: a systematic qualitative review"
},
{
"paperId": "698057e0ce4fdacc40eb3f28b50670b2ceafaea2",
"title": "Low frequency of treatable pediatric disease alleles in gnomAD: An opportunity for future genomic screening of newborns"
},
{
"paperId": "1d7b5af69837570297f15632e8adab36141582a4",
"title": "Psychosocial Effect of Newborn Genomic Sequencing on Families in the BabySeq Project: A Randomized Clinical Trial."
},
{
"paperId": "d5db5e591db09e3250b1a8c06d739d15d7d1b47e",
"title": "Variant interpretation using population databases: Lessons from gnomAD"
},
{
"paperId": "5d2428be0a021099776be0070a6b5c3abb144c90",
"title": "The Use of Whole Genome and Exome Sequencing for Newborn Screening: Challenges and Opportunities for Population Health"
},
{
"paperId": "a44e312cf7dda561903ca26899ecc9ae9cfe17ed",
"title": "Principles of Genomic Newborn Screening Programs"
},
{
"paperId": "144846532e1ae756e587caefd6a8609f69508106",
"title": "Recommendations for reporting of secondary findings in clinical exome and genome sequencing, 2021 update: a policy statement of the American College of Medical Genetics and Genomics (ACMG)"
},
{
"paperId": "e6ed7ac24641e609a1bb11f739b7414267c396b8",
"title": "We Have “Gifted” Enough: Indigenous Genomic Data Sovereignty in Precision Medicine"
},
{
"paperId": "a2a5cfed759b38ffcae2ccba118bb563ac8936c0",
"title": "Discordant results between conventional newborn screening and genomic sequencing in the BabySeq Project"
},
{
"paperId": "f72fb1ce300d461ab6126b64b5fc52f9e8b633e0",
"title": "Newborn screening in Latin America: A brief overview of the state of the art"
},
{
"paperId": "3bbc5725937253b30dba9a6892fa155f50ad280f",
"title": "An appraisal of the Wilson & Jungner criteria in the context of genomic-based newborn screening for inborn errors of immunity."
},
{
"paperId": "af59a050c51c33a63ad9f07d2a00e904da3891f6",
"title": "Does the law require reinterpretation and return of revised genomic results?"
},
{
"paperId": "efc3b5c2715c9eae64ea2733a89f5abd2cb341ae",
"title": "Opportunistic genomic screening. Recommendations of the European Society of Human Genetics"
},
{
"paperId": "4979ae1a00710252a48ad0271b05b74018dd8eae",
"title": "Variant Classification Concordance using the ACMG-AMP Variant Interpretation Guidelines across Nine Genomic Implementation Research Studies."
},
{
"paperId": "3c121084a7f9403f9b380027767b560f27e70b78",
"title": "Regulatory landscape of providing information on newborn screening to parents across Europe"
},
{
"paperId": "c5bbd964a693f32fb00d8d8a88d81dc7fb9efd3a",
"title": "State and Local Government Expenditures and Infant Mortality in the United States"
},
{
"paperId": "07f2275dece3e247c2008b4699f425be2cfc148c",
"title": "Infants with Congenital Disorders Identified Through Newborn Screening — United States, 2015–2017"
},
{
"paperId": "bba46eff638095df0a5088e0f1548f640f7bbbcf",
"title": "Genetic counseling service delivery models: A study of genetic counselors’ interests, needs, and barriers to implementation"
},
{
"paperId": "6c85f4505fcc165a29503a418238a2383711675e",
"title": "The role of exome sequencing in newborn screening for inborn errors of metabolism"
},
{
"paperId": "61421ebce87423363bdee3d7881a97c0a65c1619",
"title": "Empowering newborn screening programs in African countries through establishment of an international collaborative effort"
},
{
"paperId": "b46ad4399ff6b93b2c9135409cbbe12419e812d7",
"title": "Mandatory newborn screening in the United States: History, current status, and existential challenges"
},
{
"paperId": "f6e36e2b2d57e300ecb6c3e9fc0f8808801d74a4",
"title": "Genomic Sequencing for Newborn Screening: Results of the NC NEXUS Project."
},
{
"paperId": "eab019a98b0807c8757f40ce4ff0e6ea9478b5f2",
"title": "Ethical Issues in Newborn Sequencing Research: The Case Study of BabySeq"
},
{
"paperId": "3af43fabfaa7fe89f0c8ffe666b5fd0ba104dd27",
"title": "Is there a duty to reinterpret genetic data? The ethical dimensions"
},
{
"paperId": "0b6c17207b908152de37e94077b8a2a1efad2c0f",
"title": "Inequities in newborn screening: Race and the role of medicaid☆"
},
{
"paperId": "6530c6954e593d8e8f6f528ce8059ced57f39655",
"title": "Genome-wide Association Studies in Ancestrally Diverse Populations: Opportunities, Methods, Pitfalls, and Recommendations"
},
{
"paperId": "4779a28455293dfbcf983e4dad3dd1feb7b4ad6b",
"title": "Effects of the Convention on the Rights of the Child on child mortality and vaccination rates: a synthetic control analysis"
},
{
"paperId": "f3f4ed9df39872e034aaf7226d43b9814bb1b736",
"title": "An Age-Based Framework for Evaluating Genome-Scale Sequencing Results in Newborn Screening."
},
{
"paperId": "5164cf3f0956b8fa72bb00c71605edc817462c78",
"title": "Council for International Organizations of Medical Sciences"
},
{
"paperId": "3293396095ab63c52fb3e91eea6bd4ec64ed01fe",
"title": "Perceived Benefits, Risks, and Utility of Newborn Genomic Sequencing in the BabySeq Project"
},
{
"paperId": "53551c819dcf126bd471a63ef26bf1cb42070131",
"title": "Interpretation of Genomic Sequencing Results in Healthy and Ill Newborns: Results from the BabySeq Project."
},
{
"paperId": "255287819382eb783e86ae2d2d18ea32d61259f5",
"title": "Newborn screening in the developing countries"
},
{
"paperId": "35cd91892a4caccdd33d7714bf390113dc751906",
"title": "The value of genomic variant ClinVar submissions from clinical providers: Beyond the addition of novel variants"
},
{
"paperId": "c7682b0ae406134e1e2b0c0010a1c90a3c444568",
"title": "Newborn dried blood spot samples in Denmark: the hidden figures of secondary use and research participation"
},
{
"paperId": "9ebf5eac882d850473e87c71aa1f4973ac4ee0c2",
"title": "Newborn Screening Programs: Next Generation Ethical and Social Issues"
},
{
"paperId": "6c3f8ac4bc5b9225d66746c5e442d70f9d4711fe",
"title": "The BabySeq project: implementing genomic sequencing in newborns"
},
{
"paperId": "2189d1feadbc865a02562ffe901a68fb7ed1435f",
"title": "Parental Interest in Genomic Sequencing of Newborns: Enrollment Experience from the BabySeq Project"
},
{
"paperId": "2a310c8dd600ab712124380f53c96cf458bb103b",
"title": "Sequencing Newborns: A Call for Nuanced Use of Genomic Technologies."
},
{
"paperId": "a31050670c2a827d7d9b0592bb8211b14d74d2b1",
"title": "Universal newborn screening: A roadmap for action."
},
{
"paperId": "7c5973d9817bc9d086f61c8058e8296c8fc975c1",
"title": "Periodic reanalysis of whole-genome sequencing data enhances the diagnostic advantage over standard clinical genetic testing"
},
{
"paperId": "216d4b7ce528a0a4c8ac2e4fe372732f01311e8c",
"title": "An Assessment of a Shortened Consent Form for the Storage and Research Use of Residual Newborn Screening Blood Spots"
},
{
"paperId": "281f1fdfdba052a33cbba45d336539d77bcd5e7f",
"title": "Genomic newborn screening: public health policy considerations and recommendations"
},
{
"paperId": "09d16bb20223d9d49e0c7ef0942144bfa537e227",
"title": "International differences in the evaluation of conditions for newborn bloodspot screening: a review of scientific literature and policy documents"
},
{
"paperId": "94c627d2d8b631fabb20cb816ab50095deb6952c",
"title": "Recontacting Pediatric Research Participants for Consent When They Reach the Age of Majority."
},
{
"paperId": "f0f44716e502fbe14fad5d0b71533ab6d1abb3e6",
"title": "Genomics is failing on diversity"
},
{
"paperId": "77fff587e39e6e8a9487e649498fe275ad3c7b5f",
"title": "Systematic reanalysis of clinical exome data yields additional diagnoses: implications for providers"
},
{
"paperId": "0356b8549b733372b5d4fecc0d22a9f26d986c8b",
"title": "Parental Views on Expanded Newborn Screening Using Whole-Genome Sequencing"
},
{
"paperId": "faef92d7d702868cb174cd9faaabe5d096b7b6f6",
"title": "Supporting Parental Decisions About Genomic Sequencing for Newborn Screening: The NC NEXUS Decision Aid"
},
{
"paperId": "dccb4c6f9f0b382a750e9678a952219c5398e38f",
"title": "Principle of proportionality in genomic data sharing"
},
{
"paperId": "27d27c729555f50bccf997be5a58a6e96d62c877",
"title": "Return of Results from Research Using Newborn Screening Dried Blood Samples"
},
{
"paperId": "59abff3409f7691d6b0536ebcc57dda24ff777a8",
"title": "Points to Consider: Ethical, Legal, and Psychosocial Implications of Genetic Testing in Children and Adolescents."
},
{
"paperId": "30fc5297c5d0c28ee282b7b9ed52f4c5f3924720",
"title": "ClinGen--the Clinical Genome Resource."
},
{
"paperId": "6ba430a944df5ec5b280bb4881272ee2dd0eb133",
"title": "Current status of newborn screening worldwide: 2015."
},
{
"paperId": "3769aa115f75024483d9f696f48ca9d40ee55483",
"title": "Whole-genome sequencing in newborn screening? A statement on the continued importance of targeted approaches in newborn screening programmes"
},
{
"paperId": "bc3bf1606f8c34014d584ca7bb2a273d544b8151",
"title": "Development of DNA Confirmatory and High-Risk Diagnostic Testing for Newborns Using Targeted Next-Generation DNA Sequencing"
},
{
"paperId": "e6675636454dd6580191061a524b4c9be43801de",
"title": "Newborn screening."
},
{
"paperId": "7e416fe71f417d634d422cedb6e3c4b39262acaa",
"title": "Newborn screening controversy: past, present, and future."
},
{
"paperId": "04865ed09a45bb846dcae8f7507efafe301d1630",
"title": "Public views on participating in newborn screening using genome sequencing"
},
{
"paperId": "796238ac5178f93c1581b8b8280279b0eb30768b",
"title": "Attitudes of parents toward the return of targeted and incidental genomic research findings in children"
},
{
"paperId": "540d930637ff7249d34730a74f1e00e3eb1fd1f4",
"title": "The ‘thousand-dollar genome’: an ethical exploration"
},
{
"paperId": "40f343a8f6d8a5ca4a3eea214b328cc030bee1e7",
"title": "Tri-Council Policy Statement: Ethical Conduct for Research Involving Humans"
},
{
"paperId": "5ced25089ba767b9a4182727a39449b43556c4fd",
"title": "Decision analysis, economic evaluation, and newborn screening: challenges and opportunities"
},
{
"paperId": "132f7daf6f0778d1967b76c310c463db3f6ba6f1",
"title": "The ethical hazards and programmatic challenges of genomic newborn screening."
},
{
"paperId": "07d1a2f016f413fb786f7c476a50fb385e55614d",
"title": "Newborn screening and maternal diagnosis: rethinking family benefit."
},
{
"paperId": "896a00302ecef2c77aed9c15544e7b7b250fbb79",
"title": "Evaluation of population newborn screening practices for rare disorders in member states of the European Union"
},
{
"paperId": "ed1ff66f9edf97ec06238434a0d24f49a4414077",
"title": "Parents' Decision-Making in Newborn Screening: Opinions, Choices, and Information Needs"
},
{
"paperId": "d3a49b64933b8e4af4cd49ba6da3d60ecea49111",
"title": "The Prospect of Genome-guided Preventive Medicine: A Need and Opportunity for Genetic Counselors"
},
{
"paperId": "9f5bf739ac7a0e0ca14c942dacef3f45b67eb61b",
"title": "Conference report: Second conference of the Middle East and North Africa newborn screening initiative: Partnerships for sustainable newborn screening infrastructure and research opportunities"
},
{
"paperId": "876061a13749f98be99b5260dd2f49a662d1414b",
"title": "Newborn Screening"
},
{
"paperId": "bbfa1aa85bbac3b427f45e218220b02fd7c26172",
"title": "Revisiting Wilson and Jungner in the genomic age: a review of screening criteria over the past 40 years."
},
{
"paperId": "50d806f6f054d860326ef0584f0b54b0c517f69b",
"title": "Comprehensive Cost-Utility Analysis of Newborn Screening Strategies"
},
{
"paperId": "1df1172e6bbb5f2923ca149280dadcc14560e8e3",
"title": "Population screening in the age of genomic medicine."
},
{
"paperId": "a7e5a72374ca7224bd90e1f5e8259cab21d3dc81",
"title": "Serving the family from birth to the medical home. Newborn screening: a blueprint for the future - a call for a national agenda on state newborn screening programs"
},
{
"paperId": "242187731d65c888430df594d75436db90366b0d",
"title": "Genetic dilemmas and the child's right to an open future."
},
{
"paperId": "5d301630d781b16e082fcbed9e9bd00003274beb",
"title": "The United Nations convention on the rights of the child"
},
{
"paperId": "d01508c5496069303c28363a295d7058fef1c162",
"title": "On retributive justice"
},
{
"paperId": null,
"title": "WMA - the World Medical Association-WMA Declaration of Helsinki – Ethical Principles for Medical Research Involving Human Subjects"
},
{
"paperId": null,
"title": "United States Common Rule"
},
{
"paperId": null,
"title": "Screening: Between Hope and Hype"
},
{
"paperId": "d932f6af927d5e62d11e189b826ab56c05605f8a",
"title": "Serving the family from birth to the medical home. A report from the Newborn Screening Task Force convened in Washington DC, May 10-11, 1999."
},
{
"paperId": null,
"title": "Canadian Child Health Law: Health Rights and Risks of Children"
},
{
"paperId": "9978d23a1ab3194e555f405c605b557625a1e9ac",
"title": "Principles and practice of screening for disease"
},
{
"paperId": null,
"title": "National Organization for Rare Disorders Newborn screening State report card (2021)"
}
] | 18,071
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Mathematics",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/0291665270a93c509ecde8b2c50642b3adcfe176
|
[
"Computer Science",
"Mathematics"
] | 0.839306
|
Automated abstraction by incremental refinement in interpolant-based model checking
|
0291665270a93c509ecde8b2c50642b3adcfe176
|
IEEE/ACM International Conference on Computer-Aided Design
|
[
{
"authorId": "1743867",
"name": "G. Cabodi"
},
{
"authorId": "1730731",
"name": "P. Camurati"
},
{
"authorId": "3251152",
"name": "M. Murciano"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"IEEE/ACM Int Conf Comput Des"
],
"alternate_urls": [
"http://ieeexplore.ieee.org/xpl/mostRecentIssue.jsp?punumber=345"
],
"id": "b87983f2-c272-4ada-8161-52fb86e00bfb",
"issn": "1063-6757",
"name": "IEEE/ACM International Conference on Computer-Aided Design",
"type": null,
"url": "http://ieeexplore.ieee.org/xpl/mostRecentIssue.jsp?punumber=212"
}
| null |
# Automated Abstraction by Incremental Refinement in Interpolant-based Model Checking
## G. Cabodi, P. Camurati, M. Murciano
Dipartimento di Automatica ed Informatica
Politecnico di Torino - Torino, Italy
Email: gianpiero.cabodi, paolo.camurati, marco.murciano @polito.it
_{_ _}_
**_Abstract—This paper addresses the field of Unbounded Model_**
**Checking (UMC) based on SAT engines, where Craig interpolants**
**have recently gained wide acceptance as an automated abstrac-**
**tion technique.**
**We start from the observation that interpolants can be quite**
**effective on large verification instances. As they operate on**
**SAT-generated refutation proofs, interpolants are very good at**
**automatically abstract facts that are not significant for proofs.**
**In this work, we push forward the new idea of generating**
**abstractions without resorting to SAT proofs, and to accept**
**(reject) abstractions whenever they (do not) fulfill given ade-**
**_quacy constraints. We propose an integrated approach smoothly_**
**combining the capabilities of interpolation with abstraction and**
**over-approximation techniques, that do not directly derive from**
**SAT refutation proofs.**
**The driving idea of this combination is to incrementally**
**generate, by refinement, an abstract (over-approximate) image,**
**built up from equivalences, implications, ternary and localization**
**abstraction, then (eventually) from SAT refutation proofs.**
**Experimental results, derived from the verification of hard**
**problems, show the robustness of our approach.**
I. INTRODUCTION
Symbolic model checking [1] is a method for verifying
temporal properties of finite state systems, which relies on
a symbolic representation of sets, typically through Binary
Decision Diagrams (BDDs) [2]. Symbolic representations of
state sets have been by far the key factor for BDD success in
symbolic model checking [1]. Given the canonicity of BDDs,
and the efficient implementations of quantification, image
and pre-image operators could produce Boolean functions
representing state sets. Unfortunately, scalability problems
practically limit BDD usage at circuits with tens to few
hundreds latches.
By contrast, bounded model checking (BMC) [3] can falsify
temporal properties using Boolean satisfiability (SAT). BMC is
efficient at producing counter-examples and it has been shown
to be more robust and scalable than BDD-based symbolic
model checking. As a matter of fact, SAT tools are the core
technology to attack larger problems. And Inverted Graphs
(AIGs), or similar (non-canonical) circuit-based representations [4] are used in order to model circuit unrollings. Circuits
are either converted to CNF clauses, to be processed by
SAT solvers, or directly manipulated by circuit-based solvers.
However, BMC is not complete, as it only guarantees the
correctness of a property up to a given bound.
Therefore, specific techniques are required in order to
support Unbounded Model Checking (UMC) in SAT based
environments. For the sake of detail, the ability to check
reachability fix-points is the main difference between BMC
and UMC. All UMC approaches basically rely on one or more
methods able to detect that the forward, backward or mixed
analysis they perform is complete.
_A. Related Works_
Our work follows UMC approaches based on SAT rather
than BDDs.
Inductive proofs are at the base of the most of such
approaches [5], [6], [7], [8], all following the seminal work
of Sheeran et al. [9]. Fix-point checks are proved inductively,
whereas completeness is based on uniqueness constraints,
expressing loop-free paths between states. Unfortunately, the
longest loop-free path can be exponentially longer than the
diameter of the reachable state space, thus most of the research
in this field has concentrated on finding tight sets of inductive
invariants, i.e. over-approximations of reachable states, quite
often sufficient for inductive proofs.
In order to avoid the exponential depth, symbolic representations of state sets are an alternative approach. But they are
generally difficult to manipulate within non BDD-based frameworks, as both CNF and circuit-based representations can
lead to memory explosion. Williams et al. [10] first adopted
Boolean Expression Diagrams (BEDs), for the removal of
quantifiers. Abdulla et al. [11] adopted Reduced Boolean
Circuits (RBCs), i.e., a variant of BEDs, to represent formulas
on which they performed existential quantifier elimination
through substitution, scope reduction, etc. McMillan [12], later
followed by Kang and Park [13], proposed quantifier elimination through the enumeration of SAT solutions (all-solutions
_SAT). Ganai et al. [14] extended the previous approaches by_
using “circuit co-factoring”. The authors adopted a circuit to
represent state sets, and they used circuit-based co-factoring
to capture a large set of states in every SAT enumeration step.
All the above methods potentially converge faster than [9],
although they share the common problem of possibly exponential state set representations.
Abstraction techniques represent an orthogonal direction to
tackle complexity, as they seek and remove those parts of
a circuit/system that are not relevant for the proof. Within
this general path of research, our work follows the ideas first
-----
introduced by McMillan in [15], who used Craig interpolants
for Unbounded Model Checking. Craig interpolants exploit
the ability of Modern SAT solvers to generate proofs of
unsatisfiability. Over-approximations of the reachable states
are computed, starting from refutation proofs of (unsatisfied)
BMC-like runs. The approach can be viewed as an iterative
refinement of proof-based abstractions to narrow down a proof
to relevant facts and its convergence is bound to the state
graph diameter. Craig interpolant’s most interesting features
are its completeness and the automated abstraction mechanism,
while the drawbacks are (again) mainly related to the potential
size blow-up of SAT-based interpolant circuits, and to the
convergence of over-approximate reachability.
New applications and improvements over the base
method [15] were proposed in [16], [17], [18], [19], in order
to push forward applicability and scalability of the technique.
This paper is a direct follow-up of [18], as we push forward our
previous idea, to compute abstractions at the image level, generating interpolant-like over-approximations, without resorting
to SAT refutation proofs.
_B. Motivations_
Our experience [20], [21] shows that SAT-based Craig interpolants, combined with preliminary computations of inductive
invariants, can prove a broad range of verification instances.
A careful analysis of the unproved problems lead us to the
following observations:
_• Craig interpolants tend to produce highly redundant_
circuit representations of state sets, derived from SATgenerated proofs (often of unpredictable size). Circuits
can be compacted by means of logic synthesis optimizers,
that are limited in terms of scalability, and thus poorly
working with big interpolants
_• Over-approximation can drastically reduce the forward_
diameter (as long state transition paths can by bypassed
by over-approximation). But it can also trigger state space
explorations within unreachable areas, with direct consequences in terms of both traversal depth and reachable
state set size
_• Craig interpolants combine the power of forward and_
backward reachability, by over-approximating forward
states within the region left “free” by backward reachability. Whenever the backward unrolling representation
is complex, and the “free” region shrinks, interpolant
representations tend to explode.
Due to the previously mentioned problems, we seek for
_alternative ways to generate state set over-approximations,_
whose circuit size can be better kept under control, and
we focus our efforts on the generation of tighter over_approximations._
_C. Contributions_
In this paper, we propose a set of abstraction techniques
not directly derived from SAT refutation proofs, although
controlled by SAT-based “adequacy” checks. We incrementally
compute sets of atomically simple over-approximations, with
increasing complexity and computational effort.
Our approach is a follow-up of our previous experience [18],
where we already explored the combination of interpolants and
localization abstraction. With the current work we go much
beyond that idea, as we introduce an automated abstraction
procedure, based on incremental learning and simplification
steps. We learn small atomic constraints such as equivalences
and implications between state variables. We simplify circuits representing state sets and combinational unrollings, by
exploiting equivalences and implications, plus ternary and
localization abstractions. Compared to [19], our method works
on abstractions at the image level, whereas they seek for
abstraction/refinement steps to be used for entire interpolantbased traversals.
We adopt:
_• cube-based over-approximations, by detecting state vari-_
ables temporarily stuck at constant values
_• equivalence classes and implications among state vari-_
ables, able to provide simple over-approximations, at the
base of powerful circuit optimizations
_• abstractions based on ternary logic models._
Starting from the above considerations, our approach proposes two main novel contributions:
1) We adopt state set over-approximation techniques that
are novel for interpolant-based Model Checking
2) We propose an integrated approach for image computation, incrementally combining and tuning the above
over-approximation techniques within a unified SATbased (complete) Unbounded Model Checking approach.
Our experimental results concentrate on proving correct
properties, and they show that the proposed methods improve
the original one by making it faster, more robust and scalable.
We show experiments where in some cases we are able
to complete difficult instances, not achievable with previous
techniques.
_D. Outline_
Section II introduces background notions on notation, BMC
and UMC, and SAT-based Craig interpolant Model Checking.
Sections III and IV present our main contributions. Section V
discusses the experiments we performed. Section VI concludes
with some summarizing remarks.
II. BACKGROUND
_A. Model and Notation_
We address systems modeled by labeled state transition
structures, and represented implicitly by Boolean formulas.
The state space and the (free) inputs are defined by indexed sets of Boolean variables V = _{v1, . . ., vn} and_
_W = {w1, . . ., wn}, respectively. States correspond to the_
valuations of variables in V, whereas transition labels correspond to the valuations of variables in W . We indicate
next states with the primed variable set V _[′]_ = {v1[′] _[, . . ., v]n[′]_ _[}][.]_
-----
Whenever we explicitly need time frame variables, we use
_V_ _[i]_ = {v1[i] _[, . . ., v]n[i]_ _[}][ and][ W][ i][ =][ {][w]1[i]_ _[, . . ., w]n[i]_ _[}][ for variable]_
instances at the i−th time frame. We also adopt the short
notation[1] _V_ _[i..j]_ for V _[i], V_ _[i][+1], ..., V_ _[j]. Similarly for W_ .
A set of states is expressed by a state predicate S(V )
(or S(V _[′]) for the next state space). I(V ) is the initial state_
predicate. We use P (V ) to denote an invariant property, and
_F_ (V ) = ¬P (V ) for its complement (as it is often used as
target for bug search). With abuse of notation, in the rest of
this paper, we make no distinction between the characteristic
function of a set and the set itself.
_T_ (V, W, V _[′]) is the transition relation, that we assume given_
by a circuit graph, with state variables mapped to latches.
Present and next state variables correspond to latch outputs
and inputs, respectively. The input of the i−th latch is fed by
a combinational circuit, described by the δi(V, W ) Boolean
function. Hence, the transition relation can be expressed as
� �
_T_ (V, W, V _[′]) =_ _ti(V, W, vi[′][) =]_ (vi = δi(V, W ))
_i_ _i_
A state path of length k is a sequence of states σ0, . . ., σk
such that T (σi, νi, σi+1) is true, given some input νi, for all
0 ≤ _i < k._
A state set S[′] is reachable from state set S in k steps if
there exists a path of length k, in the labeled state transition
structure, connecting a state belonging to S to another one
belonging to S[′]; equivalently
_S(V_ [0]) ∧ [�]i[k]=0[−][1] _[T]_ [(][V][ i][, W][ i][, V][ i][+1][)][ ∧] _[S][′][(][V][ k][)]_
The image operator IMG(T, From) computes the set of states
_T o reachable in one step from the states in From_
_T o(V_ _[′])_ = IMG(T (V, W, V _[′]), From(V ))_
= _∃V,W (From(V ) ∧_ _T_ (V, W, V _[′]))_
An over-approximate image (or pre-image) is any state set
including the exact image
_T o[+]_ = IMG[+](T, From) _⊇_ IMG(T, From)
Pre-image is dual, with the only difference that existential
quantification of functionally computed state variables can be
operated by composition:
_T o(V )_ = PREIMG(T (V, W, V _[′]), From(V_ _[′]))_
= _∃W,V ′(From(V_ _[′]) ∧_ _T_ (V, W, V _[′]))_
= _∃W From(δ(V, W_ ))
_B. Bounded Model Checking_
SAT-based Bounded Model Checking (BMC) [3] considers
only k−bounded reachability, as expressed by the propositional formula
BMC[k]0 = _I(V_ [0]) ∧ [�]i[k]=0[−][1] _[T][ (][V][ i][, W][ i][, V][ i][+1][)][ ∧]_ _[F]_ [(][V][ k][)]
A bounded proof is thus translated into a SAT problem. BMC[k]0
is satisfiable iff there is a counter-example (a path from I to
1V i..j is defined if i ≤ _j, otherwise we conventionally define it as a void_
variable set
_F_ ) of length k. In the case of circuits, existential quantification
is conveniently applied to intermediate sets of state variables
BMC[k]0 [=] _I(V_ [0]) ∧∃V 1..k ([�][k]i=0[−][1] _[T]_ [(][V][ i][, W][ i][, V][ i][+1][)][ ∧] _[F]_ [(][V][ k][))]
= _I(V_ [0]) ∧ _Conek(V_ [0], W [0][..k][−][1])
where Conek represents a combinational single output circuit
unrolling, formally defined exploiting quantification by functional composition
_Conek_ = _Conek(V_ [0], W [0][..k][−][1])
= _∃V 1..k_ ([�][k]i=0[−][1][(][V][ i][+1][ =][ δ][i][(][V][ i][, W][ i][))][ ∧] _[F]_ [(][V][ k][))]
= _F_ (δ(...(δ(W [0], V [0]))), W _[k][−][1])_
The main advantage of using Conek (a combinational circuit)
instead of transition relation instances ([�]i _[T][ i][), is that several]_
circuit-based simplifications, besides Cone-Of-Influence (COI)
reductions, are possible on Conek, from constant propagations
to combinational optimizations, before moving to CNF- or
circuit-based SAT.
_C. State sets and Fix-points in Unbounded Model Checking_
Although reachability is usually formulated in terms of
the image and/or pre-image operators, we will here express
backward reachability using Conek, as previously defined.
SAT-based model checking approaches generally keep explicit
representations of circuit unrollings instead of (exact) state
set representations, due to the inherent complexity of quantification operators. The set of states (backward) reachable
from F in (exactly) k steps can be obtained by primary input
quantification over a circuit unrolling
_BckRk(V ) = ∃W 0..k−1_ _Conek(V, W_ [0][..k][−][1])
The overall set of backward reachable states is the union of
all reachable states up to depth k
_BckR0..k(V )_ = �ki=0 _[BckR][i][(][V][ )]_
= �ki=0 _[∃][W][ 0][..i][−][1]_ _[Cone][i][(][V, W][ 0][..i][−][1][)]_
= _∃W 0..k−1_ [�]i[k]=0 _[Cone][i][(][V, W][ 0][..i][−][1][)]_
where distributivity of existential quantification over union has
been applied. We introduce the short notation Cone0..k for
�k
_i=0_ _[Cone][i][. So backward reachable states are defined as]_
_BckR0..k(V ) = ∃W0..k−1_ _Cone0..k(V, W0..k−1)_
A backward reachability fix-point could be checked by a SAT
run on the following Boolean formula
_BckRk+1(V ) ∧¬BckRk(V )_
or in a simpler one, with quantified state sets on the second
term only
_Conek+1(V, W0..k) ∧¬BckRk(V )_
Unfortunately, both the above formulations are difficult to
manipulate in many practical cases, due to the complexity of
SAT-based quantification.
-----
_D. Craig Interpolants in Model Checking_
Given two inconsistent formulas A and B (A ∧ _B = 0), an_
interpolant C is a formula such that:
1) It is implied by A
2) It is inconsistent with B, i.e., C ∧ _B is unsatisfiable_
3) It is expressed over the common variables of A and B.
A Craig interpolant C = ITP (A, B) is an AND/OR circuit
that can be computed in linear time from the refutation proof
of A∧B. Albeit the computation is linear, the refutation proof
itself can be exponentially larger than A and B.
A _k-adequate_ over-approximate image is an
IMG[+](T, From) that does not intersect any state on
paths of length k to F . Using the Cone0..k circuit unrolling,
a k-adequate over-approximate image is
IMG[+]Adq[(][T, From, Cone][0][..k][)]
defined as follows[2]:
IMG[+]Adq [is][ undefined][ iff]
_From(V ) ∧_ _T_ (V, W0, V _[′]) ∧_ _Cone0..k(V_ _[′], W_ [1][..k]) ̸= 0
Otherwise, it is computed by interpolation:
IMG[+]Adq[(][T, S, Cone][0][..k][)] = ITP(S(V ) ∧ _T_ (V, W0, V _[′]),_
_Cone0..k(V_ _[′], W1..k))_
An image is called adequate if it is k-adequate for any k, i.e.,
no path of any length can lead from a state within the image
to states in F . Since the model is finite, a k-adequate image
is adequate if k ≥ _d, where d is the diameter of the state_
transition graph.
McMillan [15] proposed an effective fully SAT-based Unbounded Model Checking algorithm, exploiting interpolants,
as sketched in Figure 1.
INTERPOLANTMC (I, T, F )
k = 0
**do**
_Cone0..k = CIRCUITUNROLL(F, δ, k)_
res = FINITERUN (I, T, Cone0..k)
k = k + 1
**while (res = undecided)**
FINITERUN (I, T, Cone)
**if (SAT(I ∧** _T ∧_ _Cone)_
**return (reachable)**
R = I
**while (true)**
_To = IMG[+]Adq_ [(][T] [, R,][ Cone][)]
**if (To = undefined)**
**return (undecided)**
**if (To ⇒** R)
**return (unreachable)**
R = R ∨ _To_
Fig. 1. Interpolant-based Verification.
While INTERPOLANTMC is the entry point of the algorithm, routine FINITERUN takes care of the interpolant-based
2Indexes of input variable sets have been shifted up in Conek, from
_W_ [0][..k][−][1] to W _[i..k], in order to use W_ [0] variables in the forward T instance.
over-approximate traversal. The latter function may end up
with three possible results:
_• “reachable”, if it proves F reachable in k steps, hence_
the property has been disproved
_• “unreachable”, if the approximate traversal using the_
IMG[+]Adq [image computation reaches a fix-point. In this]
case the property is proved
_• “undecided”, if F is intersected by the over-approximate_
state sets. Then, k in increased for a new FINITERUN
call.
McMillan [15] proved that the previous algorithm is sound
and complete. In synthesis, let us assume I and F mutually
unreachable: if k < d, a k-adequate set can produce a non
_k-adequate image. In this case, “undecided” is returned and_
_k is increased. Otherwise, when k ≥_ _d, IMG[+]Adq_ [is always]
adequate. At this point, the algorithm will terminate with an
approximate reachability fix-point.
According to [17], k can be incremented by the depth of
the last FINITERUN execution to avoid a quadratic number of
image computations.
III. SAT-BASED OVER-APPROXIMATE IMAGE
The over-approximate images we are concerned with are
defined by the following two conditions:
_• An over-approximate image includes the exact image._
Avoiding existential quantification, a given T o[+](V _[′]) is_
an over-approximate image if:
_T_ (V, W0, V _[′]) ∧_ _From(V ) ⇒_ _T o[+](V_ _[′])_
_• We look for k−adequate images, as adequacy (by increas-_
ing k values) is the condition adopted to incrementally
tighten abstractions. Given Cone0..k
_T o[+](V_ _[′]) ∧_ _Cone0..k(V_ _[′], W_ [1][..k]) = 0
Given the above observations, we look for an overapproximate image, computed as a Boolean conjunction of
several atomic over-approximations
_T o[+](V_ _[′]) =_ [�]i _[to][i][+][(][V][ ′][)]_
characterized by the property that each toi[+] factor includes
the exact image, but only the [�]i _[to][i][+][ product is][ k][−][adequate]_
(whereas each toi[+] is not required to be k−adequate)
_∀i, T_ (V, W0, V _[′]) ∧_ _From(V )_ _⇒_ _toi[+](V_ _[′])_
�i _[to][i][+][(][V][ ′][)][ ∧]_ _[Cone][0][..k][(][V][ ′][W][ 1][..k][)]_ = 0
Hence, we look for a product of toi[+] factors, where each
one is an over-approximate image in itself, and the overall
product is k−adequate. We choose to possibly find several
(small) toi[+], instead of a single T o[+], as we adopt an iterative algorithm to select among multiple (small) candidate
over-approximations. Whereas it might be computationally
infeasible to enumerate all functions of the (V’) variables as
candidate over-approximations, we explore given classes of
atomic functions.
-----
The second criterion we adopt is the incremental refinement
of an initially coarse T o[+] abstraction:
_• We consider toi[+]_ candidates, selected by incremental
complexity. We group them by iterating through classes
of functions, so that we first capture the simplest/easiest
ones, before moving to more complex candidates
_• Whenever we find an over-approximation, we possibly_
use it to simplify the data structure for the next steps
(e.g. the transition relation and the backward cone used
for adequacy checks)
_• We_ stop the iterative process whenever the overapproximation is adequate. As the selection we operate
is not complete, we possibly end up computing a Craig
interpolant (on the simplified backward cone and forward
transition relation) from a SAT refutation proof.
Figure 2 shows the skeleton of our iterative image overapproximation algorithm.
IMG[+]Adq [(][T] [,][ From][,][ Cone][)]
_To[+]_ = 1
Class = constClass
**do**
Candclass = GETCANDIDATES (Class)
Abstrclass = { }
**foreach cj (V’) ∈** CandClass
**if ACCEPT (ci, T**, From, Cone)
Abstrclass = Abstrclass ∪{cj _}_
_To[+]class_ [=][ OVER][A][PPR][S][ET][ (Abstr][class][)]
_To[+]_ = To[+] _∧_ _To[+]class_
**if ¬SAT (To[+]** _∧_ _Cone)_
**return (To[+])**
SIMPLIFY (Cone, Abstrclass)
SIMPLIFY (T, Abstrclass)
Class = NEXTCLASS()
**while (Class ̸= emptyClass)**
**return (To[+]** _∧_ ITP (From∧ _T_, Cone))
Fig. 2. Over-approximate k-adequate image computation.
Let Class represent the choice for the type of overapproximations to be considered: we start by constClass,
the equivalence class of _constant_ _variables,_ then
NEXTCLASS() returns, at each new iteration, equivalences,
implications, ternary abstraction and localization abstraction.
For each class, we iterate through the cj candidates, in
order to gather the accepted ones. A candidate is accepted
if it is a valid over-approximation, i.e. adequacy conditions
are kept after applying the over-approximation. Once all
accepted candidates for a given class have been selected,
we generate the class over-approximate image (T o[+]class [=]
OVERAPPRSET(Abstrclass)). T o[+]class [is a new refinement for]
(and it is and-ed to) the overall image T o[+]. If the refined
_T o[+]_ is adequate (¬SAT (T o[+] _∧_ _Cone)), it is returned as_
a result, otherwise we exploit the set of atomic abstractions
(Abstrclass) to simplify Cone and T, and we move to the
next class.
When we have looped through all classes, without returning
an adequate T o[+] set, this means that the over-approximation is
still too coarse, and we end up calling SAT-based interpolation
(which is possibly easier due to all previous simplifications on
_T and Cone)._
All of the above mentioned classes will be discussed in
section IV, where, for each one, we motivate it, we discuss
the abstraction obtained, and how to efficiently compute
_k−adequacy and to simplify Cone and T_ .
IV. ABSTRACTION CLASSES AND ADEQUACY CHECKS
Abstraction classes are now individually discussed, as well
as the related acceptance criteria, set over-approximation and
simplification strategies.
_A. Constant variables_
We first look for variables equivalent to constant values,
i.e. variables that are implied to constant values in the next
state. More formally, let variable x[′]j _[∈]_ _[X]_ _[′][ be the generic next]_
state variable, we consider two possible over-approximation
literal candidates: ¬x[′]j [and][ x]j[′] [. Therefore, the candidate class]
returned by GETCANDIDATES(constClass) is the set of all
(direct and complemented) literals:
GETCANDIDATES(constClass) = {x[′]1[,][ ¬][x][′]1[, . . ., x][′]n[,][ ¬][x][′]n[}]
We efficiently accept/reject individual candidates by an incremental SAT procedure that iterates through all literals, and
detects the implied ones. An accepted literal lj is such that:
_S ∧_ _T ⇒_ _lj_
The next step is the explicit computation of T o[+]constClass
(function OVERAPPRSET), as the conjunction of all implied
literals.
Given a cube, the simplification task (performed by the
function SIMPLIFY(T, AbstrconstClass)) is straightforward.
We just need to replace variables by the corresponding constant values, and this action generally reduces the overall AIG
node count for Cone. SIMPLIFY(T, AbstrconstClass)) is easier,
as we simply need to remove the T components corresponding
to constant values.
It is worth noticing that constant state variables typically
appear at the initial steps of forward traversals and possibly
throughout the traversal of externally constrained (by external
assumptions) systems.
Keeping exact values for implied variables is a way to
tighten over-approximation vs. other abstraction techniques.
Craig interpolants as well as localization abstraction could,
for instance, abstract away implied variables (whenever they
are not relevant for the proof). But this might result in
a looser over-approximation, and possibly trigger visits of
unreachable states. By explicitly expressing T o[+]constClass[, we]
also remove implied variables from further computations (as
in other abstraction techniques), but the T o[+]constClass [factor]
in the overall over-approximate image will constrain the corresponding present state variables at their exact value, at the
next forward traversal iteration.
-----
_B. Equivalences and Implications_
After detecting constant state variables, we look for equivalences (modulo complementation) between couples of next
state variables.
Let literals li and lj be two literals to be considered for
equivalence, the corresponding candidate equivalence is eqij =
_li ⇔_ _lj._
The set of candidate equivalences is obviously quadratic in
the number of variables, as well as the iteration for acceptance
checks. Again we can exploit efficient solutions based on
equivalence classes and incremental SAT, inspired by the ones
proposed in [7], [8].
_T o[+]eqClass_ [(function][ OVER][A][PPR][S][ET][) is straightforwardly]
computed as the conjunction of all proved equivalences.
Function simplification given a set of equivalences is performed by means of variable merging (for each equivalence
class we substitute each variable literal with a class representative literal).
Once equivalences have been detected and used to simplify
_Cone and T_, we consider variable implications, again, modulo complementation. Implications are accepted similarly to
equivalences, but their computation cannot rely on equivalence
classes any more. Furthermore, implications cannot be used
for direct function simplification. We just need to explicitly
represent them (T o[+]Impl[) in order to tighten the candidate][ T o][+][.]
We also explicitly add the T o[+]Impl [constrain to][ Cone][ as]
an additional redundant factor, for space search restriction
purposes.
_C. Ternary abstraction_
Three-Valued Logic is at the base of several synthesis and
verification approaches, following the general idea to encode
an additional logic value, representing either the unknown, or
the {0, 1} set. Our approach directly follows an idea proposed
by Bres et al. [22], following works by Malik [23] and Shiple
et al. [24]. Other related works can be found in the field of
equivalence checking [25], where ternary logic is used in order
to handle circuit initialization sequences.
The cited works are inspired from Scott’s three-valued logic,
built upon the usual two-valued Boolean logic by adding a
third value, that denotes the undefined or unknown value.
_⊥_
Bres et al. [22] adopt a two bit encoding (sometimes called
dual rail) for ternary constants ((0, 1) for false, (1, 0) for
_true, (0, 0) for ⊥). A Boolean function f is represented by_
two bit functions f0 and f1, such that f0 (resp. f1) is the
characteristic function of the set for which f evaluates to 0
(resp. 1). f⊥ = ¬f0 ∧¬f1 is the set for which the function
is undefined (the don’t care set). The function is completely
defined if f1 ∨ _f0 = 1, and f⊥_ = 0. Let us introduce the
notation f [3] = (f0, f1) for such a dual rail encoding.
Boolean operators for over the ternary encoding are defined
by the following rules:
_¬f_ [3] = _¬(f0, f1)_ = (f1, f0)
_f_ [3] _∨_ _g[3]_ = (f0, f1) ∨ (g0, g1) = (f0 ∧ _g0, f1 ∨_ _g1)_
_f_ [3] _∧_ _g[3]_ = (f0, f1) ∧ (g0, g1) = (f0 ∨ _g0, f1 ∧_ _g1)_
Let si be a Binary variable, and σi the corresponding ternary
one. Abstraction of variable σi in f [3](σ1, . . ., σi, . . ., σn) was
done in [22], by setting variable σi to the unknown value:
_f_ [3](σ1, . . ., ⊥, . . ., σn)
In the dual rail encoding, the σi variable is assigned the
_⊥_ = (0, 0) ternary constant, whereas all other σj variables
are encoded by the symbolic ternary value (sj, ¬sj ) (using
the corresponding binary variable).
Given the above assignments, function f [3] has now a
possibly non-void don’t care set f⊥. And a binary overapproximation of f can be obtained by ¬f0 = f1 ∨ _f⊥_ _⊇_ _f_ .
_T ernaryAbs (f_ (s1, . . ., si, . . ., sn), si) =
_¬f0((s1, ¬s1), . . ., (0, 0), . . ., (sn, ¬sn))_
Informally, we have replaced the characteristic function of the
set ”on which f is true” (the onset of f ), by a superset ”on
which f is certainly not false”.
We apply ternary abstraction to our over-approximate image
computation, by looping through all primary input and state
variables, and iteratively selecting them as possible ternary
abstraction candidates. Ternary abstraction of the generic
state/input variable si could violate an adequacy condition.
This is the reason why it is accepted just if it still guarantees
adequacy:
_From(V )_ _∧_ _ternaryAbs(T (V, W_ [0], V _[′]), si)_
_∧_ _ternaryAbs(Conek(V_ _[′], W_ [1][..k]), si) = 0̸
The si variable can be one of the V _[′]_ state variables, or
the W [0][..k] input variables. An accepted ternary abstraction
does not directly produce any over-approximation of the
image (T o[+]) (unless all W [0] variables are quantified). Hence,
OVERAPPRSETternaryClass generally returns no abstraction
(T o[+]ternaryClass [= 1][. On the other hand, ternary abstraction]
is directly applied to simplify T and/or Cone.
_D. Localization abstraction_
Localization abstraction is our last attempt to produce
an over-approximation. A candidate abstraction corresponds
to letting a state variable be free at the forward/backward
boundary, i.e. simply re-labeling the chosen vi[′] [variable in]
_Cone(V_ _[′], W_ [1][..k]) by a fresh new variable.
The abstraction process is inspired by our previous
work [18], and is managed analogously to ternary abstraction.
V. EXPERIMENTAL RESULTS
We implemented our algorithms on top of the PdTrav tool,
a state-of-the-art verification framework which won two of the
sub-categories at the 2007 Model Checking competition [20].
We compared results of our tool with and without the proposed
methodology.
Our experiments ran on a Dual-Core Pentium IV 3 GHz
Workstation with 3 GByte of main memory, running Debian
Linux. We performed extensive tests, by specifically addressing proofs of correctness. For each verification instance, we
used a 900 seconds time limit.
-----
900
800
700
600
500
400
300
200
100
0
0 100 200 300 400 500 600 700 800 900
Original time [sec]
problems, as shown by the results on the large industrial
benchmarks.
VI. CONCLUSIONS
This paper addresses improvements to Interpolant-based
model checking by means of an integrated approach exploiting
over-approximation techniques that are novel for this field.
We describe an integrated approach for image computation, incrementally combining and tuning different techniques
within a unified SAT-based (complete) Unbounded Model
Checking approach.
The method we propose adopts the general skeleton of
interpolant-based model checking procedure, and exploits preliminary SAT-based exploration of candidate atomic abstractions, within a global effort to tighten over-approximations and
keep state set sizes under control.
Experimental results, specifically oriented to hard verification problems, show the robustness of our approach implemented on a state-of-the-art verification framework.
REFERENCES
Fig. 3. Verification time on circuits coming from [20], [21].
We present results on circuits derived from the Model
Checking competition [20], [21], and a few standard benchmarks from the VIS distribution [26], with particular emphasis
on true hard-to-solve instances. The reason for reporting only
true properties verification data is the ability of BMC in being
the most effective technique for checking falsifications (i.e.,
false properties), as showed by the competition results.
The scattered plot in Figure 3 shows verification times on
152 benchmarks from the Model Checking competition. It
compares a standard interpolant-based UMC technique (time
on the X-axis) against the same strategy improved as suggested
in this paper (time on the Y-axis). Figure 3 clearly shows a
set of “easy” benchmarks, i.e., the ones which are solved in
few seconds/minutes for both the techniques. The plot also
highlights a set of problems that were not solved within the
900 seconds time limit with the standard interpolant computation, while they are solved with the proposed optimization.
The overall results clearly show the robustness of the proposed
approach.
Table I reports more detailed data on a few selected hard-tosolve verification instances. The meaning of columns follows:
Model is instance name (industrial designs names have intentionally been hidden), # PI, # FF and # NODES represent
the number of primary inputs, memory elements and AIG
nodes of the circuit respectively; finally, Std ITP, Method A
and Method B provide the verification time in seconds, with
different strategies. Ternary abstraction was disabled in column
Method A, while it was enabled in column Method B. The
data clearly show that most of the problems could already be
completed without resorting to ternary abstraction (see column
Method A), whereas the latter plays an important role in a few
cases. An inner look at those experiments showed that ternary
abstraction requires a time overhead for acceptance checks,
which is not worth the advantage gained with the abstraction,
in several cases.
Though we still need to further enhance the self-tuning
capability of our approach, it is already able to attack large
[1] J. R. Burch, E. M. Clarke, D. E. Long, K. L. McMillan, and D. L. Dill,
“Symbolic Model Checking for Sequential Circuit Verification,” IEEE
_Trans. on Computer-Aided Design, vol. 13, no. 4, pp. 401–424, Apr._
1994.
[2] R. E. Bryant, “Graph–Based Algorithms for Boolean Function Manipulation,” IEEE Trans. on Computers, vol. C–35, no. 8, pp. 677–691, Aug.
1986.
[3] A. Biere, A. Cimatti, E. M. Clarke, M. Fujita, and Y. Zhu, “Symbolic
Model Checking using SAT procedures instead of BDDs,” in Proc.
_36th Design Automat. Conf._ New Orleans, Louisiana: IEEE Computer
Society, Jun. 1999, pp. 317–320.
[4] P. Bjesse and A. Boralv, “DAG-Aware Circuit Compression For Formal
Verification,” in Proc. Int’l Conf. on Computer-Aided Design. San Jose,
California: IEEE Computer Society, Nov. 2004.
[5] P. Bjesse and K. Claessen, “SAT–Based Verification without State Space
Traversal,” in Proc. Formal Methods in Computer-Aided Design, ser.
LNCS, vol. 1954. Austin, TX, USA: Springer, 2000.
[6] M. L. Case, A. Mishchenko, and R. K. Brayton, “Inductively Finding
a Reachable State Space Over-Approximation,” in Proc. Int’l Workshop
_on Logic Synthesis, Lake Tahoe, California, May 2006._
[7] F. Lu and K. T. Cheng, “IChecker: An Efficient Checker for Inductive
Invariants,” in High-Level Design Validation and Test Workshop, 2006,
pp. 176–180.
[8] G. Cabodi, S. Nocco, and S. Quer, “Boosting the Role of Inductive
Invariants in Model Checking,” in Proc. Design Automation & Test in
_Europe Conf._ Nice, France: IEEE Computer Society, Apr. 2007.
[9] M. Sheeran, S. Singh, and G. St˚almarck, “Checking Safety Properties Using Induction and SAT Solver,” in Proc. Formal Methods in
_Computer-Aided Design, ser. LNCS, W. A. Hunt and S. D. Johnson,_
Eds., vol. 1954. Austin, Texas, USA: Springer, Nov. 2000, pp. 108–
125.
[10] P. F. Williams, A. Biere, E. M. Clarke, and A. Gupta, “Combining
Decision Diagrams and SAT Procedures for Efficient Symbolic Model
Checking,” in Proc. Computer Aided Verification, ser. LNCS, E. A.
Emerson and A. P. Sistla, Eds., vol. 2102. Chicago, Illinois: SpringerVerlag, Jul. 2000, pp. 124–138.
[11] P. A. Abdulla, P. Bjesse, and N. Een, “Symbolic Reachability Analysis
based on SAT-Solvers,” in Tools and Algorithms for the Construction
_and Analysis of Systems, M. I. S. Susanne Graf, Ed., vol. 1785._ Berlin,
Germany: Springer-Verlag, Apr. 2000, pp. 411–425.
[12] K. L. McMillan, “Applying SAT Methods in Unbounded Symbolic
Model Checking,” in Proc. Computer Aided Verification, ser. LNCS,
E. Brinksma and K. G. Larsen, Eds., vol. 2404. Copenhagen, Denmark:
Springer, 2002, pp. 250–264.
[13] H. J. Kang and L. C. Park, “SAT-based unbounded symbolic model
checking,” in Proc. 40th Design Automat. Conf. Anaheim, CA: IEEE
Computer Society, 2003, pp. 840–843.
-----
TABLE I
VERIFICATION DATA ON HARD-TO-SOLVE EXPERIMENTS (− MEANS OVERFLOW, BOLD FONTS ARE USED FOR best RESULTS)
Model #PI #FF #NODES Std ITP Method A Method B
intel 005.blif 165 170 1776 (-) 26.88 **22.15**
intel 006.blif 345 350 3265 (-) **518.82** (-)
intel 020.blif 349 354 5735 (-) 757.51 **646.22**
intel 021.blif 360 365 5882 (-) **410.02** (-)
intel 024.blif 352 357 5710 (-) (-) **437.55**
intel 026.blif 486 492 6263 (-) 215.04 **211.10**
intel 029.blif 559 564 8816 (-) (-) **347.52**
sfeistel-inv.blif 68 296 6837 304.44 **111.51** 555.18
blackjack-inv.blif 5 103 3979 (-) 96.51 **35.58**
31 2 batch 1.blif 24 122 1506 (-) 23.12 **23.00**
soap-inv.blif 11 140 3605 (-) 199.15 **172.48**
intel 049.blif 136 77 1305 254.29 (-) **193.83**
nusmvguidancep9.blif 84 86 1902 359.59 **249.91** 357.99
pdtvisns3p09.blif 21 101 3770 386.14 **315.51** 747.10
pdtvisvsa16a29.blif 32 172 7016 663.12 753.26 **306.39**
visprodcellp22.blif 30 63 2771 147.09 216.99 **116.42**
cmu.periodic.N.blif 32 34 1555 230.59 75.84 **35.12**
nusmv.guidance[∧]2.C.blif 84 86 1920 251.45 **134.33** 220.22
nusmv.guidance[∧]6.C.blif 84 86 1901 347.76 185.84 **149.77**
nusmv.guidance[∧]7.C.blif 84 86 2001 **91.26** 140.41 134.88
nusmv.guidance[∧]8.C.blif 84 86 1919 424.77 678.26 **392.52**
nusmv.reactor[∧]6.C.blif 74 76 1396 (-) (-) **475.04**
vis.coherence[∧]2.E.blif 6 29 1216 111.35 **60.01** 82.17
vis.coherence[∧]3.E.blif 6 29 1214 649.73 60.57 **49.18**
industrial1.blif 120 76 1089 (-) **93.43** 289.62
industrial2.blif 119 79 1103 (-) **96.90** 224.40
industrial3.blif 119 78 1100 (-) **255.72** 261.15
industrial4.blif 138 97 2172 (-) 422.33 **326.60**
industrial5.blif 113 459 7666 (-) 112.32 **108.67**
industrial6.blif 52 187 3600 126.08 **59.31** 75.16
[14] M. K. Ganai, A. Gupta, and P. Ashar, “Efficient SAT-based Unbounded
Symbolic Model Checking Using Circuit Cofactoring,” in Proc. Int’l
_Conf. on Computer-Aided Design. San Jose, California: IEEE Computer_
Society, Nov. 2004.
[15] K. L. McMillan, “Interpolation and SAT-Based Model Checking,” in
_Proc. Computer Aided Verification, ser. LNCS, W. A. H. Jr. and_
F. Somenzi, Eds., vol. 2725. Boulder, CO, USA: Springer, 2003, pp.
1–13.
[16] K. L. McMillan and R. Jhala, “Interpolation and SAT-Based Model
Checking,” in Proc. Computer Aided Verification, ser. LNCS, T. Ball
and R. B. Jones, Eds., vol. 3725. Edimburgh, Scotlan, UK: Springer,
2005, pp. 39–51.
[17] J. Marques-Silva, “Improvements to the implementation of Interpolant–
Based Model Checking,” in Proc. Computer Aided Verification, ser.
LNCS, D. Borrione and W. Paul, Eds., vol. 3725. Edimburgh, Scotlan,
UK: Springer, 2005, pp. 367–370.
[18] G. Cabodi, S. Nocco, M. Murciano, and S. Quer, “Stepping Forward
with Interpolants in Unbounded Model Checking,” in Proc. Int’l Conf.
_on Computer-Aided Design._ San Jose, California: ACM Press, Nov.
2006.
[19] B. Li and F. Somenzi, “Efficient Abstraction Refinement in
Interpolation-Based Unbounded Model Checking,” in Tools and Algo_rithms for the Construction and Analysis of Systems, vol. 3920, 2006,_
pp. 227–241.
[20] A. Biere and T. Jussila, “The Model Checking Competition Web Page,
http://fmv.jku.at/hwmcc07/organizers.html,” 2007.
[21] ——, “The Model Checking Competition Web Page,
http://fmv.jku.at/hwmcc08/organizers.html,” 2008.
[22] A. B. Y. Bres, G. Berry and E. M. Sentovich, “State Abstraction
Techniques for the Verification of Synchronous Circuits,” in dcc02:
_Designing Correct Circuits 2002, Grenoble, France, Apr. 2002._
[23] S. Malik, “Analysis of Cyclic Combinational Circuits,” vol. 13, no. 7,
1994, pp. 950–956.
[24] G. B. T. R. Shiple and H. Touati, “Constructive Analysis of Cyclic e
e Circuits,” in IDTC’96: International Design and Testing Conference,
Paris, France, 1996.
[25] Z. Khasidashvili and Z. Hanna, “Sat-based methods for sequential
hardware equivalence verification without synchronization,” in BMC’03:
_First International Workshop on Bounded Model Checking, Boulder,_
Colorado, Jul. 2003, pp. 593–607.
[26] R. K. B. et al., “VIS,” in Proc. Formal Methods in Computer-Aided
_Design, ser. LNCS, M. Srivas and A. Camilleri, Eds., vol. 1166._ Palo
Alto, California: Springer, Nov. 1996, pp. 248–256.
-----
|
{
"disclaimer": "Notice: The following paper fields have been elided by the publisher: {'abstract', 'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1109/ICCAD.2008.4681563?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/ICCAD.2008.4681563, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "GREEN",
"url": "http://www.cecs.uci.edu/%7Epapers/iccad08/PDFs/Papers/02B.2.pdf"
}
| 2,008
|
[
"JournalArticle",
"Conference"
] | true
| 2008-11-10T00:00:00
|
[] | 12,619
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Engineering",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/0291fb253d7567a106da907b2fe867cec86be541
|
[
"Computer Science"
] | 0.872326
|
A directory service for configuring high-performance distributed computations
|
0291fb253d7567a106da907b2fe867cec86be541
|
Proceedings. The Sixth IEEE International Symposium on High Performance Distributed Computing (Cat. No.97TB100183)
|
[
{
"authorId": "143767485",
"name": "Steven M. Fitzgerald"
},
{
"authorId": "1698701",
"name": "Ian T Foster"
},
{
"authorId": "8682509",
"name": "C. Kesselman"
},
{
"authorId": "1745570",
"name": "G. Laszewski"
},
{
"authorId": "2581430",
"name": "Warren Smith"
},
{
"authorId": "1720669",
"name": "S. Tuecke"
}
] |
{
"alternate_issns": null,
"alternate_names": null,
"alternate_urls": null,
"id": null,
"issn": null,
"name": null,
"type": null,
"url": null
}
| null |
**Government retains lor itself, and others act-**
**ing on its behalf, a paid-up, nonexclusive,**
**irrevocable worldwide license in said article**
**to reproduce, prepare derivative works, dis-**
**tribute copies to the public, and perform pub-**
**licly and display publicly, by or** **on behalf of**
###### A Directory Service for Configuring the Government. High-Performance Distributed Computations
#### 1V
###### Steven Fitzgerald,' Ian Foster: Carl Kesselman,' Gregor von Laszewski?
JUL 0 7
Warren Smith? Steven Tuecke2
```
0 ST.1
Information Sciences Institute Mathematics and Computer Sc,;nce University of Southern California Argonne National Laboratory Marina del Rey, CA 90292 Argonne, IL 60439
http://www.globus.org/
Abstract standard default protocols, interfaces, and so on. The situ-
```
ation is also quite different in traditional high-performance
_High-pelfotmance execution in distributed computing_ computing, where systems are usually homogeneous and
_envimnments ofren requires careful selection and configu-_ hence can be configured manually. But in high-performance
_ration not only of computers, networks, and other resources_ distributed computing, neither defaults nor manual config-
_but also of the protocols and algorithms used by applica-_ uration is acceptable. Defaults often do not result in ac-
_tions. Selection and configuration in turn require access_ ceptable performance, and manual configuration requires
**_to_** _accurate, up-to-date information on_ _the structure and_ low-level knowledge of remote systems that an average
_state_ _of_ _available resources. Unfortunately, no standard_ programmer does not possess. We need an _infonnation-_
_mechanism exists for organizing or accessing such i n f o m -_ _rich approach to configuration in which decisions are made_
_tion. Consequently, diferent tools andapplications adopt ad_ (whether at compile-time, link-time, or run-time [ 191) based
_hoc mechanisms, or t h q compromise their portability and_ upon information about the structure and state of the system
_pegormance by using default configurations. We propose_ on which a program is to run.
_a Metacomputing Directory Service that provides efficient_ An example from the I-WAY networking experiment il-
_and scalable access to diverse, dynamic, and distributed_ lustrates some of the difficulties associated with the configu-
_information about resource structure and state. We define_ ration of high-performance distributed systems. The I-WAY
_an extensible data model to represent required information_ was composed of massively parallel computers, worksta-
_and present a scalable, high-perfotmance, distributed im-_ tions, archival storage systems, and visualization devices 161.
_plementation. The data representation and application pro-_ These resources were interconnected by both the internet
_gramming interface are adopted from the Lightweight Di-_ and a dedicated 155 Mbisec IP over ATM network. In this
_rectory Access Protocol; the data model and implementation_ environment, applications might run on a single or multi-
_are new. We use the Globus distributed computing toolkit to_ ple parallel computers, of the same or different types. An
_illustrate how this directory service enables the development_ optimal communication configuration for a particular situa-
`of more flexible and efficient distributed computing services` tion might use vendor-optimized communication protocols
_and applications._ within a computer but TCP/IP between computers over an
**ATM** network (if available). A significant amount of infor-
mation must be available to select such configurations, for
###### 1 Introduction example:
High-performance distributed computing often requires **_0_** What are the network interfaces (i.e., IP addresses) for
careful selection and configuration of computers, networks, the ATM network and Internet?
application protocols, and algorithms. These requirements
do not arise in traditional distributed computing, where con- What is the raw bandwidth of the **ATM** network and
figuration problems can typically be avoided by the use of the Internet, and which .
###### I
-----
###### DISCLAIMER
This report was prepared as an account of work sponsored by an agency of the United
States Government. Neither the United States Government nor any agency thereof, nor
any of their employees, make any warranty, express or implied, or assumes any legal liabili-
ty or responsibility for the accuracy, completeness, or usefulness of any information, appa-
ratus, product, or process disclosed, or represents tbat its use would not infringe privately
owned rights. Reference herein to any specific commercial product, process, or service by
trade name, trademark, manufacturer, or otherwise does not necessarily **comtitute or**
imply its endorsement, recommendation, or favoring by the United States Government or
any agency thereof. The views and opinions of authors expressed herein do not necessar-
ily state or reflect those of the United States Government or any agency thereof.
-----
##### Portions of this document may be iiiegiile
###### in electronic image products. Images are produced from the best available original document.
-----
**_0_** Is the ATM network currently available? **_0_** a demonstration of the use of the information provided
by MDS to guide resource and communication config-
**_0_** Between which pairs of nodes can we use vendor pro- uration within a distributed computing toolkit.
tocols to access fast internal networks?
The rest of this article is organized as follows. In Sec-
**_0_** Between which pairs of nodes must we use TCP/IP? tion 2, we explain the requirements that a distributed com-
puting information infrastructure must satisfy, and we pro-
Additional information is required if we use a resource lo- pose MDS in response to these requirements. We then de-
cation service to select an "optimal" set of resources from scribe the representation (Section 3), the data model (Sec-
among the machines available on the I-WAY at a given time. tion 4), and the implementation (Section **_5) of MDS. In_**
In our experience, such configuration decisions are not Section 6, we demonstrate how MDS information is used
difficult if the right information is available. Until now, within Globus. We conclude in Section 7 with suggestions
however, this information has not been easily available, and for future research efforts.
this lack of access has hindered application optimization.
Furthermore, making this information available in a useful
###### 2 Designing a Metacomputing Directory Ser-
fashion is a nontrivial problem: the information required to
###### vice
configure high-performance distributed systems is diverse
in scope, dynamic in value, distributed across the network,
and detailed in nature. The problem of organizing and providing access to in-
In this article, we propose an approach to the design formation is a familiar one in computer science, and there
of high-performance distributed systems that addresses this are many potential approaches to the problem, ranging from
need for efficient and scalable access to diverse, dynamic, database systems to the Simple Network Management Proto-
and distributed information about the structure and state col (SNMP). The appropriate solution depends on the ways
of resources. The core of this approach is the definition in which the information is produced, maintained, accessed,
and implementation of a Metacomputing Directory Service and used.
(MDS) that provides a uniform interface to diverse infor-
mation sources. We show how a simple data representa- `2.1` **Requirements**
tion and application programming interface (API) based
on the Lightweight Directory Access Protocol (LDAP) Following are the requirements that shaped our design
meet requirements for uniformity, extensibility, and dis- of an information infrastructure for distributed computing
tributed maintenance. We introduce a data model suitable applications. Some of these requirements can be expressed
for distributed computing applications and show how this in quantitative terms (e.g., scalability, performance); others
model is able to represent computers and networks of inter- are more subjective (e.g., expressiveness, deployability).
est. We also present novel implementation techniques for
this service that address the unique requirements of high- **Performance.** The applications of interest to us frequently
performance applications. Finally, we use examples from operate on a large scale (e.g., hundreds of proces-
the Globus distributed computing toolkit [9] to show how sors) and have demanding performance requirements.
MDS data can be used to guide configuration decisions with Hence, an information infrastructure must permit rapid
realistic settings. We expect these techniques to be equally access to frequently used configuration information. It
useful in other systems that support computing in distributed is not acceptable to contact a server for every item:
environments, such as Legion [12], NEOS [5], NetSolve [4], caching is required.
Condor [16], Nimrod [I], PRM [18], AppLeS [2], and het-
**Scalability and cost.** The infrastructure must scale to large
erogeneous implementations of MPI [ 131.
numbers of components and permit concurrent access
The principal contributions of this article are
by many entities. At the same time, its organization
**_0_** a new architecture for high-performance distributed must permit easy discovery of information. The human
computing systems, based upon an information service and resource costs (CPU cycles, disk space, network
called the Metacomputing Directory Service; bandwidth) of creating and maintaining information
must also be low, both at individual sites and in total.
**_0_** a design for this directory service, addressing issues of
**Uniformity. Our goal is to simplify the development of**
data representation, data model, and implementation;
tools and applications that use data to guide config-
**_0_** a data model able to represent the network structures uration decisions. We require a uniform data model
commonly used by distributed computing systems, in- `as well as an application programming interface (API)`
cluding various types of supercomputers; and for common operations on the data represented via that
###### 2
-----
model. One aspect of this uniformity is a standard rep- and the Network Information Service (NIS) both permit re-
resentation for data about common resources, such as mote access but are defined within the context of the IP
processors and networks. protocol suite, which can add significant overhead to a high-
performance computing environment. Furthermore, SNMP
**Expressiveness. We require a data model rich enough to**
does not define an API, thus preventing its use as a compo-
represent relevant structure within distributed comput-
nent within other software architectures.
ing systems. **A** particular challenge is representing
High-performance computing systems such as PVM [ 111,
characteristics that span organizations, for example net-
p4 [3], and MPICH [ 131 provide rapid access to configura-
work bandwidth between sites.
tion data by placing this data (e.g., machine names, network
interfaces) into files maintained by the programmer, called
**Extensibility. Any data model that we define will be in-**
“hostfiles.” However, lack of support for remote access
complete. Hence, the ability to incorporate additional
means that hostfiles must be replicated at each host, compli-
information is important. For example, an applica-
cating maintenance and dynamic update.
tion can use this facility to record specific information
The Domain Name Service (DNS) provides a highly dis-
about its behavior (observed bandwidth, memory re-
tributed, scalable service for resolving Internet addresses to
quirements) for use in subsequent runs.
values (e.g., IP addresses) but is not, in general, extensible.
**Multiple information sources. The information that we** Furthermore, its update strategies are designed to support
require may be generated by many different sources. values that change relatively rarely.
Consequently, an information infrastructure must inte- The X.500 standard [ 14, 201 defines a directory service
grate information from multiple sources. that can be used to provide extensible distributed directory
services within a wide area environment. A directory service
**Dynamic data.** Some of the data required by applications
is a service that provides read-optimized access to general
###### is highly dynamic: for example, network availability
data about entities, such as people, corporations, and com-
or load. An information infrastructure must be able to
puters. X.500 provides a framework that could, in principle,
make this data availabIe in a timely fashion.
be used to organize the information that is of interest to us.
However, it is complex and requires IS0 protocols and the
**Flexible access.** We require the ability to both read and up-
date data contained within the information infrastruc- heavyweight ASN. 1 encodings of data. For these and other
ture. Some form of search capability is also required, reasons, it is not widely used.
###### to assist in locating stored data. The Lightweight Directory Access Protocol [24] is a
streamlined version of the X.500 directory service. It re-
**Security.** It is important to control who is allowed to update moves the requirement for an IS0 protocol stack, defining
configuration data. Some sites will also want to control a standard wire protocol based on the IP protocol suite. It
access. also simplifies the data encoding and command set of X.500
and defines a standard API for directory access [ 151. LDAP
**Deployability. An information infrastructure is useful only**
is seeing wide-scale deployment as the directory service of
if is broadly deployed. In the current case, we require
choice for the World Wide Web. Disadvantages include its
techniques that can be installed and maintained easily
only moderate performance (see Section 5), limited access
at many sites.
to external data sources, and rigid approach to distributing
**Decentralized maintenance.** It must be possible to dele- data across servers.
gate the task of creating and maintaining information Reviewing these various systems, we see that each is
about resources to the sites at which resources are lo- in some way incomplete, failing to address the types of
cated. This delegation is important for both scalability information needed to build high-performance distributed
and security reasons. computing systems, being too slow, or not defining an API
to enable uniform access to the service. For these reasons,
###### 2.2 Approaches we have defined our own metacomputing information in-
frastructure that integrates existing systems while providing
a uniform and extensible data model, support for multiple
It is instructive to review, with respect to these require-
information service providers, and a uniform API.
ments, the various (incomplete) approaches to information
infrastructure that have been used by distributed computing
systems. **2.3** **A Metacomputing Directory Service**
Operating system commands such `as` **mame** and
sysinf o can provide important information about a partic- Our analysis of requirements and existing systems leads
ular machine but do not support remote access. SNMP [21] us to define what we call the Metacomputing Directory Ser-
_3_
-----
vice (MDS). This system consists of three distinct compo- entry represents. This type information, which is encoded
nents: within the MDS data model, is encoded in MDS by associ-
ating an object class with each entry. We now describe how
###### 1. Representation and data access: The directory struc- entries are named and then, how attributes are associated
ture, data representation, and API defined by LDAP. with objects.
###### 2. Data model: A data model that is able to encode
3.1 Naming MDS Entries
the types of resources found in high-performance dis-
tributed computing systems.
Each MDS entry is identified by a unique name, called its
**3.** **Implementation: A set of implementation strategies** **_distinguished name. To simplify the process of locating an_**
designed to meet requirements for performance, mul- MDS entry, entries are organized to form a hierarchical, tree-
tiple data sources, and scalability. structured name space called a directory information tree
(DIT). The distinguished name for an entry is constructed
We provide more details on each of these components in the by specifying the entries on the path from the DIT root to
following sections. the entry being named.
Figure 1 illustrates the structure of MDS and its role in a Each component of the path that forms the distinguished
high-performance distributed computing system. An appli- name must identify a specific DIT entry. To enable this, we
cation running in a distributed computing environment can require that, for any DIT entry, the children of that entry
access information about system structure and state through must have at least one attribute, specified a priori, whose
a uniform API. This information is obtained through the value distinguishes it from its siblings. (The X.500 repre-
MDS client library, which may access a variety of services sentation actually allows more than one attribute to be used
and data sources when servicing a query. to disambiguate names.) Any entry can then be uniquely
named by the list of attribute names and values that identify
###### 3 Representation its ancestors up to the root of the DIT. For example, consider
the following MDS distinguished name:
The MDS design adopts the data representations and API < `hn =` `dark.mcs.anl.gov,`
defined by the LDAP directory service. This choice is driven `ou =` MCS,
by several considerations. Not only is the LDAP data rep- o = Argonne National Laboratory,
resentation extensible and flexible, but LDAP is beginning o = Globus,
to play a significant role in Web-based systems. Hence, we c = u s >
can expect wide deployment of LDAP information services,
familiarity with LDAP data formats and programming, and The components of the distinguished name are listed in little
the existence of LDAP directories with useful information. **_endian order, with the component corresponding to the root_**
Note that the use of LDAP representations and API does not of the DIT listed last. Within a distinguished name, abbrevi-
constrain us to use standard LDAP implementations. As we ated attribute names are typically used. Thus, in this exam-
explain in Section 5, the requirements of high-performance ple, the names of the distinguishingattributes are: host name
distributed computing applications require alternative im- (HN), organizational unit (Ow, organization (0), and coun-
plementation techniques. However, LDAP provides an at- try (C). Thus, a country entry is at the root of the DIT, while
tractive interface on which we can base our implementation. host entries are located beneath the organizational unit level
LDAP also provides a mechanism to restrict the types of of the DIT (see Figure 2). In addition to the conventional set
operations that can be performed on data, which helps to of country and organizational entries (US, ANL, USC, etc.),
address our security requirements. we incorporate an entry for a pseudo-organization named
In the rest of this section, we talk about the “MDS repre- “Globus,” so that the distinguished names that we define do
sentation,” although this representation comes directly from not clash with those defined for other purposes.
LDAP (which in turn “borrows” its representation from
X.500). In this representation, related information is or- `3.2` **Object Classes**
ganized into well-defined collections, called entries. MDS
contains many entries, each representing an instance of some Each DIT entry has a user-defined type, called its object
type of object, such as an organization, person, network, or **_class._** (LDAP defines a set of standard object class defi-
computer. Information about an entry is represented by one nitions, which can be extended for a particular site.) The
or more attributes, each consisting of a name and a cor- object class of an entry defines which attributes are associ-
responding value. The attributes that are associated with ated with that entry and what type of values those attributes
a particular entry are determined by the type of object the may contain. For example, Figure 3 shows the definition
4
-----
###### Figure 1. Overview of the architecture of the Metacomputing Directory Service
of the object classes `GlobusHost and` `GlobusResource,` **4** **DataModel**
and Figure 4 shows the values associated with a particular
host. The object class definition consists of three parts: a To use the MDS representation for a particular purpose,
parent class, a list of required attributes, and a list of optional we must define a data model in which information of interest
attributes. can be maintained. This data model must specify both a DIT
The SUBCLASS section of the object class definition en- hierarchy and the object classes used to define each type of
ables a simple inheritance mechanism, allowing an object entry.
class to be defined in terms of an extension of an existing In its upper levels, the DIT used by MDS (see Figure 2)
object class. The MUST `CONTAII and MAY CONTAIN sec-` is typical for LDAP directory structures, looking similar to
tions specify the required and optional attributes found in the organization used for multinational corporations. The
an entry of this object class. Following each attribute name root node is of object class country, under which we place
is the type of the attribute value. While the set of attribute first the _organization entry representing Globus and then_
types is extensible, a core set has been defined, including the _organization and_ _organizational unit @e., division or_
case-insensitive strings (cis) and distinguished names (dn). department) entries. Entries representing people and com-
puters are placed under the appropriate organizational units.
In Figure 3, GlobusHost inherits from the object class
The representation of computers and networks is central
```
GlobusResource. This means that a GlobusHost entry
```
to the effective use of MDS, and so we focus on this issue
(i.e., an entry of type GlobusHost) contains all of the at-
in this section.
tributes required by the `GlobusResource class, as well as`
the attributes defined within its own MUST `CONTAIN sec-`
###### 4.1 Representing Networks and Computers
tion. `In Figure` 4, the administrator attribute is inherited
from `GlobusResource. A GlobusHost entry may also`
We adopt the framework for representing networks intro-
optionally contain the attributes from both its parent’s and
its own MAY CONTAIN section. duced in RFC 1609 [17] as the starting point for the repre-
sentation used in MDS. However, the RFC 1609 framework
Notice that the administrator attribute in Figure 4 con-
provides a network-centric view in which computers are ac-
tains a distinguished name. This distinguished name acts as
cessible only via the networks to which they are connected.
a pointer, linking the host entry to the person entry represent-
We require a representation of networks and computers that
ing the administrator. One must be careful not to confuse
allows us to answer questions such as
this link, which is part of an entry, with the relationshipsrep-
resented by the DIT, which are not entry attributes. The DIT **_0_** Are computers A and B on the same network?
should be thought of as a separate structure used to organize
**_0_** What is the latency between computers C and D?
an arbitrary collection of entries and, in particular, to enable
the distribution of these entries over multiple physical sites.
**_0_** What protocols are available between computers E and
Using distinguished names as attribute values enables one to
F?
construct more complex relationships than the trees found in
the DIT. The ability to define more complex structures is es- In answering these questions, we often require access to
sential for our purposes, since many distributed computing information about networks, but questions are posed most
structures are most naturally represented as graphs. often from the perspective of the computational resource.
_5_
-----
**_nn=WAN_** **_O'USC_**
- **_OU=MCS_**
**_IS1 at USC_** [. ]\
**_(California)_** '--
###### Figure 2. A subset of the DIT defined by MDS, showing the organizational nodes for Globus, ANL, and USC; the organizational units IS1 and MCS; and a number of people, hosts, and networks.
That is, they are computer-centric questions. Our datamodel see, logical information, such as the network protocol being
reflects this perspective. used, is not specified in the GlobusNetwork object but is
A high-level view of the DIT structure used in MDS is associated with a GlobusNetworkImage object. Networks
shown in Figure 2. **As indicated in this figure, both people** that span organizations can be represented by placing the
and hosts are immediate children of the organizations in `GlobusNetwork object higher in the DIT.`
which they are located. For example, the distinguished Networks and hosts are related to one another via
name `GlobusNetworkInt erf ace objects: hosts contain network`
interfaces, and network interfaces are attached to networks.
< `hn=dark.mcs.anl.gov,`
A network interface object represents the physical charac-
```
ou=MCS, o=Argonne National Laboratory,
```
`o=Globus, c=US >` teristics of a network interface (such as interface speed) and
the hardware network address (e.g. the 48-bit Ethernet ad-
dress in the case of Ethernet). Network interfaces appear
identifies a computer administered by the Mathematics and
under hosts in the DIT, while a network interface is `as-`
Computer Science (MCS) Division at Argonne National
sociated with a network via an attribute whose value is a
Laboratory.
distinguished name pointing to a GlobusNetwork object.
Communication networks are also explicitly represented
```
A reverse link exists from the GlobusNetwork object back
```
in the DIT as children of an organization. For example, the
to the interface.
distinguished name
###### To illustrate the relationship between GlobusHost,
< `nn=mcs-lan,` `GlobusNetwork, and GlobusNetworkInt erf aceobjects,`
**ou=MCS,** `o=Argonne National Laboratory,` we consider the configuration shown in Figure 5. This con-
`o=Globus, c=US >` figuration consists of an IBM SP parallel computer and two
workstations, all associated with MCS. The SP has two net-
represents the local area network managed by MCS. works: an internal high-speed switch and an Ethernet; the
This distinguished name identifies an instance of a workstations are connected only to an Ethernet. Although
`GlobusNetwork object.` The attribute values of a the SP Ethernet and the workstation Ethernet are connected
`GlobusNetwork object provides information about the` via a router, we choose to represent them `as a single net-`
_physical network link. such as the link protocol (e.g., ATM_ work. An alternative, higher-fidelity MDS representation
or Ethernet), network topology (e.g., bus or ring type), and would capture the fact that there are two interconnected
physical media (e.g., copper or fiber). As we shall soon Ethernet networks.
**_6_**
-----
```
GlobusHost OBJECT CLASS GlobusResource OBJECT CLASS
SUBCLASS OF GlobusResource SUBCLASS OF GlobusTop
MUST CONTAIN { MUST CONTAIN c
```
`hostflame` :: `cis,` `administrator :: dn`
`type` : : `cis,` 3
`vendor` :: `cis,` `MAY CONTAIE <`
`model` ; : `cis,` `manager` :: dn,
`OStype` .. . . `cis,` `provider` :: dn,
`OSversion` :: cis `technician` :: dn,
###### 3 description :: cis,
```
MAY CONTAIN < documentation :: cis
```
`net workNode` :: dn, 3
`tot alMemory` : : cis,
`tot alSwap` : : cis,
`dat aCache` : : cis,
`instructioncache` : : cis
###### 3
Figure 3. Simplified versions of the MDS object classes GlobusHost and GlobusResource
```
dn: hn=dark.mcs.anl.gov, ou=MCS,
o=Argonne National Laboratory , o=Globus , c=US
objectclass: GlobusHost
objectclass: GlobusResource
administrator: cn=John Smith, ou=MCS,
o=Argonne National Laboratory, o=Globus, c=US
hostName: dark.mcs.anl.gov
type : sparc
```
`vendor :` **SUI**
```
model ; SPARCstation-IO
```
`OStype :` sunos
```
OSversion: 5.5.1
###### Figure 4. Sample data representation for an MDS computer
```
The MDS representation for Figure 5 is shown in Fig- each node has at least two network interfaces: one to the
ure 6. Each host and network in the configuration appear high-speed switch and one to an Ethernet. Finally, we see
in the DIT directly under the entry representing MCS at Ar- that distinguished names are used to complete the repre-
gonne National Laboratory. Note that individual SP nodes sentation, linking the network interface and network object
are children of MCS. This somewhat unexpected represen- together.
tation is a consequence of the SP architecture: each node
is a fully featured workstation, potentially allowing login. **4.2** **Logical Views and Images**
Thus, the MDS representation captures the dual nature of
the SP as a parallel computer (via the switch network object)
At this point, we have described the representation of a
and as a collection of workstations.
physical network: essentially link-level aspects of the net-
As discussed above, the GlobusNetuorkInt erf ace ob- work and characteristics of network interface cards and the
jects are located in the DIT under the GlobusHost objects. hosts they plug into. However, a physical networkmay sup-
Note that a GlobusHost can have more than one network port several “logical” views, and we may need to associate
interface entry below it. Each entry corresponds to a dif- additional information with these logical views. For exam-
ferent physical network connection. In the case of an SP, ple, asingle network might be accessible viaseveral different
_7_
-----
###### Figure 6. The MDS representation of the configuration depicted in Figure 5, showing host (HN), network (NN), and network interface (NIN) objects. The dashed lines correspond to “pointers” represented by distinguished name attributes
cept of **_images_** `as a mechanism for representing multiple`
logical views of the same physical network. We apply
the same concept in our data model. Where physical net-
works are represented by `GlobusHost, Globusletwork,`
and `GlobusNetworkInterf ace object classes, network`
images are repre-
sented by GlobusHost Image, GlobusNetnorkImage, and
`GlobusNetworkInterf aceImage object classes.` Each
image object class contains new information associated
with the logical view, `as` well `as a distinguished name`
pointing to its relevant physical object. In addition, a
physical object has distinguished name pointers to all of
the images that refer to it. For example, one may use
both IP and IPX protocols over a single Ethernet interface
card. We would represent this in **MDS** by creating two
###### Figure 5. A configuration comprising two net- GlobusNetworkInterf aceImage objects. One image ob- works and N+2 computers ject would represent the IP network and contain the IP ad-
dress of the interface, as well as a pointer back to the object
class representing the Ethernet card. The second image ob-
ject would contain the IPX address, as well as adistinguished
protocol stacks: IP, Novel1 IPX, or vendor-provided libraries name pointing back to the same entry for the Ethernet card.
such as MPI. Associated with each of these protocols can The GlobusNetworkInterf ace object would include the
be distinct network interface and performance information. distinguished names of both interface images.
Additionally, a “partition” might be created containing a
The structure of network images parallels that of the cor-
subset of available computers; scheduling information can
responding physical networks, with the exception that not
be associated with this object.
all network interfaces attached to a host need appear in an
The RFC 1609 framework introduces the valuable con- image. To see why, consider the case of the IBM SP. One
**8**
-----
might construct a network image to represent the “parallel turn our attention to the MDS implementation. Since our
computer” view of the machine in which IBM’s proprietary data model has been defined completely within the LDAP
message-passing library is used for communication. Since framework, we could in principle adopt the standard LDAP
this protocol cannot be used over the Ethernet, this image of implementation. This implementation uses a TCP-based
the network will not contain images representing the Ether- wire protocol and a distributed collection of servers, where
net card. Note that we can also produce a network image each server is responsible for all the entries located within a
of the SP representing the use of IP protocols. This view complete subtree of the DIT. While this approach is suitable
may include images of both the switch and Ethernet network for a loosely coupled, distributed environment, it has three
interfaces. significant drawbacks in a high-performance environment:
###### 4.3 Questions Revisited 0 Single information provider. The LDAP implemen-
tation assumes that all information within a DIT subtree
At this stage we have gone quite deeply into the repre- is provided by a single information provider. (While
sentation of computers and networks but have strayed rather some LDAP servers allow alternative “backend” mech-
far from the issue that motivated the MDS design, namely, anisms for storing entries, the same backend must be
the configuration of high-performance distributed compu- used for all entries in the DIT subtree.) However, re-
tations. To see how MDS information can be used, let us stricting all attributes to the same information provider
revisit the questions posed in Section **1 with respect to the** complicates the design of the MDS data-model. For
use of multiple computers on the I-WAY example, the IP address associated with a network in-
terface image can be provided by a system call, while
**_What are the nenvork intelfaces (i.e., IP addresses)_** the network bandwidth available through that interface
**_for the ATM network and Internet?_** A host’s IP ad- is provided by a service such as the Network Weather
dress on the ATM network can be found by look- Service ( N W S ) [23].
ing for a `GlobusBetaorkInterface that is point-`
ing to a `GlobusBetwork with a link protocol at-` **0** **Clienffserver architecture.** The LDAP implementa-
tribute value of ATM. From the interface, we find the tion requires at least one round-trip network commu-
`GlobusNetworkInterf aceImage representing an IP` nication for each LDAP access. Frequent MDS ac-
network, and the IP address will be stored as an attribute cesses thus becomes prohibitively expensive. We need
in this object. a mechanism by which MDS data can be cached locally
for a timely response.
**_What is the raw bandwidth of the ATM network and_**
**_the Internet, and which is higher? Is the ATM network_** **0** **Scope of Data. The LDAP implementation assumes**
**_currently available?_** The raw bandwidth of the ATM that any piece of information may be used from any
network will be stored in the I-WAY GlobusNetwork point in the network (within the constraints of access
object. Information about the availability of the ATM control). However, a more efficient implementation of
network can also be maintained in this object. attribute update can be obtained if one can limit the
locations from which attribute values can be accessed.
**_Between which pairs of_** **_nodes can we use vendorproto-_**
The introduction of scope helps to determine which
**_cols to access fast internal networks? Between which_**
information must be propagated to which information
**_pairs of_** **_nodes must we use_** `TCP/IP? Two nodes can` providers, and when information can be safely cached.
communicate using a vendor protocol if they both point
to GlobusHostImage objects that belong to the same Note that these drawbacks all relate to the LDAP im-
`GlobusNetworkImage object.` plementation, not its API. Indeed, we can adopt the LDAP
API for MDS without modification. Furthermore, for those
Note that the definition of the MDS representation, API, and
DIT subtrees that contain information that is not adversely
data model means that this information can be obtained via
affected by the above limitations, we can pass the API calls
a single mechanism, regardless of the computers on which
straight through to an existing LDAP implementation. In
an application actually runs.
general, however, MDS needs a specialized implementation
of the LDAP API to meet the requirements for high perfor-
###### 5 Implementation mance and multiple information providers.
The most basic difference between our MDS implemen-
We have discussed how information is represented in tation and standard LDAP implementations is that we allow
MDS, and we have shown how this information can be used information providers to be specified on a per attribute ba-
to answer questions about system configuration. We now sis. Referring to the above example, we can provide the IP
**9**
-----
address of an interface via SNMP, the current available band- first uses MDS information to determine which low-level
width viaNWS, and the name of the machine into which the mechanisms are available between the processors. Then, it
interface card is connected. Additionally, these providers selects from among these mechanisms, currently on the basis
can store information into MDS on a periodic basis, thus of built-in rules (e.g., “ATM is better than Internet”); rules
allowing refreshing of dynamic information. The specifi- based on dynamic information (“use **ATM** if current load
cation of which protocol to use for each entry attribute is is low”), or programmer-specified preferences (“always use
stored in an object class metadata entry. Metadata entries Internet because I believe it is more reliable”) can **also be**
are stored in MDS and accessed via the LDAP protocol, supported in principle. The result is that application source
In addition to specifying the access protocol for an at- code can run unchanged in many different environments,
tribute, the MDS object class metadata also contains a time- selecting appropriate mechanisms in each case.
to-live (TIL) for attribute values and the update scope of the These method-selection mechanisms were used in the
attribute. The “L data is used to enable caching; a TIL of I-WAY testbed to permit applications to run on diverse het-
###### 0 indicates that the attribute value cannot be cached, while a erogeneous virtual machines. For example, on a virtual ?TL of - 1 indicates that the data is constant. Positive TIL machine connecting IBM SP and SGI Challenge comput-
values determine the amount of time that the attribute value ers with both ATM and Internet networks, Nexus used three
is allowed to be provided out of the cache before refreshing. different protocols (IBM proprietary MPL on the SP, shared-
The update scope of an attribute limits the readers of an memory on the Challenge, and TCP/IP or AAL5 between
updated attribute value. Our initial implementation consid- computers) and selected either ATM or Internet network
ers three update scopes: process, computation, and global. interfaces, depending on network status [8].
Process scope attributes are accessible only within the same Another application for MDS information that we are
process as the writer, whereas computation scope attributes investigating is resource location _[22]._ **A** “resource bro-
can be accessed by any process within a single computation, ker” is basically a process that supports specialized searches
and global scope attributes can be accessed from any node against MDS information. Rather than incorporate these
or process on a network. search capabilities in MDS servers, we plan to construct
resource brokers that construct and maintain the necessary
###### 6 MDS Applications in Globus indexes, querying MDS periodically to obtain up-to-date
information.
We review briefly some of the ways in which MDS in-
formation can be used in high-performance distributed com- **7 Summary**
puting. We focus on applications within Globus, an infras-
tructure toolkit providing a suite of low-level mechanisms We have argued that the complex, heterogeneous, and
designed to be used to implement a range of higher-level dynamic nature of high-performance distributed computing
services [9]. These mechanisms include communication, systems requires an **_information-rich approach to system_**
authentication, resource location, resource allocation, pro- configuration. In this approach, tools and applications do
cess management, and (in the form of MDS) information not rely on defaults or programmer-supplied knowledge to
infrastructure. make configuration choices. Instead, they base choices on
The Globus toolkit is designed with the configuration information obtained from external sources.
problem in mind. It attempts to provide, for each of its With the goal of enabling information-rich configuration,
components, interfaces that allow higher-level services to we have designed and implemented a Metacomputing Direc-
manage how low-level mechanisms are applied. As an ex- **_tory Service. MDS is designed to provide uniform, efficient,_**
ample, we consider the problem referred to earlier of select- and scalable access to dynamic, distributed, and diverse in-
ing network interfaces and communication protocols when formation about the structure and state of resources. MDS
executing communication code within a heterogeneous net- defines a representation (based on that of LDAP), a data
work. The Globus communication module (a library called model (capable of representing various parallel computers
Nexus [lo]) allows a user to specify an application’s com- and networks), and an implementation (which uses caching
munication operations by using a single notation, regardless and other strategies to meet performance requirements). Ex-
of the target platform: either the Nexus API or some library periments conducted with the Globus toolkit (particularly in
or language layered on top of that API. At run-time, the the context of the I-WAY) show that MDS information can
Nexus implementation configures a communication struc- be used to good effect in practical situations.
ture for the application, selecting for each communication We are currently deploying MDS in our GUSTO dis-
link (a Nexus construct) the communication method that is tributed computing testbed and are extending additional
to be used for communications over that link [7]. In mak- Globus components to use MDS information for configura-
ing this selection for a particular pair of processors, Nexus tion purposes. Other directions for immediate investigation
10
-----
include expanding the set of information sources supported, [9] **I. Foster and C.** Kesselman. Globus: A metacomputing in-
evaluating performance issues in applications, and develop- frastructure toolkit. International Journal of Supercomputer
ing optimized implementations for common operations. In **_Applications, 1997. To appear._**
the longer term, we are interested in more sophisticated ap- [lo] **1. Foster, c. Kesselm, and s. necke- The Nexus approach**
to integrating multithreading and communication. **_Journal_**
plications (e.g., source routing, resource scheduling) and in
**_of Parallel and Distributed Computing, 37:70-82,1996._**
the recording and use Of aPP1ication-generated performance
[ 1 11 A. G&, A. Bepelin, J. Dongma, W. Jiang, B. Manchek,
metrics.
and V. Sunderam. PVM: Parallel Krtml Machine-A User's
**_Guide and Tutorial for Network Parallel Computing._** MlT
###### Acknowledgments Press, 1994.
[I21 A. Chimshaw, J. Weissman, E. West, and E. Lyot, Jr. Meta-
systems: **_An_** approach combining parallel processing and
We gratefully acknowledge the contributions made by heterogeneous distributed computing systems. **_Journal of_**
Craig Lee, Steve Schwab, and Paul Stelling to the design and **_Parallel and Distributed Computing, 21(3):257-270,1994._**
implementation of Globus components. This work was sup- [13] W. Gropp, E. Lusk, N. Doss, and A. Skjellum. **A** high-
performance, portable implementation of the MPI message
ported by the Defense Advanced Research Projects Agency
passing interface standard. Parallel Computing, 22:789-828,
under contract N66001-96-C-8523 and by the Mathemati-
1996.
cal, Information, and Computational Sciences Division sub-
[14] **S. Heker, J. Reynolds, and C. Weider. Technical overview**
program of the Office of Computational and Technology
of directory services using the x.500 protocol. `RFC 1309,`
Research, **U.S. Department of Energy, under Contract W-**
Ey14,03/12 92.
3 1-109-Eng-38. [15] T. Howes and M. Smith. The ldap application program in-
terface. RFC 1823,08/09 95.
[16] M. Litzkow, M. Livney, and M. Mutka. Condor - a hunter
###### References
of idle workstations. In Proc. 8th Intl Conf: on Distributed
**_Computing Systems, pages 104-1 11,1988._**
D. Abramson, `R. Sosic, J. Giddy, and B. Hall. Nimrod:` [ 171 G. Mansfield, T. Johannsen, and M. Knopper. Charting net-
**_A tool for performing parameterised simulations using dis-_** works in the x.500 directory. RFC 1609,03125 94. (Experi-
tributed workstations. In **_Pmc. 4th IEEE Symp._** **_on High_** mental).
**_Performance Distributed Computing. IEEE Computer Soci-_** [ 181 B. C. Neumann and S . Rao. The Prosper0 resource manager:
ety Press, 1995. A scalable framework for processor allocation in distributed
**E Berman, R.** Wolski, S . Figueira, J. Schopf, and G. Shao. systems. **_Concurrency: Practice_** & Experience, 6(4):339-
355,1994.
Application-level scheduling on distributed heterogeneous
[ 191 D. Reed, C. Elford, T. Madhyastha,E. Smimi, and S . Lamm.
networks. In **_Proceedings_** `of Supercomputing` **_'96. ACM_**
The Next Frontier: Interactive and Closed Loop Performance
Press, 1996.
Steering. In **_Proceedings_** **_of_** **_the_** **_1996_** **_ICPP Workshop on_**
[3 I **R.** Butler and E. Lusk. Monitors, message, and clusters:
**_Challengesfor ParallelProcessing,pages20-31,Aug. 1996._**
The p4 parallel programming system. **_Parallel Computing,_**
[20] J. Reynolds and C. Weider. Executive introduction to direc-
20547-564, April 1994.
tory services using the x.500 protocol. RFC 1308, FYI 13,
H. Casanova and J. Dongarra. Netsolve: A network server for
03/12 92.
solving computational science problems. Technical Report
[21] M. Rose. `The Simple Book. Prentice Hall, 1994.`
CS-95-3 13, University of Tennessee, Nov. 1995.
[22] G. von Laszewski. **_A Parallel Data Assimilatim System_**
J. Czyzyk, M. P. Mesnier, and J. J. More. The Network- **_and Its Implications on a Metacomputing Environment. PhD_**
Enabled Optimization System (NEOS) Server. Preprint thesis, Syracuse University, Dec. 1996.
MCS-P6 15-0996, Argonne National Laboratory, Argonne, [23] R. Wolski. Dynamically forecasting network performance
Illinois, 1996. using the network weather service. Technical Report TR-
T. DeFanti, I. Foster, M. Papka, R. Stevens, and T. Kuh- CS96-494, U.C. San Diego, October 1996.
**fuss. Overview of the I-WAY: Wide area visual supercomput-** [24] W. Yeong, T. Howes, and **_S . Kille. Lightweight directory_**
ing. **_International Journal of Supercomputer Applications,_** access protocol. RFC 1777,03/28 95. Draft Standard.
10(2):123-130,1996.
I. Foster, J. Geisler, C. Kesselman, and S . Tuecke. Manag-
ing multiple communication methods in high-performance
networked computing systems. Journal of Parallel and Dis-
**_tributed Computing, 40:35-48, 1997._**
I. Foster, J. Geisler, W. Nickless, W. Smith, and S . Tuecke.
Software infrastructure for the I-WAY high-performance dis-
tributed computing experiment. In Proc. 5th IEEE Symp. on
**_High Performance Distributed Computing, pages 562-57_** 1.
IEEE Computer Society Press, 1996.
11
-----
|
{
"disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references', 'abstract'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1109/HPDC.1997.626445?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/HPDC.1997.626445, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "other-oa",
"status": "GREEN",
"url": "https://digital.library.unt.edu/ark:/67531/metadc691395/m2/1/high_res_d/508115.pdf"
}
| 1,997
|
[
"JournalArticle"
] | true
| 1997-08-05T00:00:00
|
[] | 11,504
|
en
|
[
{
"category": "Business",
"source": "s2-fos-model"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/02920271ab40ff82676325adcb340f8a60a63eb2
|
[] | 0.822347
|
Improving Operational Efficiency through Quality 4.0 Tool: Blockchain Implementation and Subsequent Market Reaction
|
02920271ab40ff82676325adcb340f8a60a63eb2
|
Kvalita Inovácia Prosperita
|
[
{
"authorId": "2180117178",
"name": "Vladimíra Gimerská"
},
{
"authorId": "37871549",
"name": "M. Šoltés"
},
{
"authorId": "119015195",
"name": "Rajmund Mirdala"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"Quality, Innovation, Prosperity",
"Qual Innov Prosper",
"Kval Inovácia Prosper"
],
"alternate_urls": [
"https://www.qip-journal.eu/index.php/QIP/about"
],
"id": "85d3209d-e1d9-4246-a5fa-843669e87d00",
"issn": "1335-1745",
"name": "Kvalita Inovácia Prosperita",
"type": "journal",
"url": "http://www.qip-journal.eu/"
}
|
Purpose: This article aims to observe and measure how modern and innovative blockchain technology improves the data quality and transparency and thus affect the stock prices of publicly traded companies after announcing its implementation in their operations. Additionally, the objective is to compare the results with control group of non-adopters.
Methodology/Approach: We selected 30 public companies across various sectors, obtained daily stock price data, identified peer companies, and employed an event study approach to examine the statistical impact of blockchain adoption announcements.
Findings: A significant negative reaction (-0.4%) was observed in stock prices the day following a blockchain adoption announcement, but overall, the market response was unsystematic, indicating no consistent reaction in stock prices post-announcement.
Research Limitation/Implication: The event study approach assumes that markets are always efficient. This methodology has some limitations because we live in a world that is not perfect, and stock prices do not necessarily fully reflect all available information.
Originality/Value of paper: Blockchain implementation is a current and intriguing subject that has attracted limited scholarly research. Each new study contributes valuable insights to the understanding of how this innovative technology impacts corporate operations. Furthermore, this research endeavours to draw comparisons between companies that have announced their adoption of blockchain and their non-adopters counterparts.
|
# Improving Operational Efficiency through Quality 4.0 Tool: Blockchain Implementation and Subsequent Market Reaction
## DOI: 10.12776/QIP.V27I2.1877
Vladimíra Gimerská, Michal Šoltés, Rajmund Mirdala
Received: 2023-06-21 Accepted: 2023-06-28 Published: 2023-07-31
## ABSTRACT
**Purpose: This article aims to observe and measure how modern and innovative**
blockchain technology improves the data quality and transparency and thus affect
the stock prices of publicly traded companies after announcing its
implementation in their operations. Additionally, the objective is to compare the
results with control group of non-adopters.
**Methodology/Approach: We selected 30 public companies across various**
sectors, obtained daily stock price data, identified peer companies, and employed
an event study approach to examine the statistical impact of blockchain adoption
announcements.
**Findings: A significant negative reaction (-0.4%) was observed in stock prices**
the day following a blockchain adoption announcement, but overall, the market
response was unsystematic, indicating no consistent reaction in stock prices postannouncement.
**Research Limitation/Implication: The event study approach assumes that**
markets are always efficient. This methodology has some limitations because we
live in a world that is not perfect, and stock prices do not necessarily fully reflect
all available information.
**Originality/Value of paper: Blockchain implementation is a current and**
intriguing subject that has attracted limited scholarly research. Each new study
contributes valuable insights to the understanding of how this innovative
technology impacts corporate operations. Furthermore, this research endeavours
to draw comparisons between companies that have announced their adoption of
blockchain and their non-adopters counterparts.
**Category: Conceptual paper**
**Keywords: quality 4.0, blockchain; event studies; digitalisation**
ISSN 1335 1745 ( i t) ISSN 1338 984X ( li )
-----
## 1 INTRODUCTION
According to Global Data Management Research organisations need to improve
their data quality. Report shows that failing to improve the data can cause
increased costs, unreliable analytics, negative impact on customer trust,
experience and company reputation which lead to slow digital transformation.
(Reno, 2022)
The American Society of Quality defines Quality 4.0 as the term which
references “the future of quality and organisational excelence within the context
of Industry 4.0” (American Society for Quality, n.d.). It is confirmed that quality
management and Industry 4.0 directly influence performance (Nguyen et al.,
2021). Technologies 4.0 such as Internet of Things, Artificial Inteligence or
Blockchain are utilised to improve quality of products and services for the
customer and at the same time increase value for shareholders. It is
unquestionable that using Technologies 4.0 as part of Quality 4.0 “provides
numerous benefits to quality management, including increased speed and
transparency, increased adaptability to new situations and continual improvement
across businesses plus increased awareness, skills and inteligence” (Mtotywa,
2022). It also enables early error detection and reduces downtime through
anticipatory maintenance planning (Mtotywa, 2022).
Blockchain technology as one of the Quality 4.0 tools has substantially advanced
since its inception, and companies across multiple industries have widely adopted
it. While most of the attention surrounding blockchain relates to its use in
cryptocurrency, recent literature and applications show its vast potential for
various applications in many industries, especially within the finance sector and
the supply chain. It is an innovative technology that brings significant
optimisation and automatisation when implemented in the company’s various
operations. Blockchain as a quality toll can help company to perform better as it
helps gaining operational excellence, and as a result, foster process innovation.
“Moreover, new forms of collaboration and traceability, such as, block chain, are
very important in this period, especially when factors affecting competitiveness
can vary” (Santos et al., 2021). On the other hand, its adoption is complex and
expensive, so exploring existing use cases is important for companies to help
them in their strategic decision-making process whether to invest in this
technology or not.
This paper focuses on observing and measuring how this Quality 4.0 tool affects
the stock prices of publicly traded companies that announced its implementation
in their operations. We conducted an event study analysis on 30 selected publicly
traded companies from various areas and sectors which announced blockchain
adoption and how this announcement as an event impacted the price
development. We use SPSS software and market model to test the abnormal
returns and their significance on 41 days, 20 days prior and 20 days after the
announcement. Additionally, through the platform Infront Analytics, we searched
for peer companies for each analysed firm from our sample to compare the
ISSN 1335 1745 ( i t) ISSN 1338 984X ( li )
-----
development during the event window. The objective is to determine whether
and to what extent the market reacts to such announcements about blockchain
implementations.
## 2 LITERATURE REVIEW
It is worth analysing blockchain as a technology in the context of consequent
market reactions after a new technology is announced. Such new technological
changes could be e-commerce platforms (Subramani and Walden, 2001; Dehning
et al., 2004), mobile apps (Boyd, Kannan and Slotegraaf, 2019) or ERP systems
(Hendricks, Singhal and Stratman, 2006; Ranganathan and Brown, 2006).
A study conducted by Chen et al. (2022) shares similarities with our objectives
but focuses exclusively on China and Chinese businesses. The researchers
examined two categories of firms – those in high-tech industries and those
outside- intending to embrace blockchain technology in the future. In total, 302
companies listed on the Shanghai and Shenzhen Stock Exchanges between 2016
and 2020 were chosen. The analysis was conducted over 41 and 11 trading days
over two timeframes. The findings revealed that high-tech firms’ blockchain
announcements gained greater interest from investors, eliciting more significant
stock price reactions as investors deemed these companies more trustworthy
(Chen et al., 2022).
There is evidence that blockchain can potentially reduce costs. In the airspace
industry, companies like Honeywell, Moog and Air New Zealand reported up to
30% savings by using blockchain to create secure digital marketplaces for 3Dprinted aircraft parts (Tampi, 2020). In the IT sector, a positive relationship
between technological initiatives and financial performance was observed (e.g.,
Bose and Man Leung, 2019; Bradley et al., 2018), where the emphasis was also
placed on operational efficiency improvements, revenue generation and firms
value (Bose and Man Leung, 2019; Melville, Kraemer and Gurbaxani, 2004).
Additionally, blockchain has the potential to promote innovation in business
models leading to cost reduction and providing new sources of revenue (Lacity,
2018).
Although studies on blockchain application announcements exist, companies’
returns are often compared with Bitcoin returns (Cheng et al., 2019; Cahill et al.,
2020). Only some consider the market value that can be created by implementing
blockchain. In such cases, an event study methodology is usually used to assess
the short-term value investors assign to recently revealed IT initiatives based on
future cash flow anticipation (Boyd, Kannan and Slotegraaf, 2019).
The closest study to ours was published by Klockner, Schmidt and Wagner
(2022), where 175 blockchain announcements from 100 companies were
analysed. The study was well diversified in 11 industries and 15 countries, and
data were additionally tested for robustness. Here, a positive market reaction was
identified for announcements in the context of operations and supply chain
ISSN 1335 1745 ( i t) ISSN 1338 984X ( li )
-----
management. Furthermore, this sample confirmed a significant average abnormal
return of 0.30% on the announcement day. However, when an external IT
provider is used to implement blockchain, a significantly less positive reaction is
observed. Klockner’s research (2022) also provides a comprehensive summary of
recent research which involves blockchain and its influence on cost-efficient
processes. The researches include the following use cases: effect on supply chain
and traceability, enhancement of data and knowledge sharing between supply
chain participants, security and acceleration of inter-organisational payments and
order processing (Klockner, Schmidt and Wagner, 2022)
An investigation of blockchain-related announcements was carried out by Cahill
and colleagues in 2020 on a sample of 713 companies in year between 2016 and
2018 that explored the relationship between Bitcoin development and blockchain
announcement. An average abnormal return of 5.3% was observed on
announcement days, and smaller companies experienced greater abnormal
returns than larger ones. Furthermore, lower returns occurred by non-speculative
announcements than by speculative ones (Cahill et al., 2020).
Cheng et al. (2019) also explored the connection between 79 publicly traded
companies’ initial 8-K filings on blockchain activities and investors’ reactions.
They classified the activities detailed in these disclosures as either existing or
speculative (“existing” were firms with a well-defined strategy for blockchain
implementation, and “speculative” firms outlining ambiguous plans for
blockchain). Their research showed that speculative information had 7.5%
positive abnormal returns while existing disclosures experienced almost zero
abnormal returns. These favourable responses are undone within a month,
suggesting investor overreaction to speculative disclosures (Cheng et al., 2019).
Another event study looks at financial corporations that use blockchain and how
their stocks performed during the COVID-19 pandemic. The common parameter
is that high-tech companies, whether they are members of blockchain
consortiums or have some technological advantage, have better positive stock
development results, avoiding potential losses during pandemic-related
announcements (Paul, Adhikari and Bose, 2022).
Liu et al. (2022) examined market reactions to blockchain announcements,
focusing on a company with 143 announcements. The researchers employed
event study methodology and multivariate regression to analyse market responses
and determine factors affecting these changes. They found a positive market
reaction on announcement days and noted that strategic-level announcements
elicited a stronger positive response from the market (Liu et al., 2022).
## 3 METHODOLOGY AND DATA
The event study methodology is gaining popularity in business and marketing
disciplines to measure the impact of significant events at the firm. This technique
can be used to assess the effect of some important event or corporate
ISSN 1335 1745 ( i t) ISSN 1338 984X ( li )
-----
announcement on a company’s financial performance, profitability, and market
valuation over a defined event window, ranging from a few days to a few years.
The methodology is flexible and can be adapted to measure different events,
making it useful for researchers in various fields (Ullah et al., 2021). Within our
study, we aim to answer following research question (RQ) and thus we create the
null hypothesis (H0):
RQ: Is there any reaction in stock prices after the company officially announces
the application of blockchain technology in its operations?
H0: There is no reaction in stock prices after the company’s announcement
regarding blockchain implementation.
Within the null hypothesis, we will test abnormal returns of companies that
announced blockchain and compare them to the peer group of similar companies
that have not announced any blockchain application in the time around event
window. The null hypothesis will be confirmed when abnormal returns are equal
to zero, and we will reject the null hypothesis when abnormal returns are not
equal to zero. We will also analyse whether the announcement of blockchain’s
adoption had a positive or negative impact on the stock price.
In order to test the hypothesis, we first gathered two main types of data,
announcements of blockchain, which are publicly available, and stock prices.
Then, based on Infront Analytics (2023), we created a control group of similar
companies that had not publicly communicated any blockchain adoption in that
time period. When the same company from the analysed group appeared as a
peer to some other company (mostly in the case of industry car producers), we
took the second or the third listed international company as peer (Infront
Analytics, 2023).
We chose thirty globally active corporations from various industries and obtained
daily stock close prices for the last ten years from the Yahoo.com platform. In
addition, we chose the MSCI World Index to compare prices with general market
performance. Because certain companies and indices representing benchmarks
are traded in different countries, the problem of non-trading days arose, a
common issue in event studies. To solve this, we follow the methodology
mentioned by Campbell, Cowan and Salotti (2010), which completely omits nontrading days from the analysis.
Simultaneously during the phase of choosing the companies for our analysis, we
searched for specific announcements regarding real blockchain implementation
projects. We did not consider any press releases about exploring the technology,
only the real adoption of blockchain in the company’s operations. These
announcements were set in our event study approach as event days (t0). In almost
all cases, the t0 was between 2016 and 2020 except for a few early adopters who
have worked on adoption since 2015, for instance, IBM and Microsoft. If the
announcement was made during a non-trading day, as the event day (t0) we took
the first following trading day.
ISSN 1335 1745 ( i t) ISSN 1338 984X ( li )
-----
Tab. 1 summarises our sample and corresponding announcement days and
sources. Selected corporations come from the Automotive, Finance, Food &
Beverages, Supply Chain and IT sector. For each of the selected companies, we
found peer company and this group was also tested within our study (furthermore
called as “blockchain group” and “control group”).
_Table 1 – Companies which Announced Blockchain Adoption and Their Peers_
**Company** **Official** **Peer company** **Sources of**
**Announcement** **(Infront Analytics, 2023)** **blockchain**
**announcement**
1 Walmart 19 Oct 2016 Pan Pacific Int. Holding Coindesk.com,
Accenture,
2 Anheuser-Busch 14 March 2018 Boston Beer Company LedgerInsights.com,
SAP.com, group
3 Allianz 07 Nov 2017 Unipol
media.mercedesbenz.com,
4 AT&T 26 Sept 2018 Verizon Communications
Volkswagen.com,
5 SAP 16 May 2017 Oracle Corp Porsche.com,
Microsoft.com,
6 Mercedes -Benz 28 June 2017 Stellantis Hyperledger.com,
Reuters.com,
7 Volkswagen 22 Apr 2019 Kia
carrefour.com,
Computerworld.com,
8 BMW 13 Feb 2019 Honda
prnewswire.com,
9 Porsche 22 Feb 2018 Renault bnnbloomberg.ca,
Yahoo.com
10 Microsoft 09 Nov 2015 Adobe
11 IBM 17 Dec 2015 HP
12 Foxconn 06 March 2017 Pegatron
13 Nestle 22 Aug 2017 Danone
14 Carrefour 06 March 2018 Tesco PLC
15 MasterCard 21 Oct 2016 Visa
16 Honeywell International 17 Dec 2018 General Electric Company
17 JPMorgan Chase & Co. 03 March 2016 Bank of America
18 Tyson Foods 22 Aug 2017 Hormel Foods
19 Wells Fargo 24 Oct 2016 Regions Financial Corp.
20 Coca-Cola 05 Nov 2019 Keurig Dr. Pepper
21 FedEx 14 May 2018 Deutsche Post
22 Cisco Systems 11 July 2017 Ciena Corp
23 HSBC Bank 3 Oct 2017 Credit Agricole
24 Deutsche Bank 16 Sept 2019 Commerzbank
25 UBS Bank 11 Dec 2017 BNP Paribas
26 Maersk 16 Jan 2018 Hapag Lloyd
ISSN 1335 1745 ( i t) ISSN 1338 984X ( li )
|Col1|Company|Official Announcement|Peer company (Infront Analytics, 2023)|Sources of blockchain announcement|
|---|---|---|---|---|
|1|Walmart|19 Oct 2016|Pan Pacific Int. Holding|Coindesk.com, Accenture, LedgerInsights.com, SAP.com, group- media.mercedes- benz.com, Volkswagen.com, Porsche.com, Microsoft.com, Hyperledger.com, Reuters.com, carrefour.com, Computerworld.com, prnewswire.com, bnnbloomberg.ca, Yahoo.com|
|2|Anheuser-Busch|14 March 2018|Boston Beer Company||
|3|Allianz|07 Nov 2017|Unipol||
|4|AT&T|26 Sept 2018|Verizon Communications||
|5|SAP|16 May 2017|Oracle Corp||
|6|Mercedes -Benz|28 June 2017|Stellantis||
|7|Volkswagen|22 Apr 2019|Kia||
|8|BMW|13 Feb 2019|Honda||
|9|Porsche|22 Feb 2018|Renault||
|10|Microsoft|09 Nov 2015|Adobe||
|11|IBM|17 Dec 2015|HP||
|12|Foxconn|06 March 2017|Pegatron||
|13|Nestle|22 Aug 2017|Danone||
|14|Carrefour|06 March 2018|Tesco PLC||
|15|MasterCard|21 Oct 2016|Visa||
|16|Honeywell International|17 Dec 2018|General Electric Company||
|17|JPMorgan Chase & Co.|03 March 2016|Bank of America||
|18|Tyson Foods|22 Aug 2017|Hormel Foods||
|19|Wells Fargo|24 Oct 2016|Regions Financial Corp.||
|20|Coca-Cola|05 Nov 2019|Keurig Dr. Pepper||
|21|FedEx|14 May 2018|Deutsche Post||
|22|Cisco Systems|11 July 2017|Ciena Corp||
|23|HSBC Bank|3 Oct 2017|Credit Agricole||
|24|Deutsche Bank|16 Sept 2019|Commerzbank||
|25|UBS Bank|11 Dec 2017|BNP Paribas||
|26|Maersk|16 Jan 2018|Hapag Lloyd||
-----
|Col1|Company|Official Announcement|Peer company (Infront Analytics, 2023)|Sources of blockchain announcement|
|---|---|---|---|---|
|27|Northern Trust|22 Feb 2017|Key Corp||
|28|Tata Motors|16 Dec 2020|Ashok Leyland||
|29|Morgan Stanley|28 Nov 2018|State Street Corp||
|30|Deutsche Telekom|24 June 2019|Telefonica DE||
In our analysis, we decided to explore an event window of 41 days in total (20
days prior to and after the event day). As an estimation window, we take 200
days, starting from 250 days before the announcement and ending 51 days before
the event (Fig. 1).
Estimation Period Event Period
_Figure 1 – Event Timeline – Estimation Period and Event Period_
Furthermore, we conduct calculations of actual and expected returns of each
company from blockchain and control group using the market model described
by formulas below:
��,� � ln ��[�]����[��]
� (1)
����,����� ��� ∙��� (2)
���,� � ��,� � ����,�� (3)
��
����,� � ����,�
����
(4)
Firstly, we calculate daily returns as natural logarithms (formula 1). The expected
return of company _i_ on day _t is represented by_ ����,�� �formula 2�, and ���
represents the return of the MSCI World, our benchmark index at time t.
ISSN 1335 1745 ( i t) ISSN 1338 984X ( li )
-----
The company’s abnormal returns _ARi,t_ (formula 3) re calculated as a difference
between actual and expected returns. In the next step, cumulative abnormal
return is calculated as a sum of previous abnormal returns during the event
window between t1 and t2. We get the final CAR (formula 4) of all companies as
the sum of average abnormal returns for each day in the event window.
## 4 RESULTS
Using the methodology described earlier, we created charts showing abnormal
returns and cumulative abnormal returns of 30 observed blockchain companies
and 30 companies belonging to the control group (Fig. 2 and Fig. 3).
Statistical tests were then conducted using SPSS Software. The following two
tables display our results for the Blockchain group, indicating that returns on the
event day t0 were slightly negative. Negative performance can also be observed
for three days after the announcement. On the contrary, the control group
performed on t0 positively but in a small magnitude of 0.00172. For the
consequent four days, results change to negative, ranging from -0.00073 to
-0.00318.
Before the announcement, only six out of 20 trading days recorded negative
results. However, the development changed after the announcement when nine
out of 20 trading days ended with negative returns.
As Tab. 3 presents, after the announcement, there was only one statistically
significant day, and it was the day after the day of the announcement. A negative
average abnormal return on the day t+1 (0.4%) can be interpreted as some fear or
insecurity of investors about adopting new technology into a company’s
operations. According to the data, we could observe another three days, which
showed statistical significance: days t-13, t-12 and t-11. Here the abnormal
returns were positive, which can be interpreted as result of some insider
information coming to the market before the announcement.
We tested statistical significance also in the case of the control group. Only the
day t-5 was tested as significant on the 5% level of significance (average ARt-5
was +0.00622) and by 10% level of significance, there were another three days
showing significance, t-13 (average ARt-13 was +0.00479) and t-10 (average
ARt-10 was -0.00536).
In this event study analysis, we considered only blockchain announcements to be
exclusive events that could affect the stock price, while other factors that may
impact the stock price were not considered.
ISSN 1335 1745 ( i t) ISSN 1338 984X ( li )
-----
_Table 2 – One-Sample Statistics – Blockchain Group_
**N** **Mean** **Std.** **Std. Error** **N** **Mean** **Std.** **Std.**
**Deviation** **Mean** **Deviation** **Error**
**Mean**
t-20 30 0.0010 0.0159 0.0029 t+1 30 -0.0036 0.0094 0.0017
t-19 30 0.0037 0.0194 0.0035 t+2 30 -0.0018 0.0071 0.0013
t-18 30 0.0014 0.0104 0.0019 t+3 30 -0.0021 0.0180 0.0033
t-17 30 -0.0005 0.0086 0.0016 t+4 30 0.0008 0.0058 0.0011
t-16 30 0.0030 0.0104 0.0019 t+5 30 0.0013 0.0131 0.0024
t-15 30 -0.0016 0.0070 0.0013 t+6 30 0.0032 0.0155 0.0028
t-14 30 -0.0002 0.0104 0.0019 t+7 30 -0.0009 0.0129 0.0023
t-13 30 0.0075 0.0145 0.0026 t+8 30 -0.0018 0.0100 0.0018
t-12 30 0.0044 0.0093 0.0017 t+9 30 0.0009 0.0093 0.0017
t-11 30 0.0070 0.0188 0.0034 t+10 30 0.0020 0.0089 0.0016
t-10 30 0.0009 0.0125 0.0023 t+11 30 -0.0015 0.0101 0.0018
t-9 30 -0.0017 0.0111 0.0020 t+12 30 0.0022 0.0142 0.0026
t-8 30 0.0007 0.0087 0.0016 t+13 30 0.0030 0.0153 0.0028
t-7 30 -0.0027 0.0121 0.0022 t+14 30 -0.0002 0.0124 0.0023
t-6 30 0.0019 0.0112 0.0020 t+15 30 0.0029 0.0105 0.0019
t-5 30 -0.0008 0.0109 0.0020 t+16 30 -0.0014 0.0102 0.0019
t-4 30 0.0011 0.0137 0.0025 t+17 30 0.0018 0.0246 0.0045
t-3 30 -0.0027 0.0136 0.0025 t+18 30 0.0041 0.0187 0.0034
t-2 30 0.0000 0.0112 0.0020 t+19 30 0.0012 0.0102 0.0019
t-1 30 0.0014 0.0091 0.0017 t+20 30 -0.0013 0.0158 0.0029
t0 30 -0.0013 0.0091 0.0017
Notes: N – Number of observations; Std. Deviation – Standard Deviation; Std. Error Mean – Standard
Error Mean.
_Table 3 – One-Sample T-Test – Blockchain Group_
**t** **Sig.** **Mean Difference** **95% Confidence Interval of the**
**(2-tailed)** **Difference**
Lower Upper
t-20 0.355 0.725 0.00103 -0.00492 0.00698
t-19 1.032 0.311 0.00366 -0.00360 0.01092
t-18 0.754 0.457 0.00144 -0.00246 0.00533
ISSN 1335 1745 ( i t) ISSN 1338 984X ( li )
|Col1|N|Mean|Std. Deviation|Std. Error Mean|Col6|N|Mean|Std. Deviation|Std. Error Mean|
|---|---|---|---|---|---|---|---|---|---|
|t-20|30|0.0010|0.0159|0.0029|t+1|30|-0.0036|0.0094|0.0017|
|t-19|30|0.0037|0.0194|0.0035|t+2|30|-0.0018|0.0071|0.0013|
|t-18|30|0.0014|0.0104|0.0019|t+3|30|-0.0021|0.0180|0.0033|
|t-17|30|-0.0005|0.0086|0.0016|t+4|30|0.0008|0.0058|0.0011|
|t-16|30|0.0030|0.0104|0.0019|t+5|30|0.0013|0.0131|0.0024|
|t-15|30|-0.0016|0.0070|0.0013|t+6|30|0.0032|0.0155|0.0028|
|t-14|30|-0.0002|0.0104|0.0019|t+7|30|-0.0009|0.0129|0.0023|
|t-13|30|0.0075|0.0145|0.0026|t+8|30|-0.0018|0.0100|0.0018|
|t-12|30|0.0044|0.0093|0.0017|t+9|30|0.0009|0.0093|0.0017|
|t-11|30|0.0070|0.0188|0.0034|t+10|30|0.0020|0.0089|0.0016|
|t-10|30|0.0009|0.0125|0.0023|t+11|30|-0.0015|0.0101|0.0018|
|t-9|30|-0.0017|0.0111|0.0020|t+12|30|0.0022|0.0142|0.0026|
|t-8|30|0.0007|0.0087|0.0016|t+13|30|0.0030|0.0153|0.0028|
|t-7|30|-0.0027|0.0121|0.0022|t+14|30|-0.0002|0.0124|0.0023|
|t-6|30|0.0019|0.0112|0.0020|t+15|30|0.0029|0.0105|0.0019|
|t-5|30|-0.0008|0.0109|0.0020|t+16|30|-0.0014|0.0102|0.0019|
|t-4|30|0.0011|0.0137|0.0025|t+17|30|0.0018|0.0246|0.0045|
|t-3|30|-0.0027|0.0136|0.0025|t+18|30|0.0041|0.0187|0.0034|
|t-2|30|0.0000|0.0112|0.0020|t+19|30|0.0012|0.0102|0.0019|
|t-1|30|0.0014|0.0091|0.0017|t+20|30|-0.0013|0.0158|0.0029|
|t0|30|-0.0013|0.0091|0.0017||||||
|Col1|t|Sig. (2-tailed)|Mean Difference|95% Confidence Interval of the Difference|Col6|
|---|---|---|---|---|---|
|||||Lower|Upper|
|t-20|0.355|0.725|0.00103|-0.00492|0.00698|
|t-19|1.032|0.311|0.00366|-0.00360|0.01092|
|t-18|0.754|0.457|0.00144|-0.00246|0.00533|
-----
|Col1|t|Sig. (2-tailed)|Mean Difference|95% Confidence Interval of the Difference|Col6|
|---|---|---|---|---|---|
|||||Lower|Upper|
|t-17|-0.316|0.754|-0.00049|-0.00370|0.00271|
|t-16|1.572|0.127|0.00299|-0.00090|0.00689|
|t-15|-1.26|0.218|-0.00161|-0.00423|0.00101|
|t-14|-0.116|0.909|-0.00022|-0.00411|0.00367|
|t-13|2.843|0.008a|0.00753|0.00211|0.01295|
|t-12|2.607|0.014a|0.00443|0.00096|0.00791|
|t-11|2.054|0.049a|0.00705|0.00003|0.01406|
|t-10|0.413|0.683|0.00094|-0.00373|0.00561|
|t-9|-0.858|0.398|-0.00173|-0.00586|0.00240|
|t-8|0.445|0.659|0.00071|-0.00254|0.00396|
|t-7|-1.22|0.234|-0.00268|-0.00720|0.00183|
|t-6|0.916|0.367|0.00187|-0.00231|0.00606|
|t-5|-0.420|0.678|-0.00084|-0.00493|0.00325|
|t-4|0.423|0.675|0.00105|-0.00404|0.00615|
|t-3|-1.10|0.280|-0.00274|-0.00782|0.00235|
|t-2|0.011|0.991|0.00002|-0.00415|0.00420|
|t-1|0.827|0.415|0.00137|-0.00202|0.00476|
|t0|-0.781|0.441|-0.00130|-0.00472|0.00211|
|t+1|-2.081|0.046a|-0.00355|-0.00705|-0.00006|
|t+2|-1.400|0.172|-0.00180|-0.00444|0.00083|
|t+3|-0.636|0.530|-0.00209|-0.00881|0.00463|
|t+4|0.723|0.476|0.00076|-0.00140|0.00293|
|t+5|0.552|0.585|0.00132|-0.00356|0.00619|
|t+6|1.140|0.263|0.00322|-0.00255|0.00899|
|t+7|-0.392|0.698|-0.00092|-0.00573|0.00388|
|t+8|-0.980|0.335|-0.00179|-0.00552|0.00194|
|t+9|0.554|0.584|0.00094|-0.00252|0.00439|
|t+10|1.246|0.223|0.00203|-0.00130|0.00536|
|t+11|-0.814|0.422|-0.00150|-0.00528|0.00227|
|t+12|0.848|0.403|0.00219|-0.00310|0.00749|
|t+13|1.071|0.293|0.00299|-0.00272|0.00870|
ISSN 1335 1745 ( i t) ISSN 1338 984X ( li )
-----
|Col1|t|Sig. (2-tailed)|Mean Difference|95% Confidence Interval of the Difference|Col6|
|---|---|---|---|---|---|
|||||Lower|Upper|
|t+14|-0.097|0.923|-0.00022|-0.00486|0.00442|
|t+15|1.507|0.143|0.00289|-0.00103|0.00680|
|t+16|-0.743|0.464|-0.00138|-0.00519|0.00243|
|t+17|0.410|0.685|0.00184|-0.00735|0.01103|
|t+18|1.186|0.245|0.00406|-0.00294|0.01106|
|t+19|0.641|0.527|0.00120|-0.00263|0.00502|
|t+20|-0.463|0.647|-0.00133|-0.00722|0.00455|
Notes: Test value – 0; a, b indicate the 5 and 10 percent significance levels; T – value of t-statistic; Sig.
(2-tailed) – two-tailed significance.
The following figures demonstrate abnormal returns of both analysed groups of
companies and cumulative abnormal returns. As seen in Fig. 2, the day of the
announcement caused negative abnormal returns – according to our data, this
occurred by 18 out of 30 companies, and this trend continued for the next three
days. The biggest loss suffered by Volkswagen (abnormal return ARt0 was 2.56%, and the actual return on that day Rt0 was -1.92. The biggest decline
abnormal on the day t+1 had Deutsche Bank (ARt+1 -2.69%) and Morgan
Stanley (ARt+1 -2.15%)
For the selected sample of blockchain companies and the control group, the data
in the selected period showed a similar direction of stock price movements. The
only difference was the magnitude of abnormal returns.
In Figure 3, we can track the development of cumulative abnormal returns. We
see that CAR was more or less positive during the event window, which is also
similar to the general market development between 2015 and 2019, where we
observed an increasing trend. Additionally, we can see outperformance between
the blockchain-adopting companies and their peer companies between days t-16
and t0, which we can interpret as positive expectations of investors about coming
announcements since insider information is common practice on the market.
However, these results show that investors do not yet assign such an important
role in this technology probably because they cannot estimate the long-term
impact. Thus they approach this information rather more cautiously.
ISSN 1335 1745 ( i t) ISSN 1338 984X ( li )
-----
_Figure 2 – Abnormal Returns_
_Figure 3 – Cumulative Abnormal Returns_
To check the robustness of our data, we also calculated the cumulative abnormal
returns of each company and tested different event windows to find out whether
there were some statistically significant periods. Those observed intervals were
<-20,+20>; <-15,+15>; <-10,+10>; <-5,+5>; <-2,+2>; <-1,+1>; <-5,+10>; <5,+15>; <-5,+20>. Within the blockchain group, five out of nine intervals were
slightly negative on average (Tab. 4), while the control group’s results were
slightly positive on average.
Shorter periods around the event day, mostly between t-10 and t+10, were
negative. However, longer intervals above ten days prior to and after the event
showed positive results of abnormal returns.
ISSN 1335 1745 ( i t) ISSN 1338 984X ( li )
-----
As presented in Tab. 5, no interval showed statistical significance. The same
procedure we have done with the control group. After performing statistical tests
in SPSS software, results showed that no event window was statistically
significant.
_Table 4 – One-Sample Statistics – Various Event Windows – Blockchain Group_
**N** **Mean** **Std. Deviation** **Std. Error Mean**
<-20,+20> 30 0.031321 0.118237 0.021587
<-15,+15> 30 0.018313 0.073109 0.013348
<-10,+10> 30 -0.005217 0.049348 0.009010
<-5,+5> 30 -0.007796 0.035402 0.006464
<-2,+2> 30 -0.005267 0.018841 0.003440
<-1,+1> 30 -0.003487 0.014663 0.002677
<-5,+10> 30 -0.004325 0.037792 0.006900
<-5,+15> 30 0.002025 0.054756 0.009997
<-5,+20> 30 0.006407 0.077646 0.014176
Notes: N – Number of observations; Std. Deviation – Standard Deviation; Std. Error Mean – Standard
Error Mean.
_Table 5 – One-Sample T-Test – Various Event Windows – Blockchain Group_
**Test Value** **df** **Sig.** **Mean** **95% Confidence Interval of the**
**= 0** **(2-tailed)** **Difference** **Difference**
**T**
Lower Upper
<-20.+20> 1.451 29 0.158 0.031321 -0.012829 0.075472
<-15.+15> 1.372 29 0.181 0.018313 -0.008987 0.045612
<-10.+10> -0.579 29 0.567 -0.005217 -0.023644 0.013210
<-5.+5> -1.206 29 0.238 -0.007796 -0.021015 0.005423
<-2.+2> -1.531 29 0.137 -0.005267 -0.012302 0.001768
<-1.+1> -1.303 29 0.203 -0.003487 -0.008962 0.001988
<-5.+10> -0.627 29 0.536 -0.004325 -0.018436 0.009787
<-5.+15> 0.203 29 0.841 0.002025 -0.018421 0.022472
<-5.+20> 0.452 29 0.655 0.006407 -0.022587 0.035400
Notes: Test value – 0; T – value of t-statistic; Sig. (2-tailed) – two-tailed significance.
ISSN 1335 1745 ( i t) ISSN 1338 984X ( li )
|Col1|N|Mean|Std. Deviation|Std. Error Mean|
|---|---|---|---|---|
|<-20,+20>|30|0.031321|0.118237|0.021587|
|<-15,+15>|30|0.018313|0.073109|0.013348|
|<-10,+10>|30|-0.005217|0.049348|0.009010|
|<-5,+5>|30|-0.007796|0.035402|0.006464|
|<-2,+2>|30|-0.005267|0.018841|0.003440|
|<-1,+1>|30|-0.003487|0.014663|0.002677|
|<-5,+10>|30|-0.004325|0.037792|0.006900|
|<-5,+15>|30|0.002025|0.054756|0.009997|
|<-5,+20>|30|0.006407|0.077646|0.014176|
|Col1|Test Value = 0|df|Sig. (2-tailed)|Mean Difference|95% Confidence Interval of the Difference|Col7|
|---|---|---|---|---|---|---|
||T||||||
||||||Lower|Upper|
|<-20.+20>|1.451|29|0.158|0.031321|-0.012829|0.075472|
|<-15.+15>|1.372|29|0.181|0.018313|-0.008987|0.045612|
|<-10.+10>|-0.579|29|0.567|-0.005217|-0.023644|0.013210|
|<-5.+5>|-1.206|29|0.238|-0.007796|-0.021015|0.005423|
|<-2.+2>|-1.531|29|0.137|-0.005267|-0.012302|0.001768|
|<-1.+1>|-1.303|29|0.203|-0.003487|-0.008962|0.001988|
|<-5.+10>|-0.627|29|0.536|-0.004325|-0.018436|0.009787|
|<-5.+15>|0.203|29|0.841|0.002025|-0.018421|0.022472|
|<-5.+20>|0.452|29|0.655|0.006407|-0.022587|0.035400|
-----
## 5 CONCLUSIONS
The popularity of blockchain as one of the Quality 4.0 instruments has grown
rapidly in business and academic communities. It has the potential to enhance the
data transparency and quality which optimises company operations and thus
increase value for shareholders. This paper aimed to analyse the impact of
blockchain announcements on selected international companies using an event
study approach. Our objective was to answer the following research question set
at the beginning of our analysis.
RQ: Is there any reaction in stock prices after the company officially announces
the application of blockchain technology in its operations?
H0: There has been no reaction in stock prices after the company’s
announcement regarding blockchain implementation.
Using the event study approach and SPSS software we analysed our data sample
consisting of two groups of sixty companies in total. Within the blockchain
group. three days before the event were statistically significant with positive
results (t-13. t-12 and t-11). which can be interpreted as some insider information
or signals spread on the market about planned announcements. However, after
the event only the first day after the announcement (t+1) was tested as
statistically significant with a negative reaction (-0.4%). This can be connected to
cautious investors on the market when talking about some new technology as
blockchain. where current knowledge is probably not sufficient yet. and the
technology needs to be explored more.
According to our data. there was no reaction on the market after the
announcement. and the significance of t+1 day was rather random as systematic.
Thus we do not reject the null hypothesis. and we can summarise that there has
been no reaction in stock prices after the company’s announcement regarding
blockchain implementation.
The importance and maturity of blockchain will rise in the next years. and thus
every additional study around this topic will be important to extend the pool of
knowledge. Once it is properly established in the market. evaluating its long-term
impact on companies will be interesting. Therefore. we hereby encourage
researchers to analyse in the future the results on a long-term basis to conclude
whether the blockchain positively influences companies’ operations and whether
this technology is worth investment of such considerable financial resources.
## REFERENCES
American Society for Quality, n.d. _Quality Glossary Definition: Quality 4.0._
[online] Available at: <https://asq.org/quality-resources/quality-40#:~:text=%22Quality%204.0%22%20is%20a%20term.%2C%20digital%2C%2
0and%20disruptive%20technologies> [Accessed 04 April 2023].
ISSN 1335 1745 ( i t) ISSN 1338 984X ( li )
-----
Bose, I. and Man Leung, A.C., 2019. Adoption of identity theft countermeasures
and its short- and long-term impact on firm value. _MIS Quarterly, [e-journal]_
43(1), pp.313-327. DOI: 10.25300/misq/2019/14192.
Boyd, D.E., Kannan, P.K. and Slotegraaf. R.J., 2019. Branded apps and their
impact on firm value: A design perspective. Journal of Marketing Research, [ejournal] 56(1), pp.76-88. DOI: 10.1177/0022243718820588.
Bradley, R.V. and Esper, T.L., 2018. The joint use of RFID and EDI:
Implications for hospital performance. Production and Operations Management,
[e-journal] 27(11), pp.2071-2090. DOI: 10.1111/poms.12955.
Cahill, D., Baur. D.G., Liu, Z. and Yang, J.W., 2020. I am a blockchain too: How
does the market respond to companies' interest in blockchain?. _Journal of_
_Banking_ _and_ _Finance,_ [e-journal] 113, 105740. DOI:
10.1016/j.jbankfin.2020.105740.
Campbell, C.J., Cowan, A.R. and Salotti, V., 2010. Multi-country event-study
methods. Journal of Banking & Finance, [e-journal] 34(12), pp.3078-3090. DOI:
10.1016/j.jbankfin.2010.07.016.
Chen, K., Lai, T.L., Liu, Q. and Wang, C., 2022. Beyond the blockchain
announcement: Signaling credibility and market reaction. International Review of
_Financial Analysis, [e-journal] 82, 102209. DOI: 10.1016/j.irfa.2022.102209._
Cheng, S.F., De Franco, G., Jiang, H. and Lin, P., 2019. Riding the blockchain
mania: Public firms’ speculative 8-K Disclosures. _Management Science, [e-_
journal] 65(12), pp.5901-5913. DOI: 10.1287/mnsc.2019.3357.
Dehning, B., Richardson, V.J., Urbaczewski, A. and Wells, J.D., 2004.
Reexamining the value relevance of E-commerce initiatives. _Journal of_
_Management_ _Information_ _Systems,_ [e-journal] 21(1), pp.55-82. DOI:
10.1080/07421222.2004.11045788
Hendricks, K.B., Singhal, V.R. and Stratman, J.K., 2006. The impact of
Enterprise Systems on corporate performance: A study of ERP, SCM, and CRM
system implementations. _Journal of Operations Management, [e-journal] 25(1),_
pp.65-82. DOI: 10.1016/j.jom.2006.02.002.
Infront Analytics, 2023. _Find. Compare. Evaluate. [online] Available at:_
<https://www.infrontanalytics.com/> [Accessed 13 March 2023].
Klockner, M., Schmidt, C.G. and Wagner, S.M., 2022. When Blockchain Creates
Shareholder Value: Empirical Evidence from International Firm Announcements.
_Production and Operations Management, [e-journal] 31(1), pp.46-64. DOI:_
10.1111/poms.13609.
Lacity, M.C., 2018. Addressing key challenges to making enterprise blockchain
applications a reality. MIS Quarterly Executive, 17(3), pp.201-222.
ISSN 1335 1745 ( i t) ISSN 1338 984X ( li )
-----
Liu, W., Wang, J., Jia, F. and Choi, T., 2022. Blockchain announcements and
stock value: A Technology Management Perspective. _International Journal of_
_Operations & Production Management, [e-journal] 42(5), pp.713-742. DOI:_
10.1108/ijopm-08-2021-0534.
Melville, N., Kraemer, K. and Gurbaxani, V., 2004. Information technology and
organisational performance: An integrative model of IT business value. _MIS_
_Quarterly, [e-journal] 28(2), pp.283-322. DOI: 10.2307/25148636._
Mtotywa, M., 2022. Developing a Quality 4.0 Maturity Index for Improved
Business Operational Efficiency and Performance. _Quality Innovation_
_Prosperity, [e-journal] 26(2), pp.101-127. DOI: 10.12776/QIP.V26I2.1718._
Nguyen, N., Nguyen, Ch., Nguyen, H. and Nguyen, V., 2021. The Impact of
Quality Management on Business Performance of Manufacturing Firms: The
Moderated Effect of Industry 4.0. Quality Innovation Prosperity, 25(3), pp.120135. DOI: 10.12776/QIP.V25I3.1623.
Paul, S., Adhikari, A. and Bose, I., 2022. White Knight in dark days? Supply
chain finance firms. blockchain. and the COVID-19 pandemic. _Information &_
_Management, [e-journal] 59(6), 103661. DOI: 10.1016/j.im.2022.103661._
Ranganathan, C. and Brown, C.V., 2006. ERP investments and the market value
of firms: Toward an understanding of influential ERP project variables.
_Information_ _Systems_ _Research,_ [e-journal] 17(2), pp.145-161. DOI:
10.1287/isre.1060.0084.
Reno, G., 2022. 12 _Things You Can Do to Improve Data Quality. [online]_
Napperville: FirstEigen. Available at: <firsteigen.com/blog/12-things-you-cando-to-improve-data-quality/> [Accessed 10 March 2023].
Santos, G., Sá, J.C., Félix, M.J., Barreto, L., Carvalho, F., Doiro, M., Zgodavová,
K. and Stefanović, M., New Needed Quality Management Skills for Quality
Managers 4.0. _Sustainability,_ [e-journal] 13(11), 6149. DOI:
10.3390/su13116149.
Subramani, M. and Walden, E., 2001. The impact of e-commerce announcements
_on_ _the_ _market_ _value_ _of_ _firms._ Available at SSRN:
<https://ssrn.com/abstract=269668 or http://dx.doi.org/10.2139/ssrn.269668>
[Accessed 10 March 2023].
Tampi, T., 2020. VeriTX to use blockchain to transform aerospace 3D printing
supply chain. _3Dprint.com,_ [online] 02 November. Available at:
<https://3dprint.com/274744/veritx-to-use-blockchain-to-transform-aerospace3d-printing-supply-chain/> [Accessed 10 March 2023].
Ullah, S., Zaefarian, G., Ahmed, R. and Kimani, D., 2021. How to apply the
event study methodology in Stata: An overview and a step-by-step guide for
authors. _Industrial Marketing Management, [e-journal] 99(November 2021),_
pp.A1-A12. DOI: 10.1016/j.indmarman.2021.02.004.
ISSN 1335 1745 ( i t) ISSN 1338 984X ( li )
-----
## ABOUT AUTHORS
**Vladimíra Gimerská[0000-0001-7108-7201 ](V.G.) – PhD student, Technical University**
of Košice, Faculty of Economics, Slovak Republic, e-mail:
gimerska.vladimira@gmail.com.
**Michal Šoltés[0000-0002-1421-7177 ]** (M.S.) – Assoc. Prof., Dean of the Faculty of
Economics, Technical University of Košice, Slovak Republic, e-mail:
michal.soltes@tuke.sk.
**Rajmund Mirdala[0000-0002-9949-3049 ]** (R.M.) – Prof., Department of Economics,
Faculty of Economics, Technical University of Košice, Slovak Republic, e-mail:
rajmund.mirdala@tuke.sk.
## AUTHOR CONTRIBUTIONS
Conceptualisation. V.G.; Methodology, V.G.; Formal analysis, V.G.;
Investigation, V.G.; Original draft preparation, V.G.; Review and editing, M.S.
and R.M.; Visualization, V.G.; Supervision, M.S. and R.M.
## CONFLICTS OF INTEREST
The authors declare no conflict of interest. The funders had no role in the design
of the study; in the collection. analyses. or interpretation of data; in the writing of
the manuscript. or in the decision to publish the results.
© 2023 by the authors. Submitted for possible open access publication under the
terms and conditions of the Creative Commons Attribution (CC-BY) license
(http://creativecommons.org/licenses/by/4.0/).
ISSN 1335 1745 ( i t) ISSN 1338 984X ( li )
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.12776/qip.v27i2.1877?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.12776/qip.v27i2.1877, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "GOLD",
"url": "https://www.qip-journal.eu/index.php/QIP/article/download/1877/1369"
}
| 2,023
|
[
"JournalArticle"
] | true
| 2023-07-31T00:00:00
|
[
{
"paperId": "7edfb72b0a06d49eca54f19b92f0fad67868a4a1",
"title": "Developing a Quality 4.0 Maturity Index for Improved Business Operational Efficiency and Performance"
},
{
"paperId": "16c463725c5093fd4e7f3048a86f2a96f6f7b419",
"title": "Beyond the blockchain announcement: Signaling credibility and market reaction"
},
{
"paperId": "c277fd4a221725eee919eb655af30b88e9b4fd2e",
"title": "White knight in dark days? Supply chain finance firms, blockchain, and the COVID-19 pandemic"
},
{
"paperId": "eec034c0d306b6bcd33d83f0188cc9c2571521c2",
"title": "Blockchain announcements and stock value: a technology management perspective"
},
{
"paperId": "7af62dcc5c954d63c2673b8cc40929619a84dc04",
"title": "When Blockchain Creates Shareholder Value: Empirical Evidence from International Firm Announcements"
},
{
"paperId": "2bf57fbdc071ec75415192bad3fc329771fcaad0",
"title": "How to apply the event study methodology in STATA: An overview and a step-by-step guide for authors"
},
{
"paperId": "03454a5e070413c3974a2bffdab16a8a085dffac",
"title": "Riding the Blockchain Mania: Public Firms’ Speculative 8-K Disclosures"
},
{
"paperId": "512566eecd9189e6d76828affee72c8de20dcdf9",
"title": "Branded Apps and Their Impact on Firm Value: A Design Perspective"
},
{
"paperId": "1fb9bec1ae8850c35549dc11345f07ccd32e3fbb",
"title": "Adoption of Identity Theft Countermeasures and its Short- and Long-Term Impact on Firm Value"
},
{
"paperId": "1bf7de009e2dc8d8c5e077a77ebba1b44eef9476",
"title": "The Joint Use of RFID and EDI: Implications for Hospital Performance"
},
{
"paperId": "135d16b93ee0b407446740d7651c36393182a0c2",
"title": "I Am a Blockchain Too..."
},
{
"paperId": "2e29693babcc1f7734b58f7de0f89fc56fa2ed1a",
"title": "Multi-Country Event Study Methods"
},
{
"paperId": "41eb2c8e05a7566e4d16f76255f202eb260efe44",
"title": "ERP Investments and the Market Value of Firms: Toward an Understanding of Influential ERP Project Variables"
},
{
"paperId": "ceb847ccef950d8f53c9a3f2cb1a10669dafa46e",
"title": "Reexamining the Value Relevance of E-Commerce Initiatives"
},
{
"paperId": "fa02855ba4688a077b22607f420f5be7fd4f9139",
"title": "Review: Information Technology and Organizational Performance: An Integrative Model of IT Business Value"
},
{
"paperId": "8282feec96a60372129f8915c788e7d2abb8eda5",
"title": "The Impact of E-Commerce Announcements on the Market Value of Firms"
},
{
"paperId": null,
"title": "2022. 12 Things You Can Do to Improve Data Quality. [online] Napperville: FirstEigen"
},
{
"paperId": "28019620225469236f764fa91f326582eff57c10",
"title": "The Impact of Quality Management on Business Performance of Manufacturing Firms: The Moderated Effect of Industry 4"
},
{
"paperId": "1773a51be5142a516a7622a5f8672bd7310949fc",
"title": "Addressing Key Challenges to Making Enterprise Blockchain Applications a Reality"
},
{
"paperId": "3732d243e534fed43de5065d39bfcc4a8d6de3bd",
"title": "The impact of enterprise systems on corporate performance: A study of ERP, SCM, and CRM system implementations"
},
{
"paperId": null,
"title": "12 Things You Can Do to Improve Data Quality. [online] Napperville: FirstEigen. Available at: <firsteigen.com/blog/12-things-you-cando-to-improve-data-quality/> [Accessed"
},
{
"paperId": null,
"title": "VeriTX to use blockchain to transform aerospace 3D printing supply chain. 3Dprint.com"
},
{
"paperId": null,
"title": "There has been no reaction in stock prices after the company’s announcement regarding blockchain implementation"
}
] | 13,152
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Mathematics",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/0293ee9829d6cd3d9a774e7486b068205ca1179e
|
[
"Computer Science",
"Mathematics"
] | 0.866338
|
Attack on the Edon-kKey Encapsulation Mechanism
|
0293ee9829d6cd3d9a774e7486b068205ca1179e
|
International Symposium on Information Theory
|
[
{
"authorId": "40323363",
"name": "Matthieu Lequesne"
},
{
"authorId": "1764813",
"name": "J. Tillich"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"International Symposium on Information Technology",
"Int Symp Inf Theory",
"Int Symp Inf Technol",
"ISIT"
],
"alternate_urls": null,
"id": "234ccdc0-f58f-4f94-b86a-428d11a0c5ad",
"issn": null,
"name": "International Symposium on Information Theory",
"type": "conference",
"url": "http://www.wikicfp.com/cfp/program?id=1719"
}
|
The key encapsulation mechanism $\text{EDON}-\mathcal{K}$ was proposed in response to the call for post-quantum cryptography standardization issued by the National Institute of Standards and Technologies (NIST). This scheme is inspired by the McEliece scheme but uses another family of codes defined over $\mathbb{F}_{2^{128}}$ instead of $\mathbb{F}_{2}$ and is not based on the Hamming metric. It allows significantly shorter public keys than the McEliece scheme. In this paper, we give a polynomial time algorithm that recovers the encapsulated secret. This attack makes the scheme insecure for the intended use. We obtain this result by observing that recovering the error in the McEliece scheme corresponding to $\text{EDON}-\mathcal{K}$ can be viewed as a decoding problem for the rank-metric. We show that the code used in $\text{EDON}-\mathcal{K}$ is in fact a super-code of a Low Rank Parity Check (LRPC) code of very small rank (1 or 2). A suitable parity-check matrix for the super-code of such low rank can be easily derived from for the public key. We then use this parity-check matrix in a decoding algorithm that was devised for LRPC codes to recover the error. Finally we explain how we decapsulate the secret once we have found the error.
|
## Attack on the EDON-K Key Encapsulation Mechanism
##### Matthieu Lequesne
Sorbonne Université, UPMC Univ Paris 06
Inria, Team SECRET,
2 rue Simone Iff, CS 42112,
75589 Paris Cedex 12, France
Email: matthieu.lequesne@inria.fr
Abstract—The key encapsulation mechanism EDON-K was
proposed in response to the call for post-quantum cryptography
standardization issued by the National Institute of Standards and
Technologies (NIST). This scheme is inspired by the McEliece
scheme but uses another family of codes defined over F2128
instead of F2 and is not based on the Hamming metric. It allows
significantly shorter public keys than the McEliece scheme.
In this paper, we give a polynomial time algorithm that
recovers the encapsulated secret. This attack makes the scheme
insecure for the intended use. We obtain this result by observing
that recovering the error in the McEliece scheme corresponding
to EDON-K can be viewed as a decoding problem for the rankmetric. We show that the code used in EDON-K is in fact a
super-code of a Low Rank Parity Check (LRPC) code of very
small rank (1 or 2). A suitable parity-check matrix for the supercode of such low rank can be easily derived from for the public
key. We then use this parity-check matrix in a decoding algorithm
that was devised for LRPC codes to recover the error. Finally
we explain how we decapsulate the secret once we have found
the error.
I. INTRODUCTION
The syndrome decoding problem is a fundamental problem in complexity theory, since the original paper of
Berlekamp, McEliece and van Tilborg [BMvT78] proving its
NP-completeness for the Hamming distance. The same year,
McEliece proposed a public-key cryptosystem based on this
problem [McE78] and instantiated it with binary Goppa codes.
This scheme was for a long time considered inferior to RSA
due to its large key size. However, this situation has changed
drastically when it became apparent in [Sho94] that RSA
and actually all the other public-key cryptosystems used in
practice could be attacked in polynomial time by a quantum
computer. There are now small prototypes of such computers
that lead to think that they will become a reality in the
future and in 2016, the National Institute of Standards and
Technology (NIST) announced a call for standardization of
cryptosystems that would be safe against an adversary equiped
with a quantum computer. Four families of cryptosystems
are often mentioned as potential candidates: cryptosystems
based on error correcting codes, lattices, hash functions and
multivariate quadratic equations [BBD09]. All of these are
based on mathematical problems that are expected to remain
hard even in the presence of a quantum computer.
##### Jean-Pierre Tillich
Inria, Team SECRET,
2 rue Simone Iff, CS 42112,
75589 Paris Cedex 12, France
Email: jean-pierre.tillich@inria.fr
The key encapsulation mechanism EDON-K [GG17] was
proposed by Gligoroski and Gjøsteen in response to the call
issued by the NIST. This scheme is inspired by the McEliece
scheme but uses another family of codes defined over F2128
instead of F2. This choice leads to very short keys for a codebased scheme. The metric used for the decoding is not properly
defined and the security relies on an ad-hoc problem named
finite field vector subset ratio problem supposedly hard on
average.
In this paper, we show that the metric used for EDON-K is
in fact equivalent to the well-known rank metric. This metric
was first introduced in 1951 as “arithmetic distance” between
matrices over a field Fq [Hua51]. The notion of rank distance
and rank codes over matrices was defined in 1978 by Delsarte
[Del78]. He introduced a code family, named maximum rank
distance (MRD) codes, that attains the analogue of the MDS
(maximum distance separable) bound for the rank metric.
Gabidulin suggests in [Gab85] to consider a subfamily of
such codes that are linear over an extension field Fqm. It
provides a vectorial representation of these codes and allows
to represent them in a much more compact way. This is the
main reason why the rank metric based McEliece schemes
achieve significantly smaller key sizes. Moreover this vectorial
representation allows to view the known families of MRD
codes as rank metric analogues of Reed-Solomon codes and
to obtain an efficient decoding algorithm for them [Gab85].
There are also rank metric analogues for other families of
codes. For instance, the Low Rank Parity-Check (LRPC) codes
introduced in [GMRZ13] can be considered as analogues
of Low Density Parity-Check (LDPC) codes. Just like their
binary cousins, they enjoy an efficient decoding algorithm that
is based on a low rank parity-check matrix of such a code.
Here, we prove that the code used in EDON-K is a actually
a super-code of an LRPC code of rank 2. What is more, this
LRPC code is itself a subspace of codimension 1 of another
LRPC code of rank 1. It turns out that parity-check matrices
of rank 2 for the first super-code and rank 1 for the second
one can easily be derived from the public key. In both cases,
this allows us to decode the ciphertext without the secret key.
This gives a way to recover the encapsulated secret and breaks
completely the EDON-K system.
-----
The paper is organized as follows. First, we recall some
basic definitions and properties of rank metric and LRPC codes
in Section II. In Section III we present the scheme of EDON-K.
Then we explain the general idea of our attack in section IV. In
Section V, we detail how we reconstruct a parity-check matrix
of the code and in Section VI how we decode the ciphertext.
In Section VII, we explain how we derive the encapsulated
secret from the error. Finally in Section VIII we discuss the
cost of this attack and its consequences.
II. RANK METRIC CODES
A. Notation
In the following document, q denotes a power of a prime
number. In the case of EDON-K, we will have q = 2. Fq
denotes the finite field with q elements and, for any positive
integer m, Fqm denotes the finite field with q[m] elements. We
will sometimes view Fqm as an m-dimensional vector space
over Fq.
We use bold lowercase and capital letters to denote vectors
and matrices respectively.
We denote ⟨x1, . . ., xk⟩K the K-vector space generated by
the elements {x1, . . . xk}.
B. Definitions
Definition 1 (Rank metric over F[n]q[m][)][.][ Let][ x][ = (][x][1][, . . ., x][n][)][ ∈]
F[n]q[m][ and][ (][β][1][, . . ., β][m][)][ be a basis of][ F][q][m][ viewed as an][ m][-]
dimensional vector space over Fq. Each coordinate xj ∈
Fqm is associated to a vector of F[m]q in this basis: xj =
�m
i=1 [m][i,j][β][i][. The][ m][ ×][ n][ matrix associated to][ x][ is given by]
M(x) := (mi,j)1≤i≤m,1≤j≤n.
The rank weight wt(x) of x is defined as :
wt(x) := Rank M(x).
The associated distance d(x, y) between elements x and y
of F[n]q[m][ is defined by][ d][(][x][,][ y][) :=][ wt][(][x][ −] [y][)][.]
Definition 2 (Support of a word). Let x = (x1, . . ., xn) ∈
F[n]q[m] [. The support of][ x][, denoted][ Supp(][x][)][, is the][ F][q][-subspace]
of Fqm generated by the coordinates of x:
Supp(x) := ⟨x1, . . ., xn⟩Fq .
We have dim(Supp(x)) = wt(x).
Definition 3 (Fqm-linear code). An Fqm-linear code C of
dimension k and length n is a subspace of dimension k of
F[n]q[m] [.][ C][ can be represented in two equivalent ways: by a]
generator matrix G ∈ Fq[k][m][×][n] such that C = {xG | x ∈ F[k]q[m] [}]
and by a parity-check matrix H ∈ Fq[(][n][m][−][k][)][×][n] such that
C = {x ∈ F[n]q[m][ |][ Hx][⊺] [=][ 0][n][−][k][}][.]
The decoding problem in the rank metric can be described
as follows.
Problem 1 (Decoding problem for the rank metric). Let C
be an Fqm -linear code of dimension k and length n. Given
y = c + e where c ∈C and e ∈ F[n]q[m][ is of rank weight][ ≤] [r]
find c and e.
C. LRPC codes
Definition 4 (LRPC code). A Low Rank Parity Check (LRPC)
code of rank d, length n and dimension k over Fqm is a code
that admits a parity-check matrix H = (hi,j ) ∈ F[(]q[n][m][−][k][)][×][n]
such that the vector space of Fqm generated by its coefficients
hi,j has dimension at most d.
LRPC codes can be viewed as analogues of LDPC codes
for the rank metric. In particular, they enjoy an efficient
decoding algorithm based on their low rank parity-check
matrix. Roughly speaking, Algorithm 1 of [GMRZ13] decodes
up to d errors when rd ≤ n − k in polynomial time (see
[GMRZ13, Theorem 1]). It uses in a crucial way the notion
of the linear span of a product of subspaces of Fqm
Definition 5. Let U and V be two Fq subspaces of Fqm . We
denote by U · V the linear span of the product of U and V :
U · V := ⟨uv : u ∈ U, v ∈ V ⟩Fq .
III. THE EDON-K KEM
EDON-K [GG17] is a key encapsulation mechanism proposed by Gligoroski and Gjøsteen for the NIST post-quantum
cryptography call. Here we describe the key generation, encapsulation and decapsulation, omitting some details that are
not relevant for the attack. We refer to [GG17] for the full
description.
A. Parameters and notations
The parameters for EDON-K are given in Table I. In this
paper we often refer to the parameters of edonk128ref, the
reference version proposed for 128 security-bits.
TABLE I
PARAMETERS PROPOSED FOR EDON-K
Name m N K R ν L
edonk128ref 128 144 16 40 8 6
edonk128K16N80nu8L6 128 80 16 40 8 6
edonk128K08N72nu8L8 128 72 8 40 8 8
edonk128K32N96nu4L4 128 96 32 40 4 4
edonk128K16N80nu4L6 128 80 16 40 4 6
edonk192ref 192 112 16 40 8 8
edonk192K48N144nu4L4 192 144 48 40 4 4
edonk192K32N128nu4L6 192 128 32 40 4 6
edonk192K16N112nu4L8 192 112 16 40 4 8
The scheme makes use of a hash function H (·) corresponding to standard SHA2 functions (SHA-256 or SHA384 depending on the parameters). We will denote H[i](·) :=
H(. . . H(·))
� i times�� �
Given a binary matrix P = (pi,j ) and two non-zero
elements a ̸= b of F2m, Pa,b = (˜pi,j) denotes the matrix
of the same size with coefficients in F2m where ˜pi,j = a if
pi,j = 0 and ˜pi,j = b if pi,j = 1.
-----
In particular, if P is orthogonal:
Pc,d⊺ = (Pa,b)−1 (1)
where c := a[2]+a b[2][ and][ d][ :=] a[2]+b b[2][ .]
For two vectors (or matrices) x and y, we will denote x||y
their concatenation.
B. Key generation
Given the security level and the appropriate parameters.
$
- a, b ← F2m non-zero elememts such that a ̸= b.
$ N ×N
- P ← F2 an orthogonal matrix.
$ R×N
- H ← F2 such that H = [HT ||HB][⊺] where HB is an
R × R orthogonal matrix and HT has columns of even
Hamming weight.
a b
- c := a[2]+b[2][,][ d][ :=] a[2]+b[2][ .]
$ ν
- ˜g ← F2[m] [.]
- Vg := Support(˜g).
$ K×N
- G ←Vg such that
GH[⊺] = 0K×R. (2)
⊺
- Gpub := GPc,d .
- Return (PublicKey := Gpub, SecretKey := (a, b, P, H)).
C. Encapsulation
Given the PublicKey and the public parameters.
$ K
- m ← F2[m][.]
- ˜e ∈ F[L]2[m][ generated as follows:]
$
– (˜e0, ˜e1) ← F2m;
– for 1 ≤ i ≤ [L]2 [−] [1][,][ (˜][e][2][i][,][ ˜][e][2][i][+1][) =][ H][ (˜][e][2][i][−][2][||][e][˜][2][i][−][1][)][.]
- Ve := Support(˜e).
$ N
- e ←Ve [.]
- c := mGpub + e.
- (s0, s1) := H (˜eL−2||e˜L−1).
- SharedSecret := H (s0||s1||H (c)).
- h := H (s1||so||H (c)).
- Ciphertext := (c, h).
- Return (Ciphertext, SharedSecret).
D. Decapsulation
Given Ciphertext, SecretKey and the public parameters.
- Recover e by decoding the c using the private matrix
H[′] := HPa,b⊺.
- Deduce Ve the vector space spaned by the coefficients of
the vector e.
- For all (λ, ν) ∈Ve × Ve, for 1 ≤ i ≤ [L]2 [−] [1][:]
– (s[′]0[, s][′]1[) :=][ H][i][ (][λ][||][µ][||H][ (][c][))][;]
– if H (s[′]1[||][s][′]0[||][c][) =][ h][:]
Return SharedSecret := H (s[′]0[||][s][′]1[||H][ (][c][))][.]
IV. OUTLINE OF THE ATTACK ON EDON-K
Our attack is based on three observations
- The ciphertext is a vector c such that
c = mGpub + e. (3)
This error e is of low rank, since its rank is at most L.
- This code Cpub generated by Gpub is a subcode of an
LRPC code, namely the code C[′] with parity-check matrix
H[′] := HPa,b⊺. This code is indeed an LRPC code of
rank 2 since all the entries of H[′] belong to ⟨a, b⟩F2. We
have
Cpub ⊂C[′] (4)
since
GpubH[′][⊺] = GPc,d⊺(HPa,b⊺)[⊺]
= GPc,d⊺Pa,bH⊺
= GH[⊺] (from (1))
= 0K×R (from (2)).
This equation also appears as Corollary 1 of [GG17,
p.19]. We have given its proof here for the convenience
of the reader. Let K [′] = N − R be the dimension of C[′].
- If we recover a parity-check matrix of rank 2 for C[′] we
will be able to recover mGpub and e from c. Indeed,
mGpub ∈C[′] and we can decode C[′] using a variation
of Algorithm 1 of [GMRZ13] and the knowledge of
the parity-check matrix, provided wt(e) ≤ L < (N −
K [′])/2 = R/2 is verified, which is the case for the
parameters of EDON-K.
Hence we will proceed in three steps:
1) constructing and solving a linear system of equations to
find a parity-check matrix for the code C[′] (detailed in
Section V);
2) decoding the ciphertext using a slight variation of Algorithm 1 of [GMRZ13] (see Section VI);
3) recovering the secret from the error vector (explained in
Section VII).
V. RECONSTRUCTING THE PARITY-CHECK MATRIX
A. Compressed public key
In order to reduce the public key size, the designers
of EDON-K chose to represent the public key in a compressed form. They took advantage of the fact that all the
coefficients of Gpub live in the vector space Vg,c,d :=
⟨cg˜1, . . ., cg˜ν, dg˜1, . . ., dg˜ν⟩F2 of dimension 2ν. Hence, the
compressed public key consists in two parts: first the basis
g˜c,d := (cg˜1, . . ., cg˜ν, dg˜1, . . ., dg˜ν) ∈ F[2]2[m][ν] [ of the vector-]
space Vg,c,d, then the entries of the matrix Gpub such that
each entry is represented by its coefficients in the basis ˜gc,d.
For example, if an entry x of Gpub is equal to c [�]i[ν]=1 [γ][i][g][˜][i][ +]
d i=1 [δ][i][g][˜][i][ with][ γ][i][, δ][i] ∈ F2, x will be represented by
[�][ν]
(γ1, . . ., γν, δ1, . . ., δν) ∈ F[2]2[ν][. There is another subtlety in]
the compression that we will not mention here.
-----
B. Finding a basis
The attacker does not have access to the value of a and b but
can deduce the value of ab[−][1] = cd[−][1] = (cg˜1)(dg˜1)[−][1] from
g˜c,d as mentioned in paragraph 7.2.2 of the documentation of
EDON-K [GG17].
Let us bring in
α := ab[−][1].
We notice that H” := b[−][1]H[′] is also a parity-check matrix
of the LRPC code C[′]. This matrix has all its coefficients
in ⟨1, α⟩F2. We use this information to reconstruct such a
parity-check matrix of the code C[′] by solving a linear system,
similarly to what is done in [GRS16, Section IV B]. This
system is derived from the following facts:
(i) Gpub H[′′][⊺] = 0K×R;
(ii) the entries of H[′′] belong to ⟨1, α⟩F2.
In other words, the possible rows x = (x1, . . ., xN ) of H[′′]
are solutions of the following system
� Gpubx⊺ = 0K (5)
xi ∈ ⟨1, α⟩F2 for all i ∈{1, . . ., N }.
This system is obviously linear over F2 and the solution set
is an F2-linear subspace. A basis of this subspace can then be
used as rows for H[′′]. We now show that solving this system
can be done by solving a linear system over F2.
C. Recovering H[′′] by solving a linear system over F2 and an
affine system in a more general case
Actually in this section we will consider a more general
version of (5). Given a system
Ax[⊺] = b[⊺] (6)
where A = (aij )1≤i≤r,1≤j≤N is a given matrix in F[r]2[×][m][N] and b
is a given vector in F[r]2[m][, and given][ V][ a subspace of dimension]
t of F2m (viewed as vector space over F2 of dimension m),
how to find the affine set of the solutions x = (xi)1≤i≤N ∈
V [N] of the system?
We can rewrite the system (6) as
a11x1 + · · · + a1N xN = b1
- · · = - · · (7)
ar1x1 + · · · + arN xN = br.
We introduce a basis {v1, . . ., vt} of V and express each
unknown xj in this basis in terms of t other unknowns
xj1, . . ., xjt ∈ F2:
Let {β1, . . ., βm} be an F2-basis of F2[m], we introduce for
1 ≤ ℓ ≤ m the projection πℓ from F2m to F2 defined by:
πℓ : a = Fj2=1m [a][j][β][j] �−→−→ aFℓ2. (9)
[�][m]
The r equations of system (8) defined over F2m lead to rm
affine equations over F2 by applying πℓ for ℓ ∈{1, . . ., m}:
N t
�j=1 �i=1 [π][ℓ][(][a][1][j] [v][i][)][x][ji] = πℓ(b1)
. . . = . . . (10)
�Nj=1 �ti=1 [π][ℓ][(][a][rj][v][i][)][x][ji] = πℓ(br).
We can solve this affine system in F2 to recover the solution
of (6). The system has rm binary equations and tN unknowns,
hence a complexity of O(rmt[2]N [2]). If we apply this technique
to (5), where t = 2 and r = K we obtain a basis of the vector
space in time O(KmN [2]).
VI. DECODING STEP
The previous step recovers an R × N matrix H[(3)] whose
entries all belong to ⟨1, α⟩F2. The matrices H[(3)] and H[′′] share
the property that their rows form a basis of solutions of (5).
Therefore, there exists an R × R binary invertible matrix Q
such that
H[(3)] = QH[′′]. (11)
We use H[(3)] to decode and recover e from the ciphertext
c. The vectors are linked by the equation
c = mGpub + e. (12)
We use here a slight variation of Algorithm 1 of [GMRZ13]
to decode. Algorithm 1 would consist in performing the
following steps:
1) Compute s[⊺] := H[(3)]c[⊺] and then V := Support(s). Here
we typically have V = Support(e) - 1, α⟩Fq when H[(3)]
is a random matrix.
2) Compute V [′] := V ∩ α[−][1]V . This step typically recovers
Support(e) when V = Support(e) · ⟨1, α⟩Fq .
3) Once we have Support(e) we recover e = (e1, . . ., eN )
by solving the linear equation H[(3)]e[⊺] = s[⊺] with the additional constraints ei ∈ Support(e) for i ∈{1, . . ., N }.
This is done by using the technique given in Subsection
V-C.
In our case, due to the special structure of H which contains
only a’s and b’s V is not equal to Support(e) · ⟨1, α⟩Fq . This
is due to the following result.
Proposition 1. We have for every e ∈ F[N]2[m][:]
xj =
t
�
xjivi.
i=1
�
.
F2
� N
� ei
i=1
In other words, the system (6) is equivalent to
N t
�j=1 �i=1 [a][1][j][v][i][x][ji] = b1
. . . = . . . (8)
�Nj=1 �ti=1 [a][rj][v][i][x][ji] = br.
Support(H[(3)]e[⊺]) ⊂ (1 + α)Support(e) +
Proof. From (11), we deduce that
Support(H[(3)]e[⊺]) = Support(QH[′′]e[⊺]) = Support(H[′′]e[⊺]).
-----
Let s[⊺] := H[′′]e[⊺]. Denote the i-entry of s by si and the entry
of H[′′] in row i and column j by h[′′]ij [. We have:]
si =
N
�
h[′′]ij[e][j]
j=1
� �
= ej + αej
j s.t. h[′′]ij [=1] j s.t. h[′′]ij [=][α]
=
N
� �
ej + (1 + α) ej.
j=1 j s.t. h[′′]ij [=][α]
This implies the proposition.
This proposition directly gives a subspace of dimension
L + 1 that contains Support(e) since we deduce from it that
Support(e) ⊂ (1 + α)[−][1]Support(H[′′]e). (13)
A slight modication of Algorithm 1 of [GMRZ13] yields
therefore e:
1) compute the syndrome s[⊺] := H[(3)]c[⊺] and then V :=
(1 + α)[−][1]Support(s);
2) The space V contains Support(e), so we can recover e = (e1, . . ., eN ) by solving the linear equation
H[(3)]e[⊺] = s[⊺] with the additional constraints ei ∈ V for
i ∈{1, . . ., N }. This is done by using the technique
given in Subsection V-C.
Note that we can also skip step 2 and directly look for s0
and s1 in the space V of dimension L +1 instead of decoding
exactly the value of e. In fact, this is what is specified in the
decapsulation of EDON-K.
VII. RECOVERING THE SHARED SECRET
Once we have recovered the error vector e ∈ F[N]q[m] [, we need]
to recover s0 and s1 to obtain the value of SharedSecret.
We know that the elements of e were picked randomly in
Ve = Support(˜e).We proceed just like in the decapsulation
algorithm.
We generate Support(e) which is equal to Ve with high
probability. More exactly, the probability that Support(e) is of
dimension < L is � LL−1 �N . For the parameters of edonk128ref
this probability is 2[−][37]. In such a case, the attack might fail,
but the decapsulation would fail too.
Then, among the 2[L] elements of Ve, we need to identify a
couple of consecutive elements of ˜e to deduce the secret. For
all pairs of candidates (λ, µ) ∈Ve × Ve, for 1 ≤ i ≤ [L]2 [−] [1]
we compute (s[′]0[, s][′]1[) :=][ H][i][ (][λ][||][µ][||H][ (][c][))][. If][ H][ (][s][′]1[||][s][′]0[||][c][) =][ h]
then we have (s[′]0[, s][′]1[) = (][s][0][, s][1][)][. Finally we recover the secret]
SharedSecret = H (s0||s1||c). In total this operation requires
O(L2[2][L]) operations, just like the decapsulation. This is the
reason why the value of L needs to remain small, otherwise
the decapsulation is not possible.
VIII. CONCLUDING REMARKS
A. Cost of the attack
Let us analyze the cost of the three steps of the attack
mentioned in Section IV.
Step 1 and 2 are polynomial in terms of the parameters
of the code. Step 1 only uses linear algebra operations and
has a complexity at most O(KmN [2]). The complexity of
step 2 is given by Theorem 1 of [GMRZ13] (using n =
N, k = N − R, r = L and d = 2), hence is equal to
L[2](16m + N [2]). The complexity of step 3 is O(L2[2][L]). This
is not polynomial in L but L is a very small parameter
(4 ≤ L ≤ 8 in the proposal). Moreover this third step is
the same as the decapsulation algorithm, so L needs to stay
small, otherwise the decapsulation would become too costly or
even impossible. So L can be considered as a constant ≤ 10
to allow a reasonable decapsulation. Hence the most costly
operation appears to be step 1.
B. Without compression of the public key
Our attack takes advantage of the compressed form of the
public key that allows a direct access to the value α = ab[−][1].
One could think that this is the origin of the attack, and decide
to express the public key in its uncompressed form to fix the
attack. As a consequence, the public key would be of size
K × N × m bits instead of K × N × ν bits in the compressed
form. In practice the public key for edonk128ref would be 16
times longer (around 288 kbits). This inflation of the key size
could be avoided by sending out a random basis of the space
Vg,c,d.
However, this is not enough. There is an even more direct
way to proceed, without the value of α. Instead of looking for
a matrix H[(3)] with entries liyng in ⟨1, α⟩F2, we can use the
following result.
Proposition 2. There exists a full rank (R − 1) × N binary
matrix H[(4)] that satisfies
GpubH[(4)][⊺] = 0K×(R−1).
Proof. Let T be a binary full-rank matrix (R − 1) × R matrix
that has rows of even Hamming weight. For instance we can
choose
1 1 0 - · · 0
0 1 1 0 ...
T = .
... ... ... ... ...
0 - · · 0 1 1
We observe now that TH has all its entries in {0, a + b}.
This follows directly from the fact that if we sum an even
number of elements in {a, b} we either get 0 (if the number
of a’s is even, and therefore also the number of b’s) or a+b (if
the number of a’s is odd). From this, it follows immediately
that
1
H[(4)] :=
a + b [TH]
-----
satisties the property. First, it is clear that this is a binary
matrix and we also have
1
GpubH[(4)][⊺] =
a + b [G][pub][H][⊺][T][⊺]
= 0K×(R−1).
Obtaining such a matrix H[(4)] is straightforward. We just
have to use the algorithm given in Section V to recover a
basis of dimension R − 1 of binary vectors x satisfying
Gpubx[⊺] = 0K.
We then use this matrix H[(4)] to compute the syndrome
s = H[(4)]c[⊺]. Since H[(4)]c[⊺] = H[(4)]e[⊺] we directly obtain with
very high probability that
Support(e) = Support(H[(4)]c[⊺]).
This reveals the support of the error and from there we can
go directly to the last step of the attack to reconstruct the
shared secret.
C. Security of the scheme
Considering the attack that we described, there is a way to
recover the secret of the edonk128ref scheme from a public
key without the private key in polynomial time. In practice, the
attack implemented with Sage on a personal computer recovers
the secret in less than a minute, so the scheme is far from
achieving the 128-bits security claimed in [GG17]. Hence this
scheme is insecure for the intended use. Moreover, the cost of
this attack is polynomial in terms of the parameters, so there
is no proper way to increase the parameters to achieve the
intended security level while keeping a reasonably small key
size.
REFERENCES
[BBD09] Daniel J. Bernstein, Johannes Buchmann, and Erik Dahmen,
editors. Post-Quantum Cryptography. Springer-Verlag, 2009.
[BMvT78] Elwyn Berlekamp, Robert McEliece, and Henk van Tilborg. On
the inherent intractability of certain coding problems. IEEE
Trans. Inform. Theory, 24(3):384–386, May 1978.
[Del78] Philippe Delsarte. Bilinear forms over a finite field, with
applications to coding theory. J. Comb. Theory, Ser. A, 25(3):226–
241, 1978.
[Gab85] Ernest Mukhamedovich Gabidulin. Theory of codes with maximum rank distance. Problemy Peredachi Informatsii, 21(1):3–16,
1985.
[GG17] Danilo Gligoroski and Kristian Gjøsteen. Edon-k. first round submission to the NIST post-quantum cryptography call, November
2017.
[GMRZ13] Philippe Gaborit, Gaétan Murat, Olivier Ruatta, and Gilles Zémor. Low rank parity check codes and their application to
cryptography. In Proceedings of the Workshop on Coding and
Cryptography WCC’2013, Bergen, Norway, 2013. Available on
www.selmer.uib.no/WCC2013/pdfs/Gaborit.pdf.
[GRS16] Philippe Gaborit, Olivier Ruatta, and Julien Schrek. On the
complexity of the rank syndrome decoding problem. IEEE Trans.
Information Theory, 62(2):1006–1019, 2016.
[Hua51] Loo-Keng Hua. A theorem on matrices over a sfield and its
applications. J. Chinese Math. Soc., 1(2):109–163, 1951.
[McE78] Robert J. McEliece. A Public-Key System Based on Algebraic
Coding Theory, pages 114–116. Jet Propulsion Lab, 1978. DSN
Progress Report 44.
[Sho94] Peter W. Shor. Algorithms for quantum computation: Discrete
logarithms and factoring. In S. Goldwasser, editor, FOCS, pages
124–134, 1994.
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/1802.06157, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "GREEN",
"url": "https://arxiv.org/pdf/1802.06157"
}
| 2,018
|
[
"JournalArticle"
] | true
| 2018-02-16T00:00:00
|
[
{
"paperId": "d791016d78b1054ce6c756a55ac78909ede25fdb",
"title": "Low Rank Parity Check codes and their application to cryptography"
},
{
"paperId": "684cfa5805f671ec1a24dc0b8e529259241693e4",
"title": "On the Complexity of the Rank Syndrome Decoding Problem"
},
{
"paperId": "99291ce0b97a31c786560241fea62604332afbf5",
"title": "Post-quantum cryptography"
},
{
"paperId": "2273d9829cdf7fc9d3be3cbecb961c7a6e4a34ea",
"title": "Algorithms for quantum computation: discrete logarithms and factoring"
},
{
"paperId": "5ea62d87fedfe0374fca3d5852c820e49d98ed7e",
"title": "Bilinear Forms over a Finite Field, with Applications to Coding Theory"
},
{
"paperId": "5e29000d24d5ded11e7a32216a91bdadaa9877f1",
"title": "On the inherent intractability of certain coding problems (Corresp.)"
},
{
"paperId": null,
"title": "Attack on the Edon-K Key Encapsulation Mechanism"
},
{
"paperId": null,
"title": "Edon-k. first round sub-mission to the NIST post-quantum cryptography call"
},
{
"paperId": null,
"title": "Theory of codes with maximum rank distance"
},
{
"paperId": null,
"title": "A Public-Key System Based on Algebraic Coding Theory, pages 114–116"
},
{
"paperId": "50508720c488f87598f832874906ea510d42f7da",
"title": "A THEOREM ON MATRICES OVER A SFIELD AND ITS APPLICATIONS"
},
{
"paperId": null,
"title": "constructing and solving a linear system of equations to find a parity-check matrix for the code C ′"
},
{
"paperId": null,
"title": "decoding the ciphertext using a slight variation of Algorithm 1 of [GMRZ13] (see Section VI)"
}
] | 8,887
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/0294a22ca84a83792d5deedf651d2ddfca7d3a79
|
[
"Computer Science"
] | 0.88685
|
Optimized information discovery using self-adapting indices over Distributed Hash Tables
|
0294a22ca84a83792d5deedf651d2ddfca7d3a79
|
IEEE International Performance, Computing, and Communications Conference
|
[
{
"authorId": "153347544",
"name": "F. Memon"
},
{
"authorId": "1770669",
"name": "Daniel Tiebler"
},
{
"authorId": "145046960",
"name": "Frank Dürr"
},
{
"authorId": "1700118",
"name": "K. Rothermel"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"International Phoenix Conference on Computers and Communications",
"IPCCC",
"Int Phoenix Conf Comput Commun",
"International Performance, Computing, and Communications Conference",
"Int Perform Comput Commun Conf",
"IEEE Int Perform Comput Commun Conf"
],
"alternate_urls": null,
"id": "8f125553-3fd5-4370-a175-f7179db29048",
"issn": null,
"name": "IEEE International Performance, Computing, and Communications Conference",
"type": "conference",
"url": "http://www.ipccc.org/"
}
| null |
# Optimized Information Discovery using Self-adapting Indices over Distributed Hash Tables
## Faraz Memon, Daniel Tiebler, Frank D¨urr, Kurt Rothermel IPVS – Distributed Systems Department, Universit¨at Stuttgart Universit¨atsstraße 38, 70569 Stuttgart, Germany Email: faraz.memon, tiebledl, frank.duerr, kurt.rothermel @ipvs.uni-stuttgart.de { }
## Abstract
_Distributed Hash Table (DHT)-based peer-to-peer in-_
_formation discovery systems have emerged as highly scal-_
_able systems for information storage and discovery in mas-_
_sively distributed networks._ _Originally DHTs supported_
_only point queries. However, recently they have been ex-_
_tended to support more complex queries, such as multi-_
_attribute range (MAR) queries. Generally, the support for_
_MAR queries over DHTs has been provided either by cre-_
_ating an individual index for each data attribute or by cre-_
_ating a single index using the combination of all data at-_
_tributes. In contrast to these approaches, we propose to_
_create and modify indices using the attribute combinations_
_that dynamically appear in MAR queries in the system._
_In this paper, we present an adaptive information dis-_
_covery system that adapts the set of indices according to the_
_dynamic set of MAR queries in the system. The main con-_
_tribution of this paper is a four-phase index adaptation pro-_
_cess. Our evaluations show that the adaptive information_
_discovery system continuously optimizes the overall system_
_performance for MAR queries. Moreover, compared to a_
_non-adaptive system, our system achieves several orders of_
_magnitude improved performance._
## 1. Introduction
During the past decade, DHTs have led the way for
distributed, scalable and fault-tolerant information discovery systems. DHTs have been extended from their original form, where they supported only point queries, to meet
modern application demand of supporting multi-attribute
range (MAR) queries. Queries such as, “find all comput_ers with RAM from 2 to 6 GB and CPU speed from 1.0 to_
_4.0 GHz” or “find all restaurants open from 10 to 11 PM_
_and with seating capacity for 8 to 10 people”, are typical_
examples of MAR queries.
DHTs have been extended using three different indexing
approaches to provide the support for MAR queries. The
first approach maps the value ranges of individual data attributes to a network of peers [4, 5, 19, 21]. A MAR query is
resolved by dividing the query into multiple single-attribute
range queries and then by joining the results at the query
initiator. The second approach indexes the combination of
all data attributes [6, 10, 18]. Data attributes that are not
included in a MAR query are considered to be wild-cards
in this approach. The third type of approach, employed by
our Optimized Information Discovery (OID) system [16],
indexes several attribute combinations with each combination different from the other. A MAR query is resolved by
selecting the most efficient index for performing the query.
Although the third type of indexing approach outperforms the other two approaches in terms of individual query
efficiency [16], the overall system performance still depends on the attribute combinations used for defining each
index. The efficiency of the overall system increases with
increasing number of queries being able to find a closely
matching index, in terms of the used attribute combination.
In [15], we presented a tool that assists the designer of
a distributed application in defining a useful set of indices
for the third type of DHT indexing approach. Given a limit
for the maximum number of indices and a representative set
of MAR queries (workload), our tool recommends a set of
indices that produces close-to-optimal system performance
for the workload within the given limit.
The index recommendation tool is an offline tool, i.e., it
is assumed that the workload provided to the tool is somehow collected from an already existing DHT-based information discovery system. Further, it is assumed that the
recommended set of indices is manually installed over the
DHT by the designer of the distributed application. In order
to carry out such an installation, the information discovery
system would have to be taken offline, which is highly undesirable for large-scale peer-to-peer (P2P) systems. In this
paper, we relax these assumptions to present an adaptive
OID system. The adaptive OID system performs the task of
## Published in Proceedings of 29th International Performance C i d C i i C f ( CCC 10) 1 9
-----
index recommendation and index installation online, eliminating the need for manually updating the set of indices.
The main contribution of this paper is the index adaptation process. The index adaptation process in a DHT, including online index recommendation and index installation, is carried out in four phases. During the first phase,
a workload of MAR queries is collected from several peers
in the network using uniform random sampling. The second phase involves execution of the index recommendation
tool to determine an optimal set of indices for the collected
workload. During the third phase, the cost and the benefit
of installing the recommended set of indices is calculated.
If it is beneficial to install the recommended set of indices,
the installation is carried out during the fourth phase.
The rest of the paper is organized as follows: in Section
2 we give an overview of the related work, the architecture
of the adaptive OID system is discussed in Section 3, in
Section 4 we describe the index adaptation process in detail,
evaluation results are presented in Section 5, and finally we
conclude the paper with an overview of our future work in
Section 6.
## 2. Related Work
A number of adaptive P2P information discovery systems have been proposed in the past. In this section, we
discuss some of them in relation to our system.
### 2.1. Unstructured P2P Systems
Several unstructured P2P information discovery systems
have been suggested that improve the efficiency of future
queries based on the past query workload in the system
[3, 13, 14, 17]. The major difference between these systems
and the structured P2P systems such as ours is that, each
peer in these systems tries to optimize the performance of
queries individually by modifying local data index. This
does not necessarily lead to the optimization of overall system performance. Moreover, given a query, these systems
perform only a best-effort search in the network, i.e., not all
matching data objects are always retrieved.
### 2.2. Structured P2P Systems
In order to improve the search efficiency of queries in
structured P2P information discovery systems, several DHT
extensions have been proposed [7, 8, 20].
Deng et al. [7] introduce learning-aware blind search for
range queries in DHTs. Each peer in their system stores information about previously retrieved results from each link
of the DHT using a local index structure. Queries are forward to regions of the DHT that had previously returned the
highest number of results. Unlike our system, their system
performs only best-effort search since each peer tries to optimize the query performance individually.
Skobeltsyn et al. [20] present a system that stores the
results of frequently issued queries at certain peers in the
DHT. The choice of queries whose results are cached is
based on the dynamic workload of queries in the system.
A query is resolved first by looking up the results in the local cache. If no results are found, the peer tries to find a
neighboring cache with results. If still no results are found,
the query is sent to all peer using broadcast. In our system,
queries are never resolved using broadcast since it is highly
unscalable to resolve queries in such manner. Instead, we
optimize indices for efficient query processing.
The HiPPIS system [8] indexes the data in a DHT using hierarchical indices. Each peer in the system logs each
query that it issues. If the granularity of a queried attribute
changes locally at a peer, e.g., more queries contain “city”
attribute instead of the “state” attribute, the peer checks if
the index has to be adapted accordingly. The peer performs
the adaptation check by asking every peer in the system for
the query statistics on the attribute using flooding. If adaptation is needed, the peer locks all the peers in the system
by flooding a lock message. During this period, queries are
answered also using flooding. Finally, the adaptation message is sent to all peers in the system using flooding as well.
Unlike the HiPPIS system, our system has a flooding-free
scalable index adaptation process.
## 3. System Architecture
The adaptive OID system has a layered architecture (see
Fig. 1(a)). The top layer consists of distributed applications
that require support for MAR queries. The bottom layer is
the DHT layer that provides the service for looking up a
key, broadcasting a message, and aggregating a value. The
middle layer, known as the OID framework layer, consists
of four major components: data index space, data placement
controller, query engine, and adaptation engine.
The data index space of OID framework layer consists
of several indices. The data placement controller uses these
indices to route each data object to the peer responsible for
hosting it. The query engine is responsible for distributed
query resolution, while the adaptation engine participates
in the index adaptation process.
Each data index in the OID framework layer is a Hilbert
Space-filling curve (SFC) [11]. Due to locality preserving
properties of Hilbert SFC, data objects that are close in a
multi-dimensional attribute space tend to be mapped to sets
of neighboring peers in a DHT. This enables efficient processing of MAR queries. A Hilbert SFC is defined as:
**Definition 1 A continuous function h : (a1, a2, . . ., ad) �→**
_x ∈_ N, where (a1, a2, . . ., ad) is a point in a d-dimensional
_Euclidean space and N is the set of natural numbers._
-----
|0|Col2|3|
|---|---|---|
|1|||
Distributed Applications
(b)
CPU Speed
2560
2048
Adaptation Engine
Index
Recommendation
Workload Tool
Data Placement Query Engine
Controller
Data Index Space
SFC1 SFC2 SFC3 … SFCo
1536
cess are: distributed workload collection, index recommendation, adaptation decision, and index installation.
We define following three types of peer roles to carry out
the index adaptation process:
**Adaptation Peer – An adaptation peer is a peer that period-**
ically initiates the index adaptation process. The length of
a period is set by the designer of the distributed application.
In order to avoid conflicting index updates, there can only
be a single adaptation peer at a time in the network. We assume that the location of the adaptation peer is pre-selected
by the designer of the distributed application. This could
be done by deciding that the peer that is the successors of a
certain key would be the adaptation peer in the network.
If a new peer joins at the location of the adaptation peer,
the state of the adaptation peer is transferred to it, making
it the new adaptation peer. Moreover, if the adaptation peer
fails during the first three phases of the adaptation process,
the process is restarted by the new adaptation peer. We assume a correctly functioning DHT where any peer that fails
or leaves the network is automatically replaced.
**Monitoring Peer – Each peer in our system is a monitoring**
peer. Monitoring peers are involved in the local collection
of the query workload, i.e., each monitoring peer logs each
query that it resolves. This log is emptied when a new set
of indices is installed in the data index space of the peer.
**Sampling Peer – A sampling peer is a peer that is involved**
in distributed workload collection discussed in the next section. Any peer can take the role of a sampling peer.
### 4.1. Distributed Workload Collection
Distributed Hash Table
(a)
1024
5 6 9 10
512
1.0 1.5 2.0 2.5 3.0
(c)
|0|Col2|1|Col4|14|Col6|15|Col8|
|---|---|---|---|---|---|---|---|
|||||||||
|3||2||13||12||
|4||7||8|||11|
|||||||||
|5||||9||||
**Figure 1. System Architecture**
A Hilbert SFC divides a d-dimensional euclidean space
into 2[k][·][d] cubes, called zones. A line then passes through
all zones defining an order among them. The result is a k[th]
order SFC, where k, known as the approximation level, defines the granularity of the space sub-division. Figure 1(b)(c) show a 2[nd] and a 3[rd] order Hilbert SFC respectively.
A data object in our system is indexed using each SFC
defined in the data index space of the OID framework layer.
If the SFC shown in Fig. 1(c) is one such index, then a data
object defined as (CPU Speed = 2.7 GHz, Mem Size =
1792 MB) would receive an identifier 12 from this index.
After a data object receives an identifier from each SFCbased index, a copy of the data object is routed to the DHT
peers responsible for the object identifiers. For a detailed
description of the data indexing process, see [16].
A MAR query is resolved in two steps. First, the query
is mapped to each SFC defined in the data index space. For
example, a MAR query defined as “(CPU Speed >= 1.3
_GHz)_ (CPU Speed <= 2.3 GHz) (Mem Size >=
_∧_ _∧_
640 MB) (Mem Size <= 2304 MB)” can be mapped
_∧_
to 11 zones on the SFC shown in Fig. 1(c). In the second
step, the query is routed to the peers responsible for the zone
identifiers obtained using the least expensive index [16].
## 4. Index Adaptation
The goal of the index adaptation process is to update the
set of indices in the OID framework layer of each peer according to the dynamic workload of MAR queries in the
system. In order to achieve this goal, we introduce a fourphase index adaptation process that is periodically executed
in the system. The four phases of the index adaptation pro
Ideally, if the complete set of past queries were collected
from all peers in the network, an optimal set of indices could
be obtained. However, collecting queries from all peers is
neither efficient nor scalable. Therefore, the goal of distributed workload collection is to collect a subset of the
complete set of queries by sampling some random peers.
The idea is to sample a sufficiently large subset of peers at
different locations in the network to get an approximation
of the complete set of queries.
The adaptation peer could directly collect a workload of
MAR queries by randomly sampling some monitoring peers
in the network. However, in this case, the adaptation peer
will have to issue a large number of sampling requests and
handle a large number of sampling responses, making the
sampling process unscalable. Therefore, in order to limit
the fanout of the adaptation peer and to make the sampling
process scalable, we use a two-level sampling process.
The adaptation peer initiates the first level of the sampling process by generating β random keys from the identifier space of the DHT, i.e., [0, 2[m]), where m is the number of identifier bits. A DHT lookup is then performed for
each random key in order to identify the peer responsible
-----
for it. Here, we assume a basic DHT lookup functionality
that, given a key, returns the identity of the peer responsible
for the key. Once the identity of a random peers is learned,
a sampling request with parameter γ is sent to it, where γ
indicates the number of peers to be sampled at the second
level of the sampling process.
Upon receiving a sampling request from the adaptation
peer, a peer assumes the role of a sampling peer. The sampling peer then forwards the sampling request to γ random
monitoring peers in the same manner as the adaptation peer.
After receiving a sampling request from a sampling peer, a
monitoring peer responds with the local query workload.
A sampling peer accumulates all the workloads received
from γ random monitoring peers into a single workload.
Since the same query could have been resolved by several
monitoring peers, it could appear multiple times in the accumulated workload. Therefore, duplicates are eliminated
during the accumulation process. Note that the same query
issued twice is not considered as a duplicate query since
each query has a globally unique identifier. Finally, the
accumulated workload including the workload of the sampling peer is sent to the adaptation peer where the accumulation process is repeated.
In order to detect the failures of the monitoring or the
sampling peers, the process of distributed workload collection includes timeouts at each level. At the level of a sampling peer, if a response is not received from a monitoring
peer before the timeout, the sampling request is re-issued
assuming that the faulty monitoring peer has been replaced
by the DHT. Similarly, at the level of the adaptation peer, if
a response is not received from a sampling peer before the
timeout, the sampling request is re-issued.
The distributed workload collection phase requires
_O ((β · γ) · (log2N + 2)) messages in the worst-case to_
collect a workload of MAR queries. N is the total number
of peers in the network and log2N is the maximum number
of messages required for a DHT lookup. Two additional
messages are needed to send a sampling request to a peer
and receive a sampling response from it.
### 4.2. Index Recommendation
Once a workload of MAR queries has been collected at
the adaptation peer, the next step in the adaptation process
is to search for an optimal set of indices for the collected
workload. For this purpose, we utilized the index recommendation tool previously introduced by us.
Given a workload of MAR queries and a limit o for the
maximum number of indices, the index recommendation
tool recommends a close-to-optimal set of indices Ir for
the given workload. For a detailed description of the index
recommendation tool and the index recommendation algorithms, see [15].
ti-j ti-3 ti-2 ti-1 ti ti+1 ti+2 ti+3 ti+j
Ti, i-j Ti, i+j
**Figure 2. Adaptation Decision**
### 4.3. Adaptation Decision
After obtaining a recommended set of indices from the
index recommendation tool, a na¨ıve approach would be to
directly install this set of indices in the network. However,
it is possible that the cost of installing the recommended set
of indices outweighs the benefit of installing it. Therefore,
the goal of the adaptation decision phase is to determine
whether the installation of the recommended set of indices
is beneficial or not. This is done by comparing the estimated
cost of the workload over the current set of indices with the
estimated cost of the workload over the recommended set
of indices. The installation cost of the recommended set of
indices is also taken into account.
Let ti mark the current periodic execution of the index
adaptation process, Ic be the current set of indices, and Ir
be the recommended set of indices. Then, we define the following quantities in our system (see Fig. 2):
**Ti,i−j – Time interval between ti and ti−j ∀j ∈** N[+] where,
_ti−j marks the index adaptation process where Ic was in-_
stalled. Note that this time interval is dynamic since a new
set of indices is not installed during each periodic execution
of the index adaptation process.
**Wi−j – Complete set of MAR queries during the time in-**
terval Ti,i−j.
**SWi−j – Sampled workload, from the complete set of**
MAR queries during the time interval Ti,i−j.
**costin – Estimated cost of installing the recommended set**
of indices Ir.
The adaptation peer considers the installation of the
recommended set Ir beneficial, if the following condition
holds:
_cost(SWi−j, Ic) > cost(SWi−j, Ir) + costin_ (1)
i.e., if the cost of the sampled workload SWi−j over the
current set of indices Ic is greater than the cost of the same
workload over the recommended set of indices Ir plus the
installation cost of the recommended set of indices. The
assumption behind Condition 1 is that the complete set of
MAR queries Wi−j would be repeated for a similar interval
of time in the future, i.e., for Ti,i+j. This is the most general assumption for predicting the cost of future queries. If
|Col1|Col2|Col3|Col4|Col5|Col6|
|---|---|---|---|---|---|
|W i-j||||||
|||||||
|||t t t t t t t|||t|
-----
Condition 1 is satisfied, the next phase of index adaptation
is carried out. Otherwise, the index adaptation process is
halted until the next periodic execution.
The cost functions cost(SWi−j, Ic) and cost(SWi−j, Ir)
in Condition 1 can be generalized as a cost function
_cost(Q, I). If Q = {q1, q2, q3, . . ., ql} is a set of queries_
and I = {SFC1, SFC2, SFC3, . . ., SFCo} is a set of
indices, then cost(Q, I) is calculated as:
Note that Equations 3 and 4 require the knowledge of
global parameters such as N and λ, which are generally not
known to a peer in a DHT network. However, an estimate
for these parameters could be obtained by installing a reliable broadcast/aggregation tree in the network. The root
of this tree will be the adaptation peer in our system. Such
a broadcast/aggregation tree could be installed and maintained using the approaches discussed in [9] and [12].
### 4.4. Index Installation
Once it is determined that installing a recommended set
of indices is beneficial, the adaptation peer initiates the index installation phase. The goal of the index installation
phase is to broadcast the new set of indices Ir, and to reindex the data on each peer accordingly.
A na¨ıve way of carrying out the index installation phase
is to broadcast the recommended set of indices using the
DHT broadcast/aggregation tree, and let each peer re-index
the data according to the new set of indices. However,
queries issued in the system during the re-indexing of the
data may not be able to recall the matching data objects
completely. This could happen in cases where, e.g., a query
issued using the new set of indices searches for matching
data objects at a peer where the data has not been placed yet
using the new set of indices. In order to avoid this shortcoming, we introduce a 3-step index installation phase.
During the first step, a broadcast message containing
the new set of indices Ir is sent by the adaptation peer to
each peer in the system. For this purpose, the DHT broadcast/aggregation tree is used. Upon receiving the broadcast
message, each peer begins to re-index its data. Note that
the old set of indices Ic and the corresponding data is not
yet removed from the system. Hence, the queries that are
issued during this step continue to be resolved using Ic.
A data object in the OID system is indexed using o number of indices, i.e., |Ic| = o. This means that there are o
copies of the same data object in the system. Therefore, it
has to be made sure that each copy of the data object is not
re-indexed using each index in Ir. For example, if Ic and Ir
are as shown in Fig. 3, a data object indexed using Ic would
be located at four locations in the system. Now if the same
data object is re-indexed at each location using every index
in Ir, it would be sent four times to each new location in the
network. To avoid this, the data is re-indexed as follows.
Data re-indexing at a peer starts with the comparison of
the installed set of indices Ic with the new set of indices Ir
(see Fig. 3). First, the common elements in both sets are
ignored. Next, a mapping is defined from each element in
_Ic to each corresponding element in Ir. The data objects_
that had been previously indexed using an element of Ic
are now re-indexed only using the corresponding element
in Ir. For example, in Fig. 3, the data objects that had been
_|Q|_
�
_cost(Q, I) =_ _cost(qi, SFCj) such that_
_i=1_
_cost(qi, SFCj) < cost(qi, SFCk) where_
_j, k : 1_ (j, k) _I_ and j = k
_∀_ _≤_ _≤|_ _|_ _̸_
(2)
i.e., the cost of a set of queries Q over a set of indices I is a
sum of the cost of each query in Q over the least expensive
index in I.
In order to determine the least expensive index for a
query, the network cost of the query over each index needs
to be calculated. Due to highly dynamic nature of P2P systems, this cost cannot be accurately anticipated. However,
if the cost of routing a message in the network is known, the
maximum cost for resolving a query can be calculated.
Let z be the total number of zones a query maps to, on
an SFC-based index. In order to resolve this query, the peer
responsible for each zone has to be queried. If a basic query
routing strategy is considered, where first a lookup is performed to determine the peer responsible for each zone,
then the maximum cost of a query q on an index SFC is
calculated as:
_cost(q, SFC) = z · (log2N + 2) [messages]_ (3)
where N is the total number of peers in the network and
_log2N is the maximum number of messages needed for_
looking up a peer responsible for a zone. Two additional
messages are needed to send a query request to a peer and
receive a query response from it.
In order to check if Condition 1 holds, the cost of installing the recommended set of indices costin has to be
calculated. Similar to the cost calculation above, only the
maximum cost of installation can be calculated. Let λ be
the total number of unique data objects in the system, then
the maximum cost for installing the recommended set of
indices Ir is calculated as:
_costin = 3 · (N −_ 1) + (|Ir| −|Ic ∩ _Ir|)·_ (4)
_λ · (log2N + 2) [messages]_
where 3 (N 1) is the cost of broadcasting the recom_·_ _−_
mended set of indices Ir, (|Ir| −|Ic ∩ _Ir|) is the total num-_
ber of new indices, and λ · (log2N + 2) is the cost of reindexing the data. The reason for the broadcast cost being
almost three-times the network size is discussed in the next
section.
-----
Ic SFC1 SFC4 SFC7 SFC9
Ir SFC7 SFC6 SFC8 SFC1
**Figure 3. Data Re-Indexing**
indexed using SFC4 in Ic are only re-indexed using SFC6
in Ir. The re-indexing process involves a look-up for the
new location of a data object and then transfer of the data
object to this location.
Once the re-indexing of the data is finished at a peer, it
sends an acknowledgement to the parent node in the DHT
broadcast/aggregation tree. During the second step of the
index installation process, the acknowledgements from all
peers in the network are aggregated until the adaptation
peer receives the aggregated acknowledgement. Queries
still continue to be resolved using the old set of indices Ic.
Upon receiving an aggregated acknowledgement from
the child nodes in the DHT broadcast/aggregation tree, the
adaptation peer starts the third step of index installation by
broadcasting a use index message. When this message
is received at a peer, the peer removes Ic, discards the corresponding data, empties the monitored query log, and starts
using Ir for query resolution. Note that the data common
between Ic and Ir is not discarded. During this step of index installation, if a query is issued from a peer that has
not received the use index message yet, then there are
two possibilities. First, the query will be resolved using Ic,
if all peers involved in query resolution have not discarded
the data corresponding to Ic. Second, even if a single peer
involved in query resolution has discarded Ic, then the peer
that issued the query will be asked to re-issue it using Ir.
If the adaptation peer fails before the first step of the
index installation phase, then the process of index adaptation is repeated by the new adaptation peer. However, if the
adaptation peer fails after the first step of index installation,
the new adaptation peer is already aware of the state of index installation due to the broadcast of new indices in the
network. Therefore, the new adaptation peer executes the
next steps of the index installation phase.
## 5. System Evaluation
In this section, we present the results from the performance evaluation of the adaptive OID system. We simulated our system using the PeerSim [1] simulator. The simulations were performed on an AMD Opteron machine with
4 GB of RAM.
Considering resource discovery in grid computing as an
example scenario, we represent the data objects in our simulations as resource specifications. Each resource specification consists of attributes shown in Table 1. The value for
|SFC 1|SFC 4|Col3|SFC 7|SFC 9|
|---|---|---|---|---|
||||||
|SFC 7|SFC 6||SFC 8|SFC 1|
|Attribute|Value Domain|Definition|
|---|---|---|
|CPU Speed|1.0 – 4.0|CPU clock speed in gigahertz|
|Busy CPU|0 – 100|Percentage of CPU(s) in use|
|Mem Size|1.0 – 8.0|Total Memory size in gigabytes|
|Mem Used|0 – 100|Percentage of Memory in use|
|HDD Size|100.0 – 3000.0|Total HDD size in gigabytes|
|DL Bandwidth|0.5 – 100|Bandwidth of down link in mbits/sec|
**Table 1. Attribute List**
each attribute in a resource specification is randomly generated from the value domain of the attribute.
Unlike the database management systems where benchmark workloads are made available by the TPC [2], no such
workload of MAR queries is readily available for P2P systems. Hence, we generate the workloads using the attributes
in Table 1 for simulating different scenarios of our system.
For each point on the graphs displayed in this section,
the corresponding experiment is repeated 10 times with different workloads, and an average value is plotted.
### 5.1. Varying Number of Attributes
In this section, we present the results from the performance evaluation of our system using a workload of queries
with varying number of attributes. We show that an adaptive
OID system is essential for continuous optimization of overall system performance for MAR queries. Table 2 shows the
parameter values used for this simulation.
**Parameter** **Value** **Definition**
_N_ 1000 Total number of peers in the DHT
_n_ 1600 Total number of queries in the workload
_o_ 3 Maximum number of indices
_λ_ 5000 Total number of data objects
_β_ 33 First level sampling parameter
_γ_ 2 Second level sampling parameter
**Table 2. Simulation Parameters**
The workload is generated in a manner that the start of
the workload contains queries with 4 attributes followed by
queries with 3, 2, and 4 attributes again. To simulate a slow
change in the workload over time, the attributes in queries
are varied slowly, i.e., the change from queries with 4 attributes to queries with 3 attributes and so on, is not sudden.
Each attribute in a query is randomly selected from the
list shown in Table 1. Similarly, the range for an attribute
in a query is randomly selected from the domain of the attribute. The values for parameters β and γ are set so that
almost 10% of peers in the network are sampled.
We simulate the adaptive OID system, the non-adaptive
system, and a system with only a single adaptation (partially adaptive system), by executing the generated workload from random peers in the DHT over a period of time.
The non-adaptive system is a system with only a single data
index over all 6 attributes shown in Table 1. For the partially
|Parameter|Value|Definition|
|---|---|---|
|N|1000|Total number of peers in the DHT|
|n|1600|Total number of queries in the workload|
|o|3|Maximum number of indices|
|λ|5000|Total number of data objects|
|β|33|First level sampling parameter|
|γ|2|Second level sampling parameter|
-----
10[9]
10[8]
10[7]
10[6]
10[5]
10[4]
10[3]
10[2]
0 250 500 750 1000 1250 1500
Simulation Time
0 250 500 750 1000 1250 1500
Simulation Time
10[8]
10[7]
10[6]
10[5]
10[4]
10[3]
**Figure 4. Varying Number of Attributes**
adaptive system, the adaptation takes place after 10 simulation time units. Moreover, for the adaptive OID system, the
index adaptation process is scheduled to run after every 10
simulation time units. A single simulation time unit is long
enough to allow execution of a single query.
For every 5 simulation time units, we plot the average
number of messages in all three systems during that 5-timeunit-window (see Fig. 4). The number of messages represents all the messages in the system including messages for
the index adaptation process. For the adaptive OID system,
the peaks in the number of messages (see Fig. 4) mark the
points where index installation takes place. The higher the
peak, the larger the number of indices that are exchanged.
Similar to the non-adaptive system, the adaptive OID
system and the partially adaptive system start with one index over all attributes. However, the first adaptation happens very soon in both the systems and 2 additional indices
are installed (see Fig. 4). This improves the performance
of MAR queries in both systems because the queries are
able to find less expensive indices for resolution. Since the
first adaptation is based on a very small workload, the second adaptation follows soon in the adaptive OID system.
The system continues to adapt itself over time according to
the workload of queries. After each adaptation, the performance of MAR queries improves as the average number of
messages in the system are reduced.
Figure 4 shows that the partially adaptive system produces 99.2% less messages than the non-adaptive system.
Moreover, the adaptive OID system produces 83.6% less
messages than the partially adaptive system. Therefore, the
adaptive OID system is several orders of magnitude better
than the non-adaptive system. Figure 4 also shows that, in
order to optimize the overall system performance for MAR
queries, a system with continuous adaptations is essential.
The performance of the non-adaptive system worsens
with decreasing number of attributes in queries (see Fig. 4).
This happens because with decreasing number of query attributes, more attributes have to be considered as wild-cards
on a single large index. The performance of the system gets
better towards the end of the simulation because the number
**Figure 5. Fixed Number of Attributes**
of attributes in queries increases from 2 to 4 attributes.
In order to further analyze the impact of the number of
attributes in queries, we perform another simulation where
the number of attributes in the workload is kept constant
to 3 attributes. Other simulation parameters have the same
values as in Table 2. Figure 5 shows the performance of all
three systems with respect to the average number of message in a 5-time-unit-window.
Figure 5 shows that the adaptive OID system quickly
adapts its indices to the changing workload of queries. Major adaptations come close to the start of the simulation.
After that, even though some small adaptations happen in
the system, the performance of the system remains roughly
constant. This happens because the indices adapted during
the start of the simulation remain beneficial for the complete
simulation. The performance of the non-adaptive system
remains almost constant, and several orders of magnitude
worse than the adaptive system, throughout the simulation.
With a constant number of attributes in queries, the performance of the partially adaptive system comes close to the
performance of the adaptive system (see Fig. 5). However,
the adaptive system still produces 3.1% less messages compared to the partially adaptive system. This difference in the
number of messages grows larger over time. Therefore, in
a long running system, the adaptive system would perform
significantly better than a partially adaptive system.
### 5.2. Varying Number of Indices
In this section, we present the performance evaluation of
the adaptive OID system by showing the impact of varying
number of indices on the system. We perform 3 different
simulations using the same workload as in the first simulation discussed in Sec. 5.1. The maximum number of indices
_o is varied from 3 to 5 across these simulations. Other sim-_
ulation parameters have the same values as in Table 2. For
each simulation we plot the average number of messages in
a 10-time-unit-window.
Generally, the larger the set of indices, the better the performance of the system after an adaptation (see Fig. 6), be
-----
5k 10k 20k 40k 80k 160k
Num. of Data Objects
10[5]
10[4]
10[3]
10[2]
0 250 500 750 1000 1250 1500
Simulation Time
800
700
600
500
400
300
200
**Figure 6. Varying Number of Indices**
cause with increasing number of indices, more queries find
an optimal index for resolution. Since more queries are
optimized, the overall system performance also improves
slightly with increasing number of indices, e.g., the system
with 4 indices produces 2.3% less messages than the system with 3 indices. Similarly, the system with 5 indices
produces 1% less messages than the system with 4 indices.
### 5.3. Varying Number of Data Objects
In this section, we present the performance evaluation of
the adaptive OID system by showing the impact of varying
number of data objects on the system. We perform 6 different simulations using the same workload as in the first
simulation discussed in Sec. 5.1. The total number of data
objects in the system λ is doubled across the simulations,
starting from 5000 and going up to 160,000. Other simulation parameters have the same values as in Table 2. For
each simulation we plot the average adaptation window size
defined as: average number of simulation time units needed
for an adaptation to happen in the system.
Figure 7 shows the performance of the adaptive OID system with respect to the average adaptation window size.
The larger the number of data objects in the system, the
longer it takes for an adaptation to happen. The reason is
that with increasing number of data objects, the index installation cost also increases. Hence, a larger and more diverse
workload is needed for the adaptation to be beneficial.
### 5.4. Distributed Workload Collection
In this section, we discuss the results from the performance evaluation of the distributed workload collection (see
4.1) phase of the index adaptation process. We perform 16
simulations using the same workload as in the first simulation discussed in Sec. 5.1. For a fixed DHT network size,
we vary the values of β and γ across 4 simulations, such
that the total number of peers sampled in the network vary
between 6% and 12% (in steps of 2%) of the total network
size. This simulation scenario is repeated for varying DHT
**Figure 7. Varying Number of Data Objects**
network sizes of N = (10[2], 10[3], 10[4], 10[5]). Other simulation parameters have the same values as in Table 2.
During each simulation, after a distributed workload collection phase ends, we measure the cost deviation metric
defined as:
� _|cost(W, IrSW_ ) − _cost(W, Ir[W]_ [)][|] �
100
_∗_
_cost(W, Ir[W]_ [)]
where W is the complete set of MAR queries from all peers,
_Ir[SW]_ is the recommended set of indices obtained using the
sampled workload SW, and Ir[W] is the recommended set of
indices obtained using the complete set of queries W .
The cost deviation indicates how good the recommended
set of indices is (in percentage), if it is obtained using the
sampled workload, compared to the recommended set of
indices obtained using the global workload. The lower the
cost deviation, the better the performance of the system because the indices are more optimized for future queries.
For each simulation, the average cost deviation is plotted
in Fig. 8. For the network size of 10[2], the calculation for the
number of peers to sample using β and γ, was rounded-off
to the same value (7% of the network size) in case of 6%
and 8% sampled peers.
Figure 8 shows that for a fixed network size, the larger
the number of sampled peers, the smaller is the cost deviation. This happens because with increasing number of sampled peers, a better approximation of the complete set of
queries is acquired. Hence, the recommended set of indices
obtained using the sampled workload is more similar to the
recommended set of indices obtained using the complete
set of queries. Figure 8 also portrays that with increasing
network size, sampling a smaller percentage of peers in the
network is sufficient for having a low cost deviation.
## 6. Conclusion and Future Work
In this paper, we presented the design and evaluation of
the adaptive OID system. The adaptive OID system optimizes the overall system performance for MAR queries by
dynamically adapting the set of indices in a DHT. The set of
-----
16
14
12
10
2[468]
10[2]
10[3]
10[4]
10[5] 6 [ 7 8 9 10 11 12 13]
**Figure 8. Distributed Workload Collection**
indices is adapted using a four-phase index adaptation process. During the first phase, a workload of MAR queries
is collected from the DHT network using uniform random
sampling of peers. This workload is then used in the second phase for obtaining a new set of indices using the index recommendation tool [15]. During the third phase the
cost and the benefit of installing a new set of indices is estimated. If it is beneficial to install the new set of indices, the
installation is carried out during the fourth phase of index
adaptation process.
Our evaluations show that the adaptive OID system continuously adapts the set of indices in the system according to
the dynamic workload of MAR queries. The adaptations are
most useful when there is a variety of different queries in the
system. Nonetheless, the adaptive OID system shows several orders of magnitude improved performance compared
to a non-adaptive system.
Currently, the complete log of MAR queries is retrieved
from a peer during the distributed workload collection
phase. In future, we plan to change this phase so that it
is possible to retrieve the query log until a specified point
in time in the past. This would limit the amount of network
information flow during the sampling process, making the
distributed workload collection phase more scalable.
## References
[1] PeerSim: A P2P Simulator. http://peersim.sourceforge.net/.
[2] Transaction Processing Performance Council .
http://www.tpc.org/.
[3] W. Acosta and S. Chandra. Exploiting the Properties of
Query Workload and File Name Distributions to Improve
P2P Synopsis-based Searches. In Proc. of Intl. Conf. on
_Computer Communications. IEEE, 2008._
[4] A. Andrzejak and Z. Xu. Scalable, Efficient Range Queries
for Grid Information Services. In Proc. of Intl. Conf. on P2P
_Computing. IEEE, 2002._
[5] M. Cai, M. Frank, J. Chen, and P. Szekely. MAAN: A
Multi-Attribute Addressable Network for Grid Information
Services. In Proc. of Intl. Workshop on Grid Computing.
IEEE, 2003.
[6] Y. Chawathe, S. Ramabhadran, S. Ratnasamy, A. LaMarca,
S. Shenker, and J. Hellerstein. A Case Study in Building
Layered DHT Applications. In Proc. of Conf. on Applica_tions, Technologies, Architectures, and Protocols for Com-_
_puter Communications. ACM, 2005._
[7] Z. Deng, D. Feng, K. Zhou, Z. Shi, and C. Luo. Range
Query Using Learning-Aware RPS in DHT-Based Peer-toPeer Networks. In Proc. of Intl. Symp. on Cluster Computing
_and the Grid. IEEE, 2009._
[8] K. Doka, D. Tsoumakos, and N. Koziris. HiPPIS: An Online
P2P System for Efficient Lookups on d-dimensional Hierarchies. In Proc. of Workshop on Web Information and Data
_Management. ACM, 2008._
[9] S. El-Ansary, L. O. Alima, P. Brand, and S. Haridi. Efficient
Broadcast in Structured P2P Networks. In Peer-to-Peer Sys_tems II. Springer, 2003._
[10] P. Ganesan, B. Yang, and H. Garcia-Molina. One Torus to
Rule Them All: Multi-dimensional Queries in P2P Systems.
In Proc. of Intl. Workshop on the Web and Databases. ACM,
2004.
[11] D. Hilbert. Uber die stetige Abbildung einer Linie auf ein[¨]
Fl¨achenst¨uck. In Mathematische Annalen, 1891.
[12] K. Huang and D. Zhang. DHT-based Lightweight Broadcast Algorithms in Large-scale Computing Infrastructures.
_Future Gener. Comput. Syst., 2010._
[13] V. Kalogeraki, D. Gunopulos, and D. Zeinalipour-Yazti. A
Local Search Mechanism for Peer-to-Peer Networks. In
_Proc. of Conf. on Information and Knowledge Management._
ACM, 2002.
[14] G. Koloniari, Y. Petrakis, E. Pitoura, and T. Tsotsos. Query
Workload-aware Overlay Construction using Histograms. In
_Proc. of Intl. Conf. on Information and Knowledge Manage-_
_ment. ACM, 2005._
[15] F. Memon, F. D¨urr, and K. Rothermel. Index Recommendation Tool for Optimized Information Discovery Over Distributed Hash Tables. In Proc. of Intl. Conf. on Local Com_puter Networks. IEEE, 2010._
[16] F. Memon, D. Tiebler, F. D¨urr, K. Rothermel, M. Tomsu, and
P. Domschitz. OID: Optimized Information Discovery using
Space Filling Curves in P2P Overlay Networks. In Proc. of
_Intl. Conference on Parallel and Distributed Systems. IEEE,_
2008.
[17] L. T. Nguyen, W. G. Yee, and O. Frieder. Query Workload
Driven Summarization for P2P Query Routing. In Proc. of
_Intl. Conf. on Peer-to-Peer Computing. IEEE, 2008._
[18] C. Schmidt and M. Parashar. Flexible Information Discovery in Decentralized Distributed Systems. In Proc. of Intl.
_Symp. on High Performance Distributed Computing. IEEE,_
2003.
[19] Y. Shu, B. C. Ooi, K.-L. Tan, and A. Zhou. Supporting
Multi-dimensional Range Queries in Peer-to-Peer Systems.
In Proc. of Intl. Conf. on P2P Computing. IEEE, 2005.
[20] G. Skobeltsyn and K. Aberer. Distributed Cache Table: Efficient Query-driven Processing of Multi-term Queries in P2P
Networks. In Proc. of Intl. Workshop on Information Re_trieval in P2P Networks. ACM, 2006._
[21] P. Triantafillou and T. Pitoura. Towards a Unifying
Framework for Complex Query Processing over Structured
Peer-to-Peer Data Networks. In Proc. of Intl. Workshop
_on Databases, Information Systems and P2P Computing._
Springer, 2003.
-----
|
{
"disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references', 'abstract'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1109/PCCC.2010.5682330?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/PCCC.2010.5682330, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "GREEN",
"url": "https://www2.informatik.uni-stuttgart.de/bibliothek/ftp/ncstrl.ustuttgart_fi/INPROC-2010-116/INPROC-2010-116.pdf"
}
| 2,010
|
[
"JournalArticle",
"Conference"
] | true
| 2010-12-01T00:00:00
|
[] | 11,311
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/02987c79896dfedd7d494deaead3e78eac241f65
|
[
"Computer Science"
] | 0.89571
|
Unlinkable Collaborative Learning Transactions: Privacy-Awareness in Decentralized Approaches
|
02987c79896dfedd7d494deaead3e78eac241f65
|
IEEE Access
|
[
{
"authorId": "30682012",
"name": "S. Rahmadika"
},
{
"authorId": "1708489",
"name": "K. Rhee"
}
] |
{
"alternate_issns": null,
"alternate_names": null,
"alternate_urls": [
"http://ieeexplore.ieee.org/servlet/opac?punumber=6287639"
],
"id": "2633f5b2-c15c-49fe-80f5-07523e770c26",
"issn": "2169-3536",
"name": "IEEE Access",
"type": "journal",
"url": "http://www.ieee.org/publications_standards/publications/ieee_access.html"
}
|
Smart contracts (SCs) and collaborative learning (CL) are disclosed publicly, in which most transactions and activities that occur by the parties can be bared in real-time. Both are strengthened in a decentralized manner. CL allows numerous clients to collectively build deep learning models privately by aggregating the gradient values from clients’ devices, yet it lacks the incentive mechanism for the contributing clients. On the other hand, the merits of SCs can be a plausible solution as an incentive mechanism in the CL system because self-executing contracts with immutable data records are resistant to failure. The clients can claim the rewards by stating their contribution arbitrarily in the SCs and tendering a proof transaction function. Nevertheless, directly adopting SCs in the CL system could breach clients’ privacy because the transactions are exposed openly. The observer can infer the properties of the clients’ resources. Therefore, we designed schemes that can overcome observers’ ability to link clients’ information with their associated devices during training. In essence, our schemes are unbiased. We also provide a secure incentive mechanism for the parties in the CL system by obscuring the information values. Finally, the numerical results indicate that the proposed schemes satisfy the design goals.
|
Received December 30, 2020, accepted April 19, 2021, date of publication April 28, 2021, date of current version May 6, 2021.
_Digital Object Identifier 10.1109/ACCESS.2021.3076205_
# Unlinkable Collaborative Learning Transactions: Privacy-Awareness in Decentralized Approaches
SANDI RAHMADIKA 1 AND KYUNG-HYUNE RHEE 2, (Member, IEEE)
1Department of Information Security, Graduate School, Pukyong National University, Busan 48513, South Korea
2Department of IT Convergence and Application Engineering, Pukyong National University, Busan 48513, South Korea
Corresponding author: Kyung-Hyune Rhee (khrhee@pknu.ac.kr)
This work was supported in part by the National Research Foundation of Korea (NRF) grant funded by the Korea Government (MSIT)
under Grant NRF-2018R1D1A1B07048944, and in part by the Ministry of Science and ICT (MSIT), South Korea, under the Information
Technology Research Center (ITRC) support program, supervised by the Institute of Information and Communications Technology
Planning and Evaluation (IITP), under Grant IITP-2020-0-01797.
**ABSTRACT Smart contracts (SCs) and collaborative learning (CL) are disclosed publicly, in which most**
transactions and activities that occur by the parties can be bared in real-time. Both are strengthened in
a decentralized manner. CL allows numerous clients to collectively build deep learning models privately
by aggregating the gradient values from clients’ devices, yet it lacks the incentive mechanism for the
contributing clients. On the other hand, the merits of SCs can be a plausible solution as an incentive
mechanism in the CL system because self-executing contracts with immutable data records are resistant
to failure. The clients can claim the rewards by stating their contribution arbitrarily in the SCs and tendering
a proof transaction function. Nevertheless, directly adopting SCs in the CL system could breach clients’
privacy because the transactions are exposed openly. The observer can infer the properties of the clients’
resources. Therefore, we designed schemes that can overcome observers’ ability to link clients’ information
with their associated devices during training. In essence, our schemes are unbiased. We also provide a
secure incentive mechanism for the parties in the CL system by obscuring the information values. Finally,
the numerical results indicate that the proposed schemes satisfy the design goals.
**INDEX TERMS Blockchain, collaborative learning, decentralized approach, smart contracts, unlinkability.**
**I. INTRODUCTION**
The satisfactory applications of internet-based information
systems that rely on a dispersed manner have been extensively
researched by academia, developers, and industries. The foremost objective of the decentralized approach is to address
the communication bottleneck issues and memory usage of
the conventional centralized system [1]. The paradigm of a
centralized system for various major implementations also
shifted toward dispersed manners such as for financial applications, medical records, digital rights, and intellectual property, among others. Blockchain technology through Bitcoin
cryptocurrency [2] and federated learning [3] are the most
prominent practical adoptions of decentralized approaches.
The appearance of blockchain technology in 2008 that
represented a thoroughly peer-to-peer version of electronic
cash called Bitcoin is the trigger for further development of
a decentralization-based system. Blockchain 1.0 is the first
The associate editor coordinating the review of this manuscript and
approving it for publication was Chi-Yuan Chen .
generation [4] with simple ledgers that record transactions,
followed by generation 2.0 with smart contract (SC) features
pioneered by the Ethereum platform [5]. Blockchain 3.0 is
the latest version of the decentralized generation by combining more features that support cloud nodes, open-chain
access, and incentives for self-evolution. Regardless of the
blockchain merits, such as tamper-resistance, paperless,
immutability, append-only based data structure, blockchain
suffers from privacy issues [6] because the process is transparent and disclosed publicly. Even though there is a private version of the blockchain, the validators of consensus
are semi-trusted parties that can jeopardize the sustainability
of the system. Hence, blockchain-based service alone without several additional protocols is likely inappropriate for
application in systems that hold numerous types of sensitive
information.
Comparable to the Bitcoin blockchain, federated learning
(FL) also relies on the decentralized approach in building
the deep learning model collectively from multiple devices.
In contrast to conventional machine learning, where the
-----
clients process the training model centrally, FL allows the
clients to build the artificial intelligence (AI) model by sending the updated gradient values to the aggregation server
without revealing the dataset [7]. In this sense, private data
remain confidential (the FL preserves privacy for clients by
design). Nevertheless, FL-based schemes, such as collaborative learning, lack the proper incentive mechanism that can
motivate clients to improve AI models. Several applications
do not even provide a reward for clients. Blockchain with SC
features can be a solution to tackle the incentive mechanism
issues. However, directly adopting SCs threatens the clients’
privacy because it is transparent and openly available in the
network. Accordingly, SCs in CL will be a serious consideration if implemented in a system with profoundly confidential data. Specifically, complementary protocols need to be
implemented.
Privacy-awareness in the smart contract blockchain and
collaborative learning appears as part of a flaw that
must be overcome. For instance, in collaborative learning,
where the primary objective is to provide clients’ privacy,
Melis et al. [8] surprisingly argued that the observer could
infer the presence of exact data points of clients’ datasets
with certain assumptions. On the other hand, SCs make
transactions visible to the public. The stored data can be
accessed at any time, and the value of data managed by the
individuals is noticeable. Transparency is one of the concrete
features of the SC blockchain. However, this feature is not
desirable for various cases, especially in cross-silo FL with
highly confidential data such as medical records, biometric data, employee data, sexual orientation [9], philosophical beliefs, and so forth. For these reasons, the relationship
between the data used in training and the owner needs to be
obscured.
To support an unlinkable incentive mechanism in crosssilo FL, SCs combined with supplementary protocols can be
a credible solution to address privacy and linkability issues
for decentralized applications. For example, a well-known
decentralized cryptocurrency called Monero (stock symbol
XMR) provides several features to obfuscate the information
of every transaction. The core technology of Monero (XMR)
is based on the CryptoNote algorithm [10] with the elliptic
curve parameters, ring signatures, and stealth address as the
principal protocols. However, its application has been limited
to only the financial sector. Therefore, an unlinkable and
secure incentive scheme in CLs can be created by referring
to the CryptoNote protocol as the supplementary values in
the arbitrary functions of Ethereum SCs.
To evaluate the objective of this research, we implemented
a prototype of the federated learning scheme by utilizing
a convolutional neural network installed on devices fully
controlled by clients. Our system is built by referring to
FL’s principles. Clients who have contributed by using their
resources can tender reward claims transaction through SCs,
which will then be verified by the validators. If the claim is
verified successfully, the incentives are propagated through
the Ethereum SC platform. Every transaction that occurs is
obscured using several protocols (elaborated in Section IV) so
that the observer has no knowledge of the transaction values.
In summary, this research provides the main following
contributions:
(i) We present the cross-silo FL framework by referring to
the FL principles (developed by the Google AI team) as a
case study for an incentive scheme based on blockchain
SCs.
(ii) We introduce used-model-only services with the obscur_ing transactions feature. This use case is intended for_
clients who only want to use the model without expecting the incentives from model providers, e.g., due to an
insufficient amount of the dataset.
(iii) We introduce unlinkable rewarding and training
activities without revealing the information values of
transactions. It is part of the extended version of used_model-only services._
The road map of this paper is organized as follows.
Section II investigates the existing decentralized model
and techniques that leverage the FL-based approach and
blockchain technology. In Section III, we provide the essential information related to the CL models, conventional
incentives mechanism, concerns, and benefits from this
research. The model and operations of our proposed schemes
are presented in Section IV. In Section V, we outline the
fundamental points of this study. The opportunities and challenges are described in Section VI. Finally, we conclude this
paper in Section VII.
**FIGURE 1. General overview of the collaborative learning model.**
**II. RELATED WORK**
The growing interest in adopting blockchain-based incentive
mechanisms in various fields has produced several prominent
schemes. In this section, we focus on federated, collaborative, and decentralized learning terms from prior works. The
general model of the federated learning approach is depicted
in Figure 1. A joint optimization approach that combines the
party’s reputation and contract theory as the incentive mechanism in the FL system was introduced in [11]. The party’s
reputation is generated by calculating the multi-weight subjective logic model to motivate users to always participate
in training. Correspondingly, a reliable and accountable FL
-----
that relies on the blockchain is outlined in [12]. The proof-ofconcept is applied, and it is associated with practical methods
for secure aggregation of local model updates in the FL.
A system called DeepChain was proposed in [13] to preserve
privacy for clients during training in the FL. The incentive
is designed to rely on a smart contract. The system forces the
parties to behave honestly with a designed punishment policy.
In line with this, an auditable FL with trust and blockchainbased incentive, namely FLChain, has been discussed in [14]
which also proposed a protocol to reduce the time cost of
blockchain queries.
In short, the previously mentioned related works combine blockchain with an FL system for different objectives.
Blockchain is primarily utilized to propagate incentives to the
parties. Various methods to preserve privacy have also been
discussed. However, the linkability concerns within the system, especially in smart contracts, have not been thoroughly
reviewed in previous works. The uncertainty of leaking personal data during training needs to be taken into account.
Thus, we raise this topic to be discussed in this research.
Another comparative approach with different purposes is
described in [15]. Instead of using blockchain as an incentive scheme, Sharma et al. (2020) proposed a distributed
computing defense framework by utilizing blockchain merits
in the sustainable society fields. Recently, efforts to utilize
blockchain and FL were elaborated in [16] and [17]. In fog
and crowd computing environments, several troublesome
problems, such as network congestion, overhead, and communication delays are addressed. This limits the discussion of
the trade-off between privacy and efficiency of the proposed
schemes.
In terms of security from a federated learning perspective,
the research in [18] outlines that malicious users might perform poisoning attacks against the updated models targeting
specific devices in the FL network. With the same intentions, the authors in [19] presented Sybil-based poisoning
attacks in the FL and introduced a novel defense to address
these problems. However, recent studies have been broadly
explored to cover the poisoning attacks in the FL, such as
those presented in [7] and [20]. The methods addressing
targeted model poisoning using a simple improvement of
some FL protocols can be feasible solutions. However, concerns about the probability of data leakage during training are
still significant. Moreover, the incentive mechanism through
SCs in transparent mode with clear communication between
parties that cause the linkability concerns should be dealt with
in the first place.
In this paper, we use the term collaborative learning (see
Figure 2) as a scheme in which multiple computing devices
conjointly build a deep learning model over shared memory, whereby multiple machines with specific computational
capabilities accomplish tasks independently. The linkability
of the smart contract is based on the private Ethereum platform. Our objective is not only to preserve users’ privacy, but
also to drop the linkage between the flow of information in
the network and clients’ identity.
**FIGURE 2. The general form of collaborative learning structures**
(it consists of 7 minimum operations (1-7 OPx)).
**III. PROBLEM STATEMENT**
This section provides an overview of the conventional training model with a cloud-trained mode (centralized logging)
that is heavily adopted in machine learning fields. We also
outline existing incentive schemes based on a centralized
approach in several applications. Finally, the merits of combining these technologies are highlighted concisely along
with the requirements of this research.
_A. CONCENTRATED LEARNING WITH A CLOUD-TRAINED_
_STRATEGY (CENTRALIZED LOGGING)_
An advanced stage of extracting information from a set quantity of data is known as a traditional server-in-the-loop with a
centralized logging approach [21]. This type of architecture
allows the synapse server to gather log data from multiple log
files (on devices) to be later sent to a specific address on the
network. As a result, legitimate parties can carry out several
transactions and activities such as troubleshooting, malicious
detection, and analyzing the behavior of the learning process.
The conventional training with a cloud-trained approach is
shown in Figure 3. Because the log data has been increasing
significantly over time, it complicates the manual maintenance of the logs by operators and developers (e.g., using the
matching protocol).
Regardless of these matters, conventional training with a
cloud-trained approach is straightforward to be employed
in the real-world. This method does not burden the clients’
machines because the training process is conducted in the
cloud. The overhead on the devices can also be automatically
resolved. In this sense, everything is accomplished through
cloud services. However, the security trade-off cannot be
neglected by directly adopting this approach without thorough consideration. Users unknowingly break their privacy
-----
**FIGURE 3. Conventional training with a cloud-trained approach.**
by sending their valuable data to cloud services [22]. The
cloud environment is a semi-trusted party [23]. The server
might be a malicious party that utilizes users’ information to
make a profit. Moreover, it becomes a serious concern when a
malicious server publicly reveals users’ sensitive data. Hence,
a credible solution in regular training with a cloud-trained
approach is necessary.
Improvement of the conventional training model has been
extensively researched in recent years. The upgraded version is called on-device interference with cloud-based training [24]. The training is still conducted on the cloud, but users
can create a type of data bundle that is used for training within
a defined time frame. In this respect, the users are no longer
required to send their data gradually to the model providers.
This scheme reduces the burden on the device even better
than the previous approach. The users also do not have to
be online during the process, making the process more agile
and faster. These merits are essential in systems that adopt
a dispersed approach because the resources and memory are
limited and likely cannot be regularly updated. Nevertheless,
users’ privacy also becomes a concern in this scheme. Users’
valuable data are sent to the cloud services in order to be able
to use the model. The model providers could be malicious,
which might expose users’ data. Therefore, federated learning
is popular for eliminating these concerns because training is
carried out privately within devices.
_B. CONVENTIONAL INCENTIVE SCHEMES_
Practically every incentive scheme is in a centralized model
that relies on the middleman for each transaction. The incentive mechanism over the Internet with a different type of system has a related objective (to motivate the parties to behave
honestly). The nature of a centralized incentive scheme is
still suffering from bottleneck and single point of failure
(SPoF) issues that can jeopardize the entire root of the system. A trust-based incentive scheme with a big data field
as a use case is presented in [25]. The authorized mobile
users are assigned to allocate the tasks of big data with a
reverse auction game model. There is a score function that
determines the highest score to be a winner. However, candidate selection based on trust is concentrated (vulnerable to
bottlenecks and SPoF issues). Similarly, the authors in [26]
designed an incentive mechanism for opportunistic cloud
computing services to address the free-riding problem where
the users are selfish and unwilling to share their resources in
the network. The incentive scheme is based on game theory
using the Nash equilibrium principle. However, avoiding the
involvement of the middleman is likely to be challenging in
real implementation.
Since the advent of blockchain technology publicly,
the paradigm of propagating incentives has deliberately
changed into a decentralized form. Bitcoin (BTC), Ethereum
(ETH), and Monero (XMR) are prominent decentralized
cryptocurrencies built on blockchain technology. A middleman is no longer required to manage the transactions.
As a result, the overall costs of transactions can be reduced
(removing intermediaries’ fees). Incentive-based blockchain
technology provides irrevocable and tamper-resistant activities that enable the parties to monitor the process effectively.
Nevertheless, the transparency feature on the blockchain is
not desired for several cases.
The research papers in [27]–[29], and [30] presented truthful schemes based on blockchain technology to provide a
secure incentive mechanism for different use cases. Even
though the implementation based on blockchain is secure by
design as well as the addition of several protocols, the systems
still suffer from linkability issues that might have a significant
effect on the sensitive data. Therefore, the supplementary
protocol is needed so that it can eliminate the linkage between
information data and the identity of the user.
_C. COLLABORATIVE LEARNING WITH DECENTRALIZED_
_INCENTIVE MECHANISM_
The Google AI team in 2017 presented distributed machine
learning techniques that enable users to improve a machine
learning model privately without exposing their dataset. The
system is called FL as a strategy to improve communication
efficiency (its architecture is explained in detail in [31]). It is
implemented on a smart device mobile keyboard that can
predict the most likely next-words or phrases. Since then,
further research has been gradually carried out by concentrating on specific features such as incentive mechanisms
for the parties involved. Existing federated learning schemes
lack incentive mechanisms. The free-rider problem is still
likely to be the greatest concern in such a system. Besides
being able to motivate users, incentives can also force users
to behave honestly. In line with this, blockchain merits can be
a plausible solution to be adopted within federated learning.
A recent study on FL and blockchain-based incentive
mechanisms was presented in [32]. These technologies are
implemented in the edge network field involving internetof-things devices (network edge). A similar objective was
also presented in [33]. Overall, key design aspects and incentive mechanisms are achieved. However, as in most previous studies, the offered system still suffers from traceability,
-----
linkability, and transparency issues. It is likely not desirable
to be implemented straightforwardly for users that hold a
large amount of sensitive information.
Determining the type of blockchain adopted in the private
cross-silo federated learning system is paramount because the
new platforms that have emerged with the unique features
offered are unsuitable and likely to be vulnerable to viewing
by unauthorized observers. The effectiveness of verifying
transactions, including the process of distributing incentives,
is also a vital detail that must be thoroughly considered. BTC
with proof-of-work (PoW) as a consensus mechanism could
be a barrier in federated learning that requires rapid incentive
procedures. The difficulty level in BTC always increases over
time, as can be seen in Figure 4. By directly adjusting the
difficulty level (either by slowing down or speeding up) can
affect the security [34].
**FIGURE 4. Statistic of BTC’s difficulties over the years [35].**
_D. REQUIREMENTS_
Linkability-awareness in collaborative learning and smart
contract transactions becomes essential when these approaches are linked to valuable user data. The fundamental objective is to provide a tamper-proof incentive mechanism in
collaborative learning by relying on smart contracts. It can
resolve disputes between parties by design. However, it is
not advisable to directly adopt the transparency process.
Additional protocols are required to break the linkage information during transactions. More precisely, our design objective must meet the following requirements:
(i) (RQ-1) Sustainable private learning activities. A collaborative learning scheme should be able to provide
sustainable private learning activities for the parties. Our
system preserves users’ privacy by design because the
training is conducted confidentially without revealing
the data. We propose plausible techniques that allow
users with sensitive data to carry out training without
suffering linkability issues.
(ii) (RQ-2) Compatible decentralized incentive mecha**nism. The blockchain-based incentive provides tangible**
benefits for enterprises. Incentive schemes based on
smart contracts should be able to provide a compatible
incentive for the parties within the collaborative learning
system. The system must also satisfy the fairness of
revenues associated with users’ resources.
(iii) (RQ-3) Unlinkable and untraceable transactions.
Accuracy, high-speed, and trust quotient are among
the advantages offered by blockchain smart contracts.
Nevertheless, in cases of private distributed learning
with sensitive data, the transparency process is not
preferable. Therefore, the system should be able to
provide unlinkable and untraceable transactions. The
observer still can see the information in the blockchain
network since the SC is visible publicly, but the observer
has zero knowledge about the transactions that occurred.
(iv) (RQ-4) Comprehensiveness. The designed system
must be able to ensure the completeness of the entire
transaction process, starting from learning activities up
to the propagation of incentives.
**IV. MODELING AND OPERATIONS**
First, we provide information about the federated learning
model as a backbone use case in this research. The concept
development model is adopted by referring to FL principles.
The protocol for group signatures is detailed in this section.
We also design the used-model-only feature and the secure
decentralized rewarding schemes as plausible approaches to
preserve users’ privacy and unlinkable transactions while
using the services offered.
_A. COLLABORATIVE LEARNING FRAMEWORKS_
In this research, collaborative learning frameworks are signified by the number of clients Cx1, Cx2, . . ., Cxn _Cxi,j in_
∈
different groups G1, G2, . . ., Gn _Gi,j that conduct training_
∈
on a global model ψglb(x) with a private dataset δn. For ease
of presentation, we set five groups of clients G1 _G5 with_
−
20 devices for each group G(1 − 5) − _Cx(01−20) (a total_
of 100 clients). Each client possesses the same dataset with
comparable capability settings on all devices (in the realworld implementation, the dataset and device capabilities are
distinct from one another).
Every group with predetermined settings jointly conducts training and sends the results of the gradient values
_ψup[group][1], ψup[group][2], . . ., ψup[group][5]_ to the leader in each
group. Leaders are sorted based on the number of transactions
recorded by the system. In round 1 r1, the leaders perform
aggregation in order to obtain the updated gradient values
from the clients within the group. The final aggregation value
is determined by the aggregation server Svr _[Ag], which also_
has the role of a model provider. Figure 5 illustrates the
updated filtering of each group ψup[group][1], . . ., ψup[group][5] to
detect suspicious updates from malicious clients. The server
_Svr_ _[Ag]_ makes a prediction about which malicious clients have
tendered one of the gradients. Eventually, the server compares
the gradient values and flags the outliers that are significantly distant from the rest {[′]ψup[group][1], . . .,[′] _ψup[group][5]} ⊂_
{[′]ψup[group][1], . . .,[′] _ψup[group][5]}. However, the filtering update_
operation is out of the scope of our research because we
-----
**FIGURE 5. Groups of collaborative learning with an aggregation server.**
assume that the clients are behaving honestly. For more
details about the filtering process, we recommend the readers
refer to the following references [20] and [36]. The clients
with their associated datasets can be signified in (1) as
follows:
_Cxn_
�
_r1 →_ (ψglb(x)) = {Cx1(δ1), . . ., Cxn(δn)}, (1)
_n=0_
_Tr1 ≤_ _Tmax[Cx][n][;]_ ∀G(1−5) − _Cx(01−20) →_ _ψglb(x))_ (2)
_MAE(ψglb(x))_
_n_
�
= [1]n |yi − _f (xi)|_ (3)
_n=1_
The clients are bounded with a maximum training Tmax[r][n]
time for each round. The training time is set by the aggregation server Svr _[Ag], which also has a role as a global model_
_ψglb(x) provider. The maximum training time is adjustable_
and can be determined by the model provider as needed
(denoted in (2)). Within Tmax[r][n] [, the clients collectively trans-]
fer the updated gradients to the group leader. The updated
gradient groups ψup[group][1], . . ., ψup[group][5] are derived from
the aggregation value from every participant in the group.
Specifically, there are five gradient values from the total
number of the group that will be filtered by the aggregation
server {[′]ψup[group][1], . . .,[′] _ψup[group][5]} →_ _Svr_ _[Ag]. The accuracy_
is calculated by the mean absolute error (MAE), as shown
in (3) [37]. Eventually, the final aggregated group updates
_ψup[final]_ are calculated by the server by excluding the updates
identified as suspicious clients. The accuracy is calculated
by the MAE, as shown in (3), where f (xi) is the prediction
value of model ψglb(x), and yi is the actual value of the
records. In short, the lower the MAE values in model ψglb(x),
the higher its accuracy.
_B. GROUP SIGNATURES OF CLIENTS AND TERMINOLOGY_
The main objective of group signatures in the collaborative
learning transaction is to hide the real identity of the parties
involved in a transaction. To perform group signatures in
a transaction, the members do not require a group manager or middleman. The identity of the signer is disguised
by design because the transaction is signed on behalf of the
group. The use of the ring signature in collaborative learning
transactions is inspired by CryptoNote v 2.0 [38]. Untraceability in a transaction can be accomplished by implementing
a ring signature in a transaction, which is also a core piece
of technology behind the CryptoNote protocol. Once a group
signature of clients has been created, every member of the
group is free to use the signature by combining it with his/her
private key to disguise the signers’ identity.
The signer allows choosing the number of signatures to
be included in the transaction in order to be able to disguise
the signers’ information. The signatures can be chosen freely
as long as they are part of the group signatures of clients.
_Rsgn ∈_ _Rmb ⩾_ 1 (Rsgn is a group signature chosen by
the clients, and Rmb is the ring members/total number of
signatures available in the group). The parent keys for each
party are derived from the trapdoor permutation function,
such as Rivest-Shamir-Adleman (RSA), Rabin cryptosystem,
and elliptic curve cryptography (ECC).
**TABLE 1. Summary of notations used.**
Suppose the aggregation server/model provider (Svr ),
client 1 Cx1, client 2 Cx2, client n Cxn, and reward manager
_Scx are in the same group of the collaborative learning system._
Each party has a pair of parent keys (Pubn, Privn). The public
key is computed by encryption yn _gn(xn), where gn is an_
=
extended trapdoor permutation function and gn(xn) defines
_fi(xi) = x[2]_ _mod ni over {0, 1}[b]. A summary of the notations_
used in this paper is defined in Table 1. Eventually, the signature key for every member in our collaborative learning
scheme can be defined as follows:
-----
(i) Aggregation server (Svr _[Ag]) →_ _hash(Pubvr_ _, Privvr_ ) →
_Pubvr = yvr = gvr_ (xvr )
(ii) Client 1 (Cx1) → _hash(Pubx1, Privx1) →_ _Pubx1 =_
_yx1 = gx1(xx1)_
(iii) Client 2 (Cx2) → _hash(Pubx2, Privx2) →_ _Pubx2 =_
_yx2 = gx2(xx2)_
(iv) Client n (Cxn) → _hash(Pubxn, Privxn) →_ _Pubxn =_
_yxn = gxn(xxn)_
(v) Reward Manager (Scx) → _hash(Pubsc, Privsc)_ →
_Pubsc = ysc = gsc(xsc)_
Since the clients use the signature on behalf of the group,
the observer cannot infer any information within the transaction. The signature can be used for any transaction without procuring permission from every member of the group.
For instance, clients Cxi,j leverage the following keys (4)
to sign a message hash(msg, yvr _, yx1, yxn . . ., ysc). After the_
group signatures of clients are generated, the parties can unite
new members to the group by using members’ public keys
_Update_Rsgn →_ _Get_New_PubKey ⊕_ _Rsgn. The parties can_
also exclude the members’ public key straightforwardly out
of the ring members DelRsgn → _Del(Get_PubKeyn) ∈_ _Rsgn._
_Rsgn →_ _ycx1 ⊕_ _ycx2 ⊕_ _ycx3⊕, . . ., ⊕ycxn;_
{gcx1(xcx1) ⊕ _gcx2(xcx2)⊕, . . ., ⊕gcxn(xcxn)};_
_Rtot_
�
_gi(xi) = gcxn(xcxn) ⊕_ _gvr_ (xvr ) ⊕ _gsc(xsc);_ (4)
_i=1_
Enhancing privacy in CL activities entails terminology that
should not be confused with similar entities in a different
environment. The CL’s base signature algorithm refers to
the elliptic curve discrete logarithm problem. Secretec _key_
−
for each party is a number sα ∈ [1, l − 1], where l is a
prime order of the base point in ECC. Publicec _key is_
−
defined as a point pubα = secα · G, with G as a generator.
_One_ _timekeypair in the CL’s transaction is a set of secret_
−
and publicec _keys. Intuitively, each participant possesses_
−
a pair of secretuserkeys(sα, secβ) from a couple of different
_secretec−keys. There is also a pair of trackingkeys(sα, pubβ),_
which is derived from secret and publicec − _key (pubβ =_
sec β · G and secα ̸= _secβ). In conclusion, a pair of_
_publicuserkeys(pubα, pubβ) is obtained from the associated_
private key (sα, secβ).
The public key of the clients is enforced as a one-time
destination key and a one-time private key in order to use the
funds. We elaborate on this point in detail in Section IV-C
and Section IV-D. The structure of the transaction generally persists, comparable to the Bitcoin and Ethereum fields.
Every participant in CL requests a global model by collecting several independent transaction outputs and signing the
transaction with the corresponding secret keys.
_C. USED-MODEL-ONLY TRANSACTIONS_
In Section IV-B, we present the group signatures of clients
and the terminology used in the CL system. Clients may
desire to use the deep learning model provided by the aggregation server without acquiring cryptocurrencies as a reward.
→ _Ck,v(yx1, yx2, . . ., yxn),_
_Ek_ (yxn ⊕ _Ek_ (yn − 1 ⊕ _Ek_ (yx1 ⊕ _v)) ≡_ _v_ (7)
Algorithm 1 presents the sequence of collaborative learning with the distribution of incentives through the blockchain
network. A transaction request Txδ1_req consists of global
model info, dynamic rule, and request statement Txδ1_req =
_ψglb(i,j)||rdc_info||‘‘req.’’; and this transaction is signed_
using Cx1’s private key combined with group signatures of
clients that have been generated in advance as can be seen
in (6). The value of yx1 = gx1(xx1) within the group is
calculated using Cx1’s private key. Simultaneously, the ring
equation of the total group generated in yx1 can be solved
using Cxi’s private keys xi = g[−][1](yi). The client expresses
the desired global model in the form of type and version.
In another case, the client may not satisfy dynamic rules in the
sense of the minimum number of datasets set by the model
provider, so rewards are not given. Even though no reward
is distributed, clients still require their identity to be hidden
during training for several purposes. The CL system must
disguise communication between the clients and the aggregation server. We call this case used-model-only transactions
_(UMO-Tx)._
In the case of UMO-Tx, earnings incentives are not the ultimate goal for clients. The CL framework (in Section IV-A) is
designed to be private and requires an authentication process
to join the system. We assume that the clients are legitimate
parties who have passed through the official authentication
process. In short, the authentication method is beyond the
scope of our research. The initial process is identical in that
the client makes a transaction request to the model provider
by using the group signatures of clients in the transaction. The
provider will send the global model if the clients’ transaction
is labeled ‘‘true’’.
To use the desired global model ψglb(x), the client Cxn ∈
_Cxi,j is required to make a transaction request Txδ1_req_
addressed to the model provider Svr _[Ag]. For instance, client_
1 Cx1 applies the UMO-Tx mode for a particular purpose.
_Cx1 then conceives a group signature of clients by selecting_
a number of members from the total available. As an illustration, Cx1 takes 25 public keys of clients in sequence combined
with the public key of the aggregation server Svr _[Ag]_ and the
smart contract manager Scx (available: Rsgn 100 clients).
=
In this case, the ring signature can be constructed as follows:
_Rsgn →_ _gvr_ (xvr )⊕gsc(xsc)⊕gcx1(xcx1)⊕, . . ., ⊕gcx25(xcx25).
The final group of signatures used by clients in the Txδ1_req
transaction is signified as (5). Client Cx1 is free to choose the
number of members as long as Rsgn ∈ _Rmb ⩾_ 1.
_Rsgn_
�
_Clientsgcxn(xcxn ⊕_ _Agg.Servergvr_ (xvr )
_i=1_
⊕SCManagergsc(xsc) ∈ _Rmb ⩾_ 1); (5)
_ψglb(i,j)||rdc_info||‘‘req.’’_
_Txδ1_req =_ _,_ (6)
_signwithRsgn ∈_ _Rmb ⩾_ 1||privx1
_Combiningfunction_
-----
**Algorithm 1 The Global Model ψglbx Is Provided by the Aggregation Server Svr** _[Ag]. The Model Is Gradually Trained by a Set_
Number of Clients Locally and Privately in Their Respective Groups Cx(i,j) in G(i,j)
1: procedure MODEL PROVIDER Svr _[Ag]_ PERFORMS:
2: _Svr_ _[Ag]_ the provider publishes several global models ψglb1, . . ., ψglb2, . . ., ψglbn;
3: _Svr_ _[Ag]_ conceives a group of FL’s signature(ex. 25 members)
4: Estimates Cxn _Cx(i,j) in groups Gn_ _G(i,j)_ _*roughly mapping available devices_
∈ ∈
5: Publishes rdc_info →∀ _ψglbn_ _*minimum requirements and rewarding’ info_
6: Set Tmax[r][n] [→∀] _[ψ][glb]n_
7: 20 devices for each of the five groups → _G(1 −_ 5) − _Cx(01−20)_
8: **for group signatures of clients do**
9: Parents private keys of the parties → (Pubn, Privn)
10: Signature for one party is calculated → _yn = gn(xn)_
11: Ex: Aggregation server (Svr _[Ag]) →_ _hash(Pubvr_ _, Privvr_ ) → _Pubvr = yvr = gvr_ (xvr )
12: _*(pair of parent keys from trapdoor permuatation functions)_
13: Ex: One group signatures → _Rsgn →_ _ycx1 ⊕_ _ycx2 ⊕_ _ycx3⊕, . . ., ⊕ycxn;_
14: _(every member of the group is free to use the signature by combining it with his private key ..._
15: _to disguise the signers’ identity)_
16: **end for**
17: **for used-model-only transaction (UMO-Tx) do**
18: _(For example, Cx1 as requester & Svr_ _[Ag]_ _as a model provider & an aggregation server)_
19: _Cx1 determines the desired global model ψglb(i,j)_
20: _Cx1 generates a group of signature Rsgn ∈_ _Rmb ⩾_ 1
21: Submits Txδ1_req = ψglb(i,j)||rdc_info||‘‘req.’’; _*sign it with Rsgn_
22: _Svr_ _[ag]_ checks clients transaction Txδ1_req
23: **end for**
24: **for rewarding mechanism (performed by Svr** _[Ag]) do_
25: _(Cx1 believes his updated gradient value ψup[δ][1]_ _meets the requirements to be incentivized)_
26: _Cx1 tenders a new transaction Txδ1_ETH_
27: _Svr_ _[Ag]_ confirms clients’ transaction with respective updated gradient value ψup[δ][1]
28: _Svr_ _[Ag]_ unpacks Cx1’s public keys (Pubα1, Pubβ1)
29: _Svr_ _[Ag]_ generates a random r ∈ [1, l − 1] & computes a one-time destination OTDcx1
30: _OTDcx1 is sent over the blockchain network_ _*Cx1 checks every passing txs using Privα1, Privβ1_
31: _Cx1 can recover the corresponding one-time private key OPKcx1_ _*one-time private key for spending_
_the reward_
32: **end for**
33: _(the process is carried out repeatedly until it reaches the maximum round determined by the provider)_
34: _(each model may require different dynamic rules, and it might be distinct)_
35: end procedure
Concurrently, dynamic rules contain information about system and device requirements, network, and conditions for
reward provisions. The transaction is computed by the system
that combines functions Ck,v, where k is a hash value of
_Txδ1_req and v is a random glue value picked by the client._
This computation assumes a random oracle for a cryptographic hash function, while the client will use k as a key for
_Ek as signified in (7). Clients also enable the addition of a new_
signature Update_Rsgn or remove the member’s signature
_DelRsgn as needed without having to obtain approval from_
the model provider.
A transaction request is addressed to the model provider
directly through a secure channel that is resistant to eavesdropping and tampering. The model provider unpacks the
transaction Txδ1_req sent by Cx1 and verifies whether the
signature co.mes from the group signatures of clients. The
integrity of transactions is also verified promptly by the
provider. Finally, the desired global model will be sent to
_Cx1 by the provider if Txδ1_req meets the following con-_
dition: Rsgn ∈ _Rmb ≡_ _v AND Txδ1_req’’ = True and_
described in (8).
- sgni,j ∈ _Rmb ≡_ _v ‘‘AND’’_
‘‘Txδ1_req’’ = True → _then‘‘Approve’’_
- sgni,j /∈ _Rmb ≇_ _v ‘‘OR’’_
‘‘Txδ1_req’’ = False → _then‘‘Decline’’_
_Txδ1_req_
(8)
Conclusively, in the event of UMO-Tx, clients can still
take advantage of the models provided by the system
without having to reveal their identities to the public.
-----
Obscuring a transaction request can be achieved because the
transaction is signed on behalf of the group. The initial stage
for a transaction request is always the same for each client.
The difference between them is that the transaction is signed
by the private key of each requester combined with the group
signatures generated beforehand. From the client’s perspective, receiving incentives is not the primary purpose; this
could be due to the client not meeting the dynamic rules set
by the model provider or any other limitations. Nevertheless,
the clients’ identities remain anonymous, and transactions
can be carried out securely.
_D. SECURE DECENTRALIZED REWARDING SCHEMES_
In Section IV-B, we elaborate the used-model-only transaction scheme (UMO-Tx), where the client does not intend
to receive incentives but still wants to be anonymous while
using the global model. To receive incentives from the model
provider, the clients have to meet all the dynamic rule requirements set by the model provider. One of the requirements
that should be met by clients is the least number of datasets
_δ1 knowledge that corresponds to the updated gradient values_
from training ψup. When all conditions are satisfied, and the
provider has verified the transaction, incentives can be given
to clients securely. This section presents an unlinkable and
untraceable incentive mechanism by utilizing the CryptoNote
protocol and blockchain technology through smart contract
features.
For ease of understanding, we consider that client Cx1
possesses a sufficient amount of dataset δ1 to be incentivized.
The client also meets the dynamic rules in terms of device and
network requirements. First, the client deploys a transaction
request addressed to the model provider. The client generates
a group of signatures by choosing a number of members’
public keys [�]i[R]=[tot]1 _[g][i][(][x][i][)][ =][ gcx][n][(][x][cxn][)][ ⊕]_ _[g][vr]_ [(][x][vr] [)][ ⊕] _[g][sc][(][x][sc][).]_
The clients state the desired global model in the transaction,
along with the dynamic rules information. Cx1’s transaction
request is then signed with the group signatures combined
with Cx1’s private key. The model provider checks every
incoming transaction and unpacks the transaction request
of Cx1. The provider will send the global model to Cx1 if
the only transaction satisfies a particular condition, which is
signified in (8). The conditions applied by the provider can
vary significantly for each global model.
1) DEPLOYING TRANSACTION Txδ1_ETH VIA SCs
In order to be incentivized, the Cx1 is required to submit
a new transaction through a private Ethereum SC, which
is denoted as Txδ1_ETH . This transaction represents the
global model type and version used ψglb(info), the gradient
value of the training results ψup[δ][1], and the dataset’s knowledge δ1knowledge. The primary difference in the transaction
_Txδ1_ETH is that Client Cx1 inserts a pair of public keys_
(Pubα1, Pubβ1). Public key Pubα1 is generated based on
_Cx1’s private key Privα combined with a base point/generator_
_Gα as follows: Pubα1 →_ _Privα1 · Gα. Whilst, public key_
_Pubβ1 is derived from another Cx1’s private key Privβ with_
its own base point/generator Pubβ → _Privβ1 · Gβ; where_
_Privα1_ _Privβ1 ‘‘AND’’Gα_ _Gβ as given in (9)._
̸= ̸=
� �
_ψglb(info)||ψup[δ][1]||δ1knowledge_
_Cx1[′]sPubKey →_ _Pubα1, Pubβ1_
_Txδ1_ETH =_
{signwithRsgn ∈ _Rmb ⩾_ 1||privx1}
_where →_ _Privα1 ̸= Privβ1‘‘AND’’Gα ̸= Gβ;_
(applied to all clients with respective ‘‘G’’)
(9)
A pair of public keys (Pubα1, Pubβ1) that are attached to
an Ethereum SC transaction Txδ1_ETH have their respective purposes. The first Cx1’s public key (Pubα1) is used
together with the model providers’ random data r
=
_Svr_ _[Ag][′]srandomdata_ _R_ _r_ _G as a part of the Diffie–_
→ = Hellman key exchange concept in a transaction. In this sense,
the sender and receiver both use half of the information that
can be decrypted using the recipients’ secret key. On the
other hand, the second public key (Pubβ1) is employed as a
tracking key. The client Cx1 will search for a transaction in the
blockchain network sent by the provider by attaching Pubβ1,
so that the client can recognize that the transaction is intended
for him. Transactions sent by the provider contain funds or a
cryptocurrency that can be used by legitimate clients only.
2) ONE-TIME DESTINATION KEY AND
ONE-TIME PRIVATE KEY
Before a specified amount of cryptocurrency is transferred
through the blockchain network, the provider first checks the
transactions Txδ1_ETH claimed by the client. If all conditions are met (labeled as ‘‘True’’), then a pair of the public
key of Cx1 is unpacked. The provider then performs a random
base point r ∈ [1, l − 1] and computes a one-time destination
_OTDcx1 key addressed to the client Cx1, as shown in (10):_
_OTDcx1 = Hs(rPubα1) · G + Pubβ1_
_where r = Svr_ _[Ag][′]srandomdata →_ _R = r · G,_ (10)
_OPKcx1 = Hs(Privα1 · R) + Privβ1_ (11)
When transaction OTDcx1 is sent over the blockchain network, client Cx1 checks every passing transaction using the
private key Privα1, Privβ1. Cx1 can recover the corresponding one-time private key to use the funds/cryptocurrency
because only Cx1 has knowledge about Privα1 and Privβ1.
The clients Cx1’s one-time private key is signified in (11).
In the original CryptoNote protocol, the one-time private key
is also being used as part of a ring signature to disguise the
signers’ identity. Eventually, a key-image can prevent doublespending intentions from malicious clients.
_E. RESEARCH DESIGN LIMITATIONS AND ASSUMPTIONS_
1) A HARD FORK (RADICAL CHANGE) REQUIREMENT
The Ethereum platform provides a decentralized ecosystem
(see Figure 6) for developers to create products using the
Ethereum Virtual Machine (EVM). The EVM is powerful and
-----
**FIGURE 6. Ethereum VM on blockchain.**
embedded within each full blockchain network by design.
The smart contract byte-codes are executed through an EVM.
Interacting with the EVM via smart contracts is likely to
be more costly than with traditional servers. In other words,
numerous use cases are favored using the EVM rather than
conventional servers. However, our proposed schemes cannot
be employed entirely in the Ethereum network because it
requires a hard fork (radical change) to be applied in the
entire network. A hard fork in the Ethereum requires a radical
change to the network protocol, which can alter the entire
course of the transaction. Therefore, we provide several performances estimates for the transaction through the EVM,
which are detailed in Section V.
2) SYNCHRONOUS DISTRIBUTED LEARNING
AND COMPUTING
Theoretically, collaborative learning requires a massive number of devices with various device capabilities to generate
an aggregation value. Likewise, the datasets used in local
training are distinct from one another. In terms of simulation,
it is not straightforward to achieve all of these requirements.
For these reasons, we set the simultaneous computation with
multiple computing devices over shared memory with the
same dataset derived from [39]. Regardless of the outlined
limitations and assumptions, the critical points of the simulation output were successfully collected.
**V. PERFORMANCE RESULTS, COMPARISONS,**
**AND LESSONS LEARNED**
_A. EXPERIMENTAL SETUP_
As discussed in the previous sections, we propose the model
provider to have another role as an aggregation server, which
calculates the most recent gradient values obtained from
multiple groups. The model provider constructs an AI model
using a convolutional neural network (ConvNet) to analyze
visual imagery from private datasets of several clients. The
deep learning model used was not our primary focus in
this research. Thus, we implement a straightforward ConvNet model with a convolution layer and rectified linear unit
(ReLu) as an activation function and a pooling layer (added
after the convolutional layer). Finally, in the classification
part of ConvNet, several core components are adopted, such
as flattened, fully connected, and softmax functions.
Inside the first layer of ConvNet, we set the input to be 1,
the output is 25, the kernel size is 5, and stride is 1. A fully
connected layer has a fixed size input image of 4 4 for
×
feature size with 50 channels combined. Meanwhile, the ConvNet training loader’s batch size is arranged to be 50 data
samples with 1,000 samples for testing (out of 60,000 training
samples [39]) that have been size-normalized by default.
Finally, collaborative learning and blockchain performance
tests were carried out on an Intel(R) Core(TM) i7-7700 CPU
@ 3.60GHz and 3.60GHz with 16.0 GB RAM modules.
For the blockchain network, we use the Ethereum smart
contract through the Ganache - Truffle Suite platform to
run inquiries, execute commands, and inspect the transactions’ state with all required dependencies installed. The
clients’ and providers’ account addresses are obtained from
Ganache, similar to the parties’ private keys. The blockchain
network is implemented on the remote procedure call server
HTTP://127.0.0.1:7545 with automining mode. The gas price
was set to be 20,000,000,000 Wei, and the gas limit was
set to 6,721,975 by default. Ganache provides 100 Ether
by default; with these amounts of Ether, clients are able to
conduct various transactions in the Ethereum. The parties’
identities are managed by a crypto wallet called MetaMask,
which is an extension for accessing Ethereum and a gateway
to blockchain applications.
_B. UNLINKABLE COLLABORATIVE_
_LEARNING PERFORMANCE_
Collaborative learning activities in this research are performed in a synchronous distributed learning and computing
environment, where multiple computing devices are shared
over the same memory and computing resources. This can be
described as a single machine that executes some commands
simultaneously [40], where the execution tasks are divided
separately. Therefore, training activities can be completed
nearly at the same time in a parallel form. Detailed information on training methods can be found in Section IV-A, and
the environmental setup is presented in Section V-A.
Collaborative learning activities are carried out within each
group using the same model and training data. In real-world
implementations, the dataset is closed publicly. To use the
global model, each client first makes a Txδ1_req transaction
addressed to the provider. The leader of each group collects
the updated gradient values from all clients within the group.
The leader eventually calculates the group’s updated gradient values ψupgroup. The final aggregated gradient values
_ψup[final]_ are computed by the provider, which later redistributes the new updated model back to the clients.
The unlinkable collaborative training activity is performed
for one round. Figure 7 (a) illustrates the log loss for each
group during training activities, while Figure 7 (b) shows
the length of time taken to complete the training. Comprehensively, training activities improve with increasing cycles,
as is the nature of training in a single deep learning model.
However, we find that the training activities in Group 1
are slightly better than those in the other groups in terms
of log loss (1.1785, 1.4285, 1.5585, 1.6485, and 1.7585,
respectively).
-----
**FIGURE 7. (a) Visualization of the log loss for each group in CL; (b) the**
length of time taken to complete the training in one round.
Correspondingly, the time required to complete the collaborative learning activities also varies for each group, even
though the command is executed simultaneously over shared
memory. Based on the performance results collected, Group
1 completed training moderately fast (272.92 minutes) compared to Group 2, which also completed the task a little faster
than Group 3 (276.62 and 280.91 minutes, respectively). The
longest time was experienced by Group 4 and Group 5, which
required nearly 295 minutes to complete the task. This sort of
phenomenon can occur because we placed Group 1, Group 2,
and other groups in the order that allows them to be a priority
in execution on the computer. However, the results might be
significantly different for a real-world implementation with
different datasets, types of networks, and computing device
capabilities.
**FIGURE 8. (a) Accuracy (out of 10k training data); (b) Distribution points**
of the average loss of collaborative learning; (c) Heatmap of average loss
per cycle.
The average training accuracy for all groups is illustrated
in Figure 8 (a), where the maximum size of training data is
10,000 images. The green line separates the higher half from
the lower half of the training sample. Figure 8 (b) depicts the
distribution points of the average loss of collaborative learning activities, which are derived from the calculation results
of all groups. The red horizontal line is the middle number
in the sorted log loss derived from training. Figure 8 (b)
shows the heatmap of the average loss per cycle. Figure 8 (c)
shows the distribution values of the heatmap of the average
loss per cycle. The average log loss value is considerably high
from the first cycle up to the 40th percentile, ranging from
2.0115 to 2.7365. Nevertheless, the accuracy increased gradually as the number of cycles increased, reaching 91.776% on
average from clients’ combined total. The log loss among the
groups is not enormously distinct because the environmental
setup is comparable from one to another. For better data analysis from performance results, a variance of network types
and device capabilities are required (as future endeavors).
_C. GROUP SIGNATURES OF CLIENTS AND_
_DECENTRALIZED REWARDING PERFORMANCE_
A secure decentralized reward scheme is constructed by
deploying Txδ1_ETH through the Ethereum smart contract.
The client must meet all the requirements set by the model
provider. The client also publishes a pair of public keys
within the transaction. The clients’ public key (Pubα1) will
be used together with the providers’ random data r
=
_Svr_ _[Ag][′]srandomdata →_ _R = r · G and the public key (Pubβ1)_
as a tracking key in the blockchain network. We describe
this scheme in detail in Section IV-D. When the model
provider confirms the Txδ1_ETH transaction, the incentive
is given to the client by unpacking the clients’ public key
(Pubα1, Pubβ1). A fair incentive system can motivate clients
to behave honestly without worrying about getting incentives appropriate to their contribution. Furthermore, incentives are propagated securely with unlinkable and untraceable
transactions.
To estimate the amount of gas and Ether used in a transaction, we recorded 10 different transactions from 10 consecutive clients. Transactions between clients are differentiated by the amount of arbitrary value input in the smart
contract. The input in transaction Txδ1_ETH is contrasted
with information about the global model used, the training
result (gradient values), information about the dataset used,
a pair of public keys from the client, and signed with the group
signatures of clients selected as coveted. Therefore, for ease
of presentation, we set Client 1 to input the least arbitrary
values in the smart contract, followed by Client 2 with more
input than Client 1. In particular, Client 10 is the client with
the longest arbitrary input.
**TABLE 2. Summary of the cumulative gas used in the transaction.**
A summary of the cumulative gas used in the transaction is
displayed in Table 2. We record the gas used by the clients
and the smart contract manager, which is also a part of
the model providers’ role in distributing the reward for the
clients. The cumulative gas from the clients is used in the
-----
transaction Txδ1_ETH, while the gas of the smart contract
manager is used to distribute the reward. The lowest gas consumption in the transaction occurred in the first transaction
by Client 1, which is 96,388 units. Meanwhile, the highest
gas consumption occurred in the 10th transaction, amounting
to 106,921 units with an average overall transaction rate
of 100,960 units with the gas limits being automatically
adjusted by the system. Various arbitrary inputs cause this
difference with different sizes for each client. Therefore,
the amount of gas used varied. On the smart contract managers’ side, the difference is not significant because there is
no notable difference in distributing rewards to clients.
In terms of a group signature of clients directly affecting
the gas fee, Ether is used because the clients are required
to state the information of the signature used in transaction
_Txδ1_ETH_ . The more combinations of public keys that are
used in the transaction, the more reliable the transaction will
be from the observer’s efforts to find out the contents of the
transaction. However, the use of a large number of signatures
raises trade-offs in terms of execution times, which is longer,
and vice versa. There are numerous researches seeking a
better way in using a ring signature scheme such as proposed
in [41], [42], and [43].
**TABLE 3. Benchmark of the spent Ether in transaction (Tx).**
The amount of Ether spent deploying a transaction is
summarized in Table 3. Every client and provider in the CL
possesses 100.00 Ether by default. The Ether can be used to
deploy a transaction in the Ethereum network, so the amount
of Ether issued and earned can be inspected through the
Ganache interface. The lowest amount of ether consumption
occurred in the first transaction, which is 4.43 × 10[−][3] ETH.
The last transaction with the greatest input spent 5.67 × 10[−][3]
ETH. The amount of Ether spent on each transaction is associated with the amount of gas consumed. Meanwhile, Ether
on the smart contract manager side is shown to be larger
because the results have been accumulated with rewards for
each client (a combination of rewards and transaction costs).
The number of rewards given can be set freely by the model
provider (from 0.001 ETH to 0.1 ETH). Therefore, we focus
only on transactions made by clients. Eventually, all transaction information is shown in Figure 9.
_D. COMPARATIVE ANALYSIS_
With the emergence of federated learning that can overcome
privacy issues by design and blockchain technology, which is
immutable, the merger of these two technologies has begun
to be studied by researchers and industries. Research on the
use of blockchain in federated learning regularly influences
**FIGURE 9. (a) Comparison of gas units between clients and smart**
contract manager; (b) The amount of ether spent in a transaction.
methods of distributing incentives wherein blockchain technology and third parties are not required to be involved
in the transaction. By design, a blockchain-based incentive
mechanism is suitable for use in the federated learning system
because the data structure in a decentralized ledger is appendonly, such that the data record cannot be altered or deleted.
Furthermore, blockchain relies upon protected cryptography
to secure data records (in chronological order with a timestamp). We analyze our research results with various prior
studies, with multiple platforms and objectives summarized
in Table 4. Previous studies are selected in terms of combining
collaborative learning technology and blockchain as the backbone of the incentive scheme in the decentralized learning
approach.
Other advantages of blockchain, such as transparency and
traceability of transactions, are not expected in a distributed
learning system that processes sensitive data. Many studies
have shown that decentralized training allows for data leakage
by observers with certain assumptions. The observers with
marked assumptions as presented in [49] can deploy whitebox and black-box attacks by enrolling the engineered term,
which is capable of producing a good performance (primary
task), but also capable of leaking clients’ training data (malicious task). Moreover, the implementation of blockchain
technology that is transparent and traceable exacerbates the
-----
**TABLE 4. Performance benchmark with several of the existing approaches on unlinkable collaborative learning and decentralized rewarding in CL.**
possibility of data leakage allowing the observer to associate
each transaction with an openly viewable data owner. Therefore, a secure, unlinkable, and untraceable incentive distribution mechanism is necessary for a collaborative learning
system as part of our objective in this research.
Table 4 presents a comparable approach to several studies in utilizing blockchain-based incentives in decentralized
learning systems. The types of platforms and methods used
vary, but we focus on three key points, namely additional pri_vacy, unlinkable transactions, and untraceable transactions_
provided by the authors. Research in [44] proposed a similar
approach and platform to our research, but the transactions
are still traceable and can be linked by the observer. Likewise,
the schemes proposed by [13], [17] and [33] also preserves
additional privacy to protect the client’s identity and to make
sure the transaction is carried out securely. Yet, the authors do
not focus on the linkable and traceable transaction concerns
in decentralized learning with blockchain-based incentives.
Meanwhile, the research papers in [11], [14], [45]–[48] proposed a similar approach, but there is no information about
the additional privacy in decentralized learning, nor about the
unlinkable and untraceable transaction that can jeopardize the
clients’ identity and data.
Our proposed approach successfully satisfies three key
points. First, we designed additional privacy to protect the
identity of the clients. Second, we also designed an unlinkable
transaction that provides a sense of security for the client, and
finally, we designed transactions that cannot be tracked by
observers. In terms of training accuracy, our results are no
better than those of previous studies because the decentralized
learning and algorithms used are not our primary focus in this
research. We concentrate on designing transactions that are
secure, unlinkable, and untraceable by observers.
**VI. CHALLENGES AND CONSIDERATIONS**
We differentiate this section into two points: the considerations of aggregation servers’ roles and future directions of
collaborative learning with the blockchain-based incentive
mechanism for a variety of similar applications.
_A. AGGREGATION SERVERS CHALLENGES_
_AND CONSIDERATIONS_
Protecting users’ privacy is the fundamental premise of our
proposed schemes. Aggregation servers collect the updated
gradient values derived from multiple users. The server then
calculates the aggregation values as representative of the
updated model. The aggregation servers also play a role
in determining the amount of cryptocurrency for the contributed users. In this sense, a large number of transactions
might burden the servers that can reduce system effectiveness
because the server is still in a centralized form affected by
bottleneck issues. Moreover, aggregation servers are assumed
to be semi-trusted parties [50]. The whole transaction process
is in a decentralized form, but the aggregating process of the
updated gradient values. Hence, for the potential directions,
we put forward the role of the aggregation server to be
empowered by the blockchain approach that does not rely on
a single third party to manage a transaction.
_B. CHALLENGES AND CONSIDERATIONS_
_IN COLLABORATIVE LEARNING WITH_
_BLOCKCHAIN-BASED INCENTIVE_
The objective of consolidating collaborative learning and
blockchain-based incentives is to enable multiple users at different geographical locations to improve an AI model with a
reasonable incentive for the contributing users. Fundamental
schemes of collaborative learning with smart contracts can
be directly implemented for general data that are deemed
insensitive to the users. Our proposed schemes empower
users to manage training activities confidentially, even for
private data. However, the following challenges should be
considered.
(i) The availability of end-users. Even though a commensurate incentive scheme has been designed to
-----
motivate users in conjointly building models (using their
resources), technical difficulties might still appear, especially on the users’ side. The user may not complete all
processes or violate defined protocols, causing failure in
the building of the model.
(ii) System heterogeneity. The capabilities of each device
might differ considerably in terms of storage, computing power, communication capabilities, CPU, memory,
network connectivity [51], and the battery level.
(iii) Costly communication. Collaborative learning potentially comprises an extensive number of devices. This
makes the communication and computing process
slower, more costly, and time-consuming [52] by several
orders of magnitude.
(iv) Smart contract adoption and learning curve. A highvolume stream of transaction records must be overcome
by the system. On-chain and off-chain transactions with
filters have various features that are suitable for adoption
in collaborative learning.
(v) Human readable execution [53]. Our objective is to
preserve privacy and avoid linkability of transactions,
yet the flow of the execution process is still readable
by the observer, for example, utilizing byte-code in the
EVM. Nevertheless, the value remains confidential.
**VII. CONCLUSION**
We have presented privacy-awareness in decentralized
approaches as plausible solutions to address the linkability concerns in collaborative learning and Ethereum smart
contracts. These points become essential because the existing schemes provide various privacy techniques, yet the
linkability issues within the systems are beyond the focus.
Hence, we design supplementary protocols to be adopted
within collaborative learning and blockchain smart contracts.
We have completed the main requirements in this research,
such as providing sustainable private learning activities, compatible decentralized incentives with unlinkable and untraceable transactions. We have also shown that our schemes can
eliminate the worries of observers’ knowledge in associating
clients’ resources with their devices’ identity by obscuring the
transaction values that can only be recognized by a legitimate
party. Finally, the overall results positively recommend that
our schemes satisfy the design goals.
Apart from the merits of the given scheme, the role of
the centralized aggregation server in computing the gradient
values is another interest in the long run. The aggregation
server is likely to suffer from bottleneck issues and become an
SPoF that is inherent to the centralized approach. Therefore,
in the near future, we emphasize the replacement of centralized aggregation servers with distributed computing parties
based on blockchain technology.
**REFERENCES**
[1] A. Koloskova, S. Stich, and M. Jaggi, ‘‘Decentralized stochastic optimization and gossip algorithms with compressed communication,’’ in Proc. Int.
_Conf. Mach. Learn., 2019, pp. 3478–3487._
[2] S. Nakamoto, ‘‘Bitcoin: A peer-to-peer electronic cash system,’’ Manubot,
Madison, WI, USA, Tech. Rep., 2019.
[3] J. Konečný, H. B. McMahan, F. X. Yu, P. Richtárik, A. T. Suresh,
and D. Bacon, ‘‘Federated learning: Strategies for improving communication efficiency,’’ 2016, arXiv:1610.05492. [Online]. Available:
http://arxiv.org/abs/1610.05492
[4] P. Garcia, ‘‘Biometrics on the blockchain,’’ Biometric Technol. Today,
vol. 2018, no. 5, pp. 5–7, May 2018.
[5] V. Buterin, ‘‘Ethereum white paper,’’ GitHub Repository, vol. 1, pp. 22–23,
2013.
[6] B. K. Mohanta, D. Jena, S. S. Panda, and S. Sobhanayak, ‘‘Blockchain
technology: A survey on applications and security privacy challenges,’’
_Internet Things, vol. 8, Dec. 2019, Art. no. 100107._
[7] A. N. Bhagoji, S. Chakraborty, P. Mittal, and S. Calo, ‘‘Analyzing federated
learning through an adversarial lens,’’ in Proc. Int. Conf. Mach. Learn.,
2019, pp. 634–643.
[8] L. Melis, C. Song, E. De Cristofaro, and V. Shmatikov, ‘‘Exploiting unintended feature leakage in collaborative learning,’’ in Proc. IEEE Symp.
_Secur. Privacy (SP), May 2019, pp. 691–706._
[9] M. D. Mansh, A. Nguyen, and K. A. Katz, ‘‘Improving dermatologic care
for sexual and gender minority patients through routine sexual orientation
and gender identity data collection,’’ JAMA Dermatol., vol. 155, no. 2,
pp. 145–146, 2019.
[10] A. Mackenzie, S. Noether, and M. C. Team, ‘‘Improving obfuscation in the
cryptonote protocol,’’ Monero Res. Lab, Tech. Rep., 2015.
[11] J. Kang, Z. Xiong, D. Niyato, S. Xie, and J. Zhang, ‘‘Incentive mechanism
for reliable federated learning: A joint optimization approach to combining
reputation and contract theory,’’ IEEE Internet Things J., vol. 6, no. 6,
pp. 10700–10714, Dec. 2019.
[12] S. Awan, F. Li, B. Luo, and M. Liu, ‘‘Poster: A reliable and accountable
privacy-preserving federated learning framework using the blockchain,’’
in Proc. Conf. Comput. Commun. Secur. (ACM SIGSAC), Nov. 2019,
pp. 2561–2563.
[13] J. Weng, J. Weng, J. Zhang, M. Li, Y. Zhang, and W. Luo, ‘‘DeepChain:
Auditable and privacy-preserving deep learning with blockchain-based
incentive,’’ IEEE Trans. Dependable Secure Comput., early access,
[Nov. 8, 2019, doi: 10.1109/TDSC.2019.2952332.](http://dx.doi.org/10.1109/TDSC.2019.2952332)
[14] X. Bao, C. Su, Y. Xiong, W. Huang, and Y. Hu, ‘‘FLChain: A blockchain
for auditable federated learning with trust and incentive,’’ in Proc. 5th Int.
_Conf. Big Data Comput. Commun. (BIGCOM), Aug. 2019, pp. 151–159._
[15] P. K. Sharma, J. H. Park, and K. Cho, ‘‘Blockchain and federated learningbased distributed computing defence framework for sustainable society,’’
_Sustain. Cities Soc., vol. 59, Aug. 2020, Art. no. 102220._
[16] Y. Qu, L. Gao, T. H. Luan, Y. Xiang, S. Yu, B. Li, and G. Zheng,
‘‘Decentralized privacy using blockchain-enabled federated learning in
fog computing,’’ IEEE Internet Things J., vol. 7, no. 6, pp. 5171–5183,
Jun. 2020.
[17] Z. Li, J. Liu, J. Hao, H. Wang, and M. Xian, ‘‘CrowdSFL: A secure
crowd computing framework based on blockchain and federated learning,’’
_Electronics, vol. 9, no. 5, p. 773, May 2020._
[18] M. Shayan, C. Fung, C. J. M. Yoon, and I. Beschastnikh, ‘‘Biscotti:
A ledger for private and secure peer-to-peer machine learning,’’ 2018,
_arXiv:1811.09904. [Online]. Available: http://arxiv.org/abs/1811.09904_
[19] C. Fung, C. J. M. Yoon, and I. Beschastnikh, ‘‘Mitigating sybils in federated learning poisoning,’’ 2018, arXiv:1808.04866. [Online]. Available:
http://arxiv.org/abs/1808.04866
[20] E. Bagdasaryan, A. Veit, Y. Hua, D. Estrin, and V. Shmatikov, ‘‘How to
backdoor federated learning,’’ in Proc. Int. Conf. Artif. Intell. Statist., 2020,
pp. 2938–2948.
[21] S. Savazzi, M. Nicoli, and V. Rampa, ‘‘Federated learning with cooperating
devices: A consensus approach for massive IoT networks,’’ IEEE Internet
_Things J., vol. 7, no. 5, pp. 4641–4654, May 2020._
[22] S. A. Osia, A. S. Shamsabadi, A. Taheri, K. Katevas, H. R. Rabiee,
N. D. Lane, and H. Haddadi, ‘‘Privacy-preserving deep inference for rich
user data on the cloud,’’ 2017, arXiv:1710.01727. [Online]. Available:
http://arxiv.org/abs/1710.01727
[23] D. Sangeetha and V. Vaidehi, ‘‘A secure cloud based personal health record
framework for a multi owner environment,’’ Ann. Telecommun., vol. 72,
nos. 1–2, pp. 95–104, Feb. 2017.
[24] S. Dhar, J. Guo, J. Liu, S. Tripathi, U. Kurup, and M. Shah, ‘‘On-device
machine learning: An algorithms and learning theory perspective,’’ 2019,
_arXiv:1911.00623. [Online]. Available: http://arxiv.org/abs/1911.00623_
-----
[25] Q. Xu, Z. Su, S. Yu, and Y. Wang, ‘‘Trust based incentive scheme to allocate
big data tasks with mobile social cloud,’’ IEEE Trans. Big Data, early
[access, Oct. 23, 2017, doi: 10.1109/TBDATA.2017.2764925.](http://dx.doi.org/10.1109/TBDATA.2017.2764925)
[26] E. Kuada and H. Olesen, ‘‘Incentive mechanisms for opportunistic cloud
computing services,’’ in Proc. 8th IEEE Int. Conf. Collaborative Comput.,
_Netw., Appl. Worksharing, Oct. 2012, pp. 127–136._
[27] S. Zou, J. Xi, S. Wang, Y. Lu, and G. Xu, ‘‘Reportcoin: A novel blockchainbased incentive anonymous reporting system,’’ IEEE Access, vol. 7,
pp. 65544–65559, 2019.
[28] Y. Wang, Z. Su, and N. Zhang, ‘‘BSIS: Blockchain-based secure incentive
scheme for energy delivery in vehicular energy network,’’ IEEE Trans. Ind.
_Informat., vol. 15, no. 6, pp. 3620–3631, Jun. 2019._
[29] B. Jia, T. Zhou, W. Li, Z. Liu, and J. Zhang, ‘‘A blockchain-based location privacy protection incentive mechanism in crowd sensing networks,’’
_Sensors, vol. 18, no. 11, p. 3894, Nov. 2018._
[30] L. Li, J. Liu, L. Cheng, S. Qiu, W. Wang, X. Zhang, and Z. Zhang,
‘‘CreditCoin: A privacy-preserving blockchain-based incentive announcement network for communications of smart vehicles,’’ IEEE Trans. Intell.
_Transp. Syst., vol. 19, no. 7, pp. 2204–2220, Jul. 2018._
[31] K. Bonawitz, H. Eichner, W. Grieskamp, D. Huba, A. Ingerman, V. Ivanov,
C. Kiddon, J. Konečný, S. Mazzocchi, H. B. McMahan, T. Van Overveldt,
D. Petrou, D. Ramage, and J. Roselander, ‘‘Towards federated learning
at scale: System design,’’ 2019, arXiv:1902.01046. [Online]. Available:
http://arxiv.org/abs/1902.01046
[32] L. U. Khan, S. R. Pandey, N. H. Tran, W. Saad, Z. Han, M. N. H. Nguyen,
and C. S. Hong, ‘‘Federated learning for edge networks: Resource optimization and incentive mechanism,’’ 2019, arXiv:1911.05642. [Online].
Available: http://arxiv.org/abs/1911.05642
[33] Y. Qu, S. R. Pokhrel, S. Garg, L. Gao, and Y. Xiang, ‘‘A blockchained
federated learning framework for cognitive computing in industry 4.0
networks,’’ IEEE Trans. Ind. Informat., vol. 17, no. 4, pp. 2964–2973,
Apr. 2021.
[34] S. Rahmadika, S. Noh, K. Lee, B. J. Kweka, and K.-H. Rhee,
‘‘The dilemma of parameterizing propagation time in blockchain P2P
network,’’ J. Inf. Process. Syst., vol. 16, no. 3, pp. 699–717, 2020.
[35] B. Developers. Difficulty in Bitcoin (BTC). Accessed: Aug. 5, 2020.
[Online]. Available: https://btc.com/stats/diff
[36] K. Sharad, G. Karame, and G. A. Marson, ‘‘System for secure federated
learning,’’ US Patent App. 16 296 380, Sep. 10, 2020.
[37] Y. Lu, X. Huang, Y. Dai, S. Maharjan, and Y. Zhang, ‘‘Blockchain and
federated learning for privacy-preserved data sharing in industrial IoT,’’
_IEEE Trans. Ind. Informat., vol. 16, no. 6, pp. 4177–4186, Jun. 2020._
[38] S. Noether, S. Noether, and A. Mackenzie, ‘‘A note on chain reactions in
traceability in cryptonote 2.0,’’ Res. Bull. MRL-0001. Monero Res. Lab,
vol. 1, pp. 1–8, Sep. 2014.
[39] L. Deng, ‘‘The MNIST database of handwritten digit images for machine
learning research [best of the Web],’’ IEEE Signal Process. Mag., vol. 29,
no. 6, pp. 141–142, Nov. 2012.
[40] B. Barney, ‘‘Introduction to parallel computing,’’ Lawrence Livermore Nat.
_Lab., vol. 6, no. 13, p. 10, Dec. 2010._
[41] T. H. Yuen, J. K. Liu, M. H. Au, W. Susilo, and J. Zhou, ‘‘Efficient
linkable and/or threshold ring signature without random oracles,’’ Comput.
_J., vol. 56, no. 4, pp. 407–421, Apr. 2013._
[42] F. Zhang and K. Kim, ‘‘ID-based blind signature and ring signature from
pairings,’’ in Proc. Int. Conf. Theory Appl. Cryptol. Inf. Secur. Berlin,
Germany: Springer, 2002, pp. 533–547.
[43] C. A. Melchor, P.-L. Cayrel, P. Gaborit, and F. Laguillaumie, ‘‘A new
efficient threshold ring signature scheme based on coding theory,’’ IEEE
_Trans. Inf. Theory, vol. 57, no. 7, pp. 4833–4842, Jul. 2011._
[44] S. Rahmadika and K.-H. Rhee, ‘‘Reliable collaborative learning with
commensurate incentive schemes,’’ in Proc. IEEE Int. Conf. Blockchain
_(Blockchain), Nov. 2020, pp. 496–502._
[45] Y. Liu, Z. Ai, S. Sun, S. Zhang, Z. Liu, and H. Yu, ‘‘Fedcoin: A peerto-peer payment system for federated learning,’’ in Federated Learning.
Cham, Switzerland: Springer, 2020, pp. 125–138.
[46] S. Fan, H. Zhang, Y. Zeng, and W. Cai, ‘‘Hybrid blockchain-based resource
trading system for federated learning in edge computing,’’ IEEE Internet
_Things J., vol. 8, no. 4, pp. 2252–2264, Feb. 2021._
[47] L. U. Khan, S. R. Pandey, N. H. Tran, W. Saad, Z. Han, M. N. H. Nguyen,
and C. S. Hong, ‘‘Federated learning for edge networks: Resource optimization and incentive mechanism,’’ IEEE Commun. Mag., vol. 58, no. 10,
pp. 88–93, Oct. 2020.
[48] S. Rahmadika and K.-H. Rhee, ‘‘Merging collaborative learning and
blockchain: Privacy in context,’’ in Proc. Korea Inf. Process. Soc.
_Conf. Seoul, South Korea: Korea Information Processing Society, 2020,_
pp. 228–230.
[49] C. Song, T. Ristenpart, and V. Shmatikov, ‘‘Machine learning models that
remember too much,’’ in Proc. ACM SIGSAC Conf. Comput. Commun.
_Secur., Oct. 2017, pp. 587–601._
[50] S. Truex, N. Baracaldo, A. Anwar, T. Steinke, H. Ludwig, R. Zhang, and
Y. Zhou, ‘‘A hybrid approach to privacy-preserving federated learning,’’ in
_Proc. 12th ACM Workshop Artif. Intell. Secur. (AISec), 2019, pp. 1–11._
[51] T. Li, A. K. Sahu, A. Talwalkar, and V. Smith, ‘‘Federated learning:
Challenges, methods, and future directions,’’ IEEE Signal Process. Mag.,
vol. 37, no. 3, pp. 50–60, May 2020.
[52] J. Mills, J. Hu, and G. Min, ‘‘Communication-efficient federated learning
for wireless edge intelligence in IoT,’’ IEEE Internet Things J., vol. 7, no. 7,
pp. 5986–5994, Jul. 2020.
[53] Z. Zheng, S. Xie, H.-N. Dai, W. Chen, X. Chen, J. Weng, and M. Imran,
‘‘An overview on smart contracts: Challenges, advances and platforms,’’
_Future Gener. Comput. Syst., vol. 105, pp. 475–491, Apr. 2020._
SANDI RAHMADIKA received the dual master’s degree in engineering from Institut Teknologi
Bandung (ITB), Indonesia, and Pukyong National
University (PKNU), South Korea, in 2016, where
he is currently pursuing the Ph.D. degree with
the Laboratory of Information Security and Internet Applications (LISIA). His research interests
include applied cryptography, privacy preservation in the decentralized systems, and AI with
blockchain integration.
KYUNG-HYUNE RHEE (Member, IEEE) received the M.S. and Ph.D. degrees from the Korea
Advanced Institute of Science and Technology
(KAIST), South Korea, in 1985 and 1992, respectively. He worked as a Senior Researcher with
the Electronic and Telecommunications Research
Institute (ETRI), South Korea, from 1985 to 1993.
He also worked as a Visiting Scholar with The University of Adelaide, The University of Tokyo, and
the University of California, Irvine. He has served
as the Chairman of the Division of Information and Communication Technology, Colombo Plan Staff College for Technician Education in Manila,
Philippines. He is currently a Professor with the Department of IT Convergence and Application Engineering, Pukyong National University, South
Korea. His research interests include security and evaluation of blockchain
technology, key management and its applications, and AI-enabled security
evaluation of cryptographic algorithms.
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.1109/ACCESS.2021.3076205?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/ACCESS.2021.3076205, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "GOLD",
"url": "https://ieeexplore.ieee.org/ielx7/6287639/9312710/09417207.pdf"
}
| 2,021
|
[
"JournalArticle"
] | true
| null |
[
{
"paperId": "a208331a249f082c6b19076754dad9cb129e12aa",
"title": "A Blockchained Federated Learning Framework for Cognitive Computing in Industry 4.0 Networks"
},
{
"paperId": "f6d4d82357206a36ec3423314bae65913896bc1c",
"title": "Hybrid Blockchain-Based Resource Trading System for Federated Learning in Edge Computing"
},
{
"paperId": "e027677ee8e1c12b2da1a662309ad53ce2a994f8",
"title": "Reliable Collaborative Learning with Commensurate Incentive Schemes"
},
{
"paperId": "ba0bdd244e57f941be9d190e5a8d46785d7e2e1a",
"title": "Blockchain and federated learning-based distributed computing defence framework for sustainable society"
},
{
"paperId": "4f9477b4e894e05b6570e4ad7da82420085eb8c0",
"title": "Communication-Efficient Federated Learning for Wireless Edge Intelligence in IoT"
},
{
"paperId": "2a367204cc142f4a810ee5bdcc8c55dd4d2fc40e",
"title": "The Dilemma of Parameterizing Propagation Time in Blockchain P2P Network"
},
{
"paperId": "d792ce75ae10d0534cada7fb9c8d6ef316e35a9f",
"title": "Blockchain and Federated Learning for Privacy-Preserved Data Sharing in Industrial IoT"
},
{
"paperId": "dd62be03be2b747c9f09672a056d14a05c7c3e54",
"title": "CrowdSFL: A Secure Crowd Computing Framework Based on Blockchain and Federated Learning"
},
{
"paperId": "d2da19b2ae5af19df02c27862f403883a71c82f5",
"title": "Decentralized Privacy Using Blockchain-Enabled Federated Learning in Fog Computing"
},
{
"paperId": "2865e822f43ca92125fd2a19e5f15b946539419c",
"title": "FedCoin: A Peer-to-Peer Payment System for Federated Learning"
},
{
"paperId": "0d7e26c623068f7119878f12f5ee1a49b20b9c9d",
"title": "Federated Learning With Cooperating Devices: A Consensus Approach for Massive IoT Networks"
},
{
"paperId": "e0bc89f5776804bc2be27f1945f900d1ac8f1e7f",
"title": "An Overview on Smart Contracts: Challenges, Advances and Platforms"
},
{
"paperId": "75c91d2cf7e926de17bb0c9f501d4183ddf22dc6",
"title": "Blockchain technology: A survey on applications and security privacy Challenges"
},
{
"paperId": "07f229cc6e5b80fc8ffbbb3e4db85142466f55a9",
"title": "DeepChain: Auditable and Privacy-Preserving Deep Learning with Blockchain-Based Incentive"
},
{
"paperId": "2a3d09bbdfe21418ce75d6973f71028fa9192b89",
"title": "Federated Learning for Edge Networks: Resource Optimization and Incentive Mechanism"
},
{
"paperId": "503d39b221099bed1f583585e94e0087bf7656f3",
"title": "Poster: A Reliable and Accountable Privacy-Preserving Federated Learning Framework using the Blockchain"
},
{
"paperId": "6d8db61577a021981ccf58d0c5ced167674b389c",
"title": "On-Device Machine Learning: An Algorithms and Learning Theory Perspective"
},
{
"paperId": "9251ad8137f9aea4596aeccc656f351cfeced551",
"title": "Incentive Mechanism for Reliable Federated Learning: A Joint Optimization Approach to Combining Reputation and Contract Theory"
},
{
"paperId": "49bdeb07b045dd77f0bfe2b44436608770235a23",
"title": "Federated Learning: Challenges, Methods, and Future Directions"
},
{
"paperId": "f7c20aef687e3157596e9d56c10c43ee9cb2ef38",
"title": "FLChain: A Blockchain for Auditable Federated Learning with Trust and Incentive"
},
{
"paperId": "5643004ceb08739e657977a4bc105d50f5bfe29b",
"title": "Reportcoin: A Novel Blockchain-Based Incentive Anonymous Reporting System"
},
{
"paperId": "b59b37f21261e008ecb3ad8f0658ee790b42ea7e",
"title": "BSIS: Blockchain-Based Secure Incentive Scheme for Energy Delivery in Vehicular Energy Network"
},
{
"paperId": "79cf9462a583e1889781868cbf8c31e43b36dd2f",
"title": "Towards Federated Learning at Scale: System Design"
},
{
"paperId": "44b3b3bb40a9055eccdf86ea1702f6ae8b38934c",
"title": "Decentralized Stochastic Optimization and Gossip Algorithms with Compressed Communication"
},
{
"paperId": "827e8d3891dc6e6c0cc868e9161c867c3fa8868f",
"title": "Improving Dermatologic Care for Sexual and Gender Minority Patients Through Routine Sexual Orientation and Gender Identity Data Collection."
},
{
"paperId": "67498fdf77fd036a09a4593c37b012d6cf34f3f6",
"title": "A Hybrid Approach to Privacy-Preserving Federated Learning"
},
{
"paperId": "6c66108edb9af0533309055e7b2ecb8922db03d8",
"title": "Analyzing Federated Learning through an Adversarial Lens"
},
{
"paperId": "1df492149bce34a88c60fa19e01c25c77e3733af",
"title": "Biscotti: A Ledger for Private and Secure Peer-to-Peer Machine Learning"
},
{
"paperId": "a5e0e841f79ea97ca32a10c302b9d14f26a8272a",
"title": "A Blockchain-Based Location Privacy Protection Incentive Mechanism in Crowd Sensing Networks"
},
{
"paperId": "333420606f059a7d5574a6fb9e35591346d3f957",
"title": "Mitigating Sybils in Federated Learning Poisoning"
},
{
"paperId": "14d8b4fdb0262c30ae9afe20ea8e7227b115c63e",
"title": "How To Backdoor Federated Learning"
},
{
"paperId": "30e0ffeb519a4df2d4a2067e899c5fb5c5e85e70",
"title": "Exploiting Unintended Feature Leakage in Collaborative Learning"
},
{
"paperId": "ff48135f2681a2f5ed596a8e050080a50402c467",
"title": "Biometrics on the blockchain"
},
{
"paperId": "19dc966f0fb70abffaea682b81d3554c15551fda",
"title": "CreditCoin: A Privacy-Preserving Blockchain-Based Incentive Announcement Network for Communications of Smart Vehicles"
},
{
"paperId": "a56f5aa03b1058158a71432b79d2e595bcac4408",
"title": "Trust Based Incentive Scheme to Allocate Big Data Tasks with Mobile Social Cloud"
},
{
"paperId": "6cefb70f4668ee6c0bf0c18ea36fd49dd60e8365",
"title": "Privacy-Preserving Deep Inference for Rich User Data on The Cloud"
},
{
"paperId": "18cfd4b9e35fb12fbebedb0fdc3f7811090372bf",
"title": "Machine Learning Models that Remember Too Much"
},
{
"paperId": "7a26bd8a284b027a3783f113782cdc7b9eb1504d",
"title": "Introduction to Parallel Computing"
},
{
"paperId": "7fcb90f68529cbfab49f471b54719ded7528d0ef",
"title": "Federated Learning: Strategies for Improving Communication Efficiency"
},
{
"paperId": "f913c3a47106b51ca6566dcc37b2e52c0d549dc9",
"title": "A secure cloud based Personal Health Record framework for a multi owner environment"
},
{
"paperId": "97ed4d9379ba1008bb1a0dacdc76f9df9f4079be",
"title": "Efficient Linkable and/or Threshold Ring Signature Without Random Oracles"
},
{
"paperId": "46f74231b9afeb0c290d6d550043c55045284e5f",
"title": "The MNIST Database of Handwritten Digit Images for Machine Learning Research [Best of the Web]"
},
{
"paperId": "63f68e7c6a41c89972c9dcb7a598586f2dfc1f06",
"title": "Incentive mechanisms for Opportunistic Cloud Computing Services"
},
{
"paperId": "8997d7151509eb0534dcc654b159921c3d8dc39c",
"title": "A New Efficient Threshold Ring Signature Scheme Based on Coding Theory"
},
{
"paperId": "5475bd4f5a58f6bf171243e72e0eccb8eb3d0506",
"title": "ID-Based Blind Signature and Ring Signature from Pairings"
},
{
"paperId": null,
"title": "‘‘Merging collaborative learning and blockchain: Privacy in context"
},
{
"paperId": null,
"title": "Difficulty in Bitcoin (BTC)"
},
{
"paperId": "5917a3dfa83f8aff6a10539da236b15fe06956da",
"title": "MRL-0004 Improving Obfuscation in the CryptoNote Protocol"
},
{
"paperId": null,
"title": "‘‘A note on chain reactions in traceability in cryptonote 2.0,’’"
},
{
"paperId": null,
"title": "Ethereum white paper"
},
{
"paperId": "4e9ec92a90c5d571d2f1d496f8df01f0a8f38596",
"title": "Bitcoin: A Peer-to-Peer Electronic Cash System"
},
{
"paperId": null,
"title": "South Korea, in 2016, where he is currently pursuing the Ph.D. degree with the Laboratory of Information Security and Internet Applications (LISIA)"
},
{
"paperId": null,
"title": "ii) System heterogeneity"
},
{
"paperId": null,
"title": "Unlinkable CL Transactions: Privacy-Awareness in Decentralized Approaches"
},
{
"paperId": null,
"title": "Accessed"
},
{
"paperId": null,
"title": "iii) We introduce unlinkable rewarding and training activities without revealing the information values of transactions"
}
] | 20,245
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Engineering",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/029a9046f114af95c0ec83e13e144704b7f77868
|
[
"Computer Science"
] | 0.885171
|
A survey on parallel and distributed multi-agent systems for high performance computing simulations
|
029a9046f114af95c0ec83e13e144704b7f77868
|
Computer Science Review
|
[
{
"authorId": "4747047",
"name": "A. Rousset"
},
{
"authorId": "2501805",
"name": "B. Herrmann"
},
{
"authorId": "1910857",
"name": "Christophe Lang"
},
{
"authorId": "2005060",
"name": "L. Philippe"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"Comput Sci Rev"
],
"alternate_urls": null,
"id": "3aa92b7f-af7a-4ebd-8925-1152710bfbc7",
"issn": "1574-0137",
"name": "Computer Science Review",
"type": "journal",
"url": "http://www.elsevier.com/wps/find/journaldescription.cws_home/710138/description#description"
}
| null |
# A survey on parallel and distributed Multi-Agent Systems
## Alban Rousset, Bénédicte Herrmann, Christophe Lang, Laurent Philippe
To cite this version:
Alban Rousset, Bénédicte Herrmann, Christophe Lang, Laurent Philippe. A survey on parallel and
distributed Multi-Agent Systems. Padabs 2014, 2nd Workshop on Parallel and Distributed AgentBased Simulations, in conjunction with Euro-Par 2014, 2014, Porto, Portugal. pp.371–382. hal01230768
## HAL Id: hal-01230768
https://hal.science/hal-01230768
Submitted on 19 Nov 2015
**HAL is a multi-disciplinary open access**
archive for the deposit and dissemination of scientific research documents, whether they are published or not. The documents may come from
teaching and research institutions in France or
abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est
destinée au dépôt et à la diffusion de documents
scientifiques de niveau recherche, publiés ou non,
émanant des établissements d’enseignement et de
recherche français ou étrangers, des laboratoires
publics ou privés.
-----
# A survey on parallel and distributed Multi-Agent Systems
### Alban ROUSSET[∗][1], Bénédicte Herrmann[†][1], Christophe LANG[‡][1], and Laurent PHILIPPE[§][1]
1Femto-ST Institute – University of Franche-Comté, 16 Route de
### Gray 25030 Besançon cedex - France
**Abstract**
Simulation has become an indispensable tool for researchers to explore systems without having recourse to real experiments. Depending
on the characteristics of the modeled system, methods used to represent the system may vary. Multi-agent systems are, thus, often used
to model and simulate complex systems. Whatever modeling type
used, increasing the size and the precision of the model increases the
amount of computation, requiring the use of parallel systems when it
becomes too large. In this paper, we focus on parallel platforms that
support multi-agent simulations. Our contribution is a survey on existing platforms and their evaluation in the context of high performance
computing. We present a qualitative analysis, mainly based on platform properties, then a performance comparison using the same agent
model implemented on each platform.
## 1 Introduction
In the field of simulation, we often seek to exceed limits, that is to say analyse
larger and more precise models to be closer to the reality of a problem.
Increasing the size of a model has however a direct impact on the amount
of needed computing resources and centralised systems are often no longer
sufficient to run these simulations. The use of parallel resources allows us to
overcome the resource limits of centralised systems and also to increase the
size of the simulated models.
_∗alban.rousset@femto-st.fr_
_†bherrman@femto-st.fr_
_‡clang@femto-st.fr_
_§lphilipp@femto-st.fr_
1
-----
There are several ways to model a system. For example, the time behavior of a large number of physical systems is based on differential equations.
In this case the discretization of a model allows its representation as a linear
system. It is then possible to use existing parallel libraries to take advantage
of many computing nodes and run large simulations. On the other hand it
is not always possible to model any time dependent system with differential
equations. This is for instance the case of complex systems. A complex
system is defined in [25] as "A system that can be analyzed into many com_ponents having relatively many relations among them, so that the behavior_
_of each component depends on the behavior of others". Thus the complexity_
of the dependencies between the phenomena that drive the entities behavior
makes it difficult to define a global law that models the entire system. For
this reason multi-agent systems are often used to model complex systems
because they rely on an algorithmic description of agents that interact and
simulate the expected behavior. From the viewpoint of increasing the size of
simulations, multi-agent systems are constrained to the same rules as other
modelling techniques but there exists less support for parallel execution of
the models.
In this article, we focus on multi-agent platforms that provide parallel
distributed programming environments for multi-agent systems. Recently,
the interest for parallel multi-agent platforms has increased. This is because
parallel platforms offer more resources to run larger agent simulations and
thus allows to obtain results or behavior that was not possible to obtain
with smaller number of agents (eg. simulation of individual motions in a
city/urban mobility).
The contribution of this article is a survey on parallel distributed multiagent platforms. This survey is based on an extensive bibliographical work
done to identify the existing platforms, a qualitative analysis of these platforms in terms of ease of development, distribution management or proposed
agent model, and a performance evaluation based on a representative model
run on a HPC cluster.
The article is organised as follows. First, we give the context of multiagent system (MAS) in general and parallel distributed multi-agent systems
(PDMAS) in particular. We then introduce the different multi-agent platforms found in our bibliographical research. In the third section, we describe
the method used to classify platforms and we describe the model implemented in each platform to evaluate its performance. In the fourth section,
we present the qualitative comparison of the different PDMAS followed by
the benchmark based on the implemented model. We finish the paper with
conclusion and future work.
2
-----
## 2 Related works
The concept of agent has been studied extensively for several years and in
different domains. It is not only used in robotics and other fields of artificial
intelligence, but also in fields such as psychology [6] or biology [23]. One of
the first definitions of the agent concept is due to Ferber [13] :
"An agent is a real or virtual autonomous entity, operating
in a environment, able to perceive and act on it, which can communicate with other agents, which exhibits an independent behavior, which can be seen as the consequence of his knowledge,
its interactions with other agents and goals it need to achieved".
A multi-agent system, or MAS, is a platform that provides the mandatory
support to run simulations based on several autonomous agents. These
platforms implement functions that provide services such as agent life cycle management, communication between agents, agent perception or environment management. Among well known platforms we can cite Repast
Simphony [21], Mason [19], NetLogo [28] and Gama [1]. These platforms
however do not natively implement a support to run models in parallel and
it is necessary to develop a wrapper from scratch, in order to distribute or
parallelize a simulation. There exists several papers that propose survey on
these multi-agent platforms [29, 5, 4, 16].
Some platforms like RepastHPC [10], D-Mason [12], Pandora [2], Flame
[8] or JADE [3] provide a native support for parallel execution of models.
This support usually includes the collaboration between executions on several physical nodes, the distribution of agents between nodes and so on.
During our analysis of the literature, we did not find any survey about parallel multi-agent platforms except the paper written by Coakley and al. [8].
This comparison is based on qualitative criteria such as the implementation
language but the paper does not provide any performance comparison of the
studied platforms.
After an extensive bibliographical work, we identified 10 implementations
or projects of parallel multi-agent platforms. For each platform we tried
to download the source or executable code and we tried to compile it and
test it with the provided examples and templates. Some of the platforms
cannot be included in our study because there is no available source code
or downloadable executable (MACE3J [15], JAMES[17], SWAGES [24]), or
because only a demonstration version is available (PDES-MAS [22, 27]), or
because there is a real lack of documentation (Ecolab [26]). It was thus not
possible to build a new model in these platforms and thus to assess their
parallel characteristics and performance. These platforms have subjected to
a qualitative analysis which is not included in this paper.
For the 5 remaining platforms, on which we were able to implement our
model, we can consider that they truly offer a functioning parallel multi-agent
3
-----
support. We succinctly present each of these platforms in the following.
**D-Mason (Distributed Mason) [12] is developed by the University of**
Salerno. D-Mason is the distributed version of the Mason multi-agent platform. The authors choose to develop a distributed version of Mason to
provide a solution that does not require users to rewrite their already developed simulations and also to overcome the limitations on maximum number
of agents. D-Mason uses ActiveMQ JMS as a base to implement communications. D-Mason uses the Java language to implement the agent model.
**Flame [8] is developed by the University of Sheffield. Flame was designed**
to allow a wide range of agent models. Flame provides specifications in the
form of a formal framework that can be used by developers to create models
and tools. Flame allows parallelization using MPI. Implementing a Flame
simulation is based on the definition of X-Machines [9] which are defined as
finite state automata with memory. In addition, agents can send and receive
messages at the input and the output of each state.
**Jade [3] is developed by the Telecom laboratory of Italia. The aims of**
Jade are to simplify the implementation of distributed multi-agent models
across a FIPA compliant [3] middleware and to provide a set of tools that
support the debugging and the deployment phases. The platform can be distributed across multiple computers and its configuration can be controlled
from a remote GUI. Agents are implemented in Java while the communications relay on the RMI library.
**Pandora [2] is developed by the Supercomputing center of Barcelona.**
It is explicitly programmed to allow the execution of scalable multi-agent
simulations. According to the literature, Pandora is able to treat thousands
of agents with complex actions. Pandora also provides a support for a geographic information system (GIS) in order to run simulations where spatial
coordinates are used. Pandora uses the C + + language to define and to implement the agent models. For the communications, Pandora automatically
generates MPI code from the Pandora library.
**RepastHPC [10] is developed by the Argone institute of USA. It is a**
part of a series of multi-agent simulation platforms: RepastJ and Repast
Simphony. RepastHPC is specially designed for high performance environments. RepastHPC use the same concepts as the core of RepastSimphony,
that is to say it uses also the concept of projections (grid, network) but this
concept is adapted to parallel environments. The C + + language is used
to implement an agent simulation but the ReLogo language, a derivative
of the NetLogo language, can also be used. For the communications, the
RepastHPC platform relays on MPI using the Boost library [11].
From these descriptions we can note that some platforms have already
been designed to target high performance computing systems such as clusters
whereas others are more focused on distribution on less coupled nodes such
as a network of workstations.
4
-----
## 3 Survey methodology
In this section we explain the methodology used to make this survey. As
already stated we started by a bibliographical search (using keywords on
search engines and following links cited in the studied articles). This bibliographical search allowed us to establish a first list of existing platforms.
By testing the available platforms we established a second list of functioning
platforms. To our knowledge this list is complete and their is no other available and functional platform that provide a support for parallel distributed
MAS. Note we only concentrate on distributed platforms and that the list
excludes shared memory parallel platforms and many-cores (as GPU or Intel
Xeon Phi) platforms. After we defined different criteria to compare and analyse each platform. We finished by implementing a reference model on each
platform and executing it in order to compare the platform performance.
These evaluation steps are detailed in the following.
This survey mainly focuses on the implementation, more precisely the
development, of models and their execution efficiency. To classify the platforms we defined two sets of criteria: first, implementation and execution
based criteria and, second, criteria about classical properties of parallel systems. We briefly explain in which correspond each criteria.
For the implementation and execution criteria, all platforms have their
own constraints that impact on the ease of the model implementation. The
chosen criteria are:
1. Programming language,
2. Agent representation
3. Simulation type, time-driven or event-driven
4. Reproductibility, do several executions of a simulation give the same
results?
For the classical properties of parallel systems, we focus on:
1. Scalability of platform, in terms of agents and nodes,
2. Load balancing, agent distribution,
3. MultiThread execution, to take benefit of multicore processors,
4. Communication library.
To further compare the platforms, we have defined a reference agent
model that we implemented on each platform. The reference model is based
on three important behaviors for each agent: the agent perception, the communications between agents and/or with the environment and agent mobility. The reference model simulates each of these behaviors.
5
-----
Figure 1: AML representation of the reference agent model
Figure 1 gives an AML [7] (Agent Modeling Language) representation
of our reference model. The Environment is represented by a square grid.
_Agents are mobile and move randomly on the grid. A vision characterised_
by the "radius" property is also associated with each agent. It represents
the limited perception of the agent on the environment.
Each agent is composed of 3 sub-behaviors :
1. The walk behavior allows agents to move in a random direction on the
environment. This behavior is used to test the mobility and the perception of the agents. As the agents walk through their environment to
discover other agents and other parts of the environment, interactions
and communications with the environment are also tested with this
behavior.
2. The interact behavior allows agents to interact and send messages to
other agents in their perception fields. This behavior intends to simulate communications between agents and to evaluate the communication support of the platforms.
3. The compute behavior allows agents to compute a "Fast Fourier Transform (FFT)" [14] in order to represent a workload. This behavior intends to simulate the load generated by the execution of the agent
inner algorithms.
The global agent behavior consists in performing each of this three behaviors at each time step. The reference model has several parameters that
determine the agent behavior and also the global model properties. For instance, the model allows to vary the workload using different sizes of input
for the FFT calculus. It is also possible to generate more or less communications between agents by setting the number of contacted agents in the
interact behavior or to assess the agent mobility by setting the agent speed
in the walk behavior.
6
-----
## 4 Qualitative analysis
In this section we expose two levels of comparisons between the studied platforms: first a qualitative comparison using the previously presented criteria
and second a performance comparison using the reference model.
Table 1 gives a synthetic representation of the comparison for the implementation and execution criteria. Most platforms use classical languages
such as C-C++ or Java to define agents, except the Flame platform which
uses the XMML language. The XMML language is an extension of the XML
language designed to define X-Machines. Note that the RepastHPC platform
implements, in addition to the C++ programming language, the widespread
Logo agent language. The Repast-Logo or R-Logo is the Repast implementation of Logo for C++. It allows to simplify the simulation implementation
at the price of a lower power of expression compared to C++.
RepastHPC D-Mason Flame Pandora Jade
Prog. lang. C++/R-Logo Java XMML/C C/C++ Java
Agent repre- Object Object X-Machine Object Object
sent.
Simu. type event-driven time-driven time-driven time-driven time-driven
ReproductibilityYes Yes No Yes No
Table 1: Comparison of implementation and execution properties
Agents are usually defined as objects with methods representing behaviors. An agent container gathers all the agents. This container is cut and
distributed in the case of parallel execution. The agent implementation is different for the Flame platform that does not use the object concept to define
a agent but rather uses automatas called X-Machines. In a X-Machine, a behavior is represented by a state in the automata and the order of execution
between behaviors are represented by transitions. This difference changes
the programming logic of a model but induces no limitation compared with
other platforms because agents are in fact encoded in C language.
For the simulation type, event or time driven, all platforms use the timedriven approach except RepastHPC which is based on the event-driven approach. RepastHPC however allows to fix a periodicity to each scheduled
event, so that we can reproduce the behavior of time-driven simulations.
Finally all platforms allow agents to communicate. This communication
can be performed either internally with agents that are on the same node,
or externally, with agents that are on different nodes. The D-Mason and
Pandora platforms propose remote method invocations to communicate with
other agents while the other platforms use messages to communicate between
agents.
Table 2 summarises the criteria of the platforms about classical properties
of parallel systems. Globally we can note that all studied platforms meet
the demands for the development of parallel simulations. Note that we did
7
|Col1|RepastHPC|D-Mason|Flame|Pandora|Jade|
|---|---|---|---|---|---|
|Prog. lang.|C++/R-Logo|Java|XMML/C|C/C++|Java|
|Agent repre- sent.|Object|Object|X-Machine|Object|Object|
|Simu. type|event-driven|time-driven|time-driven|time-driven|time-driven|
|Reproductibil|ityYes|Yes|No|Yes|No|
-----
not find any information on the scalability property of the Pandora and Jade
platforms, so they are marked as Not Available (NA) for this property. To
efficiency exploit the power of several nodes the computing load must be
balanced among them. There is different ways to balance the computing
load . The load can be balanced at the beginning of the simulation (Static)
or adapted during the execution (Dynamic). A dynamic load balancing is
usually better as it provides a better adaptation in case of load variation
during the model execution, but it can also be subject to instability. Most
platforms use dynamic load balancing except the Jade and Flame platforms.
In [20] the authors propose a way to use dynamic load balancing with the
Flame platform.
RepastHPC D-Mason Flame Pandora Jade
Scalability 1028 36 nodes [8] 432 proc. [8] NA NA
proc. [18]
Load Balancing Dynamic Dynamic Static [8] Dynamic Static [3]
Multithread exec Yes [8] Yes [12, 8] No [8] Yes Yes
Com. library MPI [11, 10] JMS [12] MPI [18] MPI [2] RMI
Table 2: Comparison classical properties of parallel systems
Note that only Flame does not support multi-threaded executions. The
platform however relays on the MPI messaging library. As most MPI libraries
provide optimised implementations of message passing functions when the
communicating processes are on the same node, using processes located on
the same node instead of threads does not lead to large overhead. In the
implementation of a multi-agent system this probably leads to equivalent
performance as the simplification of synchronisation issues may compensate
the cost of using communication functions.
Last, the communication support for most platforms is MPI. This is not
surprising for platforms targeting HPC systems as this library is mainly
used on these computers. Note that the D-Mason platform relays on the
JMS communication service despite it is not the most scalable solution for
distributed environments. An MPI version of D-MASON is in development.
Finally, the Jade platform is based on the java Remote Method Invocation
(RMI) library which is not very adapted to parallel applications as it is based
on synchronous calls. During the model implementation we also noted that
the Jade platform seems to be more oriented for equipment monitoring and
cannot be run on HPC computers due to internal limitations. Jade is thus
not included in the rest of the comparisons.
## 5 Performance evaluation
For the performance evaluation we have implemented the reference model defined in section 3 on the four functional platforms: RepastHPC, D-MASON,
8
|Col1|RepastHPC|D-Mason|Flame|Pandora|Jade|
|---|---|---|---|---|---|
|Scalability|1028 proc. [18]|36 nodes [8]|432 proc. [8]|NA|NA|
|Load Balancing|Dynamic|Dynamic|Static [8]|Dynamic|Static [3]|
|Multithread exec|Yes [8]|Yes [12, 8]|No [8]|Yes|Yes|
|Com. library|MPI [11, 10]|JMS [12]|MPI [18]|MPI [2]|RMI|
-----
Flame, Pandora. During this model implementation, we did not encounter
noticeable difficulties expect with the RepastHPC platform for which we
have not been able to implement external communications, communications
between agents running on different nodes. RepastHPC does not have the
native mechanisms to make it whereas it is possible to implement it on the
other platforms. RepastHPC actually offers the possibility to interact with
an agent on an other node but not to report the modifications.
Although we have been able to run the four platforms, D-Mason, Flame,
Pandora, RepastHPC, on a standard workstation, only two of them (RpastHPC,
Flame) have successfully run on our HPC system. The D-Mason platform
uses a graphical interface that cannot be disconnected. We are thus not able
to run D-MASON on our cluster, only accessible through its batch manager.
The Pandora simulations have deadlock problems even if we use examples
provided with the platform. For these reasons the presented results only
consider the Flame and RepastHPC platforms.
We have realised several executions in order to exhibit the platform behaviors concerning scalability (Figures 2 and 3) and workload (Figure 4).
To assess scalability we vary the number of nodes used to execute the simulations while we fix the number of agents. We then compute the obtained
speedup. For workload we fix the number of nodes to 8 and we vary the
number of agents in the simulation. Each execution is realised several times
to assess the standard variation and the presented results are the mean of
the different execution durations. Due to a low variation in the simulation
runtime, the number of executions for a result is set to 10.
60
40
20
0
0 50 100
number of cores
**Legend** Ideal speedUp Max speedUp Min speedUp
Figure 2: Scalability of FLAME simulations using 10 000 agents, FFT 100
and 200 cycles
About the HPC experimental settings, we have run the reference model
on a 764 cores cluster using the SGE batch system. Each node of the cluster
is a bi-processors, with Xeon E5 (8*2 cores) processors running at 2.6 Ghz
9
-----
Figure 3: Scalability of RepastHPC simulations using 10 000 agents, FFT
100 and 200 cycles
frequency and with 32 Go of memory. The nodes are connected through a
non blocking DDR infinyBand network organised in a fat tree. The system
is shared with other users but the batch system garanties that the processes
are run without sharing their cores.
Execution results for scalability for a model with 10 000 agents are given
on Figure 2 and 3, with the ideal speedup reference. Note that the reference
time used to compute the speedup is based on a two core run of the simulations. This is due to RepastHPC which cannot run on just one core so that
its reference time must be based on two core runs. The speedup is therefore
limited to half the number of nodes. We can note that both platforms scale
well up to 32 cores but the performance does not progress so well after, becoming 2/3 of the theoretical speedup for 128 cores. In addition on Figure 3
we can see that RepastHPC results are above the theorical speedup for simulations with less than 50 cores. As we suspected that these better results
come from cache optimizations in the system, we did more tests to confirm
this hypothesis. The realized tests increase the number of agents and the
load on each agent to saturate the cache and force memory accesses. As the
results for these new tests are under the theorical speedup the hypothesis is
validated.
Figures 4 represents the workload behavior of the two platforms. The inner load of agents (FFT) is here set to 100. The figure shows that RepastHPC
really better reacts to load increasing than Flame. The same behavior has
also been noted for a load of 10 (for 20 000 agents the ratio is 0.92). On the
opposite for a load of 1000 the difference is less noticeable (for 20 000 agents
the ratio is 0.81). Obviously the used model does not use all the power of
Flame as it is limited in term of inter-agent communications. The question
to answer is: is it due to the use of the concept of X-Machines or synchro
10
-----
Figure 4: Workload behavior for simulation using 8 cores
nisation mechanisms in the underlying parallelism? Another possible reason
that could justify this difference is the cost of the synchronisations provided
by Flame when using remote agents and that is not managed in RepastHPC.
## 6 Conclusion
In this article we have presented a comparison of different parallel multi-agent
platforms. This comparison is performed at two levels, first at a qualitative
level using criteria on the provided support, and second at a quantitative
level, using a reference agent model implementation. The qualitative comparison shows the properties of all the studied platforms. The quantitative
part shows an equivalent scalability for both platforms but better performance results for the RepasHPC platform.
When implementing our reference model we have noticed that the synchronisation support of the platforms does not provide the same level of
service: the RepastHPC platform does not provide communication support
for remote agents while Flame do it. This support seems to be a key point
in the platform performance.
For this reason, in our future work, we intend to better examine the
efficiency of synchronisation mechanisms in parallel platforms. For example
how are the synchronizations made during an execution and is there a way
to improve synchronization mechanisms in parallel multi-agent systems?
## Acknowledgment
Computations have been performed on the supercomputer facilities of the
Mésocentre de calcul de Franche-Comté.
11
-----
## References
[1] Edouard Amouroux, Thanh-Quang Chu, Alain Boucher, and Alexis
Drogoul. Gama: an environment for implementing and running spatially explicit multi-agent simulations. In Agent computing and multi_agent systems, pages 359–371. Springer, 2009._
[2] Elaini S Angelotti, Edson E Scalabrin, and Bráulio C Ávila. Pandora: a
multi-agent system using paraconsistent logic. In Computational Intel_ligence and Multimedia Applications, 2001. ICCIMA 2001., pages 352–_
356. IEEE, 2001.
[3] Fabio Bellifemine, Agostino Poggi, and Giovanni Rimassa. Jade–a
fipa-compliant agent framework. In Proceedings of PAAM, volume 99,
page 33. London, 1999.
[4] Matthew Berryman. Review of software platforms for agent based models. Technical report, DTIC Document, 2008.
[5] Rafael H Bordini, Lars Braubach, Mehdi Dastani, Amal El FallahSeghrouchni, Jorge J Gomez-Sanz, Joao Leite, Gregory MP O’Hare,
Alexander Pokahr, and Alessandro Ricci. A survey of programming
languages and platforms for multi-agent systems. Informatica (Slove_nia), 30(1):33–44, 2006._
[6] Gregory Carslaw. _Agent based modelling in social psychology._ PhD
thesis, University of Birmingham, 2013.
[7] Radovan Červenka, Ivan Trenčansk`y, Monique Calisti, and Dominic
Greenwood. Aml: Agent modeling language toward industry-grade
agent-based modeling. In Agent-Oriented Software Engineering V, pages
31–46. Springer, 2005.
[8] Simon Coakley, Marian Gheorghe, Mike Holcombe, Shawn Chin, David
Worth, and Chris Greenough. Exploitation of hpc in the flame agentbased simulation framework. In Proceedings of the 2012 IEEE 14th Int.
_Conf. on HPC and Communication & 2012 IEEE 9th Int. Conf. on Em-_
_bedded Software and Systems, HPCC ’12, pages 538–545, Washington,_
DC, USA, 2012. IEEE Computer Society.
[9] Simon Coakley, Rod Smallwood, and Mike Holcombe. Using x-machines
as a formal basis for describing agents in agent-based modelling. SIM_ULATION SERIES, 38(2):33, 2006._
[10] Nicholson Collier and Michael North. Repast HPC: A platform for large_scale agentbased modeling. Wiley, 2011._
[11] Nick Collier. Repast hpc manual, 2010.
12
-----
[12] Gennaro Cordasco, Rosario Chiara, Ada Mancuso, Dario Mazzeo, Vittorio Scarano, and Carmine Spagnuolo. A Framework for Distributing Agent-Based Simulations. In Euro-Par 2011: Parallel Processing
_Workshops, volume 7155 of Lecture Notes in Computer Science, pages_
460–470, 2011.
[13] Jacques Ferber and Jean-François Perrot. Les systèmes multi-agents:
_vers une intelligence collective. InterEditions Paris, 1995._
[14] Matteo Frigo and Steven G Johnson. The design and implementation
of fftw3. Proceedings of the IEEE, 93(2):216–231, 2005.
[15] Les Gasser and Kelvin Kakugawa. Mace3j: fast flexible distributed simulation of large, large-grain multi-agent systems. In Proceedings of the
_first inter. joint Conf. on Autonomous agents and multiagent systems:_
_part 2, pages 745–752. ACM, 2002._
[16] Brian Heath, Raymond Hill, and Frank Ciarallo. A survey of agentbased modeling practices (january 1998 to july 2008). JASSS, 12(4):9,
2009.
[17] Jan Himmelspach and Adelinde M. Uhrmacher. Plug’n simulate. In
_Proceedings of the 40th Annual Simulation Symposium, ANSS ’07, pages_
137–143, Washington, DC, USA, 2007. IEEE Computer Society.
[18] Mike Holcombe, Simon Coakley, and Rod Smallwood. A general framework for agent-based modelling of complex systems. In Proceedings of
_the 2006 European Conf. on Complex Systems, 2006._
[19] Sean Luke, Claudio Cioffi-Revilla, Liviu Panait, and Keith Sullivan. MASON: A New Multi-Agent Simulation Toolkit. _Simulation,_
81(7):517–527, July 2005.
[20] Claudio Márquez, Eduardo César, and Joan Sorribes. A load balancing schema for agent-based spmd applications. In International
_Conf. on Parallel and Distributed Processing Techniques and Applica-_
_tions (PDPTA), Accepted, 2013._
[21] Michael J North, Nicholson T Collier, Jonathan Ozik, Eric R Tatara,
Charles M Macal, Mark Bragen, and Pam Sydelko. Complex adaptive systems modeling with repast simphony. Complex adaptive systems
_modeling, 1(1):1–26, 2013._
[22] Ton Oguara, G Theodoropoulos, B Logan, M Lees, and C Dan. Pdesmas: A unifying framework for the distributed simulation of multi-agent
systems. School of computer science research - University of Birming_ham, 6, 2007._
13
-----
[23] Vincent Rodin, Abdessalam Benzinou, Anne Guillaud, Pascal Ballet,
Fabrice Harrouet, Jacques Tisseau, and Jean Le Bihan. An immune
oriented multi-agent system for biological image processing. _Pattern_
_Recognition, 37(4):631–645, 2004._
[24] M Scheutz, P Schermerhorn, R Connaughton, and A Dingler. Swagesan extendable distributed experimentation system for large-scale agentbased alife simulations. Proceedings of Artificial Life X, pages 412–419,
2006.
[25] Herbert A Simon. The architecture of complexity. Springer, 1991.
[26] Russell K Standish and Richard Leow. Ecolab: Agent based modeling
for c++ programmers. arXiv preprint cs/0401026, 2004.
[27] Vinoth Suryanarayanan, Georgios Theodoropoulos, and Michael Lees.
Pdes-mas: Distributed simulation of multi-agent systems. _Procedia_
_Comp. Sc., 18:671–681, 2013._
[28] Seth Tisue and Uri Wilensky. Netlogo: Design and implementation of
a multi-agent modeling environment. In Proceedings of Agent, volume
2004, pages 7–9, 2004.
[29] Robert Tobias and Carole Hofmann. Evaluation of free java-libraries
for social-scientific agent based simulation. JASS, 7(1), 2004.
14
-----
|
{
"disclaimer": "Notice: The following paper fields have been elided by the publisher: {'abstract', 'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1016/j.cosrev.2016.08.001?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1016/j.cosrev.2016.08.001, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "BRONZE",
"url": "https://link.springer.com/content/pdf/10.1007/978-3-319-14325-5_32.pdf"
}
| 2,014
|
[
"JournalArticle",
"Review"
] | true
| 2014-08-25T00:00:00
|
[] | 7,828
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Engineering",
"source": "s2-fos-model"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Environmental Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/029e0f6017a14e5059d2f130da4c6b7992414533
|
[
"Computer Science"
] | 0.884818
|
An Efficient Privacy-Preserving Authentication Scheme for Energy Internet-Based Vehicle-to-Grid Communication
|
029e0f6017a14e5059d2f130da4c6b7992414533
|
IEEE Transactions on Smart Grid
|
[
{
"authorId": "1791276",
"name": "P. Gope"
},
{
"authorId": "48440849",
"name": "B. Sikdar"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"IEEE Trans Smart Grid"
],
"alternate_urls": null,
"id": "1c2f3998-b5ca-48ca-9991-94b71c71ecb7",
"issn": "1949-3053",
"name": "IEEE Transactions on Smart Grid",
"type": "journal",
"url": "http://ieeexplore.ieee.org/servlet/opac?punumber=5165411"
}
|
The energy Internet (EI) represents a new electric grid infrastructure that uses computing and communication to transform legacy power grids into systems that support open innovation. EI provides bidirectional communication for analysis and improvement of energy usage between service providers and customers. To ensure a secure, reliable, and efficient operation, the EI should be protected from cyber attacks. Thus, secure and efficient key establishment is an important issue for this Internet-based smart grid environment. In this paper, we propose an efficient privacy-preserving authentication scheme for EI-based vehicle-to-grid communication using lightweight cryptographic primitives such as one-way non-collision hash functions. In our proposed scheme, a customer can securely access services provided by the service provider using a symmetric key established between them. Detailed security and performance analysis of our proposed scheme are presented to show that it is resilient against many security attacks, cost effective in computation and communication, and provides an efficient solution for the EI.
|
## This is a repository copy of An efficient privacy-preserving authentication scheme for energy internet-based vehicle-to-grid communication.
White Rose Research Online URL for this paper: http://eprints.whiterose.ac.uk/156053/
Version: Accepted Version
Article: Gope, P. orcid.org/0000-0003-2786-0273 and Sikdar, B. (2019) An efficient privacy-preserving authentication scheme for energy internet-based vehicle-to-grid communication. IEEE Transactions on Smart Grid, 10 (6). 6. pp. 6607-6618. ISSN 1949-3053
https://doi.org/10.1109/tsg.2019.2908698
© 2019 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other users, including reprinting/ republishing this material for advertising or promotional purposes, creating new collective works for resale or redistribution to servers or lists, or reuse of any copyrighted components of this work in other works. Reproduced in accordance with the publisher's self-archiving policy.
**Reuse**
Items deposited in White Rose Research Online are protected by copyright, with all rights reserved unless
indicated otherwise. They may be downloaded and/or printed for private study, or other acts as permitted by
national copyright laws. The publisher or other rights holders may allow further reproduction and re-use of
the full text version. This is indicated by the licence information on the White Rose Research Online record
for the item.
**Takedown**
If you consider content in White Rose Research Online to be in breach of UK law, please notify us by
emailing eprints@whiterose.ac.uk including the URL of the record and the reason for the withdrawal request.
[eprints@whiterose.ac.uk](mailto:eprints@whiterose.ac.uk)
[https://eprints.whiterose.ac.uk/](https://eprints.whiterose.ac.uk/)
-----
# An Efficient Privacy-preserving Authentication Scheme for Energy Internet-based Vehicle-to-Grid Communication
### Prosanta Gope, Member, IEEE and Biplab Sikdar, Senior Member, IEEE
Abstract—The Energy Internet (EI) represents a new electric
grid infrastructure that uses computing and communication to
transform legacy power grids into systems that support open
innovation. EI provides bidirectional communication for analysis
and improvement of energy usage between service providers and
customers. To ensure a secure, reliable and efficient operation,
the EI should be protected from cyber attacks. Thus, secure and
efficient key establishment is an important issue for this Internetbased smart grid environment. In this paper, we propose an
efficient privacy-preserving authentication scheme for EI-based
Vehicle-to-Grid Communication using lightweight cryptographic
primitives such as one-way non-collision hash functions. In our
proposed scheme, a customer can securely access services provided by the service provider using a symmetric key established
between them. Detailed security and performance analysis of
our proposed scheme are presented to show that it is resilient
against many security attacks, cost effective in computation and
communication, and provides an efficient solution for the EI.
Index Terms—Energy internet, mutual authentication, advanced metering infrastructure, smart grids.
I. INTRODUCTION
Energy, science and economy can mutually reinforce each
other through new synergies and bring about greater efficiencies. From the perspective of sustainable development of
society, the exploitation and utilization of the renewable energy
and replacing traditional fossil fuels are important directions
for reforming the energy landscape. However, the traditional
grid structure makes it difficult to meet the requirements
associated with integrating renewables and distributed generation, and incorporate other mechanisms to improve energy
efficiency. In order to address these issues, the concept of
Energy Internet has been proposed that seeks to integrate
Information and Communications Technologies (ICT), cyberphysical systems and power system technologies to develop
the next-generation of smart grids [1], [31]. Analogous to the
conventional Internet, the idea of EI has been introduced to
allow energy to be shared similar to information sharing in the
Internet [30]. The fundamental idea behind the EI is to combine economics, information and energy using the power grid
as the backbone network to provide an open and egalitarian
framework for exchanging energy and associated information.
The EI is designed to facilitate the seamless integration of
P. Gope, is with the Department of Computer Science, University of
Sheffield, United Kingdom. (E-mail: prosana.nitdgp@gmail.com)
B. Sikdar is with the National University of Singapore, (Email: bsikdar@nus.edu.sg)
diverse energy sources with the grid, and facilitate the interaction between various elements of the power grid to achieve
increase in energy efficiencies [2]. All aspects of a power grid
such as generation, transmission, distribution, service provider,
operations, markets, and customers will benefit from secure
and efficient communication on decisions about energy and
information flow [25-26]. Finally, compared with smart grids,
EI further integrates other energy networks such as gas for
improved energy operations.
Vehicle to Grid (V2G) technology broadly consists of systems that facilitate the bi-directional flow of electrical energy
between vehicles and the electrical grid. Electrical energy may
flow from the grid to the vehicle to charge the battery and it
may also flow in the reserve direction when the grid requires
energy (e.g., to provide peaking power). With bi-directional
chargers, electric vehicles (EVs) can become participants in
the V2G eco-system, and such vehicles are energy assets for
the smart grid. EVs need to charge and draw power from
the grid when the State of Charge (SOC) of their batteries
becomes low. The V2G property of EVs would also allow
EVs to deliver power back to the grid and the concept of
EI in V2G networks can be used to allow energy to be
transported from vehicles to a location where it is used to
perform useful work. One of the key benefits of EI in V2G
environments is that it allows individuals (e.g., EV owners,
households etc.) to trade energy without the need to build
their own transmission and distribution networks [31]. With
EI-based V2G, the unstable and intermittent energy generated
by renewable energy sources (mainly solar and wind energy
sources) can be used by EVs two provide two benefits. First,
it provides a way to address the large energy demand of EVs
through renewable energy sources, thus reducing the potential
adverse impact of EVs on power grids. Second, it prevents
renewable energy from being wasted when they are generated
during low demand periods of traditional (non-EV) loads. This
allows more efficient use of energy and can hence facilitate the
wider adoption of renewable energy. EI-based V2G systems
also have other applications, including power dispatch between
cities, power transfer from renewable energy sources to end
users, etc.
In addition to the routing of energy between various entities,
the exchange of information is an important aspect of the
EI and EI-based V2G systems. A number of protocols have
been proposed for information exchange in EI based systems.
The ISO/IEC/IEEE 18880 standard defines communication
protocols and architectures for the EI It defines the data
-----
Table I
COMPARATIVE ANALYSIS OF THE RELATED SCHEMES
|Scheme|Primitive Used|Mobilty Support|Location Privacy Support|
|---|---|---|---|
|[4]|ECC, Bilinear Pairing|No|No|
|[5]|ECC, Bilinear Pairing|No|No|
|[6]|Bilinear Pairing|No|No|
|[7]|AES-CBC, Hash function|No|No|
|[8]|Bilinear Pairing, Hash function|No|No|
|[9]|Bilinear Pairing, Hash function|No|No|
|[10]|Bilinear Pairing,Hash function|No|No|
|[11]|PUF, Hash function|No|No|
|[12-15]|Public-key sign-encryption|Yes|Most of them cannot|
|[21-25]|Public-key sign-encryption|Yes|Most of them cannot|
|IEC15118|ECDSA|Yes|No|
|OCPP|ECDSA|Yes|No|
|Proposed Scheme|Hash function|Yes|Yes|
exchange protocols and the network architecture for integrating various components and participants in the grid, data
storage, and application services. ISO/IEC/IEEE 18880 uses
wide area communications using TCP/IP, and existing nonTCP/IP networks can connect through multi-protocol gateways. ISO/IEC/IEEE 18881 and ISO/IEC/IEEE 18883 have
been developed to address network management and network
security issues that are neglected in ISO/IEC/IEEE 18880. In
contrast, V2G communications typically just focus on the communication between the vehicles, charging stations, and the
grid. In V2G, Electric Vehicle Supply Equipment (EVSE) (i.e.,
EV chargers), such as those in charging stations, can be shared
by many customers. Therefore, a temporal association between
the EV and the EVSE has to be initiated for the charging and
billing process when the charging cable is inserted. In this
regard, some EVs and EVSEs use the IEC 15118 standard
for communication. Similarly, EVSEs use Open Charge Point
Protocol (OCPP) for communication between the EVSE and
the energy management systems. While EI-based V2G has
many benefits (as mentioned above), cyber-security of the
components and data are big concerns [28-29]. The vehicles
themselves face increasingly complex attacks that target not
only the vehicle’s operation but also its privacy. Threats to
privacy include the exposure of the vehicle user’s real identity,
the vehicle’s driving path, location, and disclosure of other
private information. Thus, there are significant challenges to
the design of security mechanisms for V2G environments
and these are further complicated by the topological structure
autonomy and fast rate of transformations due to vehicular
movement. A number of organizations are working on the
development of security solutions for the EI [2-3].
A. Related Work
Secure communication is one of the most important requirements for the EI environment in order to guarantee
secure exchange of data at all times. For secure and efficient
data exchange between the components, protocols with high
security and performance are required To address this issue
many researchers have proposed several mutual authentication
and key establishment schemes suitable for the Advanced
Metering Infrastructure (AMI) with various security considerations and goals. Mohammadali et al. [4] proposed two
ECC-based identity-based key establishment protocols. The
protocols reduce the computational overhead at the smart
meter side of the AMI, and they are resilient against replay
and desynchronization attacks. However, they are vulnerable
to man-in-the-middle, impersonation, and false data injection
attacks, and they incur high computational cost during key
establishment. Nicanfar et al. [5] introduced two key exchange
protocols that are based on the use of a symmetric-key algorithm and ECC. The protocols provide security and scalability
for key exchange in smart grids. However, they are vulnerable
to false data injection attacks. Moreover, both the protocols
incur large computational costs which makes them unsuitable
for resource limited devices in smart grids.
Wu and Zhou [6] presented an authentication and key
distribution scheme by combining symmetric key and public
key cryptographic systems, and the authors claim that their
scheme can eliminate man-in-the-middle and reply attacks.
Subsequently, Xia and Wang showed that [6] cannot ensure
security against man-in-the-middle attacks and they also proposed a new data aggregation scheme [7]. However, Park
et al. reported that the scheme presented in [7] is insecure
against impersonation attacks [8]. Besides, it cannot address
the customer’s privacy requirements. Tsai et al. combined an
identity-based signature scheme and an identity-based encryption scheme [9] for key distribution in smart grids. Odelu et
al. investigated the protocol presented in [9] and demonstrated
that it cannot guarantee the security of the session key and
the strong credentials privacy of the smart meter [10]. They
also introduced a new scheme with a claim that is can
reduce computation overheads. However, Chen et al. proved
that the scheme presented in [10] is vulnerable to several
attacks, and it has large computational and communication
costs. Moreover, our analysis shows that the scheme in [10] is
weak against man-in-the-middle attacks which may lead to
DoS attacks at the server end In this context an attacker
-----
Table II
SYMBOLS AND CRYPTOGRAPHIC FUNCTION
Symbol Definition
IDCS Identity of charging station
PIDi Pseudo identity of Useri
ki Secret key of the Useri
pswi Password of the Useri
βi Thumbprint of the Useri
SK Session key (Useri -CSj )
Kcu Shared secret key between the CSj and USP
LAIx Location area identifier of the entity x
h(·) One-way hash function
⊕ Exclusive-OR operation
|| Concatenation operation
(say Eve) can capture the initial message (Msg1 in [10]),
alter the message, and then send the altered message Msg1[e]
to the service provider (SP). The SP can only decide about
the validity of a request message (Msg1[e] [) after completing the]
whole process, i.e., after receiving the response message (Msg3
in [10]). Consequently, each request is stored in a buffer, where
several intensive pairings first need to be computed, followed
by submission response. This buffer needs to be kept open
until a response from the smart meter (SM) is received. As
a consequence, the memory can easily overflow if a large
number of invalid requests are sent, since the invalid requests
cannot be distinguished due to the late detection of the forged
messages. In [11], the authors have considered the physical
security of the smart meter and they proposed an authentication scheme by using the concept of physical unclonable
functions (PUFs). In addition to [4-11], some recent studies
on privacy issues in V2G communications have appeared in
literature [12-15,21-25]. In these schemes, privacy of the car
owner is considered as an important concern. However, in
these schemes an EV needs to perform several computationally
inefficient cryptographic primitives such as group signature,
sign-encryption, etc. Besides, most of these schemes cannot
ensure the location privacy of the EV user, which is essential
for securely monitoring the status of the EV and efficiently
providing services to the EV user. Table I compares related
work to our approach with respect to the primitive used, ability
for mobility support, and location privacy.
B. Our Contribution
In this paper, we first introduce a new model for EI-based
V2G communication. Subsequently, we propose a lightweight
authentication and key establishment scheme for EI-based
V2G communication. The major contributions of this paper
can be summarized as follows:
- A new model for EI-based V2G communication, which
allows an EV user to seamlessly charge or discharge
the battery of his/her vehicle from the charging stations
located in different geographical locations. However, the
charging/discharging rate may vary based on the location
of the charging station
- An efficient privacy-preserving authentication protocol,
which provides several key security properties including
Authentication Key Exchange (AKE) security, privacy of
the user, protection against eavesdropping or interception
attacks, protection against man-in-the middle attacks, and
location privacy, which are all requirements for secure EIbased V2G communication [32-33]. There are some existing schemes which can ensure most of the security requirements for EI-based V2G communication. However,
they use computationally expensive public-key cryptographic primitives. On the contrary, the proposed scheme
is based lightweight cryptographic primitives such as oneway hash function and exclusive-OR operation, which
creates significantly less computational overhead on the
resource limited user’s device (as shown in Table V).
- Most of the existing schemes including the existing
underlying communication protocols such as IEC15118
and OCPP for V2G communications are vulnerable to
some of the well-known security attacks such as man-inthe middle attacks, impersonation attacks, etc. Therefore,
we provide a rigorous formal security analysis of our
proposed scheme using the BR93-model [18] to show
that it is secure against such attacks.
- A comparative study of the proposed scheme with closely
related existing schemes. It is shown that the proposed
scheme is secure and computationally efficient, and requires significantly lower overhead for establishing a
session key between an user’s device and the charging
station, as compared to the related existing schemes.
The remainder of this paper is organized as follows. In
Section II, we present our system and adversary model. In
Section III we introduce the proposed scheme. The formal
security analysis and performance analysis of the proposed
scheme are presented in Section IV and Section V, respectively. Finally, conclusions are drawn in Section VI. The
symbols and cryptographic functions of the proposed scheme
are defined in Table II.
II. SYSTEM AND ADVERSARY MODEL
A. System Model for EI-based V2G Communication
Fig. 1 shows the system model for an EI-based V2G
environment, which consists of three major components: a set
of EV users each with a mobile device (MD) connected to the
Internet, a set of charging stations (CSs), and a utility service
provider (USP). The USP consists of two components: power
generation, distribution, and management center (PGDMC)
and data center (DC). Each user is required to register their
EV with the USP. Then, the USP maintains all the user
information in its data center. In this network model, the USP
is an organization that is responsible for procuring electricity
from various vendors. The USP also supplies electricity to
charging stations in different locations. These charging stations
may be owned by several private companies. A user may
charge/discharge the batteries of his/her EV from/to any of
the CSs. However, the charging/discharging rate may vary
based on the location of the CS. For example, the charging/discharging rate of the CSs located at commercial area
|Symbol|Definition|
|---|---|
|ID CS|Identity of charging station|
|PID i|Pseudo identity of User i|
|k i|Secret key of the User i|
|psw i|Password of the User i|
|β i|Thumbprint of the User i|
|SK|Session key (User -CS j) i|
|K cu|Shared secret key between the CS and USP j|
|LAI x|Location area identifier of the entity x|
|h(·)|One-way hash function|
|⊕|Exclusive-OR operation|
||||Concatenation operation|
-----
Figure 1. System model for the proposed scheme.
networks (CANs) may be higher than others. On the other
hand, the charging/discharging rate of public area networks
(PANs) may be lower than residential area networks (RANs).
We assume that a secure channel is available between an EV
user and the USP during the initial registration. Subsequently,
each user with a mobile device communicates with the CS
through the Internet. A CS may communicate with the USP
through the public Internet or private networks. In this model,
two types of flows, i.e., energy flows (shown by dotted lines)
and data flows (shown by solid lines) have been considered.
All the entities (user, CS and USP) need to authenticate
themselves before sharing any information. Because of the
public network based communication used in the system
environment, there is a possibility of various attacks, such as
replay, man-in-the middle, and impersonation attacks. In our
scheme, users use biometrics (e.g., fingerprints) in addition to
a password for two-factor authentication.
B. Adversary Model
During user registration, a user and the USP interact through
a secure channel. On the other hand, during the execution
of the proposed authenticated key agreement scheme, all
parties communicate through an insecure public channel. In
this context, we consider the Dolev-Yao threat model (DY
model) [29], where an adversary may eavesdrop, modify, or
delete the messages exchanged during transmission. Now, due
to the usage of public networks and wireless communication
in this EI-based V2G environment, there is a possibility of
several attacks, such as impersonation, man-in-the middle,
replay attacks, etc. The user’s privacy is another important
issue in this environment. Also, an adversary can impersonate
as a legitimate user and try to obtain services. Similarly, a
charging station may impersonate as others and ask for higher
charges from a user. Hence, there is a need for an authenticated
key agreement scheme by which the legitimacy of the entities
can be verified, and also both the user and CS can establish a
session key
III. PROPOSED SCHEME
In this section, we present our proposed lightweight authentication protocol for EI-based V2G communication, where
a user (Useri ) who has mobile device MDi with Internet
connectivity requests charging of his/her EV’s battery from
a charging station CSj . In this regard, both Useri and CSj
need to authenticate each other with the help of the USP.
After successful mutual authentication between Useri and
CSj, both entities will establish a session key SK for their
secure communication. Our proposed scheme consists of the
following two phases: user registration and authentication.
A. User Registration Phase
Each user first needs to register with the USP. The registration process consists of the following steps:
Step R1: Useri sends the registration request along with
its identity IDi to the USP through the secure (out-of-band)
channel.
Step R2: Upon receiving the request, the USP creates
an account and inserts a new row in its database. It then
randomly generates a unique pseudo identity PIDi, a secret
key ki, and also generates a set of shadow identities SID =
{sid1, sid2, · · ·, sidn }, which are later used in case of loss
of synchronization between the USP and Useri . Next, the
USP composes a message with {PIDi, ki, SID} and sends it
to Useri through the secure channel. Finally, the USP stores
{PIDi, ki, SID} in its database for further interaction with
Useri .
Step R3: Upon receiving {PIDi, ki, SID} from the USP,
the user inputs his/her biometrics (e.g., thumbprint) βi and
password pswi and computes ki[∗] [=][ k][i][ ⊕] [h][(][β][i][||][psw][i] [)][. Finally,]
Useri stores {PIDi, ki[∗][,][ SID][}][ in his/her mobile device for]
further communication with the USP.
B. Authentication Phase
To accomplish communication security, Useri has to go
through an authentication process each time before obtaining
services from charging station CSj . The authentication phase
of the proposed scheme comprises of the following steps:
-----
Figure 2. Steps and computations in the key agreement phase of proposed scheme.
Step 1: Useri inputs his/her thumbprint βi and password
pswcomputesi into his/her mobile device αi = h(βi) and ∂i′ MD[=][ h]i[(]. The mobile device then[α][i][||][psw][i] [)][, and validates]
the user’s legitimacy. If the user’s validation is successful,
then the device calculates ki = ki[∗] [⊕] [h][(][β][i][||][psw][i] [)][. After that,]
the user generates a nonce Nu and finds his/her location area
identity, LAIu, using the MD’s location service. Next, Useri
�
computes EL = LAIu h(ki ||Nu ), a key-hash response
V1 = h(PIDi ||Nu ||ki ||EL), and subsequently composes a
message MA1 : {PIDi, Nu, EL, V1 } and sends it to charging
station CSj .
Step 2: Upon arrival of message MA1, charging station CSj generates a nonce Nc and computes V2 =
h(IDcs ||Nc||Kcu ||LAIcs ), where LAIcs denotes the location
area identifier of charging station CSj . Next, CSj composes a
message MA2 : {MA1, IDcs, Nc, LAIcs, V2 } and sends it to
the USP.
Step 3: Upon arrival of message MA2, the USP first locates
PIDi in its database and then computes and validates the key
hash responses V1 and V2 . Next, the USP decodes LAIu from
EL and then compares and validates LAIu with LAIcs . If the
validation is successful, the USP generates a key SK and a
new pseudo identity� PIDi[new] . It then computes PIDi[new] [∗] =
PIDi[new] h(PIDi ||ki ), SKu = h(IDu ||ki ||Nu ) [�] SK,
SKcs = h(IDcs ||Kcu ||Na ) [�] SK, V3 = h(SKcs ||Kcu ||Nc),
and V4 = h(SKu ||ki ||PIDi[new] [∗]). Next, the USP composes
a message MA3 : {(PIDi[new] [∗], SKu, V4 )||(SKcs, V3 )} and
sends MA3 to charging station CSj .
Step 4: Upon arrival of the response message MA3 from
the USP, the charging station first computes and validates the
key-hash response V3 . If the validation is successful, CSj
decodes the session key SK = h(IDcs ||Kcu ||Na ) [�] SKcs and
composes a new message MA4 : {(PIDi[new] [∗], SKu, V4 )} and
then sends it to Useri .
Step 5: Upon arrival of message MA4, Useri first
verifies the key-hash response V4 . If the validation is
successful, Useri computes and decodes the session key
SK h(ID ||k ||N ) SK and the new pseudo identity
[�]
�
PIDi[new] = PIDi[new] [∗] h(PIDi ||ki ) for the next round.
The entities involved in the protocol will stop the execution of the scheme if any of the above verification steps is
unsuccessful. For dealing with the loss of synchronization
problem, instead of the pseudo identity PIDi, Useri needs
to select one of the unused shadow identities sidx from
SID = {sid1, sid2, · · ·, sidn } and send it in message MA1 .
On receiving this message and after successfully validating the
user, the USP generates a new pseudo identity and securely
sends it in message MA3 by using the secret key ki . At the
end of the authentication process, both Useri and the USP
delete the used shadow identity sidx from their storage. Also,
in the proposed scheme, Useri can only use almost t shadow
identities, where t < n − 1. After that, the user needs to
request for reloading. In this context, the user sends a “ReLoad” message to the USP. On receiving that message, the
USP generates a new set of shadow identities and then securely
sends it in message MA3 by using the secret key ki . Details
of this phase are depicted in Fig. 2.
Remark 1: In our proposed scheme, if a user needs to
charge or discharge his/her vehicle multiple times in a day,
then he/she needs to go through the authentication process
each time, even if the same EV is used. Besides, since one
of the goals of the proposed scheme is to achieve location
privacy, we do not keep any footprint of the CSs. Therefore,
even if the EV uses the same CS multiple times, it needs to
execute the proposed anonymous authentication process. Since
our proposed scheme is based on lightweight cryptographic
primitives such as hash functions, it has a lower computational
cost (execution times are shown in Table III and Table IV).
Besides, from Table IV we can see that the communication
cost of the proposed scheme is significantly less than the other
schemes. On the other hand, in our proposed scheme, we allow
a user to have a single account for multiple EVs, which will
avoid any increase in the credential storage requirement.
Remark 2: Now, we consider the scenario where two
users Useri and Userj share a vehicle. In such cases, during registration the USP will generate two sets of security
-----
credentials {PIDi, ki, SIDi } and {PIDj, kj, SIDj } under the
same account and send them to Useri and Userj, respectively.
After receiving their credentials, both the users securely store
them in their respective mobile devices (as shown in Step R3).
Now, when Useri uses the vehicle then he/she needs to use
{PIDi, ki, SIDi } to get through the authentication process.
Similarly, when user Userj uses the vehicle then he/she is
required to use {PIDj, kj, SIDj } in order to authenticate with
the USP. In this way, the proposed scheme can support the
scenario where a vehicle is shared among multiple users.
However, in this context, the storage complexity at the USP
will increase linearly with the number of shared users.
IV. FORMAL SECURITY ANALYSIS
This section presents the formal security proof of the proposed scheme. We first demonstrate that our proposed scheme
is secure.
A. Definitions and Assumptions
Bellare and Rogaway introduced a theoretical security proof
for an authentication and key exchange protocol for a symmetric two-party case, which we refer to as the BR93-Model [18].
During the authentication process only the USP can authenticate a user, and a CS needs to forward the authentication
request of the user to the USP. Thus, we assume that the
communication between the CS and USP is secure, so that
the USP and the CS can be regarded as a single participant
and we call it the service agent (SA).
1) Complexity Assumptions: The security of our proposed
scheme is based on the secure one-way hash function, which
can be regarded as a pseudorandom function [19]. Therefore,
we first introduce the security definitions of pseudorandom
functions and show their game environments that will be used
for the security proofs of the proposed scheme.
Definition 1: Let f be a polynomial-time computable
function and AdvH = |Pr[H [f] = 1] − Pr[H [f][ ′] = 1]|
denote the advantage of an algorithm H, controlled by a
probabilistic polynomial-time adversary A, in distinguishing
f from another function f [′]. We say that f is a (n, q, ε)-secure
pseudorandom function if there is no feasible algorithm H
that can distinguish f from f [′] with advantage AdvH ≥ ε,
while making at most q oracle queries to f or a truly random
function f [′] and running at most n times by playing the
following game:
Initialization: A challenger C interacting with A picks a
random bit b ∈{0, 1} to determine the function fb, where f0
is a pseudorandom function and f1 is a truly random function.
Training Phase: A issues q queries, x1, · · ·, xq to C, where
xi ∈{0, 1}[∗] are binary strings of arbitrary length. The
challenger responds to these queries by sending fb(xi) to A
for i = 1, · · ·, q, where fb(xi) ∈{0, 1}[l] and l is a fixed
positive integer.
Guess: A outputs b[′] ∈{0, 1} as a guess of b. A wins this
game if b[′] = b. We define the advantage of A winning the
game as Advf0,A = |Pr[b[′] = b] − [1]2 [|][.]
According to the pseudorandom function assumption, no
probabilistic polynomial-time adversary can win the above
game with non negligible advantage
2) Security Model and Notations: Protocol Participants:
�s
A,B [denotes the oracle which plays the role of][ A][ to interact]
with B in session s, and [�]A,B[t] [denotes the oracle which plays]
the role B to interact with A in session t, where A, B ∈ I,
s, t ∈ N, I is the set of identities of the players such as a user
and the service agent who participate in the protocol, and N
is the set of positive integers.
Protocols: The proposed authentication scheme uses a
three-party authentication and key exchange scheme. However,
the protocol can be reduced to a de facto two-party setting
protocol. Therefore, we define a two-party authentication and
key exchange protocol as follows.
Definition 2: A two-party authentication and key exchange
protocol P, is formally specified by an efficiently computable
function on the following inputs:
[�]
k: The length of the security parameter used in the protocol.
A: The identity of the initiator of P, where A ∈ I.
B: The identity of the intended partner of P, where B ∈ I.
x: The secret information, where x ∈{0, 1}[∗].
K: The conversation in P so far.
r: The random coin flips of the sender or initiator, where
r ∈{0, 1}[+].
The output of (k, A, B, x, K, r) = (m, δ, α) is defined as
[�]
follows:
m: The next message to be sent, where m ∈{0, 1} {∗},
[�]
where {∗} specifies that the initiator sends no message.
δ: The decision, where δ ∈{A, R, ∗}, and A, R, and *
denote accept, reject, and no decision, respectively.
α : The private output, where α ∈{0, 1}[∗] [�]{∗} and {∗}
denotes that the initiator does not have any private output.
3) Adversary Model: An adversary A is a probabilistic
polynomial-time Turing machine during the execution of protocol P . A can control the channel between A and B by
eavesdropping on the messages sent by A and B, modifying
these messages, and compromising the session secrets shared
between A and B. These behaviors can be modeled by the
following queries.
Execute([�][s]A,B[,][ �][t]B,A[)][: This query models all kinds of]
passive attacks, where a passive adversary can intercept all
the data exchanged between [�]A,B[s] [and][ �]B,A[t] [in a session of]
P .
Send([�][s]A,B[, m][)][: This query models active attacks, where]
an adversary sends a message m to [�]A,B[s] [and obtains a]
response message according to the proposed scheme.
Reveal([�][s]A,B[)][: This query models the exposure of session]
keys (known session key attacks) in a particular session s.
Corrupt([�][s]A,B[)][: This query models the revelation of long-]
term secret keys. This query models passive attacks.
Test([�][s]A,B[)][: When][ �]A,B[s] [has accepted and shared a]
session key, adversary A can make this query and try to
distinguish a real session key from a random string.
4) Security Definitions: Before defining the notion of mutual authentication security, we first briefly review the definition of a matching conversation.
Definition 3 (Matching Conversations): An authenticated
key exchange protocol P is a message-driven protocol and the
goal of P is to achieve a matching conversation. We first define
a protocol session of a party A as (A B s role) where B
-----
is the identity of A’s partner, s is the session identifier, and
role can be either initiator or responder. A P with two
protocol sessions between a party A and a party B are of
the form (A, B, s, initiator) and (A, B, t, responder), respectively. Two sessions are said to be a matching conversation
involving A and B if their session identifiers are identical and
the initiator and responder parties are A and B. If a protocol
P consists of more than two sessions and each pair of sessions
in sequence is a matching conversation, then P is said to be
a protocol of matching conversations.
We define mutual authentication based on the definition of
matching conversation as follows. P is a mutual authentication
protocol if for any polynomial time adversary A: (1) matching
conversation implies acceptance and (2) acceptance implies
matching conversation. The first condition says that if the sessions of two parties consists of a matching conversation, then
the parties accept the authentication of each other. The second
condition says that if each party accepts the authentication
with the other party in a conversation, then the probability that
there is no matching conversation between them is negligible.
Formally, mutual authentication (MA) security is defined as:
Definition 4: An authentication protocol P is MA-Secure
(i.e., P satisfies MA-Security) if:
(1) Matching conversation implies acceptance: If oracles
�s
A,B [and][ �]B,A[t] [have matching conversations, then both]
oracles accept the authentication of each other, AND
(2) Acceptance implies matching conversations: The probability of event NoMatching[A](k) is negligible, where k is a
security parameter and NoMatching[A](k) is the event that
there exist i, j, A, and B such that [�]A,B[i] [is accepted but]
there is no oracle [�]B,A[j] [which is engaged in a matching]
conversation.
The event NoMatching[A](k) can also be denoted as
Succ[MA]P (A) which is the probability that a polynomial-time
adversary A can successfully impersonate one of the two
interactive entities who want to authenticate each other in P .
Authentication Key Exchange (AKE) Security: In an
execution of an MA-Secure authentication protocol P, a
polynomial-time adversary A interacts with two fresh oracles:
�s
A,B [and its partner][ �]B,A[t] [. At the end of the execution,][ A]
issues a Test query to one of the two fresh oracles. Then the
real session key or a random string is returned to A according
to the value of a random bit b. Finally, A outputs a bit b[′] and
terminates the game. The AKE-Advantage, AdvP[AKE](A), is
defined as |Pr [b = b[′]] − 1/2|. We give a formal definition of
AKE-Security below:
Definition 5 A protocol P is AKE-Secure if it satisfies the
following properties:
(1) At the beginning the adversary engages in the execution
of P with [�]A,B[s] [and its partner][ �]B,A[t] [. Then both oracles can]
accept and share the same session key with each other.
(2) P is MA-Secure.
(3) For every probabilistic polynomial-time adversary A,
AdvP[AKE](A) is negligible.
When a Test query is issued before finishing the execution
of the protocol, the game is played as per the above definition
if the session key is generated by any one of the two fresh
parties Otherwise the Test query will be rejected
B. Formal Security Analysis of the Proposed Scheme
The proposed scheme is based on hash functions, which
can be considered as secure pseudorandom functions [19]. In
this section, we show that the proposed scheme is provably
secure based on the pseudorandom function assumption. As
mentioned earlier, even though our proposed scheme is based
on a three-party authentication and key exchange protocol, it
can be reduced to a two-party authentication and key exchange
protocol.
Lemma 1: If h is a (n0, q0, ε0)-secure pseudorandom
function family with negligible ε0, then the proposed authentication scheme is MA-Secure.
Proof: Assume that there is a polynomial-time adversary
A who can break MA-Security of the proposed protocol P
with non-negligible probability Succ[MA]P (A). We construct a
polynomial time algorithm F using A to show that F can
break the pseudorandom function with non-negligible advantage, thus providing a contradiction. Also, Succ[MA]P (A) =
Pr [SuccUser] + Pr [SuccSA] − Pr [SuccUser, SuccSA] ≤
Pr [SuccUser] + Pr [SuccSA], where SuccUser and SuccSA
are the events that A successfully impersonates as a legitimate
user and SA, respectively, to pass authentication. Therefore,
we split the proof into two cases, one for SA impersonation
and the other for user impersonation.
Case 1 (SA Impersonation): Assume that A can impersonate
as a SA with probability ǫ[′]. If A wants to be successfully
authenticated by a user (say Ui ) using [�]User,SA[s] [controlled]
by F, A must correctly send V4 = h(SKu ||ki ||PIDi[new] [∗]). In
the following game, F will exploit the ability of A to break the
pseudorandom function assumption with ǫ[′] ≤ 4ǫ0+2[−][k], where
k is the security parameter. F plays the game in Definition 1
with challenger C as follows.
Initialization: Let the long-term secret key ki be k-bit long.
C picks a random bit b ∈{0, 1} and sets up a secure one-way
hash function hb where h0 = hki is a pseudorandom function
and h1 is a random function. If F simulates the game by using
h1 to interact with A, we call this game a random experiment.
On the other hand, if F uses h0 to simulate the game, we call
this game a real experiment. The goal of F is to correctly
guess if hb = h0 or hb = h1 (i.e., b = 0 or b = 1).
Training: F simulates [�]User,SA[s] [and][ �]SA,User[t] [to interact]
with A by answering the following queries:
- Execute([�][s]User,SA[,][ �][t]User,HG[)][:][ F][ uses][ h][b][ given by][ C]
as hki in the protocol. F also randomly generates kh and
PIDi[new] and then computes PIDi[new] [∗] = h(PIDi ||ki ) ⊕
PIDi[new], SKu = h(IDu ||ki ||Nu ) ⊕ SK, and V4 =
h(SKu ||ki ||PIDi[new] [∗]). Subsequently, F simulates [�]User,SA[s]
and [�]SA,User[t] [with the help of][ h][b][,][ PID]i[new] [∗], SKu, and V4.
- Send([�][s]User,SA[, m][)][:][ �][s]User,SA [sends the request mes-]
sage m = {PIDi, Nu, V1} of the protocol. [�]User,SA[s] [first]
validates V1 by querying hb and then finds PIDi in its
database and then checks the correctness of V1 by querying
hb.
- Send([�][t]SA,User[, m][)][:] If m = {PIDi, Nu, V1},
�t
then SA,User computes PIDi[new] [∗] = h(PIDi ||ki ) ⊕
PID [new] k [HG] h(ID ||k ||N ) ⊕ SK and V
-----
h(SKu ||ki ||PIDi[new] [∗]). [�]SA,User[t] [then responds by sending]
{PIDi[new] [∗], SKu, V4} to A.
Challenge: First, A queries Send([�][s]User,SA[, m][)][ to trigger]
the protocol. [�]User,SA[s] [then sends][ m][ =][ {][PID][i][, N][u][,][ V][1][}][ to]
A. Then A generates the authentication response parameter V4
with success probability Pr [SuccSA] = ε[′]. Thus, A queries
Send([�][t]SA,User[,][ {][PID]i[new] [∗], SKu, V4}). After receiving this
query, F issues a query x [∗] = h(SKu ||ki ) to hb and obtains
the output V4[∗] [=][ h][(][SK][u] [||][k][i] [||][PID]i[new] [∗]).
Guess: Finally, F outputs a guess bit b[′] ∈{0, 1}. If V4[∗] [=]
V 4 then F outputs 0; otherwise, F outputs a random bit 0 or
1.
The analysis of the probability that F can successfully
distinguish between the given hb (i.e., b = b[′]) can be divided
into two cases: under a real experiment (i.e., b = 0), and
under a random experiment (i.e., b = 1). In the case of a real
experiment, A can successfully send the correct authentication
information to win the game with probability ε[′]. Hence, F
will output b[′] = 0 with probability ǫ[′] when A sends correct
authentication information under a real experiment. However,
if A sends wrong information, F can only make a random
guess for b, and thus F will output b[′] = 0 with probability
(1 − ε[′])/2. Thus, when b = 0, Pr[b = b[′]|b = 0] =
ε[′] + (1 − ε[′])/2. In the case of random experiments, A can
only send the correct authentication information by random
guessing and the probability of a correct guess is 2[−][k]. Thus,
when b = 1, F outputs b[′] = 1 with probability (1 − 2[−][k])/2
(i.e., Pr[b = b[′]|b = 1] = (1 − 2[−][k])/2). Combining the two
cases, we have
Pr [b = b[′]] = Pr [b = b[′], b = 0] + Pr [b = b[′], b = 1]
= (ε[′] + (1 − ε[′])/2)1/2 + ((1 − 2[−][k])/2)1/2
= 1/2 + ǫ[′]/4 − 2[−][(][k][+2)].
Thus we have
ε0 ≥|Pr [b = b[′]] − 1/2|
= ǫ[′]/4 − 2[−][(][k][+2)].
⇒ ǫ[′] ≤ 4ε0 + 2[−][k].
Case2 (User Impersonation): Suppose that A can impersonate as a user with probability ε[′′]. If A wants to be accepted by
�t
SA,User[, then][ A][ has to send out the correct authentication]
information. Thus F plays the same game as in Case 1 with
C.
Initialization: C selects a hash function hb according to a
random bit b ∈{0, 1} for answering the queries from F where
h0 = hki is a pseudorandom function and h1 is a random
function.
Training: F first selects the required Nu and PIDi
�s
in the protocol. F then simulates User,SA and
�t
SA,User by answering Execute([�][s]User,SA[,][ �]SA,User[t] [)]
and Send([�][s]User,SA[, m][)][. The simulations of these oracles]
are similar to those in Case 1.
Guess: F outputs a guess b[′] ∈{0, 1} according to PIDi
and V1. If PIDi and V1 are valid, then F outputs 0, implying
hb = hki ; otherwise it outputs a random bit 0 or 1.
The probability that A successfully sends out the correct
Succ[MA]P (A) ≤ Pr [SuccSA] + Pr [SuccUser]
= ε[′] + ε[′′]
≤ 8ǫ0 + 2[−][(][k][−][1)].
From the above, ε0 is non-negligible, which contradicts the
assertion in the lemma’s statement that ε0 is negligible. Thus
we can conclude that the proposed authentication scheme is
MA-Secure.
Lemma 2: If h is a (n0, q0, ε0)-secure pseudorandom
function family with negligible ε0, then the proposed scheme
is AKE-Secure.
Proof: In Lemma 1 we have proved that the proposed protocol P is MA-Secure. Now, consider an adversary A who can
break AKE-Security of P with non-negligible AdvP[AKE](A) =
ε. We construct a simulator F using the ability of A to
break the pseudorandom function assumption [20]. F plays
the following game, as given in Definition 3, with a challenger
C.
Initialization: C picks a random bit b ∈{0, 1} and sets
up a secure hash function hb for answering the queries from
F, where h0 = hki is a pseudorandom function and h1 is a
random function.
Training: F selects the required Ng and SIDi in the protocol. F then simulates [�]User,SA[s] [, and][ �][t]SA,User [by answer-]
ing Execute([�][s]HG,SA[,][ �][t]SA,HG[)][ and][ Send][(][�][s]User,SA[, m][)][,]
respectively. The simulations of these oracles are similar to
those in the proof of Lemma 1.
- Test([�][s]User,SA[)][: If][ k][h][ of][ �][s]User,SA [is generated, then]
F randomly chooses c ∈{0, 1}, and returns the real session
key kh if c = 0 or a random string for c = 1. Otherwise, F
returns ⊥, denoting meaninglessness.
- Test([�][t]N,E[)][: The simulation is the same as the one]
above.
Challenge: After querying Execute([�][s]HG,SA[,][ �][t]SA,HG[)][,]
A sends a Test query to F.
Guess: After querying Test([�][s]User,SA[)] or
Test([�][t]SA,User[)][,][ A][ outputs a bit][ b] = 0 if it thinks
that the responding string is the real session key; otherwise, it
outputs b = 1. Finally, F outputs b[′] = 0 if b[′] = b; otherwise
F outputs b[′] = 1.
The analysis of the probability of the event b = b[′] is similar
to that in the proof of Lemma 1. A can win the game by
successfully guessing b = b[′] with probability (ε + 1/2) under
a real experiment (i.e., b = 0). Also, A can only guess if b = b[′]
with probability 1/2 under a random experiment (i.e., b = 1).
If A successfully guesses b = b[′], then F will output b[′] = 1.
Therefore the probability of b b[′] and b 0 is (ε+1/2)1/2
PIDi and V1 is ε[′′] in the real experiment and 2[−][k] in the
random experiment. Following the analysis of Case 1, we have
Pr [b = b[′]] = 1/2 + ε[′′]/4 − 2[−][(][k][+2)]
⇒ ε′′ ≤ 4ǫ0 + 2−k.
Combining Case 1 and Case 2,
-----
Figure 3. Attack Tree.
and the probability of b = b[′] and b = 1 is 1/4. Thus we have
Pr [b = b[′]] = Pr [b = b[′], b = 0] + Pr [b = b[′], b = 1]
= (ε + 1/2)1/2 + 1/4
= 1/2 + ǫ/2
⇒ ǫ0 ≥ Pr [b = b[′]] − 1/2
= ǫ/2
From the above, ε0 is non-negligible, and thus a contradiction occurs. Therefore, AdvP[AKE](A) is negligible for each
polynomial-time adversary A and P is AKE-Secure.
C. Informal Security Analysis
So far, we have formally proved that the proposed scheme
can ensure AKE-security, which is imperative to achieve security against impersonation attacks or replay attacks, session
key security, etc. In this subsection we use the attack tree
shown in Fig. 3 to show how the proposed scheme ensures
some of the important security properties which are necessary
for EI-based V2G communications.
1) Protection Against Impersonation or Forgery Attacks:
In the proposed scheme, if an adversary tries to
impersonate as a legitimate user Useri, then he/she
needs to send a valid authentication request MA1 :
{PIDi, Nu, EL, V1 }. However, the adversary cannot
provide the thumbprint βi and password pswi . Therefore, he/she cannot use the mobile device and compute
ki = ki[∗] [⊕] [h][(][β][i][||][psw][i] [)][,][ EL][ =][ LAI][u] � h(ki ||Nu ), and
a valid key-hash response V1 = h(PIDi ||Nu ||ki ||EL),
which are essential to authenticate with the USP. On the
other hand, if the adversary tries to impersonate as a
legitimate service provider, then he/she must know the
secret keys Kcu and ki . Without knowing the secrets
Kcu and ki, the adversary cannot generate valid keyhash responses V3 = h(SKcs ||Kcu ||Nc) and V4 =
h(SKu ||ki ||PIDi[new] [∗]). In our EI-based V2G communications model, charging/discharging rates vary based on
the location. A charging station CSj may try to cheat
the Useri by providing a false location identity LAIcs
to the USP and demand an inaccurate amount from the
user. The proposed scheme will be able to detect such
forgery attempts in the following way: the USP decodes
LAIu from the EL and then compares and validates
LAI with LAI If the validation is successful then
only the USP will proceed with the execution of the
further steps. Otherwise, the USP will terminate the
execution of the protocol and take necessary against
the CS. Similarly, an user may intentionally provide
a forged LAIu in order to pay a lower amount for
charging or ask for a higher amount for discharging.
The USP will similarly be able to detect such attempts.
Next, we consider a scenario where the user’s mobile
device is lost or stolen. The adversary may try to use
this device to impersonate as a legitimate user. However,
in our proposed scheme we have considered multi-factor
security and the adversary cannot provide the valid
thumbprint βi and password pswi . Hence, he/she will
not be able to proceed with further execution of the
protocol. In this way, we can ensure security against
impersonation and forgery attacks.
2) Privacy of the User: In the proposed scheme, the user
needs to use a valid pseudo identity PIDi for each
session, which cannot be used twice. Therefore, no one
except the service provider can a recognize the activity
of the user. Besides, in case of loss of synchronization,
the user needs to use one of the unused shadow identities
sidj from SID = {sid1, · · ·, sid n }. After that, the user
deletes sidj from its memory. Therefore, changing the
pseudonym in each session ensures identity intractability. This approach of the proposed scheme is quite useful
for achieving privacy against eavesdropper (PAE).
3) Protection Against Eavesdropping or Interception Attacks: In the proposed scheme, an adversary cannot
reuse the message MA1 : {PIDi, Nu, EL, V1 } since
PIDi changes in each session. The adversary cannot
reuse message MA2 since a new random number Nc
is used in each session. Similarly, an adversary also
cannot resend the messages MA3 and MA4 since keyhash response messages V3 and V4 change in each
session and they are generated based on the challenges
Nu and Nc, respectively. In this way, we ensure security
against replay attacks.
4) Protection Against Compromised User’s Device: Next,
we consider a scenario when an attacker hijacks the
car with the user’s device and forces the legitimate
user to input his/her password and thumbprint and then
change the password and the thumbprint. After that,
the adversary may try to ask for charging/discharging
services from the USP. In order to address this issue, the
-----
Table III
PERFORMANCE COMPARISON BASED ON SECURITY FEATURES
SP1 SP2 SP3 SP4 SP5
Yes Yes No No Yes
No No No No No
No No No No No
No No Yes No No
Yes Yes Yes Yes No
Yes Yes Yes Yes Yes
Yes Yes Yes Yes Yes
SP2: Privacy against eavesdropper; SP3: Resilience against man-in-the-middle attacks;
SP4: Forward secrecy; SP5: Session key security; SP6: Resilience against DoS attacks
Table IV
EXECUTION TIME OF VARIOUS CRYPTOGRAPHIC OPERATIONS
User’s Device (HTC One Smartphone) USP/CS (Intel Core i5-4300 Machine)
5.12 ms 2.6 ms
21.86 ms 14.5 ms
8.67 ms 3.78 ms
gen 55.946 ms
ver - 17.237 ms
0.0186 ms 0.011 ms
7.235 ms 2.338 ms
0.0584 ms 0.041 ms
Table V
PERFORMANCE COMPARISON BASED ON COMPUTATION COST (IN MS) AND COMMUNICATION COST
User’s Device USP/CS
2Tmp+Tm +Tcertgen +3Th ≈88.15 3Tmp+Tm +Tcertver +4Th ≈57.87
3Tmp+Tm +Tcertgen +Th ≈93.24 4Tmp+Tm +Tcertver +4Th +Ts ≈63.77
2Tmp+Tm +Tcertgen +Th +Ts ≈92.38 3Tmp+Tm +Tcertver +3Th +Ts ≈57.88
Ts + 4Th ≈0.13 Ts + 4th ≈0.085
4Tmp+Te +5Th ≈27.85 3Tmp+Te + 2Tb+5Th ≈23.22
3Tmp+Te + 6Th ≈22.74 2Tmp+Te + 2Tb+6Th ≈15.32
6Th ≈0.15 8Th ≈0.88
: Time Required for a multiplication point operation; Tm : Time Required for a multiplication operation;
:Time Required for of a modular exponential operation; Ts :Time Required for a symmetric encryption/decryption;
Tb:Time Required for a bilinear pairing;Th : Time Required for a hash operation;
Tcertgen/ver : Time Required for a certificate generation/verification operation
|Scheme|SP1|SP2|SP3|SP4|SP5|SP6|
|---|---|---|---|---|---|---|
|Mohammadali et al. [4]|Yes|Yes|No|No|Yes|No|
|Nicanfar et al. [5]|No|No|No|No|No|No|
|Wu et al. [6]|No|No|No|No|No|No|
|Xia et al. [7]|No|No|Yes|No|No|No|
|Tsai et al. [9]|Yes|Yes|Yes|Yes|No|No|
|Odelu et al. [10]|Yes|Yes|Yes|Yes|Yes|No|
|Proposed Scheme|Yes|Yes|Yes|Yes|Yes|Yes|
|SP1: Privacy of customer; SP2: Privacy against eavesdropper; SP3: Resilience against man-in-the-middle attacks;|||||||
|SP4: Forward secrecy; SP5: Session key security; SP6: Resilience against DoS attacks|||||||
|Operation|User’s Device (HTC One Smartphone)|USP/CS (Intel Core i5-4300 Machine)|
|---|---|---|
|T mp|5.12 ms|2.6 ms|
|T m|21.86 ms|14.5 ms|
|T b|8.67 ms|3.78 ms|
|T certgen|55.946 ms|-|
|T certver|-|17.237 ms|
|T h|0.0186 ms|0.011 ms|
|T e|7.235 ms|2.338 ms|
|T s|0.0584 ms|0.041 ms|
|Scheme|User’s Device|USP/CS|Communication Cost|
|---|---|---|---|
|Mohammadali et al. [4]|2T +T +T +3T ≈88.15 mp m certgen h|3T mp+T m+T +4T h≈57.87 certver|2340-bits|
|Nicanfar et al. [5]|3T +T +T +T ≈93.24 mp m certgen h|4T mp+T m+T +4T h+T s≈63.77 certver|2176-bits|
|Wu and Zhou [6]|2T mp+T m+T certgen+T h+T ≈92.38 s|3T mp+T m+T +3T h+T s≈57.88 certver|4064-bits|
|Xia and Wang [7]|T + 4T ≈0.13 s h|T + 4t ≈0.085 s h|3296-bits|
|Tsai and Lo [9]|4T +T +5T ≈27.85 mp e h|3T +T + 2T +5T ≈23.22 mp e b h|6880-bits|
|Odelu et al. [10]|3T +T + 6T ≈22.74 mp e h|2T +T + 2T +6T ≈15.32 mp e b h|2912-bits|
|Proposed Scheme|6T ≈0.15 h|8T ≈0.88 h|1802-bits|
|T : Time Required for a multiplication point operation; T : Time Required for a multiplication operation; mp m||||
|T :Time Required for of a modular exponential operation; T :Time Required for a symmetric encryption/decryption; e s||||
|T :Time Required for a bilinear pairing;T : Time Required for a hash operation; b h||||
|T : Time Required for a certificate generation/verification operation certgen/ver||||
legitimate user needs to inform such an incident to the
USP as soon as possible. After that, the USP will block
the user’s account. In addition, the USP can also place
a limit on the weekly or monthly charging/discharging
amount for an user. In this way, we can address the
scenario of compromised user devices.
5) Protection Against Physical Attacks: In the proposed
scheme we assume that all the devices (such as user’s
mobile device, EV, EVSE) are tamper proof. Therefore,
if an adversary attempts to perform any physical attacks,
they can be resisted by the hardware. In addition,
in order to deal with physical attacks, devices with
embedded physical uncloneable functions (PUFs) [11]
can also be used. Any attempt to tamper with the PUF
changes the behavior of the device and renders the
PUF useless, thereby making it possible to detect any
tampering attempts.
V. PERFORMANCE EVALUATION
This section evaluates and compares the performance of the
proposed scheme with respect to other authentication schemes
for smart grids. We first consider several imperative security
properties such as forward secrecy, session key security, etc.
for analyzing the performance of our proposed authentication
scheme on the security front with respect to other schemes
([4] [5] [6] [7] [9]) Table III shows that the schemes
-----
presented in [4], [5], [6], [7], [9], and [10] fail to guarantee
all the imperative security properties. Although Odelu et
al.’s scheme can provide various security features, it is not
robust against DoS attacks (as discussed in Section 1). In
contrast, the proposed scheme can ensure all the important
security features (as shown in Table III). For instance, in
our proposed scheme, the USP can quickly make a decision
against an invalid authentication request, which helps our
scheme to be resilient against DoS attacks. Next, we evaluate
the performance of the proposed scheme in terms of the
computation and communication costs. In this regard, we first
conduct simulations of the cryptographic operations used by
all the schemes on an Ubuntu 12.04 virtual machine with
an Intel Core i5-4300 dual-core 2.60 GHz CPU (operating
as the USP/CS). To simulate a customer’s mobile device, we
use a HTC One smartphone with ARM Cortex-A9 MPCore
processor operating at 890 MHz. We use the JPBC library
Pbc-05.14 [21] and the JCE library [22] for evaluating the
computation times of different cryptographic operations used
in the proposed scheme and [4], [5], [6], [7], [9], and [10].
From Table IV we can see that the performance of the
proposed scheme in terms of computation and communication
costs is better than the others. Next, if we consider the existing
standards such as IEC 15118 and OCPP protocol for V2G
communications, then we find that like [4], [5], and [6], their
authentication and key-establishment schemes are based on
the computationally expensive ECDSA crypto-system, where
each signature generation takes 23.81 ms (at the user’s device)
and each signature verification (at the USP) takes 17.56 ms.
Besides, according to [32] and [33], these protocols also suffer
from several security issues (such as insecure against manin-the middle attacks, network impersonation attacks, DoS
attacks, etc.) and challenges. These protocols also expose some
important information such as customer name, vehicle identification number, charging location, and charging schedule,
which affects the customer’s privacy. Here, we argue that our
lightweight authentication and key establishment scheme can
easily be used by these underlying communication protocols
(such as IEC 15118 and OCPP) so that they can address all
the underlying security issues and ensure an enhanced security
level along with higher degree of efficiency.
Next, in order to comprehensively evaluate the practicality
of the proposed scheme, we consider the scalability of the
proposed scheme when deployed by organizations that own
charging stations. Since companies with large number of
charging stations do not exist yet, we use traditional gasoline refueling companies to obtain representative numbers.
In the USA, the biggest service providers are Shell (13727
stations), Chevron (6075 stations) and Exxon (5800 stations)
[34]. Current battery charging technologies for EVs may be
classified as either slow (energy flow rates of 2-6 KW) or
rapid charging (upto 150 KW) [35]. We consider EV models
Nissan LEAF (2018), Tesla Model S 100D and Mitsubishi
Outlander PHEV (2018) that come with battery capacities of
40 KHh, 100 KWh and 13.8 KWh, respectively. Assuming
a fast charging station with energy flow rate of 50 KWh, the
empty to full charging time for these vehicles is 1 hour, 2 hours
and 40 minutes respectively While a 150 KW rapid charger
takes 1 hour to charge the Tesla Model S 100D battery, the
Nissan and Mitsubishi models do not support this technology.
Thus we use one hour as a representative time for charging
current EVs in charging stations. The number of charging
points in CSs varies. For traditional (petrol) filling stations,
even in larger stations, studies indicate the average number is
18 (in Florida, [36]), i.e., 18 vehicles can fill up at the same
time. We use 18 as the number of charging points in a CS
and thus, 18 authentication requests are generated from a CS
every hour. Now, based on Table V, the communication cost
for the proposed protocol is 1802 bits = 226 bytes. On the
other hand, TCP + IP + Ethernet overhead = 20 + 20 + 24 =
64 bytes and during the authentication process 4 messages are
required to be exchanged. Therefore, the total communication
overhead = 226 + 4×64 = 482 bytes (approx. 500). The
computation time required at the USP is 0.00088 sec for
verifying an authentication request. Using Shell as an example,
we have 13800×18 = 248400 authentication requests per hour
(Shell has 13800 refilling stations). Therefore, the amount of
CPU time required every hour for verifying these transactions
is 0.00088 × 248400 = 219 seconds. A simple personal
computer or low end server can easily handle such computational requirements. The communication requirement of these
authentication requests is 500 × 248400 = 124200000 bytes
every hour = 276000 bits/sec = 276 Kbps. Thus, we conclude
that the proposed scheme can provide all the important security
properties and has lower (and practical) computation and
communication costs, and is hence suitable for EI-based V2G
communication.
VI. CONCLUSION
Secure and efficient key exchange is critical for ensuring secure data exchange in the Energy Internet. Aiming
at the problem of safe communication between EV users,
the USP and CSs, this paper proposed an efficient privacypreserving authentication scheme for EI-based Vehicle-to-Grid
communication. In this regard, only lightweight cryptographic
primitives such as one-way non-collision hash functions have
been considered. We quantified the performance of our scheme
using theoretical analysis and simulation tools. Our scheme is
resilient against many security attacks, efficient in computation
and communication, and compares favorably with existing
related schemes.
ACKNOWLEDGMENT
This research was supported by the National Research
Foundation, Prime Minister’s Office, Singapore under its Corporate Laboratory@University Scheme, National University
of Singapore, and Singapore Telecommunications Ltd. This
research was supported in part by Singapore Ministry of
Education Academic Research Fund Tier 1 (R-263-000-C13112). The authors would like to thank all the five reviewers
for their insightful comments and valuable suggestions.
REFERENCES
[1] J. Rifkin, "The third industrial revolution," Engineering and Technology,
l 3 7 26 27 2008
-----
[2] Z. Y. Dong, "Towards an intelligent future energy grid," The University
of Sydney, New South Wales, 2016.
[3] V. C. Gungor et al., "A Survey on Smart Grid Potential Applications
and Communication Requirements," IEEE Transactions on Industrial
Informatics, vol. 9, no. 1, pp. 28-42, 2013.
[4] Mohammadali, M. Sayad Haghighi, M. H. Tadayon, and A. Mohammadi
Nodooshan, "A Novel Identity-Based Key Establishment Method for
Advanced Metering Infrastructure in Smart Grid," IEEE Transactions
on Smart Grid, pp. 1-1, 2016.
[5] H. Nicanfar and V. C. M. Leung, "Multilayer Consensus ECC-Based
Password Authenticated Key-Exchange (MCEPAK) Protocol for Smart
Grid System," IEEE Transactions on Smart Grid, vol. 4, no. 1, pp. 253264, 2013.
[6] D. Wu and C. Zhou, "Fault-tolerant and scalable key management for
smart grid," IEEE Trans. Smart Grid, vol. 2 no. 2 pp. 371-378 Jun.
2011.
[7] J. Xia and Y. Wang, "Secure key distribution for the smart grid," IEEE
Trans. Smart Grid, vol. 3 no. 3 pp. 1437-1443 Aug. 2012.
[8] J. H. Park, M. Kim, and D. Kwon, "Security weakness in the smart grid
key distribution proposed by Xia and Wang" IEEE Trans. Smart Grid,
vol. 4 no. 3 pp. 1613-1614 Sep. 2013.
[9] J.-L. Tsai and N.-W. Lo, “Secure Anonymous Key Distribution Scheme
for Smart Grid,” IEEE Transactions on Smart Grid, vol. 7, no. 2, pp.
906–914, 2016.
[10] V. Odelu, A. K. Das, M. Wazid, and M. Conti, “Provably Secure Authenticated Key Agreement Scheme for Smart Grid,” IEEE Transactions
on Smart Grid, 2016, DOI: 10.1109/TSG.2016.2602282.
[11] P. Gope, and B. Sikdar, “Privacy-Aware Authenticated Key Agreement
Scheme for Secure Smart Grid Communication,” IEEE Transactions on
Smart Grid DOI:10.1109/TSG.2018.2844403, 2018.
[12] M. Mustafa, N. Zhang, and Z. Fan, “Smart electric vehicle charging:
Security analysis,” in Proc. IEEE PES ISGT, Washington, DC, USA,
Feb. 2013, pp. 1–6.
[13] H. Guo, Y. Wu, and M. Ma, “UBAPV2G: A unique batch authentication
protocol for vehicle-to-grid communications,” IEEE Trans. Smart Grid,
vol. 2, no. 4, pp. 707–714, Nov. 2011.
[14] H. Liu, H. Ning, and L. Yang, “Role-dependent privacy preservation
for secure V2G networks in the smart grid,” IEEE Trans. Inf. Forensics
Security, vol. 9, no. 2, pp. 208–220, Feb. 2014.
[15] Z. Yang, S. Yu, and C. Liu, “P [2] : Privacy-preserving communication
and precise reward architecture for V2G networks in smart grid,” IEEE
Trans. Smart Grid, vol. 2, no. 4, pp. 697–706, Dec. 2011.
[16] .Pbc library. Tech. rep. http://crypto.standford.edu/pbc/(accessed on 16
April 2017).
[17] Oracle Technology Network. Java Cryptography Architecture (JCA).
[Online].
[18] M. Bellare and P. Rogaway, “Entity Authentication and Key Distribution,” Advances in Cryptology - Crypto 1993, D. Stinson, ed. pp. 110125, Springer-Verlag, 1993.
[19] B. Schneier, Applied Cryptography (2nd edn), pp. 197-211, John Wiley
& Sons, New York, 1996.
[20] A. Menezes and S. Wanstone, Handbook of Applied Cryptography. Boca
Raton, FL, USA: CRC Press, 1996.
[21] H. Liu, H. Ning, Y. Zhang and L-T. Yang, “Aggregated-Proofs Based
Privacy-Preserving Authentication for V2G Networks in the Smart
Grid,” IEEE Trans. Smart Grid, vol. 66, no. 3, pp. 1722–1733, 2012.
[22] A. Abdallah and X. Shen, “Lightweight Authentication and PrivacyPreserving Scheme for V2G Connections,” IEEE Transactions on Vehicular Technology vol. 3, no. 4, pp. 2615–2629, 2017.
[23] N. Saxena and B.-J. Choi, “Authentication Scheme for Flexible Charging
and Discharging of Mobile Vehicles in the V2G Networks,” IEEE
Transactions on Information Forensics and Security vol. 11, no. 11,
pp. 1438–1452, 2017.
[24] D. He, S. Chan and M. Guizani, “A Privacy-friendly and efficient secure
communication framework for V2G networks,” IET Communications
vol. 12, no. 3, pp. 304–309, 2018.
[25] Y. Zhang, S. Gjessing, H. Liu, H. Ning, L-T. Yang and M. Guizani,
“Securing vehicle-to-grid communications in the smart grid,” IEEE
Wireless Communications vol. 20, no. 6, pp. 66–73, 2018.
[26] P. Gope, and B. Sikdar, "An Efficient Privacy-Friendly Hop-by-Hop
Data Aggregation Scheme for Smart Grids," IEEE Systems Journals,
10.1109/JSYST.2019.2899986, 2019.
[27] P. Gope, and B. Sikdar, "Lightweight and Privacy-Friendly Spatial Data
Aggregation for Secure Power Supply and Demand Management in
Smart-Grids," IEEE Transactions on Information Forensics & Security,
DOI 10 1109/TIFS 2018 2881730 2018
[28] A.-S. Sani et al., “Cyber security framework for Internet of Thingsbased Energy Internet,” Future Generation Computing Systems,
doi.org/10.1016/j.future.2018.01.029, 2018.
[29] A. Jindal, N. Kumar and M. Singh, “Internet of energy-based
demand response management scheme for smart homes and
PHEVs using SVM,” Future Generation Computing Systems,
doi.org/10.1016/j.future.2018.04.003, 2018.
[30] J. Shen, T. Zhou, F. Wei, X. Sun and Y. Xiang, “Privacy-Preserving and
Lightweight Key Agreement Protocol for V2G in the Social Internet of
Things,” IEEE Internet of Things Journal vol. 4, no. 4, pp. 2526–2536,
2017.
[31] K. Zhou, S. Yang and Z. Shaoa, “Energy Internet: The business
perspective,” Applied Energy vol. 178, no. 15, pp. 212–222, 2016.
[32] S. Lee et al., “Study on Analysis of Security Vulnerabilities and
Countermeasures in ISO/IEC 15118 Based Electric Vehicle Charging
Technology,” International Conference on IT Convergence and Security
(ICITCS), doi: 10.1109/ICITCS.2014.7021815, 2014.
[33] C. Alcaraz, J. Lopez and S. Wolthusen, “OCPP Protocol: Security
Threats and Challenges,” IEEE Trans. Smart Grid, vol. 8, no. 5, pp.
2452–2459, 2017.
[34] https://247wallst.com/retail/2017/04/21/10-retailers-that-controlamericas-gasoline-sales
[35] Y. Ligen, H. Vrubel and H. Girault, “Mobility from Renewable Electricity: Infrastructure Comparison for Battery and Hydrogen Fuel Cell
Vehicles,” World Electric Vehicle Journal, vol. 9, no. 1, 2018.
[36] Florida Department of Transportation, Trip Generation Characteristics
of Large Gas Stations/Convenience Stores and Student Apartments,
https://fdotwww.blob.core.windows.net/sitefinity/docs/defaultsource/content/planning/systems/programs/sm/tripgen/trip-generationof-convenience-stores.pdf.
Prosanta Gope (M’18) received the PhD degree in
computer science and information engineering from
National Cheng Kung University (NCKU), Tainan,
Taiwan, in 2015. He is currently working as a Lecturer in the department of Computer Science (Cyber
Security) at the University of Sheffield, UK. Prior
to this, Dr. Gope was working as a Research Fellow
in the department of Computer Science at National
University of Singapore (NUS). His research interests include lightweight authentication, authenticated
encryption, access control system, security in mobile
communication and cloud computing, lightweight security solutions for smart
grid and hardware security of the IoT devices. He has authored over 50 peerreviewed articles in several reputable international journals and conferences,
and has four filed patents. He received the Distinguished Ph.D. Scholar
Award in 2014 from the National Cheng Kung University, Tainan, Taiwan.
He currently serves as an Associate Editor of the IEEE INTERNET OF
THINGS JOURNAL, IEEE SENSORS JOURNAL, the SECURITY AND
COMMUNICATION NETWORKS and the MOBILE INFORMATION SYSTEMS JOURNAL.
Biplab Sikdar (S’98-M’02-SM’09) received the
Ph.D. degree in electrical engineering from the Rensselaer Polytechnic Institute, Troy, NY, USA, in 2001.
He was on the faculty of Rensselaer Polytechnic
Institute from 2001 to 2013, first as an Assistant
and then as an Associate Professor. He is currently an Associate Professor with the Department
of Electrical and Computer Engineering, National
University of Singapore, Singapore. His research
interests include computer networks, and security
for IoT and cyber physical systems. Dr. Sikdar is
a member of Eta Kappa Nu and Tau Beta Pi. He served as an Associate
Editor for the IEEE Transactions on Communications from 2007 to 2012. He
currently serves as an Associate Editor for the IEEE Transactions on Mobile
Computing.
-----
|
{
"disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1109/TSG.2019.2908698?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/TSG.2019.2908698, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "mit",
"status": "GREEN",
"url": "http://eprints.whiterose.ac.uk/156053/1/Final%20TSG.pdf"
}
| 2,019
|
[
"JournalArticle"
] | true
| 2019-04-02T00:00:00
|
[] | 19,266
|
en
|
[
{
"category": "Engineering",
"source": "s2-fos-model"
},
{
"category": "Environmental Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/02a06d4d6f9118145f72870c7e1a548707a44f04
|
[] | 0.863393
|
Improved Virtual Synchronous Generator Principle for Better Economic Dispatch and Stability in Grid-Connected Microgrids with Low Noise
|
02a06d4d6f9118145f72870c7e1a548707a44f04
|
Energies
|
[
{
"authorId": "2145892508",
"name": "Shruti Singh"
},
{
"authorId": "32235486",
"name": "D. Gao"
}
] |
{
"alternate_issns": null,
"alternate_names": null,
"alternate_urls": [
"http://nbn-resolving.de/urn/resolver.pl?urn=urn:nbn:ch:bel-155563",
"https://www.mdpi.com/journal/energies",
"http://www.mdpi.com/journal/energies"
],
"id": "1cd505d9-195d-4f99-b91c-169e872644d4",
"issn": "1996-1073",
"name": "Energies",
"type": "journal",
"url": "http://www.e-helvetica.nb.admin.ch/directAccess?callnumber=bel-155563"
}
|
The proper operation of microgrids depends on Economic Dispatch. It satisfies all requirements while lowering the microgrids’ overall operating and generation costs. Since distributed generators constitute a large portion of microgrids, seamless communication between generators is essential. While guaranteeing a reliable microgrid operation, this should be achieved with the fewest losses as possible. The distributed generator technology introduces noise into the system by design. To find the best economic dispatch strategy, noise was considered in this research as a limitation in grid-connected microgrids. The microgrid’s performance was improved, and the proposed technique also showed increased resilience. A virtual synchronous generator (VSG) control approach is proposed with a noiseless consensus-based algorithm to improve the power quality of microgrids. Voltage and frequency regulation modules are the foundation of the VSG paradigm. The synchronous generator’s second-order equation (hidden-pole configuration) was also used to represent the voltage of the stator and rotor motion. This study compared changes in power, frequency, and voltage for the microgrid by utilizing the described control approach using MATLAB. According to the findings, this method aids in controlling load and noise variations and offers distributed generators an efficient control strategy.
|
# energies
_Article_
## Improved Virtual Synchronous Generator Principle for Better Economic Dispatch and Stability in Grid-Connected Microgrids with Low Noise
**Shruti Singh and David Wenzhong Gao ***
Department of Electrical and Computer Engineering, University of Denver, Denver, CO 80210, USA;
shruti.singh@du.edu
*** Correspondence: wenzhong.gao@du.edu**
**Abstract: The proper operation of microgrids depends on Economic Dispatch. It satisfies all re-**
quirements while lowering the microgrids’ overall operating and generation costs. Since distributed
generators constitute a large portion of microgrids, seamless communication between generators
is essential. While guaranteeing a reliable microgrid operation, this should be achieved with the
fewest losses as possible. The distributed generator technology introduces noise into the system by
design. To find the best economic dispatch strategy, noise was considered in this research as a limitation in grid-connected microgrids. The microgrid’s performance was improved, and the proposed
technique also showed increased resilience. A virtual synchronous generator (VSG) control approach
is proposed with a noiseless consensus-based algorithm to improve the power quality of microgrids.
Voltage and frequency regulation modules are the foundation of the VSG paradigm. The synchronous
generator’s second-order equation (hidden-pole configuration) was also used to represent the voltage
of the stator and rotor motion. This study compared changes in power, frequency, and voltage for the
microgrid by utilizing the described control approach using MATLAB. According to the findings, this
method aids in controlling load and noise variations and offers distributed generators an efficient
control strategy.
**Citation: Singh, S.; Gao, D.W.**
Improved Virtual Synchronous
Generator Principle for Better
Economic Dispatch and Stability in
Grid-Connected Microgrids with
Low Noise. Energies 2023, 16, 4670.
[https://doi.org/10.3390/en16124670](https://doi.org/10.3390/en16124670)
Academic Editors: Favuzza Salvatore
and Jaser Sa’Ed
Received: 15 May 2023
Revised: 5 June 2023
Accepted: 11 June 2023
Published: 12 June 2023
**Copyright:** © 2023 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
[Attribution (CC BY) license (https://](https://creativecommons.org/licenses/by/4.0/)
[creativecommons.org/licenses/by/](https://creativecommons.org/licenses/by/4.0/)
4.0/).
**Keywords: microgrids; virtual synchronous generator; consensus-based algorithm; economic**
dispatch; power systems; distributed generators
**1. Introduction**
Economic Dispatch is an optimization problem used to reduce a system’s costs. It
is a significant issue in the world of power systems. All of the system’s constraints are
considered when determining the system’s minimum cost. Economic dispatch problems
have been solved using a variety of techniques, most frequently the quadratic convex
function [1,2]. The Lagrangian relaxation approach and the quadratic programming, respectively, were applied in [3,4]. The equal increment cost condition was taken into
consideration when studying consensus-based algorithms [5–8]. Economic dispatch and
demand-side management issues were resolved to reduce the overall costs [9–15]. Particle
Swarm Optimization was used in [16] for effective demand response in islanded microgrids.
Refs. [17,18] used the Dragonfly algorithm and the Cuckoo Search Algorithm to solve the
demand response in economic dispatch problems respectively. Ref. [19] introduced an
improved Genetic Algorithm for optimal dispatch. The system/microgrids was assumed
to be noiseless for the sake of the traditional economic dispatch problem. Noise from both
system components and the environment is present in real-time. The effectiveness and resilience of the microgrids are impacted by this. It restricts their stability as well. To stabilize
microgrids and improve their performance and resilience, noise must be incorporated into
the consensus-based algorithm economic dispatch problem.
-----
_Energies 2023, 16, 4670_ 2 of 19
Noise was taken into account in several analyses [20–22]. These studies created a
power-sharing strategy in microgrids that was parameter-independent and a noise-less
algorithm for better voltage and frequency synchronization. However, there has not been
much research in this particular area, which this study explored using the mentioned strategy. This approach was presented by [23] for isolated microgrids; however, grid-connected
microgrids were not considered. This study outlines the grid-connected microgrids’ noiseless economic dispatch problem. Additionally, this method does not require a central
controller, making the system cheaper and more cost-effective. Because a distributed strategy was used, a central controller was not required, which minimized the communication
complexity [24–29]. Ref. [30] proposed a multi-agent consensus control-based economic
dispatch algorithm allowing the microgrid to switch from isolated to grid-connected modes
more reliably.
A key role is played by inverters in the interaction between the distribution network
and the microgrid [31]. Conventionally, the droop control technique [32] is employed,
although it is extremely vulnerable to changes in load. An improved droop control approach has been suggested in many publications. Some of them are dependent on the
inverter’s output voltage. The drawback is that the droop coefficient causes the frequency
to be too unstable. In DC microgrids, a discrete consensus-based adaptive droop control
technique has also been put forth [33,34]. In several articles, the P/Q control method has
been applied. The U/f control approach has primarily been employed in island mode. It is
possible to create two sets of control systems using the P/Q and U/f control methods in conjunction with switching control components between the two [35]. However, because the
two strategies have a complex structure, this system is challenging to create.
This paper introduces the novel concept of economic dispatch with noise effects on
a grid-connected microgrid’s performance. A consensus-based algorithm along with a
virtual synchronous generator strategy was used to reduce the fluctuations in voltage and
frequency of the microgrids due to various noise levels. Two economic dispatch algorithms,
i.e., the Lagrange formulation and the particle swarm optimization technique were compared to analyze their effect on the grid-connected microgrid’s overall performance. This
paper is divided into multiple parts. The economic dispatch problem and PSO algorithm
are defined in Section 2. Section 3 introduces the microgrid structure. The distributed noiseresilient economic dispatch approach is presented in Section 4 [23]. The VSG model and
control approach are introduced in Section 5. The results and the discussion are explained
in Section 6, and the conclusions are stated in Section 7.
**2. Economic Dispatch Formulation**
_2.1. Lagrange Formulation_
The economic dispatch issue for a microgrid that is connected to a grid is defined
using the Lagrangian method. The goal purpose of the microgrid is first established. The
most prevalent use of this function is to address economic dispatch issues. The cost of a
generator in a microgrid system can be expressed using the following equation [36], taking
into account all the generation units:
_n_ _n_
### ∑i=1 [P][i][C][i][ =][ ∑]i=1 [x][i][C]i[2] [+][ y][i][C][i] [+][ z][i] (1a)
where
_PiCi is the generator cost_
_xi, yi, zi are the cost coefficients_
_Ci is the generator’s total power output._
To solve the economic dispatch issue, we aimed at lowering the microgrid’s generation
costs. Equation (1a) becomes:
_n_ _n_
min ∑i=1 _[P][i][C][i][ =][ min]_ [∑]i=1 _[x][i][C]i[2]_ [+][ y][i][C][i] [+][ z][i] (1b)
-----
_Energies 2023, 16, 4670_ 3 of 19
Additionally, the generator’s total electric output can be defined as [36]:
_n_
### ∑i=1 [C][i][ =][ C][D][ +][ C][loss][, for][ C]i[min]< Ci < Ci[max] (2)
where
_CD = total load; Closs = losses during transmission_
_Ci[min]_ = minimum generation limit of generator i
_Ci[max]_ = maximum generation limit of generator i.
In formulating the Lagrangian function, the above three equations together become [36]:
_n_ _n_ _n_ _n_ � �
_L(C1, C2, . . . Cn) = ∑i=1_ _[P][i][C][i][ +][ λ][(][C][D][ +][ C][loss][ −]_ [∑]i=1 _[C][i][)+]_ [∑]i=1 _[u][x][(][C][i][ −]_ _[C]i[max])+_ ∑i=1 _[u][y]_ _Ci[min]_ _−_ _Ci_ (3)
where λ, ux, uy are Lagrange multipliers.
Calculating each generator’s incremental cost (IC1, IC2, . . ., ICn) is necessary to discover a solution to the aforementioned economic dispatch problem. These incremental
costs for various generators should be equal to determine the microgrid’s minimal cost, i.e.,
_IC1 = IC2 = . . . = ICn_
where n defines the number of generation units
This problem’s most popular solution was used here [36]:
_λi=_ _[∂]∂[P]C[i][C]i_ _[i]_ [=][ 2][x][i][C][i][ +][ y][i][ =][ λ][∗][, for][ C]i[min]< Ci < Ci[max]
_λi=_ _[∂]∂[P]C[i][C]i_ _[i]_ [=][ 2][x][i][C][i][ +][ y][i][ <][ λ][∗][, for][ C][i][ =][ C]i[max]
_λi=_ _[∂]∂[P]C[i][C]i_ _[i]_ [=][ 2][x][i][C][i][ +][ y][i][ >][ λ][∗][, for][ C][i][ =][ C]i[min]
(4)
where
_λi = incremental cost_
_λ[∗]_ = optimal incremental cost.
To determine an economic dispatch schedule for the microgrid, the economic dispatch
issue must take into consideration the generation restrictions for each unit. The economic
dispatch problem is rather simple to resolve and takes into account all the restraints of
generators. However, when addressing the economic dispatch issue for microgrids, the
majority of problems have some limitations that must be taken into account. For any issues
relating to economic dispatch, the aforementioned equations serve as the fundamental
problem formulation.
_2.2. Particle Swarm Optimization (PSO) Algorithm_
Particle Swarm Optimization is a computational method that was inspired by the
movement of bird flocks and other organisms/particles by Kennedy, Eberhart, and Shi [16].
It is a population-based optimization tool in which particles change position by taking
into account their velocity, their own experience, and the experience of their neighboring
particles. The position and velocity of particle j in N-dimensional space are represented
as aj = (aj1, aj2, . . . ajN) and bi = (bj1, bj2, . . . bjN). The best position for this particle can
be represented as Abestj = (aj[A]1 [,][ a]j[A]2 [,][ . . .][ a]jN[A] [). The best position for the neighboring particle]
can be represented as Bbest = (a1[B] [,][ a]2[B] [,][ . . .][ a][B]N[). New modified position and velocity can be]
formulated as:
� � �
_b[k]jN[+][1][=][ ξ][.][ b][k]jN[+][m][1][r][1][ ×]_ _AbestjN −_ _a[k]jN_ [) +][ m][2][r][2][ ×] _BbestN −_ _a[k]jN_ and,
_a[k]jN[+][1][=][ a][k]jN[+][b][k]jN[+][1]_
(4a)
where
-----
_Energies 2023, 16, 4670_ 4 of 19
_k = number of iterations_
Ξ = inertia weight factor
_m1, m2 = acceleration constant_
_r1, r2 = random number within the range [0, 1]_
The inertia weight factor and the acceleration constant affect the performance significantly. The weight factor provides the required momentum for particles to move around
in N-dimensional space. The acceleration constant indicates the weight of stochastic acceleration terms that help in pulling all particles towards Abestj and Bbest positions. This
algorithm is used iteratively to find convergence in optimal dispatch solutions. The best
incremental cost was determined using this method and was then sent to the agents to
either accept or modify the output power of generators to minimize the effect of noise on
system parameters’ fluctuations.
**3. Microgrid Structure**
Four generator units constituted the microgrid that was the subject of this paper’s
investigation. It featured two coal-based generator units, a wind generator, and a solar/photovoltaic (PV) generator and was connected to the grid. The quadratic Equation
(1a) expresses the cost function of the units. Closs was estimated to make up 7% of the total
load. Table 1 below provides the cost-coefficients values for each unit, the minimum power
generation limits, and the maximum power generation limits for all the units [36].
**Table 1. List of parameters for generators [36].**
**Unit** **Cmin (kW)** **Cmax (kW)** **x** **y** **z**
1 4 18 0.070 2.15 56
2 8 40 0.080 1.15 50
3 5 25 0.070 3.3 41
4 5 40 0.056 3.4 36
**4. Economic Dispatch with Consensus-Based Approach for Noise-Less Communication [19]**
The strategy described in [23] is explained in this section. A microgrid’s communication link was developed. Each generator unit had a corresponding agent that gathered
data from its corresponding unit. These data were read by a certain agent. All the agents
that were part of this communication system could share data and communicate with
each other [34]. We had four agents in total, each of which was connected to a different
generation unit on our microgrid in grid-connected mode; there were four generation units.
The information data received, collected, and processed by an agent was also exchanged
with other agents. This exchange helped understand the current status of each unit. To
reduce the overall cost of the microgrid system, the information received from the agent(s)
was used to modify the output power. Noise from the components, surroundings, and
electric/magnetic interference was taken into account for this analysis. This method includes noise that accumulated as a result of the communication between the units and that
resulting from the communication between the units and the agents; it was considered
in modeling as Gaussian noise [16]. The communication links between the agents were
indicated as c12, c21, c23, c32, c34, c43, c13, c31, and so on.
Each agent determined the corresponding incremental cost of each unit before exchanging it with the others. Based on the data, the set point of the output power was
determined and supplied to the appropriate generation units. To address the economic dispatch issue, the units modify their power generation capacity to have an equal incremental
cost. This reduces the cost of microgrids. According to Refs. [23,36]:
-----
_Energies 2023, 16, 4670_ 5 of 19
Z[k + 1] = Z[k] + µ[k][P z[k] + WN[k]]
P = H[′]NH
_−_
W = H[′]N
H = H2 − H1
where
Z[k] = incremental cost of a unit at the kth iteration,
Z[k + 1] = incremental cost of a unit at the (k + 1)th iteration,
µ[k] = recursive step size,
N = r r diagonal matrix with link control gain as its diagonal elements,
_×_
H1 and H2 = r × n matrix where the rows are elementary vectors,
N[k] = communication link noise.
(5)
1 1 0 0
_−_
1 1 0 0
_−_
0 1 1 0
_−_
0 1 1 0
_−_
0 0 1 1
_−_
0 0 1 1
_−_
������������
������������
������������
H1 =
������������
0 1 0 0
1 0 0 0
0 0 1 0
0 1 0 0
0 0 0 1
0 0 1 0
; H2 =
������������
1 0 0 0
0 1 0 0
0 1 0 0
0 0 1 0
0 0 1 0
0 0 0 1
and H =
������������
= H2 − H1
N (small noise) = diag [0.2 0.2 0.2 0.2 0.2 0.2]
N (medium noise) = diag [0.5 0.5 0.5 0.5 0.5 0.5]
N (large noise) = diag [0.8 0.8 0.8 0.8 0.8 0.8]
Similarly, P and W can be determined from (5).
To lessen the effects of noise, we averaged the additional costs of the units. This
produced a microgrid that was more durable, stable, and free of (or, with less)
communication noise [23].
_Zavg[k + 1] =_ _k+1_ 1 [∑][k]j=[+]1[1] _[Z][[][j][]=]_ _k+1_ 1 [∑][k]j=1 _[z][[][j][] +][ Z][[][k][ +][ 1][]]_
(6)
= Zavg[k] − _k+1_ 1 _[Z][avg][[][k][] +]_ _k+1_ 1 _[Z][[][k][ +][ 1][]]_
The noise-less economic dispatch with the consensus-based strategy using (5) and
(6) is [23,36]:
_Z[k + 1] = Z[k] + µ[k][P z[k] + WN[k]]_
1
_Zavg[k + 1] = Zavg[k] +_ �Z[k + 1] − _Zavg[k]�_ (7)
_k + 1_
where Zavg[k + 1] are the desired set points for the unit incremental costs.
This method is iterative, and an estimate was created using the step size. It was
then averaged in the subsequent stages to limit the effect of noise. For each step size, the
consensus problem was iteratively solved. The flowchart for the consensus-based economic
dispatch algorithm is shown in Figure 1.
-----
_Energies 2023, 16, x FOR PEER REVIEW_ 6 of 19
_Energies 2023, 16, 4670_ 6 of 19
**Figure 1. Figure 1. Algorithm flowchart.Algorithm flowchart.**
**5. Virtual Synchronous Generator (VSG)**
**5. Virtual Synchronous Generator (VSG)**
The VSG control system [37] is a comprehensive system that combines several modules
The VSG control system [37] is a comprehensive system that combines several mod
to enable an efficient and effective electricity management. It is based on the VSG strategy,
ules to enable an efficient and effective electricity management. It is based on the VSG
which is responsible for simulating the system’s performance and determining its optimal
strategy, which is responsible for simulating the system’s performance and determining
power output, and contains the Frequency Regulation Module, which adjusts the frequency
its optimal power output, and contains the Frequency Regulation Module, which adjusts
of the output power to match that of the grid, the Voltage Regulation Module, which
the frequency of the output power to match that of the grid, the Voltage Regulation
regulates the voltage of the output power, the grid-connected mode of the Control Module,
Module, which regulates the voltage of the output power, the grid-connected mode of the
which ensures the output power is synchronized with the grid, the SPWM Modulation
Control Module, which ensures the output power is synchronized with the grid, the
Module, which adjusts the output current amplitude, and the Sampling Calculation Module,
SPWM Modulation Module, which adjusts the output current amplitude, and the Sam
which calculates the output power by sampling the input signal. All of these modules
pling Calculation Module, which calculates the output power by sampling the input work together to provide a reliable and secure electricity management system as shown in
signal. All of these modules work together to provide a reliable and secure electricity Figure 2 below.
management system as shown in Figure 2 below. When exposed to disturbances and load variations, power electronic inverters have
poor system stability [37]. The traditional SG rotation has a significant output inductance
and moment of inertia. Therefore, the microgrid’s power supply can be compared to
the prime mover by reproducing the exterior characteristics of the microgrid into an SG.
The microgrid inverter’s inverter and filter modules provide the electric energy produced
-----
_Energies 2023, 16, 4670_ 7 of 19
_Energies 2023, 16, x FOR PEER REVIEW_ 7 of 19
by distributed sources to the load, while the energy storage system stores the residual
electric energy.
###### Figure 2. VSG control strategy block diagram.
##### When exposed to disturbances and load variations, power electronic inver poor system stability [37]. The traditional SG rotation has a significant output in and moment of inertia. Therefore, the microgrid’s power supply can be compar prime mover by reproducing the exterior characteristics of the microgrid into an microgrid inverter’s inverter and filter modules provide the electric energy pro distributed sources to the load, while the energy storage system stores the resid tric energy. Ref. [38] presented the SG second-order equation modeling, which include lowing equations.
**Figure 2. Figure 2.Stator voltage equation: VSG control strategy block diagram.VSG control strategy block diagram.**
###### Figure 2. VSG control strategy block diagram.
##### When exposed to disturbances and load variations, power electronic inver poor system stability [37]. The traditional SG rotation has a significant output in and moment of inertia. Therefore, the microgrid’s power supply can be compar prime mover by reproducing the exterior characteristics of the microgrid into an microgrid inverter’s inverter and filter modules provide the electric energy pro distributed sources to the load, while the energy storage system stores the resid tric energy. Ref. [38] presented the SG second-order equation modeling, which include lowing equations.
7 of 19
7 of 19
by distributed sources to the load, while the energy storage system stores the residual
electric energy.
When exposed to disturbances and load variations, power electronic inverters have Ref. [38] presented the SG second-order equation modeling, which includes the fol-U[�] ������ = E[�] −ΔU[�]
poor system stability [37]. The traditional SG rotation has a significant output inductance lowing equations.
##### where and moment of inertia. Therefore, the microgrid’s power supply can be compared to the Stator voltage equation: . . . Uprime mover by reproducing the exterior characteristics of the microgrid into an SG. The [�] ������ = three-phase reference voltage Urefabc = E − ∆U (8) Ewheremicrogrid inverter’s inverter and filter modules provide the electric energy produced by [�] [= ]electromotive force
distributed sources to the load, while the energy storage system stores the residual elec-.
##### ΔUtric energy. U. refabc[�] = voltage drop caused by virtual synchronous impedance. = three-phase reference voltage
E = electromotive forceRef. [38] presented the SG second-order equation modeling, which includes the fol-The output current I0 of the inverter is equal to the synchronous genera
.
##### current; rlowing equations. ∆U = voltage drop caused by virtual synchronous impedance.a and Xd, respectively, are armature resistance and synchronous reac obtain ΔU, a vector multiplication is used for (rStator voltage equation: The output current I0 of the inverter is equal to the synchronous generator statora + jXd) and I0. The module in provides a corresponding control signal in line with Ucurrent; ra and Xd, respectively, are armature resistance and synchronous reactance. ToU[�] ������ = E[�] −ΔU[�] refabc after E is corrected f(8)
obtain ∆U, a vector multiplication is used for (ra + jXd) and I0. The module in Figure 3
##### tion.
provides a corresponding control signal in line with Uwhere refabc after E is corrected for deviation.
U[�] ������ = three-phase reference voltage
E[�] [= ]electromotive force
ΔU[�] = voltage drop caused by virtual synchronous impedance.
The output current I0 of the inverter is equal to the synchronous generator stator
current; ra and Xd, respectively, are armature resistance and synchronous reactance. To
obtain ΔU, a vector multiplication is used for (ra + jXd) and I0. The module in Figure 3
provides a corresponding control signal in line with Urefabc after E is corrected for deviation.
###### Figure 3. Figure 3. Stator voltage model.Stator voltage model.
The rotor motion model promotes system stability, as shown in Figure 4. When Pm
and Pe do not match, it obviates this by adding J and D; dθ is the correction angle.
E[�] [= ]electromotive force
ΔU[�] = voltage drop caused by virtual synchronous impedance.
The output current I0 of the inverter is equal to the synchronous generator stator
current; ra and Xd, respectively, are armature resistance and synchronous reactance. To
obtain ΔU, a vector multiplication is used for (ra + jXd) and I0. The module in Figure 3
provides a corresponding control signal in line with Urefabc after E is corrected for devia-
tion.
-----
D = damping co-efficientThe rotor motion model promotes system stability, as shown in Figure 4. When Pm
_Energies 2023, 16, 4670_ 8 of 19
ω = angular velocity and Pe do not match, it obviates this by adding J and D; dθ is the correction angle.
ωR = rated angular velocity For the rotor motion model [39]:
∆ω = [1] −D∆ω) dt (9)
J [�(x][�] ω[−x][�]
ω = ∆ω + ω� (10)
where
Δω = angular velocity difference
Xm = mechanical power
Xe = electromagnetic power
**Figure 4. J = moment of inertia Figure 4. Rotor motion model.Rotor motion model.**
D = damping co-efficient
The frequency module in Figure 5 includes the grid-connected sinusoidal wave SS, For the rotor motion model [39]:
ω = angular velocity
the system frequency fωR = rated angular velocity V, the reference active power P� ref, the reference frequency fref, and
the grid side frequency fg. The frequency regulation module chooses its reference value ∆ω = [1] ( [x][m][ −] [x][e] D∆ω) dt (9)
_−_
J ω
based on the fg range once the grid-connected signal SS has been sent by Judger2. The
reference value is chosen as fg if it falls within the typical range and as fref if it does not. fref
ω = ∆ω + ωr (10)
is used as the reference value while the system is in islanded mode.
where
∆ω = angular velocity difference
Xm = mechanical power
Xe = electromagnetic power
J = moment of inertia
**Figure 4. D = damping co-efficientRotor motion model.**
ω = angular velocity
ωR = rated angular velocityThe frequency module in Figure 5 includes the grid-connected sinusoidal wave SS,
the system frequency fThe frequency module in FigureV, the reference active power P 5 includes the grid-connected sinusoidal wave SS, theref, the reference frequency fref, and
the grid side frequency fsystem frequency fV, the reference active power Pg. The frequency regulation module chooses its reference value ref, the reference frequency fref, and the
**Figure 5. based on the fgrid side frequency fFrequency regulation module. g range once the grid-connected signal SS has been sent by Judgerg. The frequency regulation module chooses its reference value based2. The**
reference value is chosen as fon the fg range once the grid-connected signal SS has been sent by Judgerg if it falls within the typical range and as fref if it does not. f2. The referenceref
value is chosen as fis used as the reference value while the system is in islanded mode. To Judger1, the frequency deviation Δf is provided. Depending on the interval in g if it falls within the typical range and as fref if it does not. fref is used
which the frequency difference is situated, Judger1 passes on to the regulator in the next
as the reference value while the system is in islanded mode.
stage. The secondary frequency regulation is simulated by PI1, and the frequency module
regulates the main frequency per the co-efficient kp. It also regulates and switches to a
secondary frequency, if necessary. The Synchronous Generator maintains system frequency stability, both primary and secondary.
Qref and Q0 are the inputs to the virtual voltage regulation module. The difference is
multiplied by the voltage-reactive co-efficient kU to obtain the electro-motive force for
power adjustment (reactive), ΔE1. To determine ΔE2, which is the terminal voltage electro-motive force, the effective capacitor voltage Uc in the filter module’s reference voltage
**Figure 5. Frequency regulation module.**
**Figure 5. Frequency regulation module.**
To Judger1, the frequency deviation ∆f is provided. Depending on the interval in
which the frequency difference is situated, JudgerTo Judger1, the frequency deviation Δf is provided. Depending on the interval in 1 passes on to the regulator in the next
which the frequency difference is situated, Judgerstage. The secondary frequency regulation is simulated by PI1 passes on to the regulator in the next 1, and the frequency module
stage. The secondary frequency regulation is simulated by PIregulates the main frequency per the co-efficient kp. It also regulates and switches to a1, and the frequency module
secondary frequency, if necessary. The Synchronous Generator maintains system frequency
regulates the main frequency per the co-efficient kp. It also regulates and switches to a
stability, both primary and secondary.
secondary frequency, if necessary. The Synchronous Generator maintains system frequency stability, both primary and secondary. Qref and Q0 are the inputs to the virtual voltage regulation module. The difference
is multiplied by the voltage-reactive co-efficient kQref and Q0 are the inputs to the virtual voltage regulation module. The difference is U to obtain the electro-motive force for
power adjustment (reactive), ∆E1. To determine ∆E2, which is the terminal voltage electro
multiplied by the voltage-reactive co-efficient kU to obtain the electro-motive force for
motive force, the effective capacitor voltage Uc in the filter module’s reference voltage
power adjustment (reactive), ΔE1. To determine ΔE2, which is the terminal voltage electro-motive force, the effective capacitor voltage UUref differential value is translated into amplitude. When the synchronous generator isc in the filter module’s reference voltage
operating in no-load mode, Eref is the reference electro-motive force, whereas dE is the
corrected electromotive force when in grid-connected mode, as shown in Figure 6. The
U–Q relationship is as follows:
∆ω = angular velocity difference
Xm = mechanical power
Xe = electromagnetic power
J = moment of inertia
**Figure 4. D = damping co-efficientRotor motion model.**
ω = angular velocity
ωR = rated angular velocityThe frequency module in Figure 5 includes the grid-connected sinusoidal wave SS,
the system frequency fThe frequency module in FigureV, the reference active power P 5 includes the grid-connected sinusoidal wave SS, theref, the reference frequency f
[�]
∆ω = −D∆ω) dt
J [�(x][�] ω
ω = ∆ω + ω�
where
Δω = angular velocity difference
Xm = mechanical power
-----
##### (Q Q) ( )
_Energies 2023, 16, 4670_ Uref differential value is translated into amplitude. When the synchronous generator is operat-9 of 19
##### where
ing in no-load mode, Eref is the reference electro-motive force, whereas dE is the corrected
##### kSGU = SG voltage-reactive coefficient electromotive force when in grid-connected mode, as shown in Figure 6. The U–Q rela- USGref = reference values of voltage tionship is as follows: QSGref = reference values of reactive power U −U − UUSGrefSGref = k = kSGUSGU (Q (QSGrefSGref − − Q) Q) (11)(11)
where
where
kkSGUSGU = SG voltage-reactive coefficient = SG voltage-reactive coefficient
UUSGrefSGref = reference values of voltage = reference values of voltage
QQSGrefSGref = reference values of reactive power = reference values of reactive power
###### Figure 6. Voltage regulation module. [Where Q]ref = [ reference reactive power, Q]0 = [ control output ] reactive power.
where
kkSGUSGU = SG voltage-reactive coefficient = SG voltage-reactive coefficient
UUSGrefSGref = reference values of voltage = reference values of voltage
QQSGrefSGref = reference values of reactive power = reference values of reactive power
**Figure 6. Figure 6. Voltage regulation module. Where QVoltage regulation module. [Where Q]refref = = reference reactive power, Q[ reference reactive power, Q]0 =0 = control output[ control output ]**
##### Figure 7 depicts the control module in grid-connected mode. The grid-connected mod-reactive power. reactive power. ule completes pre-synchronization when the sinusoidal wave SS switches from ‘0’ position to
Figure 7 depicts the control module in grid-connected mode. The grid-connected
Figure 7 depicts the control module in grid-connected mode. The grid-connected mod
##### ‘1’ position. The PI3 regulator and Judgermodule completes pre-synchronization when the sinusoidal wave SS switches from ‘0’4 receive the difference between [φ]g [and ][φ]V (voltage
ule completes pre-synchronization when the sinusoidal wave SS switches from ‘0’ position to
##### phase angles)[. The rotor motion model receives the value provided by the PI]position to ‘1’ position. The PI‘1’ position. The PI3 regulator and Judger3 regulator and Judger4 receive the difference between 4 receive the difference between[3][ as ][dθ][φ]g[. Judger][and ][φ]V (4voltage ϕg
and ϕV (voltage phase angles). The rotor motion model receives the value provided by
##### chooses the next input value based on the interval in which the difference is found. The PIphase angles)[. The rotor motion model receives the value provided by the PI][3][ as ][dθ]2, a [. Judger]4
the PI3 as dθ. Judger4 chooses the next input value based on the interval in which the
##### regulator, and Judgerchooses the next input value based on the interval in which the difference is found. The PI3 receive the difference between [|U][�][�|][ and ] [U]amp. [The virtual voltage ] 2., a
difference is found. The PI2, a regulator, and Judger3 receive the difference between Ug
##### regulator module receives the value from the PIregulator, and Judger3 receive the difference between2 as dE. Similar to Judger[|U][�][�|][ and ]4[U], Judgeramp. [The virtual voltage ]3 chooses ��� ���
regulator module receives the value from the PIand Uamp. The virtual voltage regulator module receives the value from the PI2 as dE. Similar to Judger4, Judger3 chooses 2 as dE.
##### the following input value based on the interval in which the difference is placed. It is deter-Similar to Judger4, Judger3 chooses the following input value based on the interval in
the following input value based on the interval in which the difference is placed. It is deter
##### mined by the difference between the two sides when determining the frequency. When all which the difference is placed. It is determined by the difference between the two sides
mined by the difference between the two sides when determining the frequency. When all
##### three Judgers are selected as 1 and the switch signal is changed from the “0” position to the when determining the frequency. When all three Judgers are selected as 1 and the switchthree Judgers are selected as 1 and the switch signal is changed from the “0” position to the
signal is changed from the “0” position to the “1” position, the pre-synchronization phase
##### “1” position, the pre-synchronization phase is said to be complete. “1” position, the pre-synchronization phase is said to be complete.
is said to be complete.
**Figure 7. Grid control module.**
###### Figure 7. Grid control module. Figure 7. Grid control module.
**Figure 7. Grid control module.**
###### Voltage regulation module. [Where Q]ref = [ reference reactive power, Q]
-----
_Energies 2023, 16, 4670_ 10 of 19
The system specifications for the LCL filter and the three-phase bridge inverter are
provided in Table 2.
The system specifications for the LCL filter and the three-phase bridge inverter are
**Table 2. List of the components.**
provided in Table 2.
**Components** **Values**
**Table 2. List of the components.L1** 6 mH
L2 1.5 mH
**Components** **Values**
C 6 micro-F
J L1 0.15 kg.m6 mH[2]
L2 1.5 mH
Kp, kUC 800 kW/Hz, 0.8 Hz/kVar 6 micro-F
PWM freq J 25 kHz 0.15 kg·m[2]
P at constant load Kp, kU 800 kW/Hz, 0.8 Hz/kVar10 kW
PWM freq 25 kHz
Q at constant load 8 kVar
P at constant load 10 kW
Q at constant loadra 0.05 ohm 8 kVar
Xd ra 0.05 H 0.05 ohm
P variable Xd 5 kW 0.05 H
P variable 5 kW
Q variable 3 kVar
Q variable 3 kVar
**6. Results and Discussion**
**6. Results and Discussion**
Four different scenarios were used to examine the grid-connected microgrid. The
Four different scenarios were used to examine the grid-connected microgrid. The
microgrid was initially evaluated when there was no noise in the system. The Lagrange
microgrid was initially evaluated when there was no noise in the system. The Lagrange
method described in the preceding section and the PSO algorithm were evaluated to see
method described in the preceding section and the PSO algorithm were evaluated to see
how well they operated in the absence of noise. In the second scenario, the system was
how well they operated in the absence of noise. In the second scenario, the system was
subjected to noise with a variance of 0.2, and the performance was tracked. The noise
subjected to noise with a variance of 0.2, and the performance was tracked. The noise
variance was raised to 0.5 in the third test, and it was set to 0.8 in the final condition ex
variance was raised to 0.5 in the third test, and it was set to 0.8 in the final condition
amined. MATLAB was used to examine how well the microgrid performed under vari
examined. MATLAB was used to examine how well the microgrid performed under
ous noise circumstances with and without the VSG control approach. Figure 8 shows the
various noise circumstances with and without the VSG control approach. Figure 8 shows
network figure of the algorithm used. The incremental costs from each generator were
the network figure of the algorithm used. The incremental costs from each generator were
shared with the agents. These agents shared data and decided if the incremental cost was
shared with the agents. These agents shared data and decided if the incremental cost was
optimal. If this was not the case, the information was passed to the generator to adjust its
optimal. If this was not the case, the information was passed to the generator to adjust
output power until the optimal incremental cost criterion was met. Once an optimal
its output power until the optimal incremental cost criterion was met. Once an optimal
economic dispatch solution was found, the total output power was sent to the VSG which
economic dispatch solution was found, the total output power was sent to the VSG which
was then used to meet the load demand or sent to the grid to fulfill any power deficit. The
was then used to meet the load demand or sent to the grid to fulfill any power deficit. The
use of a consensus-based algorithm and of the VSG strategy helped reduce the noise
use of a consensus-based algorithm and of the VSG strategy helped reduce the noise effects
effects and stabilize the microgrid.
and stabilize the microgrid.
**Figure 8. Figure 8.Network figure. Network figure.**
With the introduction of various noise levels, the power output of the four generating
units was observed, and we tried to model the ideal dispatch schedule in all instances.
-----
ing units was observed, and we tried to model the ideal dispatch schedule in all instancing units was observed, and we tried to model the ideal dispatch schedule in all instanc
_Energies 2023, 16, 4670_ es. To demonstrate the system’s stability and the incremental cost for various noise levels, es. To demonstrate the system’s stability and the incremental cost for various noise levels, 11 of 19
a comparison was evaluated. Figures 9 and 10 demonstrate, respectively, the fluctuating a comparison was evaluated. Figures 9 and 10 demonstrate, respectively, the fluctuating
power output of the four generating units during a 60 s period with and without the VSG power output of the four generating units during a 60 s period with and without the VSG
using the Lagrange formulation. It took about 15 s without the VSG to stabilize the power using the Lagrange formulation. It took about 15 s without the VSG to stabilize the power
output of all generators, whereas with the VSG, it took about 12 s. For a low (0.2) and output of all generators, whereas with the VSG, it took about 12 s. For a low (0.2) and To demonstrate the system’s stability and the incremental cost for various noise levels, a
medium (0.5) noise variance, it the system required about 20 s and 25 s, respectively to medium (0.5) noise variance, it the system required about 20 s and 25 s, respectively to comparison was evaluated. Figures 9 and 10 demonstrate, respectively, the fluctuating
establish a consistent producing power output. Figures 11 and 12 show this observation. establish a consistent producing power output. Figures 11 and 12 show this observation. power output of the four generating units during a 60 s period with and without the VSG
For a 0.8 noise level, shown in Figure 13, the microgrid required about 45 s to reach a For a 0.8 noise level, shown in Figure 13, the microgrid required about 45 s to reach a using the Lagrange formulation. It took about 15 s without the VSG to stabilize the power
stable power output. It was observed that the economic dispatch solution was not stable stable power output. It was observed that the economic dispatch solution was not stable
output of all generators, whereas with the VSG, it took about 12 s. For a low (0.2) and
and was less reliable for noise levels without the VSG strategy. It is clear from the graphs and was less reliable for noise levels without the VSG strategy. It is clear from the graphs
that the system required a few seconds to reach the desired constant value. that the system required a few seconds to reach the desired constant value. medium (0.5) noise variance, it the system required about 20 s and 25 s, respectively to
establish a consistent producing power output. FiguresThe system required lesser time to each the intended value of optimal power when The system required lesser time to each the intended value of optimal power when 11 and 12 show this observation.
was connected with the VSG, as shown in Figure 14. For low, medium, and high noise was connected with the VSG, as shown in Figure 14. For low, medium, and high noise For a 0.8 noise level, shown in Figure 13, the microgrid required about 45 s to reach a stable
variance for all four units, it can be seen in the graph that there was a 5–10 s improvement variance for all four units, it can be seen in the graph that there was a 5–10 s improvement power output. It was observed that the economic dispatch solution was not stable and was
for each unit in noise variance when the Lagrange method was used with the VSG control for each unit in noise variance when the Lagrange method was used with the VSG control less reliable for noise levels without the VSG strategy. It is clear from the graphs that the
strategy. strategy.
system required a few seconds to reach the desired constant value.
3030
2525
2020
1515
1010
55
00
00 1010 2020 3030 4040 5050 6060
Time (sec)Time (sec)
**Figure 9. Figure 9. Generator output power in kW without the VSG in the absence of noise. Generator output power in kW without the VSG in the absence of noise.**
**Figure 9. Generator output power in kW without the VSG in the absence of noise.**
3030
2525
2020
1515
1010
55
00
00 1010 2020 3030 4040 5050 6060
_Energies 2023, 16, x FOR PEER REVIEW_ 12 of 19
Time (sec)Time (sec)
**Figure 10. Figure 10. Figure 10.Generator output power in kW with the VSG in the absence of noise. Generator output power in kW with the VSG in the absence of noise. Generator output power in kW with the VSG in the absence of noise.**
30
25
20
15
10
5
0
0 10 20 30 40 50 60
Time (sec)
**Figure 11. Generator output power in kW with a 0.2 noise variance and without the VSG.**
**Figure 11. Generator output power in kW with a 0.2 noise variance and without the VSG.**
-----
0 10 20 Time (sec)30 40 50 60
Time (sec)
_Energies 2023, 16, 4670_ 12 of 19
**Figure 11. Generator output power in kW with a 0.2 noise variance and without the VSG.**
**Figure 11. Generator output power in kW with a 0.2 noise variance and without the VSG.**
30
30
25
25
20
20
15
15
Unit 1
Unit 1Unit 2
1010 Unit 2Unit 3Unit 3
5
5
0
0 0 10 20 30 40 50 60
0 10 20 30 40 50 60
Time (sec)
Time (sec)
|Col1|Col2|Col3|Col4|Col5|
|---|---|---|---|---|
||||||
||||||
||||||
||||||
|Unit 1 UUnint i1t 2 UUnint i2t 3 UUnint i3t 4 Unit 4|||||
||||||
||||||
||||||
**Figure 12. Generator output power in kW with a 0.5 noise variance and without the VSG.**
**Figure 12. Figure 12.Generator output power in kW with a 0.5 noise variance and without the VSG. Generator output power in kW with a 0.5 noise variance and without the VSG.**
3030
2525
2020
1515
1010
55
00
00 1010 2020 3030 4040 5050 6060
Time (sec)Time (sec)
|UUnint i1t 1 UUnint i2t 2 UUnint i3t 3 UUnint i4t 4|UUnint i1t 1 UUnint i2t 2 UUnint i3t 3 UUnint i4t 4|Col3|
|---|---|---|
||||
||||
||||
||||
||||
||||
||||
**Figure 13. Figure 13. Figure 13.Generator output power in kW with a 0.8 noise variance and without the VSG. Generator output power in kW with a 0.8 noise variance and without the VSG. Generator output power in kW with a 0.8 noise variance and without the VSG.**
The system required lesser time to each the intended value of optimal power when
was connected with the VSG, as shown in Figure 14. For low, medium, and high noise
variance for all four units, it can be seen in the graph that there was a 5–10 s improve
_Energies 2023, 16, x FOR PEER REVIEW_ 13 of 19
ment for each unit in noise variance when the Lagrange method was used with the VSG
control strategy.
30
25
20
15
10
5
0
0 10 20 30 40 Unit 3 0.8 noise50 60
Time (sec) Unit 4 0.8 noise
**Figure 14. Figure 14. Comparison of the generator units’ output power in kW with all noise levels with theComparison of the generator units’ output power in kW with all noise levels with the**
Lagrange method and the VSG. Lagrange method and the VSG.
|Unit 1 0.2 noise Unit 2 0.2 noise Unit 3 0.2 noise Unit 4 0.2 noise Unit 1 0.5 noise Unit 2 0.5 noise Unit 3 0.5 noise Unit 4 0.5 noise Unit 1 0.8 noise Unit 2 0.8 noise|Col2|Col3|
|---|---|---|
||Unit 1 0.2 noise Unit 2 0.2 noise Unit 3 0.2 noise Unit 4 0.2 noise Unit 1 0.5 noise Unit 2 0.5 noise Unit 3 0.5 noise Unit 4 0.5 noise Unit 1 0.8 noise Unit 2 0.8 noise||
|0 10 20 30 40 Time (sec)|Un5it0 3 0.8 noise Unit 4 0.8 noise|6|
-----
#### Figure 15 compares the output power of all generating units for a 0.8 noise level
_Energies 2023, 16, 4670_ 13 of 19
#### using the Lagrange method and the PSO algorithm with the VSG control strategy. The resulting graph shows that the PSO algorithm performed better than the Lagrange method. With the PSO algorithm, convergence occurred faster than with the Lagrange Figure 15 compares the output power of all generating units for a 0.8 noise level method. For Unit 1, the PSO achieved stability 3 s earlier at the 27 s mark. For unit 2, the using the Lagrange method and the PSO algorithm with the VSG control strategy. The
resulting graph shows that the PSO algorithm performed better than the Lagrange method.
#### PSO algorithm performed better by 10 s, as indicated by the red legend. For Units 3 and 4, With the PSO algorithm, convergence occurred faster than with the Lagrange method. the PSO algorithm performed slightly better than the Lagrange method. The processing For Unit 1, the PSO achieved stability 3 s earlier at the 27 s mark. For unit 2, the PSO
algorithm performed better by 10 s, as indicated by the red legend. For Units 3 and 4,
#### time for the PSO algorithm was 15.648 s, whereas the Lagrange method required 22.343 s.
the PSO algorithm performed slightly better than the Lagrange method. The processing
#### Overall, it can be concluded that the PSO algorithm solved the economic dispatch prob-time for the PSO algorithm was 15.648 s, whereas the Lagrange method required 22.343 s. lem much quicker and efficiently than the Lagrange method. Overall, it can be concluded that the PSO algorithm solved the economic dispatch problem
much quicker and efficiently than the Lagrange method.
30
25
20
15
10
5
###### gies 2023, 16, x FOR PEER REVIEW
0
0 10 20 30 40 50 60
Time (sec)
|Col1|Col2|Col3|
|---|---|---|
|Unit 1 - Lag Unit 2 - Lag Unit 3 - Lag Unit 4 - Lag Unit 1 - PSO Unit 2 - PSO R PEER REVIEW Unit 3 - PSO|||
||Unit 1 - Lag Unit 2 - Lag Unit 3 - Lag Unit 4 - Lag Unit 1 - PSO Unit 2 - PSO Unit 3 - PSO||
||||
||||
||Unit 4 - PSO||
##### Figure 15. Comparison of the generator units’ output power in kW with a 0.8 noise levels using the Figure 16. In Figure 17, we can see that it took about 50 s to reach the opFigure 15. Comparison of the generator units’ output power in kW with a 0.8 noise levels using the Lagrange and the PSO algorithm with the VSG. Lagrange and the PSO algorithm with the VSG.
### cost for a 0.2 noise level. In Figure 18, it took 55 s to reach the optimal v level, whereas it took more than 60 s to reach the optimal value for aThe consensus-based approach aided in setting the incremental cost of each generator
#### The consensus-based approach aided in setting the incremental cost of each gener-unit more quickly when there was noise. However, it is observed in Figures 9–15 that the
### shown in Figure 19. consensus-based approach required more time as the noise variance increased. The graphs
#### ator unit more quickly when there was noise. However, it is observed in Figures 9–15 that
in Figures 16–19 show that the average incremental cost ($/kWh) was approximately 5.91.
#### the consensus-based approach required more time as the noise variance increased. The graphs in Figures 16–19 show that the average incremental cost ($/kWh) was approxi- mately 5.91. In Figures 16–19, the incremental costs for all generator units were compared under various noise situations using the Lagrange method. As can be observed from the graph, it was difficult to stabilize the microgrid when it was connected to the grid, due to a larger noise variance (pink legend). For low to medium noise levels, it functioned well. Under no noise condition, the optimal incremental cost was reached in 45 s, as shown in
Figure 16. Figure 16. Incremental Cost (IC) of the generating units compared in the absence of noise with theIncremental Cost (IC) of the generating units compared in the absen
VSG using the Lagrange method.
#### VSG using the Lagrange method.
-----
**Figure 16. Incremental Cost (IC) of the generating units compared in the absence of noise with the**
_Energies 2023, 16, 4670_ **Figure 16. Incremental Cost (IC) of the generating units compared in the absence of noise with the 14 of 19**
VSG using the Lagrange method.
VSG using the Lagrange method.
**Figure 17.** Incremental Cost (IC) of the generating units compared in the presence of a 0.2 noise
**Figure 17. Figure 17. Incremental Cost (IC) of the generating units compared in the presence of a 0.2 noiseIncremental Cost (IC) of the generating units compared in the presence of a 0.2 noise**
variance with the VSG, using the Lagrange method.
variance with the VSG, using the Lagrange method.
variance with the VSG, using the Lagrange method.
_Energies 2023, 16, x FOR PEER REVIEW_ 15 of 19
**Figure 18. Figure 18. Figure 18. Incremental Cost (IC) of the generating units compared in the presence of a 0.5 noiseIncremental Cost (IC) of the generating units compared in the presence of a 0.5 noise Incremental Cost (IC) of the generating units compared in the presence of a 0.5 noise**
variance with the VSG, using the Lagrange method.
variance with the VSG, using the Lagrange method. variance with the VSG, using the Lagrange method.
**Figure 19. Figure 19. Incremental Cost (IC) of the generating units compared in the presence of a 0.8 noiseIncremental Cost (IC) of the generating units compared in the presence of a 0.8 noise**
variance with the VSG, using the Lagrange method. variance with the VSG, using the Lagrange method.
In Figures 20 and 21, it can be seen that with the VSG strategy, the generator units in In Figures 16–19, the incremental costs for all generator units were compared under
the presence of higher noise levels stabilized more quickly. On average, 0.45 s were re-various noise situations using the Lagrange method. As can be observed from the graph, it
quired for the system to stabilize with a high noise level of 0.8. In the absence of VSG, it was difficult to stabilize the microgrid when it was connected to the grid, due to a larger
took more than 0.9 s for the frequency to stabilize, as shown in Figure 20. noise variance (pink legend). For low to medium noise levels, it functioned well. Under no
noise condition, the optimal incremental cost was reached in 45 s, as shown in Figure 16.
In Figure60.6 17, we can see that it took about 50 s to reach the optimal incremental cost for a
Unit 1
0.2 noise level. In Figure60.4 18, it took 55 s to reach the optimal value for a 0.5 noise level,Unit 2Unit 3
Unit 4
-----
### p g q y g
_Energies 2023, 16, 4670_ quired for the system to stabilize with a high noise level of 0.8. In the a15 of 19
### took more than 0.9 s for the frequency to stabilize, as shown in Figure 2Figure 19. Incremental Cost (IC) of the generating units compared in the presence of a 0.8 noise
variance with the VSG, using the Lagrange method.
whereas it took more than 60 s to reach the optimal value for a 0.8 noise level, as shown
in Figure60.6 19.
###### the presence of higher noise levels stabilized more quickly. On average, 0.45 s were re-in the presence of higher noise levels stabilized more quickly. On average, 0.45 s were60.4 In FiguresIn Figures 20 and 21, it can be seen that with the VSG strategy, the generator units in 20 and 21, it can be seen that with the VSG strategy, the generator unitsUnit 1Unit 2Unit 3 quired for the system to stabilize with a high noise level of 0.8. In the absence of VSG, itUnit 4
required for the system to stabilize with a high noise level of 0.8. In the absence of VSG, it
###### took more than 0.9 s for the frequency to stabilize, as shown in Figure 20. took more than 0.9 s for the frequency to stabilize, as shown in Figure60.2 20.
60.660
|gure 19.|Col2|Col3|Col4|Col5|Col6|
|---|---|---|---|---|---|
60.4
59.8
60.2
59.6
60
59.459.8
59.259.6
59.4
59
59.2
58.8
59
58.658.8
0 0.1 0.2 0.3 0.4 0.5 0.6
Time (sec)
58.6
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
Time (sec)
#### Figure 20. Comparison of the generator units’ frequency changes with a 0.8 noi
**Figure 20. Figure 20. Comparison of the generator units’ frequency changes with a 0.8 noise level without theComparison of the generator units’ frequency changes with a 0.8 noise level without the**
#### VSG, using the Lagrange method.
VSG, using the Lagrange method. VSG, using the Lagrange method.
60.160.1
60
60
59.9
59.9
59.8
59.8
59.7
59.759.6
59.5
59.6
|Unit 1 Unit 2 Unit 3 Unit 4|Col2|Col3|Col4|
|---|---|---|---|
|||||
|||||
|||||
||Unit 1 Unit 2 Unit 3 Unit 4|||
||||Unit 1 Unit 2 Unit 3 Unit 4|
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
Time (sec)
**Figure 21. Comparison of the generator units in the presence of a 0.8 noise level with the VSG in**
59.5
terms of frequency change, using the Lagrange method. 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
Time (sec)
**Figure 21. Comparison of the generator units in the presence of a 0.8 noise level with the VSG in**
#### Figure 21. Comparison of the generator units in the presence of a 0.8 noise le
terms of frequency change, using the Lagrange method.
#### terms of frequency change, using the Lagrange method.
In Figures 22 and 23 it is observed that with a load change, the system was more stable
and reached its maximum limit faster with the VSG strategy. The system oscillated more
and had more THD without the VSG, as observed in Figure 22. The microgrid stabilized in
0.35 s with a noise variance of 0.8 when both the Lagrange method and the VSG were in
operation. This can be seen in Figure 23.
-----
more and had more THD without the VSG, as observed in Figure 22. The microgrid sta-more and had more THD without the VSG, as observed in Figure 22. The microgrid sta
_Energies 2023, 16, 4670_ bilized in 0.35 s with a noise variance of 0.8 when both the Lagrange method and the VSG bilized in 0.35 s with a noise variance of 0.8 when both the Lagrange method and the VSG 16 of 19
were in operation. This can be seen in Figure 23. were in operation. This can be seen in Figure 23.
4040
3535
3030
2525
2020
1515
1010
55
00
00 0.10.1 0.20.2 0.30.3 0.40.4 0.50.5 0.60.6 0.70.7
Time (sec)Time (sec)
**Figure 22. Figure 22.Figure 22. Comparison of the generating units’ maximum power with a 0.8 noise variance and a loadComparison of the generating units’ maximum power with a 0.8 noise variance and a Comparison of the generating units’ maximum power with a 0.8 noise variance and a**
load change without the VSG, using the Lagrange method. change without the VSG, using the Lagrange method.load change without the VSG, using the Lagrange method.
4040
3535
3030
2525
2020
1515
1010
55
00
00 0.10.1 0.20.2 0.30.3 0.40.4 0.50.5 0.60.6 0.70.7
Time (sec)Time (sec)
**Figure 23. Figure 23.Figure 23. Comparison of the generating units’ maximum power with a 0.8 noise variance and a loadComparison of the generating units’ maximum power with a 0.8 noise variance and a Comparison of the generating units’ maximum power with a 0.8 noise variance and a**
load change with the VSG, using the Lagrange method. change with the VSG, using the Lagrange method.load change with the VSG, using the Lagrange method.
Table 3 compares the Lagrange method and the PSO algorithm performances when TableTable 3 compares the Lagrange method and the PSO algorithm performances when 3 compares the Lagrange method and the PSO algorithm performances when
used with the VSG control strategy in relation to the incremental cost. For all noise vari-used with the VSG control strategy in relation to the incremental cost. For all noise variances,used with the VSG control strategy in relation to the incremental cost. For all noise variances, the PSO algorithm performed better and provided stability more quickly. In the the PSO algorithm performed better and provided stability more quickly. In the absence ofances, the PSO algorithm performed better and provided stability more quickly. In the
absence of noise, the PSO algorithm required 27.45 s to reach the optimal incremental noise, the PSO algorithm required 27.45 s to reach the optimal incremental cost, whereasabsence of noise, the PSO algorithm required 27.45 s to reach the optimal incremental
cost, whereas the Lagrange method required about 38.21 s. With a 0.2 noise variance, the the Lagrange method required about 38.21 s. With a 0.2 noise variance, the PSO algorithmcost, whereas the Lagrange method required about 38.21 s. With a 0.2 noise variance, the
PSO algorithm required 10 s less than the Lagrange method to reach the optimal incre-required 10 s less than the Lagrange method to reach the optimal incremental cost. For aPSO algorithm required 10 s less than the Lagrange method to reach the optimal incremental cost. For a medium noise variance of 0.5, the PSO algorithm was faster by about 13 medium noise variance of 0.5, the PSO algorithm was faster by about 13 s and for a highmental cost. For a medium noise variance of 0.5, the PSO algorithm was faster by about 13
s and for a high noise variance of 0.8, it was faster by about 39 s. It is observed in Table 3 noise variance of 0.8, it was faster by about 39 s. It is observed in Tables and for a high noise variance of 0.8, it was faster by about 39 s. It is observed in Table 3 3 that the PSO
that the PSO algorithm performed better for all levels of noise variance and required algorithm performed better for all levels of noise variance and required much less time tothat the PSO algorithm performed better for all levels of noise variance and required
much less time to stabilize the system than the Lagrange method. stabilize the system than the Lagrange method.much less time to stabilize the system than the Lagrange method.
Table 4 compares the Lagrange method and the PSO algorithm performances when
used with the VSG control strategy in relation to frequency and maximum power. For all
noise variances, the PSO algorithm performed better and provided stability more quickly.
For the 0.8 noise condition, the PSO algorithm required 0.2 s to reach frequency stability,
whereas the Lagrange method required about 0.45 s. Similarly, for the maximum power,
the PSO algorithm required half the time to reach stability, as seen in Table 4.
-----
_Energies 2023, 16, 4670_ 17 of 19
**Table 3. Comparison of the time to reach the average optimal incremental cost by the two examined methods.**
**Noise Variance** **Lagrange Method** **PSO Algorithm**
No noise 38.21 s 27.45 s
0.2 variance 48 s 38.20 s
0.5 variance 52.57 s 40.19 s
0.8 variance 90 s 51.85 s
**Table 4. Comparison of the time to reach the optimal levels of the shown parameters for a 0.8 noise variance.**
**Method/Algorithm** **Frequency (Hz)** **Max. Power (kW)**
Lagrange 0.45 s 0.30 s
PSO 0.20 s 0.15 s
**7. Conclusions**
For islanded microgrids, the suggested consensus-based approach for economic dispatch performs well [23]. This algorithm was utilized in this study to examine how the
microgrid operated in grid-connected mode.
The VSG strategy was also introduced to enhance the system’s stability. The microgrid’s performances of the Lagrange method and the PSO algorithms were compared, with
and without the use of the VSG strategy. It is concluded that with the inclusion of the VSG
control strategy, the system could reach stabilization much faster in the presence of different
levels of noise and load changes, as described in the Results section. This was observed
for both the Lagrange method and the PSO algorithm. The consensus-based economic
dispatch algorithm worked efficiently in conjunction with the VSG control strategy. It can
also be concluded from the results obtained that the PSO algorithm performed better in
stabilizing the frequency, output power, and load changes in the microgrid. The optimal
incremental cost was also achieved faster with the PSO algorithm.
The results clearly showed that a consensus-based economic dispatch solution with
the VSG strategy yielded a better stabilization in microgrids in the presence of low, medium,
and high noise variances. Future research should be carried out to assess the performance
of different algorithms on the noise effect in both grid-connected and islanded microgrids.
Reactive power compensation can also be included in future studies for a better overall
performance of microgrids.
**Author Contributions: Conceptualization, S.S.; Methodology, D.W.G.; Software, S.S.; Validation,**
S.S. and D.W.G.; Investigation, S.S. and D.W.G.; Resources, D.W.G.; Writing—original draft, S.S.;
Writing—review & editing, D.W.G. All authors have read and agreed to the published version of
the manuscript.
**Funding: This research received no external funding.**
**Data Availability Statement: Data is unavailable due to privacy or ethical restrictions.**
**Acknowledgments: The noiseless consensus-based economic dispatch algorithm was created in [23],**
and its performance for islanded microgrids was examined. To the best of the authors’ knowledge,
no other researchers have examined the impact of this method on grid-connected microgrids.
**Conflicts of Interest: The authors declare no conflict of interest.**
**References**
1. Liu, D.; Cai, Y. Taguchi method for solving the economic dispatch problem with nonsmooth cost functions. IEEE Trans. Power
_[Syst. 2005, 20, 2006–2014. [CrossRef]](https://doi.org/10.1109/TPWRS.2005.857939)_
2. Park, J.B.; Jeong, Y.W.; Shin, J.R.; Lee, K. An improved particle swarm optimization for nonconvex economic dispatch problems.
_[IEEE Trans. Power Syst. 2010, 25, 156–166. [CrossRef]](https://doi.org/10.1109/TPWRS.2009.2030293)_
3. Guo, T.; Henwood, M.; Van Ooijen, M. An algorithm for combined heat and power economic dispatch. IEEE Trans. Power Syst.
**[1996, 11, 1778–1784. [CrossRef]](https://doi.org/10.1109/59.544642)**
-----
_Energies 2023, 16, 4670_ 18 of 19
4. Fan, J.Y.; Zhang, L. Real-time economic dispatch with line flow and emission constraints using quadratic programming. IEEE
_[Trans. Power Syst. 1998, 13, 320–325. [CrossRef]](https://doi.org/10.1109/59.667345)_
5. Olfati-Saber, R.; Murray, R.M. Consensus problems in networks of agents with switching topology and time-dalys. IEEE Trans.
_[Autom. Control 2004, 49, 1520–1533. [CrossRef]](https://doi.org/10.1109/TAC.2004.834113)_
6. Ren, W.; Beard, R.W.; Atkins, E.M. Information consensus in multivehicle cooperative control. IEEE Control Syst. Mag. 2007,
_27, 71–82._
7. Jadbabaie, A.; Lin, J.; Morse, A.S. Coordination of groups of mobile autonomous agents using nearest neighbor rules. IEEE Trans.
_Autom. Control 2005, 50, 169–182._
8. Moreau, L. Stability of multiagent systems with time-dependent communication links. IEEE Trans. Autom. Control 2005,
_[50, 169–182. [CrossRef]](https://doi.org/10.1109/TAC.2004.841888)_
9. Ma, Y.; Zhang, W.; Liu, W.; Yang, Q. Fully distributed social welfare optimization with line flow constraint consideration. IEEE
_[Trans. Ind. Informat. 2015, 11, 1532–1540. [CrossRef]](https://doi.org/10.1109/TII.2015.2475703)_
10. Rahbari-Asr, N.; Ojha, U.; Zhang, Z.; Chow, M.-Y. Incremental welfare consensus algorithm for cooperative distributed genera[tion/demand response in smart grid. IEEE Trans. Smart Grid 2014, 5, 2836–2845. [CrossRef]](https://doi.org/10.1109/TSG.2014.2346511)
11. Xu, Y.; Li, Z. Distributed optimal resource management based on the consensus algorithm in a microgrid. IEEE Trans. Ind. Electron
**[2015, 62, 2584–2592. [CrossRef]](https://doi.org/10.1109/TIE.2014.2356171)**
12. Xu, Y.; Yang, Z.; Gu, W.; Li, M.; Deng, Z. Robust real-time distributed optimal control based energy management in a smart grid.
_[IEEE Trans. Smart Grid 2017, 8, 1568–1579. [CrossRef]](https://doi.org/10.1109/TSG.2015.2491923)_
13. Zheng, W.; Wu, W.; Zhang, B.; Lin, C. Distributed optimal residential demand response considering operational constraints of
[unbalanced distribution networks. IET Gener. Transm. Distrib. 2018, 12, 1970–1979. [CrossRef]](https://doi.org/10.1049/iet-gtd.2017.1366)
14. Guo, F.; Wen, C.; Li, Z. Distributed optimal energy scheduling based on a novel PD pricing strategy in smart grid. IET Gener.
_[Transm. Distrib. 2017, 11, 2075–2084. [CrossRef]](https://doi.org/10.1049/iet-gtd.2016.1722)_
15. Rahbari-Asr, N.; Zhang, Y.; Chow, M.-Y. Consensus-based distributed scheduling for cooperative operation of distributed energy
[resources and storage devices in smart grids. IET Gener. Transm. Distrib. 2016, 10, 1268–1277. [CrossRef]](https://doi.org/10.1049/iet-gtd.2015.0159)
16. Jordehi, A.R.; Javadi, M.S.; Catalao, J.P.S. Dynamic Economic Load Dispatch in Isolated Microgrids with Particle Swarm
Optimisation considering Demand Response. In Proceedings of the 55th International Universities Power Engineering Conference
(UPEC), Turin, Italy, 1–4 September 2020; pp. 1–4.
17. Imtiaz, B.; Cui, Y.; Zafar, I. Economic Dispatch of Microgrid Incorporating Demand Response Using Dragonfly Algorithm.
In Proceedings of the 2021 IEEE International Conference on Advances in Electrical Engineering and Computer Applications
(AEECA), Dalian, China, 27–28 August 2021; pp. 59–68.
18. Singh, J.; Poddar, S.; Ramalingam, S.P.; Shanmugam, P.K.; Kalam, A. Investigation on Dynamic Economic Dispatch Problem
of Microgrid Using Cuckoo Search Algorithm—Grid Connected and Island Mode. In Proceedings of the 2019 IEEE Region 10
Conference (TENCON), Kochi, India, 17–20 October 2019; pp. 1886–1891.
19. Luo, N.; Liu, J.; Zhang, P. Optimal Dispatching of Active Distribution Network based on Improved Genetic Algorithm. In
Proceedings of the 2022 44th International Conference on Frontiers Technology of Information and Computer (ICFTIC), Qingdao,
China, 2–4 December 2022; pp. 551–554.
20. Abhinav, S.; Schizas, I.D.; Lewis, F.L.; Davoudi, A. Distributed noise-resilient networked synchrony of active distribution systems.
_[IEEE Trans. Smart Grid 2018, 9, 836–846. [CrossRef]](https://doi.org/10.1109/TSG.2016.2569602)_
21. Abhinav, S.; Schizas, I.D.; Ferrese, F.; Davoudi, A. Optimization based Ac microgrid synchronization. IEEE Trans. Ind. Informat.
**[2017, 13, 2339–2349. [CrossRef]](https://doi.org/10.1109/TII.2017.2702623)**
22. Dehkordi, N.M.; Baghaee, H.R.; Sadati, N.; Guerrero, J.M. Distributed noise-resilient secondary voltage and frequency control for
[islanded microgrids. IEEE Trans. Smart Grid 2018, 10, 3780–3790. [CrossRef]](https://doi.org/10.1109/TSG.2018.2834951)
23. Chen, F.; Chen, M.; Zhao, X.; Guerrero, J.M.; Wang, L.Y. Distributed Noise-resilient economic dispatch strategy for islanded
[microgrids. IET Gener. Transm. Distrib. 2019, 13, 3029–3039. [CrossRef]](https://doi.org/10.1049/iet-gtd.2018.5740)
24. Yazdanian, M.; Mehrizi-Sani, A. Distributed control techniques in microgrids. IEEE Trans. Smart Grid 2014, 5, 2901–2909.
[[CrossRef]](https://doi.org/10.1109/TSG.2014.2337838)
25. Molzahn, D.K.; Dorfler, F.; Sandberg, H.; Low, S.H.; Chakrabarti, S.; Baldick, R.; Lavaei, J. A survey of distributed optimization
[and control algorithms for electric power systems. IEEE Trans. Smart Grid 2017, 8, 2941–2962. [CrossRef]](https://doi.org/10.1109/TSG.2017.2720471)
26. Han, Y.; Zhang, K.; Hong, L.; Coelho, E.A.A.; Guerrero, J.M. MAS-based distributed coordinated control and optimization in
[kicrogrid and microgrid clusters: A comprehensive review. IEEE Trans. Power Electron 2018, 33, 6488–6508. [CrossRef]](https://doi.org/10.1109/TPEL.2017.2761438)
27. Xu, T.; Wu, W.; Sun, H.; Wang, L. Fully distributed multi-area dynamic economicdispatch method with second-order convergence
[for active distribution networks. IET Gener. Transm. Distrib. 2017, 11, 3955–3965. [CrossRef]](https://doi.org/10.1049/iet-gtd.2016.1945)
28. Kouveliotis-Lysikatos, I.; Hatziargyriou, N. Fully distributed economic dispatch of distributed generators in active distribution
[networks considering losses. IET Gener. Transm. Distrib. 2017, 11, 627–636. [CrossRef]](https://doi.org/10.1049/iet-gtd.2016.0616)
29. Zheng, W.; Wu, W.; Zhang, B.; Li, Z.; Liu, Y. Fully distributed multi-area economic dispatch method for active distribution
[networks. IET Gener. Transm. Distrib. 2015, 9, 1341–1351. [CrossRef]](https://doi.org/10.1049/iet-gtd.2014.0904)
30. Tu, Y.; Su, J.H.; Du, Y.; Yang, X.Z.; Xu, H.D. Analysis of microgrid inverter paralleling system based on virtual oscillator. Electr.
_Power Autom. Equip. 2017, 37, 24–30._
-----
_Energies 2023, 16, 4670_ 19 of 19
31. Cheng, Q.M.; Gao, J.; Cheng, Y.M. An inverter control method suitable for islanding operation. Power Syst. Technol. 2018,
_42, 203–209._
32. Xu, Y.Q.; Ma, H.J. Inverter parallel operation technology based on improved droop control. Power Syst. Prot. Control. 2015,
_43, 103–107._
33. Lü, Z.Y.; Wu, Z.J.; Dou, X.B. Adaptive discrete droop control of isolated DC microgrid based on discrete consistency. Proc. CSEE
**2015, 35, 4397–4407.**
34. Chen, W.; Li, T. Distributed Economic Dispatch for Energy Internet based on Miultiagent Consensus Control. IEEE Trans. Autom.
_[Control 2021, 66, 137–152. [CrossRef]](https://doi.org/10.1109/TAC.2020.2979749)_
35. Meng, J.; Shi, X.; Wang, Y.; Fu, C. A virtual synchronous generator control strategy for distributed generation. In Proceedings
of the 2014 China International Conference on Electricity Distribution (CICED), Shenzhen, China, 23–26 September 2014;
pp. 495–498.
36. Singh, S.; Gao, D.W. Noiseless Consensus based Algorithm for Economic Dispatch problme in Grid-connected Microgrids to
enahnce Stability among Distributed Generators. In Proceedings of the 2019 North American Power Symposium (NAPS), Wichita,
KS, USA, 13–15 October 2019; pp. 1–5.
37. Li, H.; Gu, R.-Z. Research on Grid-coonected Control and Simulation of Microgrid Inverter based on VSG. In Proceedings of the
2018 China International Conference on Electricity Distribution (CICED), Tianjin, China, 17–19 September 2018.
38. Miškovi´c, M.; Miroševi´c, M.; Milkovi´c, M. Analysis of synchronous generator angular stability depending on the choice of the
[excitation system. Angew. Chem. 2009, 39, 4555. [CrossRef]](https://doi.org/10.37798/2009584308)
39. Wang, K.; Qi, C.; Huang, X.; Li, G. Large disturbance stability evaluation of interconnected multi-inverter power grids with VSG
[model. J. Eng. 2017, 2017, 2483–2488. [CrossRef]](https://doi.org/10.1049/joe.2017.0775)
**Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual**
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.3390/en16124670?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.3390/en16124670, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "CLOSED",
"url": "https://www.mdpi.com/1996-1073/16/12/4670/pdf?version=1686621916"
}
| 2,023
|
[
"JournalArticle"
] | true
| 2023-06-12T00:00:00
|
[
{
"paperId": "efad9615a2ae16f704f4f14e0522e5f6b0541c48",
"title": "Distributed Noise-Resilient Secondary Voltage and Frequency Control for Islanded Microgrids"
},
{
"paperId": "829bf44d870635e6708918fd193fead00215afef",
"title": "Distributed noise‐resilient economic dispatch strategy for islanded microgrids"
},
{
"paperId": "8c9ac526021d13fda27220c98e2894a27ba446a6",
"title": "Distributed Economic Dispatch for Energy Internet Based on Multiagent Consensus Control"
},
{
"paperId": "05651c5dd4ead4996aab47c1e54341ff23fafdc7",
"title": "MAS-Based Distributed Coordinated Control and Optimization in Microgrid and Microgrid Clusters: A Comprehensive Overview"
},
{
"paperId": "3282ffbca9a87b4a313992ad9fa99a7c3895d1b4",
"title": "Distributed Noise-Resilient Networked Synchrony of Active Distribution Systems"
},
{
"paperId": "4d935d24433b237ee3ba44eea34ecc45522f7d5b",
"title": "Distributed optimal residential demand response considering operational constraints of unbalanced distribution networks"
},
{
"paperId": "baf0d2448403e98c6bcd89ca9db5cb483eb1a9c7",
"title": "Large Distribution Stability Evaluation of Interconnected Multi-Inverter Power Grids with Virtual Synchronous Generator Model"
},
{
"paperId": "6685d35225152d56a6d234e6f4e7159da89a3709",
"title": "A Survey of Distributed Optimization and Control Algorithms for Electric Power Systems"
},
{
"paperId": "e6e567ab99fe573c72a8494055db287c88caf61c",
"title": "Robust Real-Time Distributed Optimal Control Based Energy Management in a Smart Grid"
},
{
"paperId": "8862e5e5788f0e04b4655888ac3bb0d33bde3591",
"title": "Fully distributed multi-area dynamic economic dispatch method with second-order convergence for active distribution networks"
},
{
"paperId": "87c0587853c4f8ed61557b49cf7569e579869f26",
"title": "Optimization-Based AC Microgrid Synchronization"
},
{
"paperId": "98c73bb37f51ce3f972e11c1d35e010807c41d90",
"title": "Distributed optimal energy scheduling based on a novel PD pricing strategy in smart grid"
},
{
"paperId": "313a19291882db4fc3a2bee4d851bc28363cbb15",
"title": "Fully distributed economic dispatch of distributed generators in active distribution networks considering losses"
},
{
"paperId": "bf715411d4145dced796edd2962b33a8f78ad56f",
"title": "Consensus-based distributed scheduling for cooperative operation of distributed energy resources and storage devices in smart grids"
},
{
"paperId": "fbf3a61ce06e5e8669c213025a8487c637cb810f",
"title": "Fully Distributed Social Welfare Optimization With Line Flow Constraint Consideration"
},
{
"paperId": "ff236c180d66d616cc4b9f7d24c6df50297e52c4",
"title": "Fully distributed multi-area economic dispatch method for active distribution networks"
},
{
"paperId": "1260805dc54fb86c893f94df3298622e1f1c2c39",
"title": "Distributed Optimal Resource Management Based on the Consensus Algorithm in a Microgrid"
},
{
"paperId": "e7d8f5aa3794745dd8a68af2064cbd1bf9de0cad",
"title": "Incremental Welfare Consensus Algorithm for Cooperative Distributed Generation/Demand Response in Smart Grid"
},
{
"paperId": "1d84be92cbeab857d079fb639dee39da93bac911",
"title": "Distributed Control Techniques in Microgrids"
},
{
"paperId": "d1210000a5d35f96dd1df161ca2968aaa7c14193",
"title": "An Improved Particle Swarm Optimization for Nonconvex Economic Dispatch Problems"
},
{
"paperId": "e274acb96305dd122815565afb9e3edf747557cf",
"title": "ANALYSIS OF SYNCHRONOUS GENERATOR ANGULAR STABILITY DEPENDING ON THE CHOICE OF THE EXCITATION SYSTEM"
},
{
"paperId": "c92917c75a596fd27351cb70a5694b61f622863f",
"title": "Information consensus in multivehicle cooperative control"
},
{
"paperId": "dea24d1daa97f262e723a62a866fd40d619248f5",
"title": "Taguchi method for solving the economic dispatch problem with nonsmooth cost functions"
},
{
"paperId": "7fe0ef2ddacd193101dc5ba3df97b0241a5e8fc6",
"title": "Stability of multiagent systems with time-dependent communication links"
},
{
"paperId": "9839ed2281ba4b589bf88c7e4acc48c9fa6fb933",
"title": "Consensus problems in networks of agents with switching topology and time-delays"
},
{
"paperId": "20be6a4e06295792977d2d6e4c9eb9a8405226e9",
"title": "Coordination of groups of mobile autonomous agents using nearest neighbor rules"
},
{
"paperId": "4219380acc45645484a5b6ad10d0932a4df82d65",
"title": "Real-time economic dispatch with line flow and emission constraints using quadratic programming"
},
{
"paperId": "fdd93d5fa6cdf685901bec57aa7f9af1e046d811",
"title": "An algorithm for combined heat and power economic dispatch"
},
{
"paperId": null,
"title": "An inverter control method suitable for islanding operation"
},
{
"paperId": null,
"title": "Analysis of microgrid inverter paralleling system based on virtual oscillator"
},
{
"paperId": null,
"title": "Inverter parallel operation technology based on improved droop control"
},
{
"paperId": null,
"title": "Adaptive discrete droop control of isolated DC microgrid based on discrete consistency"
}
] | 19,814
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/02a1e357408df84ddf78c6d1270caaa9b4f1013e
|
[
"Computer Science"
] | 0.800405
|
A CIPHERTEXT-POLICY ATTRIBUTE-BASED SEARCHABLE ENCRYPTION SCHEME IN NON-INTERACTIVE MODEL
|
02a1e357408df84ddf78c6d1270caaa9b4f1013e
|
Journal of Computer Science and Cybernetics
|
[
{
"authorId": "2004098326",
"name": "Van Anh Trinh"
},
{
"authorId": "2241026",
"name": "V. Trinh"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"J Comput Sci Cybern"
],
"alternate_urls": null,
"id": "819d1840-7bd3-4c53-b5cc-5b9fb861c981",
"issn": "1813-9663",
"name": "Journal of Computer Science and Cybernetics",
"type": "journal",
"url": null
}
|
We address the problem of searching on encrypted data with expressive searching predicate and multi-writer/multi-reader, a cryptographic primitive which has many concrete application scenarios such as cloud computing, email gateway application and so on. In this paper, we propose a public-key encryption with keyword search scheme relied on the ciphertext-policy attribute-based encryption scheme. In our system, we consider the model where a user can generate trapdoors by himself/herself, we thus can remove the Trusted Trapdoor Generator which can save the resource and communication overhead. We also investigate the problem of combination of a public key encryption used to encrypt data and a public-key encryption with keyword search used to encrypt keywords, which can save the storage of the whole system
|
_Journal of Computer Science and Cybernetics, V.35, N.3 (2019), 233–249_
DOI 10.15625/1813-9663/35/3/13667
# A CIPHERTEXT-POLICY ATTRIBUTE-BASED SEARCHABLE ENCRYPTION SCHEME IN NON-INTERACTIVE MODEL
VAN ANH TRINH[1], VIET CUONG TRINH[2][,][∗]
1Thanh Hoa University of Culture, Sports and Tourism
2Hong Duc University, Thanh Hoa, Viet Nam
_[∗Trinhvietcuong@hdu.edu.vn](mailto:Trinhvietcuong@hdu.edu.vn)_
�
**Abstract. We address the problem of searching on encrypted data with expressive searching predi-**
cate and multi-writer/multi-reader, a cryptographic primitive which has many concrete application
scenarios such as cloud computing, email gateway application and so on. In this paper, we propose
a public-key encryption with keyword search scheme relied on the ciphertext-policy attribute-based
encryption scheme. In our system, we consider the model where a user can generate trapdoors by
himself/herself, we thus can remove the Trusted Trapdoor Generator which can save the resource and
communication overhead. We also investigate the problem of combination of a public key encryption
used to encrypt data and a public-key encryption with keyword search used to encrypt keywords,
which can save the storage of the whole system.
**Keywords. Attribute-based Encryption; Searchable Encryption; Searching on Encrypted Data.**
**1.** **INTRODUCTION**
Searching on encrypted data is an important task, which can be applicable to many
practical contexts such as cloud computing or email gateway application. In the context of
cloud computing, the user’s data is first encrypted and then outsourced to the cloud server.
When user would like to find some specific data, he/she needs to ask help from the cloud
server, user however doesn’t want the cloud server to know about his/her original data.
In the email gateway application, when anyone would like to securely send email to Alice,
he/she encrypts the content of email under Alice’s public key before sending. On the other
hand, Alice would like to set priority order of receiving her emails. To this aim, Alice gives
the email gateway the ability to check the priority order of incoming emails and then send
to her emails in the order she wants. However, Alice also doesn’t want the email gateway to
know about the content of such emails.
Searchable Encryption (SE) was introduced in [2, 16] to deal with such aforementioned
problems. In a nutshell, in a system which supports SE, we append encrypted keywords
with corresponding encrypted data. User then relies on his/her secret key in SE scheme
and chosen keywords to generate a trapdoor for the cloud server (or the email gateway) to
perform the search. The trapdoor is generated in such a way that the cloud server (or the
email gateway) using this trapdoor can perform successfully the search, but doesn’t get any
information about the original data in the resulted ciphertexts. On the other hand, since
keywords are encrypted, unauthorized users (called outsiders) as well as cloud server (called
c 2019 Vietnam Academy of Science & Technology
_⃝_
-----
234 VAN ANH TRINH, VIET CUONG TRINH
insider) ideally also don’t know any information about keywords in each ciphertext. We can
category SE in two types:
1. SE in the private-key setting [16], where there is only one writer (data owner who
encrypts the data as well as the corresponding keywords) and one/multi reader (user
who would like to search and then should be able to decrypt the resulted ciphertexts).
This type of SE has obviously limited applications in practice. For example, it cannot
apply to the context of sending email above since anyone should have the capacity of
encrypting the content of emails sent to Alice.
2. SE in the public-key setting [2], where there are multi-writer and one/multi reader. A
full searchable encryption system in practice includes two components: the first is a
Public Key Encryption (PKE) scheme used to encrypt data; the second is a PublicKey Encryption with keyword Search (PEKS) used to encrypt keywords. Such full
system is called a PKE-PEKS scheme. In a PKE-PEKS scheme, a full ciphertext,
including both the encrypted keywords and encrypted data, should be in the form
PKEAlicepk (data)||PEKSAlice′pk [(keywords).]
There are two cases: PKE and PEKS are independent, that means Alice’s publickey/secret-key in PKE is different to the ones in PEKS; and otherwise, where Alice’s
public-key/secret-key could be the same in both PKE and PEKS. Obviously, such full
system will become more efficient in the latter case. However, in this case we have to
consider carefully the security of the full system [10] since the adversary is now more
powerful than the one in the former case. When PKE and PEKS are independent, we
often only care about PEKS scheme and omit the PKE scheme for simplicity. In some
schemes [6, 11, 14], Alice cannot generate the trapdoor by himself/herself, he/she needs
to contact with a Trusted Trapdoor Generator (TTG), that will obviously increase the
communication overhead of the user, and moreover the Trusted Trapdoor Generator
should be always online. We call such schemes interactive schemes.
In summary, there are several following important properties one should take into
account when estimating a system which supports searching on encrypted data:
Efficiency: Performance of encryption/decryption/searching algorithm, key-size/
_•_
ciphertext-size, PKE and PEKS are independent or not, interactive or noninteractive, etc;
Expressive searching predicate: Whether or not the PEKS scheme supports con
_•_
junctive keywords or even boolean formulas of keywords for searching. Obviously,
this property is more desirable than simple equality keyword search in practice;
Trapdoor security: Cloud server with a trapdoor in hand knows nothing about
_•_
the keywords in the ciphertext and trapdoor, even when the trapdoor “matches”
the ciphertext. We note that this property is very hard to achieve in the publickey setting, to the best of our knowledge there is only one scheme [6] that can
_partially achieve this property._
Keyword security: Ideally, unauthorized users and cloud server cannot derive any
_•_
information about keywords in the ciphertext.
-----
A CIPHERTEXT-POLICY ATTRIBUTE-BASED SEARCHABLE ENCRYPTION SCHEME 235
**1.1.** **Related work**
Over past decade, substantial progress has been made on problem of searching on encrypted
data [1, 2, 3, 6, 8, 9, 11, 14, 16, 17, 18, 19], to name a few. These papers use different techniques and consider different situations for searching on encrypted data. SE in private-key
setting and supports only single-writer/single-reader was first introduced in [16]. Continue
this line of research, the authors in [3] investigated searchable encryption with conjunctive
keyword searches and boolean queries. The authors in [17, 18] went further to investigate
searchable encryption scheme in single-writer/multi-reader setting and partially trapdoor
security. Partially non-interactive which can reduce the communication overhead was also
investigated in [17].
SE in the public-key setting was first introduced by Boneh et al. [2], but their schemes only support multi-writer/single-reader and equality queries. In [1, 19], the authors
extended to support multi-writer/multi-reader but their schemes still only support equality queries. Expressive searching predicate and multi-writer/multi-reader were investigated
in [6, 8, 11, 12, 14], where authors manage to transform from a key policy/ciphertext policy
attribute-based encryption scheme to a PEKS scheme, these schemes are thus called key
policy/ciphertext policy attribute-based searchable encryption scheme. The authors in [6]
went further to consider partially trapdoor security in the sense that, they split a keyword
into two parts: The keyword name and the keyword value, where one keyword name can
have many keyword values. They then showed that in their scheme, the cloud server with
the trapdoor in hand can only know keyword names but nothing about keyword values in the
ciphertext. This interesting property is useful in some specific practical contexts. However,
the downside of this technique is that the searching time is only acceptable if the keyword
names are included in the ciphertext, this leads to the fact that anyone can also know the
keyword names in the ciphertext. On the other hand, the combination of PKE and PEKS
was investigated in [10] where [10] also investigated the non-interactive property, however
this scheme does not support expressive searching predicate.
**1.2.** **Our contribution and organization of the paper**
In this paper, we propose a PKE-PEKS scheme supporting both the expressive searching
predicate and multi-writer/multi-reader, our scheme is built from the CP-ABE scheme in [13],
we thus name our scheme CP-ABSE scheme for short. Our scheme has following properties:
Our scheme is a combination of an existing PKE scheme (which is exactly the CP-ABE
_•_
in [13]) and a new proposed PEKS scheme. In our scheme, user has only one pair of
public key/secret key for both PKE and PEKS, and moreover user can use the CP-ABE
setting to encrypt/decrypt data;
Our scheme is non-interactive: User can generate the trapdoor by himself/herself, we
_•_
thus can remove the Trusted Trapdoor Generator in our system. On the other hand,
since trapdoor is generated from user’s secret key, user is able to decrypt all resulted
ciphertexts which can save time and communication overhead of the system;
Efficiency: Since our CP-ABSE scheme is built from the CP-ABE scheme in [13], our
_•_
scheme naturally inherits the efficiency and properties of this CP-ABE scheme such as
-----
236 VAN ANH TRINH, VIET CUONG TRINH
constant-size of user’s secret key, optimized ciphertext size, multi-authority and fast
decryption. Note that the CP-ABE scheme in [13] is still one of the most efficient
CP-ABE schemes to date.
We also note that our scheme does not achieve trapdoor security. We emphasize that
_•_
this property is very hard to achieve in the public-key setting, to the best of our
knowledge there is only one scheme [6] that can partially achieve this property.
In the Section 5, we give the details comparison among our scheme and several schemes
which also support both the expressive searching predicate and multi-writer/multi-reader.
The paper includes 6 sections. The first section presents the definition and security model
of a CP-ABSE scheme. In Section 3, we present the construction of our CP-ABSE scheme and
prove that it is secure in the following section. The comparison and discussions are given in
Section 5. Finally, the conclusion is in Section 6.
**2.** **PRELIMINARIES**
In this section, we first give the system workflow and the threat model of our system,
then we present the definition and security model for our CP-ABSE scheme.
**2.1.** **Ciphertext policy attribute based searchable encryption**
**2.1.1.** **System workflow and threat model**
Our CP-ABSE scheme is a combination of a traditional CP-ABE scheme and a PEKS scheme
supporting expressive searching predicate. In our scheme, there are four entities: data owner;
user; cloud server and Private Key Generator (PKG). More precisely:
1. PKG: Play the role of PKG in traditional CP-ABE scheme, generates secret keys for
users.
2. Data owner: Encrypt data as well as corresponding keywords, upload them to a public
cloud.
3. User: Rely on his/her secret key to generate a trapdoor, send this trapdoor to the
cloud server and get back resulted ciphertexts. Finally, decrypt resulted ciphertexts to
recover data.
4. Cloud server: Receive a trapdoor from a user, perform the search based on the trapdoor
and send back resulted ciphertexts to the user.
**Threat model.** Similar to the threat model in recent schemes [6, 8, 9, 11, 12, 14], in our
system, there are two goals for which an adversary would like to achieve: getting information
about encrypted data and getting information about encrypted keywords.
-----
A CIPHERTEXT-POLICY ATTRIBUTE-BASED SEARCHABLE ENCRYPTION SCHEME 237
**2.1.2.** **System algorithms**
Formally, our CP-ABSE scheme includes seven following probabilistic algorithms.
**Setup(1[ν],** ): The inputs of this algorithm are security parameter ν and the description
_B_
of attribute universe, the outputs are master key MSK and the public parameters
_B_
param of the system.
**Extract(u,** (u), MSK, param): The inputs of this algorithm are attribute set (u) of user
_B_ _B_
_u, param and MSK, the output is the user’s secret key du._
**Encrypt(M, A, param): The inputs of this algorithm are param, a message M and an**
access policy A over the universe of attributes, the output is ciphertext ct along with
a description of the access policy A.
**Decrypt(ct, du, param): The inputs of this algorithm are param, the ciphertext ct and the**
secret key du of user u, the output is the message M if and only if B(u) satisfies A.
Otherwise, the output is .
_⊥_
**Trapdoor(du, Wi, param): The inputs of this algorithm are param, secret key du of user u**
and a set of keywords user would like to search Wi, the output is the trapdoor tds.
**EncryptKW(KF, A[′], param): The inputs of this algorithm are param, an access policy A[′]**
over the universe of attributes and an access policy over the universe of keywords,
_KF_
the output is the ciphertext ct[′] along with a description of the access policy A[′].
**Search(tds, ct[′], param): The inputs of this algorithm are param, a trapdoor tds and a**
ciphertext ct[′], the output is 1 if the keyword set Wi embedded in tds matches the
access structure KF in ct[′] and B(u) satisfies A[′]. Otherwise, the output is 0.
We note that the full ciphertext should be the couple (ct, ct[′]). In addition, in order for
user to be able to decrypt the resulted ciphertext, we choose A[′] in ct[′] such that if B(u)
satisfies A[′] then B(u) satisfies A in ct.
**2.1.3.** **Security model**
**Selective semantic security.** The selective semantic security game is similar to the one
in [13], except that the adversary can ask additional corruption trapdoor query. Due to the
space limitations we refer the reader to [13] for details.
**Insider security. Assume that** is the attacker, is the challenger. The insider security
_A_ _C_
game is defined as follows.
**Setup(1[ν],** ). At the beginning of the game, the adversary provides a target access
_B_ _A_
policy A[∗] over universe of attributes, and two equal target access policy KF 0[∗][,][ KF] _[∗]1_
over universe of keywords for which she intends to attack, where “equal access policy”
means that if KF 0[∗] [and][ KF] 1[∗] [are described in the DNF form, they have the same]
number of clauses. runs the Setup(1[ν], ) algorithm to obtain param and MSK. She
_C_ _B_
then gives param to .
_A_
-----
238 VAN ANH TRINH, VIET CUONG TRINH
**Query phase 1. A chooses a set of attributes B(u) as well as a set of keywords Wi and**
asks corruption trapdoor query corresponding to these sets. The challenger computes
and returns corresponding tds to the adversary.
**Challenge. C chooses b** _←{[$]_ 0, 1} and runs EncryptKW(A[∗], KF _b[∗][,][ param][) to generate][ ct][′∗][.]_
Finally, outputs ct[′∗].
_C_
**Query phase 2. The same as phase 1.**
**Guess.** finally outputs b[′] 0, 1 as its guess for b.
_A_ _∈{_ _}_
_A wins the game if b[′]_ = b, and if A never asks on B(u), Wi such that both B(u) satisfies A[∗]
and Wi satisfies either KF 0[∗] [or][ KF] 1[∗][. The advantage of][ A][ to win the game is defined]
**Adv[IS]A** [= Pr] �b = b[′][�] _−_ 2[1] _[.]_
**Definition 1. A ciphertext-policy attribute-based searchable encryption scheme achieves**
insider security if all polynomial time adversaries have at most a negligible advantage in the
above game.
**Outsider security. The outsider security game is similar to the insider game, the difference**
is that the adversary can ask corrupted secret key queries, instead of corrupted trapdoor
queries.
Due to the space limitations, we refer the reader to [13] for the definitions of Access
Structures, LSSS Matrices, Bilinear Maps and (P, Q, f ) GDDHE Assumptions and so on.
_−_
**3.** **CIPHERTEXT POLICY ATTRIBUTE BASED SEARCHABLE**
**ENCRYPTION**
In this paper, we rely on the CP-ABE scheme in [13] to build our CP-ABSE scheme.
Concretely, our CP-ABSE scheme is a combination of CP-ABE scheme in [13] and a new
PEKS scheme, where the later scheme is also built from the former scheme. User in our
scheme, therefore, can use the same public key and secret key for both CP-ABE scheme and
PEKS scheme.
In our scheme, user relies on his/her secret key and a set of chosen keywords W =
(w1, . . ., wt) to generate the trapdoor. More precisely, from a set of chosen keywords W,
user has to indicate exactly which combinations of keywords he/she would like to search.
Consider the example in [6], W = (w1, w2, w3) where w1 = “Diabetes”, w2 = “Age : 30”
and w3 = “Weight : 150 − 200”, user has to indicate the set of combinations of keywords
he/she would like to search Wi = (w1||w2, w1||w3). The advantage of this point is that we
can save the searching time and the communication overhead, since cloud server only needs
to find and then send back ciphertexts user really wants. In other schemes [5, 6, 11, 14],
user doesn’t indicate exactly which combinations of keywords he/she would like to search,
the cloud server thus searches on all possible combinations of keywords.
-----
A CIPHERTEXT-POLICY ATTRIBUTE-BASED SEARCHABLE ENCRYPTION SCHEME 239
**3.1.** **Detailed construction**
Our scheme is described as follows.
**Setup(ν,** ): Denote N = the maximal number of attributes in the system,
_B_ _|B|_
(p, G, GT, e(·, ·)) a bilinear group system. The algorithm first picks a random generator g ∈ G, random scalars a, α, λ ∈ Zp, computes g[a], g[α], g[λ]. The algorithm continues to generate 2N group elements in G associated with N attributes in the system
_h1, . . ., hN_ _,_ _h[˜]1, . . .,_ _h[˜]N_ . Let H, _H[˜] be hash functions such that H : {0, 1}[∗]_ _→_ G and
_H˜ : GT × {0, 1}[∗]_ _→_ Zp.
Suppose that the keyword universe in the system is W = (w1, w2, w3, . . . ), where each
_wi ∈{0, 1}[∗]. In our system the set W is unbounded, we can add any new keyword into_
the system at anytime. For simplicity, we omit W in the global parameters. Finally,
we set the master secret key and global parameters as MSK = (g[α], λ) and
param = (g, g[a], g[λ], e(g, g)[α], h1, . . ., hN _,_ _h[˜]1, . . .,_ _h[˜]N_ _, H,_ _H[˜])._
**Extract(u,** (u), MSK, param): Assume (u) is the attribute set of user u. The algorithm
_B_ _B_
$
chooses su _←_ Zp, then computes user u’s secret key as du = (du0, d′u0[,][ {][d][u]i[}]i∈B(u)[, λ][),]
where
_du0 = g[α]_ _· g[a][·][s][u], d[′]u0_ [=][ g][s][u][,][ {][d][u]i [=][ h]i[s][u][}][i][∈B][(][u][)][.]
User u then keeps du0 and λ secret and publishes the rest of his/her secret key to the
public domain. That means the secret key of user is of the constant-size.
**Encrypt(M, A, param): The inputs are a message M, an access policy A, as well as param.**
Assume that A is a boolean formula β and that the size of β is |β|. At first, encryptor
describes β in the form of DNF access policy as β = (β1 _βm), where each βi is_
_∨· · · ∨_
a set of attributes, i = 1, . . ., m.
$
The encryptor chooses a scalar s _←_ Zp, then computes C, C0 as
_C = M · e(g, g)[α][·][s], C0 = g[s]._
Next, encryptor compares between m and _β_, if m _β_ he/she computes
_|_ _|_ _≤|_ _|_
_C1 = (g[a][ �]_ _hi)[s], . . ., Cm = (g[a][ �]_ _hi)[s]._
_i∈β1_ _i∈βm_
Else, the encryptor constructs an LSSS matrix M representing the original boolean
formula β, and a map function ρ such that (M, ρ) ∈ (Z[ℓ]p[×][n], F([ℓ] → [N ])). She then
chooses a random vector _[−→]v = (s, y2, . . ., yn) ∈_ Z[n]p [.] For i = 1, . . ., ℓ she computes
_λi =_ _[−→]v · Mi, where Mi is the vector corresponding to the i’th row of M_ . She computes
_Ci = g[a.λ][i]h[−]ρ([s]i)[, i][ = 1][, . . ., ℓ.]_
Eventually, the output is either ct = (C, C0, . . ., Cm) along with a description of β or
_ct = (C, C0, . . ., Cℓ) along with a description of (M, ρ)._
-----
240 VAN ANH TRINH, VIET CUONG TRINH
**Decrypt(ct, du, param): The decryptor u first parses the ct and checks the number of**
elements in ct. If it is equal to m + 1, decryptor parses the ct as (C0, C1, . . . Cm), then
finds j such that βj ⊂B(u), and computes
_e(C0, du0_ �i∈βj _[d][u]i[)]_ = _e(g[s], g[α](g[a][ �]i∈βj_ _[h][i][)][s][u][)]_ = e(g, g)[α][·][s] = K.
_e(d[′]u0, Cj)_ _e(g[s][u], (g[a][ �]i∈βj_ _[h][i][)][s][)]_
Finally, computes = C _K[−][1]._
_M_ _·_
Else, she defines the set I ⊂{1, 2, . . ., ℓ} such that I = {i : ρ(i) ∈B(u)}. Let {ωi ∈
Zp}i∈I be a set of constants such that if {λi} are valid shares of any secret s according
to M then [�]i∈I _[ω][i][λ][i][ =][ s][. Note that from the relation][ �]i∈I_ _[ω][i][M][i][ = (1][,][ 0][, . . .,][ 0) where]_
_Mi is the i-th row of the matrix M_, she can determine these constants. She parses the
_ct as (C0, C1, . . . Cℓ) and computes_
� �
_e(_ _Ci[−][ω][i], d[′]u0[)][ ·][ e][(][C][0][, d][u]0_ _d[−]uρ[ω](i[i])[) =][ K.]_
_i∈I_ _i∈I_
Then computes = C _K[−][1]._
_M_ _·_
**Trapdoor(du, Wi = ( ˜wi1, . . ., ˜wik** ), param): Suppose that each ˜wij ∈{0, 1}[∗], j ∈ [k], is a
concatenation of set of keywords, for example “Diabetes _Age : 30”._
_||_
The user randomly chooses scalars r1, . . ., rk ∈ Zp, computes the trapdoor
�
tds = �{tds0,j, tds1,j, {tds2,j,ℓ}ℓ∈Bu}j∈[k], tds0, {tdsi}i∈B(u), _W[˜]_ _i_
� �
= _{g[α]g[as][u]g[ar][j]_ (g[a]H( ˜wij ))[λ], g[r][j] _, {h[˜][r]ℓ[j]_ _[}][ℓ][∈B][u][}][j][∈][[][k][]][, g][s][u][,][ {][h]i[s][u][}][i][∈B][(][u][)][,][ ˜][W][i]_ _._
where W[˜] _i is a short description of Wi. User then sends ({tds0,j}j∈[k],_ _W[˜]_ _i) to the cloud_
server, he/she publishes the rest of tds to the public domain. That means the trapdoorsize is linear in the number of combinations of keywords user would like to search.
**EncryptKW(KF, A[′], param, ): Assume that access policy A[′]** = β = (β1 ∨· · · ∨ _βm) and_
_KF = (kf1 ∨· · · ∨_ _kfm′), where each βi is a set of attributes and kfi is a concatenation_
of set of keywords. Note that βi ̸= βj, kfi′ ̸= kfj′, ∀i, j ∈ [m], i[′], j[′] _∈_ [m[′]].
$
The encryptor picks a scalar s _←_ Zp, then computes
_C0 = g[s], C1 = (g[a][ �]_ _hi)[s], . . ., Cm = (g[a][ �]_ _hi)[s],_
_i∈β1_ _i∈βm_
_C˜1 = (g[a][ �]_
_i∈β1_
_h˜i)[s], . . ., ˜Cm = (g[a][ �]_
_i∈βm_
_h˜i)[s]._
Next, he/she computes
_Xi = e(g, g)[α][·][s]_ _· e(g, g[a]H(kfi))[λ][·][s], i = 1, . . ., m[′],_
then computes
_K1 = H[˜](X1, kf1), . . ., Km′ = H[˜](Xm′, kfm′)._
-----
A CIPHERTEXT-POLICY ATTRIBUTE-BASED SEARCHABLE ENCRYPTION SCHEME 241
Eventually, encryptor outputs
_ct[′]_ = (C0, . . ., Cm, _C[˜]1, . . .,_ _C[˜]m, K1, . . ., Km′)_
along with a description of β.
**Search(tds, ct[′], param): The cloud server first finds ℓ** _∈_ [m] such that βℓ _⊂B(u), then_
computes (Xj, Yj), j = 1, . . ., k
_Xj_ = _e(C0, tds0,j_ �i∈βℓ [tds][i][ ·][ tds][2][,j,i][)] = _e(g[s], g[α]g[as][u]g[ar][j]_ _g[aλ]H( ˜wij_ )[λ][ �]i∈βℓ _[h]i[s][u][h][˜]i[r][j]_ [)]
_e(tds0, Cℓ) · e(tds1,j,_ _C[˜]ℓ)_ _e(g[s][u], (g[a][ �]i∈βℓ_ _[h][i][)][s][)][ ·][ e][(][g][r][j]_ _[,][ (][g][a][ �]i∈βℓ_ _[h][˜][i][)][s][)]_
= _e(g, g)[α][·][s]_ _· e(g, g[a]H( ˜wij_ ))[λ][·][s],
_Yj_ = _H˜(Xj, ˜wij_ ).
If there exists a pair (i, j), i ∈ [m[′]], j ∈ [k] such that Ki = Yj then the cloud server
outputs “yes”. Otherwise, the cloud server outputs “no”. Note that, the cloud server
doesn’t need to compute all pairs (Xj, Yj), j = 1, . . ., k, as long as he/she finds a pair
(i, j), i ∈ [m[′]], j ∈ [k] such that Ki = Yj, he/she outputs “yes” and stops.
**Correctness: It is easy to verify that if there exists at least one pair ˜wij** _Wi and kft_
_∈_ _∈KF_
such that ˜wij = kft, then
_Xt_ = _e(g, g)[α][·][s]_ _· e(g, g[a]H(kft))[λ][·][s]_ = e(g, g)[α][·][s] _· e(g, g[a]H( ˜wij_ ))[λ][·][s] = Xj,
that means
_Kt = H[˜](Xt, kft) = H[˜](Xj, ˜wij_ ) = Yj.
**Remark 1.** As in [13] all sets βi must be disjoint subsets to resist the simple attack,
_i = 1, . . ., m. This leads to the fact that attributes in the system cannot be reused in the_
access formula. To deal with this problem, they allow each attribute to have kmax copies of
itself as in [4, 7]. Note that the user’s secret key is still of constant-size.
**4.** **SECURITY**
In this section, we show that our scheme is secure in the model defined in Subsection 2.1.3.
We first refer the reader to the modified BDHE assumption defined in [13], and then we
define a new modified BDHE assumption. We finally prove our scheme achieves the selective
semantic security under the new modified BDHE assumption, and achieves the insider and
outsider security under the modified BDHE assumption.
**Definition 2. (New Modified-BDHE problem) Let (p, G, GT, e) be a bilinear group system,**
$
pick a, t, s, q, θ, r1, . . ., rθ _←_ Zp, a generator g ∈ G. Given
�
_⃗Y =_ _g, g[s], g[a], . . ., g[a][q]_ _, g[a][q][+2], . . ., g[a][2][q]_ _, g[s][(][at][+][a][)], g[at], . . ., g[a][q][t],_
�
_g[a][q][+2][t], . . ., g[a][2][q][t], g[a][q][+1]g[ar][1], . . ., g[a][q][+1]g[ar][θ]_ _, g[r][1], . . ., g[r][θ]_
$
it is hard to distinguish between T = e(g, g)[a][q][+1][s] _∈_ GT and T _←_ GT .
-----
242 VAN ANH TRINH, VIET CUONG TRINH
Assume that is an adversary that outputs b 0, 1 with advantage ϵ in solving new
_A_ _∈{_ _}_
Modified-BDHE problem in G if
� � �
Pr (Y, T[⃗] = e(g, g)[a][q][+1][s]) = 0 Pr (Y, T[⃗] = R) = 0 _ϵ._
_A_ _−_ _A_ _≥_
��� ����
**Definition 3. The new Modified-BDHE assumption is secure if no polytime adversary has**
a non-negligible advantage in solving the new Modified-BDHE problem.
Intuitively, to compute e(g, g)[a][q][+1][s] one should know one of the values g[a][q][+1] or g[a][q][+1][t] or
_e(g, g)[sar][i], i_ [θ], but these elements are not provided in _Y[⃗] ._
_∈_
**4.1.** **Selective semantic security**
**Theorem 1. Let β[∗]** _be the challenge access policy, from β[∗]_ _we construct the corresponding_
_challenge LSSS matrix L’ of size ℓ[′]_ _n[′]_ _and map function ρ[′]._ _We next describe β[∗]_ =
_×_
_β1[∗]_ _m_ _[where][ β]i[∗][, i][ = 1][, . . ., m][ are disjoint sets and then construct the corresponding]_
_[∨· · · ∨]_ _[β][∗]_
_challenge LSSS matrix L[∗]_ _of size ℓ[∗]_ _n[∗]_ _and map function ρ[∗]. If those LSSS matrices satisfy_
_×_
_ℓ[′], n[′], ℓ[∗], n[∗]_ _q, and if θ_ _k[∗]_ _q[∗]_ _where k[∗]_ _and q[∗]_ _are maximum number of combinations_
_≤_ _≥_ _·_
_of keywords in a trapdoor and maximum number of trapdoor queries corresponding to β[∗]_
_adversary can make, respectively, our scheme is selectively semantic secure under the new_
_Modified-BDHE assumption._
Compare to the proof in [13], here the simulator needs to answer additional corruption
trapdoor query. To answer this kind of query, the simulator uses the elements g[a][q][+1]g[a.r][1], . . .,
_g[a][q][+1]g[a.r][θ]_ _, g[r][1], . . ., g[r][θ]_ . Note that these elements only appear in new Modified-BDHE assumption, not in Modified-BDHE assumption. The rest of the proof of this theorem is similar
to the one in [13].
**4.2.** **Keyword security**
**4.2.1.** **Insider security**
**Theorem 2. Assume that β[∗]** = β1[∗] _[∨· · · ∨]_ _[β]m[∗]_ _[is the challenge access policy and from][ β][∗]_
_construct a corresponding challenge LSSS matrix L[∗]_ _of size ℓ[∗]_ _n[∗]_ _and map function ρ[∗]._
_×_
_If this LSSS matrix satisfies ℓ[∗], n[∗]_ _q, our scheme achieves insider security under the_
_≤_
_Modified-BDHE assumption in the random oracle model._
_Proof_
In this proof we show that the simulator who attacks Modified - BDHE assumption
_S_
can simulate an adversary who attacks our scheme in the insider security game as defined
_A_
in the Subsection 2.1.3. As a result, if wins with non-negligible advantage then also can
_A_ _S_
win with non-negligible advantage. More precisely:
At the setup phase, is given an instance of Modified-BDHE assumption, and then receives
_S_
the challenge access policy β[∗] = β1[∗] _[∨· · · ∨]_ _[β]m[∗]_ [as well as][ KF] 0[∗] [= (][kf]0[∗],1[, . . ., kf]0[∗],m[′][) and]
_KF_ 1[∗] [= (][kf]1[∗],1[, . . ., kf]1[∗],m[′][) from][ A][. Note that][ β]i[∗][, i][ = 1][, . . ., m][ are disjoint sets. From]
challenge access policy β[∗] = β1[∗] _[∨· · · ∨]_ _[β]m[∗]_ [, simulator builds LSSS matrix (][M]ℓ[∗][∗]×n[∗][, ρ][∗][)]
-----
A CIPHERTEXT-POLICY ATTRIBUTE-BASED SEARCHABLE ENCRYPTION SCHEME 243
such that both ℓ[∗], n[∗] _q. To program the parameters for the system, simulator picks_
_≤_
_α[′]_ _←$_ Zp and implicitly sets α = α′ + _aq+1, then computes e(g, g)α = e(ga, gaq_ )e(g, g)α′.
The simulator finds sets of rows of matrix M _[∗]: I1, . . ., Im where {ρ(i), i ∈_ _Ij} = βj[∗]_
(note that Ij, j = 1, . . ., m are disjoint sets since βj[∗] [are disjoint sets). Now,][ β][∗] [can be]
rewritten as (∧ρ(i))i∈I1 ∨ (∧ρ(i))i∈I2 ∨· · · ∨ (∧ρ(i))i∈Im.
To program h1, . . ., hN _,_ _h[˜]1, . . .,_ _h[˜]N_, the simulator implicitly defines the vector
_−→y = (t, ta, ta2, . . ., tan[∗]−1)⊥_ _∈_ Znp _[∗][.]_
Let _[−→]Λ_ = (λ1, . . ., _λℓ∗)_ = _M_ _[∗]_ _· [−→]y_ be the vector shares, for j = 1, . . ., ℓ[∗],
_λj =_ [�]i∈[n[∗]] _[M]j,i[∗]_ _[ta][i][−][1][.]_
He/she next finds {ωi}1≤i≤ℓ∗ where for all j = 1, . . ., m, [�]i∈Ij _[ω][i][ ·][ λ][i][ =][ t][. Note that]_
we can find {ωi}1≤i≤ℓ∗ since from the property of LSSS matrix, there exists {ωi}1≤i≤ℓ∗
such that for all j = 1, . . ., m, [�]i∈Ij _[ω][i][ ·][ M]i[∗]_ [= (1][,][ 0][, . . .,][ 0)][.]
For each hj, _h[˜]j, 1 ≤_ _j ≤_ _N, where there exists an index i ∈_ [ℓ[∗]] such that j = ρ[∗](i)
$
(note that the function ρ[∗] is injective), the simulator chooses zj, ˜zj _←_ Zp and compute:
Note that the simulator knows matrix M _[∗]_ and g[ta][k] where k [n[∗]] from the instance
_∈_
of Modified-BDHE assumption.
� �
_hj = g[z][j]_ _· g[ω][i]_ _k∈[n[∗]]_ _[M]i,k[∗]_ _[ta][k]_ = g[z][j] · g[aω][i][λ][i]; ˜hj = g[z][˜][j] · g[ω][i] _k∈[n[∗]]_ _[M]i,k[∗]_ _[ta][k]_ = g[z][˜][j] · g[aω][i][λ][i].
Otherwise, the simulator chooses zj, ˜zj _←$_ Zp and computes hj = gzj _, ˜hj = gz˜j_ . We
note that {hj, _h[˜]j}j=1,...,N are distributed randomly due to choosing randomly zj, ˜zj._
To program g[λ], the simulator implicitly sets λ = _a[q]_ and computes g[λ] = (g[a][q] )[−][1].
_−_
Simulator also chooses hash functions _,_ [˜] and in this proof simulator models _,_ [˜]
_H_ _H_ _H_ _H_
as random oracles. At the end of this phase, the simulator gives following param to
_A_
(g, g[a], g[λ], e(g, g)[α], h1, . . ., hN _,_ _h[˜]1, . . .,_ _h[˜]N_ _, H,_ _H[˜])._
Query phase 1. In this phase, the simulator needs to answer five types of query:
1. The hash query.
2. The corrupted trapdoor query (Bu, Wi = ( ˜wi1, . . ., ˜wik )) where Wi doesn’t “satisfy” KF 0[∗] [or][ KF] 1[∗][, that means there doesn’t exist any triple (][i][j][, b, b][′][) such that]
_w˜ij = kfb,b[∗]_ _[′][.]_
3. The corrupted trapdoor query (Bu, Wi) where Wi “satisfies” KF 0[∗] [or][ KF] 1[∗][, but]
_Bu doesn’t “satisfy” β[∗]._
4. The partially corrupted trapdoor query (Bu, Wi) where Wi “satisfies” KF 0[∗] [or]
_KF_ 1[∗] [and][ B][u] [“satisfies”][ β][∗][. Note that user only keeps (][{][tds][0][,j][}]j∈[k][,][ ˜][W][i][) secret]
and publishes the rest of tds to the public domain. That means can know
_A_
({tds1,j, {tds2,j,ℓ}ℓ∈Bu}j∈[k], tds0, {tdsi}i∈Bu) for any (Bu, Wi = ( ˜wi1, . . ., ˜wik )).
-----
244 VAN ANH TRINH, VIET CUONG TRINH
5. The partially corrupted secret key query Bu for any Bu. The reason is that user
only keeps du0 secret.
Regarding the hash query: Simulator creates two lists _,_ ˜, at the beginning
_•_ _L_ _L_
_L,_ _L[˜] are empty. For each hash query corresponding to ˜wi which doesn’t satisfy_
_KF_ 0[∗] [or][ KF] 1[∗][, the simulator first checks whether ˜][w][i] [has been queried before. If]
not, he/she chooses yi _←$_ Zp and adds triple ( ˜wi, gyi, yi) ∈ ({0, 1}∗, G, Zp) into
_L and returns g[y][i]_ to A. Otherwise, he/she simply finds the triple ( ˜wi, g[y][i], yi)
and returns g[y][i] to A. In the case ˜wi “satisfies” KF 0[∗] [or][ KF] 1[∗][, the simulator first]
$
checks whether ˜wi has been queried before. If not, he/she chooses yi _←_ Zp and
adds triple ( ˜wi, g[−][a] _· g[y][i], yi) into L and returns g[−][a]_ _· g[y][i]_ to A. Otherwise, he/she
simply finds the triple ( ˜wi, g[−][a] _· g[y][i], yi) and returns g[−][a]_ _· g[y][i]_ to A. For each hash
query corresponding to (Ki, kfj) where Ki ∈ GT, kfj ∈{0, 1}[∗], simulator first
$
checks whether (Ki, kfj) has been queried before. If not, he/she chooses yij _←_ Zp
and adds triple (Ki, kfj, yij ) into L[˜]. Otherwise, he/she simply finds the triple
(Ki, kfj, yij ) and returns yij to A.
_• Regarding the second type of query: A first sends Wi = ( ˜wi1, . . ., ˜wik_ ) and B(u) to
simulator with the requirement that Wi doesn’t satisfy KF 0[∗] [or][ KF] 1[∗][. To program]
each tds0,j, j ∈ [k], the simulator first checks whether ˜wij _, j ∈_ [k] has been queried
before. If not, he/she chooses yij _←$_ Zp and adds triple ( ˜wij _, gyij, yij_ ) into L. In
both ways, simulator knows yij from L, and H( ˜wij ) = g[y][ij] since Wi doesn’t
$
satisfy KF 0[∗] [or][ KF] 1[∗][. Next, simulator first chooses][ s][u][, r][j] _←_ Zp then computes
tds0,j = g[α][′]g[as][u]g[ar][j] (g[λ])[y][ij] = g[α]g[as][u]g[ar][j] (g[a]H( ˜wij ))[λ].
Note that g[α][′] = g[α][′] _g[a][q][+1]_ _g[−][a][q][+1]_ = g[α] _g[aλ], since g[λ]_ = g[−][a][q] . Since simulator
_·_ _·_ _·_
knows su, rj, j ∈ [k], he/she can easily computes the rest of the trapdoor for any
set (u). Finally, simulator returns tds to .
_B_ _A_
_• Regarding the third type of query: A first sends Wi = ( ˜wi1, . . ., ˜wik_ ) and B(u) to
simulator with the requirement that B(u) doesn’t satisfy β[∗] and Wi “satisfies”
_KF_ 0[∗] [or][ KF] 1[∗][. The simulator first finds a vector][ −→][x][ = (][x][1][, . . ., x][n][∗][)][ ∈] [Z][n]p _[∗]_ such
that x1 = −1 and for all i where ρ[∗](i) ∈B(u) the product ⟨[−→]x · Mi[∗][⟩] [= 0. The]
$
simulator continues to pick ζ _←_ Zp and implictly define the value su as
_su = ζ + x1a[q]_ + x2a[q][−][1] + · · · + xn[∗]a[q][−][n][∗][+1].
The simulator computes
� �
_du0 = g[α][′]g[aζ]_ (g[a][q][+1][−][i])[x][i] = g[α] _· g[a][·][s][u]; d[′]u0_ [=][ g][ζ] (g[a][q][+1][−][i])[x][i] = g[s][u].
_i=2,...,n[∗]_ _i=1,...,n[∗]_
For j (u) such that there is no i [ℓ[∗]] satisfying ρ[∗](i) = j. The simulator
_∈B_ _∈_
knows values zj and computes h[s]j[u] = (g[s][u])[z][j] . For j ∈B(u) such that there is an
indice i [ℓ[∗]] satisfying ρ[∗](i) = j. The simulator computes
_∈_
�
_h[s]j[u]_ = (g[s][u])[z][j] _· g[(][ζ][+][x][1][a][q][+][···][+][x][n][∗]_ _[a][q][−][n][∗][+1][)][ω][i]_ _k∈[n[∗]]_ _[M]i,k[∗]_ _[ta][k]_ _._
-----
A CIPHERTEXT-POLICY ATTRIBUTE-BASED SEARCHABLE ENCRYPTION SCHEME 245
Note that the product ⟨[−→]x · _Mi[∗][⟩]_ [= 0 thus the simulator doesn’t need to know the]
unknown term of form g[a][q][+1][t] to compute h[s]j[u][, all other terms he/she knows from]
the assumption. Simulator simply sets g[s][u] and {h[s]j[u][}][j][∈B]u [as tds][0][ and][ {][tds][i][}][i][∈B]u[,]
respectively. To program {tds0,j, tds1,j, {tds2,j,ℓ}ℓ∈Bu}j∈[k], the simulator considers two cases:
1. ˜wij “satisfies” KF 0[∗] [or][ KF] 1[∗][, that means there exists at least a triple (][i][j][, b, b][′][)]
such that ˜wij = kfb,b[∗] _[′][.]_ Simulator checks whether ˜wij has been queried
before. If not, he/she chooses yij _←$_ Zp and adds triple ( ˜wij _, g−agyij, yij_ )
into L. In both ways, simulator knows yij from L, and H( ˜wij ) = g[−][a]g[y][ij] .
$
Next, simulator chooses rj _←_ Zp then computes
tds0,j = g[α]g[as][u]g[ar][j] (g[−][a][q] )[y][ij] = g[α]g[as][u]g[ar][j] (g[a]H( ˜wij ))[λ].
Note that λ = −a[q]. With known rj, simulator can easily compute tds1,j,
_{tds2,j,ℓ}ℓ∈Bu._
2. ˜wij doesn’t “satisfy” KF 0[∗] [or][ KF] 1[∗][. Simulator checks whether ˜][w][i]j [has been]
queried before. If not, he/she chooses yij _←$_ Zp and adds triple ( ˜wij _, gyij, yij_ )
into L. In both ways, simulator knows yij from L, and H( ˜wij ) = g[y][ij] . Next,
$
simulator picks ζj _←_ Zp and implicitly defines the value rj as
_rj = ζj + x1a[q]_ + x2a[q][−][1] + · · · + xn[∗]a[q][−][n][∗][+1],
then similarly computes g[α]g[ar][j] _, g[r][j]_ _, {h[˜]ℓrj_ _}ℓ∈Bu as above (note that rj now_
plays the role as su). He/she then sets g[−][r][j] _, {h[˜]ℓ−rj_ _}ℓ∈Bu as tds1,j, {tds2,j,ℓ}ℓ∈Bu_
respectively, and computes
tds0,j = g[α]g[as][u]g[−][α]g[−][ar][j] _g[α][′](g[−][a][q]_ )[y][ij] = g[α]g[as][u]g[−][ar][j] (g[a]H( ˜wij ))[λ].
Note that g[α] = g[α][′]g[a][q][+1] and (g[a]H( ˜wij ))[λ] = g[−][a][q][+1](g[−][a][q] )[y][ij] _._
Finally, simulator returns tds to .
_A_
Regarding the fourth and fifth types of query: It is straightforward since the
_•_
unknown value g[α] only appears in the tds0,j and du0, therefore simulator can
simply choose su, {rj}j∈[k] _←$_ Zp and then computes tds, or choose su, $← Zp and
then computes du.
Challenge: The simulator picks a random bit _b,_ computes _C0[∗]_ = _g[s]_ and
(C1[∗][, . . ., C]m[∗] [) =][ I,][ ( ˜][C]1[∗][, . . .,][ ˜][C]m[∗] [) =][ J][, where]
_I =_ �g[s][(][a][+][at][)]g�i∈I1 _[sz][ρ][∗][(][i][)], . . ., g[s][(][a][+][at][)]g�i∈Im_ _[sz][ρ][∗][(][i][)]�_
_,_
=
g[s], (g[a][ �] _hi)[s], . . ., (g[a][ �]_ _hi)[s]_
_i∈β1[∗]_ _i∈βm[∗]_
-----
246 VAN ANH TRINH, VIET CUONG TRINH
_J =_ �g[s][(][a][+][at][)]g�i∈I1 _[s][z][˜][ρ][∗][(][i][)], . . ., g[s][(][a][+][at][)]g�i∈Im_ _[s][z][˜][ρ][∗][(][i][)]�_
_h˜i)[s], . . ., (g[a][ �]_
_i∈βm[∗]_
_._
=
(g[a][ �]
_i∈β1[∗]_
_h˜i)[s]_
To computes {Ki[∗] _b,i_ [has been queried before.]
_[}][i][∈][[][m][′][]][, simulator first checks whether][ kf]_ _[∗]_
$
If not, he/she chooses yi _←_ Zp and adds triple (kfb,i∗ _[, g][−][a][g][y][i][, y][i][) into][ L][. Otherwise,]_
he/she finds (kfb,i[∗] _[, g][−][a][g][y][i][, y][i][) from][ L][. In both ways, simulator knows][ y][i][ from][ L][, and]_
_H(kfb,i[∗]_ [) =][ g][−][a][g][y][i][. He/she computes]
_Xi[∗]_ [=][ T][ ·][ e][(][g][s][, g][α][′][)][ ·][ e][(][g][s][,][ (][g][λ][)][y][i][) =][ T][ ·][ e][(][g][s][, g][α][′][)][ ·][ e][(][g][λ][, g][a][H][(][kf]b,i[∗] [))][s]
then computes Ki[∗] [= ˜][H][(][X]i[∗][, kf]b,i[∗] [). Finally, he/she outputs]
_ct[′∗]_ = (C0[∗][,][ {][C]i[∗][}]i∈[m][,][ {][ ˜][C]i[∗][}]i∈[m][,][ {][K]i[∗][}]i∈[m[′]][)][.]
Note that if T = e(g, g)[a][q][+1][s] then ct[′∗] is in valid form.
Query phase 2. Similar to phase 1.
Guess: gives his guess b[′] to, will output its guess 0 corresponding to T = e(g, g)[a][q][+1][s]
_A_ _S_ _S_
if b[′] = b; otherwise, outputs its guess 1 corresponding to T is a random element.
_S_
When T = e(g, g)[a][q][+1][s], gives a perfect simulation so
_S_
� �
Pr _S(Y, T[⃗]_ = e(g, g)[a][q][+1][s]) = 0 = [1]2 [+][ Adv]A[IS][.]
When T is a random element {Ki[∗][}][i][∈][[][m][′][]] [are completely hidden from the][ A][, so]
_Pr[_ (Y, T[⃗] = R) = 0] = [1]
_S_
2 [. As a result,][ S][ is able to play the Modified -][ BDHE][ game]
with non-negligible advantage (equal to Adv[IS]A [) or][ S][ is able to break the security of]
Modified-BDHE assumption.
**Outsider Security. Outsider security is similar to the case of Insider Security, due to the**
space limitations we omit it here.
**5.** **COMPARISON**
Regarding PEKS scheme supporting both expressive searching predicate and multi-writer
/multi-reader, to our knowledge [6, 8, 9, 11, 14] are the best schemes to date. In these
schemes, the authors in [9] proposed a PEKS schemes scheme with constant size of ciphertext,
however their scheme only supports limited AND-gates access policy. The authors in [6, 8,
11, 14] managed to transform from existing KP-ABE or CP-ABE schemes to PEKS schemes,
these schemes enjoy some interesting properties such as fast keyword search or outsourcing
decryption or partially trapdoor security. Other properties such as fined-grain access policy,
public key-size, ciphertext-size are similar to ones in our scheme, but our scheme has constant
-----
A CIPHERTEXT-POLICY ATTRIBUTE-BASED SEARCHABLE ENCRYPTION SCHEME 247
size of secret key while they haven’t, moreover our model is different to their model. We give
in Fig 1 the comparison between our model and the model in [6, 8, 11, 14]. We can easily
see from Fig 1 that there is no TTG in our model, user relies solely on his/her secret key to
generate a trapdoor. In contrast, in their model, TTG takes charge of generating trapdoors
and therefore should be always online. Compare to their model, our model has two following
advantages:
There is no TTG in our model, we therefore can save the system resource and the
_•_
communication overhead between user and TTG.
In our model, user uses his/her secret key to generate trapdoors, cloud server relies on
_•_
such trapdoors to search corresponding ciphertexts, user therefore is able to decrypt all
the resulted ciphertexts. In contrast, in their model, TTG takes charge of generating
trapdoors and user’s secret key is not involved in this process, this leads to the fact that
the resulted ciphertexts may contain ciphertexts for which the user cannot decrypt. We
argue that our model is more useful in practice since it is waste time and communication
overhead if user receives the ciphertexts for which he/she cannot decrypt.
Data Encrypt documents Cloud Data Encrypt documents Cloud
owner and corresponding server owner and corresponding server
keywords keywords
3.Trapdoor
1.Trapdoor 4.Resulted
ciphertexts
2.Resulted
Trusted
ciphertexts
User User 1.Trapdoor request Trapdoor
Generator
2.Trapdoor
_Figure 1. Our model on the left and their model on the right_
We also note that the scheme in [15] can deal with well the problem of trapdoor security, and moreover this scheme is very efficient in term of both communication and computation. However, this scheme is in the different type to our scheme since our scheme
supports expressive searching while the scheme in [15] supports equality queries. Consider the example in [6], W = (w1, w2, w3) where w1 = “Diabetes[′′], w2 = “Age : 30[′′] and
_w3 = “Weight : 150 −_ 200[′′], user in our scheme can search for ciphertexts which has keyword either “Diabetes[′′] or “Weight : 150 200[′′]. While the user in the scheme in [15] must
_−_
specify the keyword search is ”Diabetes” or “Weight : 150 200[′′] and then receives only
_−_
the ciphertext corresponding to the keyword search. For example, if the keyword search is
“Diabetes[′′], user only receives the ciphertext corresponding to “Diabetes[′′]. This is similar
to the difference between traditional public key encryption and attribute-based encryption.
Encrypt documents
and corresponding
keywords
1.Trapdoor request
2.Trapdoor
2.Resulted
ciphertexts
3.Trapdoor
2.Trapdoor
User
Data
owner
Cloud
server
-----
248 VAN ANH TRINH, VIET CUONG TRINH
**6.** **CONCLUSION**
In this paper, we propose a CP-ABSE scheme which supports both expressive searching
predicate and multi-writer/multi-readers. To our knowledge, our scheme has some interesting
properties such as constant-size of secret key and in the non-interactive model. Our scheme
will become very efficient when the number of combinations of keywords to which a user
would like to search is small. Our scheme is therefore very suitable for a large class of
applications in practice for which the aforementioned case falls into.
**ACKNOWLEDGMENTS**
This research is funded by Vietnam National Foundation for Science and Technology
Development (NAFOSTED) under grant number 102.01-2018.301.
**REFERENCES**
[1] M. R. Asghar, G. Russello, B. Crispo, and M. Ion, “Supporting complex queries and access
policies for multi-user encrypted databases,” in Proceeding CCSW ’13 Proceedings of the
_2013 ACM workshop on Cloud computing security workshop, Berlin, Germany, November_
4, 2013, pp 77–88. Doi 10.1145/2517488.2517492
[2] D. Boneh, G. Di Crescenzo, R. Ostrovsky, and G. Persiano, “Public key encryption with keyword
search,” Advances in Cryptology - EUROCRYPT 2004 International Conference on the
_Theory and Applications of Cryptographic Techniques, Interlaken, Switzerland, May 2-6,_
2004, pp 506–522.
[3] D. Cash, S. Jarecki, C. S. Jutla, H. Krawczyk, M.-C. Rosu, and M. Steiner, “Highly-scalable
searchable symmetric encryption with support for Boolean queries”, in Advances in Cryptology
_CRYPTO 2013 33rd Annual Cryptology Conference, Proceedings, Part I, Santa Barbara,_
CA, USA, August 18-22, 2013.
[4] S. Canard and V. C. Trinh, “Constant-size ciphertext attribute-based encryption from multichannel broadcast encryption,” Information Systems Security. 12th International Confe_rence, ICISS 2016, Proceedings, Jaipur, India, December 16-20, 2016._
[5] H. Cui, R. Deng, J. Liu, and Y. Li, “Attribute-based encryption with expressive and authorized
keyword search,” Information Security and Privacy. 22nd Australasian Conference, ACISP
_2017, Proceedings, Auckland, New Zealand, July 35, 2017, Part I, pp. 106–126._
[6] H. Cui, Z. Wan, R. Deng, G. Wang, and Y. Li, “Efficient and expressive keyword search over
encrypted data in the cloud,” IEEE Transactions on Dependable and Secure Computing,
vol. 15, no. 3, pp. 409–422, 2018. Doi 10.1109/TDSC.2016.2599883
[7] S. Hohenberger and B. Waters, “Attribute-based encryption with fast decryption,” Public_Key Cryptography_ _PKC 2013. 16th International Conference on Practice and Theory_
_in Public-Key Cryptography, Proceedings, Nara, Japan, February 26_ March 1, 2013, pp.
162–179.
[8] Jianting Ning, Zhenfu Cao, Xiaolei Dong, Kaitai Liang, Hui Ma, Lifei Wei, “Auditable σTime Outsourced Attribute-Based Encryption for Access Control in Cloud Computing,” IEEE
_Transactions on Information Forensics and Security, vol. 13, no. 1, pp. 94–105, 2018. Doi_
10.1109/TIFS.2017.2738601
-----
A CIPHERTEXT-POLICY ATTRIBUTE-BASED SEARCHABLE ENCRYPTION SCHEME 249
[9] Jinguang Han, Ye Yang, Joseph K. Liu, Jiguo Li,Kaitai Liang,Jian Shen, “Expressive attributebased keyword search with constant-size ciphertext,” In Soft Computing Journal, vol. 22, no.
15, pp 5163-5177, August 2018.
[10] A. Kiayias, O. Oksuz, A. Russell, Q. Tang, and B. Wang, “Efficient encrypted keyword search for
multi-user data sharing,” Computer Security ESORICS 2016. 21st European Symposium
_on Research in Computer Security, Proceedings, Heraklion, Greece, September 26-30, 2016,_
Part I, pp 173–195.
[11] J. Lai, X. Zhou, R. H. Deng, Y. Li, and K. Chen, “Expressive search on encrypted data,” ASIA
_CCS ’13 Proceedings of the 8th ACM SIGSAC Symposium on Information, Computer and_
_Communications Security, Hangzhou, China, May 8-10, 2013, pp 243–251._
[12] Z. Lv, C. Hong, M. Zhang, and D. Feng, “Expressive and secure searchable encryption in the
public key setting,” Information Security 17th International Conference, ISC 2014, Pro_ceedings, Hong Kong, China, October 12-14, 2014, pp 364–376._
[13] Q. M. Malluhi, A. Shikfa, and V. C. Trinh, “A ciphertext-policy attribute-based encryption
scheme with optimized ciphertext size and fast decryption,” Proceeding ASIA CCS ’17 Procee_dings of the 2017 ACM on Asia Conference on Computer and Communications Security,_
Abu Dhabi, United Arab Emirates, April 02 - 06, 2017, pp. 230–240.
[14] R. Meng, Y. Zhou, J. Ning, K. Liang, J. Han, and W. Susilo, “An efficient key-policy attributebased searchable encryption in prime-order groups,” Provable Security 11th International
_Conference, ProvSec 2017, Proceedings, Xi’an, China, October 23-25, 2017, pp. 39–56._
[15] Qiong Huang and Hongbo Li, “An efficient public-key searchable encryption scheme secure
against inside keyword guessing attacks,” Information Sciences Journal, vol. 403–404, pp.
1–14, September 2017. Doi.org/10.1016/j.ins.2017.03.038
[16] D. X. Song, D. Wagner, and A. Perrig, “Practical techniques for searches on encrypted data,”
_Proceeding 2000 IEEE Symposium on Security and Privacy, Berkeley, CA, USA, May 14–17,_
2000, pp. 44–55. Doi 10.1109/SECPRI.2000.848445
[17] S. Sun, J. K. Liu, A. Sakzad, R. Steinfeld, and T. H. Yuen, “An efficient non-interactive multiclient searchable encryption with support for boolean queries,” Computer Security ESORICS
_2016. 21st European Symposium on Research in Computer Security, Proceedings, Part I,_
Heraklion, Greece, September 26–30, 2016, pp. 154–172.
[18] Y. Wang, J. Wang, S. Sun, J. Liu, W. Susilo, and X. Chen, “Towards multi-user searchable encryption supporting boolean query and fast decryption,” Provable Security 11th International
_Conference, ProvSec 2017, Proceedings, Xi’an, China, October 23–25, 2017, pp. 24–38._
[19] Y. Yang, H. Lu, and J. Weng, “Multi-user private keyword search for cloud computing,”
_2011 IEEE Third International Conference on Cloud Computing Technology and Science,_
Athens, Greece, 29 Nov.- Dec. 01, 2011, pp. 264–271. Doi 10.1109/CloudCom.2011.43
_Received on March 05, 2019_
_Revised on April 18, 2019_
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.15625/1813-9663/0/0/13667?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.15625/1813-9663/0/0/13667, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "GOLD",
"url": "https://vjs.ac.vn/index.php/jcc/article/download/13667/383095"
}
| 2,019
|
[] | true
| 2019-08-06T00:00:00
|
[
{
"paperId": "31f7bbf30a0692e87c0c7874a5b38a1d916dad92",
"title": "An Efficient Key-Policy Attribute-Based Searchable Encryption in Prime-Order Groups"
},
{
"paperId": "243e6f59ebf485056ab0587909b1327e8f42663f",
"title": "Towards Multi-user Searchable Encryption Supporting Boolean Query and Fast Decryption"
},
{
"paperId": "274a85d51bff300d0c75a6c8ac8a85f6ea2b79de",
"title": "An efficient public-key searchable encryption scheme secure against inside keyword guessing attacks"
},
{
"paperId": "121858abcbf359cb9e6049c520bd48bb5f3cdf0c",
"title": "Expressive attribute-based keyword search with constant-size ciphertext"
},
{
"paperId": "b4a28bf0a9dc659c260b5d2acc37dc3afa752e65",
"title": "Attribute-Based Encryption with Expressive and Authorized Keyword Search"
},
{
"paperId": "1395e5687a9877b955e338ea54cc4867cb3b7fa6",
"title": "A Ciphertext-Policy Attribute-based Encryption Scheme With Optimized Ciphertext Size And Fast Decryption"
},
{
"paperId": "f684fa1eed729ffd06c6e970844c3b9f7265a43f",
"title": "Constant-Size Ciphertext Attribute-Based Encryption from Multi-channel Broadcast Encryption"
},
{
"paperId": "84e84bed92a5761e30f899720eca680f1a45352a",
"title": "An Efficient Non-interactive Multi-client Searchable Encryption with Support for Boolean Queries"
},
{
"paperId": "d279baa32365b2f8b7a6b259d0ec208c0bceff0e",
"title": "Efficient and Expressive Keyword Search Over Encrypted Data in Cloud"
},
{
"paperId": "254ccff069b29181b1696cdbfe06fb0886a996e1",
"title": "Expressive and Secure Searchable Encryption in the Public Key Setting"
},
{
"paperId": "8cc288bb4fa70839bc33893b30575cbe22f995b0",
"title": "Supporting complex queries and access policies for multi-user encrypted databases"
},
{
"paperId": "effd352328e57bc7b68e0802b7c91be386b86614",
"title": "Highly-Scalable Searchable Symmetric Encryption with Support for Boolean Queries"
},
{
"paperId": "e1156302e833962e9de91d8bbb9401bcdf31e7dd",
"title": "Expressive search on encrypted data"
},
{
"paperId": "fae557d8ba82575d1f7b8689080f13ff27df3f73",
"title": "Attribute-Based Encryption with Fast Decryption"
},
{
"paperId": "590bcd5e8c4d566b1f71128e0f62af63eaf215cd",
"title": "Multi-User Private Keyword Search for Cloud Computing"
},
{
"paperId": "afaa1636e44246ebf8f3c27ad32fca9a8ce5f0a1",
"title": "Public Key Encryption with Keyword Search"
},
{
"paperId": "95216caad0c07d9fae76416676ffa9184a88b674",
"title": "Practical techniques for searches on encrypted data"
},
{
"paperId": null,
"title": "Auditable σ- Time Outsourced Attribute-Based Encryption for Access Control in Cloud Computing"
},
{
"paperId": "210463e45c96847bd12cac858446e413877b161d",
"title": "Edinburgh Research Explorer Efficient Encrypted Keyword Search for Multi-user Data Sharing"
}
] | 17,602
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Engineering",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/02a29c5897b1fa71149d802ad7f082ae833fee5c
|
[
"Computer Science"
] | 0.861416
|
ART: sub-logarithmic decentralized range query processing with probabilistic guarantees
|
02a29c5897b1fa71149d802ad7f082ae833fee5c
|
Distributed and parallel databases
|
[
{
"authorId": "1718248",
"name": "S. Sioutas"
},
{
"authorId": "1732298",
"name": "P. Triantafillou"
},
{
"authorId": "1837376",
"name": "G. Papaloukopoulos"
},
{
"authorId": "1710324",
"name": "E. Sakkopoulos"
},
{
"authorId": "1702182",
"name": "K. Tsichlas"
},
{
"authorId": "1796253",
"name": "Y. Manolopoulos"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"Distributed and Parallel Databases",
"Distrib parallel database",
"Distrib Parallel Database"
],
"alternate_urls": null,
"id": "ceac6326-95ad-4e90-a522-b7ab052ea5a4",
"issn": "0926-8782",
"name": "Distributed and parallel databases",
"type": "journal",
"url": "https://link.springer.com/journal/10619"
}
|
We focus on range query processing on large-scale, typically distributed infrastructures, such as clouds of thousands of nodes of shared-datacenters, of p2p distributed overlays, etc. In such distributed environments, efficient range query processing is the key for managing the distributed data sets per se, and for monitoring the infrastructure’s resources. We wish to develop an architecture that can support range queries in such large-scale decentralized environments and can scale in terms of the number of nodes as well as in terms of the data items stored. Of course, in the last few years there have been a number of solutions (mostly from researchers in the p2p domain) for designing such large-scale systems. However, these are inadequate for our purposes, since at the envisaged scales the classic logarithmic complexity (for point queries) is still too expensive while for range queries it is even more disappointing. In this paper we go one step further and achieve a sub-logarithmic complexity. We contribute the ART (Autonomous Range Tree) structure, which outperforms the most popular decentralized structures, including Chord (and some of its successors), BATON (and its successor) and Skip-Graphs. We contribute theoretical analysis, backed up by detailed experimental results, showing that the communication cost of query and update operations is $O(\log_{b}^{2} \log N)$ hops, where the base b is a double-exponentially power of two and N is the total number of nodes. Moreover, ART is a fully dynamic and fault-tolerant structure, which supports the join/leave node operations in O(loglogN) expected w.h.p. number of hops. Our experimental performance studies include a detailed performance comparison which showcases the improved performance, scalability, and robustness of ART.
|
## ART : Sub-Logarithmic Decentralized Range Query Processing with Probabilistic Guarantees
S. Sioutas[1], P. Triantafillou[2], G. Papaloukopoulos[2],
E. Sakkopoulos[2], K. Tsichlas[3], and Y. Manolopoulos[3]
1 Ionian University, Department of Informatics, sioutas@ionio.gr
2 CTI and Dept. of Computer Engineering & Informatics, University of Patras,
(peter, papaloukg, sakkopul)@ceid.upatras.gr
3 Aristotle University of Thessaloniki, Department of Informatics, (tsichlas,
manolopo)@csd.auth.gr
Abstract. We focus on range query processing on large-scale, typically distributed infrastructures, such as clouds of thousands of nodes of
shared-datacenters, of p2p distributed overlays, etc. In such distributed
environments, efficient range query processing is the key for managing
the distributed data sets per se, and for monitoring the infrastructure’s
resources. We wish to develop an architecture that can support range
queries in such large-scale decentralized environments and can scale in
terms of the number of nodes as well as in terms of the data items stored.
Of course, in the last few years there have been a number of solutions
(mostly from researchers in the p2p domain) for designing such largescale systems. However, these are inadequate for our purposes, since at
the envisaged scales the classic logarithmic complexity (for point queries)
is still too expensive while for range queries it is even more disappointing. In this paper [4] we go one step further and achieve a sub-logarithmic
complexity. We contribute the ART [5] structure, which outperforms the
most popular decentralized structures, including Chord (and some of
its successors), BATON (and its successor) and Skip-Graphs. We contribute theoretical analysis, backed up by detailed experimental results,
showing that the communication cost of query and update operations is
O(log[2]b [log][ N] [) hops, where the base][ b][ is a double-exponentially power of]
two and N is the total number of nodes. Moreover, ART is a fully dynamic and fault-tolerant structure, which supports the join/leave node
operations in O(log log N ) expected w.h.p number of hops. Our experimental performance studies include a detailed performance comparison
which showcases the improved performance, scalability, and robustness
of ART.
Keywords:Distributed Data Structures, P2P Data Management.
4 A limited and preliminary version of this work has been presented as brief announcement in Twenty-Ninth Annual ACM SIGACT-SIGOPS Symposium on Principles of
Distributed Computing, Zurich, Switzerland July 25-28, 2010 [28]
5 Autonomous Range Tree
-----
2 S. Sioutas et al.
### 1 Introduction and Motivation
Decentralized range query processing is a notoriously difficult problem to solve
efficiently and scalably in decentralized network infrastructures. It has been studied in the last years extensively, particularly in the realm of p2p, which is increasingly used for content delivery among users. There are many more real-life
applications in which the problem also materializes. Consider the (popular nowadays) cloud infrastructures for content delivery. Monitoring of thousand of nodes,
where thousands of different applications from different organizations execute,
is an apparent requirement. This monitoring process often requires support for
range queries over this decentralized infrastructure: consider range queries that
are issued in order to identify which of the cloud nodes are under-utilized, (i.e.,
utilization < threshold) in order to assign to them more data & tasks and
better exploit all available resources, increasing the revenues of the cloud infrastructure, or to identify overloaded nodes, (load > threshold) in order to avoid
bottlenecks in the cloud, which hurts overall performance, and revenues.
Each node in the cloud maintains a tuple with attributes: utilization, OS,
load, NodeId, e.t.c. Collectively, these makeup a relation, CloudNodes, and we
wish to execute queries such as:
SELECT NodeId
FROM Cloudnodes
WHERE low < utilization < high
or point and range queries, e.g.
SELECT NodeId
FROM Cloudnodes
WHERE low < utilization < high and OS=UNIX
An acceptable solution for processing range queries in such large-scale decentralized environments must scale in terms of the number of nodes as well as
in terms of the number of data items stored. The available solutions for architecting such large-scale systems are inadequate for our purposes, since at the
envisaged scales (trilions of data items at millions of nodes) the classic logarithmic complexity (for point queries) offered by these solutions is still too expensive.
And for range queries, it is even more disappointing. Further, all available solutions incur large overheads with respect to other critical operations, such as
join/leave of nodes, and insertion/deletion of items. Our aim with this work is
to provide a solution that is comprehensive and outperforms related work with
respect to all major operations, such as lookup, join/leave, insert/delete, and to
the required routing state that must be maintained in order to support these operations. Specifically, we aim at achieving a sub-logarithmic complexity for all
the above operations!
Peer-to-peer (P2P) systems have become very popular, in both academia
and industry. They are widely used for sharing resources like music files etc.
Search for a given ID, is a crucial operation in P2P systems, and there has been
considerable recent work in devising effective distributed search (a.k.a. lookup)
techniques. The proposed structures include a ring as in Chord [15], a multiple
dimensional grid as in CAN [22], a multiple list as in SkipGraph [2, 10], or a
tree as in PHT [24], BATON [13] and BATON* [14]. Most search structures
(including all the ones just mentioned except for BATON* and PHT) bound
-----
Autonomous Range Tree 3
the search cost to a base 2 logarithm of the search space: for a system with
N peer nodes, the search cost is bounded by O(log N ). Relative to tree-based
indexes, a disadvantage of PHTs (Prefix Hash Trees) is that their complexity
is expressed in terms of the log of the domain size, D, rather than the size of
the data set, N and depends on distribution over D bit keys. BATON* is a
−
multi-way search tree, which reduces the search cost to O(logm N ), where m is
the tree fanout. The penalty paid is a larger update cost, but no worse than
linear in m. One of the distributed indexes with high fanout is the P-Tree [5],
where each peer maintains a B[+]-tree leaf and a path of virtual index nodes from
the root to the specific leaf. Search is very effective, but updates are expensive,
possibly requiring substantial synchronization effort. BATON* extends BATON
by allowing a fanout of m > 2. Thus, the search cost becomes O(logm N ), as
expected. Moreover, the cost of updating routing tables is O(m logm N ) only,
as compared to O(log2 N ) in BATON - an improvement that is better than
linear in m. Furthermore, BATON* has better fault tolerance properties than
BATON, and supports load balancing more efficiently. In fact, the system’s fault
tolerance, measured as the number of nodes that must fail before the network
is partitioned, increases linearly with m. Similarly, the expected cost of load
balancing decreases linearly with m.
Our Results: In this paper we present the ART structure, which outperforms the most popular decentralized structures, including Chord (and some of
its successors), BATON and BATON* and Skip-Graphs. ART is an exponentialtree structure, which remains unchanged w.h.p., and organizes a number of fullydynamic buckets of peers. We provide and analyze all relevant algorithms for
accessing ART. We contribute theoretical analysis, backed up by detailed experimental results, showing that the communication cost of query and update
operations is O(log[2]b [log][ N] [) hops, where the base][ b][ is a double-exponentially]
power of two. Moreover, ART is a fully dynamic and fault-tolerant structure,
which supports the join/leave node operations in O(log log N ) expected w.h.p
number of hops. Since ART is a tree based system, our experimental performance studies include our development of BATON* (the best current tree based
system), and a detailed performance comparison which showcases the improved
performance, scalability, and robustness of ART.
In Section 2 we present more thoroughly key previous work. Section 3 describes the ART structure and analyzes its basic functionalities. Section 4 presents
a thorough experimental evaluation; Section 5 presents some interesting heuristics and thresholds, whereas Section 6 concludes the paper.
### 2 Previous Work
Existing structured P2P systems can be classified into three categories: distributed hash table (DHT) based systems, skip list based systems, and tree based
systems. There are several P2P DHT architectures like Chord [15], CAN [22],
Pastry [23], Tapestry [31], Kademlia [20] and and Kelips [9]. Unfortunately, these
systems cannot easily support range queries since DHTs destroy data ordering.
This means that they cannot support common queries such as ”find all research
papers published from 2004 to 2008”. To support range queries, inefficient DHT
variants have been proposed (for details see [8], [25], [1], [29]).
-----
4 S. Sioutas et al.
Skip list based systems such as Skip Graph [2, 10] and Skip Net [12] are
based on the skip-list structure. To provide decentralization they use randomized
techniques to create and maintain the structure. Moreover, they can support
both exact match queries and range queries by partitioning data into ranges of
values. However, they cannot guarantee data locality (which hurts efficient range
query processing) and load balancing in the system.
Tree based systems also carry their own disadvantages. P-Grid [5] utilizes a
binary prefix tree. It can neither guarantee the bound of search steps since it
cannot control the tree height. An arbitrary multi-way tree was proposed in [19],
where each node maintains links to its parent, children, sibling and neighbors.
It also suffers from the same problem. P-Tree [5] utilizes a B[+]-tree on top of the
CHORD overlay network, and peers are organized as a CHORD ring, each peer
maintaining a data leaf and a left most path from the root to that B[+]-tree node.
This results in significant overhead in building and maintaining the consistency
of the B[+]-tree. In particular, a tree has been built for each joining node, and
periodically, peers have to exchange their stored B[+]-tree for checking consistency.
BATON [13] utilizes a binary balanced tree and as a consequence, it can control
the tree height and avoid the problem of P-Grid. Nevertheless, similarly to other
P2P systems, BATON’s search cost is bounded by O(log2 N ). BATON* [14] is
an overlay multi-way tree based on B-trees, with better searching performance.
The penalty paid is a marginally larger update cost.
Systems like MAAN [4], Mercury [3] and DIM [18] support multi-attribute
queries in a multi-dimensional space. BATON* can also effectively support queries
over multiple attributes. In addition to supporting the use of multiple attributes
in a single index, BATON* further introduces the notion of attribute classification, based on the importance of the attribute for querying, and the notion of
attribute groups. In particular, BATON* relies on the construction of multiple
independent indexes for groups of one or more attributes. For further details
about the suggested techniques for partitioning attributes into such groups, see
[14].
P2P Lookup, Insert, Maximum Size Join/
architectures Delete key of routing table Depart peer
CHORD O(log N ) O(log N ) O(log N ) w.h.p.
H-F-Chord(a) O(log N/ log log N ) O(log N ) O(log N )
LPRS-Chord O(log N ) O(log N ) O(log N )
Skip Graphs O(log N ) O(1) O(log N ) amortized
BATON O(log N ) O(log N ) O(log N ) w.h.p.
BATON* O(logm N ) O(m logm N ) O(m logm N )
ART-tree O(log[2]b [log][ N] [)] O(N [1][/][4]/ log[c] N ) O(log log N ) expected w.h.p.
Table 1. Performance comparison between ART, Chord, BATON and Skip Graphs.
For comparison purposes, in Table 1 we present a qualitative evaluation with
respect to elementary operations between ART, Skip-Graphs, Chord and its
-----
Autonomous Range Tree 5
newest variations (F-Chord(α) [26], LPRS-Chord [30]), BATON [13] and its
newest variation BATON* [14]. It is noted that c is a big positive constant.
### 3 Our Solution
First, we build the LRT (Level Range Tree) structure, one of the basic components of the final ART structure. LRT will be called upon to organize collections
of peers at each level of ART.
3.1 Building LRT structure
LRT is built by grouping peers having the same ancestor and organizing them
in a tree structure recursively. The innermost level of nesting (recursion) will be
characterized by having a tree in which no more than b peers share the same
direct ancestor, where b is a double-exponentially power of two (e.g. 2,4,16,...).
Thus, multiple independent trees are imposed on the collection of peers. Figure 1
illustrates a simple example, where b = 2.
**CI�**
.�
**LSI�**
Fig. 1. The LRT structure for b=2
The degree of the peers at level i > 0 is d(i) = t(i), where t(i) indicates the
number of peers at level i. It holds that d(0)=b and t(0)=1. Let n be w-bit keys.
Each peer with label i (where 1 i N ) stores ordered keys that belong in the
≤ ≤
range [(i 1) ln n, i ln n–1], where N = n/lnn is the number of peers. Note here
−
that the lnn (and not logn) factor is due to a specific combinatorial game ([16])
we invoke in the next subsection.
We also equip each peer with a table named Left Spine Index (LSI), which
stores pointers to the peers of the left-most spine (see pointers starting from
peer 5).
-----
6 S. Sioutas et al.
Furthermore, each peer of the left-most spine is equipped with a table named
Collection Index (CI), which stores pointers to the collections of peers presented
at the same level (see pointers directed to collections of last level). Peers having
the same father belong to the same collection. For example, in Figure 1, peers
8, 9, 10, and 11 constitute a certain collection.
Lookup Algorithm Assume we are located at peer s (we mean the peer labeled
by integer number s) and seek a key k. First, we find the range where k belongs
in. Let say k [(j 1) ln n, j ln n 1]. The latter means that we have to search
∈ − −
for peer j. The first step of our algorithm is to find the LRT level where the
desired peer j is located. For this purpose, we exploit a nice arithmetic property
of LRT. This property says that for each peer x located at the left-most spine
of level i, the following formula holds:
label(x) = label(father(x)) + b[2][i][−][2] (1)
For example, peer 4 is located at level 2, thus 4 = father(4) + 2 or peer 8 is
located at level 3, thus 8 = father(8)+4 or peer 24 (not depicted in the Figure 1)
is located at level 4, thus 24 = father(24) + 16. The last equation is true since
father(24) = 8.
Thus, for each level i (in the next subsection we will prove that 0 i
≤ ≤
log log N ), we compute the label x of its left most peer by applying Equation (1).
Then, we compare the label j with the computed label x. If j x, we continue
≥
by applying Equation (1), otherwise we stop the loop process with current value
i. The latter means that peer j is located at the i-th level. So, first we follow the
i-th pointer of the LSI table located at peer s so as to reach the leftmost peer
x of level i. Then, we compute the collection in which the peer j belongs. Since
the number of collections at level i equals the number of peers located at level
(i 1), we divide the distance between j and x by the factor t(i 1). Let m (in
− −
� j−x+1 �
particular m = t(i−1) ) be the result of this division. The latter means that we
have to follow the (m + 1)-th pointer of the CI table so as to reach the desired
collection. Since the collection indicated by the CI[m+1] pointer is organized in
the same way at the next nesting level, we continue this process recursively.
Analysis The degree of the peers at level i > 0 is d(i) = t(i), where t(i)
indicates the number of peers at level i. It is defined that d(0)=b and t(0)=1.
It is apparent that t(i) = t(i 1)d(i 1), and, thus, by putting together the
− −
various components, we can solve the recurrence and obtain d(i) = t(i) = b[2][i][−][1]
for i 1. This double exponentially increasing fanout guarantees the following
≥
lemma:
Lemma 1: The height (or the number of levels) of LRT is O(log logb N ) in the
worst case.
The size of the LSI table equals the number of levels of LRT. Moreover, the
maximum size of the CI table appears at last level. It is apparent from the
building of the LRT structure that at last level h, t(h) = O(N ). It holds that
t(h) = b[2][h][−][1], thus b[2][h][−][1] = O(N ) or h−1 = O(loglogbN ) or h = O(loglogbN )+1.
Since the number of collections at level h equals the number of peers located
-----
Autonomous Range Tree 7
at level (h 1) we take t(h 1) = b[2][h][−][2] = b[2][(][O][(][loglogb N] [)+1)][−][2] or b[2][O][(][loglogb N] [)][−][1] =
− −
�
b[2][O][(][loglogbN] [)][2][−][1] = b[2][O][(][loglogb N] [)][�][1][/][2] and the lemma 2 follows:
√
Lemma 2: The maximum size of the CI and LSI tables is O( N ) and O(log log N )
in worst-case respectively.
We need now to determine what will be the maximum number of nesting
trees that can occur for N peers. Observe that the maximum number of peers
with the same direct ancestor is d(h 1). Would it be possible for a second level
−
tree to have the same (or bigger) depth than the outermost one?
This would imply that [�][h]j=0[−][1] [t][(][j][)][ < d][(][h][ −] [1).]
As otherwise we would be able to fit all the d(h 1) peers within the first h 1
− −
levels. But we need to remember that d(i) = t(i), thus d(h − 1) + j=0 [d][(][j][)][ <]
[�][h][−][2]
d(h 1).
−
This would imply that the number of peers in the first h 2 levels is negative,
−
clearly impossible. Thus, the second level tree will have depth strictly lower than
the depth of the outermost tree.
The innermost (let say j[th]) level of nesting (recursion) is characterized by
having a tree in which no more than b nodes share the same direct ancestor, where
b is a double-exponentially power of two (e.g. 2,4,16,...). In this case b = N [1][/b][j]
and the lemma 3 follows:
Lemma 3: The maximum number of possible nestings in LRT structure is
O(logb log N ) in the worst case.
At each peer we pay an extra processing cost by repeating the equation (1)
O(log log N ) times at most in order to locate the desired LSI pointer. Then,
we need O(1) hops for locating the left-most peer x of the desirable level. We
must note here that the processing overhead compared to communication overhead is negligible, thus we can ignore the O(log log N ) processing factor at each
peer. Finally we need O(1) hops for locating the desirable collection of peers via
the CI[m+1] pointer. Since, the collection indicated by the CI[m+1] pointer is
organized in the same way at a next nesting level, we continue the above process recursively. According to lemma 2 the maximum number of nesting levels
is O(logb log N ), and the theorem follows:
Theorem 1: Exact-match queries in the LRT structure require O(logb log N )
hops or lookup messages in the worst case.
3.2 Building ART Structure
We define as cluster peer a bucket of Θ(polylog N [′]) ordered peers, where N [′] is
the number of cluster peers.
At initialization step we choose as cluster peers the 1st peer, the (ln n +1)-th
peer, the (2 ln n + 1)-th peer and so on. This means that each cluster peer with
label i[′] (where 1 i[′] N [′]) stores ordered peers with sorted keys belonging in
≤ ≤
the range [(i[′] 1) ln[2] n, . . ., i[′] ln[2] n 1], where N [′] = n/ ln[2] n is the number of
− −
cluster peers.
ART stores cluster peers only, each of which is structured as an independent
decentralized architecture. The backbone-structure of ART is exactly the same
with LRT (see Figure 2). Moreover, instead of the Left-most Spine Index (LSI),
which reduces the robustness of the whole system, we introduce the Random
-----
8 S. Sioutas et al.
**RSI�**
**Cluster_Peer 1�**
**1�**
**RSI�** **RSI�**
**2-level LRT�** **2�** **3�** **Decentralized Architecture of�**
**Peer_Node�1�,Peer_Node �2�,......,Peer_Node �lnn�**
**RSI�** **RSI�** **RSI�** **RSI�**
**4�.�** **5�.�** **6�** **7�.�**
**Cluster_Peer i�**
**8�** **9�** **10�** **11�**
**12�** **13�** **14�.�** **15�** **i�**
**8�**
**12�**
**9�** **10�**
**13�** **14�** **i�**
**11�**
**15�**
**Decentralized Architecture of�**
**Peer_Node�(i-1)lnn+1�Peer_Node�(i-1)lnn+2�**
**,......,Peer_Node�ilnn�**
Fig. 2. The ART structure for b=2
Spine Index (RSI) routing table, which stores pointers to randomly chosen (and
not to left-most) cluster peers (see in Figure 2 the pointers starting from peer 3).
In addition, instead of using fat CI tables, we access the appropriate collection
of cluster peers by using a 2-level LRT structure. The 2-level LRT is an LRT
structure over log[2][c] Z buckets each of which organizes logZ[2][c] Z [collections in a]
LRT manner, where Z is the number of collections at current level and c is a big
positive constant (see Figure 3)
Load Balancing We model the join/leave of peers inside a cluster peer as the
combinatorial game of bins and balls presented in [16] and the lemma 4 follows:
Lemma 4: Given a µ( ) random sequence of join/leave peer operations, the
load of each cluster peer never becomes zero and never exceeds Θ(polylog N [′])
size in expected w.h.p. case.
Routing Overhead ART stores cluster peers, each of which is structured as
an independent decentralized architecture (be it BATON*, Chord, Skip-Graph,
e.t.c.) (see Figure 2). Here, we will try to avoid the existence of CI routing tables,
√
since these tables may become very large (O( N )) in the worst case as well as
the occurrence of local hot spots in the left-most spine results in a less robust
decentralized infrastructure. Thus, instead of the Left-most Spine Index (LSI),
we introduce the Random Spine Index (RSI) routing table. The latter table
stores pointers to the cluster peers of a random spine (for example, in Figure 2
the randomly chosen cluster peers 1, 2, 6 and 10 are pointed to by the RSI table of
cluster peer 3). Furthermore, instead of CI tables, we can access the appropriate
-----
Autonomous Range Tree 9
**2nd level�** **LRT structure of�Buckets�**
**Bucket�1�** **Bucket�polylogz�**
**1st level�** **in LRT manner�** **in LRT manner�**
**C�1�** **C�z/polylogz�** **C�(�z+1-z/polylogz)�** **C�z�**
**C�i� denotes the i-th collection�**
Fig. 3. The 2-level LRT structure
collection of cluster peers by using the 2-level LRT structure discussed above (see
Figure 3). Since the larger number of collections is Z = O(N [1][/][2]) (it appears in
the last level), the overhead of routing information is dominated by the second
� Z
level structures in each of which we have an O( log[2][c] Z [) =][ O][(][N][ 1][/][4][/][ log][c][ N] [)]
routing overhead. Thus, Theorem 2 follows:
Theorem 2: The overhead of routing information in ART is O(N [1][/][4]/ log[c] N )
in the worst case.
Remark 1: If we use a k-level LRT structure, the routing information overhead
becomes O(N [1][/][2][k] / log[c] N ) in the worst case.
Lookup Algorithms Let us explain the lookup operations in ART. For example, in Figure 4 suppose we are located at cluster peer 3 and we are looking
for two keys, which are located at cluster peers 19 and 119 respectively. The
first step of our algorithm is to find the levels of the ART where the desired
cluster peers (e.g. 19 and 119) are located. In our example, the fourth and fifth
levels are the desired levels. By following the RSI[4] and RSI[5] pointers we reach
the cluster peers 10 and 87 respectively. Now, we are starting from peers 10 and
87 to lookup the peers 19 and 119 respectively in the 2-level LRT structures of
the collections in respective levels.
Generally speaking, since the maximum number of nesting levels is O(logb log N )
and at each nesting level i we have to apply the standard LRT structure in N [1][/][2][i]
collections, the whole searching process requires T1(N ) hops or lookup messages
to locate the target cluster peer, where:
T1(N ) =
logb log N logb log N
� logb log(N [1][/][2][i]) = logb( � log(N [1][/][2][i] )) (2)
i=0 i=0
where
logb log N
� log(N [1][/][2][i] ) < (log N )[log][b][ log][ N]
i=0
from which we get:
T1(N ) < logb((log N )[log][b][ log][ N] ) = O(log[2]b [log][ N] [)]
-----
10 S. Sioutas et al.
Then, we have to locate the target peer by searching the respective decentralized structure, requiring T2(N ) hops. Since each of the known decentralized
architectures requires a logarithmic number of hops, the total process requires
T (N ) = T1(N ) + T2(N ) = O(log[2]b [log][ N] [) hops or lookup messages and the theo-]
rem follows.
Theorem 3: Exact-match queries in the ART structure require O(log[2]b [log][ N] [)]
hops or lookup messages.
Having located the target peer for key kℓ and exploiting the order of keys
on each node, range queries of the form [kℓ, kr] require an O(log[2]b [log][ N][ +][ |][A][|][)]
complexity, where A is the number of node-peers between the peers responsible
| |
for kℓ, kr respectively. The theorem follows.
Theorem 4: Range queries of the form [kℓ, kr] in the ART structure require an
O(log[2]b [log][ N][ +][ |][A][|][) complexity, where][ |][A][|][ is the answer size.]
**RSI�**
**1�**
**RSI�** **RSI�**
**2�** **3�.�**
**4�.�** **RSI�** **5�.�** **RSI�** **6�** **RSI�** **7�.�** **RSI�** **2-level LRT�**
8� 9� 10� 11� 12� 13� 14�.� 15� 16� 19� 20� 23�
24�.� 39�.� 72� **87�** **88�.�** 103�.� 104�.� **119�.�**
**2-level LRT�**
Fig. 4. An example of Lookup Steps via RSI[ ] tables and 2-level LRT structures
Query Processing, Data Insertion and Data Deletion, Peer Join and
Peer Departure In the following we briefly present the basic routines for query
processing, data insertion and data deletion, peer join and peer departure.
The Range Search(s, kℓ, kr) routine (Algorithm 1) gets as input the peer s
in which the query is initiated and the respective range of keys [kℓ, kr] and
returns as output the id of the cluster peer S, which contains peer s as well
as the cluster peer W in which the key kℓ belongs. Then, it calls the basic
ART Lookup(T, S, idS, W, idW ) routine, in order to locate the target peer responsible for key kℓ, and then, exploiting the order of keys on each peer performs
-----
Autonomous Range Tree 11
Algorithm 1 Range Search(s,kℓ,kr,A)
1: Input: s, kℓ, kr (we are at peer s and we are looking for keys in range [kℓ, kr])
2: Output: idW (the identifier of cluster-peer W, which stores kℓ key), A (the answer)
3: BEGIN
4: We compute idS:the identifier of Cluster peer S, which contains peer s;
5: We compute idW :let j be the identifier of target Cluster peer W, which stores kℓ
key;
6: Let T the basic ART structure of cluster-peers;
7: W=ART Lookup(T, S, idS, W, idW ); {call of the basic routine}
8: A=Linear Scan of all Cluster peers located in and right to W until we find a
key > kr;
9: END
a right linear scan till it finds a key > kr.
The ART Lookup(T, S, idS, W, idW ) routine (Algorithm 2) gets as input the
cluster peer S (with identifier idS) in which the query is initiated and returns
as output the id (idW ) of the cluster peer W in which the key kℓ belongs. T
denotes the ART-tree structure. Moreover, Algorithm 2 requires O(log[2]b [log][ N] [)]
hops, according to first part (T1(N )) of Theorem 3. Obviously, the same complexity holds for insert/delete key operations (see Algorithms 3 and 4), since we
have to locate the target peer into which the key must be inserted or deleted.
For join (depart) peer operations (for details see Algorithm 5), we need
O(log[2]b [log][ N] [) +][ T][join][(][N] [) (][O][(log]b[2] [log][ N] [) +][ T][depart][(][N] [)) lookup messages, where]
Tjoin(N ) (Tdepart(N )) is the number of hops required from the respective decentralized structure for peer-join (peer-departure).
In the peer join algorithm we assumed that the new peer is accompanied by
a key, and this key designates the exact position in which the new peer must be
inserted. If an empty peer u makes a join request at a particular peer v (which
we call entrance peer) then there is no need to get to a different cluster peer than
the one in which u belongs. Similarly, the algorithm for the departure of a peer
u assumes that the request for departure of peer u can be made from any peer
in the ART-structure. This may not be desirable, and in many applications it
is assumed that the choice for departure of peer u can be made only from this
peer. Of course, in this way the algorithm for peer departure is simplified since
there is no need to traverse the ART structure but only the cluster peer in which
u belongs. In order to bound the size of each cluster peer we assume that the
probability of picking an entrance peer is equal among all existing peers, and
that the probability of a peer departing is equal among all existing peers in the
ART. Since the size of the cluster peer is bounded by polylogN expected w.h.p.,
the following theorem is established:
Theorem 5: The peer join/departure can be carried out in O(loglogN ) hops
or lookup messages.
Node Failure, Fault Tolerance, Network Restructuring and Load Balancing Since we have modeled the join/leave of peers inside a cluster peer as
the combinatorial game of bins and balls presented in [16], each cluster peer of an
ART structure (according to lemma 4) never exceeds a polylogarithmic number
-----
12 S. Sioutas et al.
Algorithm 2 ART Lookup(T, S, idS, W, idW )
1: Input: We are at cluster-peer S with identifier idS
2: Output: We are looking for the cluster-peer W with identifier idW
3: BEGIN
4: If (S is responsible for kℓ)
5: Return S;
6: Else
7: If W =1 then i=0;
8: Else if W ∈{2, 3, . . ., b + 1] then i=1;
9: Else
10: x=b+2;
11: For (i = 2; i < c1log logb N ; + + i)
12: x = father(x) + b[2][i][−][2] ;
13: If j < x then break( );
14: Follow the RSI[i] pointer of cluster peer S;
15: Let X the correspondent cluster peer;
16: Search for W the 2-level LRT structure starting from X;
17: Let Y the first cluster-peer of the correspondent collection;
18: Let T [′] the ART structure of the collection above at next level of nesting with root
the cluster-peer Y ;
19: S = Y ;
20: ART Lookup(T [′], S, idS, W, idW ); {recursive call of the basic routine}
21: Return W ;
22: END
Algorithm 3 ART insert(T, s, k)
1: Input: We are at peer s and we want to insert the key k
2: Output: The peer w in which k must be inserted
3: BEGIN
4: We compute idS:the identifier of Cluster peer S, which contains peer s;
5: We compute idW :let j be the identifier of target Cluster peer W, which stores the
k key;
6: ART Lookup(T, S, idS, W, idW );
7: Let W the target cluster peer;
8: Search W for peer w containing k;
9: If k does not exist into w, then insert k into it;
10: END
Algorithm 4 ART delete(T, s, k)
1: Input: We are at peer s and we want to delete the key k
2: Output: The peer w in which k must be deleted
3: BEGIN
4: We compute idS:the identifier of Cluster peer S, which contains peer s;
5: We compute idW :let j be the identifier of target Cluster peer W, which stores the
k key;
6: ART Lookup(T, S, idS, W, idW );
7: Let W the target cluster peer;
8: Search W for peer w containing k;
9: If k exists into w, then delete it;
10: END
-----
Autonomous Range Tree 13
Algorithm 5 ART join/leave peer(T, s, w)
1: Input: We are at peer s and we want to insert/delete the new peer w
2: Output: The cluster peer W in which the peer w must be inserted/deleted
3: BEGIN
4: We compute idS:the identifier of Cluster peer S, which contains peer s;
5: We compute idW :let j be the identifier of target Cluster peer W, which contains
peer w;
6: ART Lookup(T, S, idS, W, idW ); {call of the basic routine}
7: Let W the target cluster peer;
8: Insert/delete w into/from W ;
9: END
of peers and never becomes empty in expected case with high probability. The
latter means that the skeleton ART structure of cluster peers remains unchanged
in the expected case with high probability as well as in each cluster peer the algorithms for peer failure, network restructuring and load balancing are according
to the polylogarithmic-sized decentralized architecture we use.
Multi-attribute Queries As in [14], we divide the whole range of attributes
into several sections: each section is used to index an attribute (if it appears
frequently in queries) or a group of attributes (if these attributes rarely appear
in queries). Since ART can only support queries over one-dimensional data, if we
index a group of attributes, we have to convert their values into one-dimensional
values (by choosing Hilbert space filling curve or other similar methods). For
example, if we have a system with 12 attributes: a1, a2, · · ·, a12 in which only
4 attributes from a1 to a4 are frequently queried (i.e. 90% of all queries), we
can build 4 separate indexes for them. The remaining attributes can be divided
equally into two groups to index, four attributes in each group. This way, the
number of replications can be significantly reduced from 12 down to 6.
### 4 Evaluation
For evaluation purposes we used the Distributed Java D-P2P-Sim simulator
presented in [27]. The D-P2P-Sim simulator is extremely efficient delivering
- 100, 000 cluster peers in a single computer system, using 32-bit JVM 1.6
and 1.5 GB RAM and full D-P2P-Sim GUI support. When 64-bit JVM 1.6 and
5 RAM is utilized the D-P2P-Sim simulator delivers > 500, 000 cluster peers
and full D-P2P-Sim GUI support in a single computer system. When D-P2PSim simulator acts in a distributed environment with multiple computer systems
with network connection delivers multiple times the former population of cluster
peers with only 10% overhead.
Our experimental performance studies include a detailed performance comparison with BATON*, one of the state-of-the-art decentralized architectures. In
particular, we implemented each cluster peer as a BATON* [14], the best known
decentralized tree-architecture. We tested the network with different numbers of
peers ranging up to 500,000. A number of data equal to the network size multiplied by 2000, which are numbers from the universe [1..1,000,000,000] are inserted
-----
14 S. Sioutas et al.
to the network in batches. The synthetic data (numbers) from this universe were
produced by the following distributions: beta, uniform and power-law. For each
test, 1,000 exact match queries and 1,000 range queries are executed, and the
average costs of operations are taken. Searched ranges are created randomly by
getting the whole range of values divided by the total number of peers multiplies
α, where α [1..10]. Note that in all experiments the default value of parameter
∈
b is 4. The source code of the whole evaluation process is publicly available [6].
4.1 Single- and Multi-attribute Query Performance
**Cost of exact match query�** **Cost of range query�**
8�
8� BATON* (fanout=10)�
6�
6� BATON* (fanout=10)� **Number of �** ART (normal, beta,�
**Number of routing �** 4� ART (normal, beta, uniform)� **routing hops�** 4� uniform)�
**hops�**
2� ART (powlow)� 2� ART (powlow)�
0� 0�
0� 100000�200000�300000�400000�500000�600000� 0� 200000� 400000� 600000�
**Number of nodes�** **Number of nodes�**
Fig. 5. Cost of exact match query (left) and cost of range query (right).
As proved previously, the whole query performance of ART is O(log[2]b [log][ N][ ′][)]
where the N [′] cluster peers structure their internal peers according to the BATON* architecture. For normal, beta and uniform distributions each cluster peer
contains 0.75 log[2] N peers on average and for power-law distributions each cluster peer contains 2.5 log[2] N peers on average. Thus, in the former case the average number of cluster peers is N [′] = 0.75 logN [2] N [, whereas in the latter case]
the number of cluster peers becomes N [′] = 2.5 logN [2] N [on average. In all cases,]
ART outperforms BATON* by a wide margin. As depicted in Figure 5 (up),
our method is almost 2 times faster and as a consequence we have a 50% improvement. The results are analogous with respect to the cost range queries as
depicted in Figure 5 (down).
Figure 6 (up) depicts the cost of updating routing tables. Since each cluster peer
structures O(N/polylog N ) (and not O(N )) peers according to BATON* architecture, the results are as expected. We remark that BATON* requires m logm N
hops, whereas m logm polylog N hops are required by ART. In particular and as
depicted in Figure 6 (up), our method updates the routing tables 3 or 4 times
faster. Figure 6 (down) depicts the insertion cost in multi-attribute case, where
we have 6 separate indexes. BATON* requires 6 log N hops and ART requires
6 log[2]b [log(][N/][polylog][ N] [) + 6 log(polylog][ N] [) hops. We observe that the insertion]
cost of ART is the lowest for any distribution. Again, our method is almost 2
times faster. Finally, the results are analogous for multi-attribute exact-match
and range queries respectively (see Figures 7 (up) and 7 (down)).
6 http://code.google.com/p/d-p2p-sim/
-----
Autonomous Range Tree 15
**Cost of updating routing tables�** **Cost of Insertion�**
40�
200� BATON* (fanout=10)�
30�
150� BATON* (fanout=10)� **Number of �** ART (normal, beta,�
**Number of �100�** **messages�** 20� uniform)�
**messages�** 50� ART (normal, beta,�uniform)� 10� ART (powlow)�
0� ART (powlow)� 0�
0� 200000� 400000� 600000� 0� 200000� 400000� 600000�
**Number of nodes�** **Number of nodes�**
Fig. 6. Cost of updating routing tables (left) and cost of insertion (right).
**Cost of exact match query�** **Cost of range query�**
8� 10�
BATON* (fanout=10)� BATON* (fanout=10)�
6� 8�
**Number of �** ART (normal, beta,� **Number of �** 6� ART (normal, beta,�
4�
**messages�** uniform)� **messages�** 4� uniform)�
2� ART (powlow)� 2� ART (powlow)�
0� 0�
0� 200000� 400000� 600000� 0� 200000� 400000� 600000�
**Number of nodes�** **Number of nodes�**
Fig. 7. Cost of multi-attribute exact-match (left) and range queries (right).
4.2 Load Balancing
ART not only reduces the search cost but also achieves better load balancing. To
verify this claim, we test the network with a variety of distributions and evaluate
the cost of load balancing. For simplicity, in our system, we assume that the query
distribution follows the data distribution. As a result, the workload of a peer is
determined only by the amount of data stored at that peer. In BATON*, when
a peer joins the network, it is assigned a default upper and lower load limit by
its parent. If the number of stored data at the peer exceeds the upper bound,
it is considered as an overloaded peer and vice versa. If a peer is overloaded
and cannot find a lightly loaded leaf peer, it is likely that all other peers also
have the same work load; thus, it automatically increases the boundaries of
storage capability. In ART the overlay of cluster peer remains unaffected in the
expected case with high probability when peers join or leave the network. Thus,
the load-balancing performance is restricted inside a cluster peer (which is a new
BATON* structure) and as a result ART needs no more than 4 lookup-messages
(instead of 1000 messages needed from BATON* in case of 500.000 nodes). For
details see Figure 8 (up).
4.3 Fault Tolerance
To evaluate the system’s fault tolerance in case of massive failure we initialized
the system with 10,000 peers. In the sequel, we let peers randomly fail step by
step without recovering. At each step, we check to see if the network is partitioned or not. With massive peer failures, we face a massive destruction of links
-----
16 S. Sioutas et al.
**Cost of load balancing�** **Search Cost�**
**in case of massive failure�**
1200�
1000� BATON* (fanout=10)� 200�
**Number of �** 800� ART (normal, beta,� 150� BATON* (fanout=10)�
**messages�** 600�400�200� uniform)�ART (powlow)� **Number of �messages�** 100�50� ART (normal, beta,�uniform)�
0� 0� ART (powlow)�
0� 200000� 400000� 600000� 0� 2000� 4000� 6000� 8000�
**Number of nodes�** **Number of failure nodes �**
Fig. 8. Cost of load-balancing (left) and search cost in case of massive failure (right).
connected to failed peers. Since the search process has to bypass these peers,
the search query has to be forwarded forth and back several times to find a way
to the destination and as a result the search cost is expected that will increase
substantially. Since the backbone of ART structure remains unaffected w.h.p.,
meaning that there is always a peer for playing the role of cluster representative,
the search cost is restricted inside a cluster peer (which is a BATON* structure) and as a result ART needs no more than 32 lookup-messages (instead of
180 messages needed from BATON* in case of 6.000 nodes). Figure 8 (down)
illustrates this effect.
### 5 Trade-offs and Heuristics
If each collection of cluster peers is organized individually as a BATON[∗] structure (not the whole level of collections), then we can climb up the ART structure
until we reach the nearest common ancestor of the cluster peer we are located
in, and the cluster peer we are searching. Then a downwards traversal is initiated to reach this cluster peer. Since, each collection of i[th]-level is organized
according to BATON*, we can decide in O(logm n[1][/][2][i]) hops the child we must
follow for further searching. As a result, the total time becomes O(logm n) and
no improvement has been achieved.
In our solution, if we parameterize the size of the buckets (depicted in Figure
3) from O(log[2][c] N ) to O(log[2][f] [(][N] [)] N ), where f (N ) is a function of the network
size, then we can get an interesting trade-off between the routing data overhead and the number of hops for an operation. In particular, if Z is the number of collections at the current level, then each bucket contains O( Z
log[2][f] [(][N] [)] N [)]
collections. Thus, the first LRT layer organizes O(log[2][f] [(][N] [)] N ) bucket representatives and each second LRT layer organizes O( Z
log[2][f] [(][N] [)] N [) collections. In this]
case, the routing overhead is dominated by the second layer LRTs which becomes O( N [1][/][4]
log[f] [(][N] [)] N [). To achieve an optimal routing data overhead we would like]
the following: O( N [1][/][4]
log[f] [(][N] [)] N [) =][ O][(1)][ ⇔] [f] [(][N] [) =][ O][(log][ N] [). In this case the first]
LRT layer contains O(log[2][f] [(][N] [)] N ) or O(log[2 log][ N] N ) bucket representative nodes.
Therefore, a lookup operation in first layer requires O(log log(log[2 log][ N] N )) or
ω(log log N ) hops. Each of the second layer LRTs contains O( Z
log[2][f] [(][N] [)] N [) collec-]
tion representative nodes, where Z is the number of collections at current level.
-----
Autonomous Range Tree 17
Therefore, the number of hops required by a lookup operation in second layer
is O(log log N ). So, the total time becomes ω(log log N ) and the sub-logarithmic
complexity is not guaranteed. As a result, if we want an optimal routing overhead we cannot guarantee sub-logarithmic complexity. If we relax the routing
overhead to be of polynomial size then we can achieve this.
In our solution the routing data overhead (O(N [1][/][4]/ log[c] N )) is a polynomial function. However, in reality even for an extremely large number of peers
N=1.000.000.000, the routing data overhead is 6 for c = 1, which is less than the
fanout of BATON[∗] (m = 10) that we used to run our experiments. The latter
demonstrates the significance of our result.
### 6 Conclusions
We presented a new efficient decentralized infrastructure for range query processing with probabilistic guarantees, the ART structure. Theoretical analysis
showed that the communication cost of query, update and join/leave node operations scale sub-logarithmically expected w.h.p.. Experimental performance
comparison with BATON*, the state-of-the-art decentralized structure, showed
the improved performance, scalability and efficiency of our new method. Finally,
we believe that ART will enable general purpose decentralized trees to support
a wider class of queries, and then broaden the horizon of their applicability.
### References
1. Andrzejak A. and Xu Z.: “ Scalable, Efficient Range Queries for Grid Information
Services”, Proceedings 2nd International Conference on Peer-To-Peer Computing
(P2P), pp.33-40, Linkoping, Sweden, 2002.
2. Aspnes J. and Shah G.: “Skip Graphs”, Proceedings 14th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pp.384-393, Baltimore, MD, 2003.
3. Bharambe A.R., Agrawal M. and Seshan S.: “Mercury: Supporting Scalable Multiattribute Range Queries”, Proceedings ACM Conference on Applications, Technologies, Architectures, and Protocols for Computer Communication (SIGCOMM),
pp.353-366, Portland, OR, 2004.
4. Cai M., Frank M., Chen J. and Szekely P.: “Maan: a Multi-attribute Addressable
Network for Grid Information Services”, Proceedings 4th International Workshop
on Grid Computing (GRID), pp.184-191, Phoenix, AZ, 2003.
5. Crainiceanu A., Linga P., Gehrke J. and Shanmugasundaram J.: “Querying Peerto-peer Networks Using P-Trees”, Proceedings 7th International Workshop on Web
and Databases (WebDB), pp.25-30, Paris, France, 2004.
6. A. Carzaniga, E. Nitto, D. Rosenblum, and A. L. Wolf. Issues in supporting eventbased architectural styles. In 3rd Intl Software Architecture Workshop, 1998.
7. Antonio Carzaniga, David S. Rosenblum, and Alexander L. Wolf. Design and evaluation of a wide-area event notification service. ACM Trans. Comput. Syst., 19(3):332383, 2001.
8. Gupta A., Agrawal D. and El Abbadi A.: “Approximate Range Selection Queries
in Peer-to-peer Systems”, Proceedings 1st Biennial Conference on Innovative Data
Systems Research, Asilomar, CA, 2003.
9. Indranil Gupta, Ken Birman, Prakash Linga, Al Demers AND Robbert van Renesse, ”Kelips: Building an efficient and stable P2P DHT through increased memory
-----
18 S. Sioutas et al.
and background overhead”, Proceedings of the 2nd International Workshop on Peerto-Peer Systems (IPTPS ’03), Berkeley, CA, USA, 2003.
10. Goodrich M.T., Nelson M.J. and Sun J.Z.: “The Rainbow Skip Graph: a FaultTolerant Constant-Degree Distributed Data Structure”, Proceedings 14th Annual
ACM-SIAM Symposium on Discrete Algorithms (SODA), pp.384-393, Miami, FL,
2006.
11. Jun Gao, Peter Steenkiste, ”An Adaptive Protocol for Efficient Support of Range
Queries in DHT-Based Systems,” 12th IEEE International Conference on Network
Protocols (ICNP’04), pp.239-250, Berlin, Germany, 2004.
12. Harvey N.J.A., Jones M.B., Saroiu S., Theimer M. and Wolman A.: “Skipnet: a
Scalable Overlay Network with Practical Locality Properties”, Proceedings USENIX
Symposium on Internet Technologies and Systems, Seattle, WA, 2003.
13. Jagadish H.V., Ooi B.C. and Vu Q.H.: “Baton: a Balanced Tree Structure for Peerto-peer Networks”, Proceedings 31st International Conference on Very Large Data
Bases (VLDB), pp.661-672, Trondheim, Norway, 2005.
14. Jagadish H.V., Ooi B.C., Tan K.L., Vu Q.H. and Zhang R.: “Speeding up Search
in P2P Networks with a Multi-way Tree Structure”, Proceedings ACM International
Conference on Management of Data (SIGMOD), pp.1-12, Chicago, IL, 2006.
15. Karger D., Kaashoek F., Stoica I., Morris R. and Balakrishnan H.: “Chord: a
Scalable Peer-to-peer Lookup Service for Internet Applications”, Proceedings ACM
Conference on Applications, Technologies, Architectures, and Protocols for Computer Communication (SIGCOMM), pp.149-160, San Diego, CA, 2001.
16. Kaporis A., Makris Ch., Sioutas S., Tsakalidis A., Tsichlas K. and Zaroliagis Ch.:
“Improved Bounds for Finger Search on a RAM”, Proceedings 11th Annual European
Symposium on Algorithms (ESA), pp.325-336, Budapest, Hungary, 2003.
17. Kaporis A., Makris Ch., Sioutas S., Tsakalidis A., Tsichlas K. and Zaroliagis Ch.:
“Dynamic Interpolation Search Revisited”, Proceedings 33rd International Colloquium Automata, Languages and Programming (ICALP), Part I, pp.382-394, Venice,
Italy, 2006.
18. Li X., Kim Y.J., Govindan R. and Hong W.: “Multi-dimensional Range Queries
in Sensor Networks”, Proceedings 1st International Conference on Embedded Networked Sensor Systems (SenSys), pp.63-75, Los Angeles, CA, 2003.
19. Liau C.Y., Ng W.S., Shu Y., Tan K.L. and Bressan S.: “Efficient Range Queries
and Fast Lookup Services for Scalable P2P Networks”, Proceedings 2nd International Workshop on Databases, Information Systems, and Peer-to-Peer Computing(DBISP2P), pp.93-106, Toronto, Canada, 2004.
20. Maymounkov P. and Mazieres D.: “Kademlia: a Peer-to-peer Information System
Based on the XOR Metric”, Proceedings 1st International Workshop on Peer-to-Peer
Systems (IPTPS), pp.53-65, Cambridge, MA, 2002.
21. S. Naicken, B. Livingston, A. Basu, S. Rodhetbhai, I. Wakeman, and D. Chalmers,
The state of peer-to-peer simulators and simulations, SIGCOMM Comput. Commun. Rev., 2007 Vol 37, No 2, pp. 95-98
22. Ratnasamy S., Francis P., Handley M., Karp R. Shenker S.: “A Scalable Content
addressable Network”, Proceedings ACM Conference on Applications, Technologies,
Architectures, and Protocols for Computer Communication (SIGCOMM), pp.161172, San Diego, CA, 2001.
23. Rowstron A. and Druschel P.: “Pastry: a Scalable, Decentralized Object Location,
and Routing for Large-scale Peer-to-peer Systems”, Proceedings IFIP/ACM International Conference on Distributed Systems Platforms (MIDDLEWARE), pp.329-350,
Heidelberg, Germany, 2001.
24. Sriram Ramabhadran, Sylvia Ratnasamy, Joseph M. Hellerstein, Scott Shenker:
”prefix hash tree”, Proceedings of the twenty-third annual ACM symposium on
-----
Autonomous Range Tree 19
Principles of distributed computing table of contents (Brief announcement), PP.368368, Newfoundland, Canada, 2004.
25. Sahin O.D., Gupta A., Agrawal D. and El Abbadi A.: “A Peer-to-peer Framework
for Caching Range Queries”, Proceedings 20th International Conference on Data
Engineering (ICDE), pp.165-176, Boston, MA, 2004.
26. Stoica I., Morris R., Liben-Nowell D., Karger D.R., Kaashoek M.F., Dabek F. and
Balakrishnan H.: “Chord: a Scalable Peer-to-peer Lookup Protocol for Internet Applications”, IEEE/ACM Transactions on Networking, Vol.11, No.1, pp.17-32, 2003.
27. S. Sioutas, G. Papaloukopoulos, E. Sakkopoulos, K. Tsichlas and Y. Manolopoulos
”A novel Distributed P2P Simulator Architecture: D-P2P-Sim”, ACM CIKM 2009,
pp. 2069-2070, 2009.
28. S. Sioutas, G. Papaloukopoulos, E. Sakkopoulos, K. Tsichlas and Y. Manolopoulos
”Brief announcement: ART–sub-logarithmic decentralized range query processing
with probabilistic guarantees”, ACM PODC 2010, pp. 118-119, 2010.
29. Triantafillou P. and Pitoura T.: “Towards a Unifying Framework for Complex
Query Processing over Structured Peer-to-Peer Data Networks”, VLDB 03 Workshop on Databases, Information Systems, and Peer-to-Peer Computing, 2003.
30. Zhang H., Goel A. and Govindan R.: “Incrementally Improving Lookup Latency
in Distributed Hash Table Systems”, SIGMETRICS, pp.114-125, San Diego, CA,
2003.
31. Zhao B.Y., Huang L., Stribling J., Rhea S.C., Joseph A.D. and Kubiatowicz J.D.:
“Tapestry: a Resilient Global-scale Overlay for Service Deployment”, IEEE JSAC,
Vol.22, No.1, pp.41-53, 2004.
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/1201.2766, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "GREEN",
"url": "https://arxiv.org/pdf/1201.2766"
}
| 2,010
|
[
"JournalArticle",
"Book"
] | true
| 2010-07-25T00:00:00
|
[
{
"paperId": "5d8384c21ab3a1107feec9c1d5748131d89bbda4",
"title": "ART: sub-logarithmic decentralized range query processing with probabilistic guarantees"
},
{
"paperId": "1a675bf983f18f6153d72deca0c06fb15ccda100",
"title": "A novel distributed P2P simulator architecture: D-P2P-sim"
},
{
"paperId": "23991c517cc13bcce1dd714830578a6e03646f33",
"title": "Join queries in P2P DHT Systems"
},
{
"paperId": "5f05bdecda952e862bab0231c8566b0104adf739",
"title": "The state of peer-to-peer simulators and simulations"
},
{
"paperId": "8683b86a7a640c26b0c68edbdadf1de45447a74e",
"title": "Dynamic Interpolation Search Revisited"
},
{
"paperId": "aa80c6ceb4ed86306c6c985c242dc6bac2fee60b",
"title": "Speeding up search in peer-to-peer networks with a multi-way tree structure"
},
{
"paperId": "5831f1722e9513c1b79e52b0d15cf8d35c11364d",
"title": "The rainbow skip graph: a fault-tolerant constant-degree distributed data structure"
},
{
"paperId": "d257d97ac43c5cd1e2a7643bfaf039cb25b34e01",
"title": "BATON: A Balanced Tree Structure for Peer-to-Peer Networks"
},
{
"paperId": "8b76104fa372c664cf5210d4dae1c59066b6b81f",
"title": "An adaptive protocol for efficient support of range queries in DHT-based systems"
},
{
"paperId": "4e29d549fb237c33f3035b46bdbd471da993f125",
"title": "Mercury: supporting scalable multi-attribute range queries"
},
{
"paperId": "d83253377c0ddc9e4aaa1181a825d15e98a5a3ce",
"title": "Efficient Range Queries and Fast Lookup Services for Scalable P2P Networks"
},
{
"paperId": "27521785235e785f9b48d8224e075b9037b4f1af",
"title": "Brief announcement: prefix hash tree"
},
{
"paperId": "f2c4786567ebb91169753e71c015898e14936e48",
"title": "Querying peer-to-peer networks using P-trees"
},
{
"paperId": "971ca0fce11064aa3808c0e78b6bff22c1b6e0f4",
"title": "A peer-to-peer framework for caching range queries"
},
{
"paperId": "b851141730342a9613adc38e92a38fd1cb42f623",
"title": "Tapestry: a resilient global-scale overlay for service deployment"
},
{
"paperId": "c81c79b64f2b185a330132586a6fcf8c45f8a9a6",
"title": "MAAN: A Multi-Attribute Addressable Network for Grid Information Services"
},
{
"paperId": "0c3e07c5ef9a4abb64ac152e64efaca63b0fd788",
"title": "Multi-dimensional range queries in sensor networks"
},
{
"paperId": "1a7a9fd1a17db14ce591cc633ae3caef50a7abcd",
"title": "Improved Bounds for Finger Search on a RAM"
},
{
"paperId": "91ca8c26f621a3fc04ed45f68ade474d43b6e3fd",
"title": "Towards a Unifying Framework for Complex Query Processing over Structured Peer-to-Peer Data Networks"
},
{
"paperId": "1571ee0f19ff9931490c9e6720f45345a268f7cf",
"title": "Incrementally improving lookup latency in distributed hash table systems"
},
{
"paperId": "684d6cc62cee0e5ef9dc988691848104e4a5c21e",
"title": "Kelips: Building an Efficient and Stable P2P DHT through Increased Memory and Background Overhead"
},
{
"paperId": "2ebe1ffec53e63c2799cba961503f0a6abafccd3",
"title": "Skip graphs"
},
{
"paperId": "dbd62aed6beede2d157640fec3d1ec24302c0360",
"title": "Scalable, efficient range queries for grid information services"
},
{
"paperId": "eb51cb223fb17995085af86ac70f765077720504",
"title": "Kademlia: A Peer-to-Peer Information System Based on the XOR Metric"
},
{
"paperId": "cf025469b2d7e4b37c7f2d2bf0d46c6776f48fd4",
"title": "Pastry: Scalable, Decentralized Object Location, and Routing for Large-Scale Peer-to-Peer Systems"
},
{
"paperId": "680ba806b8a651e8cb2e2d64a9d6bc2325a8eea1",
"title": "A scalable content-addressable network"
},
{
"paperId": "f03db79dc2922af3ec712592c8a3f69182ec5d65",
"title": "Chord: A scalable peer-to-peer lookup service for internet applications"
},
{
"paperId": "30e1983f9b8227760452c714ee6ad62a913d6236",
"title": "Design and evaluation of a wide-area event notification service"
},
{
"paperId": "a65586b3b7c7dbf5f620444ecc8962c563b7d7f7",
"title": "Issues in supporting event-based architectural styles"
},
{
"paperId": "8dbcb2ffbd99b7255eaa32abc615cb0b5275469b",
"title": "A scalable peer-to-peer lookup protocol for Internet applications"
},
{
"paperId": "44e4c963e379448ab4a04b172b9ec659ed08096b",
"title": "A Review of Dimension Reduction Techniques"
},
{
"paperId": "f0574a77de9b6be9b3bd6d75dd4f21e2ce1113cb",
"title": "P2p Networking And Applications"
},
{
"paperId": "dc9ced0da655ff64a60ba03bf31d0eebf7e52a03",
"title": "Querying the Internet with PIER"
},
{
"paperId": "b678e1f35d01bd3cf2c9aa02b359f1f177547b4d",
"title": "Approximate Range Selection Queries in Peer-to-Peer Systems"
},
{
"paperId": "8d60ae4c2df409e4ab1d7e518b39e4d91dc6c6a7",
"title": "Proceedings of Usits '03: 4th Usenix Symposium on Internet Technologies and Systems Symphony: Distributed Hashing in a Small World"
},
{
"paperId": null,
"title": "Proceedings of the 2nd International Workshop on Peerto-Peer Systems (IPTPS '03)"
},
{
"paperId": null,
"title": "prefix hash tree”, Proceedings of the twenty-third annual ACM symposium on Autonomous Range Tree 19 Principles of distributed computing table of contents (Brief announcement), PP.368368"
}
] | 14,699
|
en
|
[
{
"category": "Environmental Science",
"source": "s2-fos-model"
},
{
"category": "Economics",
"source": "s2-fos-model"
},
{
"category": "Business",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/02a3f37b0735b61801d2b28babacf7aec3ff83b6
|
[] | 0.862854
|
Blockchain Technology as a Catalyst for Sustainable Development: Exploring Economic, Social, and Environmental Synergies
|
02a3f37b0735b61801d2b28babacf7aec3ff83b6
|
Academic Journal of Interdisciplinary Studies
|
[
{
"authorId": "2151832242",
"name": "Marsela Thanasi Boçe"
},
{
"authorId": "2290624379",
"name": "Julian Hoxha"
}
] |
{
"alternate_issns": [
"2281-3993"
],
"alternate_names": [
"Acad J Interdiscip Stud"
],
"alternate_urls": [
"https://www.mcser.org/journal/index.php/ajis/article/view/1450"
],
"id": "6c2c7db3-50d7-4c44-9697-2b6a1224d38b",
"issn": "2281-4612",
"name": "Academic Journal of Interdisciplinary Studies",
"type": "journal",
"url": "https://www.mcser.org/journal/index.php/ajis"
}
|
This paper explores the transformative potential of blockchain technology (BT) as a catalyst for sustainable development, addressing the tri-fold aspects of environmental, economic, and social sustainability. Through a comprehensive review and theoretical framework, it delves into how BT can significantly contribute to achieving the United Nations Sustainable Development Goals (SDGs). The study highlights BT's role in enhancing transparency, ensuring product traceability, and promoting resource efficiency, thereby facilitating a more equitable economic growth and environmental stewardship. By examining various applications of BT across industries including supply chain management, renewable energy, and conservation efforts, the paper illustrates BT's capability to reduce carbon emissions, improve resource allocation, and support sustainable business practices. Furthermore, it identifies challenges such as scalability, energy consumption, and regulatory hurdles, proposing strategic recommendations for overcoming these obstacles. The research emphasizes the need for collaborative efforts among stakeholders, including policymakers, practitioners, and researchers, to leverage BT effectively for sustainable development. It contributes to both theoretical understanding and practical implementation of blockchain as a powerful enabler for sustainability, offering insights for future research and policy-making in this evolving domain.
Received: 1 December 2023 / Accepted: 19 February 2024 / Published: 5 March 2024
|
.
**_ISSN 2281-3993_** **_www.richtmann.org_** **_March 2024_**
**Research Article**
© 2024 Marsela Thanasi Boçe and Julian Hoxha.
This is an open access article licensed under the Creative Commons
Attribution-NonCommercial 4.0 International License
(https://creativecommons.org/licenses/by-nc/4.0/)
Received: 1 December 2023 / Accepted: 19 February 2024 / Published: 5 March 2024
# Blockchain Technology as a Catalyst for Sustainable Development: Exploring Economic, Social, and Environmental Synergies
**Marsela Thanasi Boçe[1]**
**Julian Hoxha[2]**
_1College of Business Administration,_
_American University of the Middle East,_
_Kuwait_
_2College of Engineering,_
_American University of the Middle East,_
_Kuwait_
**DOI: https://doi.org/10.36941/ajis-2024-0041**
**_Abstract_**
_This paper explores the transformative potential of blockchain technology (BT) as a catalyst for sustainable_
_development, addressing the tri-fold aspects of environmental, economic, and social sustainability. Through_
_a comprehensive review and theoretical framework, it delves into how BT can significantly contribute to_
_achieving the United Nations Sustainable Development Goals (SDGs). The study highlights BT's role in_
_enhancing transparency, ensuring product traceability, and promoting resource efficiency, thereby_
_facilitating a more equitable economic growth and environmental stewardship. By examining various_
_applications of BT across industries including supply chain management, renewable energy, and_
_conservation efforts, the paper illustrates BT's capability to reduce carbon emissions, improve resource_
_allocation, and support sustainable business practices. Furthermore, it identifies challenges such as_
_scalability, energy consumption, and regulatory hurdles, proposing strategic recommendations for_
_overcoming these obstacles. The research emphasizes the need for collaborative efforts among stakeholders,_
_including policymakers, practitioners, and researchers, to leverage BT effectively for sustainable_
_development. It contributes to both theoretical understanding and practical implementation of blockchain as_
_a powerful enabler for sustainability, offering insights for future research and policy-making in this evolving_
_domain._
**_Keywords: Blockchain technology (BT), Environmental, Economic, Social, Sustainability_**
**1.** **Introduction**
Sustainable development, a pressing global challenge, has gained significant attention from
researchers and policymakers (Liu et al., 2011). In the business context, sustainability is increasingly
crucial for survival due to rising regulatory pressures and evolving production practices (Hahn and
Figge, 2018). Sustainability, as defined by Clark (2007), involves responsible behavior towards the
environment, society, and future generations. Corporate sustainability is viewed as meeting the needs
of various stakeholders, shareholders, employees, customers, regulatory bodies, and society at large,
-----
**_ISSN 2281-3993_** **_www.richtmann.org_** **_March 2024_**
without compromising the ability of future stakeholders to meet their own needs (Mani et al., 2016).
Sustainable development efforts currently focus on balancing environmental, social, and economic
objectives to meet present needs without compromising future generations. Key goals include
tackling climate change, promoting clean energy, ensuring economic growth, and fostering social
inclusion. However, challenges remain significant.
Environmentally, the world faces urgent issues like climate change, loss of biodiversity, and
pollution (Luna et al., 2024; Dehshiri and Amiri, 2024; Arshad et al., 2023). Efforts to reduce carbon
emissions and increase renewable energy adoption are progressing, but not at the pace required to
meet international climate goals. Economically, sustainable development aims to reduce poverty and
inequality. While there has been progress, global economic disparities persist, exacerbated by factors
like insufficient infrastructure in developing countries and the economic impacts of the COVID-19
pandemic. Socially, challenges include ensuring equal access to quality education, healthcare, and
employment opportunities. There's also a need to strengthen social safety nets and promote gender
equality.
The modern world's need for sustainability is a complex subject that necessitates a thorough,
critical examination. It includes comprehending how problems like pollution, resource depletion, and
climate change are interrelated. This entails realizing how individuals’ actions affect the environment
and how crucial it is to lessen this harmful influence (Yang et al, 2023).
Economic aspects to consider sustainability can result in long-term economic gains, despite the
misconception that it impedes economic growth. This includes the emergence of new markets and
employment prospects, especially in renewable energy. Furthermore, social aspects of sustainability
must be considered including fair labor practices, equitable resource access, and considering future
generations' needs. Analyzing the impact of our actions and activities on the community and wider,
throughout the world, now and in the future is crucial (Shayan et al., 2022). Reaching sustainability
necessitates a substantial shift in both personal conduct and societal expectations. This is a difficult
task that requires knowledge of and control of the sociological and psychological elements that shape
human behavior.
In addition, technological innovation has a dual function: it both causes sustainability
(environmental) issues and provides answers to them. To comprehend the possibilities and
constraints of these technical solutions, critical analysis is required (Sarfraz et al., 2023; Dehshiri and
Amiri, 2024).
In an era where sustainability is not merely an option but a necessity, BT emerges as a pivotal
tool in harmonizing the trinity of social, economic, and environmental dimensions of sustainability
(Carter and Easton, 2011).
At the juncture of the Fourth Industrial Revolution, blockchain stands out as a transformative
force, capable of fostering a synergy between technological advancement and sustainable
development. Blockchain is a decentralized ledger that offers an immutable, secure, and distributed
database that can facilitate the verifiable exchange of information and assets, reducing the
dependency on intermediaries and enhancing peer-to-peer transactions (Yli-Huumo et al., 2016;
Galen et al., 2018) and significantly boost the efficacy of sustainable development endeavors by
promoting transparency and trust (Horner and Ryan, 2019).
This article seeks to inform and guide stakeholders, from policymakers to practitioners, on the
effective deployment of blockchain solutions for economic, social, and environmental sustainability.
The article's theoretical contributions lie in the creation of a conceptual framework to elucidate the
intricate relationship between blockchain and sustainability dimensions. Practically, it aims to
unearth best practices, discern obstacles to implementation, and recommend pathways to leverage
blockchain for a more sustainable future.
The upcoming sections will delve into the theoretical underpinnings of blockchain in
sustainability, outline the challenges and limitations of incorporating blockchain for sustainable
solutions, present a comprehensive framework that explicates the relationship between blockchain
and environment, economic, and social sustainability, and offer conclusions with strategic
-----
**_ISSN 2281-3993_** **_www.richtmann.org_** **_March 2024_**
recommendations for future research and policymaking.
**2.** **Sustainability Definition**
Scholars have put forth numerous definitions of sustainability encompassing various aspects. Prior to
1990, literature focused on three dimensions of sustainability and sustainable development: social
justice, environmental preservation, and economic prosperity (Purvis et al., 2019). Thus, sustainability
can be conceptualized as a triple bottom line comprising social responsibility, environmental
stewardship, and economic viability (Kapfere and Denizeau, 2017).
Subsequently, Gladwin et al. (1995), through content analysis of multiple definitions, identified
additional critical elements of sustainable development, such as inclusiveness, prudence,
connectivity, security, and equity. The most widely accepted definition of sustainability in scholarly
works is "meeting the needs of the present without compromising the ability of future generations to
meet their own needs" (Brundtland Report, 1987, p. 8).
Sustainability in the modern world is a complex issue, requiring a deep understanding of the
links between environmental concerns like pollution, resource depletion, and climate change, and
the importance of mitigating our environmental impact.
Economically, it offers long-term benefits and opportunities, particularly in renewable energy,
challenging the notion that it hinders growth. Socially, it involves ensuring equitable resource
distribution, fair labor, and considering future generations, requiring a global perspective on the
impact of our actions. Technologically, it poses both challenges and solutions, necessitating critical
analysis of innovation's potential and limitations (Sarfraz et al., 2023).
In essence, sustainability is about harmonizing environmental, economic, and social interests to
ensure the well-being of current and future generations, facilitated by technological innovation,
cultural shifts, effective policy, and a balanced global-local approach (Hariram, 2023).
_2.1_ _Fundamentals of blockchain: a comprehensive analysis of its functionality and benefits to_
_sustainability_
The term "blockchain" initially surfaced on the internet in 2008 and has exerted a substantial impact
on public institutions, private enterprises, and emerging businesses. The BT is primarily employed as
an innovative approach for facilitating transactions between two entities. It functions as a
decentralized and secure ledger, enabling direct trades between two anonymous individuals without
the requirement of a trusted intermediary. This technology introduces a new operating framework
for enterprises and institutions. According to Palacio (2018), it can serve as a viable tool for tackling
worldwide difficulties and facilitating the realization of the United Nations' sustainable development
goals (SDGs) across all countries.
_2.1.1_ _The salient sustainable characteristics of BT_
Several salient facets of BT provide advantages in terms of the advancement of sustainability.
- Transparency: The concept of transparency refers to the quality or state of being open,
honest, and accountable. The ledger system of the blockchain provides a transparent
account of transactions. The utilization of this function is of utmost importance in the
surveillance of sustainable sourcing practices of products, ensuring compliance with
environmental rules, and mitigating the likelihood of unethical operations within the supply
chain.
- Authenticity: The ability to track the trajectory of a product from its initial source
contributes to the verification of sustainable practices' authenticity. This holds particular
importance in industries such as agriculture, where it is important to closely observe the
movement of products from the manufacturer to the end consumer (Prashar, et al., 2020;
-----
**_ISSN 2281-3993_** **_www.richtmann.org_** **_March 2024_**
Bhusal, 2021).
- Decentralization: Through the utilization of a decentralized network, BT mitigates the
necessity for intermediaries or centralized governing bodies. This might potentially lead to
the development of more equitable and efficient systems, hence reducing carbon emissions
and improving resource allocation (Risso et al., 2023).
- The cryptographic nature of blockchain ensures the guarantee of data integrity and security.
Ensuring the protection of sensitive environmental data and maintaining the accuracy and
reliability of sustainability records is imperative.
- Smart contracts are agreements capable of executing themselves, as they have the
agreement's conditions explicitly encoded into their code. One potential strategy to
streamline operations and reduce administrative burden is the use of automated systems for
monitoring and implementing sustainability commitments and regulations (Dal Mas et al.,
2020).
- The implementation of BT can greatly reduce the need for excessive paperwork and manual
processing involved in monitoring and verifying sustainable practices. This technological
advancement can enhance operational efficiency and mitigate environmental consequences.
- Through the implementation of BT, resource management may be significantly enhanced by
facilitating accurate monitoring and projection capabilities. This, in turn, fosters waste
reduction and promotes the efficient utilization of resources (Parmentola et al., 2022).
- The provision of a shared and dependable platform for the exchange of information can
foster cooperation among many stakeholders involved in sustainability endeavors, including
corporations, governmental bodies, and individuals (Schulz et al., 2020).
Figure 1 shows the connection between the sustainability-related features of BT with sustainable
practices. Through these attributes, BT can significantly influence several sustainability-related
concerns, such as the conservation of resources, preservation of the environment, and promotion of
sustainable business practices.
**Figure 1: Sustainability-related features of BT**
**Source: Authors’ own work**
_2.1.2_ _Historical context: blockchain in sustainability-driven projects_
The historical backdrop of using BT into sustainability-focused initiatives can be traced back to its
inception within the financial industry, specifically with the advent of Bitcoin in 2009. However, the
recognition of BT's promise for sustainability occurred a few years later when other businesses and
-----
**_ISSN 2281-3993_** **_www.richtmann.org_** **_March 2024_**
organizations commenced exploring its wider range of uses.
The implementation of BT in supply chain management has emerged as an early and
noteworthy application within the realm of sustainability. Organizations have come to recognize that
the transparency and traceability attributes of BT may effectively guarantee ethical sourcing and
production procedures. One example of blockchain use in the food industry is IBM's Food Trust,
which was developed in partnership with Walmart. This platform utilizes BT to monitor the entire
supply chain of food goods, tracing their path from the farm to the retail shop. By doing so, it aims to
enhance food safety measures and minimize wastage. Another illustration may be found in
Everledger, a company that uses BT to track the provenance and life cycle of various items. Its
primary emphasis is on diamonds, with the aim of guaranteeing their ethical sourcing and absence of
involvement in conflicts (Everledger, 2023).
BT has gained traction within the energy sector, mostly to advance renewable energy initiatives
and enhance the efficiency of energy distribution systems. Power Ledger and WePower are examples
of companies that have utilized BT to facilitate peer-to-peer energy trading. This innovative approach
enables customers to engage in the buying and selling of surplus renewable energy, thereby fostering
the adoption of sustainable energy resources (Ahmad, 2023).
BT has also been utilized in the realm of environmental conservation. One example of a
platform that aims to modernize the energy sector is VAKT. This platform is designed to establish a
safe and unchangeable digital environment specifically for the processing of physical post-trade
activities in the energy industry. The use of this solution aids in the reduction of administrative
documentation and facilitates the establishment of an effective and environmentally conscious
energy trading system (Thoughtworks, 2023). An additional illustration may be found in the Aerial
platform, which leverages BT to establish a heightened level of transparency and accountability in the
monitoring of carbon emissions and credits (Aerial, 2023).
The concept of "green bonds" and sustainable investments has also experienced advantages
through the utilization of BT in the realm of sustainable finance. Platforms like as BanQu contribute
to the establishment of economic prospects for underprivileged communities by offering a digital
identity and financial record, which are essential for gaining access to financial services and engaging
in sustainable economic endeavors.
The utilization of BT by organizations such as Plastic Bank is being employed to address the
issue of plastic waste in marine environments, namely by providing incentives for recycling activities
in developing nations. Individuals have the capacity to gather plastic waste and subsequently trade it
for digital tokens, so facilitating the establishment of a circular economy and mitigating the adverse
effects of environmental pollution (Böckel et al., 2021).
To summarize, the historical backdrop of blockchain in the realm of sustainability is
characterized by its evolution from a singularly financial instrument to a multifaceted technological
solution that tackles diverse aspects of sustainability. The impetus for this change can be attributed
to the increasing recognition of environmental concerns and the imperative for transparent and
efficient solutions across various industries. Organizations worldwide have been actively engaging in
the exploration and implementation of BT to address sustainability objectives, thereby showcasing its
adaptability and capacity for fostering favorable ecological outcomes.
_2.2_ _Review of the relevant literature_
This section synthesizes the existing blockchain technology (BT) focused literature and elaborates on
the key themes listed in Table 1 from the literature review. Studies included in this section were
identified using the Scopus database, using the following combination of keywords (TITLE
(“Blockchain technology”) AND TITLE (“Sustainability”). This approach is similar to the approach
employed by existing review articles on various topics (Dwivedi, et al., 2023)
Existing research reviewed for this article is categorized into the following major themes:
Blockchain Technology, research in the environment, economic and social domains, and application.
-----
**_ISSN 2281-3993_** **_www.richtmann.org_** **_March 2024_**
**Table 1: Sustainability theme-based categorization of BT articles**
**Theme** **Sub-theme** **Description** **References**
Alhasan and Hamdan 2023; Cui, 2023; Ahmad, 2023; Zuo,
Renewable Energy Research related to renewable energy. 2022; Bai et al., 2022; Thakur et al., 2021; Howson, 2020;
Brilliantova and Thurner, 2019; Teufel et al., 2019;
Climate Change Arshad et al., 2023; Khan et al., 2023; Truby et al., 2022; Wang,
Studies on mitigating climate change.
Mitigation 2022; Fu et al., 2021; Olivier et al., 2017;
Luna et al., 2024; Dehshiri and Amiri, 2024; Arshad et al.,
**Environment** Environmental Research focused on environmental 2023; Tyagi, 2023; Yang et al., 2023; Naqash et al., 2023;
Rodríguez 2023; Waqar, 2023; Singh et al., 2023; Sipthorpe et
Conservation conservation.
al., 2022; Jiang and Zheng, 2021; Park and Li, 2021; Kim et al.,
2021; Richardson, 2020; Allena, 2020; Schulz et al., 2020
Dehshiri and Amiri, 2024; Gazzola et al., 2023; Bosona and
Sustainable Studies pertaining to the role of BT on Gebresenbet, 2023; Dal Mas et al., 2023; Yontar. 2023; Pandey,
Agriculture developing sustainable agriculture. 2022; Bhusal, 2021
Prashar et al., 2020; Mirabelli and Solina, 2020
Research related to BT integration to enhance
Sustainable supply Dehshiri and Amiri, 2024; Jabbar and Bjørn, 2018; Bett ́ın
supply chain management, lower costs, and
chain management D ́ıaz et al., 2018; Kshetri, 2018; Tian, 2016
achieve a competitive edge
**Economic** The role of BT on ehancing transparency Cozzio, 2023; Tafuro et al., 2023; Balzarova et al., 2022;
Promoting fair trade
within fair trade practices. Cozzio, 2023; Francisco and Swanson, 2018
Employment and The influence of BT on employment and
Tartan, 2023; Shabaltina et al., 2021; Novak, 2019
Income Distribution income distribution
Research that discusses equality and social Gazzola et al., 2023; Chaudhuri et al. 2023; Fallah Shayan et
inclusion such as financial inclusion, secure al., 2022; Thanasi-Boçe et al., 2022; Al-Issa et al., 2022; Böckel
Equality and social
identity verification, transparent government, et al., 2021; Khanfar et al., 2021; Venkatesh et al., 2020; Dal
inclusion
education and credential verification, and Mas et al., 2020; Konashevych, 2020; Chen et al., 2020; Martin
decentralized markets et al., 2011; Carter and Rogers, 2008.
Ronaghi and Mosakhani, 2023; Rai et al., 2021; Tang, 2018;
Business ethics and
Research related to the relation between BT Mani et al., 2016; Lashley, 2016; Krechovská and Prochazkova,
effective corporate
and corporate ethics and performance. 2014; Krishna et al., 2011; Aguilera et al., 2009; Carter and
governance
**Social** Rogers, 2008
Enhancing quality of
life:
Research focused on improving the quality of
-Sustainable Gazzola et al., 2023; Vishwakarma, 2023; Chaudhuri et al.,
life through the impact of BT on:
production and 2023; Sikder, 2023; Al-Issa et al., 2022; Khanfar et al., 2021;
-environment
consumption Khanfar et la., 2021; Lu et al., 2020; Rai et al., 2021; Mani et al.,
-satisfying consumers’ needs for products and
-Product 2016
services
authentication and
traceability
**Source: Authors’ own creation**
_2.3_ _Blockchain and environmental sustainability_
Environmental sustainability, as outlined in the United Nations Sustainable Development Goals
(SDGs), refers to the responsible interaction with the environment to avoid depletion or degradation
of natural resources and allow for long-term environmental quality (UN, 2023). In a time when
environmental issues are becoming ever more complicated and interconnected, environmental
sustainability is not merely a requirement but a worldwide obligation. In this effort, the SDGs of the
United Nations function as a guiding framework, delineating the complex aspects of sustainability.
These goals underscore the urgent need to address critical issues such as climate action (Goal
13) to combat climate change, protect oceans and marine life (Goal 14), and sustain terrestrial
ecosystems (Goal 15). They emphasize the necessity of clean water and sanitation (Goal 6), affordable
and clean energy (Goal 7), and responsible consumption and production (Goal 12) to ensure efficient
use of resources and waste reduction. Additionally, these goals highlight the importance of global
partnerships (Goal 17) for effective implementation and policy coherence, incorporating aspects of
international cooperation and technology transfer to achieve environmental sustainability.
|Theme|Sub-theme|Description|References|
|---|---|---|---|
|Environment|Renewable Energy|Research related to renewable energy.|Alhasan and Hamdan 2023; Cui, 2023; Ahmad, 2023; Zuo, 2022; Bai et al., 2022; Thakur et al., 2021; Howson, 2020; Brilliantova and Thurner, 2019; Teufel et al., 2019;|
||Climate Change Mitigation|Studies on mitigating climate change.|Arshad et al., 2023; Khan et al., 2023; Truby et al., 2022; Wang, 2022; Fu et al., 2021; Olivier et al., 2017;|
||Environmental Conservation|Research focused on environmental conservation.|Luna et al., 2024; Dehshiri and Amiri, 2024; Arshad et al., 2023; Tyagi, 2023; Yang et al., 2023; Naqash et al., 2023; Rodríguez 2023; Waqar, 2023; Singh et al., 2023; Sipthorpe et al., 2022; Jiang and Zheng, 2021; Park and Li, 2021; Kim et al., 2021; Richardson, 2020; Allena, 2020; Schulz et al., 2020|
||Sustainable Agriculture|Studies pertaining to the role of BT on developing sustainable agriculture.|Dehshiri and Amiri, 2024; Gazzola et al., 2023; Bosona and Gebresenbet, 2023; Dal Mas et al., 2023; Yontar. 2023; Pandey, 2022; Bhusal, 2021 Prashar et al., 2020; Mirabelli and Solina, 2020|
|Economic|Sustainable supply chain management|Research related to BT integration to enhance supply chain management, lower costs, and achieve a competitive edge|Dehshiri and Amiri, 2024; Jabbar and Bjørn, 2018; Bett ́ın- D ́ıaz et al., 2018; Kshetri, 2018; Tian, 2016|
||Promoting fair trade|The role of BT on ehancing transparency within fair trade practices.|Cozzio, 2023; Tafuro et al., 2023; Balzarova et al., 2022; Cozzio, 2023; Francisco and Swanson, 2018|
||Employment and Income Distribution|The influence of BT on employment and income distribution|Tartan, 2023; Shabaltina et al., 2021; Novak, 2019|
|Social|Equality and social inclusion|Research that discusses equality and social inclusion such as financial inclusion, secure identity verification, transparent government, education and credential verification, and decentralized markets|Gazzola et al., 2023; Chaudhuri et al. 2023; Fallah Shayan et al., 2022; Thanasi-Boçe et al., 2022; Al-Issa et al., 2022; Böckel et al., 2021; Khanfar et al., 2021; Venkatesh et al., 2020; Dal Mas et al., 2020; Konashevych, 2020; Chen et al., 2020; Martin et al., 2011; Carter and Rogers, 2008.|
||Business ethics and effective corporate governance|Research related to the relation between BT and corporate ethics and performance.|Ronaghi and Mosakhani, 2023; Rai et al., 2021; Tang, 2018; Mani et al., 2016; Lashley, 2016; Krechovská and Prochazkova, 2014; Krishna et al., 2011; Aguilera et al., 2009; Carter and Rogers, 2008|
||Enhancing quality of life: -Sustainable production and consumption -Product authentication and traceability|Research focused on improving the quality of life through the impact of BT on: -environment -satisfying consumers’ needs for products and services|Gazzola et al., 2023; Vishwakarma, 2023; Chaudhuri et al., 2023; Sikder, 2023; Al-Issa et al., 2022; Khanfar et al., 2021; Khanfar et la., 2021; Lu et al., 2020; Rai et al., 2021; Mani et al., 2016|
-----
**_ISSN 2281-3993_** **_www.richtmann.org_** **_March 2024_**
Achieving environmental sustainability demands innovative strategies, with emerging
technologies such as blockchain offering novel avenues for progress. Policymakers can make more
informed decisions by comprehending the effects of market interventions and weighing their options,
mindful of potential disruptions. The insights gained can guide political institutions in shaping the
necessary political and legal frameworks to foster the creation of effective green blockchain
applications (Arshad et al., 2023).
As the exploration of blockchain's potential in promoting sustainability progresses, it becomes
essential to examine its influence on specific domains that are important for the environmental
agenda.
_2.3.1_ _Supply chain management_
Supply chain may be made much more transparent with the use of BT (Kouhizadeh et al., 2021; Risso
et al., 2023). Every transaction or movement of products may be tracked on a tamper-proof ledger by
utilizing BT (Yap et al., 2023). As a result, parties involved in the supply chain can identify
inefficiencies and sources of waste and pollution by tracking the origin, path, and present state of
products (Kim et al., 2021). By utilizing BT to document every stage of a product's supply chain, it
becomes feasible to verify the purchase of items from practices that promote biodiversity
conservation (Park and Li, 2021; Tyagi, 2023; Dehshiri and Amiri, 2024).
By fostering sustainable supply chains, this directly promotes SDG 12 (Responsible
Consumption and Production). It can also impact SDG 13 (Climate Action) by helping reduce
emissions associated with production and logistics.
_2.3.2_ _Water resource management_
The implementation of BT has the potential to safeguard the integrity and transparency of water
usage and quality data, similar to its application in managing supply chain information and energy
data (Rodríguez 2023). The recording of every water usage event, such as extraction, purification, and
distribution, has the potential to be documented on a blockchain (Naqash et al., 2023). Initiatives
such as IBM's blockchain-based water management system facilitate the establishment of transparent
systems for the administration of water rights and consumption data, thereby promoting fair and
sustainable practices in water distribution and utilization. It aligns with SDG 6 (Clean Water and
Sanitation) by promoting sustainable management of water resources.
_2.3.3_ _Energy efficiency_
Assisting in the management of the complexities of a decentralized energy system and enhancing
energy operations along the entire value chain are two ways in which BT is anticipated to transform
the energy industry (Brilliantova and Thurner, 2019). BT enables the verification, automation, and
security of energy transfers without intermediaries (Teufel et al., 2019). BT can be used to create
decentralized energy grids and systems that reward energy-saving behaviors. For example, smart
contracts on a blockchain can automatically compensate individuals or organizations that reduce
their energy consumption. This supports SDG 7 by promoting energy efficiency and SDG 13 by
reducing energy-related emissions.
_2.3.4_ _Renewable energy certificate (REC)_
In the realm of renewable energy, blockchain can play a pivotal role in issuing, tracking, and verifying
RECs (Zuo, 2022; Cui, 2023). By guaranteeing each certificate's authenticity and distinctiveness, this
system eliminates fraud and double counting. This is in line with SDG 13 and SDG 7 since it
encourages the use of renewable energy sources.
-----
**_ISSN 2281-3993_** **_www.richtmann.org_** **_March 2024_**
As blockchain redefines the paradigms of transparency and traceability in supply chains, it opens new
avenues for mitigating GHG emissions and empowering renewable energy markets. Through its
ability to verify the authenticity of RECs and carbon credits, blockchain enables a more robust and
trustworthy system for environmental accounting (Sipthorpe et al., 2022). Its role in incentivizing
energy efficiency further illustrates its utility in the pursuit of a low-carbon economy.
_2.3.5_ _Carbon credits_
The utilization of BT facilitates the establishment of a reliable yet readily transparent infrastructure
for the distribution, exchange, and suspension of this approach is designed to prevent the duplication
of carbon credits and to verify that each credit corresponds to a genuine decrease in emissions
(Richardson, 2020). This action helps the achievement of SDG 13 by facilitating the implementation of
a mechanism that allows for the compensation of greenhouse gas emissions. It serves as an incentive
for enterprises and nations to allocate resources towards programs aimed at reducing emissions
(Sipthorpe et al., 2022).
_2.3.6_ _Green building and sustainable construction:_
Similarly to monitoring REC and ensuring energy efficiency, the utilization of BT within the
construction industry facilitates the monitoring and assessment of the sustainability of building
materials and construction methodologies (Waqar, 2023; Singh et al., 2023). Information about the
sourcing, production, and transportation of materials can be stored on a blockchain. The tracking of
construction materials' origin and environmental impact is a capability that stakeholders need. The
implementation of this transparency measure guarantees compliance with specific environmental
criteria and facilitates the adoption of sustainable construction methods (Jiang and Zheng, 2021).
Sustainable materials are key to reducing the environmental footprint of buildings (Yang et al., 2023).
This relates to SDG 11 (Sustainable Cities and Communities), focusing on sustainable construction
practices.
_2.3.7_ _Sustainable agriculture and food systems:_
The application of BT in enhancing supply chain transparency is highly relevant in the context of
sustainable agriculture and food systems since it enables the seamless monitoring of products from
their origin to the final consumer (Mirabelli and Solina, 2020; Bosona and Gebresenbet, 2023; Dal
Mas et al., 2023). The entire process, spanning from production to retail, is meticulously documented,
thereby guaranteeing the implementation of food safety protocols and sustainable practices. It has
the capacity to fundamentally transform the processes of food production, distribution, and
consumption (Yontar. 2023; Pandey, 2022). This directly corresponds to the objectives of establishing
resilient agricultural practices and guaranteeing sustainable food production systems, addressing
SDG 2 and SDG 12 (Dehshiri and Amiri, 2024). AgriDigital and provenance are two important
initiatives that provide blockchain-based solutions for the traceability of agricultural products. This
approach promotes more transparency and provides valuable support for the adoption of sustainable
farming methods.
The global importance of food safety and quality has been highlighted by recent high-profile
incidents, raising public interest in food traceability (Prashar et al., 2020).
Due to its ability to track food through all stages of its life cycle, the World Health Organization
encourages a cooperative approach among governments, producers, and consumers to ensure safety
through information sharing in complex food networks. Profit-driven businesses often use
information systems to track food, enhancing safety and potentially increasing profits. Gazzola et al.
(2023) investigated how BT can contribute to this area, particularly in building positive consumer
relationships.
-----
**_ISSN 2281-3993_** **_www.richtmann.org_** **_March 2024_**
The utilization of BT has promise for establishing a more secure, environmentally conscious,
and dependable agricultural food system in the forthcoming years (Bosona and Gebresenbet, 2023).
Despite being in its nascent phases and facing obstacles including high implementation costs, privacy
concerns, security issues, scalability limitations, and performance challenges, the integration of this
technology has the potential to bring about a substantial transformation in the agricultural sector
(Bhusal, 2021).
_2.3.8_ _Environmental policy compliance and governance:_
Blockchain can monitor compliance with environmental policies and regulations, akin to verifying
carbon credits and managing RECs. It ensures a decentralized and unalterable ledger of emissions
data and adherence to environmental regulations. Governments can manage environmental subsidies
and penalties with BT, ensuring policy implementation is transparent and efficient (Luna et al., 2024;
Allena, 2020; Schulz et al., 2020). The promotion of open and accountable governance contributes to
the achievement of SDG 13 and SDG 16 (Peace, Justice, and Strong Institutions).
_2.3.9_ _Public participation and awareness in environmental conservation_
The utilization of BT has the potential to improve public involvement in environmental initiatives
using tokenization, thereby enabling anyone to engage in conservation programs through investment
or direct participation, like how blockchain encourages participation in renewable energy markets
(Alhasan and Hamdan, 2023; Bai et al., 2022; Thakur et al., 2021; Howson, 2020)
Platforms like Earth Token facilitate individuals' ability to invest in environmental assets, hence
fostering public engagement in initiatives aimed at environmental preservation. Supports various
SDGs by fostering inclusive participation in sustainable development (SDG 17 - Partnerships for the
Goals).
_2.3.10Sustainable transportation and electric vehicles (EVs):_
The utilization of BT can facilitate the effective management and optimization of electric vehicle (EV)
charging stations, as well as their easy integration into smart power grids (Fu et al., 2021; Wang, 2022;
Khan et al., 2023). Blockchain is being investigated by initiatives such as MOBI (Mobility Open
Blockchain Initiative) to determine the identity, history, and utilization of vehicles; this research may
promote the adoption of more environmentally friendly transportation methods. Aligns with SDG 11
and SDG 13 by promoting sustainable urban transportation systems and contributing to climate
action.
_2.3.11_ _Innovative climate-conscious projects:_
As the world deals with the effects of climate change, which include loss of species, extreme weather,
and an imbalance in the environment, countries must focus on long-term economic growth. Even
though the world's economy grew by an average of 3.4% per year from 2012 to 2018, rising greenhouse
gas (GHG) emissions cast a shadow over this progress. This is mostly because of practices that use a
lot of energy and resources (Olivier et al., 2017). The Intergovernmental Panel on Climate Change
(IPCC) reports and the growing support for green projects in the public and political spheres, which
led to important agreements like the Paris Agreement of 2015, show how important it is to act quickly
to stop climate change. Climate markets need to become open to the world because people are more
aware of the effects of climate change (Truby et al., 2022; Arshad et al., 2023) and to spend and come
up with new ways to make them more resilient.
Several projects utilizing BT have been developed to address the challenge of climate change.
These include various platforms designed for monitoring and mitigating emissions, projects that
-----
**_ISSN 2281-3993_** **_www.richtmann.org_** **_March 2024_**
provide incentives for adopting sustainable behaviors, and structures that facilitate climate finance
and investment in projects aimed at promoting sustainability. These projects can contribute to
multiple SDGs, including SDG 13, SDG 11, and SDG 15 by promoting actions that mitigate climate
change and its impacts.
Each of the previously discussed topics is connected to the broader theme related to a specific
SDG and is interconnected with the previously identified areas such as: monitoring carbon footprints
with blockchain; enhancing renewable energy markets; waste management and recycling and climate
change. This methodology offers a systematic framework for comprehending the interplay and
contributions of different aspects of sustainability and BT to these broad concepts. Below in Table 2
we show a categorized table that aligns each theme placed under the broader topic where its impact
is most significant.
**Table 2. Environmental sustainability-related themes**
**Broader Topics** **Sustainability-related Themes**
- Supply Chain Management
Monitoring Carbon Footprints with Blockchain
- Environmental Policy Compliance and Governance
- Renewable Energy Certificates (RECs)
Enhancing Renewable Energy Markets - Energy Efficiency
- Sustainable Transportation and Electric Vehicles (EVs)
- Sustainable Agriculture and Food Systems
Waste Management and Recycling
- Green Building and Sustainable Construction
- Public Participation and Awareness- Innovative Climate
Climate Change
- Conscious Projects
Supply Chain Management and Environmental Policy Compliance and Governance are key in
monitoring carbon footprints, as they involve tracking emissions and ensuring regulatory adherence.
RECs, Energy Efficiency, and Sustainable Transportation/EVs enhance renewable energy markets by
promoting the use of clean energy and efficient energy practices. Sustainable Agriculture and Food
Systems and Green Building and Sustainable Construction contribute to waste management and
recycling through sustainable practices and resource efficiency. Public Participation and Awareness
and Innovative Climate-Conscious Projects are crucial for addressing climate change, as they involve
engaging the public and implementing novel solutions to climate challenges.
Table 3 provides a clear overview of how BT is being applied in various sectors to address
environmental challenges and sustainability goals. Each application demonstrates the potential of
blockchain to contribute to a more sustainable and environmentally conscious world.
**Table 3. Application of blockchain for environmental sustainability**
Broader Topic Specific Application Description
Monitoring Carbon Companies use blockchain-based smart contracts
- Smart contracts for
Footprints with to automatically track and report emissions,
emission tracking
Blockchain integrating sensors and IoT devices.
|Broader Topics|Sustainability-related Themes|
|---|---|
|Monitoring Carbon Footprints with Blockchain|- Supply Chain Management - Environmental Policy Compliance and Governance|
|Enhancing Renewable Energy Markets|- Renewable Energy Certificates (RECs) - Energy Efficiency - Sustainable Transportation and Electric Vehicles (EVs)|
|Waste Management and Recycling|- Sustainable Agriculture and Food Systems - Green Building and Sustainable Construction|
|Climate Change|- Public Participation and Awareness- Innovative Climate - Conscious Projects|
|Broader Topic|Specific Application|Description|
|---|---|---|
|Monitoring Carbon Footprints with Blockchain|- Smart contracts for emission tracking|Companies use blockchain-based smart contracts to automatically track and report emissions, integrating sensors and IoT devices.|
||- Decentralized carbon - Emission trading platforms|Blockchain enables the creation of platforms for transparent and efficient trading of carbon credits, as seen with IBM and Energy Blockchain Labs.|
|Enhancing Renewable Energy Markets|- Blockchain in microgrids|Projects like Brooklyn Microgrid use blockchain to create local energy networks for buying and selling renewable energy.|
||- Tokenization of renewable energy assets|Blockchain is used for tokenizing renewable energy assets, facilitating investment in renewable energy projects (e.g., WePower).|
-----
**_ISSN 2281-3993_** **_www.richtmann.org_** **_March 2024_**
Broader Topic Specific Application Description
Blockchain-based platforms like Plastic Bank
- Recycling incentivization offer tokens for collecting and recycling
programs materials, incentivizing proper recycling
Waste Management
practices.
and Recycling
Blockchain is used to track the lifecycle of
- Supply chain transparency for
products, ensuring responsible recycling (e.g.,
recycling
Arianee project for luxury goods).
Blockchain platforms facilitate investments in
- Climate finance and investment climate change mitigation projects, such as the
platforms Poseidon Foundation supporting forest
conservation efforts.
Climate Change
Projects like the Open Earth Foundation use
- Tracking and verifying climate blockchain for transparent and tamper-proof
data climate data, aiding in climate policy and
modeling.
_2.4_ _Blockchain in economic sustainability_
Economic sustainability, as per the United Nations' SDGs, refers to the practice of managing
resources and developing economically in a way that ensures long-term economic health without
harming the environment or compromising the ability of future generations to meet their own needs.
It involves a balanced approach that integrates economic growth with social inclusion and
environmental protection (UN, 2023).
The SDGs encompass various facets of economic sustainability through distinct objectives. Goal
8 advocates for sustained, inclusive economic growth and decent work, emphasizing productivity,
innovation, entrepreneurship, and the separation of economic growth from environmental harm.
Goal 9 targets the development of resilient infrastructure, sustainable industrialization, and
innovation as key drivers of economic progress. Goal 10 is focused on reducing inequalities both
within and among countries, promoting equitable distribution of wealth and resources as a
cornerstone of sustainable economic development. Lastly, Goal 12 aims to establish sustainable
patterns of consumption and production, prioritizing efficient resource use, and waste reduction, and
minimizing the environmental impact of economic activities.
_2.4.1_ _Promoting fair trade_
Economic sustainability can be bolstered by enhancing transparency within fair trade practices. Fair
trade relies on consumer willingness to pay premiums based on the assurance of superior quality and
ethical practices within supply chains. Ecolabels must therefore demonstrate attributes such as
traceability, accountability, and ecological sustainability to build trust. Blockchain technology (BT) is
posited as an improvement over traditional marketing strategies, with Balzarova et al. (2022) noting
its net positive impact on food supply chains. Francisco and Swanson (2018) argue that BT could
significantly improve transparency and traceability in agricultural supply chains. This technology
facilitates consumer trust, providing self-verification mechanisms that can potentially replace
reliance on ecolabels (Cozzio, 2023).
In assessing the adoption of blockchain in fair trade, Balzarova et al. (2022) utilized the
Technology Readiness Index (TRI), unveiling five themes: the conditional benefits of BT, the duality
of transparency outcomes, consumer behavior factors, and implementation barriers, highlighting the
practical challenges of BT adoption. Fairtrade certification to blockchain adoption begins with the
firm's expertise and willingness to implement BT (Holmberg et al., 2022). Moreover, public-private
partnerships (PPPs) often face transparency and accountability issues, impacting trust and
collaboration. Tafuro et al. (2023) examined blockchain’s potential to address these issues in PPPs,
|Broader Topic|Specific Application|Description|
|---|---|---|
|Waste Management and Recycling|- Recycling incentivization programs|Blockchain-based platforms like Plastic Bank offer tokens for collecting and recycling materials, incentivizing proper recycling practices.|
||- Supply chain transparency for recycling|Blockchain is used to track the lifecycle of products, ensuring responsible recycling (e.g., Arianee project for luxury goods).|
|Climate Change|- Climate finance and investment platforms|Blockchain platforms facilitate investments in climate change mitigation projects, such as the Poseidon Foundation supporting forest conservation efforts.|
||- Tracking and verifying climate data|Projects like the Open Earth Foundation use blockchain for transparent and tamper-proof climate data, aiding in climate policy and modeling.|
-----
**_ISSN 2281-3993_** **_www.richtmann.org_** **_March 2024_**
suggesting that despite its complexities, blockchain can enhance the efficacy of PPPs and contribute
to sustainable development goals.
_2.4.2_ _Sustainable Supply Chain Management_
Blockchain technology (BT) has emerged as a transformative force in supply chain management,
offering unparalleled benefits in terms of transparency, security, and operational efficiency, which are
pivotal for sustainable economic growth. This technology's core advantage lies in its ability to create a
decentralized, immutable ledger that enables real-time tracking of products from their point of
production to delivery (Sarfraz et al., 2023; Waqar et al., 2023; Yontar et al., 2023; Risso et al., 2023;
Khanfar et al., 2021; Kouhizadeh et al., 2021). This capability is instrumental in reducing losses
associated with counterfeit and gray market trading, while also promoting environmentally friendly
production methodologies (Carter and Easton, 2011; Yap et al., 2023). Moreover, blockchain facilitates
the automation of sustainability agreements through smart contracts (Dal Mas, et al., 2020), ensuring
compliance with eco-friendly criteria and fostering the circular economy by meticulously
documenting products' lifecycle for future reuse (Yap et al., 2023).
Despite the promising potential of blockchain in revolutionizing supply chains, its integration
into existing systems is fraught with challenges. These include the need for substantial technology
infrastructure development, regulatory support, and overcoming the general reluctance towards
adopting new technologies. Specific hurdles such as user-friendliness, the proprietary nature of
blockchain solutions, and the seamless integration of virtual and physical tracking mechanisms are
significant (Dehshiri and Amiri, 2024; Jabbar and Bjørn, 2018; Kshetri, 2018; Tian, 2016). The
reconciliation of blockchain’s virtual capabilities with the physical tracking of items poses a complex
problem that research has yet to fully address, often focusing more on the virtual benefits than on
practical physical applications.
In addition to revolutionizing supply chain management, blockchain significantly enhances
business efficiency by simplifying processes, eliminating intermediaries, thus reducing costs and
saving time. For instance, in the energy sector, blockchain enables efficient peer-to-peer energy
trading, facilitating the use of renewable energy sources and contributing to a more sustainable
energy supply. Technology also allows for the tokenization of physical assets, such as real estate,
making these markets more liquid and efficient by enabling assets to be traded on blockchain
platforms (Carter and Easton, 2011).
The deployment of blockchain technology in supply chain management not only ensures
greater transparency, security, and cost efficiency but also facilitates the real-time tracking of the
entire production and delivery process. This aspect significantly mitigates the risks associated with
counterfeit and gray market trading, thereby endorsing sustainable business models (Thanasi-Boçe et
al., 2022). Blockchain's ability to enhance product quality, prevent counterfeits, and achieve
stakeholder transparency is pivotal for monitoring resource utilization and promoting sustainable
manufacturing practices. Furthermore, smart contracts play a critical role in ensuring suppliers
adhere to environmental standards, thereby incentivizing responsible production (Yap et al., 2023).
Despite its potential, integrating BT into supply chains poses challenges, including the necessity
for developing technology infrastructure, regulatory support, and overcoming the reluctance towards
new technologies. Issues such as ease of use, the proprietary nature of blockchain solutions, and the
integration of virtual and physical tracking systems represent significant hurdles. Bridging the gap
between blockchain's virtual capabilities and the physical tracking of items remains a complex
challenge, with current research focusing more on the technology's virtual advantages than its
application in physical item tracking (Dehshiri and Amiri, 2024; Jabbar and Bjørn, 2018; Kshetri, 2018;
Tian, 2016). Furthermore, most research has focused on blockchain's benefits for companies, with less
emphasis on consumer information. Bett ́ın-D ́ıaz and colleagues (2018) suggested a methodology for
developing traceable supply chains with consumer considerations, but detailed strategies for
conveying information to customers are limited.
-----
**_ISSN 2281-3993_** **_www.richtmann.org_** **_March 2024_**
Addressing the challenges of blockchain integration within supply chains necessitates the
development of a robust technological infrastructure and sustainable methodologies, bolstered by
collaborative investments and regulatory frameworks among supply chain partners (Dehshiri and
Amiri, 2024). The ease of use and the capacity to integrate blockchain within existing infrastructures
are essential for promoting user adoption and overcoming resistance towards new technologies
(Jabbar and Bjørn, 2018).
_2.4.3_ _Impact on Employment and Income Distribution_
The influence of BT on employment and income distribution is a fluctuating and complex matter. BT
has the potential to generate novel employment prospects and possibilities, notably in domains such
as software development, consulting, and auditing. Conversely, other authors claim that the use of BT
may result in workforce reduction within specific industries, given its capacity to automate functions
and procedures presently executed by human labor (Novak, 2019). The utilization of BT possesses the
capacity to exert influence on employment and income distribution through several avenues,
including both advantageous and detrimental outcomes, as provided is in Table 4.
**Table 4: The impact of blockchain on employment and income distribution**
Positive or
Impact Negative Explanation
(+/-)
BT could provide jobs in blockchain development, cybersecurity, data
Job creation (+) analysis, and smart contract audits. The growing acceptance of BT is
creating job opportunities for skilled workers in connected fields.
BT can simplify and streamline manual processes, reducing middlemen
Streamlining of
(+) and administrative staff. This technique may replace jobs in some
processes
locations, but it can save companies money and improve operations.
BT enables decentralized platforms and apps, boosting the gig economy.
This phenomenon has pros and disadvantages. Peer-to-peer networks
Decentralization (+)
allow people to make money by completing activities, sharing resources,
and selling services. This could improve income distribution.
Banks can use BT to provide financial services to those who don't have
Financial
(+) access to them. This allows people to work and participate in the formal
inclusion
economy, empowering them.
Blockchain wealth concentration among early adopters and significant
Income inequality (-) enterprises may aggravate economic inequality. Individuals with
significant computer power and resources may have an advantage.
The growing demand for BT skills spurs investment in educational and
Reskilling and
(+) training programs, giving people the chance to learn new skills and boost
education
their income (Fleener, 2022).
Regulatory ambiguity in the blockchain business can hinder investment
Regulatory
(-) and job growth. Businesses are hesitant to expand in uncertain legal
uncertainty
environments.
Additionally, BT has a positive impact on reducing fraud and corruption. Blockchain's immutable
ledger ensures that records cannot be altered after the fact, which can significantly reduce fraud and
corruption, particularly in public sectors such as land registries and government contracts.
_2.5_ _Blockchain for social sustainability_
BT significantly affects social sustainability, addressing systemic issues across various key areas to
foster a more equitable, transparent, and sustainable society. This transformative technology
|Impact|Positive or Negative (+/-)|Explanation|
|---|---|---|
|Job creation|(+)|BT could provide jobs in blockchain development, cybersecurity, data analysis, and smart contract audits. The growing acceptance of BT is creating job opportunities for skilled workers in connected fields.|
|Streamlining of processes|(+)|BT can simplify and streamline manual processes, reducing middlemen and administrative staff. This technique may replace jobs in some locations, but it can save companies money and improve operations.|
|Decentralization|(+)|BT enables decentralized platforms and apps, boosting the gig economy. This phenomenon has pros and disadvantages. Peer-to-peer networks allow people to make money by completing activities, sharing resources, and selling services. This could improve income distribution.|
|Financial inclusion|(+)|Banks can use BT to provide financial services to those who don't have access to them. This allows people to work and participate in the formal economy, empowering them.|
|Income inequality|(-)|Blockchain wealth concentration among early adopters and significant enterprises may aggravate economic inequality. Individuals with significant computer power and resources may have an advantage.|
|Reskilling and education|(+)|The growing demand for BT skills spurs investment in educational and training programs, giving people the chance to learn new skills and boost their income (Fleener, 2022).|
|Regulatory uncertainty|(-)|Regulatory ambiguity in the blockchain business can hinder investment and job growth. Businesses are hesitant to expand in uncertain legal environments.|
-----
**_ISSN 2281-3993_** **_www.richtmann.org_** **_March 2024_**
underpins efforts aligned with the United Nations' Sustainable Development Goals (SDGs),
enhancing quality of life and promoting social inclusion (UN, 2023; Mani et al., 2016). The key areas
of social impact where blockchain demonstrates its substantial contributions are discussed below. In
each of the discussed areas, blockchain technology not only addresses pressing social and
environmental challenges but also promotes a more equitable distribution of resources and
opportunities, underscoring its profound impact on fostering a sustainable and inclusive global
society (Lu et al., 2020; Rai et al., 2021).
_2.5.1_ _Promoting equality and social inclusion_
Blockchain technology has the potential to promote equality and social inclusion through several key
mechanisms. By providing a decentralized and transparent ledger system, it offers innovative
solutions to systemic problems that hinder equality and social inclusion. The discussion below
focuses of the ways blockchain can contribute to these goals:
_2.5.2_ _Financial inclusion_
Blockchain facilitates access to financial services for the unbanked and underbanked populations,
who are often excluded from traditional banking systems. By enabling peer-to-peer transactions
without the need for intermediaries, blockchain can lower transaction costs and make financial
services more accessible to everyone, regardless of their geographic location or socioeconomic status.
This increased access to financial services can help reduce poverty and boost economic participation
among marginalized communities.
Blockchain plays a crucial role in bridging the financial gap for unbanked and underprivileged
populations. Platforms like BanQu provide digital identities and secure, accessible financial
transactions, offering disenfranchised individuals, including refugees, access to banking and financial
services, thereby facilitating economic participation and empowerment (Fallah Shayan et al., 2022).
_2.5.3_ _Secure identity verification_
Blockchain can provide secure and immutable digital identities, offering a solution for individuals
without official documents or those whose records have been lost due to conflicts or disasters. A
blockchain-based identity system can enable these individuals to access essential services such as
healthcare, education, and banking, thereby promoting social inclusion.
_2.5.4_ _Transparent and fair governance_
Blockchain secures electoral systems and enhances civic engagement through transparent and
trustworthy voting mechanisms. By ensuring that votes are tamper-proof and accurately recorded,
blockchain can foster a more inclusive and fair political process, giving marginalized groups a
stronger voice and promoting political equality. Projects like Voatz show the potential of blockchain
to increase participation in democratic processes, strengthen democratic governance, and encourage
citizen involvement in decision-making processes (Carter and Rogers, 2008).
_2.5.5_ _Supply chain transparency_
Blockchain can track the provenance of products from origin to consumer, ensuring fair trade and
ethical practices (Balzarova et al., 2022). This transparency can empower consumers to make
informed choices that support social and economic fairness, benefiting small producers and workers
in developing countries by ensuring they receive a fair share of profits.
-----
**_ISSN 2281-3993_** **_www.richtmann.org_** **_March 2024_**
_2.5.6_ _Education and credential verification_
Blockchain can securely store and verify academic credentials, enabling individuals from
disadvantaged backgrounds to prove their qualifications and skills easily. This can improve access to
job opportunities and higher education, breaking down barriers to social mobility and promoting
equality.
_2.5.7_ _Education and employment_
Blockchain enhances the mobility and employability of individuals by authenticating academic
credentials and professional achievements. Through initiatives like the MIT Media Lab's Digital
Certificates Project, blockchain ensures the integrity and verifiability of educational records,
streamlining employment processes and supporting lifelong learning (Thanasi-Boçe et al., 2022). This
enables individuals from disadvantaged backgrounds to prove their qualifications and skills easily.
Also, it can improve access to job opportunities and higher education, breaking down barriers to
social mobility and promoting equality.
_2.5.8_ _Decentralized markets_
By facilitating decentralized marketplaces, blockchain allows small businesses and entrepreneurs
from marginalized communities to participate in the global economy directly. This can level the
playing field, reducing the dominance of large corporations and empowering individuals and small
enterprises (Dal Mas et al., 2020).
_2.5.9_ _Protection of property rights_
In countries where land and property rights are not adequately documented and enforced,
blockchain technology has the potential to offer a record of ownership that is both secure and
unchangeable. This has the potential to safeguard the rights of vulnerable communities against
encroachment and disputes, so fostering economic stability and social inclusion simultaneously
(Konashevych, 2020; Chen et al., 2020).
_2.5.10Enhancing quality of life_
Blockchain technology has the potential to enhance the quality of life through several key
mechanisms that are discussed below:
_2.5.11_ _Environmental stewardship and circular economy_
Blockchain incentivizes recycling and responsible waste management through projects like Plastic
Bank, addressing environmental challenges and promoting the principles of the circular economy. By
tokenizing waste and facilitating the trade of recyclable materials, blockchain contributes to
sustainable practices and environmental conservation.
_2.5.12_ _Establishing consumer trust and transparency through authenticity and traceability_
BT is increasingly vital for brands in various sectors to ensure product authenticity and traceability,
which is fundamental in establishing consumer trust and transparency. This decentralized and
transparent technology enables brands to authenticate their products and track their supply chain
journey, meeting the growing consumer demand for ethical and environmentally responsible
production (Thanasi-Boçe et al., 2022) especially in the growing e-commerce (Sikder, 2023).
-----
**_ISSN 2281-3993_** **_www.richtmann.org_** **_March 2024_**
Key applications of blockchain for product authenticity and traceability illustrated with
examples include:
**_Supply Chain Management: Blockchain is used to document and track every stage of a_**
product's progression, from raw material acquisition to final delivery (Venkatesh et al., 2020). For
instance, IBM Food Trust is a blockchain-based platform employed by major retailers like Walmart to
trace the origin of food products, such as leafy greens. Consumers can access this information by
scanning a QR code on the packaging, ensuring the product's legitimacy and source.
**_Food and Agriculture: Farmers can document crop-related data on blockchain, enhancing_**
food safety and integrity. An example is the BeefLedger project in Australia, which uses blockchain to
trace the source of beef products. Consumers can scan a QR code on meat packaging to access details
about the cattle's breed, origin, diet, and processing methods, thereby verifying food provenance and
quality using product labels.
**_Luxury Goods: Luxury brands are leveraging blockchain to provide digital certificates of_**
authenticity. LVMH, the conglomerate that owns luxury brands like Louis Vuitton, has launched the
AURA blockchain platform. It allows consumers to verify the authenticity and origin of luxury items,
such as designer bags and watches. Each product comes with an NFC chip linked to blockchain,
providing a digital certificate of authenticity and detailed information about the product's history.
This shift towards sustainable luxury aligns with contemporary values, setting new standards in the
luxury industry for technological innovation and ethical practices (Al-Issa et al., 2022).
**_Pharmaceuticals: In the pharmaceutical industry, blockchain is used to monitor medication_**
lifecycles and reduce counterfeit products. Chronicled is a blockchain platform employed to combat
counterfeit drugs. It enables tracking the production and distribution of pharmaceuticals, ensuring
that medications are genuine. Patients can authenticate prescriptions using QR codes.
**_Copyright and Digital Content: Creators can establish ownership of their work through_**
blockchain timestamps. For example, Verisart is a blockchain-based platform used by artists and
creators to certify the authenticity of digital art and collectibles. It provides a timestamped certificate
of authenticity on the blockchain, making it easy for buyers to verify the originality of digital content.
**_Automotive Industry: Car manufacturers are exploring blockchain applications to maintain_**
comprehensive records of vehicle maintenance, accidents, and ownership changes. BMW, for
instance, aims to create a tamper-proof history of used cars, recording maintenance, accident history,
and ownership transfers on the blockchain. Buyers can access this information to make informed
decisions when purchasing a pre-owned vehicle.
In summary, BT offers a reliable and immutable method for maintaining a verifiable record of a
product's journey, enhancing its credibility and traceability across various sectors. The Italian coffee
roaster Lavazza's successful implementation of blockchain for product tracking demonstrates the
importance of collaborative supply chain efforts and innovation in adapting to socioeconomic trends
(Gazzola et al., 2023).
_2.5.13_ _Sustainable production and consumption_
Sustainable consumption refers to the use of products and services that satisfy basic needs and
improve quality of life while minimizing the impact on the environment, so future generations can
also fulfill their needs. Chaudhuri et al. (2023)’ study highlights the importance of customer
education and engagement, along with cultivating local partnerships, as essential behavioral
strategies for enhancing social sustainability and mitigating risks in the context of BT.
The utilization of BT has a profound influence on the promotion of sustainable production and
consumption (Böckel et al., 2021) through the augmentation of transparency, traceability, and
accountability within supply chains across diverse sectors (Khanfar et al., 2021). BT enables
consumers to achieve supply chain transparency by providing them with the ability to trace the
origins of items, thereby certifying their validity and ensuring ethical sourcing practices. The
promotion of openness fosters the adoption of sustainable practices, including the responsible
-----
**_ISSN 2281-3993_** **_www.richtmann.org_** **_March 2024_**
management of resources and the establishment of fair labor conditions. The immutability of BT
presents a formidable obstacle to the proliferation of counterfeit goods in markets, thereby
guaranteeing that consumers are provided with authentic, secure, and ecologically sustainable
products.
_2.5.14_ _Renewable energy adoption_
Facilitating peer-to-peer energy trading, blockchain projects such as the Brooklyn Microgrid enable
consumers to trade locally produced renewable energy. This democratization of energy production
promotes environmental sustainability and incentivizes the shift towards renewable energy sources,
contributing to reduced carbon footprints and supporting clean energy goals.
_2.5.15_ _Humanitarian aid and disaster relief_
In the realm of humanitarian aid, blockchain improves the efficiency and transparency of aid
distribution. The World Food Programme's Building Blocks project exemplifies how blockchain
delivers food assistance directly to beneficiaries, minimizing transaction costs and fraud risks, thus
ensuring that aid reaches those in need more effectively (Martin et al., 2011).
_2.5.16Business ethics and effective corporate governance_
Business ethics and effective corporate governance are instrumental in achieving social sustainability
(Ronaghi and Mosakhani, 2023). Business ethics, fundamental to workplace social interactions, has a
significant impact on an organization's social aspect (Lashley, 2016).
Blockchain introduces unparalleled transparency in global supply chains, enabling verification
of ethical sourcing and adherence to fair labor practices. By documenting the journey of goods from
their origin, initiatives like Everledger and Fairfood International leverage blockchain to combat the
trade in conflict minerals and ensure products are produced under ethical conditions, empowering
consumers with information to make responsible choices (Tang, 2018).
Krishna et al. (2011) identified a positive relationship between business ethics and corporate
performance. Other authors have examined corporate governance, influencing organizational social
behaviors through stakeholder monitoring and power structures (Schultz et al., 2020; Aguilera et al.,
2009), and the interaction between sustainability concepts and corporate governance concerning
corporate performance (Krechovská and Prochazkova, 2014).
_2.6_ _Challenges of BT implementation and proposed solutions in sustainability programs_
In the review of scholarly literature, it has been established that the incorporation of blockchain
technology (BT) into sustainability initiatives is not without its challenges. These obstacles are
extensively examined in the works of Mulligan et al. (2023) and Khanfar et al. (2021), among others,
and warrant careful consideration in the discourse on the advancement of BT within sustainability
programs:
## • Scalability and security issues: Scalability is a critical technical challenge for blockchain, as
increased transaction volumes can reduce system performance and efficiency. Security is
also a concern; despite inherent protections, blockchain is susceptible to cyber threats like
the 51% attack, and new vulnerabilities may emerge as the technology evolves.
## • Energy consumption: Blockchain networks, especially those using proof-of-work (PoW)
mechanisms, require significant computational resources, leading to high energy
consumption, as seen in Bitcoin mining. This poses environmental concerns, particularly
regarding sustainability objectives aimed at reducing carbon emissions.
-----
**_ISSN 2281-3993_** **_www.richtmann.org_** **_March 2024_**
## • Regulatory and ethical considerations: The decentralized and transnational nature of
blockchain creates complex regulatory challenges. Compliance with varying international
regulations on digital currencies, data protection, and cross-border transactions is crucial.
Ethical issues, such as data privacy and the potential for blockchain to enable illicit
activities, also need to be addressed.
## • Resistance to change and adoption hurdles: Resistance to adopting new technologies like
blockchain is common, often due to limited understanding or concerns about potential
impacts. Integrating blockchain into existing systems can be complex and costly, posing
significant challenges for organizations, especially smaller ones with limited resources.
## • Economic implications: BT's impact on employment and income distribution is complex.
While it can offer new job opportunities, financial inclusivity, and skill development, it also
has the potential to disrupt traditional employment structures and exacerbate income
inequalities. Addressing these issues requires regulatory frameworks, educational initiatives,
and industrial adjustments to ensure equitable benefits distribution.
The successful and sustainable adoption of BT requires overcoming these challenges through
the development of more energy-efficient consensus mechanisms, robust security measures, effective
regulatory management, and stakeholder education about the technology's benefits and
implementation. Table 5 presents comprehensive strategies required to address the multifaceted
challenges of integrating blockchain technology into sustainability initiatives. These solutions span
technical innovations, regulatory adjustments, ethical considerations, educational efforts, and
economic policies, highlighting the need for collaborative efforts among all stakeholders to harness
blockchain technology's full potential for sustainable development.
**Table 5. Strategies of integrating blockchain technology into sustainability initiatives**
**Challenge Category** **Solution** **Description**
|Challenge Category|Solution|Description|
|---|---|---|
|Scalability and security|Layered architecture and off-chain solutions|Investigating off-chain solutions and implementing a layered blockchain architecture can improve scalability by offloading transaction processing, reducing congestion and increasing efficiency while maintaining security.|
||Sophisticated consensus mechanisms|Implementing streamlined consensus mechanisms like PoS or DPoS to reduce computational and energy demands, addressing scalability and security concerns simultaneously.|
|Energy utilization|Transition to energy- efficient consensus mechanisms|Moving from PoW to more energy-efficient mechanisms like PoS to significantly reduce blockchain's energy consumption and advance sustainability goals.|
||Implementation of renewable energy sources|Promoting the use of renewable energy sources in blockchain operations through regulations and incentives to minimize environmental impact.|
|Ethical and regulatory considerations|Frameworks for international regulations|Developing harmonized international frameworks to manage the decentralized nature of blockchain, ensuring privacy, data protection, and prevention of illegal activities.|
||Establishing ethical standards and guidelines|Setting up ethical guidelines and standards for blockchain applications to uphold user privacy and contribute positively to societal goals.|
|Resistance to change and adoption|Capacity building and education|Providing educational initiatives and resources on blockchain to demystify the technology and reduce opposition, emphasizing the importance of training for developers, users, and stakeholders.|
||Collaborations and pilot projects|Implementing pilot projects and fostering collaborations between industries, governments, NGOs, and blockchain developers to demonstrate the practical benefits and feasibility of blockchain, reducing integration complexity and costs.|
|Economic consequences|Inclusive economic policies|Developing policies that support financial inclusivity and SMEs to mitigate adverse effects on employment and income distribution, ensuring equitable benefits from blockchain technology.|
||Programs for skill development and job transition|Investing in skill development and retraining programs to prepare the workforce for new opportunities in the blockchain sector, positively impacting employment and income distribution.|
-----
**_ISSN 2281-3993_** **_www.richtmann.org_** **_March 2024_**
_2.7_ _Framework for BT impact on environment, economic, and social domains_
This framework outlines the multifaceted impact of blockchain technology across environmental,
economic, and social domains. For the environment, blockchain aids in monitoring carbon footprints
and enhancing renewable energy markets, among other benefits. Economically, it supports fair trade,
improves supply chain management, and promotes financial inclusion, thereby driving efficiency.
Socially, blockchain ensures product authenticity and traceability, bolsters business ethics, and
contributes to sustainable production practices. The framework also identifies challenges such as
scalability and regulatory compliance, while highlighting blockchain's sustainable features like
transparency and decentralization.
**Figure 2. BT impact framework**
**Source: Authors’ own work**
**3.** **The Future Outlook of Sustainability Blockchain Integration**
The trajectory of blockchain integration in sustainability shows promise across technological
advancements, policy considerations, and economic impacts. Innovations in BT are expected to
address issues like energy consumption and scalability, such as transitioning from Proof of Work
(PoW) to more sustainable mechanisms like Proof of Stake (PoS) or Proof of Authority (PoA),
reducing blockchain's ecological impact.
Blockchain's application across various industries could enhance supply chain transparency,
optimize renewable energy distribution, and facilitate resource management, with smart contracts
and decentralized applications enforcing sustainability standards. Governments play a critical role in
shaping the future of blockchain for sustainability, requiring legal frameworks that balance
innovation with potential risks, including data privacy and financial stability concerns. Policymaking
might incentivize blockchain adoption in sustainable practices, such as subsidies for companies using
blockchain for sustainable supply chain tracing or legislative support for blockchain-based renewable
energy solutions (Mulligan et al., 2023).
International collaboration is essential, given blockchain's global nature and sustainability
challenges. Harmonizing regulatory approaches can improve cross-border blockchain initiatives,
contributing to global sustainability goals. The economic impact of blockchain in various sectors
-----
**_ISSN 2281-3993_** **_www.richtmann.org_** **_March 2024_**
could be significant, enabling new markets and opportunities, especially in sustainable goods and
services. Blockchain's potential to disrupt market structures, reduce intermediaries, lower transaction
costs, and democratize market access could redistribute economic power. It may also influence
investment patterns, as its ability to verify and track sustainable practices could attract more
investments into sustainable projects, impacting capital flows in global markets.
In conclusion, blockchain's convergence with sustainability is a dynamic area with the potential
for significant advancements in technology, policy, and economics. Despite challenges, the ongoing
development of blockchain, supportive government policies, and its growing economic influence
could advance global sustainability efforts.
**4.** **Conclusion and Future Research Directions**
This article offers an in-depth analysis of blockchain technology's potential to significantly contribute
to sustainable development across environmental, economic, and social dimensions. It meticulously
explores how BT can act as a transformative tool, addressing global sustainability challenges by
enhancing transparency, promoting resource efficiency, and facilitating equitable economic growth.
One of the key topics discussed in this paper is how BT can serve as a powerful enabler for
achieving sustainability goals by ensuring the traceability and authenticity of products, which is
crucial for environmental stewardship and social justice. Through various applications, such as in
supply chain management, renewable energy sectors, and conservation efforts, BT has the potential
to reduce carbon emissions, improve resource allocation, and support sustainable business practices.
Furthermore, BT fosters a unique synergy among environmental, economic, and social
dimensions by enabling transparent, secure, and efficient operations across various sectors. Its
immutable ledger ensures the traceability of products, promoting environmental sustainability
through the verification of ethical sourcing and waste reduction practices. Economically, blockchain
reduces operational costs by streamlining transactions and eliminating intermediaries, while also
providing financial inclusion for underserved populations through decentralized financial services.
Socially, the technology enhances transparency and trust among consumers, businesses, and
communities, supporting fair labor practices and equitable resource distribution. This multifaceted
impact not only encourages responsible consumption and production but also empowers individuals
and communities by democratizing access to resources and services. By addressing these pillars
simultaneously, BT creates a holistic approach to sustainable development, aligning with global
efforts to achieve the UN-SDGs. Its application across industries represents a transformative shift
towards a more sustainable, equitable, and interconnected world, showcasing the potential of
technology to resolve complex global challenges harmoniously.
The paper also highlights the importance of addressing the challenges and limitations
associated with the implementation of BT for sustainable solutions, including concerns over
scalability, energy consumption, and regulatory barriers. The authors emphasize the need for
strategic recommendations and a comprehensive approach that balances the opportunities and
obstacles of using BT in sustainability efforts.
In conclusion, this research underscores the transformative potential of BT in promoting
sustainable development. It calls for a collaborative effort among stakeholders, including
policymakers, practitioners, and researchers, to leverage BT effectively. By addressing the identified
challenges and harnessing the capabilities of BT, there is a significant opportunity to advance towards
a more sustainable, equitable, and environmentally friendly future. The findings of this paper
contribute to both theoretical understanding and practical implementation of blockchain technology
as a catalyst for sustainable development, offering guidance and insights for future research and
policy-making in this evolving domain.
However, to fully harness the sustainability potential of blockchain, further technological
advancements are necessary to address and minimize its environmental impact. In anticipation of
future investigations, the forthcoming study must direct its attention toward a few crucial domains:
-----
**_ISSN 2281-3993_** **_www.richtmann.org_** **_March 2024_**
- Explore and advance blockchain technologies characterized by reduced energy consumption
to address environmental issues related to energy efficiency.
- The establishment of regulatory frameworks is crucial for the effective governance of BT,
ensuring its deployment is in line with global sustainability objectives. These frameworks
encompass the development of comprehensive rules and regulatory guidelines.
- Investigate scaling solutions that may effectively manage higher transaction volumes while
maintaining optimal levels of energy efficiency and security.
- Investigate the cross-sector applications of BT, with a specific focus on its potential
contributions to sustainable practices in non-traditional industries such as agriculture,
healthcare, and public governance.
- Evaluate the wider socio-economic implications of BT in facilitating sustainability,
specifically examining its influence on market dynamics and global trade.
- Finally, the utilization of BT shows great potential in promoting sustainability objectives.
However, it is crucial to exercise prudent oversight in its implementation to guarantee a
favorable contribution to the overall ecosystem. Ongoing advancements in innovation,
research, and policy formulation will play a crucial role in effectively leveraging this
technology to achieve a more sustainable future.
**References**
Aerial (2023). Available at: https://aerial.ai Accessed: November 16, 2023.
Aguilera, R. C., Ortiz, M. P., Ortiz, J. P., & Banda, A. A. (2021). Internet of things expert system for smart cities
using the blockchain technology. Fractals, 29(01), 2150036.
Ahmad, M. (2023). Top 10: Energy Companies Using Blockchain Technology. Energy. Available at:
https://energydigital.com/top10/top-10-energy-companies-using-blockchain-technology. Accessed:
November 16, 2023
Alhasan, H., Hamdan, A. (2023). Blockchain Technology and Environmental Sustainability. In: El Khoury, R.,
Nasrallah, N. (eds) Emerging Trends and Innovation in Business and Finance. Contributions to
Management Science. Springer, Singapore.
AL-Issa, N., Thanasi-Boçe, M., & Ali, O. (2022). Boosting Luxury Sustainability Through Blockchain Technology.
In Blockchain Technologies in the Textile and Fashion Industry(pp. 17-46). Singapore: Springer Nature
Singapore.
Allena, M. (2020). Blockchain technology for environmental compliance: Towards a 'choral' approach.
Environmental Law Review, 50(4). Retrieved from
Arshad, A., Shahzad, F., Rehman, I. U., and Sergi, B. S. (2023). A systematic literature review of blockchain
technology and environmental sustainability: Status quo and future research. International Review of
_Economics & Finance._
Bai, Y., Hu, Q., Seo, S. -H., Kang, K., and Lee, J. J. (2022). Public participation consortium blockchain for smart city
governance. IEEE Internet of Things Journal, 9(3), 2094-2108.
Balzarova, M., Dyer, C., and Falta, M. (2022). Perceptions of blockchain readiness for fairtrade
programmes. Technological Forecasting and Social Change, 185, 122086.
Bhusal, Chandra Sekhar, Blockchain Technology in Agriculture: A Case Study of Blockchain Start-Up Companies
(October 1, 2021). International Journal of Computer Science and Information Technology, 13(5).
Böckel, A., Nuzum, A. K. and Weissbrod, I. (2021). Blockchain for the circular economy: analysis of the researchpractice gap. Sustainable Production and Consumption, 25, 525-539.
Bosona, T. and Gebresenbet, G. (2023). The role of blockchain technology in promoting traceability systems in
agri-food production and supply chains. Sensors, 23(11), 5342.
Brilliantova, V., and Thurner, T. W. (2019). Blockchain and the future of energy. Technology in Society, 57, 38-45.
Carter, C. R. and Easton, P. L. (2011). Sustainable supply chain management: evolution and future directions.
_International Journal of Physical Distribution & Logistics Managemen.,41(1), 46–62._
Carter, C. R., and Rogers, D. S. (2008). A framework of sustainable supply chain management: moving toward new
theory. International journal of physical distribution & logistics management, 38(5), 360-387.
Clark, W. C. (2007). Sustainability science: a room of its own. _Proceedings of the National Academy of Sciences,_
_104(6), 1737–1738._
-----
**_ISSN 2281-3993_** **_www.richtmann.org_** **_March 2024_**
Chen, W., Zhou, K., Fang, W., Wang, K., Bi, F., & Assefa, B. (2020). Review on blockchain technology and its
application to the simple analysis of intellectual property protection. International Journal of Computational
_Science and Engineering, 22(4), 437-444._
Cozzio, C., Viglia, G., Lemarie, L., and Cerutti, S. (2023). Toward an integration of blockchain technology in the
food supply chain. Journal of Business Research, 162, 113909.
Cui, M.-l., Feng, T.-t., and Wang, H.-r. (2023). How can blockchain be integrated into renewable energy? --A
bibliometric-based analysis. Energy Strategy Reviews, 50, 101207.
Dal Mas, F., Dicuonzo, G., Massaro, M., & Dell'Atti, V. (2020). Smart contracts to enable sustainable business
models. A case study. Management Decision, 58(8), 1601-1619.
Dal Mas, F., Massaro, M., Ndou, V., and Raguseo, E. (2023). Blockchain technologies for sustainability in the
agrifood sector: A literature review of academic research and business perspectives. _Technological_
_Forecasting and Social Change, 187, 122155._
Dehshiri, S. J. H., and Amiri, M. (2024). Evaluation of blockchain implementation solutions in the sustainable
supply chain: A novel hybrid decision approach based on Z-numbers. Expert Systems with Applications, 235,
121123.
Everledger (2023) Making the commercial case for blockchain diamond tracking. Available at:
https://everledger.io/making-the-commercial-case-for-blockchain-diamond-tracking/ Accessed: November
16, 2023
Fallah Shayan, N., Mohabbati-Kalejahi, N., Alavi, S., and Zahed, M. A. (2022). Sustainable development goals
(SDGs) as a framework for corporate social responsibility (CSR). Sustainability, 14(3), 1222.
Tian, F. (2016). An agri-food supply chain traceability system for China based on RFID & blockchain technology.
In: 13th International Conference on Service Systems and Service Management, ICSSSM 2016. IEEE, June
2016, pp. 1–6.
Fleener, M. J. (2022). Blockchain Technologies: A Study of the Future of Education. Journal of Higher Education
_Theory & Practice, 22(1)._
Fu, Z., Dong, P., Li, S., Ju, Y., and Liu, H. (2021). How blockchain renovate the electric vehicle charging services in
the urban area? A case study of Shanghai, China. Journal of Cleaner Production, 315, 128172
Gazzola, P., Pavione, E., Barge, A., and Fassio, F. (2023). Using the Transparency of Supply Chain Powered by
Blockchain to Improve Sustainability Relationships with Stakeholders in the Food Sector: The Case Study of
Lavazza. Sustainability, 15(10), 7884.
Hahn, T. and Figge, F. (2018). Why architecture does not matter: on the fallacy of sustainability balanced
scorecards. Journal of Business Ethics, _150(4), 919–935._
Howson, P. (2020). Building trust and equity in marine conservation and fisheries supply chain management with
blockchain. Marine Policy, 115, 103873.
Jabbar, K. and Bjørn, P. (2018). Infrastructural Grind”. In: Proceedings of the 2018 ACM Conference on Supporting
Groupwork - GROUP ’18. New York, NY, USA: ACM, 2018, pp. 297–308. ISBN: 9781450355629. DOI:
10.1145/3148330.3148345.
Jiang, Y., and Zheng, W. (2021). Coupling mechanism of green building industry innovation ecosystem based on
blockchain smart city. Journal of Cleaner Production, 307, 126766.
Kapferer, J. N., and Michaut-Denizeau, A. (2019). Are millennials really more sensitive to sustainable luxury? A
cross-generational international comparison of sustainability consciousness when buying luxury. Journal of
_Brand Management, 1-13._
Khan, K., Hassan, F., Koo, J., Mohammed, M. A., Hasan, Y., Muhammad, D., ... and Qureshi, N. M. F. (2023).
Blockchain-based applications and energy effective electric vehicle charging–A systematic literature review,
challenges, comparative analysis and opportunities. Computers and Electrical Engineering, 112, 108959.
Khanfar, A.A., Iranmanesh, M., Ghobakhloo, M., Senali, M. G., and Fathi, M. (2021). Applications of blockchain
technology in sustainable manufacturing and supply chain management: A systematic review. Sustaina
_bility, 13(14), 7870._
Kim, H. M., Bock, G. W., and Lee, G. (2021). Predicting Ethereum prices with machine learning based on
Blockchain information. Expert Systems with Applications, 184, 115480.
Kouhizadeh, M., Saberi, S., and Sarkis, J. (2021). Blockchain technology and the sustainable supply chain:
Theoretically exploring adoption barriers. International journal of production economics, 231, 107831.
Konashevych, O. (2020). Constraints and benefits of the blockchain use for real estate and property rights. Journal
_of Property, Planning and Environmental Law,_ _12(2), 109-127._
Krechovská, M., and Procházková, P. T. (2014). Sustainability and its integration into corporate governance
focusing on corporate performance management and reporting. Procedia Engineering, 69, 1144-1151.
-----
**_ISSN 2281-3993_** **_www.richtmann.org_** **_March 2024_**
Krishna, A., Dangayach, G. S., and Jainabc, R. (2011). Business ethics: A sustainability approach. Procedia-Social
_and Behavioral Sciences, 25, 281-286._
Lashley, C. (2016). Liberating wage slaves: Towards sustainable employment practices. In The Routledge handbook
_of hospitality studies (pp. 389-400). Routledge._
Luna, M., Fernandez-Vazquez, S., Tereñes Castelao, E., and Arias Fernández, Á. (2024). A blockchain-based
approach to the challenges of EU’s environmental policy compliance in aquaculture: From traceability to
fraud prevention. Marine Policy, 159, 105892.
Lund, E. H., Jaccheri, L., Li, J., Cico, O., & Bai, X. (2019, May). Blockchain and sustainability: A systematic mapping
study. In 2019 IEEE/ACM 2nd International Workshop on Emerging Trends in Software Engineering for
_Blockchain (WETSEB) (pp. 16-23). IEEE._
Mani, V., Gunasekaran, A., Papadopoulos, T., Dubey, R., and Benjamin, H. (2016). Supply chain social
sustainability for developing nations: evidence from India. Resources, Conservation and Recycling., _111, 42–_
52.
Mirabelli, G., and Solina, V. (2020). Blockchain and agricultural supply chains traceability: Research trends and
future challenges. Procedia Manufacturing, 42, 414-421.
Mulligan, C., Morsfield, S., and Cheikosman, E. (2023). Blockchain for sustainability: A systematic literature review
for policy impact. Telecommunications Policy, 102676.
Naqash MT, Syed TA, Alqahtani SS, Siddiqui MS, Alzahrani A., and Nauman M.A (2023). Blockchain Based
Framework for Efficient Water Management and Leakage Detection in Urban Areas. Urban Science. 7(4), 99.
Nir Kshetri, N. (2018). Blockchain’s roles in meeting key supply chain management objectives. _International_
_Journal of Information Management_ _39(Apr. 2018), 80–89._
Novak, M. (2019). The Implications of Blockchain for Income Inequality Between Science and
EconomicsBlockchain Economics: Implications of Distributed Ledgers, pp. 235-250
Olivier, J. G., Schure, K. M., & Peters, J. A. H. W. (2017). Trends in global CO2 and total greenhouse gas
emissions. PBL Netherlands Environmental Assessment Agency, 5, 1-11.
Palacio, N. S. (2018). Blockchain: A technological tool for sustainable development or a massive energy
consumption network. Διαθέσιμο στο: https://www. revistabionatura. com/files/2018.03, 4.
Pandey, V., Pant, M., and Snasel, V. (2022). Blockchain technology in food supply chains: Review and bibliometric
analysis. Technology in Society, 69, 101954.
Park, A., & Li, H. (2021). The effect of blockchain technology on supply chain sustainability
performances. Sustainability, 13(4), 1726.
Parmentola, A., Petrillo, A., Tutore, I., and De Felice, F. (2022). Is blockchain able to enhance environmental
sustainability? A systematic review and research agenda from the perspective of Sustainable Development
Goals (SDGs). Business Strategy and the Environment, 31(1), 194-217.
Prashar, D., Jha, N., Jha, S., Lee, Y., and Joshi, G. P. (2020). Blockchain-based traceability and visibility for
agricultural products: A decentralized way of ensuring food safety in india. Sustainability, 12(8), 3497.
Purvis, B., Mao, Y., & Robinson, D. (2019). Three pillars of sustainability: in search of conceptual
origins. Sustainability science, 14, 681-695.
Richardson, A., and Xu, J. (2020). Carbon Trading with Blockchain. In P. Pardalos, I. Kotsireas, Y. Guo, and W.
Knottenbelt (Eds.), Mathematical Research for Blockchain Economy (pp. 105–124).
Risso, L. A., Ganga, G. M. D., Godinho Filho, M., de Santa-Eulalia, L. A., Chikhi, T., and Mosconi, E. (2023). Present
and future perspectives of blockchain in supply chain management: a review of reviews and research
agenda. Computers and Industrial Engineering, 109195.
Rodríguez Furones, A., and Tejero Monzón, J. I. (2023). Blockchain applicability in the management of urban
water supply and sanitation systems in Spain. Journal of Environmental Management, 344, 118480.
Sarfraz, M., Khawaja, K. F., Han, H., Ariza-Montes, A., & Arjona-Fuentes, J. M. (2023). Sustainable supply chain,
digital transformation, and blockchain technology adoption in the tourism sector. Humanities and Social
_Sciences Communications, 10(1), 1-13._
Sikder, A. S. (2023). Blockchain-Empowered E-commerce: Redefining Trust, Security, and Efficiency in Digital
Marketplaces in the Context of Bangladesh.: Blockchain-Empowered E-commerce. International Journal of
_Imminent Science & Technology, 1(1), 216-235._
Shabaltina, L. V., Madiyarova, D. M., & Tamer, N. (2021). Generating Employment with Equipped and Trained
Workforce Using Blockchain Technology. In Modern Global Economic System: Evolutional Development vs.
_Revolutionary Leap 11 (pp. 431-440). Springer International Publishing._
Schulz, K. A., Gstrein, O. J., and Zwitter, A. J. (2020). Exploring the governance and implementation of sustainable
development initiatives through blockchain technology. Futures, 122, 102611.
-----
**_ISSN 2281-3993_** **_www.richtmann.org_** **_March 2024_**
Singh, A. K., Kumar, V. R. P., Dehdasht, G., Mohandes, S. R., Manu, P., and Pour Rahimian, F. (2023). Investigating
the barriers to the adoption of blockchain technology in sustainable construction projects. Journal of
Cleaner Production, 403, 136840.
Sipthorpe, A., Brink, S., Van Leeuwen, T., and Staffell, I. (2022). Blockchain solutions for carbon markets are
nearing maturity. One Earth, 5(7), 779-791.
Tafuro, A., Dammacco, G., and Costa, A. (2023). A Conceptual Study on the Role of Blockchain in Sustainable
Development of Public–Private Partnership. Administrative Sciences, 13(8), 175.
Tartan, C. (2023). A Blockchain-Based Welfare Distribution Model for Digital Inclusivity. REGION, 10(1), 19-44.
Teufel, B., Sentic, A., and Barmet, M. (2019). Blockchain energy: Blockchain in future energy systems. Journal of
_Electronic Science and Technology, 17(4), 100011._
Thakur, T., Mehra, A., Hassija, V., Chamola, V., Srinivas, R., Gupta, K. K., and Singh, A. P. (2021). Smart water
conservation through a machine learning and blockchain-enabled decentralized edge computing network.
_Applied Soft Computing, 106, 107274._
Thanasi-Boçe, M., AL-Issa, N., and Ali, O. (2022). Combating Luxury Counterfeiting Through Blockchain
Technology. In Blockchain Technologies in the Textile and Fashion Industry (pp. 1-16). Singapore: Springer
Nature Singapore.
Thoughtworks (2023). VAKT_Building the world's first enterprise-level blockchain platform. Available at:
https://www.thoughtworks.com/en-au/clients/vakt Accessed: November 16, 2023
Truby, J., Brown, R. D., Dahdal, A., and Ibrahim, I. (2022). Blockchain, climate damage, and death: Policy
interventions to reduce the carbon emissions, mortality, and net-zero implications of non-fungible tokens
and Bitcoin. Energy Research & Social Science, 88, 102499.
Tyagi, K. A global blockchain-based agro-food value chain to facilitate trade and sustainable blocks of healthy lives
and food for all. Humanit Soc Sci Commun 10, 196 (2023). https://doi.org/10.1057/s41599-023-01658-2
United Nations (2023) 17 SDGs goals. Available at: https://sdgs.un.org/goals
Venkatesh, V. G., Kang, K., Wang, B., Zhong, R. Y., and Zhang, A. (2020). System architecture for blockchain based
transparency of supply chain social sustainability. _Robotics and Computer-Integrated Manufacturing, 63,_
101896.
Wang, J. (2022). A novel electric vehicle charging chain design based on blockchain technology. Energy Reports,
8(Supplement 4), 785-793.
Waqar, A., Khan, A. M., and Othman, I. (2023). Blockchain empowerment in construction supply chains:
Enhancing efficiency and sustainability for an infrastructure development. _Journal of Infrastructure_
_Intelligence and Resilience._
Yang, Z., Zhu, C., Zhu, Y., and Li, X. (2023). Blockchain technology in building environmental sustainability: A
systematic literature review and future perspectives. Building and Environment, 245, 110970.
Yap, K., Chin, H.H., and Klemeš, J.J. (2023). Blockchain technology for distributed generation: A review of current
development, challenges and future prospect. Renewable and Sustainable Energy Reviews.
Yontar, E. (2023). The role of blockchain technology in the sustainability of supply chain management: Grey based
dematel implementation. Cleaner Logistics and Supply Chain, 8, 100113.
Zuo, Y. (2022). Tokenizing renewable energy certificates (RECs)—A blockchain approach for REC issuance and
trading. IEEE Access, 10, 134477-134490.
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.36941/ajis-2024-0041?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.36941/ajis-2024-0041, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBYNC",
"status": "HYBRID",
"url": "https://www.richtmann.org/journal/index.php/ajis/article/download/13698/13257"
}
| 2,024
|
[
"JournalArticle",
"Review"
] | true
| 2024-03-05T00:00:00
|
[
{
"paperId": "83c5079a34d7b9819803f07bb7671ccab4635de1",
"title": "A blockchain-based approach to the challenges of EU’s environmental policy compliance in aquaculture: From traceability to fraud prevention"
},
{
"paperId": "6b257a7ff1605e0c2f34cac31fabc120c9d1564d",
"title": "Blockchain-based applications and energy effective electric vehicle charging - A systematic literature review, challenges, comparative analysis and opportunities"
},
{
"paperId": "5775ed1a34ba38c1449241d0e751c5d7021dbcf4",
"title": "How can blockchain be integrated into renewable energy? --A bibliometric-based analysis"
},
{
"paperId": "144d7b8ab7ca3b083fcb5775e64e9324779bf2c9",
"title": "Blockchain empowerment in construction supply chains: Enhancing efficiency and sustainability for an infrastructure development"
},
{
"paperId": "86e9a896350472b9c046dae02100324214c1b218",
"title": "Blockchain for sustainability: A systematic literature review for policy impact"
},
{
"paperId": "59aad4b562030c0f33f8886f81b29ba7a33a4463",
"title": "Blockchain technology in building environmental sustainability: A systematic literature review and future perspectives"
},
{
"paperId": "a74bfadd545e542a89da2203df07d74cc5f99da0",
"title": "A Blockchain Based Framework for Efficient Water Management and Leakage Detection in Urban Areas"
},
{
"paperId": "6495ccdaf44d49a1912e2725f145f0dd6fb16c2f",
"title": "Sustainable supply chain, digital transformation, and blockchain technology adoption in the tourism sector"
},
{
"paperId": "c215e6685bbe18c59fbe4d6bbde4fb71efc69325",
"title": "The role of blockchain technology in the sustainability of supply chain management: Grey based dematel implementation"
},
{
"paperId": "a0dc8d1c8c3f2e54340930553fe63106bc6a7172",
"title": "Evaluation of blockchain implementation solutions in the sustainable supply chain: A novel hybrid decision approach based on Z-numbers"
},
{
"paperId": "8ac15c9704ac8879438456140a3e8386baf7ef6e",
"title": "A Conceptual Study on the Role of Blockchain in Sustainable Development of Public–Private Partnership"
},
{
"paperId": "4a234597889335155fa1541f82314593a91b417c",
"title": "Blockchain applicability in the management of urban water supply and sanitation systems in Spain."
},
{
"paperId": "cea332126c771539cd44d5ef676785b62e566450",
"title": "A systematic literature review of blockchain technology and environmental sustainability: Status quo and future research"
},
{
"paperId": "2cc63bbcf11aa7b82860df88a98183e4c168c089",
"title": "Toward an integration of blockchain technology in the food supply chain"
},
{
"paperId": "c4654c462e613b0dfadad5046817037e86bb0e47",
"title": "The Role of Blockchain Technology in Promoting Traceability Systems in Agri-Food Production and Supply Chains"
},
{
"paperId": "58ea95274cd9ef4274722c791487cd528f16c1fb",
"title": "Using the Transparency of Supply Chain Powered by Blockchain to Improve Sustainability Relationships with Stakeholders in the Food Sector: The Case Study of Lavazza"
},
{
"paperId": "893c512a0f62537c56eeb2d6fe7ffec8d6137d25",
"title": "A global blockchain-based agro-food value chain to facilitate trade and sustainable blocks of healthy lives and food for all"
},
{
"paperId": "2891a2bbb6d32b9e360017b2b5e68c25eb8afde9",
"title": "Blockchain technology for distributed generation: A review of current development, challenges and future prospect"
},
{
"paperId": "4fb6300f10dc4ffd4ad4130f3a7eb36f044b6ce5",
"title": "Blockchain-Based Welfare Distribution Model for Digital Inclusivity"
},
{
"paperId": "9b8f14a8c0a64ca6539325c788e4db2b43412167",
"title": "Present and future perspectives of blockchain in supply chain management: a review of reviews and research agenda"
},
{
"paperId": "7f774056ca253c59d7366e0efde9e395c9ed0813",
"title": "Investigating the barriers to the adoption of blockchain technology in sustainable construction projects"
},
{
"paperId": "6757dfb563758c98263828dabf2e32fb5c17e91f",
"title": "Blockchain technologies for sustainability in the agrifood sector: A literature review of academic research and business perspectives"
},
{
"paperId": "b4a6d4a81fa9c5dcc5793d75027329ec6eaac4da",
"title": "Perceptions of Blockchain Readiness for Fairtrade Programmes"
},
{
"paperId": "427295959a43638eb0dd814e6efaa94ed5de979e",
"title": "Blockchain solutions for carbon markets are nearing maturity"
},
{
"paperId": "c2e710b26cf2883f0bbbdacf476176722d6bf6b5",
"title": "Blockchain, climate damage, and death: Policy interventions to reduce the carbon emissions, mortality, and net-zero implications of non-fungible tokens and Bitcoin"
},
{
"paperId": "f50aeab4fda4629387efce3411f00191a3e932cc",
"title": "Blockchain technology in food supply chains: Review and bibliometric analysis"
},
{
"paperId": "515602084236ba2a2e28226f77a01547ca47ff00",
"title": "Blockchain Technologies: A Study of the Future of Education"
},
{
"paperId": "9c2f3c2b979f9f6d7535414ee627af5e5d24dd41",
"title": "Public Participation Consortium Blockchain for Smart City Governance"
},
{
"paperId": "12fb72aa41e95d7a5ec04a5c6fdb795871983b33",
"title": "Sustainable Development Goals (SDGs) as a Framework for Corporate Social Responsibility (CSR)"
},
{
"paperId": "2ce03b0d6a0ce6efc4b612f1a4a22f9f486b71c3",
"title": "Predicting Ethereum prices with machine learning based on Blockchain information"
},
{
"paperId": "39c01c7bd740c59c923b125536532dbdd755261f",
"title": "Blockchain Technology in Agriculture: A Case Study of Blockchain Start-Up Companies"
},
{
"paperId": "359f561a61da65defadffc4807a89f0858c39897",
"title": "How blockchain renovate the electric vehicle charging services in the urban area? A case study of Shanghai, China"
},
{
"paperId": "6502e646b5be0db1f0b265a39052a0aa2071e4de",
"title": "Is blockchain able to enhance environmental sustainability? A systematic review and research agenda from the perspective of Sustainable Development Goals (SDGs)"
},
{
"paperId": "9787abbfc6ee3aa67cd9f801364229058ffe798c",
"title": "Applications of Blockchain Technology in Sustainable Manufacturing and Supply Chain Management: A Systematic Review"
},
{
"paperId": "ad5f8c9523df448e3af8633defc89f48f0db3cb7",
"title": "Coupling mechanism of green building industry innovation ecosystem based on blockchain smart city"
},
{
"paperId": "f4ca231b52e3be27e9097d9f0e71e15aa39046e1",
"title": "Smart water conservation through a machine learning and blockchain-enabled decentralized edge computing network"
},
{
"paperId": "227c7263c0a65589708ecc429b1e198931271841",
"title": "The Effect of Blockchain Technology on Supply Chain Sustainability Performances"
},
{
"paperId": "979a7afae01dd40825f0c7aecaf6ec2391d22cad",
"title": "Blockchain technology and the sustainable supply chain: Theoretically exploring adoption barriers"
},
{
"paperId": "4a6361c924191e919b322bb7f7cc88d364d40ad0",
"title": "Blockchain for the Circular Economy: Analysis of the Research-Practice Gap"
},
{
"paperId": "9e4d761cd4c68647ec93960cd2c83fe2a96ca4aa",
"title": "INTERNET OF THINGS EXPERT SYSTEM FOR SMART CITIES USING THE BLOCKCHAIN TECHNOLOGY"
},
{
"paperId": "60ec11beba5f3f63409affe028a46256e22acb68",
"title": "Review on blockchain technology and its application to the simple analysis of intellectual property protection"
},
{
"paperId": "573ea4dca564627225b4ecdcd70c1a1c155d15cb",
"title": "Exploring the Governance and Implementation of Sustainable Development Initiatives through Blockchain Technology"
},
{
"paperId": "41aa55e38496067febbbfcb8437687b2cb6f581b",
"title": "Smart contracts to enable sustainable business models. A case study"
},
{
"paperId": "599a846955e74dedb410e5dab835e3773580612e",
"title": "System architecture for blockchain based transparency of supply chain social sustainability"
},
{
"paperId": "1c18e533121100985c53fb79420f09f2df2857ab",
"title": "Carbon Trading with Blockchain"
},
{
"paperId": "7278575118b28164a20d6250d5be296b778dcbf7",
"title": "Building trust and equity in marine conservation and fisheries supply chain management with blockchain"
},
{
"paperId": "b45a3e6a7afd4ac786ffafb01b6d197e6b59ad10",
"title": "Blockchain-Based Traceability and Visibility for Agricultural Products: A Decentralized Way of Ensuring Food Safety in India"
},
{
"paperId": "1b90d2a8c1affe6b4c957b20d92d0f2654df44b1",
"title": "Constraints and Benefits of the Blockchain Use for Real Estate and Property Rights"
},
{
"paperId": "bbdf3c6c4a479501f317d34663a03b3eab8aa2d6",
"title": "Blockchain energy: Blockchain in future energy systems"
},
{
"paperId": "9da438a348b177a62dcc15d302697b366517ca76",
"title": "Blockchain-Empowered E-commerce: Redefining Trust, Security, and Efficiency in Digital Marketplaces in the Context of Bangladesh."
},
{
"paperId": "113971f6ca186f0bc7fb3d0672a7082f5f1a3dc8",
"title": "Blockchain Technology for Environmental Compliance: Towards A 'Choral' Approach"
},
{
"paperId": "0bdcb2de1d622781b95d9583ccab80191c80e707",
"title": "Are millennials really more sensitive to sustainable luxury? A cross-generational international comparison of sustainability consciousness when buying luxury"
},
{
"paperId": "ee3679d4566368a7a0fe083b658bbed33dd1bee6",
"title": "Blockchain and the future of energy"
},
{
"paperId": "8bf68ed110b8f372ae662808429dabef8ea600b7",
"title": "Blockchain and Sustainability: A Systematic Mapping Study"
},
{
"paperId": "c1465188cca1a449d2ba6245c95f7876d7936441",
"title": "Blockchain: A technological tool for sustainable development or a massive energy consumption network?"
},
{
"paperId": "ab6d3d6dd694d1210006ed5d1163dcf0d402cbf3",
"title": "Three pillars of sustainability: in search of conceptual origins"
},
{
"paperId": "f06fdd3bdd9ab615f72e73a8b7cb61562fc01b4b",
"title": "Why Architecture Does Not Matter: On the Fallacy of Sustainability Balanced Scorecards"
},
{
"paperId": "677d276996ba7a84b9078e9c413cbb1d8820a15e",
"title": "1 Blockchain's roles in meeting key supply chain management objectives"
},
{
"paperId": "c35147e0337416d10011aac7348b589732f02adf",
"title": "The Implications of Blockchain for Income Inequality"
},
{
"paperId": "430169c00bccf06de97beca2572886871cfdacdb",
"title": "Supply chain social sustainability for developing nations: Evidence from India"
},
{
"paperId": "24cdeb7d7421012c2fdd362b8e2816c105b7071f",
"title": "An agri-food supply chain traceability system for China based on RFID & blockchain technology"
},
{
"paperId": "6af2fa42a1673631b420dfe3bd4185519b8de03c",
"title": "Sustainable supply chain management: Evolution and future directions"
},
{
"paperId": "c7f1a16a3e8f6ed10d704eb2ce7428c174d3cfe2",
"title": "A framework of sustainable supply chain management: moving toward new theory"
},
{
"paperId": "a6245960b75630a0a24c7a4816551605ac3d3ab3",
"title": "Sustainability Science: A room of its own"
},
{
"paperId": null,
"title": "Combating Luxury Counterfeiting Through Blockchain Technology"
},
{
"paperId": "89356474123db591e5593d87f58b055c3c0c8be3",
"title": "Tokenizing Renewable Energy Certificates (RECs)—A Blockchain Approach for REC Issuance and Trading"
},
{
"paperId": "3b3b4567dabcaf3ce3c9c55d9318f35343b9ae84",
"title": "A novel electric vehicle charging chain design based on blockchain technology"
},
{
"paperId": "accdb4b200bf1460fda9903eaac964a2af953a0f",
"title": "Generating Employment with Equipped and Trained Workforce Using Blockchain Technology"
},
{
"paperId": "4aa925df769289815bd0dcfe2e25572705fd4e11",
"title": "Blockchain and agricultural supply chains traceability: research trends and future challenges"
},
{
"paperId": "5c5e99f1438535bc5668eafbb3a0671e13c3f988",
"title": "TRENDS IN GLOBAL CO2 AND TOTAL GREENHOUSE GAS EMISSIONS"
},
{
"paperId": null,
"title": "Infrastructural Grind”"
},
{
"paperId": null,
"title": "Liberating wage slaves: Towards sustainable employment practices"
},
{
"paperId": "df308927c925496bc1708bc709d09f97ce90b78a",
"title": "on Intelligent Manufacturing and Automation , 2013 Sustainability and its Integration into Corporate Governance Focusing on Corporate Performance Management and Reporting"
},
{
"paperId": "28cc60bcfbded0c1583b9e8c12ee72e02be51113",
"title": "Business Ethics: A Sustainability Approach"
},
{
"paperId": null,
"title": "VAKT_Building the world's first enterprise-level blockchain platform"
},
{
"paperId": null,
"title": "Aerial"
}
] | 22,128
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Engineering",
"source": "s2-fos-model"
},
{
"category": "Environmental Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/02aa857d4991ffc0fe8e1992c1f6e5b1b94d39b0
|
[
"Computer Science"
] | 0.840041
|
LSec: Lightweight Security Protocol for Distributed Wireless Sensor Network
|
02aa857d4991ffc0fe8e1992c1f6e5b1b94d39b0
|
IFIP International Conference on Personal Wireless Communications
|
[
{
"authorId": "1794236",
"name": "R. Shaikh"
},
{
"authorId": "31273100",
"name": "Sungyoung Lee"
},
{
"authorId": "2109254546",
"name": "Mohammad A. U. Khan"
},
{
"authorId": "48481767",
"name": "Y. Song"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"PWC",
"IFIP Int Conf Pers Wirel Commun"
],
"alternate_urls": null,
"id": "06efd4cf-6136-4948-b84a-68bde61dbe66",
"issn": null,
"name": "IFIP International Conference on Personal Wireless Communications",
"type": "conference",
"url": null
}
| null |
## LSec: Lightweight Security Protocol for Distributed
Wireless Sensor Network*
Riaz Ahmed Shaikh, Sungyoung Lee, Mohammad A.U. Khan, and Young Jae Song
Department of Computer Engineering, Kyung Hee University,
Sochen-ri, Giheung-eup, Yongin-si, Gyeonggi-do, 449-701, South Korea
{riaz, sylee, khan}@oslab.khu.ac.kr, yjsong@khu.ac.kr
**Abstract.** Constraint specific wireless sensor networks need energy efficient
and secure communication mechanisms. In this paper we propose Lightweight
Security protocol (LSec) that fulfils both requirements. LSec provides
authentication and authorization of sensor nodes with simple secure key
exchange scheme. It also provides confidentiality of data and protection
mechanism against intrusions and anomalies. LSec is memory efficient that
requires 72 bytes of memory storage for keys. It only introduces 74.125 bytes of
transmission and reception cost per connection.
#### 1 Introduction
Wireless sensor networks consist of a large number of small size sensor nodes
deployed in the observed environment. Sensor nodes have smaller memory (8K of
total memory and disk space) and limited computation power (8-bit, 4 MHz CPU) [1].
They usually communicate with a powerful base station which connects sensor nodes
with external networks. The limited energy at senor nodes creates hindrances in
implementing complex security schemes. There are two major factors for energy
consumption:
1. Transmission and reception of data.
2. Processing of query request.
Wireless networks are relatively more vulnerable to security attacks than wired
networks due to the broadcast nature of communication [1]. In order to implement
security mechanism in sensor networks, we need to ensure that communication
overhead is less and consumes less computation power. With these constraints it is
impractical to use traditional security algorithms and mechanism meant for powerful
workstations.
Sensor networks are vulnerable to a variety of security threats such as DoS,
eavesdropping, message replay, message modification, malicious code, etc. In order
to secure sensor networks against these attacks, we need to implement message
- This work is financially supported by the Ministry of Education and Human Resources
Development (MOE), the Ministry of Commerce, Industry and Energy (MOCIE) and the
Ministry of Labor (MOLAB) through the fostering project of the Lab of Excellency. The
corresponding author of this paper is Prof. Sungyoung Lee.
-----
confidentiality, authentication, message integrity, intrusion detection and some other
security mechanism. Encrypting communication between sensor nodes can partially
solve the problems but it requires a robust key exchange and distribution scheme.
In general, there are three types of key management schemes [2,3]: Trusted Server
scheme, self enforcing scheme and key-predistribution scheme. Trusted server
schemes relies on a trusted base station, that is responsible for establishing the key
agreement between two communicating nodes as described in [4]. It uses symmetric
key cryptography for data encryption. The main advantages of this scheme are, it is
memory efficient, nodes only need to store single secret key and it is resilient to node
capture. But the drawback of this scheme is that it is energy expensive, it requires
extra routing overhead in the sense that each node need to communicate with base
station several times [3]. Self enforcing schemes use public key cryptography for
communication between sensor nodes. This scheme is perfectly resilient against node
capture and it is fully scalable and memory efficient. But the problem with the
traditional public keys cryptography schemes such as DSA [5] or RSA [6] is the fact
that they require complex and intensive computations which is not possible to
perform by sensor node having limited computation power. Some researchers [7,8]
uses Elliptic curve cryptography as an alternative to traditional public key systems but
still not perfect for sensor networks. Third scheme is key pre-distribution scheme
based on symmetric key cryptography, in which limited numbers of keys are stored
on each sensor node prior to their deployment. This scheme is easy to implement and
does not introduce any additional routing overhead for key exchange. The degree of
resiliency of node capture is dependent on the pre-distribution scheme [3].
Quite recently some security solutions have been proposed in [9,10,11,12,13]
especially for wireless sensor networks but each suffers from various limitations such
as higher memory and power consumptions that are discussed in section 4.
Keeping all these factors in mind we propose a lightweight security protocol
(LSec) for wireless sensor networks. LSec combines the features of trusted server
scheme and Self Enforcing security schemes. Our main contribution is the designing
and implementation of LSec that provides
- Authentication and Authorization of sensor node.
- Simple Secure key exchange scheme.
- Secure defense mechanism against anomalies and intrusions.
- Confidentiality of data.
- Usage of both symmetric and asymmetric schemes.
The rest of the paper is organized as follows. Section 2 describes the details of
LSec. Section 3 presents the simulation results and evaluation of LSec. Section 4
presents the comparison of LSec with other security solutions and Section 5 consists
of conclusion and future direction.
#### 2 Light Weight Security Protocol (LSec)
The basic objective of LSec is to provide lightweight security solution for wireless
sensor networks where all nodes can communicate with each other. LSec can support
both static and mobile environment, which may contain single and multiple Base
-----
Stations (BS). Basic system architecture is shown in figure 1. LSec uses both
symmetric and asymmetric schemes for providing secure communication in wireless
sensor networks.
AzM KMM
���
���
���
��� TGM
�������
������
�������
������
**Fig. 1. LSec System Architecture**
Key Management Module (KMM) is used to store public and shared secret key of
each node with BS to the database. Token Generator Module (TGM) is used to
generate the tokens for the requesters, which will be further used by the other
communicating party for the authentication of requester node. Authorization Module
(AzM) is used to check whether a particular node is allowed to communicate with
other node or group. Lightweight mobile agents will only be installed on Cluster
heads which sends alerts messages to intrusion detection system (IDS), which is
responsible for detecting any anomaly or intrusion in the network. Basic assumptions
and rules of LSec are given below.
**2.1 Assumptions**
1. Base Station (BS) is the trusted party and it will never be compromised.
Compromising the Base station can render the entire sensor network useless,
and it is the only point from where sensor node can communicate with
external networks.
2. Only Base Station (BS) knows the Public keys (Pk) of all the sensor nodes in
the network. Communicating nodes will know each other’s public key during
the time of connection establishment.
**2.2 Rules**
- Asymmetric scheme will only be used for sharing ephemeral secret key
between communicating nodes.
- For every session new random secret key will be used.
- Data will be encrypted by using symmetric schemes because these schemes
are considered to be executed three to four times faster than asymmetric
schemes [14].
-----
**2.3 LSec Packet Format**
LSec packet format is shown in table 1. Currently LSec uses seven types of packets,
‘Request’, ‘Response’, ‘Init’, ‘Ack’, ‘Data’, ‘Update Group Key’ and ‘Alert’ packet.
All seven packets are distinguished by ‘type’ field in the LSec packet. IDsrc field
contain the id of sending node and last encrypted portion contain the information
depending upon the type of packet, as shown in table 1.
**Table 1. LSec: Type field**
Type IDsrc Encrypted Portion
Any
Request (sensor node) EK A-BS (Intended-IDdest, N)
Response BS EKA-BS (R-type, Intended-IDdest, N,
Pk, token | R)
Any +
Init (sensor node) EKB (N, Pk, token)
Any +
Ack (sensor node) EKA (N,sk)
Any
Data EKsk (data)
(sensor node)
Any CH
UpdateGroupKey sensor node EKG (GroupID, new Key), MAC
Any CH
Alert sensor node EKCH-BS (Alert-type), MAC
EKA-BS = Encrypt with the secret key shared between node A and BS
EKA+ = Encrypt with the public key of node A
EKB+ = Encrypt with the public key of node B
EKsk = Encrypt with the shared secret key
EKG = Encrypt with group key
EKCH-BS = Encrypt with the secret key shared between Cluster head and BS
R-type = Response type (positive or negative response)
R = Reason of negative acknowledgement
Intended-IDdest = ID of Intended Destination
Pk = public key
IDsrc = ID of source node
N = Nonce (Unique Random Number)
MAC = Message Authentication Code
CH = Cluster Head
The distribution of bits to different fields (as shown in table 2), introduces some
upper limits, such as, size of source address is of 2 bytes, it means our LSec works
only in the environment where number of sensor nodes not exceeding 2[16]. Length of
Nonce (unique random number) field is of 3 bytes, so LSec can allow maximum of
2[24] connections at a time. The length of public key and private key is of exactly 128
|Type|ID src|Encrypted Portion|
|---|---|---|
|Request|Any (sensor node)|EK (Intended-ID , N) A-BS dest|
|Response|BS|EK (R-type, Intended-ID , N , A-BS dest Pk, token | R)|
|Init|Any (sensor node)|EK +(N, Pk, token) B|
|Ack|Any (sensor node)|EK +(N,sk) A|
|Data|Any (sensor node)|EKsk (data)|
|UpdateGroupKey|Any CH sensor node|EK (GroupID, new Key), MAC G|
|Alert|Any CH sensor node|EK (Alert-type), MAC CH-BS|
-----
**Table 2. Distribution of bits to different fields of LSec**
**Field** **Size** **Field** **Size**
Type 4 bits Public and Private 128
key bits
IDsrc, 16 bits Secret key 64 bits
IDdest
Nonce (N) 23 bits token 4
bytes
R-type 1 bit data 30
bytes
bits and the length of secret key is of exactly 64 bits. Only stream cipher encryption
algorithms are allowed to use because of a fixed length size of packets. MAC is of
64 bits.
**2.4 Procedure**
LSec works in three phases, authentication and authorization phase, key distribution
phase, and data transmission phase. Authentication and authorization is performed
during the exchange of “Request” and “Response” packet by using symmetric
scheme. Key distribution phase involves sharing of random secret key in a secure
manner by using asymmetric scheme. In this phase “INIT” and “ACK” packets will
be exchanged. Data transmission phase involves transmission of data packet in an
encrypted manner.
Let’s suppose node A wants to communicate with the node B. It will first send
request packet to Base station, for receiving token and public key of node B. The
request packet is encrypted with the secret key shared between node A and BS. BS
first checks in the database via AzM that weather node A has rights to establish
connection with node B. If yes, it generates the token which will be further used by
the node B for the authentication of node A. That token is encrypted with secret key
shared between node B and BS, so that node A will not able to decrypt token. BS will
sent back a response packet that contains token, public key of node B and Nonce
(Unique Random Number) that was there in request packet. Nonce will ensure node A
that packet came from genuine BS. When node A gets the positive response from BS
it sent the INIT packet to node B that contains Nonce, its own public key and token
generated by BS. The whole INIT packet is encrypted with the public key of node B.
When node B gets INIT packet it first check token, if it is correct, it will generate the
secret key and sent it back to node A in an encrypted manner. When node A gets
ACK packet, it deletes the public key of node B from its memory, and sent data to
node B by using new session secret key. When data transmission complete, both
nodes delete that session key. For group communication, each node uses the group
secret key for data transmission in a secure manner. Cluster head will update this key
after periodic interval.
|Field|Size|Field|Size|
|---|---|---|---|
|Type|4 bits|Public and Private key|128 bits|
|IDsrc, IDdest|16 bits|Secret key|64 bits|
|Nonce (N)|23 bits|token|4 bytes|
|R-type|1 bit|data|30 bytes|
-----
#### 3 Simulation and Performance Analysis
We have tested our LSec protocol on Sensor Network Simulator and Emulator
(SENSE) [15]. In sensor node we introduce the middleware between application layer
and network layer as shown in figure 2.
Application Sensor
LSec Middleware
Network
FIFO
Battery Link layer
Power Physical Mobility
to Channel from Channel Position_out Data In
**Fig 2. Sensor Node Architecture**
**Table 3. Simulation Parameters**
Terrain 1000x1000
Total Number of Nodes 101 (including BS)
Initial battery of each sensor node 1x10[6]J
Power consumption for transmission 1.6W
Power consumption for reception 1.2 W
Idle power consumption 1.15W
Carrier sense threshold 3.652e-10W
Receive power threshold 1.559e-11W
Frequency 9.14e8
Transmitting & Receiving antenna gain 1.0
That middleware uses LSec for the enforcement of security in the sensor network.
At application layer we use constant bit rate component (CBR) that generate constant
traffic during simulation between two communicating sensor nodes. For the
demonstration and performance evaluation of LSec, CBR is run with and without
|Terrain|1000x1000|
|---|---|
|Total Number of Nodes|101 (including BS)|
|Initial battery of each sensor node|1x106J|
|Power consumption for transmission|1.6W|
|Power consumption for reception|1.2 W|
|Idle power consumption|1.15W|
|Carrier sense threshold|3.652e-10W|
|Receive power threshold|1.559e-11W|
|Frequency|9.14e8|
|Transmitting & Receiving antenna gain|1.0|
-----
LSec. We randomly deploy 100 sensor nodes plus one Base station (BS) in 1000 by
1000 terrain. Basic simulation parameters employed are described in table 3.
**3.1 Performance Analysis of Communication Overhead**
In our simulation scenario, application sent data packets of size 30 bytes in a periodic
interval. The overall communication overhead of LSec for one to one communication
is decreases with the increase in transfer of number of data packets as shown in figure
3. Communication Overhead (C0 %) is calculated as
#### Nc *74.125 CO(%) = ( n )*100
# ∑ NiP *30
_i=1_
(1)
Where as ‘Nc’ is the total number of connections. _NiP_ is the number of packets
transferred by node i. We multiplied 74.125 bytes to Nc because for every connection
LSec exchange four control packets (Request, Response, Init, and Ack) during the
authentication, authorization and key exchange phase whose cumulative size is
74.125 byte. Size of each data packet is 30 bytes.
**Each Data Packet Size = 30 bytes**
30
25
20
15
10
5
0
10 20 30 40 50 60 70 80 90 100
**Number of Data Packets Transfer**
**Fig. 3. Communication Overhead (%) of LSec**
**3.2 Performance Analysis of Power Computation**
Power Computation primarily depends upon the kind of symmetric and asymmetric
scheme. If we assume that computation power required for symmetric encryption and
decryption scheme is CSE and CSD respectively and computation power of
asymmetric encryption and decryption scheme as CAE and CAD respectively. Then
the total power consumption required by single node during first two phases is
_Power Computation = (CSE + CSD) + (CAE + CAD)_ (2)
-----
Computation power required by a single node during data transmission phase is
calculate as,
_Power Computation= (TNSP*CSE) + (TNRP*CSD)_ (3)
Where TNSP is the Total Number of Sent data packets and TNRP is the Total
Number of received data packets.
**3.3 Performance Analysis of Memory Consumption**
Every sensor node needs to store only six keys, three of them are permanent and three
are ephemerals. Permanent keys consist of one public key (self), one private keys and
one public key of BS. Ephemerals keys consist of group key, public key of other node
and session secret key. In order to save these keys only 72 bytes are needed. Details
are given in table 4. This approach will make sensor network memory efficient.
**Table 4. Storage Requirement of Keys**
S/No Keys Size (in bytes)
Permanent Keys
1 Public key of node 16
2 Private key of node 16
3 shared secret key b/w Node & BS 8
Ephemeral Keys
4 Group Key 8
5 Public key of other node 16
6 Session key 8
Total Storage size Required 72 bytes
**3.4 Performance Analysis of Energy Consumption**
The main source of energy consumption at sensor node is its transmission and
reception cost. We used SENSE that consumes energy in four different modes:
TRANSMIT, RECIEVE, IDLE, and SLEEP. Energy consumption rate of each mode
**Initial Energy 1x106 J**
With out any Security With LSec
1003.53
1003.51
1003.49
1003.47
1003.45
1003.43
1003.41
1003.39
1003.37
1003.35
1 8 15 22 29 36 43 50 57 64 71 78 85 92 99
**Nodes**
**Fig 4. Energy Consumptions**
|S/No|Keys|Size (in bytes)|
|---|---|---|
|Permanent Keys|||
|1|Public key of node|16|
|2|Private key of node|16|
|3|shared secret key b/w Node &|BS 8|
|Ephemeral Keys|||
|4|Group Key|8|
|5|Public key of other node|16|
|6|Session key|8|
|Total Storage size Required||72 bytes|
-----
is given in table 3. For each connection, LSec exchange four control packets (Request,
Response, Init, and Ack) of cumulative size 74.125 bytes that requires for authentication,
authorization and key exchange mechanism. That is an acceptable tradeoff between
energy and security. Simulation result of energy consumption is shown in figure 4.
**3.5 Resilience Against Node Compromise**
Single node compromised will not expose the whole communication in network. Only
the communication links that are established with compromised node will expose the
network. Let’s suppose ‘Ncn’ is the set of nodes that establish connections and ‘Ncp’
is the set of compromised nodes. Then Ncn ∩ Ncp will give us the set of nodes that
are compromised as well as connected. Then the maximum number of connections
that can be exposed only if all compromised nodes connected to uncompromised
nodes. On the other hand minimum numbers of links that can be exposed only if all
compromised nodes are connected with each other.
(4)
(5)
_M i n_
_M a x_ : _N c n_ ∩ _N c p_
⎛ ⎞
⎜ _N c n_ ∩ _N c p_ ⎟
⎜ _f o r_ ⎯⎯→ _e v e n_ ⎟
⎜ 2 ⎟
:
⎜ ⎟
⎜ _N c n_ ∩ _N c p_ + 1 ⎟
⎜⎜ ( ) _f o r_ ⎯⎯→ _o d d_ ⎟⎟
⎝ 2 ⎠
If we assume that sensor networks consists of 1000 nodes and total 500
connections established between pair of nodes then the total links that can be
minimum and maximum compromised is shown in figure 5.
**N=1000 Connections = 500**
Min Max
100
80
60
40
20
0
50 150 200 250 300 350 400 450 500
**Compromised Nodes**
**Fig. 5. Percentage of Compromised Links**
#### 4 Comparison of LSec with Other Security Solutions
Comparison of all above discussed schemes with LSec is given in table 5. We
provided comparison from the perspective of memory requirement, transmission cost,
-----
and some other basic security parameters such as authentication, authorization,
confidentiality, etc. Data integrity is generally handled at link layer with the help of
some hashing schemes such as MD5, SHA1 etc or by CRC schemes and availability
is normally handled at physical layer. LSec lies between network and application
layer that’s why it doesn’t provide explicit data integrity and availability support.
**Table 5. Comparison of LSec with other security solutions**
**SPINS** **TinySec** **LiSP** **LSec**
Memory
Requirement with Depended
3 ≥ 8 6
respect to storage on KMS[1]
of keys
During
key Depended
-- 12.6*TNN[2] 74.125*TNC[3]
exchange on KMS
(bytes)
During
Data 20% 10% - 20 8.33%
Transmission
Public Key
Cryptography No No No Yes
Support
Symmetric key
cryptography Yes Yes Yes Yes
Support
Intrusion
Detection No No Yes Yes
mechanism
Authentication
Yes Yes Yes Yes
support
Authorization
No No Yes Yes
support
Data Integrity
Yes Yes Yes No
support
Confidentiality
Yes Yes Yes Yes
support
Availability
No No Yes No
support
1 KMS: Key Management Scheme
2 KNN: Total Number of Nodes
3 KNC: Total Number of Connections
#### 5 Conclusion and Future Directions
We proposed Lightweight security protocol (LSec) for wireless sensor networks,
which provides authentication and authorization of sensor node. It also provides
|Col1|Col2|SPINS|TinySec|LiSP|LSec|
|---|---|---|---|---|---|
|Memory Requirement with respect to storage of keys||3|Depended on KMS1|≥ 8|6|
|Transmission Cost|During key exchange (bytes)|--|Depended on KMS|12.6*TNN2|74.125*TNC3|
||During Data Transmission|20%|10%|> 20|8.33%|
|Public Key Cryptography Support||No|No|No|Yes|
|Symmetric key cryptography Support||Yes|Yes|Yes|Yes|
|Intrusion Detection mechanism||No|No|Yes|Yes|
|Authentication support||Yes|Yes|Yes|Yes|
|Authorization support||No|No|Yes|Yes|
|Data Integrity support||Yes|Yes|Yes|No|
|Confidentiality support||Yes|Yes|Yes|Yes|
|Availability support||No|No|Yes|No|
-----
simple secure key exchange scheme and confidentiality of data. LSec is highly
scalable and memory efficient. It uses 6 keys, which takes only 72 bytes of memory
storage. It introduces 74.125 bytes of transmission and reception cost per connection.
It has the advantage of simple secure defense mechanism against compromised nodes.
In future, we will try to solve the issue related to the neighboring nodes of the base
station that suffered from higher communication overhead by forwarding request and
response packets during authentication and authorization phase.
#### References
1. C. Karlof and D. Wagner, “Secure Routing in Wireless Sensor Networks: Attacks and
Countermeasures”, _proc. of the First IEEE International Workshop on Sensor Network_
_Protocols and Applications (WSNA’03), May 2003, pp. 113- 127_
2. Wenliang Du, Jing Deng, Han, Y.S., Shigang Chen, Varshney P.K, “A key management
scheme for wireless sensor networks using deployment knowledge”, _proc. of INFOCOM_
_2004, Mar 2004_
3. Lydia Ray, “Active Security Mechanisms for Wireless Sensor Networks and Energy
optimization for passive security Routing”, PhD Dissertation, Dep. of Computer Science,
Louisiana State University, Aug 2005
4. J. Kohl and B. Clifford Neuman, “The Kerberos Network Authentication Service (v5)”,
RFC 1510, Sep 1993
5. W. Diffie and M.E. Hellman, “New Directions in Cryptography”, _IEEE Transaction on_
_Information Theory, vol. 22, Nov 1976, pp. 644-654._
6. R. L. Rivest, A. Shamir, L.M. Adleman, “A method for obtaining Digital Signatures and
Public key cryptosystem”, Communication of ACM, vol. 21(2), 1978, pp. 120-126
7. Erik-Oliver Blaß and Martina Zitterbart, “Towards Acceptable Public-Key Encryption in
Sensor Networks”, _proc. of 2[nd] International Workshop on Ubiquitous Computing, ACM_
_SIGMIS, May 2005_
8. John Paul Walters, Zhengqiang Liang, Weisong Shi, and Vipin Chaudhary, “Wireless
sensor network security: A Survey”, Technical Report MIST-TR-2005-007, July, 2005
9. A. Perrig, R. Szewczyk, V. Wen, D. Culler and J. D. Tygar, “SPINS: Security protocols
for sensor networks”, proc. of 7th annual international conference on Mobile computing
_and networking, Rome, Italy, Aug 2001, pp 188-189_
10. Chris Karlof, Naveen Sastry, and David Wagner, “TinySec: a link layer security
architecture for wireless sensor networks”, _Proc. of the 2[nd] international conference on_
_Embedded networked sensor systems, Baltimore, MD, USA, Nov 2004, pp 162-175_
11. K. Jones, A.Wadaa, S. Oladu, L. W|son, and M. Etoweissy, “Towards a new paradigm for
securing wireless sensor networks”, _proc. of the 2003 workshop on New security_
_paradigms, Ascona, Switzerland, Aug 2003, pp 115 - 121_
12. Taejoon Park, and Kang G. Shin, “LiSP: A Lightweight Security Protocol for Wireless
Sensor Networks”, _ACM Transactions on Embedded Computing Systems, vol. 3(3), Aug_
2004, pp. 634–660
13. Sencun Zhu, Sanjeev Setia, and Sushil Jajodia, “LEAP: Efficient Security Mechanism for
Large-Scale Distributed Sensor Networks ”, _Proc. of the 10[th] ACM conference on_
_Computer and communications security, Washington, USA, 2003, pp. 62-72_
14. Elaine Shi and Adrian Perrig, “Designing Secure Sensor Networks”, _IEEE Wireless_
_Communications, Dec 2004, pp. 38-43_
15. Sensor Network Simulator and Emulator (SENSE) http://www.cs.rpi.edu/~cheng3/sense/
-----
|
{
"disclaimer": "Notice: The following paper fields have been elided by the publisher: {'abstract', 'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.1007/11872153_32?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1007/11872153_32, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "BRONZE",
"url": "https://link.springer.com/content/pdf/10.1007/11872153_32.pdf"
}
| 2,006
|
[
"JournalArticle"
] | true
| 2006-09-20T00:00:00
|
[] | 6,634
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Business",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Business",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/02acccd3a4dbea265de8c043807c2dbb4115130c
|
[
"Computer Science",
"Business"
] | 0.884255
|
Phishing Scam Detection on Ethereum: Towards Financial Security for Blockchain Ecosystem
|
02acccd3a4dbea265de8c043807c2dbb4115130c
|
International Joint Conference on Artificial Intelligence
|
[
{
"authorId": "47482568",
"name": "Weili Chen"
},
{
"authorId": "1591110971",
"name": "Xiongfeng Guo"
},
{
"authorId": "2285217145",
"name": "Zhiguang Chen"
},
{
"authorId": "2341167621",
"name": "Zibin Zheng"
},
{
"authorId": "2258301089",
"name": "Yutong Lu"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"Int Jt Conf Artif Intell",
"IJCAI"
],
"alternate_urls": null,
"id": "67f7f831-711a-43c8-8785-1e09005359b5",
"issn": null,
"name": "International Joint Conference on Artificial Intelligence",
"type": "conference",
"url": "http://www.ijcai.org/"
}
|
In recent years, blockchain technology has created a new cryptocurrency world and has attracted a lot of attention. It also is rampant with various scams. For example, phishing scams have grabbed a lot of money and has become an important threat to users' financial security in the blockchain ecosystem. To help deal with this issue, this paper proposes a systematic approach to detect phishing accounts based on blockchain transactions and take Ethereum as an example to verify its effectiveness. Specifically, we propose a graph-based cascade feature extraction method based on transaction records and a lightGBM-based Dual-sampling Ensemble algorithm to build the identification model. Extensive experiments show that the proposed algorithm can effectively identify phishing scams.
|
# Phishing Scam Detection on Ethereum: Towards Financial Security for Blockchain Ecosystem
## Weili Chen[1][,][2], Xiongfeng Guo [1][,][3], Zhiguang Chen [1][,][2], Zibin Zheng [1][,][3] and Yutong Lu [1][,][2]
1School of Data and Computer Science, Sun Yat-sen University
2National Supercomputer Center in Guangzhou, Sun Yat-sen University
3National Engineering Research Center of Digital Life, Sun Yat-sen University
## chenwli28@mail.sysu.edu.cn, guoxf6@mail2.sysu.edu.cn, zhiguang.chen@nscc-gz.cn, zhzibin@mail.sysu.edu.cn, yutong.lu@nscc-gz.cn
## Abstract
In recent years, blockchain technology has created
a new cryptocurrency world and has attracted a lot
of attention. It also is rampant with various scams. For example, phishing scams have grabbed a lot
of money and have become an important threat to
users’ financial security in the blockchain ecosystem. To help deal with this issue, this paper proposes a systematic approach to detect phishing accounts based on blockchain transactions and take
Ethereum as an example to verify its effectiveness.
Specifically, we propose a graph-based cascade feature extraction method based on transaction records and a lightGBM-based Dual-sampling Ensemble
algorithm to build the identification model. Extensive experiments show that the proposed algorithm
can effectively identify phishing scams.
## 1 Introduction
The birth of Bitcoin has brought a whole new world of cryptocurrency. According to coinmarketcap.com, there are now
over 5,000 cryptocurrencies (or tokens) with a market capitalization larger than $200 billion (see [Chen et al., 2020]
for a detailed analysis of the token market). The key technology behind these cryptocurrencies is blockchain technology. Generally speaking, a blockchain can be described as a
distributed and trusted database maintained by a peer-to-peer
network through a special consensus mechanism [Zheng et
_al., 2018]. A blockchain usually implements a cryptocurren-_
cy (or a virtual currency) and it can be exchanged with other
cryptocurrencies or fiat money through exchanges. The financial nature of cryptocurrency makes it the target of many
scams.
Financial security is an important foundation for the
healthy development of blockchain technology. The proliferation of scams in the ecosystem will hinder users’ acceptance and use of blockchain technology, and further, hinder the progress of the technology. Thus, identification of
these scams has become an urgent and critical problem in the
blockchain ecosystem and has attracted great attention from
researchers [Bartoletti et al., 2020; Chen et al., 2018]. The
phishing scam is a new type of cybercrime that arises along
with the rise of online business [Liu and Ye, 2001], which
4506
has now been found in the blockchain ecosystem. According to the report of Chainalysis, more than 50% of all cybercrime revenue was generated from phishing scams since
2017[1]. A widely known example is the phishing scam on Bee
Token ICO [2], in which the phisher eventually gathered about
$1 million from the investors in only 25 hours. These examples show that detecting and preventing phishing scams is an
urgent problem in the blockchain ecosystem.
Traditional phishing scams typically involve setting up a
fake official website and luring users into logging in to obtain
private information, such as passwords. Thus, the main task
of the traditional phishing scam detection method is to identify fake websites through various methods so that users can
get an early warning before logging in. However, phishing
scams in the blockchain era have many new characteristics.
First of all, instead of private information, cryptocurrencies
become the phishing targets. Phishers use a variety of methods to lure ordinary users to transfer money to a designated
account (such as in the case of Bee Token ICO scam). Second, the ill-gotten cryptocurrencies have to be cashed through
exchanges for fiat money (i.e., to convert the ill-gotten cryptocurrencies into fiat money) through transactions. Third, the
transaction records of public blockchain are publicly accessible, which provides a new data source for phishing detection.
Based on these new characteristics and the fact that phishing scams are rampant in the blockchain ecosystem, we propose to build phishing scam detection methods based on
blockchain transactions and AI. These methods can be incorporated into users’ cryptocurrency wallets (i.e., tools for
management of accounts and transactions in the blockchain
ecosystem) as a function of alerting users to potential risks
when interacting with unfamiliar accounts. Figure 1 shows
the proposed framework and uses Ethereum as an example to
demonstrate the effectiveness of our approach. Specifically,
we first downloaded the Ethereum ledger using an Ethereum
client Parity and crawled etherscan.io to get all the phishing
accounts. Then, based on common sense and data analysis,
we propose several filtering rules to alleviate the class imbalance problem. On this basis, we construct the transaction
graph and propose a graph-based cascade feature extraction
1https://blog.chainalysis.com/the-rise-of-cybercrime-onethereum/
2https://theripplecryptocurrency.com/bee-token-scam/
-----
Figure 1: The framework.
method. Next, a Dual-sampling Ensemble framework is proposed to identify suspect accounts. Finally, we verify the validity of the model by comparing it with other methods, evaluate the performance of the model under different parameters,
and discuss the effectiveness of these features.
In summary, we make the following major contributions.
(1) We propose a systematic approach to detect phishing scams in the blockchain ecosystem, and take Ethereum as
an example to verify the effectiveness. The approach has
good performance, which indicates that our method can
be embedded into users’ cryptocurrency wallets to provide users with a financial risk warning function. To accelerate the research in this field and promote the healthy
development of blockchain technology, all relevant data
and code will be released after the paper is published.
(2) We propose a graph-based cascade feature extraction
method, which can conveniently extract rich transaction
structure information and form a feature set with a good
classification effect. Besides, it is very scalable and hard
to evade according to the “six-degree separation” theorem.
(3) We propose a new model integration algorithm, namely
the Dual-sampling Ensemble algorithm, which can be
used for classification problems with a high level of class
imbalance. The evaluation results show the effectiveness
of the algorithm.
## 2 Background and Related Work
Blockchain technology is a key support technology for cryptocurrencies such as Bitcoin[3]. A blockchain can be seen as
a common ledger maintained between peers that do not need
to trust each other [Zheng et al., 2017]. The ledger records
the number of users’ cryptocurrency and the history of transfer transactions between them. The user is represented in the
system as a public-private key pair. Public keys, often called
_addresses, are like accounts in a banking system that records_
the cryptocurrency they hold. (In this paper, we use the term
address and account interchangeably.) In blockchain systems,
transactions are messages sending from one account (the initiator’s address) to another (the receiver’s address) [Chen et
_al., 2018]. Typically, the initiator transfers a certain amoun-_
t of cryptocurrency to the recipient. Transactions that occur
over a period are packaged into blocks by peers and linked
to the previous block through cryptography. Each block has a
corresponding height (denoted as blockNumber in this paper),
3https://bitcoin.org/bitcoin.pdf
increasing by 1 from 0. The block height can be viewed as
the time when the transaction took place. In the bitcoin system, blocks are created roughly every ten minutes. Ethereum
is known as the second-generation blockchain technology
because it provides full support for smart contracts [Wood,
2014]. A smart contract on a blockchain can be viewed as a
piece of code that automatically executes and cannot be terminated when a given condition is met. Ethereum is now
the largest platform for blockchain smart contracts and one
of the main targets of various cyber attacks in the blockchain
ecosystem. The cryptocurrency maintained by Ethereum is
called ether.
In recent years, with the development of blockchain technology, financial security in the blockchain ecosystem has
received extensive attention, and the identification of various fraudulent behaviors has become a research hotspot. In
the Bitcoin ecosystem, [Vasek and Moore, 2015] presents
the first empirical analysis of Bitcoin-based scams. The authors identify 192 scams and point out that at least 13,000
distinct victims lost more than $11 million. [Vasek and
Moore, 2018] analyzes the supply and demand for Bitcoinbased Ponzi schemes, while [Bartoletti et al., 2018] establish an address identification model for Ponzi scheme in the
Bitcoin ecosystem. Besides, [Chen et al., 2019a] show that
there are market manipulation in the Bitcoin exchange Mt.
Gox. In the Ethereum ecosystem, on the one hand, people are concerned with the identification of various scams,
for example, smart Ponzi schemes [Bartoletti et al., 2020;
Chen et al., 2018]. On the other hand, since most smart contracts control certain digital assets, ensuring that there are no
vulnerabilities in the smart contracts is an important part of
Ethereum’s financial security [Kalra et al., 2018].
Phishing detection has been extensively studied in the
past decades and many methods have been proposed [Khonji et al., 2013; Abdelhamid et al., 2014; Zouina and Outtaj, 2017]. However, there is seldom research on phishing fraud identification considering the characteristics of
blockchain. [Andryukhin, 2019] classify the main types and
schemes of phishing attacks on the blockchain project and
suggest methods of protection against phishing attacks from
the blockchain project side’s perspective. Unlike them, we
are targeting the entire blockchain ecosystem and providing
users with an early warning against phishing scams.
## 3 Proposed Method
Identifying phishing accounts in the blockchain system faces
two challenges: 1) we only have transaction records and know
little about account functions and holder information and 2)
the number of phishing addresses is very few and other addresses are huge, identifying such a small group of accounts
in the huge account set is like looking for a needle in the
haystack. (The details of the data are described in Section
4.) To meet the challenges, the proposed method includes
two parts, the cascade feature extraction method, and the
lightGBM-based Dual-sampling Ensemble algorithm.
### 3.1 Cascade Feature Extraction Method
Since transaction records are the only information we can
use, and they give the accounts a natural graphical structure,
-----
to extract effective features, we first construct a transaction
graph (TG) based on these transaction records. Specifically,
_TG = (V, E), where V is a set of nodes (all the addresses_
in the dataset) and E = {(vi, vj)|vi, vj ∈ _V } is a set of or-_
dered edges. Each edge indicates that an address Vi transfers
a certain amount of ether to another address Vj. Each edge
has two attributes: blockNumber and amount, representing
the time when this edge emerges and the amount of the transaction. Please note that there may be multiple edges between
two nodes in TG, depending on the number of transactions
between the two related accounts. (we use account, address,
and node interchangeably in the following.) Next, we introduce the proposed feature extraction method.
Graph-based features have proven to be very effective
in many identification problems [Chatzakou et al., 2017;
Ramalingam and Chinnaiah, 2018]. Thus, we propose a
TG-based cascade feature extraction method for phishing account identification. The idea is as follows. Treat the transaction between accounts as a friend relationship, to judge the
category of an account, we can use not only the information
of the account, but also the information of its friends, even the
information of its friends’ friends, and so on. To explain more
clearly, we first define several keywords related to a node.
_Node data: Node data is the transaction history of that_
_•_
node. Each transaction contains information about the
time, direction, and amount of the transaction. The
transaction time is denoted as blockNumber, which is
an increasing integer. A transaction has two directions:
_out and in. The out-transactions of an account trans-_
fer ether from the account to other accounts and the intransactions of an account receive ether from other accounts.
_Node features: Node features are all kinds of informa-_
_•_
tion extracted from node data. In this paper, we extract
information through various statistical methods.
_N-order friend: A node’s 1-order friend is a node di-_
_•_
rectly connected to the node (i.e., there are transactions
between them). A node’s n-order friend is a node connected to the node with at least n-1 nodes.
_N-order features: The 0-order features of a node is the_
_•_
node features of that node. The n-order features are extracted in cascade from the n-order friends.
To explain how to achieve cascade feature extraction, we
show the procedure of 2-order features extraction in Figure
2. Suppose we need to compute the 2-order features of node
_A, which have 1-order friends B, C and 2-order friends D, E,_
_F, G, H. In the figure, each undirected edge represents one_
or more transactions (regardless of the directions) between two nodes, and the counterparty of the 2-order friends is not
shown. The procedure is divided into three stages. In the
first stage, we compute a statistic (i.e., the grey rectangle) for
each 2-order friends by using its node data (i.e., the transaction history). The second stage needs to calculate a statistic
for each 1-order friend by using the statistics computed in the
first stage (not the node data of the 1-order friend). Similarly, in the last stage, we still calculate a statistic whose input
comes from the second stage. This approach is very scalable.
Figure 2: Example of 2-order feature extraction procedure.
In fact, by increasing the order and using different statistic
methods at different stages, we can extract rich information
about how a node interacts with the entire network. It should
be noted that the approach we describe here does not take into account the direction of the transaction. But, for phishing
accounts, in-transactions and out-transactions are significantly different in meaning. Therefore, in this paper, we extract
features for two different directions respectively.
**Node Features**
The node features are statistics of its node data. There are two
_types of data: transaction amount and transaction times (i.e.,_
_blockNumber). In order to distinguish the nature of the trans-_
action, statistics are made in different directions (i.e., outtransactions or in-transactions). For convenience, we name
these features as direction type method. For example, a feature in block std of a node indicates the standard deviation
(i.e., the method sd) for the transaction time (i.e., the type of
data block) of all in-transactions (i.e., the transaction direction
_in). For the transaction time, we compute only the transaction_
time span (denoted as ptp) and its standard deviation (denoted as sd). For the transaction amount, we calculated the sum,
the maximum, the minimum, the mean and the standard deviation (i.e., sd). In addition, there are statistics unrelated to
transaction amount: count, unique, and unique ratio. They
represent the number of transactions (i.e., count), the number
of counterparties (i.e., unique), and the ratio of the two (i.e.,
_unique/count). By doing so, we obtained 19 features (i.e.,_
2 1 (2 + 5 + 2) + 1).
_×_ _×_
**N-order Features**
For simplicity, in this study, we extract only 1-order network features. As mentioned, the direction of the transaction
is important in identifying phishing scams. Thus, considering the transaction direction, the 1-order friends of a node
can be divided into from friends and to friends. In simple
terms, when there is a transfer transaction from node A to node B, we call node B a from friend of node A and node A
a to friend of node B. Specifically, the 1-order network features are named as friend direction statistic2 statistic1. For
example, the from in mean max feature is calculated as follows: we first compute the maximum (i.e., max) of the intransaction amounts for each from friend. Then, we compute
the mean of all statistics in the previous stage. Similarly, to
compute to out std sum, we first compute the sum of all the
_out-transaction amounts for each to friend. Then, we com-_
-----
pute the standard deviation (i.e., sd) of all statistics in the
previous stage. By doing so, we can obtain 200 features (i.e.,
2 2 2 5 5). Please note that we did not take time into
_×_ _×_ _×_ _×_
account in the 1-order network feature extraction.
### 3.2 Dual-sampling Ensemble Method
Identifying phishing scams is essentially establishing a classification model of addresses. But the phishing account identification faces a class imbalance problem. To build
a useful suspect identification model, we propose a Dualsampling Ensemble method, an identification framework integrated with many base models trained by sampling examples and features.
**Base Model**
The base models play a central role in the identification
framework. Many mature classification algorithms can be
used as base models, such as logistic regression (LR), support
vector machine (SVM), and decision tree (DT). Among these
models, the gradient boosting decision tree (GBDT) obtained
good results in many problems. There are several different
variants of GBDT, including XGBoost [Chen and Guestrin,
2016] and lightGBM [Ke et al., 2017], which are widely used
and generally accepted. In the phishing detection problem,
we found that lightGBM is more efficient, thus we select it as
our base model.
Given the supervised training set X = {(xi, yi), i =
1, 2, _, n_, lightGBM integrates a number of K regression
_· · ·_ _}_
trees f (x) = _K1_ �Ti=1 _[h][i][(][x][)][ to approximate a certain func-]_
tion f _[∗](x) that minimizes the expected value of a specific_
loss function L(y, f (x)). In each iteration of GBDT, assume
that the strong learner obtained by the previous iteration is
_ht−1(x), the loss function is L(f_ (x), ht−1(x)), then the aim
for the current iteration is to find a week learner using CART
regression tree model which denoted as ht(x), to minimize
the formula L(f (x), ht−1(x) + ht(x)). Suppose in iteration
_t, the negative gradient for sample i can be represented as_
_rti =_ _∂L(∂hyit,h−t1−(x1(ix)_ _i))_ _. By using the Log-likelihood loss as_
loss functionL(y, h(x)) = log(1 + exp( _yh(x))), where_
_−_
_y_ [ 1, 1], we can simplify the negative gradient of sam_∈_ _−_
ple as below:
_yi_
_rti = −_ _[∂L][(][y][i][, h][t][−][1][(][x][i][))]_ =
_∂ht−1(xi)_ 1 + exp(yih(xi)) _[,]_
where i = 1, 2, _, m._
_· · ·_
By using the formula, LightGBM chooses to remove these
small gradient samples from the training set to make the model pay more attention to those samples which cause great
Loss. This technique is called Gradient-based One-Side Sampling (GOSS) [Ke et al., 2017]. When constructing the CART
regression tree, LightGBM binds the mutual exclusion features so that the number of features (the leaves) can be greatly
reduced.
**Dual-sampling Ensemble**
Inspired by EasyEnsemble [Liu et al., 2008], we propose a
Dual-sampling Ensemble algorithm to solve the class imbalance problem in the phishing scam identification. The pseudocode is shown in Algorithm 1.
**Algorithm 1 The Dual-sampling Ensemble algorithm**
**Input:** The minority class example set P, the majority
example set,, the number of base models k, the
_N_ _|P| ≪|N|_
feature sample ratio r, and the number of features d, The best
parameters for the base model
**Output:** The integration result.
1: Let i 0;
_←_
2: while i < k do
3: _i_ _i + 1;_
_←_
4: Randomly sample a subset Ni from N, |Ni| = ⌊ _K[N]_
_[⌋][;]_
5: Learn a base model hi using P ∪Ni with only d × r
randomly sampled features. The parameters are sampled around the best parameters;
6: end while
7: return H(x) = _K1_ �Ti=1 _[h][i][(][x][)]_
The idea behind the Dual-sampling Ensemble is simple.
Similar to EasyEnsemble [Liu et al., 2008], we reduce the
class imbalance by sampling the majority example set (i.e.,
negative examples). The difference is that we also sample
the features of the examples in the training set since we can
obtain a large number of features by using the cascade feature extraction method. This dual sampling method allows
the base models to have better heterogeneity.
## 4 Data Collection and Preparation
### 4.1 Data Collection
We launch an Ethereum client, Parity[4], on our server to download the ledger of Ethereum. By using Parity, we obtained
all the Ethereum blocks before January 3, 2019 (to be exact,
from block height 0 to block height 7,000,000). By analyzing the transactions obtained, we get 43,783,194 accounts,
among which 1,564,580 accounts controlled by smart contracts.
One of the most important tasks in establishing a phishing
scam identification model is to find enough phishing account
examples. Fortunately, etherscan.io provides several tags for
Ethereum addresses, and by crawling the website, we obtain
all the addresses labeled with Phishing[5]. These addresses are
used in some verified phishing scams. In this way, we obtain
1,683 phishing addresses. We call these phishing addresses
as positive examples and the rest as negative examples.
### 4.2 Data Cleaning
After getting all the data, we found that the class was very imbalanced. The class imbalance ratio, i.e., the ratio of the size
of the majority class (negative examples) to minority class
(positive examples), exceeds 26,000. Given that some addresses are not phishing addresses, we recommend that some
obvious negative examples (i.e., non-phishing addresses) be
eliminated before model training in order to build a more effective model. To this end, we 1) filter transaction records involving a smart contract address, 2) eliminate addresses
4www.parity.io/ethereum/
5etherscan.io/accounts/label/phish-hack
-----
with less than 10 or more than 1,000 transaction records, and
3) ignore all transactions that appear before block height 2
million.
The above cleaning methods are based on the following
considerations. First of all, smart contracts often have complex logic and are not convenient for phishing scams. Furthermore, smart contracts account for very little in the phishing addresses (i.e., 2.6%), and they usually relate to tokens.
Thus, In this preliminary study, for the sake of simplicity, we
leave out smart contracts. Second, we want to learn the behavioral characteristics of phishing accounts through transaction records, and too few records are not good for learning.
Besides, too many records indicate that the account may be a
wallet or other type of accounts. In fact, there are many addresses (i.e., >70%) with more than 1,000 transaction records, and only one address is labeled with phishing. Finally, by
analyzing the initial activity time of phishing addresses, we
find that all phishing addresses are active after 2016-08-02.
This may be because, in the early days of Ethereum, phishing scams were relatively few, and even fewer were recorded.
Therefore, we proposed to build the model based on records after block height of 2 million (i.e., 2016-08-02). These
filtering rules allow the model to focus on learning the characteristics of phishing scams.
## 5 Experiment Result and Analysis
### 5.1 Experiment Settings
We downloaded all of Ethereum’s transaction data from its inception to January 3, 2019 (i.e., from block height 0 to
block height 7,000,000). By using the filter rules in Section
4.2, we ended up with 7,795,044 transaction records. There
are 534,820 addresses, 323 of which are phishing addresses.
The following experiments are based on this data set. In order to reflect the effectiveness of the model more accurately
and avoid the contingency caused by the partitioning of train
and test sets, the paper adopts the evaluation method of k-fold
cross-validation. Specifically, we set the parameter k=5. To
accurately evaluate the model, we select four metrics: precision, recall, F1, and AUC, which is commonly used in classification problems.
### 5.2 Method Comparison
In order to verify that our proposed model is more suitable
for this problem, we compared the single-model lightGBM,
Support Vector Machine (SVM), decision tree (DT), and their
Dual-sampling Ensemble (DE+) models. SVM and DT are
considered efficient in many classification problems of class
imbalance [Chen et al., 2019b]. Thus, we chose it as the baseline of our model. To compare the performance of these methods, we set the feature sampling rate to 70%, and the number of base models to 1600 (i.e., balance ensemble). Table
1 shows the results. As can be seen, in these single-models,
SVM performs poorly, lightGBM and DT have certain performance, but they are obviously of no practical value. On
the contrary, after adopting the ensemble strategy, the performance of each model is significantly improved, especially
lightGBM and DT (i.e., DElightGBM and DEDT). This result
Method Precision Recall F1 AUC
SVM 0.0000 0.0002 0.0000 0.4817
DT 0.0552 0.0810 0.0657 0.5630
lightGBM 0.0535 0.0745 0.0623 0.5364
DESVM 0.2222 0.0076 0.0146 0.5046
DEDT 0.7295 0.7167 0.7230 0.7183
**DElightGBM** **0.8196** **0.8050** **0.8122** **0.8097**
Table 1: The performance comparison
#models Precision Recall F1 AUC
1 0.0789 0.0991 0.0879 0.549
100 0.7583 0.3993 0.5232 0.6947
800 **0.9288** 0.7368 **0.8217** **0.8274**
1000 0.826 0.7585 0.7908 0.8206
1600 0.8196 **0.805** 0.805 0.8097
Table 2: The effect of example sampling (with lightGBM)
shows that the ensemble method is a good choice when facing the class imbalance. It is worth noting that the proposed
model (i.e., DElightGBM) performs well on all metrics (i.e.,
all larger than 0.8). It means that the proposed model can be
deployed in a real wallet for real-time warnings.
### 5.3 Example Sampling Effect Analysis
Evaluating the impact of example sampling on the model is
essentially selecting the number of base models. Table 2
shows the four evaluation metrics of the framework DElightGBM with different numbers of base models. (We set the feature sampling rate to 70% and the parameters of each model
are randomly selected around the optimal parameters.) It can
be seen that with the increase in the number of base models, all the metrics obtained different degrees of promotion.
When the number of base models reaches 800 (i.e., half balance ensemble), three metrics (i.e., precision, F1 and AUC)
reach the maximum. However, the recall keeps going up, and
it reaches its maximum when the number of base models is
1600 (i.e., balance ensemble). This result indicates that the
level of class imbalance is a very important factor affecting
the performance of base models. From the experimental results, half balance ensemble seems to be a good choice. To
make the model more practical, however, we would prefer
to find all potential phishing scams (i.e., higher recall) at the
expense of precision. Therefore, we propose the use of the
balance ensemble for phishing scam detection.
### 5.4 Feature Sampling Evaluation
Next, we analyze the effect of feature sampling by setting different sampling ratios. To eliminate the effect of the number
of base models, it is uniformly set at 1600. Table 3 shows the
evaluation results. In general, the feature sampling method
has a certain influence on the final results, however, as compared with example sampling, its influence is far less significant. From the perspective of the most preferred metric, re_call, 0.8 is the best feature sampling ratio. Compared to using_
all the features (i.e., ratio=1), recall improved 4.24%.
-----
Ratio Precision Recall F1 AUC
0.6 0.8228 0.7832 0.8025 0.8018
0.7 0.8149 0.8205 0.8177 0.8127
**0.8** **0.8258** **0.8390** **0.8324** **0.8282**
0.9 0.8055 0.7955 0.8005 0.7957
1.0 0.8282 0.8049 0.8164 0.8096
Table 3: The effect of feature sampling
Figure 3: The top 15 important features.
These results reveal a noteworthy phenomenon. It is not
necessarily correct that the more features the model has, the
better the performance. On the contrary, in the case that we
can obtain a large number of features, a certain degree of feature sampling is conducive to obtaining a better model. This
may because feature sampling can make different base models view the object from different angles, so as to obtain better
identification.
### 5.5 Feature Analysis
Since we adopted the method of cascading feature extraction,
a large number of features were obtained. Figure 3 shows the
top 15 important features in the model. Next, we analyze why
some of these features are important.
_in block std is the standard deviation of blockNumber of_
_•_
all in transaction for a node. This feature reflects the intensity of in-transactions at a certain address. If there is
a large number of in-transactions in a short period, the
_blockNumber of these transactions will be very close to_
each other, and thus the constructed to block std will be
very small. This feature is much more important than
the others, and its meaning is easy to understand. For a
phishing address, a natural phenomenon is that the number of in-transactions increased suddenly within a period
after the phishing began. However, with the phishing scam revealed, in-transactions become rare, or even nonexistent. This leads to in-transactions are concentrated
in a small period for a phishing address, and the feature
can grasp this characteristic very well.
_to out sum median is a typical 1-order network feature._
_•_
It reflects the overall situation (i.e., sum) of all the to
friends’ out-transactions. This feature is not as intuitive
as the previous one and requires some explanation to
understand its value. First of all, we can think of the
median amount of out transaction of an address as an
indicator of its financial strength. This is not difficult
to understand, because the large median means that at
least half of the address’s out transaction amounts are
large, indicating that its financial strength is stronger.
Second, for phishing addresses, to friends are the victims of the phishing scam. Thus, for phishing scams,
this feature can be seen as an indication of the overall
financial strength of all its victims.
_from in sum min is also an 1-order network feature. D-_
_•_
ifferent from the previous feature, this feature reflects
the in transaction of the node’s from friend. It is relatively easy to understand why the feature is important.
For phishing scams, money laundering is an important part before cashing out. Therefore, the from friend of
the phishing address, which is usually the intermediate
address used for money laundering, must exhibit behavior characteristics different from normal addresses. And,
this type of features captures the difference effectively.
The above analysis of the top three features shows that our
feature engineering achieves good results, fully mining the
characteristics of the node itself and different neighbors of
the node.
## 6 Conclusion and Future Work
In blockchain ecosystems, various scams are rampant, which
seriously threaten the financial security of users involved. To
help dealing with this issue, in this study, we propose a systematic approach to detect phishing scams in the Ethereum
ecosystem. First of all, by using the Parity client and
crawl etherscan.io, we collect all transactions of the Etehreun
blockchain and the labeled phishing addresses. Then, by using this data, we construct a transaction graph and propose a
graph-based cascade feature extraction method, which helps
us extract many useful features. Next, based on the extracted
features and lightGBM, we propose a Dual-sampling Ensemble model to detect phishing suspects. Finally, we evaluate
the model from many angles, and the results indicate the effectiveness of our model. In the future, we are going to further
this study to other cybercrimes and set up a blockchain scam
detection website to provide the phishing scam identification
service in the form of API. Besides, to accelerate the research
in this field, all relevant data and code will be released after
the paper is published.
## Acknowledgments
The work described in this paper was supported by
the National Key R&D Program of China (2018YFB0204303), the National Natural Science Foundation of China (61722214), the Natural Science Foundation of Guangdong (2018B030312002, 2019B020214006), China Postdoctoral Science Foundation (Grant no. 2019TQ0372,
2019M660223), the Program for Guangdong Introducing
Innovative and Entrepreneurial Teams under Grant NO.
2016ZT06D211 and Pearl River S&T Nova Program of
Guangzhou under Grant NO. 201906010008. Zhiguang Chen
and Zibin Zheng are the corresponding authors.
-----
## References
[Abdelhamid et al., 2014] Neda Abdelhamid, Aladdin
Ayesh, and Fadi Thabtah. Phishing detection based
associative classification data mining. _Expert Systems_
_with Applications, 41(13):5948–5959, 2014._
[Andryukhin, 2019] AA Andryukhin. Phishing attacks and
preventions in blockchain based projects. In Proceedings
_of the International Conference on Engineering Technolo-_
_gies and Computer Science (EnT), pages 15–19. IEEE,_
2019.
[Bartoletti et al., 2018] Massimo Bartoletti, Barbara Pes,
and Sergio Serusi. Data mining for detecting bitcoin ponzi
schemes. In Proceedings of the Crypto Valley Conference
_on Blockchain Technology, pages 75–84. IEEE, 2018._
[Bartoletti et al., 2020] Massimo Bartoletti, Salvatore Carta, Tiziana Cimoli, and Roberto Saia. Dissecting ponzi schemes on ethereum: Identification, analysis, and impact. Future Generation Computer Systems, 102:259–277,
2020.
[Chatzakou et al., 2017] Despoina Chatzakou, Nicolas Kourtellis, Jeremy Blackburn, Emiliano De Cristofaro, Gianluca Stringhini, and Athena Vakali. Mean birds: Detecting aggression and bullying on twitter. In Proceedings of
_the ACM on web science conference, pages 13–22. ACM,_
2017.
[Chen and Guestrin, 2016] Tianqi Chen and Carlos Guestrin.
Xgboost: A scalable tree boosting system. In Proceedings
_of the ACM SIGKDD International Conference on Knowl-_
_edge Discovery and Data Mining, pages 785–794. ACM,_
2016.
[Chen et al., 2018] Weili Chen, Zibin Zheng, Jiahui Cui,
Edith Ngai, Peilin Zheng, and Yuren Zhou. Detecting
ponzi schemes on ethereum: Towards healthier blockchain
technology. In Proceedings of the World Wide Web Confer_ence (WWW2018), pages 1409–1418. International World_
Wide Web Conferences Steering Committee, 2018.
[Chen et al., 2019a] Weili Chen, Jun Wu, Zibin Zheng,
Chuan Chen, and Yuren Zhou. Market manipulation of
bitcoin: Evidence from mining the mt. gox transaction
network. In IEEE INFOCOM 2019-IEEE Conference on
_Computer Communications, pages 964–972. IEEE, 2019._
[Chen et al., 2019b] Weili Chen, Zibin Zheng, Edith C-H Ngai, Peilin Zheng, and Yuren Zhou. Exploiting blockchain
data to detect smart ponzi schemes on ethereum. IEEE
_Access, 7:37575–37586, 2019._
[Chen et al., 2020] Weili Chen, Tuo Zhang, Zhiguang Chen,
Zibin Zheng, and Yutong Lu. Traveling the token world: A
graph analysis of ethereum erc20 token ecosystem. In Pro_ceedings of the World Wide Web Conference (WWW2020),_
pages 1409–1418. International World Wide Web Conferences Steering Committee, 2020.
[Kalra et al., 2018] Sukrit Kalra, Seep Goel, Mohan Dhawan, and Subodh Sharma. Zeus: Analyzing safety of
smart contracts. In Proceedings of the Network and Dis_tributed System Security Symposium, 2018._
[Ke et al., 2017] Guolin Ke, Qi Meng, Thomas Finley,
Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, and
Tie-Yan Liu. Lightgbm: A highly efficient gradient boosting decision tree. In Proceedings of the International
_Conference on Advances in Neural Information Process-_
_ing Systems, pages 3146–3154, 2017._
[Khonji et al., 2013] Mahmoud Khonji, Youssef Iraqi, and
Andrew Jones. Phishing detection: a literature survey.
_IEEE Communications Surveys & Tutorials, 15(4):2091–_
2121, 2013.
[Liu and Ye, 2001] Jiming Liu and Yiming Ye. Introduction
to e-commerce agents: Marketplace solutions, security issues, and supply and demand. In E-Commerce Agents,
pages 1–6. Springer, 2001.
[Liu et al., 2008] Xu-Ying Liu, Jianxin Wu, and Zhi-Hua
Zhou. Exploratory undersampling for class-imbalance
learning. IEEE Transactions on Systems, Man, and Cy_bernetics, Part B (Cybernetics), 39(2):539–550, 2008._
[Ramalingam and Chinnaiah, 2018] Devakunchari Ramalingam and Valliyammai Chinnaiah. Fake profile detection
techniques in large-scale online social networks: A comprehensive review. Computers & Electrical Engineering,
65:165–177, 2018.
[Vasek and Moore, 2015] Marie Vasek and Tyler Moore.
There’s no free lunch, even using bitcoin: Tracking the
popularity and profits of virtual currency scams. In Pro_ceedings of the International Conference on Financial_
_Cryptography and Data Security, pages 44–61. Springer,_
2015.
[Vasek and Moore, 2018] Marie Vasek and Tyler Moore.
Analyzing the bitcoin ponzi scheme ecosystem. In
_Proceedings of the International Conference on Finan-_
_cial Cryptography and Data Security, pages 101–112._
Springer, 2018.
[Wood, 2014] Gavin Wood. Ethereum: A secure decentralised generalised transaction ledger. Ethereum Yellow
_Paper, 2014._
[Zheng et al., 2017] Zibin Zheng, Shaoan Xie, Hongning
Dai, Xiangping Chen, and Huaimin Wang. An overview
of blockchain technology: Architecture, consensus, and
future trends. In Proceedings of the IEEE International
_Congress on Big Data, pages 557–564. IEEE, 2017._
[Zheng et al., 2018] Zibin Zheng, Shaoan Xie, Hongning
Dai, Xiangping Chen, and Huaimin Wang. Blockchain
challenges and opportunities: A survey. _International_
_Journal of Web and Grid Services, 14:352–375, 2018._
[Zouina and Outtaj, 2017] Mouad Zouina and Benaceur
Outtaj. A novel lightweight url phishing detection system
using svm and similarity index. Human-centric Comput_ing and Information Sciences, 7(1):17, 2017._
-----
|
{
"disclaimer": "Notice: The following paper fields have been elided by the publisher: {'references'}. Paper or abstract available at https://api.unpaywall.org/v2/10.24963/ijcai.2020/621?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.24963/ijcai.2020/621, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "BRONZE",
"url": "https://www.ijcai.org/proceedings/2020/0621.pdf"
}
| 2,020
|
[
"JournalArticle",
"Conference"
] | true
| 2020-07-01T00:00:00
|
[] | 9,630
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Engineering",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/02af1e17cb68c7cace6f3d38c2e767e7b5fb1e66
|
[
"Computer Science"
] | 0.863542
|
Survey on Three Components of Mobile Cloud Computing: Offloading, Distribution and Privacy
|
02af1e17cb68c7cace6f3d38c2e767e7b5fb1e66
|
[
{
"authorId": "9491665",
"name": "Anirudh Paranjothi"
},
{
"authorId": "2109251680",
"name": "M. Khan"
},
{
"authorId": "1776310",
"name": "Mais Nijim"
}
] |
{
"alternate_issns": null,
"alternate_names": null,
"alternate_urls": null,
"id": null,
"issn": null,
"name": null,
"type": null,
"url": null
}
|
Mobile Cloud Computing (MCC) brings rich computational resource to mobile users, network operators, and cloud computing providers. It can be represented in many ways, and the ultimate goal of MCC is to enable execution of rich mobile application with rich user experience. Mobility is one of the main characteristics of MCC environment where user can be able to continue their work regardless of movement. This literature review paper presents the state-of-the-art survey of MCC. Also, we provide the communication architecture of MCC and taxonomy of mobile cloud in which specifically concentrates on offloading, mobile distribution computing, and privacy. Through an extensive literature review, we found that MCC is a technologically beneficial and expedient paradigm for virtual environments in terms of virtual servers in a distributed environment, multi-tenant architecture and data storing in a cloud. We further identified the drawbacks in offloading, mobile distribution computing, privacy of MCC and how this technology can be used in an effective way.
|
**Journal of Computer and Communications, 2017, 5, 1-31**
[http://www.scirp.org/journal/jcc](http://www.scirp.org/journal/jcc)
ISSN Online: 2327-5227
ISSN Print: 2327-5219
# Survey on Three Components of Mobile Cloud Computing: Offloading, Distribution and Privacy
### Anirudh Paranjothi, Mohammad S. Khan, Mais Nijim
Department of Electrical Engineering and Computer Science, Texas A&M University, Kingsville, TX, USA
How to cite this paper: Paranjothi, A.,
Khan, M.S. and Nijim, M. (2017) Survey on
Three Components of Mobile Cloud Computing: Offloading, Distribution and Privacy. Journal of Computer and Communications, 5, 1-31.
[https://doi.org/10.4236/jcc.2017.56001](https://doi.org/10.4236/jcc.2017.56001)
Received: January 18, 2017
Accepted: April 3, 2017
Published: April 6, 2017
Copyright © 2017 by authors and
Scientific Research Publishing Inc.
This work is licensed under the Creative
Commons Attribution International
License (CC BY 4.0).
[http://creativecommons.org/licenses/by/4.0/](http://creativecommons.org/licenses/by/4.0/)
## Abstract
Mobile Cloud Computing (MCC) brings rich computational resource to mobile users, network operators, and cloud computing providers. It can be
represented in many ways, and the ultimate goal of MCC is to enable execution of rich mobile application with rich user experience. Mobility is one of
the main characteristics of MCC environment where user can be able to continue their work regardless of movement. This literature review paper presents
the state-of-the-art survey of MCC. Also, we provide the communication architecture of MCC and taxonomy of mobile cloud in which specifically concentrates on offloading, mobile distribution computing, and privacy. Through
an extensive literature review, we found that MCC is a technologically beneficial and expedient paradigm for virtual environments in terms of virtual servers in a distributed environment, multi-tenant architecture and data storing in
a cloud. We further identified the drawbacks in offloading, mobile distribution computing, privacy of MCC and how this technology can be used in an
effective way.
## Keywords
Cloud Computing, Mobile Cloud Computing, Offloading,
Distribution and Privacy
Open Access
## 1. Introduction
Smartphones are becoming popular and its users are increasing rapidly every
year. Features of smart phones include touch screen interface, Wi-Fi, high speed
processors, GPS, etc. Popularity of smartphones allows developers to develop
mobile applications in various domains like sports, games, finance, education,
etc. [1]. Still these devices are suffered from issues like limited storage space, li
DOI 10 4236/j 2017 56001 A il 6 2017
-----
A. Paranjothi et al.
2
mited bandwidth and energy due to development of complex mobile applica
tions. To solve these issues cloud computing techniques were introduced. Some
of the popular cloud service providers are Amazon, Windows, Google, etc.
Heavy Reading [2] and ABI research [3] suggested that revenue of MCC market
$68 billion in 2017.
MCC is the combination of cloud computing, mobile computing and wireless
networks. Advantages of MCC over cloud computing are:
1) Flexibility: Due to flexibility, users can access the data using their devices
from any part of the world. But the user should have proper internet connectivi
ty.
2) Data availability: Data availability allows the user to access their data at any
time. It also provides the facility of multiple users accessing the same data si
multaneously.
3) Multiple platforms: MCC also provides support for multiple platforms. It
allows users to access their data in cloud irrespective of any platform.
Current cloud computing provides following facilities to the user: 1) Execut
ing operations on cloud, 2) Large storage capacity, 3) Backup, 4) Traffic count
ing, 5) Ability to choose datacenters. Cloud providers mainly concentrate on
areas like throughput, memory, availability of server, storage, etc. Also, they are
providing three basic services for MCC:
1) Platform service: Platform as a Service (PaaS) provides hardware and soft
ware for the user to create, modify, run their applications. Main advantage of
Paas is that it allows user to execute and complete their tasks without having ap
propriate software or hardware.
2) Application services: Application services are also known as Software as a
Service (Saas). It provides software application to the user whenever they need it
over the internet. It gained more popularity in software market due to software
on demand. The main advantage of using Saas is: 1) Cost savings, 2) Efficiency.
It also eliminates the issue of individual user license and thereby reduces the ex
pense of an organization.
3) Context-rich services: Mobile applications are becoming popular and pro
viding context aware services to its users. To support this, MCC providers are
providing context rich services to the users. It includes congestion detection,
discovering parking space, etc.
Papers [4] [5] [6] [7] have not discussed in detail about various techniques
involved in MCC. This paper gives the overall idea about offloading, mobile dis
tribution, and privacy in the cloud. Further this paper gives information about
various factors affecting MCC and future of cloud computing environment.
Rest of the paper is organized as follows: Section 2 discusses about current
mobile cloud architecture and programming model; Section 3 discusses about
Offloading mobile applications in cloud; Section 4 focuses on Mobile distribu
tion computing and cloud; Section 5 discusses about Privacy in cloud and user
authentication. Finally, we presented the conclusion and future work in Section
6.
-----
A. Paranjothi et al.
## 2. Current Mobile Cloud Architecture and Programming Models
Mobile cloud architecture:
Current mobile cloud computing architecture includes following components:
1) Regional Data center (RDC), 2) Wireless core, 3) Base stations. It is repre
sented in Figure 1. MCC architecture allows users to offload their operations on
cloud [8]. Example: Global Positioning System [GPS], multiplayer games, etc.
[9]. But, it is most suitable for heterogeneous environments [10].
RDC: It is used in home computer systems and its associated elements like
storage and telecommunication systems. RDC consists of various security devic
es, power supplies, environment controls, etc. Cloud data centers are distributed
in different locations around the world [11].
Wireless core network: Routing the telephone calls across PST is the main
function of wireless core network. Also, it provides various services to users who
are connected in a network.
Programming models:
Existing programming models in MCC are: 1) Clone cloud, 2) MAUI, 3)
Odessa, 4) Orleans, 5) RESTFUL [9]. These programming models are briefly
discussed below. Table 1 illustrates comparison of programming model based
on blocking state, cloud state and remote execution unit.
Figure 1. Mobile cloud architecture.
Table 1. Programming model comparison.
Models Blocking Cloud state Remote exec. Unit
Clone Cloud Yes Full thread Thread
MAUI Yes Partial Method
Odessa Yes Partial App task
Orleans No Partial Grains
RESTful No No Cloud task
3
-----
A. Paranjothi et al.
4
1) Clone cloud: Clone cloud allows its users to have own copy of their cloud.
By providing this facility, user will have full control over their clouds. Clone
cloud consists of solver, profiler and analyzer. Solver in clone cloud is responsi
ble for offloading the data on cloud. It will be based on dynamic profiler and
static analyzer.
2) MAUI: This programming model is based on Microsoft.NET framework.
Profiler in MAUI framework makes remotable decisions. Resource demanding
process can be accessed with the help of Remote Procedure Call (RPC). This
model is platform and language independent.
3) Odessa: It is a parallel processing framework where developers have to ar
range their applications in the form of data flow graph. In graph, vertices are
called as stages and edges are called as connectors. In Odessa, connectors give
information about data dependency between stages. This programming model is
mostly suitable for media applications. Existing applications cannot be accessed
in Odessa framework.
4) Orleans: It is the reliable framework for establishing scalable, elastic appli
cations on cloud. Orleans consists of grains, which uses asynchronous messages
for communication. Application developer in Orleans mainly concentrates on
logic since it provides scalability, reliability and availability during its runtime. It
is one of the promising programming models in MCC environment.
5) RESTFUL: This programming model is developed due to media processing
applications often requires components for gesture recognition, face recognition,
etc. In this model, appropriate functions can be invoked whenever they are
needed. It can be done by using http or https protocol.
## 3. Offloading Mobile Applications in Cloud
In recent days, cloud computing research has been moved towards how to make
offloading decisions rather than concentrating on making offloading feasible.
Analytical model helps in making these decisions [12] [13]. Offloading and par
allelism are the two main factors that impact the system performance. In this
section we illustrated the existing frameworks suitable for offloading in mobile
application environment.
Offloading: Transferring computations to servers available on the cloud are
called offloading [8]. Offloading decisions can be done in two ways 1) Manually
by the developers [14] 2) Automatically using tools [15].
### 3.1. Odessa Framework
Odessa is a lightweight framework, designed for mobile applications [17]. Odes
sa makes offloading more flexible. Mobile application has three main require
ments: 1) Crisp response, 2) Continuous data processing 3) Algorithms should
be computed intensive. This framework provides three major contributions: 1)
Odessa contributes to offloading and parallelism decisions. 2) Odessa designs a
light weight mobile interactive perception applications. 3) It works well across
variety of execution environments. The authors used three different applications
-----
A. Paranjothi et al.
to measure their system performance. The applications are described below in
detail.
**Interactive Perception Applications**
Face Recognition:
Face recognition application is represented in Figure 2(a). Face detector and
classifier are the main components involved in it. Face detector is used to detect
faces using OpenCV [18], Haar classifier and face classifier will classify the faces
detected by face detector using a dedicated algorithm.
Object and Pose Recognition:
Object and Pose recognition application is represented in Figure 2(b). Four
Feature Graph
Source Copy Tiler Detect
merger Splitter
Reco
Display Classify
merge
(a)
Feature De Feature
Source Copy Scaler Tiler SIFT
merger scaler splitter
Cluster Cluster Match Model
Display RANSAC Clustering
joiner splitter Joiner matcher
(b)
Pair Motion Feature
Scaler Tiler Descaler Copy Classify
Generator SIFT manager
Source Copy Display
Face Face
Scaler Tiler Descaler Copy
Detect Merger
(c)
Figure 2. (a) Face recognition; (b) Object recognition; (c) Gesture recognition.
5
|Source|Col2|
|---|---|
|||
|Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|
|---|---|---|---|---|---|---|---|---|---|
|Scaler|||Tiler||Motion SIFT||Feature manager|Descaler||
|Classify|Col2|
|---|---|
|Col1|Col2|Pair Generator|
|---|---|---|
||||
|Source||Copy|
||||
||||
|||Scaler|
||||
-----
A. Paranjothi et al.
6
main components are involved in it. First, the image will go through a downscale
to extract SIFT features [20]. Second, the extracted SIFT features will be com
pared with previously constructed 3D models. Object features are clustered by
position to isolate different occurrences. RANdom SAmple Consensus (RANSAC)
algorithm identifies each occurrence of image with estimated 6D posture.
Recognition of Gesture:
Gesture recognition application is represented in Figure 2(c). Face detection
and motion extraction are the major components involved in it. It extracts SIFT
features to encode optical flow. These features are then filtered by positions to
compare with previously generated histograms. The histograms are used as an
input for machine identifies the control gestures.
### 3.2. Sprout
Sprout is a distributed system used for stream processing [21]. It is used to
create and execute parallel processing applications [17]. The main goal of sprout
is to support processing of high streaming data. Two important features of
sprout are: 1) Automated data transfer, 2) Parallelism support. Also, sprout pro
vides the mechanism of adjusting applications dynamically at run time, changes
the degree of parallelism, migrating processing stages between machines.
### 3.3. Odessa Design
Odessa uses the concepts of offloading, pipelining and data parallelism to im
prove its performance and accuracy. Odessa has three main goals they are
1) To satisfy the need of mobile applications Odessa should accomplish low
make span and high throughput.
2) It should concentrate on input complexity change, device capability and
network conditions.
3) It should have low communication and computation overhead.
### 3.4. ThinkAir Framework
ThinkAir is one of the simplest frameworks in Mobile Cloud Computing (MCC)
[22], represented in Figure 3. It allows developers to migrate their software to
the cloud. Smart phone virtualization and method-level computation offloading
are two main concepts adopted by ThinkAir. Offloading in ThinkAir removes
the restrictions caused by CloneCloud during the process of offloading [22]. It
also on-demand resource allocation for efficient performance of an application.
Parallelism is attained by dynamically creating, destroying virtual machines in
the cloud.
### 3.5. ThinkAir Design
ThinkAir framework is designed based on following parameters: 1) Mobile
broad band connectivity and speeds are increasing continuously, 2) Capabilities
of smart phones are increasing, 3) Cloud computing is becoming more popular,
and provide resources to users at low cost. The design goals of ThinkAir are:
-----
A. Paranjothi et al.
Figure 3. ThinkAir framework.
1) Adaptation
ThinkAir framework easily adapts according to environment and it also
avoids interference of correct executing software.
2) Ease of Use
ThinkAir framework provides simple interface for developers to avoid the is
sue of misusing framework [22] and it increased competition among developers.
3) Performance improvement
ThinkAir framework improves performance and efficiency of mobile devices
by binding smartphones to the cloud.
4) Dynamic scaling
ThinkAir provides the feature of calculating computational power dynamical
ly at server side. It also provides parallel executions to improve the performance.
This framework has three major components: 1) Execution Environment, 2)
Application servers, 3) Profilers.
### 3.6. Compilation and Execution
Compilation and Execution section of ThinkAir deals with three major areas: 1)
Programmer API, 2) Compiler, 3) Execution controller.
7
-----
A. Paranjothi et al.
8
1) Programmer API
ThinkAir contains library with compiler since the developer can access execu
tion environment indirectly. Method considered for offloading is commented
with @Remote. ThinkAir code generator generates necessary remote able me
thod with utility functions by taking source file as input and execution controller
is used for method invocation and detects the given method is suitable for of
floading or not.
2) Compiler
ThinkAir Compiler consists of two parts 1) Remoteable Code Generator, 2)
Native Development Kit (NDK). Remoteable code generator used for annotated
code translation and Native Development Kit (NDK) used for native code sup
port in cloud.
3) Execution Controller
Execution controller executes remotable methods and makes offloading deci
sions. Offloading decision depends on the data collected during past execution,
the current environment and user’s policies. There are four such policies com
bine with execution time, energy and cost. The four policies are, 1) Execution
Time, 2) Energy, 3) Execution time and energy, 4) Execution time, energy, and
cost.
4) Execution Flow
Execution Controller starts with profiler to provide data for future invocations
and it decides this invocation is suitable for offloading or not. If it is suitable for
offloading, then it can be migrated to the cloud using java reflection technique.
If it is not suitable for offloading or if connection fails, then execution will back
to local execution environment (i.e., smart phone) by eliminating the data col
lected by the profiler.
## 4. Mobile Distribution Computing and Cloud
Mobile distribution computing will provide access to widely distributed re
sources. Distribution computing has the advantages of scalability, fault toler
ance, and load balancing. In many situations processing tasks needs to be distri
buted. However, in distribution computing there is chance of communication
failure because it could fail at any time. This section gives the information about
various distribution techniques used in the cloud.
### 4.1. Clone 2 Clone (C2C)
Clone 2 Clone (C2C) [23] provides distributed peer to peer platform for smart
phones. Performance measurement of C2C in private and public clouds shows
that it is possible to implement C2C in distributed environment with 3 times
lesser cellular traffic. In addition, it also saves 99%, 80% and 30% of the battery
respectively.
### 4.2. C2C: Architecture Design
C2C platform needs a mechanism to enable peer to peer networking, to notify
-----
A. Paranjothi et al.
others about the presence of others. In C2C, CloneDS (i.e., Clone Directory Ser
vices) maps its users to clones and clones to IPs. To establish a connection in
C2C platform, request a clone with public IP and key pair. Key pair can be pub
lic key or private key. C2C architecture is represented in Figure 4 and it consists
of five basic steps: 1) DS register, 2) DS lookup, 3) C2C Connect, 4) User lookup,
5) User clone connection.
DS register: Clone id, public key, IP address and device id will be send to clone
DS from the clone.
DS lookup: Clone A obtains a list signed by CloneDS. The list contains the
details of signed clones with their IP address and public keys.
C2C connect: Peer to peer connection will be established with other clones
from clone A.
User lookup: User A can always get her clone’s IP through a CloneDS lookup.
User clone connection: It establish a connection with user through Public IP.
### 4.3. C2C and Security
In C2C, communication between the user and clone is secured by the shared
symmetric key. This architecture provides trust to the users through CloneDS
but some destructive cloud providers use this opportunity by connecting user to
harmful ones.
### 4.4. CloneDoc Framework
CloneDoc and SPORC [24] provide more complexity to the system, but the main
advantage of using such system is that it will improve the battery performance
by reducing its usage. CloneDoc receives the operations from user’s smart
phones and keeps the device updated. The clone maintains two states 1) pending
queue, 2) committed queue [25]. The clone in C2C delivers operation received
from device to server and again send backs the result to appropriate devices. To
handle these tasks, it maintains a queue. It should be managed in such a way that
Figure 4. C2C architecture and networking.
9
-----
A. Paranjothi et al.
10
delay should be minimum. CloneDoc contains a protocol to solve this problem.
It is commonly known as clone user consistency protocol. In CloneDoc, clone is
also responsible for detecting server malfunction. Detecting server malfunction
contains checking of encryption and decryption operation, sequence numbers,
etc.
### 4.5. Code in the Air
Code In The Air (CITA) [26] is a system which simplifies the rapid development
tasking applications. It can be handled by both expert users, non-expert users. In
CITA non expert users specify their tasks easily over phone and expert users
specify their tasks by writing server side scripts. Current approaches have two
major problems: 1) Poor abstraction, 2) Poor programming support. CITA helps
developers as well as end users.
### 4.6. CITA Architecture
CITA architecture has following 3 components, 1) Tasking framework, 2) Activ
ity layer, 3) Push communication and it is represented in Figure 5.
Tasking framework: It allows developers to write and compile scripts. Compilation of scripts can be done in server side. CITA also provides JavaScript interface for its developers to manipulate different devices in single program. Backend of CITA deals with device coordination and efficient execution of code on
different devices
Activity layer: Activity layer in CITA provides an extensive abstraction to high
level activities and also it provides facility of energy efficient recognition of an
activity.
Push communication: it improves on the energy and load shortcomings of ex
isting systems.
Figure 5. Code in the air architecture.
-----
A. Paranjothi et al.
### 4.7. CITA Activity Layer
CITA contains an activity layer. The main purpose of activity layer is to express
conditions.
**4.7.1. Place Hierarchies**
CITA uses a built-in location hierarchy, it identifies three types of locations: 1)
Room level, Floor level 3) Building level. These hierarchies are having different
implementations.
Room level location hierarchy: It is used to match a named location if Wi-Fi
signal strength is good.
Floor level location hierarchy: It is used during overlapping of Wi-Fi signals.
Building level location hierarchy: It is used to refer the buildings or large
bounding box on a map.
CITA contains two activity detectors: 1) enterPlace, 2) leavePlace. It will be
called when a user enters and leaves a location respectively [26].
**4.7.2. Activity Composition**
CITA allows developers and users to create high level activities using logical
predicates [26]. This is one of the advantages of CITA because it provides reusa
bility. Developers and end users can reuse activity modules created by other de
velopers to write their own activities. CITA supports AND, OR, NOT, WITHIN,
FOR, NEXT primitives.
### 4.8. Dial to Deliver Push Service
CITA provides an asynchronous message delivery service to mobile devices from
CITA server [26] but it has three major problems.
1) The information to be delivered is very less.
2) TCP connection in mobile devices leads to timeout due to long waiting
time.
3) Current push notifications limits notifications of specific types.
CITA uses standard telephone service. It contains the registered users phone
numbers. To verify the user, CITA server initiates a voice call on the other end
CITA client verify the phone number. If the number matches it wakes up the
client. The main disadvantage of using this service is, load on the network will be
increased.
### 4.9. WhereStore
WhereStore provides location based storage for mobile devices. It uses the tech
nique of filtered replication to distribute location history among mobile devices
[27]. WhereStore reduces energy consumption by exchanging data in clouds.
Location specific applications are the one differentiates computer from mobile
phones [28]. The following applications are benefitted from WhereStore frame
work: 1) Web applications, 2) Media player, 3) Live traffic and sensing applica
tions.
11
-----
A. Paranjothi et al.
12
### 4.10. Where Store Background
WhereStore is designed using two techniques: 1) location prediction, 2 Replica
tion system [27].
**4.10.1. Replication**
Replication system uses collection as its major technique. It is used to maintain
data synchronization between peers. Collection in a replication system has user
data and meta data. It will be represented as separate items. In a filtered system,
filter identifies the subset where the data is stored exactly. Consider the example
of separating the JPEG images according to geographical region. It can be done
by identify the images according to the geotag attached in it. There are two ma
jor goals involved in filtered replication system:
1) Each clone stores exactly matching item in its filter.
2) In each clone version of items should be same.
**4.10.2. Predicting Location**
GPS in mobile devices provides location based services to the user. Location
based services provides the advantages of tracking the user [29] [30]. The main
idea of using location prediction in WhereStore is to predict the future location
of the user by matching past location history of the user with present location
[31].
### 4.11. WhereStore System Framework
WhereStore provides dynamic data storage for its user’s. It will be based on two
parameters: 1) past location of the user, 2) present location of the user. The con
ceptual view of WhereStore is represented in Figure 6. WhereStore provides
complete control to its user. Here, data will be grouped according to the graphi
cal regions where the is likely to stay. It ensures availability of data in user’s cur
rent location. WhereStore provides complete transparent mechanism for data
placement when compared to other frameworks.
Figure 6. WhereStore conceptual view.
-----
A. Paranjothi et al.
WhereStore will be located top of the replication system and location service.
In which, replication layer consists of various collections. Each application in
WhereStore has separate collections. The replication layer creates clones (i.e.,
replicas) for mobile devices and cloud. Each clone has its own storage capacity
and filter. Storage capacity of each clone specifies the maximum number of
bytes. WhereStore semantics are same as cache (i.e.,) the cloud can be accessed
only when the item is not available in a local environment. The filter located in
each clone will be adjusted according to the current location of user.
**4.11.1. Types of Data**
WhereStore acts on groups, regions and items. Each item can be identified using
its key and priority associated to it. Data can be divided into several items.
Groups are set of items and regions gives information about different geograph
ical area. WhereStore has a separate interface to create application, maintain re
gions and groups. According to the geographical areas regions will be created
and each region will be associated with multiple groups.
**4.11.2. Filters**
Filters in WhereStore specify the items stored in a given location. It has set of
filters each possible future location. Future locations will be identified using
current location based on the location prediction system. Consider (l1, l2, l3, …
ln) be the future location with probability pi. WhereStore create new filters (fi)
for each possible future location. When the devices location changes new set of
filters which is computed and updated recently will be passed into the replica
tion. Each replication system has its own probability pi and maximum storage
capacity. The rank of items will be based on occurrences. If a particular item is
located in more filters, rank will be high.
**4.11.3. Cloud Synchronization**
Data exchange is performed in replication platform using synchronization estab
lished between cloud and smartphones. It provides the advantage to the smart
phone by having only items which matches exactly with filter. Smartphones suf
fers due to limited storage capacity. To utilize the space in a proper manner,
smartphones filter will be evaluated in a cloud. Each filter will be associated with
a cloud calculates the set of items matches with filter and rank will be provided
to each item. Storage capacity of each item calculated according to the rank.
### 4.12. Implementation of WhereStore
WhereStore implementation is based on clientserver architecture. Here, cloud
act as server and user’s mobile device act as client. The client has two major
components 1) location, 2) Replication. Location specifies the information about
future smartphone location and contains information of local cache memory on
cloud. Architecture of WhereStore is represented in Figure 7. Whenever appli
cation interacts with WhereStore configuration file needs to be provided. Con
figuration file specifies replication at a particular location. Filters will be updated
13
-----
A. Paranjothi et al.
14
Figure 7. Architecture of WhereStore.
based on input given by the configuration file. Later the updated filter will be
used by the replication system.
**4.12.1. Mobile Phone Data Access**
WhereStore use existing data storage for accessing data. There are numerous ex
isting applications stores the data in their own way. WhereStores uses the con
cept of Cimbiosys as replication system [32]. It uses callback mechanism for ac
cessing the data and implemented whenever needed. Cimbiosys determines the
data to be broadcasted during the synchronization process. Wherestore is re
sponsible for creating metadata whenever new item added into it.
**4.12.2. Synchronization of Cache**
In WhereStore, Cimbiosys synchronization exchange messages based on a tech
nique called pull style exchange. It will be done in one way. Here, target clone
will establish a synchronization with source clone. The connection will be estab
lished by sending a request message. Once the connection gets established,
source clone starts checking any of its tem is are not admitted by target clone. If
any of such items exist, source clone will return the corresponding item to target
clone. The current Cimbiosys model can be expressed in two ways: 1) modifica
tion of sync request from mobile device to cloud, 2) modify the filter based on
Cimbiosys.
**4.12.3. Location Prediction**
WhereStore uses location prediction technique to predict its user’s possible fu
ture location. It can be achieved by using StarTrack framework [33]. In Star
Track framework, location of a user will be captured periodically using smart
phones. The captured location will be forwarded to cloud where StarTrack serv
ers will be located. StarTrack server will convert the location it is received from
smartphones into tracks. It provides a API (Application Program Interface) to
perform operations on tracks. In order to identify where the user will be located
in a future, StarTrack uses a technique called place transition graph. This tech
nique will be created based on the tracks generated by StarTrack framework. It
-----
A. Paranjothi et al.
also has the detail of places that are visited by the users frequently. Latitude and
Longitude pair will be used to create this FrequentPlaces. It will be created based
on the tracks. Place transition graph will be constructed based on FrequentPlac
es. Usually it will be represented in the form adjacency matrix. First, all the ele
ment in the matrix will be initialized as zero. Then, set the corresponding value
of each item by combining start and end points. Finally, normalize the value in
each row according to the probability. This normalization can be done by adding
the trip frequencies of each row and divide each frequency with sum of row.
### 4.13. Virtual Machine Synchronization (VMsync)
The utility of the mobile device will increase if the user wants to switch from one
device to another device. For example, user is able to continue the operation in
the second mobile device absolutely where the user left in first mobile device
without any delay. VMsync [34] used to synchronize virtual machines (VMs).
This synchronization will take place among mobile devices. In device switching,
VMs encapsulates computation state and data for a complete operating system
and applications associated with it. In VMs application state also getting syn
chronized along with mobile devices. System level VMs are used in now days to
provide improved security and manageability. In order to reduce the delay and
make image consistent, VMsync is used. It will transfer the changes made in ac
tive VM to other mobile devices. The most important component in VMsync
architecture is known as daemon. The main purpose of daemon is to monitor
the memory and filesystem of VM. If any changes made, it will report the
changes to the server. Server will be located on the cloud, and it will send the
changes to devices.
### 4.14. Preliminary Design of VMsync
The main function of VMsync is to handle VM images across various mobile
devices and reducing time between switching devices. It uses method called
Switch Penalty to perform device migration. The disadvantage of using this me
thod is, data transfer cost is high. VMsync architecture is represented in Figure
8. It contains multiple hosts and provides following facilities to the user: 1) Vir
tualization support, 2) Resource rich server, 3) Synchronization between devices.
Initially, VMsync had only one active device. This device is used to update server
regarding memory, file system changes over a periodic time. It can be done only
when the device is active. This process is known as checkpoint.
VMs other than active VM are known as standby VM. These VMs are getting
updated with the help of synchronization server periodically. But, during the
updating process device should be connected to network. Synchronization dae
mon used to monitor this process. VMs should be designed in a way that it can
balance data transfer and computational overhead. Modern mobile operating
systems like Windows, Android, iOS, provides support for different hardware
manufactured by various companies. This functionality can provide the facility
adapting changes in hardware during runtime in future.
15
-----
A. Paranjothi et al.
16
Figure 8. VMsync architecture.
### 4.15. Wireless Mesh Networks (WMN)
Wireless Mesh Networks (WMN) provides a low cost, next generation wireless
networking, and also it provides a high speed internet access. Wireless Mesh
Network supports wide range of mobile applications. Wireless Mesh Networking
with Mobile Cloud Computing (WM-MCC) is considered to be a best solution
for large scale big data applications [35]. In Wireless Mesh Networks mobile
client is connected to a Base Transceiver Station (BTS) and it access the mesh
network via mesh router. While mesh routers will be connected to each other
and it will communicate with cloud through internet. The cloud service platform
in wireless mesh networks provides data query services.
## 5. Privacy in Cloud
Privacy is a major component in MCC. The user needs to understand the standards and procedures provided by the cloud provider to protect their data from
threats. The number of businesses and individuals that are moving their data
and performing computation on cloud is increasing. Although the cloud computing provides numerous benefits, security remains as one of the major challenge
when data and computation are utilized by untrusted third parties [36]. The following section provides the information about different security approaches
used in the cloud to protect user data.
### 5.1. Secure Outsourcing of Collective Sensing and Systematic Applications to the Cloud (p-Cloud)
Two main approaches were proposed to provide security. i) StreamForce, ii)
CloudMine.
Streamforce: It is an access control system for sharing of data over malicious
and untrusted clouds. This approach is designed with three goals:
1. To provide support to specify and impose fine grained access control policies.
2. To outsource data to cloud if access control methods are enforced.
-----
A. Paranjothi et al.
3. When handling most expensive computations system will be efficient.
CloudMine: It is an on-demand and cloud-based service with which different
data owners achieve secure analysis over their collective data. This approach
supports three essential functions: 1) sum, 2) set union and intersection, 3) sca
lar product. CloudMine attains three security promises,
1. It provides data confidentiality in contrast to colluding, semi-honest data
owners and semi-honest clouds.
2. Protection is provided to outputs of joint computation against semi-honest
clouds.
3. Data owners can accurately identify if the cloud has been lazy.
### 5.2. System Model for p-Cloud Approach
Figure 9 represents our system model for deploying collective applications to
un-trusted clouds. It includes two entities: 1) client, 2) cloud. On the cloud, a
collective task that consists of joint data and computation from different clients
is performed. The attacker cloud consists of the un-trusted cloud and clients.
There are three levels of un-trustworthiness:
1. Curious but Honest
2. Curious and Lazy
3. Fully Malicious
Curious and lazy model allows the attackers to compromise while operating
carrying out outsourced undertakings. Particularly, the cloud attempts to learn
sensitive information and does not effectively corrupts the computation, but ra
ther it tries to do as limited as possible while charging the customers for the
same. This model is legitimized by the economic incentives to overcharge clients
without being distinguished, however, there are three security properties relating
to this framework.
1. The outsourced data should be protected for input privacy.
2. The outputs should be protected for output privacy.
3. Three parameters are included in integrity. Namely: a) Correctness, b) Com
pleteness and c) Freshness.
Streamforce approach accomplishes input and output privacy as well as cor
rectness in the curious but honesty model. CloudMine accomplishes similar
properties yet in the curious and lazy adversary model.
Figure 9. Collaborative applications on the un-trusted cloud.
17
-----
A. Paranjothi et al.
18
### 5.3. Streamforce Approach
Streamforce approach utilizes a fine grained access control system to share data
in un-trusted clouds. Implement fine-grained access control in collective appli
cations that are out-sourced to an un-trusted cloud. There are two roles assigned
to client. Namely a) Data Users, b) Data Owners. An access agreement P is given
to user and the owner is provided with private data x along with related attribute
I. When the data attribute fulfills the strategy, i.e. P(I) = True, then an autho
rized client can get access to x. First, the owner sends c = f(x;I) to cloud by using
a encoding function f. Later, the cloud changes the encoded data as t = π(c).
Lastly, the client assesses a function g(t). In this setting, input privacy infers that
the cloud cannot learn x from c. From output privacy and correctness, it is im
plied that the access control methodology is secure, that is the unauthorized
access is not permitted: g(t) = x ↔P(I) = True.
Figure 10 illustrates the design space for access control implementation on a
cloud domain. It is described into three measurement strategy. Namely: a) fine
graininess, b) cloud reliability and c) cloud/customer work proportion. Trusted
cloud can accomplish best fine-graininess. It supports an extensive variety of
approaches and accomplishes best work proportion. Streamforce is particularly
intended for stream data. It is outlined with three goals:
1) It supports specification and enforcement of fine-grained access control
policies.
2) Access control policies are enforced when data is outsourced to the cloud.
3) System is efficient when handling most expensive computations.
Streamforce security depends on three main encryption strategies. Namely: a)
Deterministic εd, b) proxy attribute-based εp and c) sliding window εw.
Figure 10. State of the art in outsourced access control.
-----
A. Paranjothi et al.
Deterministic Encryption technique (DET): It is a private-key strategy that is
semantically secure while encoding multiple plaintexts. εd = (Gen, EncDec),
εd∙Enc(m) = εd. _Enc m(_ ′) ↔ _m_ = _m′_ . The Proxy Attribute Based Encryption
technique (PABE) broadens the idea of Key Policy Attribute Based Encryption
(KP-ABE). εp = (Gen, KeyGen, Enc; Trans, Dec) [37]. Specifically, a master key
MK is generated by Gen(.), a transformation key TK is generated by KeyGen
(MK, P) and a predicate P is given by decryption key SK. By utilizing the
attributes A, Enc (m, A) encrypts m. Trans (TK; CT) partly decodes the cipher
text, which is later decrypted by Dec(SK,CT’).
Decryptions and transformation are effective if P(A) = True. Streamforce util
ize the strategy provided in [38]. The Sliding Window Encryption technique
(WE) permits a client to decrypt just the aggregate of window of ciphertexts but
not the individual ciphertexts, εw = (Gen, Enc, Dec). Assume that p(M, ws)[i]
and s(M, ws)[i] are the product and sum of i[th] window sliding windows upon a
sequence M. The general public parameters and the private keys are created by
Gen(k). M is encrypted by Enc (M = (m0,m1, ….,mn-1),W) by utilizing a set of
window sizes W, whose outcome is CT = (c0, c1, c2,…). CT is decrypted by Dec
(ws,CT, SKws) for the window size ws by utilizing the private key SKws. s(M; ws)
[i] for all I is the outcome, which is the aggregate of the sliding window.
Secure query administrator
Encryption is determined as a strategy that is used to secure data confidential
ity in contrast to the cloud and unauthorized client access. However, straightly
presenting encryption details to system entities is not considered as a perfect
reflection for access control. Rather, Streamforce models and authorizes access
control strategies by means of an arrangement of secure inquiry administrators
like: 1) secure Map, 2) Filter, 3) Join and 4) Aggregate.
Evaluation
Execute a model of Streamforce over Epser (An open source stream process
ing engine proficient of handling millions of data items every second). Make a
benchmark dataset similar to stock market data and containing one million
tuples that belong to 100 streams[7]. Throughput and latency of Streamforce are
examined through the experiments conducted on Amazon EC2 with 6 multiple
policies (T1-T6). The throughputs for various strategies on a single cloud server
are demonstrated in Figure 11. The maximum throughput is observed for sim
ple strategy using Map operator, which is 250 tuples/sec. This analyzes ineffec
tively against Esper's execution on plaintext data. In addition, experiment is also
carried with different cloud servers and the outcomes demonstrated scalability
for both latency and throughput. We shared the workload in- two ways when
more cloud servers are added. They are: a) Simple: based on stream, b) Balanced:
based on computation load.
### 5.4. Practical Confidentiality in Preserving Big Data Analysis
Cloud Computing provides support to Big Data Analysis [39] via data flow lan
guages known as Pig Latin [40]. It is of great value to manage sensitive data only
19
-----
A. Paranjothi et al.
20
Figure 11. Throughput on single server.
in an encrypted from in the cloud and to perform reasonable data analysis.
Crypsis, is a runtime system for Pig Latin which allows corresponding scripts to
be executed efficiently by utilizing cloud resources however without exposing
input data in the same form. Crypsis can broaden the scope of encryption empowered big data analysis depending on the following point of view:
i) Perspective of Extended Program
Multiple opportunities to operate in encrypted mode are identified by Crypsis
by evaluating entire data flow programs.
ii) Perspective of Extended System
Cloud resources can be performed by Crypsis instead of giving up and driving
users to run the entire data flow programs on their end. This can be done by
considering the chance of performing small computations on user end.
Three main Contributions for preserving big data analysis
1) Without sacrificing confidentiality propose an architecture for executing
Pig Latin scripts
2) Outline a novel-field sensitive program to study and transform to Pig Latin
scripts that can distinguish operation with effects.
3) Current fundamental assessment results for implementing the solution depending on run time Pig Latin scripts obtained from an open source Apache
PigLatin.
### 5.5. Background: PigLatin
Apache Pig is a data examination platform. It incorporates the Pig runtime
framework for high-level data flow language Pig Latin [40]. Pig permits data experts to query big data without the complication of writing MapReduce programs. No fixed schema is required to be operated by Pig. All these properties of
Pig Latin and in addition its wide reception made it to be chosen as the data flow
language for Crypsis.
-----
A. Paranjothi et al.
Data types and Statements
Pig Latin includes simple types (e.g., int, long), and complex types (e.g., bag,
tuple, map). In addition, field can be a data item like a tuple, bag, map. Pig
Latin statements also work with relations; relations are simply a bag of tuples.
Expressions and Operators
Relations are established by loading an input file or by applying relational op
erators to different relations. Examples of relational operators are JOIN,
GROUPBY, FOREACH. etc. Operators in Pig Latin can likewise incorporate
casts, arithmetic operators (e.g., +, −, \, *), comparisons, as well as LOAD and
STORE operators.
Functions
Pig Latin incorporates built-in functions (e.g., ABS, COS AVG) and allows
users to define their own user defined functions (UDFs) if needed.
### 5.6. Architecture and System Overview
Crypsis is having an adversary capable of fully manipulating the cloud infra
structure. The adversary can see encrypted data and Pig Latin scripts that oper
ate on the data and it can control the computation software and control the
cloud infrastructure. Crypsis ensures confidentiality in the presence of adver
sary. Figure 12 illustrates the architecture of Crypsis prototype.
1. Transformation of program
The client presents a source Pig Latin script that works on unencrypted in
formation. This is evaluated by Crypsis in order to find the suitable encryption
scheme through which the input data must be encrypted. Calls to Crypsis UDFs
which implement operations on encrypted data are used to replace the operators
in source script. The constants are supplanted using their encrypted values in
order to create an objective script that can be executed on encrypted data.
2. Encryption techniques missing in cloud
The parts of input data that are encrypted previously and stored in cloud are
tracked by an encryption service which contains an input data encryption sche
Figure 12. Architecture of Crypsis.
21
-----
A. Paranjothi et al.
22
ma. Depending on the input data encryption schema as well as the recommended encryption schemes assumed in previous step, the encryption service determines the encryption schemas lacking in the cloud.
3. Encryption, Sending data to cloud
Various encryption schemes which are enabled through diverse cryptosystems
are used by Crypsis.
1) Randomized encryption (RAN) is the main encryption scheme which does
not support operators and best secure encryption scheme. One way to execute
RAN is by utilizing Blowfish [41] in order to encrypt integer values by exploiting
the benefits of its limited 64-bit block size and also by utilizing AES [42] in order
to encrypt the remaining.
2) Deterministic encryption (DET) let’s fairness comparisons upon encrypted
data. First, develop DET utilizing AES and Blowfish permutation block ciphers
for estimations of 12 bits and 64 bits respectively. Then, pad minute values
properly to coordinate the normal block size. The approach for values greater
than 128 bits, check the approach utilized in CryptDB [43]. Later, implement the
Order Preserving Encryption (OPE) scheme that permits to arrange correlations
utilizing the order preserving symmetric encryption usage from CryptDB. Paillier cryptosystem to implement additive homo-morphic encryption (AHE)
which allows additions over encrypted data and ElGamal [44] cryptosystem to
implement multiplicative homomorphism encryption (MHE).
4. Execution
When all required encrypted data is loaded in the cloud, the execution handler requests to start executing the job.
5. Crypsis UDFs
Crypsis does not impose any changes to the PigLatin service. Instead, operations on encrypted data are handled by a set of pre-defined UDFs stored in the
cloud storage along with the encrypted data.
6. Re-encryption
At the time of target script execution, it is possible that intermediate data are
generated after some operations are performed. Encryption scheme of the particular data relies upon the previous operation executed on that data. This situation is handled by Crypsisthrough re-encryption of intermediate data. In particular, this intermediate data is directed to the user where this data can be decrypted without any risk. Later, the decrypted data is encrypted using the specific encryption scheme and then again sent to cloud. After the re-encryption is finished, execution of target script is continued.
7. Results
Results are again sent to user when the job is finished.
### 5.7. Program and Transformation Analysis
Analysis and process involved in PigLatinCrypsis are represented briefly in Listing 1. It has two input files: input1, input2. Input1 has two arguments and input
2 has one argument. Line 3 in script is used to filter all rows less than or equal to
10. The subsequent lines (i.e., Line 4 and Line 5) perform addition operation on
-----
A. Paranjothi et al.
second argument of input1 by grouping first argument. Each group sum will be
performed in Line 6 using input 2. Finally, Line 7 displays the result stored in
output file. Figure 13 illustrates outline of the different steps and intermediate
data structures in program transformation.
Input script analysis
First, Crypsis checks the user submitted source (PigLatin) script for syntax
errors and generates a directed, acyclic data flow (DAG) representation of it. The
data flow representation uses relations as vertices and the data flow between relations as the edges. Also generate two additional data structures: 1) MET (Map
of Expression Trees) 2) SAF (Set of Annotated Fields) [40]. Source script expressions will be stored in MET. In data flow graph, each vertex has keys for all expressions. SAF contains one entry for each field specified for each relation. In
input script analysis, AF (Annotated Field) is used to represent individual entries.
Figure 13. Program Transformation in Crypsis.
Listing 1. Source pig latin script S1.
Listing 2. Transformed pig latin script.
23
-----
A. Paranjothi et al.
24
Encryption analysis
The program transformation component identifies the encryption scheme
required for each field. It identifies the encryption of each field by observing
MET. In script, all operators are already registered with encryption technique.
But, some relational operators involved in PigLatin require precise encryption
schemes.
Script transformation
After knowing encryption scheme required for each field, decision will be
made for which encrypted file to be loaded. Script checks valid encryption tech
nique, if it is not available re-encrypt operation will be initialized. It calls encryp
tion scheme to change the required fields into a specific encryption technique.
The transformed PigLatin script is represented in Listing 2.
### 5.8. Evaluation of Big Data Analysis
Micro-benchmarks
Construct a micro-benchmark that compares unencrypted data with en
crypted data based on the size and time requires to execute. It is represented in
Table 2. The evaluation of this micro-benchmark was performed on a single
machine with two 32 bit CPUs and 3 GB of RAM. While running benchmark,
one problem we faced was in PigLatin scripts that projects the value of map
fields using chararrayconstants as keys (map#’key’).
PigMix
Run the Apache PigMix2 [45] benchmark to calculate the Crypsis perfor
mance. PigMix2 is a set of 17 Pig Latin scripts that tests the latency and scalabil
ity of the Pig runtime. The experiment was performed using Amazon EC2 [46].
### 5.9. Information Leakage
Information leakage [47] gives the information about how the privacy is getting
disturbed in mobile environment. Two types of attacks are possible in mobile
environment: 1) External attack, 2) Internal attack. Both attacks are used to ex
tract the user information. It is possible to perform these attacks without the us
er using devices of attacker.
Table 2. Comparing size of data and latency of addition and multiplication operations
over plaintext and encrypted data. †^ is the number of operations performed in multiples
of 1000. ‡NE denotes no encryption or plaintext data.
Size (KB) Time (ms)
Add Multiply
^[†] NE[‡] AHE MHE NE AHE NE MHE
2 269 12,071 12,153 32 477 32 2267
4 538 24,142 24,306 63 895 62 4118
6 807 36,212 36,459 92 1314 90 5978
8 1076 48,283 48,611 121 1730 118 7818
10 1345 60,354 60,764 150 2147 147 9658
-----
A. Paranjothi et al.
### 5.10. Background of Mobile Analytic Service
Background of mobile analytic service concentrates on developers, users, appli
cations, networks, etc. This section deals with app ecosystem and mobile analyt
ics.
**5.10.1. App Ecosystem**
Application developers use a technique called ad networks to increase the profit
of applications. Recent study shows that, top applications available in Android
Market (i.e., 52.1% of applications) are enclosed with at least one ad networks.
App ecosystem is represented in Figure 14. It illustrates the flow of information
among users. Mobile applications are enclosed with analytic library. Main func
tion of analytic library is to collect attributes related user and send it to the serv
ers maintained by analytic companies. The information will be processed and
will be given to ad networks like Flurry, Google Ad, etc. to provide appropriate
ad for the user.
**5.10.2. Mobile Analytics and Tracking**
Mobile analytics are used to measure the performance of the applications based
on prior knowledge about users, applications, etc. Dashboard performance of
flurry is represented in Figure 15. It gives the information about various inter
ests of a user.
### 5.11. User Profile Extraction
User profile extraction used to extract various information about users. Different
services are used to collect distinct information about users (i.e., name, age, etc.).
In order to extract information about user, first step is to act on behalf of user.
Next step for google is to extract the user profile illustrated by google. For flurry,
target information should be send to analytics application which in turn extracts
user profile.
Figure 14. App ecosystem.
25
-----
A. Paranjothi et al.
26
Figure 15. Flurry analytics.
**5.11.1. Device id Spoofing**
Getting access to device id.
Device id of an android user can be accessed by using two different methods.
1) It can be obtained by grabbing message send by third party.
2) It can be extracted by capturing identifier of a target device.
Device id spoofing
Android users can be easily identified by combining device id with device in
formation. It is possible to device information using the methods described
above. Once the information about the device is available, device id spoofing will
be done by changing the values in identifying parameters. It is represented in
Table 3.
**5.11.2. Extracting User Information**
Google:
One main advantage of using android is, it allows users to manage their application preferences. This facility extracts user information from Google server.
Using this opportunity, anyone can access user profile.
Flurry:
Unlike Google, Flurry will not allow its users to access their information. Profile extraction of Flurry is represented in Figure 16. Here, spoofing will be done
by identifying target device id. After identifying the id, it will make the Flurry to
generate report message (appIDx). All user information can be accessed using
this technique. Another method of extracting user profile is by extracting audience report. It can be done by capturing report (Pt) at time t. Flurry also provides an additional feature to distinguish user according to age, name, group, etc.
### 5.12. Deceiving User Profiles
Second attack targets analytic results. It will attack analytic service and provides
inappropriate ads to the user. It will be done by identifying target device and de
stroy the user information by supplying irrelevant usage reports. This attack will
reduce the benefits of ad companies.
-----
A. Paranjothi et al.
Figure 16. Privacy leakage attack scenario.
Table 3. Android path identifier.
Parameter File path in Android file system
Android ID /data/data/com.android.providers.settings/databases/settings.db
ro.build.id
ro.build.version.release
ro.product.brand
/system/build.prop
ro.product.name
ro.product.device
ro.product.model
**Attacking Technique**
This attack will be done in two ways: 1) validating user information, 2) imple
menting ad influence attack. Both way uses following steps.
Training
In this method, attackers create new user profiles and train them according to
different categories. By doing this, ad developers (i.e., Google and Flurry) will be
updating user profile according to the reports they received from various catego
ries. The response time will be different for both ad developers. Google takes 6
hours to update user profile and Flurry takes one week to update its user profile.
Collecting Ads
HTTP protocol is used to deliver ads to the users who are using Google or
Flurry. Attackers will run tcpdump to extract ads from TCP but it is possible
only in Google. In Flurry, redirection methods will be done to get ads.
### 5.13. User Authentication in MCC
In MCC user authentication is used to validate the user identity. Authentication
is used to protect the user against privacy and security issues [48]. It is used to
prevent unauthorized access of the user. We have to concentrate security on
three major components in MCC. The three components are cloud, wireless
27
-----
A. Paranjothi et al.
28
communication and mobile device. The efficient algorithm will have the follow
ing qualities least possible computing, memory and storage over heads. The
purpose of authentication algorithm is to reduce the security threats in mobile
devices. Some of the threats commonly occurred in mobile devices are denial of
service, loss of device, malfunction of device etc. [48]. The authentication in
MCC varies in following scenarios when compared with cloud computing: 1)
resource limitations, 2) sensors, 3) high mobility, 4) network heterogeneity.
## 6. Conclusion and Future Work
In this paper, we have reviewed and explained in detail about offloading, mobile
distribution and privacy in MCC. Also, to implement next generation wireless
networking with low cost, we explained Wireless Mesh Networking with MCC
(WM-MCC). After reviewing various aspects, we found that MCC can be used
to provide efficient data storage and processing but the factors affecting MCC
are computation power, bandwidth, security and energy. We also found that use
of encryption method in offloading and remote execution leads to performance
degradation.
New research and expansions programs are required to make offloading deci
sions more feasible and improve security in the mobile cloud. Furthermore, us
ers want to migrate their data from smartphone to cloud but this migration pos
es some technical issues. Hence, we need a concrete effort from academia and
industry to improve these shortcomings.
## References
[1] Zhao, W., Sun, Y. and Dai, L. (2010) Improving Computer Basis Teaching through
Mobile Communication and Cloud Computing Technology. Proceedings of the 3rd
International Conference on Advanced Computer Theory and Engineering
(ICACTE’10).
[2] Heavy Reading Real World Research (2013) The Mobile Cloud Market Outlook to
2017
[3] ABI (2009) Mobile Cloud Computing Subscribers to Total Nearly One Billion by
2014, Tech. Rep., ABI Research.
[4] Mobile Cloud Computing: A Survey, State of Art and Future Directions M. Reza
Rahimi · JianRen · Chi Harold Liu · Athanasios V. Vasilakos · NaliniVenkatasubramanian.
[5] Fernando, N., Loke, S.W. and Rahayu, W. (2013) Mobile Cloud Computing: A Survey. Future Generation Computer Systems, 29, 84-106.
[6] Fernando, N., Loke, S.W. and Rahayu, W. (2013) Mobile Cloud Computing: A Survey. Future Generation Computer Systems, 29, 84-106.
[7] Wilbaut, C., Hanafi, S. and Salhi, S. (2008) A Survey of Effective Heuristics and
Their Application to a Variety of Knapsack Problems. IMA Journal of Management
[Mathematics, 19, 227-244. https://doi.org/10.1093/imaman/dpn004](https://doi.org/10.1093/imaman/dpn004)
[8] Kchaou, H., Kechaou, Z. and Alimi, A.M. (2015) Towards an Offloading Framework Based on Big Data Analytics in Mobile Cloud Computing Environments.
Procedia Computer Science, 53, 292-297.
[https://doi.org/10.1016/j.procs.2015.07.306](https://doi.org/10.1016/j.procs.2015.07.306)
-----
A. Paranjothi et al.
[9] Bahl, P., et al. (2012) Advancing the State of Mobile Cloud Computing. Proceedings
of the 3rd ACM Workshop on Mobile Cloud Computing and Services, Low Wood
[Bay, 25 June 2012, 21-28. https://doi.org/10.1145/2307849.2307856](https://doi.org/10.1145/2307849.2307856)
[10] Verbelen, T., et al. (2012) Cloudlets: Bringing the Cloud to the Mobile User. Proceedings of the 3rd ACM Workshop on Mobile Cloud Computing and Services,
[ACM. https://doi.org/10.1145/2307849.2307858](https://doi.org/10.1145/2307849.2307858)
[11] Chen, M., Jin, H., Wen, Y. and Leung, V.C.M. (2013) Enabling Technologies for
Future Data Center Networking: A Primer. IEEE Network, 27, 8-15.
[https://doi.org/10.1109/MNET.2013.6574659](https://doi.org/10.1109/MNET.2013.6574659)
[12] Kumar, K. and Lu, Y.-H. (2010) Cloud Computing for Mobile Users: Can Offloading Computation Save Energy? IEEE Computer, 43, 51-56.
[https://doi.org/10.1109/MC.2010.98](https://doi.org/10.1109/MC.2010.98)
[13] Yang, K., Ou, S. and Chen, H.-H. (2008) On Effective Offloading Services for Resource-Constrained Mobile Devices Running Heavier Mobile Internet Applications.
IEEE Communications Magazine, 46, 56-63.
[https://doi.org/10.1109/MCOM.2008.4427231](https://doi.org/10.1109/MCOM.2008.4427231)
[14] Cuervo, E., Balasubramanian, A., Cho, D., Wolman, A., Saroiu, S., Chandra, R. and
Bahl, P. (2010) MAUI: Making Smartphones Last Longer with Code Offload. ACM
[MobiSys’10, 49-62. https://doi.org/10.1145/1814433.1814441](https://doi.org/10.1145/1814433.1814441)
[15] Chun, B.-G., Ihm, S., Maniatis, P., Naik, M. and Patti, A. (2011) CloneCloud: Elastic
Execution between Mobile Device and Cloud. ACM EuroSys’11, 301-314.
[https://doi.org/10.1145/1966445.1966473](https://doi.org/10.1145/1966445.1966473)
[16] Huang, B.-K., et al. (2015) A Cloud-Based Offloading Service for Computation-Intensive Mobile Applications. IEEE 21st International Conference on Embedded and
Real-Time Computing Systems and Applications (RTCSA), 19-21 August 2015.
[https://doi.org/10.1109/rtcsa.2015.17](https://doi.org/10.1109/rtcsa.2015.17)
[17] Ra, M.R., Sheth, A., Mummert, L., Pillai, P., Wetherall, D. and Govindan, R. (2011)
Odessa: Enabling Interactive Perception Applications on Mobile Devices. Proceedings of the 9th International Conference on Mobile Systems, Applications, and Services, Bethesda, 28 June-1 July 2011, 43-56.
[https://doi.org/10.1145/1999995.2000000](https://doi.org/10.1145/1999995.2000000)
[18] Bradski, G. and Kaehler, A. (2008) Learning OpenCV: Computer Vision with the
OpenCV Library. O’Reilly Media.
[19] Romea, A.C., Berenson, D., Srinivasa, S. and Ferguson, D. (2009) Object Recognition and Full Pose Registration from Single Image for Robotic Manipulation. IEEE
International Conference on Robotics and Automation.
[20] Lowe, D. (2004) Distinctive Image Features from Scale-Invariant Keypoints. International Journal on Computer Vision (IJCV), 60, 91-110.
[https://doi.org/10.1023/B:VISI.0000029664.99615.94](https://doi.org/10.1023/B:VISI.0000029664.99615.94)
[21] Pillai, P.S., Mummert, L.B., Schlosser, S.W., Sukthankar, R. and Helfrich, C.J. (2009)
"SLIPstream: Scalable Low-Latency Interactive Perception on Streaming Data. ACM
International Workshop on Network and Operating System Support for Digital
Audio and Video.
[22] Kosta, S., Aucinas, A., Hui, P., Mortier, R. and Zhang, X. (2012) Thinkair: Dynamic
Resource Allocation and Parallel Execution in the Cloud for Mobile Code Offloading. In INFOCOM, 2012 Proceedings IEEE, 945-953.
[23] Kosta, S., Perta, V.C., Stefa, J., Hui, P. and Mei, A. (2013) Clone2clone (c2c):
Peer-to-Peer Networking of Smartphones on the Cloud. In 5th USENIX Workshop
on Hot Topics in Cloud Computing (HotCloud13).
29
-----
A. Paranjothi et al.
30
[24] Feldman, A.J., Zeller, W.P., Freedman, M.J. and Felten, E.W. Sporc: Group Collaboration Using Untrusted Cloud Resources. In Proc. OSDI’10.
[25] Ellis, C.A. and Gibbs, S.J. (1989) Concurrency Control in Groupware Systems.
[SIGMOD REC, 18, 399-407. https://doi.org/10.1145/66926.66963](https://doi.org/10.1145/66926.66963)
[26] Ravindranath, L., Thiagarajan, A., Balakrishnan, H. and Madden, S. (2012) Code in
the Air: Simplifying Sensing and Coordination Tasks on Smartphones. Proceedings
of the Twelfth Workshop on Mobile Computing Systems & Applications, p. 4,
ACM.
[27] Stuedi, P., Mohomed, I. and Terry, D. (2010) WhereStore: Location-Based Data
Storage for Mobile Devices Interacting with the Cloud. Proceedings of the 1st ACM
Workshop on Mobile Cloud Computing & Services: Social Networks and Beyond,
p. 1, ACM.
[28] Trestian, I., Ranjan, S., Kuzmanovic, A. and Nucci, A. (2009) Measuring Serendipity: Connecting People, Locations and Interests in a Mobile 3g Network. Proceedings
of the 9th ACM SIGCOMM Conference on Internet Measurement Conference,
[New York, 267-279. https://doi.org/10.1145/1644893.1644926](https://doi.org/10.1145/1644893.1644926)
[29] Ananthanarayanan, G., Haridasan, M., Mohomed, I., Terry, D. and Thekkath, C.A.
(2009) Startrack: A Framework for Enabling Track-Based Applications. Proceedings
of the 7th International Conference on Mobile Systems, Applications, and Services,
[New York, 207-220. https://doi.org/10.1145/1555816.1555838](https://doi.org/10.1145/1555816.1555838)
[30] Mun, M., Reddy, S., Shilton, K., Yau, N., Burke, J., Estrin, D., Hansen, M., Howard,
E., West, R. and Peir, P.B. (2009) The Personal Environmental Impact Report, as a
Platform for Participatory Sensing Systems Research. MobiSys’09: Proceedings of
the 7th International Conference on Mobile Systems, Applications, and Services,
[New York, 55-68. https://doi.org/10.1145/1555816.1555823](https://doi.org/10.1145/1555816.1555823)
[31] Krumm, J. and Horvitz, E. (2007) Predestination: Where Do You Want to Go Today? Computer, 40, 105-107.
[32] Ramasubramanian, V., Rodeheer, T.L., Terry, D.B., Walraed-Sullivan, M., Wobber,
T., Marshall, C.C. and Vahdat, A. (2009) Cimbiosys: A Platform for Content-Based
Partial Replication. NSDI’09: Proceedings of the 6th USENIX Symposium on Networked Systems Design and Implementation, Berkeley, 261-276.
[33] Ananthanarayanan, G., Haridasan, M., Mohomed, I., Terry, D. and Thekkath, C.A.
Startrack: A Framework for Enabling Track-Based Applications. MobiSys’09: Proceedings of the 7th International Conference on Mobile Systems, Applications, and
[Services, New York, 207-220. https://doi.org/10.1145/1555816.1555838](https://doi.org/10.1145/1555816.1555838)
[34] Bickford, J. and Cáceres, R. (2013) Towards Synchronization of Live Virtual Machines among Mobile Devices. Proceedings of the 14th Workshop on Mobile Computing Systems and Applications, p. 13, ACM.
[https://doi.org/10.1145/2444776.2444794](https://doi.org/10.1145/2444776.2444794)
[35] Lin, H., et al. (2015) A Trustworthy Access Control Model for Mobile Cloud Computing Based on Reputation and Mechanism Design. Ad Hoc Networks, 35, 51-64.
[https://doi.org/10.1016/j.adhoc.2015.07.007](https://doi.org/10.1016/j.adhoc.2015.07.007)
[36] Dinh, T.T.A. and Datta, A. (2013) Towards Secure Outsourcing of Collaborative
Sensing and Analytic Applications to the Cloud-the pCloud Approach. Proceedings
of the First International Workshop on Middleware for Cloud-Enabled Sensing, p.
[2, ACM. https://doi.org/10.1145/2541603.2541606](https://doi.org/10.1145/2541603.2541606)
[37] Goyal, V., Pandey, O., Sahai, A. and Waters, B. (2006) Attribute-Based Encryption
for Fine-Grained Access Control of Encrypted Data. In CCS’06.
[38] Green, M., Hohenberger, S. and Waters, B. (2011) Outsourcing the Decryption of
Abeciphertexts. In USENIX Security.
-----
A. Paranjothi et al.
[39] Stephen, J.J., Savvides, S., Seidel, R. and Eugster, P. (2014) Practical Confidentiality
Preserving Big Data Analysis. Proceedings of the 6th USENIX Conference on Hot
Topics in Cloud Computing, USENIX Association, 10.
[40] Olston, C., Reed, B., Srivastava, U., Kumar, R. and Tomkins, A. (2008) Pig Latin: A
Not-So-Foreign Language for Data Processing. In SIGMOD.
[41] Schneier, B. (1994) Description of a New Variable-Length Key, 64- Bit Block Cipher
(Blowfish). In Fast Software Encryption. Springer-Verlag, 191-204.
[42] Daemen, J. and Rijmen, V. (2002) The Design of Rijndael: AES—The Advanced
Encryption Standard. Springer-Verlag, Berlin, Heidelberg, New York.
[43] Popa, R.A., Redfield, C.M.S., Zeldovich, N. and Balakrishnan, H. (2011) CryptDB:
Protecting Confidentiality with Encrypted Query Processing. In SOSP.
[44] Elgamal, T. (1985) A Public-Key Cryptosystem and a Signature Scheme Based on
Discrete Logarithms. IEEE Transactions on Information Theory, 31, 4.
[45] Ouaknine, K., Carey, M. and Kirkpatrick, S. (2015) The PigMix Benchmark on Pig,
MapReduce, and HPCC Systems. IEEE International Congress on Big Data (Big
Data Congress), 27 June-2 July 2015.
[https://doi.org/10.1109/bigdatacongress.2015.99](https://doi.org/10.1109/bigdatacongress.2015.99)
[46] Marathe, A., et al. (2016) Exploiting Redundancy and Application Scalability for
Cost-Effective, Time-Constrained Execution of HPC Applications on Amazon EC2.
IEEE Transactions on Parallel and Distributed Systems, 27, 2574-2588.
[https://doi.org/10.1109/TPDS.2015.2508457](https://doi.org/10.1109/TPDS.2015.2508457)
[47] Chen, T., Ullah, I., Kaafar, M.A. and Boreli, R. (2014) Information Leakage through
Mobile Analytics Services. Proceedings of the 15th Workshop on Mobile Computing Systems and Applications, p. 15, ACM.
[https://doi.org/10.1145/2565585.2565593](https://doi.org/10.1145/2565585.2565593)
[48] Alizadeh, M., et al. (2015) Authentication in Mobile Cloud Computing: A Survey.
Journal of Network and Computer Applications.
Submit or recommend next manuscript to SCIRP and we will provide best
service for you:
Accepting pre-submission inquiries through Email, Facebook, LinkedIn, Twitter, etc.
A wide selection of journals (inclusive of 9 subjects, more than 200 journals)
Providing 24-hour high-quality service
User-friendly online submission system
Fair and swift peer-review system
Efficient typesetting and proofreading procedure
Display of the result of downloads and visits, as well as the number of cited articles
Maximum dissemination of your research work
[Submit your manuscript at: http://papersubmission.scirp.org/](http://papersubmission.scirp.org/)
[Or contact jcc@scirp.org](mailto:jcc@scirp.org)
31
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.4236/JCC.2017.56001?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.4236/JCC.2017.56001, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "GOLD",
"url": "http://www.scirp.org/journal/PaperDownload.aspx?paperID=75210"
}
| 2,017
|
[
"Review"
] | true
| 2017-04-06T00:00:00
|
[
{
"paperId": "9a674d82cd4f47477e3f1aa7b89a3d9f6e3894b8",
"title": "Learning OpenCV 3: Computer Vision in C++ with the OpenCV Library"
},
{
"paperId": "8a19fc3871bbb1841864950f25c03771b298666a",
"title": "Exploiting Redundancy and Application Scalability for Cost-Effective, Time-Constrained Execution of HPC Applications on Amazon EC2"
},
{
"paperId": "d4e8e9d36d594b39b37b42c66841ef98f0e3925d",
"title": "Authentication in mobile cloud computing: A survey"
},
{
"paperId": "6a3d2d1595aa0e5e90cfed08cd2c7ee9cb346743",
"title": "A trustworthy access control model for mobile cloud computing based on reputation and mechanism design"
},
{
"paperId": "cc504af1af05c7ac1954962a037ddcc7f70bbd13",
"title": "A Cloud-Based Offloading Service for Computation-Intensive Mobile Applications"
},
{
"paperId": "35316dbc451751b8bd61a9ebbda6889b90b02140",
"title": "The PigMix Benchmark on Pig, MapReduce, and HPCC Systems"
},
{
"paperId": "7c8a4abc7624802783bde0688969fcdf373d01e7",
"title": "Practical Confidentiality Preserving Big Data Analysis"
},
{
"paperId": "83eb1b3d8b3fe674c12ed0b09fa4a803ca16046b",
"title": "Information leakage through mobile analytics services"
},
{
"paperId": "449e836eb90eff2d055e916a410ea97f19855e6d",
"title": "Towards secure outsourcing of collaborative sensing and analytic applications to the cloud - the pCloud approach"
},
{
"paperId": "6cf7236dc94e53f31a5f265b13257382c23d9253",
"title": "Mobile Cloud Computing: A Survey, State of Art and Future Directions"
},
{
"paperId": "a3b871c78aacb36c0336dc20274931b0d65701a0",
"title": "Enabling technologies for future data center networking: a primer"
},
{
"paperId": "bea33fc20db18434b12269048c4af173075b3252",
"title": "Towards synchronization of live virtual machines among mobile devices"
},
{
"paperId": "22d41328530154267c54d436d2d59d33fd9bacdc",
"title": "Cloudlets: bringing the cloud to the mobile user"
},
{
"paperId": "f2befa373403def2a367edbe60afeaa03d1c7e61",
"title": "Advancing the state of mobile cloud computing"
},
{
"paperId": "faff5dac93f324923ccb2c5ebeb1844de1643bd9",
"title": "ThinkAir: Dynamic resource allocation and parallel execution in the cloud for mobile code offloading"
},
{
"paperId": "72d0ee37ff57ae7977d451b7e2e0bcbe522bbf44",
"title": "Code in the air: simplifying sensing and coordination tasks on smartphones"
},
{
"paperId": "acc104820d211a25b568f8677b4ad29a8667c7cd",
"title": "CryptDB: protecting confidentiality with encrypted query processing"
},
{
"paperId": "ccde4f28eac0501c7fa075d06ab3d0f01fbd09af",
"title": "Outsourcing the Decryption of ABE Ciphertexts"
},
{
"paperId": "eb27a6067d7a0416d060434e307e8cf34c11a39f",
"title": "Odessa: enabling interactive perception applications on mobile devices"
},
{
"paperId": "731304583d2e40bf6ee030e3cd81f767713f999a",
"title": "CloneCloud: elastic execution between mobile device and cloud"
},
{
"paperId": "4be2c5d63e570300f30d96d12949f826c5c9c136",
"title": "SPORC: Group Collaboration using Untrusted Cloud Resources"
},
{
"paperId": "5fe125dcbff9c3023fe66eab137f5fabadd61a87",
"title": "Improving computer basis teaching through mobile communication and cloud computing technology"
},
{
"paperId": "0d366f3522bcf503b9f0fea8a5d009ba3ecddf39",
"title": "WhereStore: location-based data storage for mobile devices interacting with the cloud"
},
{
"paperId": "0f7dd6bcaf4c2a64b3a51b7351929dab23584ea8",
"title": "MAUI: making smartphones last longer with code offload"
},
{
"paperId": "93a96e0f68038fbbe70f80952632e8f0770af56e",
"title": "Cloud Computing for Mobile Users: Can Offloading Computation Save Energy?"
},
{
"paperId": "018bff5b5b6d38ed6cd96f43adbb04a95f34695c",
"title": "Measuring serendipity: connecting people, locations and interests in a mobile 3G network"
},
{
"paperId": "1802d0f19c7e324a784c5f334463607dddf7efb2",
"title": "StarTrack: a framework for enabling track-based applications"
},
{
"paperId": "8e383be7e8a37c93435952797731f07940eac6dc",
"title": "PEIR, the personal environmental impact report, as a platform for participatory sensing systems research"
},
{
"paperId": "3741c9aea5e0eba7658826052a30d0e82de67078",
"title": "SLIPstream: scalable low-latency interactive perception on streaming data"
},
{
"paperId": "8ee174f53bf7fa3e2e43eaf60553601f7e9f8022",
"title": "Object recognition and full pose registration from a single image for robotic manipulation"
},
{
"paperId": "ee07f7bea5c3c6aa44a7257dee8b63b574ce12c9",
"title": "A Platform for Content-based Partial Replication"
},
{
"paperId": "3ac939c4ada416858e1c1bff9e6f18aaa98323f3",
"title": "Pig latin: a not-so-foreign language for data processing"
},
{
"paperId": "d25ca9e72b3ede77cf08529f2bfe44ab7b07933f",
"title": "Predestination: Where Do You Want to Go Today?"
},
{
"paperId": "30b4b7503a7004646ba4a75b6f0aaf4f626f0854",
"title": "A survey of effective heuristics and their application to a variety of knapsack problems"
},
{
"paperId": "422876e542daadefe3371091a65c5671185796e2",
"title": "Attribute-based encryption for fine-grained access control of encrypted data"
},
{
"paperId": "8c04f169203f9e55056a6f7f956695babe622a38",
"title": "Distinctive Image Features from Scale-Invariant Keypoints"
},
{
"paperId": "92a6103f2c3b84e4ad87857b86b8566b59623442",
"title": "The Design of Rijndael: AES - The Advanced Encryption Standard"
},
{
"paperId": "d1776dfb8f66cb40cadcac9bb66760ec9b7b3920",
"title": "Description of a New Variable-Length Key, 64-bit Block Cipher (Blowfish)"
},
{
"paperId": "517b1a3da76f773df920ce59ce8bf6306de2c951",
"title": "Towards an Offloading Framework based on Big Data Analytics in Mobile Cloud Computing Environments"
},
{
"paperId": "ec13c3e7119191802e6f5783d297fe7a5a05293e",
"title": "Mobile cloud computing: A survey"
},
{
"paperId": "c0a568ca43a0a24c47d77934a3e228c168c825fb",
"title": "Clone2Clone (C2C): Peer-to-Peer Networking of Smartphones on the Cloud"
},
{
"paperId": "5099369c8a4096c5c997d4ae6354515ed1d7ab30",
"title": "On effective offloading services for resource-constrained mobile devices running heavier mobile Internet applications"
},
{
"paperId": "a1cd437a924849d19e0713f042e45e79dc8b95a1",
"title": "A public key cyryptosystem and signature scheme based on discrete logarithms"
}
] | 18,297
|
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Engineering",
"source": "external"
},
{
"category": "Mathematics",
"source": "external"
},
{
"category": "Engineering",
"source": "s2-fos-model"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/02af7c7509c70c91d28cd2273002aff5e479c6ee
|
[
"Computer Science",
"Engineering",
"Mathematics"
] | 0.832977
|
MAESTRO-X: Distributed Orchestration of Rotary-Wing UAV-Relay Swarms
|
02af7c7509c70c91d28cd2273002aff5e479c6ee
|
IEEE Transactions on Cognitive Communications and Networking
|
[
{
"authorId": "9078700",
"name": "Bharath Keshavamurthy"
},
{
"authorId": "153659184",
"name": "M. Bliss"
},
{
"authorId": "2709642",
"name": "Nicolò Michelusi"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"IEEE Trans Cogn Commun Netw"
],
"alternate_urls": [
"https://ieeexplore.ieee.org/xpl/RecentIssue.jsp?punumber=6687307"
],
"id": "65e58b80-9699-4da6-bd60-929b57b8533d",
"issn": "2332-7731",
"name": "IEEE Transactions on Cognitive Communications and Networking",
"type": "journal",
"url": "http://ieeexplore.ieee.org/servlet/opac?punumber=6687307"
}
|
This work details a scalable framework to orchestrate a swarm of rotary-wing UAVs serving as cellular relays to facilitate beyond line-of-sight connectivity and traffic offloading for ground users. First, a Multiscale Adaptive Energy-conscious Scheduling and TRajectory Optimization (MAESTRO) framework is developed for a single UAV. Aiming to minimize the time-averaged latency to serve user requests, subject to an average UAV power constraint, it is shown that the optimization problem can be cast as a semi-Markov decision process, and exhibits a multiscale structure: outer actions on radial wait velocities and terminal service positions minimize the long-term delay-power trade-off, optimized via value iteration; given these outer actions, inner actions on angular wait velocities and service trajectories minimize a short-term delay-energy cost; finally, rate adaptation is embedded along the trajectory to leverage air-to-ground channel propagation conditions. A novel hierarchical competitive swarm optimization scheme is developed in the inner optimization, to devise high-resolution trajectories via iterative pair-wise updates. Next, MAESTRO is eXtended to UAV swarms (MAESTRO-X) via scalable policy replication, enabled by a decentralized command-and-control network augmented with: (1) spread maximization to proactively position UAVs to serve future requests; (2) consensus-driven conflict resolution to orchestrate scheduling decisions based on delay-energy costs including queuing dynamics; (3) adaptive frequency reuse to improve spectrum utilization across the network; and (4) a piggybacking mechanism allowing UAVs to serve multiple ground users simultaneously. Numerical evaluations show that, for user requests of 10 Mbits, generated according to a Poisson arrival process with rate 0.2 req/min/UAV, single-agent MAESTRO offers $3.8\times $ faster service than a high-altitude platform and 29% faster than a static UAV deployment; moreover, for a swarm of 3 UAV-relays, MAESTRO-X delivers data payloads $4.7\times $ faster than a successive convex approximation scheme; and remarkably, a single UAV optimized via MAESTRO outclasses 3 UAVs optimized via a deep-Q network by 38%.
|
# MAESTRO-X: Distributed Orchestration of
Rotary-Wing UAV-Relay Swarms
### Bharath Keshavamurthy[∗], Matthew A. Bliss[†], and Nicolò Michelusi[∗]
**Abstract**
This work details a scalable framework to orchestrate a swarm of rotary-wing UAVs serving as
cellular relays to facilitate beyond line-of-sight connectivity and traffic offloading for ground users.
First, a Multiscale Adaptive Energy-conscious Scheduling and TRajectory Optimization (MAESTRO)
framework is developed for a single UAV. Aiming to minimize the time-averaged latency to serve
user requests, subject to an average UAV power constraint, it is shown that the optimization problem
can be cast as a semi-Markov decision process, and exhibits a multiscale structure: outer actions on
radial wait velocities and terminal service positions minimize the long-term delay-power trade-off,
optimized via value iteration; given these outer actions, inner actions on angular wait velocities and
service trajectories minimize a short-term delay-energy cost; finally, rate adaptation is embedded along
the trajectory to leverage air-to-ground channel propagation conditions. A novel hierarchical competitive
swarm optimization scheme is developed in the inner optimization, to devise high-resolution trajectories
via iterative pair-wise updates. Next, MAESTRO is eXtended to UAV swarms (MAESTRO-X) via
scalable policy replication, enabled by a decentralized command-and-control network augmented with:
(1) spread maximization to proactively position UAVs to serve future requests; (2) consensus-driven
_conflict resolution to orchestrate scheduling decisions based on delay-energy costs including queuing_
dynamics; (3) adaptive frequency reuse to improve spectrum utilization across the network; and (4)
a piggybacking mechanism allowing UAVs to serve multiple ground users simultaneously. Numerical
evaluations show that, for user requests of 10 Mbits, generated according to a Poisson arrival process with
rate 0.2 req/min/UAV, single-agent MAESTRO offers 3.8 faster service than a high-altitude platform
_×_
and 29% faster than a static UAV deployment; moreover, for a swarm of 3 UAV-relays, MAESTRO-X
delivers data payloads 4.7 faster than a successive convex approximation scheme; and remarkably, a
_×_
single UAV optimized via MAESTRO outclasses 3 UAVs optimized via a deep-Q network by 38%.
**Index Terms**
UAV-Relays, Trajectory optimization, SMDPs, Hierarchical CSO
[A preliminary version of this work was presented at Asilomar 2022 [1]. Source code is available on GitHub [2].](https://github.com/bharathkeshavamurthy/MAESTRO-X.git)
Part of this work has been supported by NSF under grants CNS-1642982 and CNS-2129015.
_∗Electrical, Computer and Energy Engineering, Arizona State University, Tempe, AZ._
_†Electrical and Computer Engineering, Purdue University, West Lafayette, IN._
-----
I. INTRODUCTION
Enterprises across various industrial sectors have stepped-up the adoption of Unmanned Aerial
Vehicles (UAVs) to gather data, survey infrastructure, monitor operations, and automate logistics
[3], [4]. UAVs can also be leveraged to enhance troop deployments in military scenarios [5],
aid emergency response during a natural disaster [6], and facilitate data harvesting in precision
agriculture [7]. Inevitably, this has fostered varied academic research and industrial R&D on
UAV-augmented beyond line-of-sight connectivity and traffic offloading in cellular networks,
whose coverage can be enhanced by the mobility and maneuverability of UAVs [8], [9].
Yet, the pervasive potential of UAV-assisted wireless networks presents a plethora of challenges
in real-world deployments [9]: limited on-board energy of aerial platforms, Quality-of-Service
(QoS) requirements, air-to-ground channels, and computational feasibility challenges of UAV
trajectory design. Several works have tackled some of these challenges by employing tools from
optimization and artificial intelligence—however, numerous problems remain unsolved: failure
to capture uncertain system dynamics vis-à-vis random traffic arrivals [10]–[14]; restrictions
on UAV path and velocity characteristics [11], [15]; inefficient centralized swarm deployments
[16]–[18]; computationally expensive joint multi-agent formulations offering limited scalability
[19]–[22]; and failure to account for link layer effects on the QoS of the network [23], [24].
In this paper, considering these drawbacks in the state-of-the-art, we study the decentralized
orchestration of multiple power-constrained rotary-wing UAVs supplementing a terrestrial base
station by relaying data traffic dynamically generated by ground users. Incorporating waiting state
optimization, computationally feasible trajectory design, throughput-maximizing rate adaptation
to Air-to-Ground (A2G) propagation conditions, queue management, frequency reuse to enhance
spectrum utilization, multi-user service, and multi-UAV consensus-driven scheduling, we develop
a scalable framework to efficiently automate the operations of distributed UAV-relay deployments.
Ergo, specializing to single UAV-relay settings, we first propose MAESTRO, a Multiscale
Adaptive Energy-conscious Scheduling and TRajectory Optimization framework to control the
idle and service phase operations of the UAV. Seeking to minimize the average communication
delay subject to an average UAV mobility power constraint, we show that the problem can be
cast as a Semi-Markov Decision Process (SMDP) with a multiscale structure: outer decisions
on radial velocities and terminal service positions influence the long-term delay-power cost;
consequently, given these outer actions, inner actions on angular wait velocities and service
-----
|Paper|Adaptive control|Channel model|Frequency reuse|Multiuser service|UAV Motion Mobility Velocity|Col7|UAV deployment|Multi-UAV scheduling|Overall formulation|Link Layer Schedule Queue|Col12|
|---|---|---|---|---|---|---|---|---|---|---|---|
|MAESTRO-X|Yes|A2G|Yes|Yes|Dynamic|Variable|Distributed|Decoupled|Model-based|Yes|Yes|
|[10]|No|FSPL|No|No|Dynamic|Variable|Single|-|Model-based|Yes|No|
|[16]|No|A2G|Yes|Yes|Dynamic|Variable|Centralized|Joint|Model-based|Yes|No|
|[19]|No|A2G|No|Yes|Restricted|Fixed|Distributed|Joint|Model-free|No|No|
|[11]|No|FSPL|No|No|Dynamic|Fixed|Single|-|Model-based|Yes|No|
|[12]|No|FSPL|No|No|Dynamic|Variable|Single|-|Model-based|Yes|No|
|[20]|No|FSPL|No|Yes|Restricted|Fixed|Distributed|Joint|Model-free|Yes|No|
|[13]|No|A2G|No|No|Static|-|Single|-|Model-based|No|No|
|[23]|No|FSPL|No|No|Static|-|Distributed|Joint|Model-based|Yes|No|
|[24]|Yes|FSPL|No|No|Static|-|Distributed|Joint|Model-based|No|No|
|[17]|No|FSPL|No|No|Dynamic|Fixed|Centralized|Joint|Model-based|Yes|No|
|[18]|No|A2G|No|No|Static|-|Centralized|Joint|Model-based|No|No|
|[27]|No|A2G|No|No|Restricted|Fixed|Distributed|Decoupled|Model-free|No|No|
|[21]|Yes|FSPL|No|No|Static|-|Distributed|Joint|Model-free|No|Yes|
|[22]|Yes|A2G|No|No|Static|-|Distributed|Joint|Model-free|No|No|
|[14]|No|A2G|No|No|Dynamic|Variable|Single|-|Model-based|Yes|No|
|[28]|Yes|FSPL|No|No|Dynamic|Variable|Single|-|Model-free|No|No|
TABLE I: A comparison of the features of our framework with those of relevant schemes in the literature.
trajectories minimize a short-term delay-energy cost. We develop a value iteration algorithm
[25] exploiting this multiscale structure to optimize outer actions, and a hierarchical variant
of Competitive Swarm Optimization (CSO) [26], decoupled from value iteration, to optimize
high-resolution trajectories embedding a novel throughput maximizing rate adaptation scheme
for A2G channels. Next, we extend MAESTRO to a swarm of UAV-relays (MAESTRO-X) via
a scalable replication strategy, enabled by a decentralized command-and-control network and
augmented with: spread maximization to proactively position the UAVs to serve future service
requests; consensus-driven conflict resolution to orchestrate ground user scheduling decisions
based on delay-energy costs, including queuing dynamics; frequency reuse to enhance spectrum
utilization; and piggybacking to enable each UAV to serve multiple users simultaneously.
**Related Work: Table I summarizes our approach (MAESTRO-X) and contrasts it with relevant**
works in the state-of-the-art. First, we observe non-adaptive schemes, e.g., [10], [17], [18]
designed for applications where ground users possess local storage or aggregation capabilities
allowing for deterministic traffic; however, practical deployments involve dynamically generated
requests and randomly located ground users. Accommodating these uncertainties calls for the
design of adaptive UAV orchestration frameworks. Yet, existing works do so only for single
UAV-relay deployments [28] or consider static placement of UAVs (i.e., no trajectory design)
[21], [22], [24]. In contrast, we design adaptive trajectory and scheduling strategies for distributed
multi-UAV swarms, that accommodate dynamic and uncertain traffic generated by ground users.
Next, works employing Free Space Pathloss (FSPL) channel models, e.g., [10]–[12], [20],
fail to account for the A2G channel characteristics in UAV-assisted wireless networks. Existing
works that model A2G channels fail to leverage small- and large-scale A2G conditions via
-----
rate adaptation. A notable exception is [14], which differs from our rate adaptation scheme
in two ways: 1) we select the rate to maximize throughput (vs. [14], which aims to satisfy an
outage constraint), and 2) we use a probabilistic line-of-sight (LoS) and Non-LoS (NLoS) model.
Furthermore, most works surveyed neither consider spectrum reuse (with the exception of [16])
nor permit simultaneous multi-user service (with the exception of [16], [19], [20])—however,
the works that do incorporate these crucial features [16], [19], [20] fail to consider adaptation
to dynamically generated requests from randomly located users, as done in our work.
A common approach for trajectory design is Successive Convex Approximation (SCA) [10],
[14]. SCA typically relies on the FSPL channel model to devise convex relaxations of the
objective and constraints. Exceptions include [14] and [16], which apply SCA approaches under
A2G channels. In [14], a logistic approximation of the achievable rate is used under outage
constraints; in [16], only large-scale fading is considered. However, when coupling trajectory
design with our throughput-maximizing rate adaptation scheme, closed-form rate expressions
with first-order convex approximations are impractical. To tackle this challenge, we propose
a CSO [26] approach for UAV trajectory design. Unlike SCA, CSO does not rely on the
problem structure of FSPL models to work effectively, and can thus accommodate realistic A2G
propagation conditions. Particle Swarm Optimization (PSO) [29], a swarm-based optimization
method in which particle updates are driven by the global and individual best positions, has been
used to optimize static UAV placement [30], [31], or restricted UAV trajectories (e.g., moving
along a circle [15], or with fixed speed [11]). Removing these restrictions calls for the more
efficient update strategy of CSO, which exhibits superior performance on several benchmarks
[26]: it involves pair-wise particle competitions, wherein winners advance to the next iteration
and the losers learn from the winners. Moreover, we scale CSO to higher-dimensional trajectory
design by embedding it within a Hierarchical wrapper (HCSO), which iteratively optimizes
trajectories of increasing resolution, without imposing unreasonable restrictions on UAV mobility.
Next, shifting our attention to swarm orchestration frameworks, several approaches consider
centralized multi-UAV deployments [16]–[18] in which an aggregation center coordinates the
UAV-relaying operations; or either joint multi-relay solutions [16], [23], [24] or model-free
formulations constituting combined state and action spaces [19]–[22]. An exception is [27], which
considers a model-free setup with decentralized UAV deployments and decoupled scheduling.
But, [27] does not consider adaptation to randomly-generated data traffic, as we do in our
work; rather, a sense-and-send protocol is devised, wherein tasks are always ready to be sensed.
-----
Centralized swarm deployments often need additional capital and operational expenditure, and
joint multi-UAV designs lead to large solution spaces resulting in prohibitive convergence times.
Mindful of such considerations, we present an orchestration framework suitable for distributed
UAV deployments by replicating our single-agent policy across the swarm and augmenting it
with spread maximization and consensus-driven link-layer prescient conflict resolution over a
command-and-control network. This eliminates the need for a centralized aggregation center,
mitigates the computational overhead encountered by joint multi-relay models, and facilitates
the seamless incorporation of queuing dynamics into scheduling decisions. Also, as shown in
our numerical evaluations, our framework can be scaled to networks with 10 UAVs, while
_≥_
state-of-the-art approaches [10], [16], [19] become prohibitively expensive for networks with 5
UAVs. Additionally, although model-free control schemes [19]–[22], [27], [28] consider unknown
system dynamics when solving for the optimal trajectory and/or scheduling solution, they fail to
efficiently exploit the problem structure, resulting in large policy convergence times. In contrast,
we use a model-based approach, by casting the problem as an SMDP, which captures the temporal
irregularities seen in the state transitions of UAV-augmented wireless networks.
**Contributions: We develop a novel framework for the scalable orchestration of UAV-relay**
swarms. To the best of our knowledge, no other work simultaneously incorporates the practical
features of 1) dynamic traffic from randomly located ground users; 2) efficient exploitation of
A2G channel conditions via a throughput-maximizing rate adaptation scheme; 3) easy scalability
to large UAV swarms via policy replication, coupled with multi-agent coordination mechanisms
over a distributed command-and-control network; and 4) waiting state optimization to position
idle UAVs for potential new requests. In a nutshell, the contributions of this paper are:
_• MAESTRO: For a single UAV, we construct an adaptive scheduling and trajectory design_
framework to minimize the communication latencies in serving dynamic transmission requests
generated by randomly located ground users, subject to an average UAV power constraint.
We show that the problem can be solved as a Semi-Markov Decision Process (SMDP). A
multiscale decomposition facilitates efficient computation of rate adaptation, scheduling and
trajectory solutions, and energy-conscious orchestration of the UAV during idle periods.
_• HCSO: To enable computationally tractable design of high-resolution UAV trajectories under_
A2G propagation conditions, we propose Hierarchical CSO (HCSO), a variant of CSO wherein
iterative pair-wise cost comparisons devise trajectories of increasingly higher resolution.
_• MAESTRO-X: Coupled with decentralized command-and-control operations over a distributed_
-----
(a) Deployment Model. (b) Throughputs under A2G channel.
Fig. 1: (a) A terrestrial BS aided by UAVs serving as relays for a diverse set of GNs: traffic offloading for cellular
UEs, and coverage extensions for livestock monitors and soil sensors; (b) rate-adapted throughputs (see Table II
for the numerical parameters) along the GN BS link (direct), GN UAV link (decode), UAV BS link (forward),
_→_ _→_ _→_
and GN UAV BS link (decode-and-forward, with the UAV relay stationed above the BS or the GN).
_→_ _→_
mesh network, we augment the single-UAV trained policy with multi-UAV mechanisms to
orchestrate waiting phase operations (spread maximization), coordinate scheduling decisions
incorporating queuing dynamics (consensus-driven conflict resolution), enable simultaneous
multi-user service (piggybacking), and enhance spectrum utilization (frequency reuse).
The rest of the paper is organized as follows: Sec. II introduces the system model; Sec. III
elucidates the design of MAESTRO; Sec. IV describes the main algorithms; Sec. V details policy
replication and multi-UAV mechanisms to manage distributed swarms (MAESTRO-X); Sec. VI
chronicles our numerical evaluations; and finally, Sec. VII lists our concluding remarks.
II. SYSTEM MODEL
Consider the deployment scenario depicted in Fig. 1a: a swarm of NU rotary-wing Unmanned
Aerial Vehicles (UAVs) operate as cellular relays to supplement a terrestrial Base Station (BS)
by relaying data traffic dynamically generated by Ground Nodes (GNs). The BS is located at the
center of the circular cell (of radius a), at height HB. The UAVs operate at a fixed height HU . The
GNs are distributed uniformly at random throughout the cell, with density λG [GNs/unit area].
Multi-user communication is enabled via OFDMA over a spectrum of bandwidth W, discretized
into NC orthogonal data channels (possibly, obtained by grouping multiple subcarriers together),
each with bandwidth B≜ _N[W]C_ [. We assume the system operates in the uplink, i.e., traffic requests]
generated by the GNs are transmitted to the BS, either directly or by using one UAV as a relay.
It can be extended to both uplink/downlink via a state variable differentiating between the two.
-----
**Communication Model: Each GN generates uplink transmission requests of L bits, according to**
a Poisson process with rate λR|G [requests/GN/unit time]. Coupled with the random deployment
of GNs, uplink requests arrive in time according to a Poisson process with rate Λ≜λG·λR|Gπa[2]
[requests/unit time] over the circular cell. Since a new request is uniformly distributed in the cell
area, the position (r, θ) of the source GN—expressed in polar coordinates with respect to the
BS—has angular coordinate θ uniform in [0, 2π), and radial coordinate with probability density
function given by fR(r)= _a[2][r][2]_ [I][(][r][≤][a][)][, where][ I][(][·][)][ is the indicator function.]
A fully-connected mesh network overlaying the BS and UAVs enables command-and-control
using the band-edges of the allocated spectrum as control channels. Since control packets
constitute short frames relative to the large GN-generated data payloads (communicated over
data channels), the control operation latencies are neglected. To request uplink transmission to
the BS, a GN sends a service request with its location; the BS broadcasts this need-for-service
to the UAV swarm. Next, a consensus-driven conflict resolution process occurs among the BS
and all UAVs (Sec. V), based on assessed delay-energy costs for this request, culminating in
a scheduling decision. If direct-BS transmission is chosen, the BS chooses an available data
channel, or queues the request until one becomes available (see Sec. V). The BS then instructs
the GN to begin direct transmission over the data channel. Otherwise, if UAV relay i is selected,
the new GN request is served via a Decode-and-Forward (D&F) strategy on an available data
channel (or queued until one becomes available), as detailed in Sec. V. While executing the D&F
protocol, the UAV moves along a pre-designed energy-conscious trajectory, i.e., a sequence of
way-points and velocities (see Sec. IV). In Sec. V, we also discuss a frequency reuse mechanism
to improve spectrum utilization efficiency, and a piggybacking mechanism allowing the scheduled
UAV to serve multiple requests simultaneously. As evident from this communication model, the
GN BS, GN UAV, and UAV BS links must be characterized, as detailed next.
_→_ _→_ _→_
**A2G Channel Model: For a generic link, we denote the flat-fading channel coefficient as**
_h≜[√]βg, where β captures the large-scale channel variations, and g with E [|g|[2]] =1 denotes_
the small-scale fading component. We model the large-scale component as β=βLoS(d)≜β0d[−][α]
for LoS and β=βNLoS(d)≜κβ0d[−][α][˜] for NLoS links, where β0 is the pathloss at a reference
distance of 1 m, 2 _α_ _α˜ are the LoS and NLoS pathloss exponents, κ_ (0, 1] captures the
_≤_ _≤_ _∈_
additional NLoS attenuation, and d denotes the Tx-Rx Euclidean distance [10]. Following [32],
we use a probabilistic LoS model, with LoS probability PLoS(ϕ)=[1+z1 exp{−z2(ϕ−z1)}][−][1],
where ϕ∈(0[o], 90[o]] is the Tx-Rx elevation angle, and z1, z2 are parameters specific to the
-----
propagation environment (e.g., urban, suburban, rural) [32]. The distribution of the small-scale
fading component g also depends on the LoS or NLoS link state [33]: for LoS links, as in [14],
we model g as Rician fading with a ϕ-dependent K-factor K(ϕ)=k1 exp{k2ϕ}, where k1, k2
are specific to the propagation environment; for NLoS links, we model g as Rayleigh fading
� �
(Rician with K=0) [33]. Given h, the link capacity is C(h)=B· log2 1+ _[|]N[h][|]0[2]B[P]_, where P is the
transmission power, N0 is the noise power spectral density at the receiver, and B is the channel
bandwidth. We assume that other sources of signal degradation, such as the Doppler effect, are
well-compensated at the receiver (for example, see the approaches in [34]).
Since the large-scale fading components typically vary slowly relative to the acquisition rate
of Channel State Information (CSI), we assume that the current large-scale parameters (β, K) are
known at the transmitter’s side throughout the communication process, using CSI feedback over
the control channel. Conversely, small-scale fading conditions vary at a much faster timescale and
cannot be tracked at the transmitter. Hence, given (β, K) and a transmission rate Υ [bits/second],
we define the outage probability as Pout(Υ, β, K)≜P(C([√]βg)<Υ)|β, K)=P (|g|[2]<u(Υ, β)),
Υ
where u(Υ, β)≜ _[N]βP[0][B]_ [(2] _B −1). The expected throughput is then R(Υ, β, K)=Υ· (1−Pout(Υ, β, K)),_
assuming that the small-scale fading is averaged out across time. The rate Υ is then selected
to maximize the expected throughput (as opposed to the approach in [14], which imposes an
outage probability constraint) as Υ[∗](β, K)≜ arg maxΥ≥0 R(Υ, β, K), solved in Proposition 1.
**_Proposition 1. Given the large-scale parameters (β, K) and γ≜_** _[N]βP[0][B]_ [, the optimal throughput-]
maximizing rate is Υ[∗](β, K)=B log2 �1+ _[Z]2[∗]_ �, where Z _[∗]_ is the unique solution in (0, ∞) of
_h[′](Z) ≜_ (2+Z) ln1 �1+ _[Z]2_ � _−_ _[γ][(][K][+1)]2_ _[e][−][K]_ exp{−γQ(K1( + 1)√2K,[Z]2 �[}][I][0]γ[(](�K2+1)γKZ(K) +1)Z) = 0, (1)
where I0(x) is the modified Bessel function of first kind of order 0, Q1(·, ·) is the standard
Marcum Q-function [14]. Z _[∗]_ is solvable via the bisection method. The expected throughput is
_√_
_R[∗](β, K) ≜_ max
Υ≥0 _[R][(Υ][, β, K][) = Υ][∗][(][β, K][)][ ·][ Q][1][(]_
�
2K, 2(K + 1)u(Υ[∗](β, K), β)). (2)
_Proof. See Appendix A._
�
When K=0 (Rayleigh fading for NLoS), Q1 specializes to Q1(0,
2u(Υ, β))= exp _u(Υ, β)_,
_{−_ _}_
while the condition h[′](Z)=0 becomes (1+ _[Z]_
2 [) ln(1+] _[Z]2_ [)=][ 1]γ [. Finally, with the LoS and NLoS]
conditions averaged out in the temporal and spatial dimensions, the average link throughput is
-----
_R¯(d, ϕ) ≜_ _PLoS(ϕ) · R[∗](βLoS(d), K(ϕ)) + (1 −_ _PLoS(ϕ)) · R[∗](βNLoS(d), 0)._ (3)
This expression is then specialized to the three distinct communication links by expressing the
transmission powers, the environment-specific parameters (z1,z2,k1,k2), the large-scale parameters
(β, K), and the LoS/NLoS probabilities based on the spatial configuration, i.e., d and ϕ. For
the GN→BS link, we let _R[¯]GB(r) be the throughput with the GN in position (r, θ), computed_
by setting the GN-BS distance as d=�HB[2] [+][r][2][ and the elevation angle as][ ϕ][= sin][−][1][ �] _[H]d[B]_ � in
(3). Similarly, for the GN→UAV link, we let _R[¯]GU_ (rGU ) be the throughput when the GN-UAV
distance (projected onto the x−y plane) is rGU, computed by setting the GN-UAV Euclidean
distance as d=�rGU[2] [+][H]U[2] [and the elevation angle as][ ϕ][= sin][−][1][ �] _[H]d[U]_ � in (3). Finally, for the
UAV→BS link, we let _R[¯]UB(rUB) be the throughput when the x−y projected UAV-BS distance_
�
is rUB, computed by setting the GN-UAV Euclidean distance as d= _rUB[2]_ [+(][H][U] _[−][H][B][)][2][ and the]_
elevation angle as ϕ= sin[−][1][ �] _[H][U]_ _[−][H][B]_ � in (3). As shown in Figs. 1b, the poor QoS experienced
_d_
by GNs farther away from the BS, caused by deterioration in LoS probabilities with distance,
motivates the need for UAV-relays to improve coverage throughout the cell.
**UAV Mobility Power Model: For a rotary-wing UAV, since its communication power needs**
( 10 W) are dwarfed by its mobility power requirements ( 1000 W), we model the on-board
_≈_ _≈_
energy expenditure as a function of the horizontal flying velocity V [10], i.e.,
�0.5
�
+ P2
��
1 + _[V][ 4]_
_−_ _[V][ 2]_
4v0[4] 2v0[2]
+ P3V [3], 0 ≤ _V ≤_ _Vmax,_ (4)
_Pmob(V ) = P1_
�
1 + [3][V][ 2]
_Utip[2]_
where Pi are the scaling constants, Utip is the rotor blade tip velocity, v0 is the mean rotor
induced velocity while hovering, and Vmax is the maximum UAV flying speed [10]. We let
_Pmax ≜_ max min
0≤V ≤Vmax[P][mob][(][V][ )][ and][ P][min][ ≜] 0≤V ≤Vmax[P][mob][(][V][ )][ be the maximum and minimum power]
consumption of the UAV, respectively. From [10], hovering requires Pmob(0) =1371 W, while
flying at 22 m/s only consumes Pmin =936 W. This suggests that the mobility of the UAVs
can be exploited to reduce power consumption, while simultaneously improving coverage across
the cell. Our goal is to define an energy-conscious adaptive service scheduling and trajectory
optimization scheme to minimize the time-averaged communication delay experienced by GNs
in the cell, subject to an average per-UAV mobility power constraint, studied next.
-----
Fig. 2: The single-agent specialization of our generalized deployment depicted in Fig. 1a.
III. MAESTRO: A SEMI-MARKOV DECISION PROCESS FORMULATION
We now specialize the system model to a single UAV relay (illustrated in Fig. 2) via an SMDP
formulation. The effective traffic rate experienced by a single UAV is Λ[′]≜ _N[Λ]U_ [[requests/unit]
time/UAV], assumed in this section in place of the overall rate Λ. Let qU (t)=(rU (t), θU (t)) be
the polar coordinate of the UAV at time t, projected onto the x−y plane, where rU (t)∈R+ and
_θU_ (t)∈[0, 2π) denote the UAV’s radius and angle with respect to the BS. The system operates
with the following phases. In the waiting phase, no GN requests are being served by the UAV,
which moves according to a waiting policy. When a new GN request originates in position (r, θ),
the system transitions to the request scheduling phase, where it is determined whether the GN
should transmit its data payload directly to the BS, or relay it through the UAV. In case of direct
transmission, the system immediately re-enters the waiting phase, as the UAV remains free to
serve other requests; else, the system enters the UAV relay phase, in which the data payload is
relayed through the UAV using the D&F protocol; upon completion, the system re-enters the
waiting phase. In this section, we conservatively assume that: 1) when the UAV is serving a
request, it is unable to serve other incoming requests, which are thus directly served by the
BS; and 2) data channels are always available at the BS to serve incoming requests. We defer
to Sec. V for the description of a piggybacking mechanism to simultaneously serve multiple
transmission requests, and of a queuing mechanism when data channels are unavailable.
**Communication Delay and UAV Energy Consumption: Here, we formulate the average**
communication delay and UAV energy consumption under a given policy µ that defines the
request scheduling, communication strategy, and UAV trajectory (formally defined later). We
define a decision interval as the time duration spanning the start of a waiting phase, the subsequent
request scheduling phase when a GN request is received, until the system re-enters the waiting
phase after scheduling a direct transmission to the BS, or following the UAV relay phase.
-----
Consider the uth such decision interval of duration ∆u, split into the time ∆u[(][w][)] to wait for
a new request, and the time ∆[(]u[s][)] to serve it, either through the BS (scheduling decision ξu=0)
or through the UAV (ξu=1). Then, ∆u=∆[(]u[w][)][+][ξ]u[∆]u[(][s][)][, since the UAV enters the waiting phase]
immediately (and the decision interval terminates) in case of direct-BS transmission. Let Nu≥0 be
the number of additional requests received during the UAV relay phase of the uth decision period:
since these are served directly by the BS, we denote their delays as ∆[(]u,i[bs][)][, i][=][{][1][,][ 2][, . . ., N][u][}][. Let]
_Eu be the UAV mobility energy expended during the uth decision interval, and let Mt be the_
total number of decision intervals completed up to time t. We define the expected long-term
average communication delay per request (D[¯] _µ) and average UAV power (P[¯]µ), under µ, as_
�
_,_ _P[¯]µ ≜_ lim
_t→∞_ [E][µ]
�
¯
_Dµ ≜_ lim
_t→∞_ [E][µ]
� _M1t_ �Mu=1t [(∆]u[(][s][)] [+][ ξ]u �Ni=1u [∆]u,i[(][bs][)][)]
1 �Mt
_Mt_ _u=1[(1 +][ ξ][u][N][u][)]_
� 1 �Mt
_Mt_ _u=1_ _[E][u]_
1 �Mt
_Mt_ _u=1_ [∆][u]
_._ (5)
Note that _D[¯]_ _µ in (5) captures the delays of all requests, i.e., those relayed through the UAV_
(ξu=1), those transmitted directly to the BS (ξu=0), as well as the Nu additional requests served
directly by the BS during the UAV relay phase. Thus, the objective is to solve
_D¯_ _[∗]_ = minµ _D¯_ _µ, s.t. ¯Pµ ≤_ _Pavg,_ (6)
where Pavg ∈ (Pmin, Pmax) is the average power constraint, and the optimal policy is denoted
as µ[∗]. To simplify, let E[¯] _µ[Cu]≜_ _tlim→∞_ [E][µ][[][ 1]Mt �Mu=1t _[C][u][]][ be a shorthand notation for the long-term]_
average cost Cu per decision interval. Let _E[¯]µ≜E[¯]_ _µ [Eu] be the average UAV energy expenditure,_
_T¯µ≜E¯_ _µ [∆u] be the average interval duration, ¯Nµ≜E¯_ _µ[1+ξuNu] be the average number of requests_
served, _W[¯]_ _µ[(][s][)][≜][E][¯]_ _µ[[∆]u[(][s][)][]][ be the average delay of requests for which a scheduling decision is made,]_
_W¯_ _µ[(][bs][)]≜E[¯]_ _µ[ξu_ �Ni=1u [∆]u,i[(][bs][)][]][ be the average delay of requests served directly by the BS during the]
UAV relay phase, per decision interval. Using Little’s Law [35], we can then express _P[¯]µ=_ _ET¯¯µµ_
and _D[¯]_ _µ=_ _W¯_ _µ[(][s][)]N+ ¯µW[¯]_ _µ[(][bs][)]_, hence the optimization problem can be recast as
¯
_D[∗]_ = min
_µ_
_W¯_ _µ[(][s][)]_ + W[¯] _µ[(][bs][)]_
¯ s.t. _E[¯]µ ≜_ _E[¯]µ −_ _PavgT[¯]µ ≤_ 0, (7)
_Nµ_
where _E[¯]µ=E[¯]_ _µ[Eu−Pavg∆u] is the excess energy cost. Note the inherent complexity to solve_
(7): as the policy varies, the delay metric changes both the numerator and denominator of the
objective function, precluding a direct application of dynamic programming tools.
**Alternative Problem Formulation: To address this challenge, we now devise a surrogate**
-----
optimization metric, by characterizing upper and lower bounds to _D[¯]_ _µ. To this end, let us define_
a "baseline" policy µBS as the one such that all requests are served by the BS and the UAV
flies around at minimum power Pmin (this policy is feasible). Since the delay to serve a request
from a GN in position (r, θ) by direct transmission to the BS is _R¯GBL_ (r) [, the expected delay]
under policy µBS is obtained by computing the expectation with respect to the radial coordinate,
_D¯_ _BS≜_ ´ a0 _R¯GBL_ (r) _[f]R[(][r][)d][r][. Clearly, optimization of the policy yields][ ¯][D][∗][≤][D][¯]_ _BS[. Under any policy]_
_µ (including µ[∗]) better than µBS (i.e., such that_ _D[¯]_ _µ_ _DBS), the following bounds hold._
_≤_ [¯]
**_Proposition 2. Let µ be such that_** _D[¯]_ _µ ≤_ _D[¯]_ _BS. Then, it holds that_
_W¯_ _µ[(][s][)]_ _≤_ _D[¯]_ _µ ≤_ _W[¯]_ _µ[(][s][)]_ 1 + Λ[′][ ¯]DBS _≤_ _D[¯]_ _BS._ (8)
1 + Λ[′][ ¯]Wµ[(][s][)]
_Proof. See Appendix B._
Noticing that both the lower and upper bounds of _D[¯]_ _µ are increasing functions of_ _W[¯]_ _µ[(][s][)][, in our]_
subsequent analyses we will focus on the alternative optimization problem
minµ _W¯_ _µ[(][s][)]_ s.t. _E[¯]µ ≤_ 0. (9)
In Sec. VI (see Table III), we show that this alternative formulation leads to a near-optimal
solution with respect to the original optimization (6). To solve (9), we define the Lagrangian
_Mt_
�
_u=1_
�
� �
∆u[(][s][)] [+][ ν][(][E][u][−][P][avg][∆][u][)]
_g(ν) = minµ_ _W¯_ _µ[(][s][)]_ + νE[¯]µ = minµ _tlim→∞_ [E][µ]
�
1
_Mt_
_,_ (10)
where ν is the dual variable, optimized by solving maxν≥0 g(ν). We now demonstrate that for
a given ν 0, (10) can be cast as a Semi-Markov Decision Process (SMDP) and solved with
_≥_
dynamic programming tools. Next, we discuss the SMDP states, actions, transitions, and policy.
**States: The state is defined by the UAV position qU**, an element of the set QUAV≜R+×[0, 2π)
(polar coordinates), and the position qG of the GN originating traffic, taking values from the
set QGN≜[0, a]×[0, 2π). The state space is then S=Swait ∪Scomm, where Swait=QUAV is the
set of waiting states and Scomm=QUAV×QGN is the set of communication states. Crucial to the
definition of the SMDP is how the system is sampled in time to define Markovian dynamics in
the evolution of the sampled states: accordingly, we define the actions available in each state
**s** and the transition probabilities, along with the time duration T (s; a), the UAV energy usage
_∈S_
_E(s; a), and the request service delay ∆(s; a) metrics accrued in state s under action a._
-----
**Waiting states’ actions, transitions, and metrics: In waiting state s=qU** _∈Swait at time t, i.e., the_
UAV is in position qU (t)=qU =(rU _, θU_ ) with no active requests, then the UAV moves with radial
and angular velocity components (vr, θc), over an arbitrarily small duration ∆0≪ Λ[1][′] [. Thus, the]
� �
� �
waiting-state action space is Await(rU )≜ (vr, θc)∈R[2][��] _vr[2][+][r]U[2]_ _[·][θ]c[2][≤][V][max]_, where vU = _vr[2][+][r]U[2]_ _[θ]c[2]_
�
is the velocity expressed using polar coordinates. Upon choosing action a=(vr, θc)∈Await(rU ),
the communication delay is ∆(s; a)=0, since there is no ongoing communication; the duration
of a waiting state is T (s; a)=∆0, and the UAV’s energy use is E(s; a)=∆0Pmob (vU ) to move
at velocity vU . The new state is then sampled at time t+∆0, with the UAV moved to the new
position qU (t+∆0)≈(rU _, θU_ )+(vr, θc)∆0. With probability e[−][Λ][′][∆][0], no new request is received
in the time interval [t, t+∆0], so that the new state is a waiting state. Otherwise, a new request
is received from a GN in position (r, θ) (communication state). The transition probabilities from
the waiting state sn=qU _∈Swait under action an=(vr, θc)∈Await(rU_ ) are thus
P(sn+1 = qU + an∆0|sn, an) = e[−][Λ][′][∆][0], (11)
P(sn+1 = (qU + an∆0, q[′]G[)][ with][ q][′]G _[∈F |][s][n][,][ a][n][) =][ A]πa[(][F][2][ ·][)]_ [ (1][ −] _[e][−][Λ][′][∆][0][)][,][ ∀F ⊆Q][GN][,]_
where A( ) is the area of region, since requests are uniformly distributed in the cell.
_F_ _F_
**Communication states’ actions, transitions, and metrics: Upon reaching a communication**
state sn=(qU _, qG)∈Scomm at time t, the system must serve a GN request at position qG=(r, θ)._
The BS first determines the scheduling decision ξ 0, 1 . If ξ=0, denoted as the action a=BS,
_∈{_ _}_
the GN transmits directly to the BS; the next state is the waiting state sn+1=qU, sampled
immediately after, resulting in the energy-time metrics E(sn; a)=T (sn; a)=0, and service delay
metric ∆(sn; a)= _R¯GBL_ (r) [(time required to transmit the payload with throughput][ ¯][R]GB[(][r][)][ between]
the GN and the BS). Instead, if ξ=1, the UAV uses the D&F protocol, while following a
trajectory starting from its current position qU and ending in position q[′]U [. We denote this action]
as a=(qU _→q[′]U_ [)][. In the][ decode][ phase of D&F (of duration][ t][p][), the GN transmits its data payload]
to the UAV; in the forward phase (of duration ∆−tp), the UAV relays it to the BS. Assuming a
_move-and-transmit strategy [10], the trajectory (qU_ _→q[′]U_ [) and the durations (][t][p][ and][ ∆][−][t][p][) must]
satisfy the data payload constraints (C.1), i.e., the entire payload of L bits is first transmitted to the
UAV with throughput _R[¯]GU_ (rGU (η)), and then relayed to the BS with throughput _R[¯]UB(rUB(η)),_
where rGU (η) and rUB(η) are the GN-UAV and UAV-BS distances (projected onto the x−y
plane) at time η along the trajectory, respectively, so that the total communication delay is ∆.
-----
∆
For this action, the cost metrics are ∆(sn; a)=T (sn; a)=∆ and E(sn; a)= ´0 _[P][mob][ (][v][U]_ [(][η][)) d][η][.]
Upon completing D&F at time t+∆, the UAV enters the waiting state (sn+1=q[′]U [). The set of]
feasible UAV trajectories from qU to q[′]U [, to serve a GN at position][ q][G][ is]
�
_QqG�qU →_ **q[′]U** � ≜ **pU : [0, ∆] �→** R+ × [0, 2π) s.t. (12)
ˆ tp ˆ ∆
_R¯GU_ (rGU (η))dη ≥ _L,_ _R¯UB(rUB(η))dη ≥_ _L,_ (C.1)
0 _tp_
�
_vU_ (η) ≤ _Vmax, pU_ (0) = qU _, pU_ (∆) = q[′]U _[,][ ∃][∆]_ _[≥]_ [0][,][ ∃] [0][ ≤] _[t][p]_ _[≤]_ [∆] _,_ (C.2)
where vU (η) is the UAV speed, C.1 reflects the data payload constraints, and C.2 the maximum
speed and trajectory constraints. Then, the action space in state (qU _, qG)∈Scomm when ξ=1 is_
the set QqG(qU )≜ _∪q[′]U_ _[∈Q][UAV][ Q][q][G]�qU_ _→q[′]U_ � of feasible trajectories starting in qU that serve the
GN at qG via the D&F protocol. The overall action space of this communication state is then
_Acomm(qU_ _, qG)≜{BS}∪{QqG(qU_ )}, including the scheduling decision ξ ∈{0, 1}.
**Policy µ: For waiting states qU** _∈Swait, the policy µ(qU_ )∈Await(rU ) selects a velocity (vr, θc)
from the respective action space. Likewise, for communication states (qU _, qG)∈Scomm, the_
policy selects the scheduling decision ξ 0, 1 and if ξ=1, the trajectory followed in the D&F
_∈{_ _}_
protocol, i.e., µ(qU _, qG)∈QqG(qU_ ). With a stationary policy µ defined, the Lagrangian metric
_L[(]µ[ν][)][≜][W][¯]_ [ (]µ[s][)][+][ν][ ¯][E]µ [in (10) is reformulated using Little’s Law [35] and is written as]
1
=
_πcomm_
�
ˆ
Πµ(s)ℓν(s; µ(s))ds, (13)
_S_
_L[(]µ[ν][)]_ = lim
_N_ _→∞_ [E][µ]
� 1 �N _−1_
_N_ _n=0_ _[ℓ][ν][(][s][n][;][ µ][(][s][n][))]_
1 �N _−1_
_N_ _n=0_ [I][(][s][n][ ∈S][comm][)]
where Πµ(s) is the steady-state probability density function of being in state s under policy µ,
_πcomm=_ ´Scomm[Π][µ][(][s][)d][s][ is the steady-state probability that the UAV is in the communication phase,]
and ℓν(s; a)≜∆(s; a)+ν�E(s; a)−PavgT (s; a)� is the Lagrangian metric in state s under action a.
In (13), [�]n[N]=0[−][1] _[ℓ][ν][(][s][n][;][ µ][(][s][n][))][ is the total Lagrangian cost accrued during the first][ N][ SMDP stages,]_
and [�]n[N]=0[−][1] [I][(][s][n][∈S][comm][)][ is the number of communication states encountered; since a new decision]
interval initiates after a communication state, this equals the number of decision intervals (Mt
in (10)). Taking the limit N _→∞, L[(]µ[ν][)]_ is the expected Lagrangian cost per decision interval, as
expressed in (10). The right-hand side expression in (13) follows because the SMDP reaches the
�
steady-state when N _→∞. Specializing, ℓν(rU_ _, θU_ ; vr, θc)=ν(Pmob( _vr[2][+][r]U[2]_ _[θ]c[2][)][−][P][avg][)∆][0]_ [for the]
waiting states, ℓν(rU _, θU_ _, r, θ; BS)=_ _R¯GBL_ (r) [for direct-BS transmission in communication states,]
∆
and ℓν(rU _, θU_ _, r, θ; pU_ )=(1−νPavg)∆+ν ´0 _[P][mob][ (][V][ (][η][)) d][η][ for a communication relayed through]_
-----
the UAV. The next proposition shows that the steady-state probability πcomm is independent of
the policy µ, i.e., it is not affected by the optimization over µ.
**_Proposition 3. We have πcomm=1 −_** (2−e[−][Λ][′][∆][0])[−][1].
_Proof. See Appendix C._
This result permits rewriting (10) as an average cost-per-stage problem
1
_g(ν) =_ min
_πcomm_ _µ_
ˆ
Πµ(s)ℓν(s; µ(s))ds, (14)
_S_
solvable through standard dynamic programming approaches (upon discretization of the state
and action spaces), followed by the dual maximization maxν≥0g(ν).
**Two-stage policy decomposition: Since GN transmission requests are uniformly distributed in**
the circular cell, the UAV radius is a sufficient statistic in decision-making for a waiting state
(rU _, θU_ ), expressed as rU _∈Swait ≜_ [0, a]. Likewise, for a communication state (rU _, θU_ _, r, θ), only_
the UAV radius, GN request radius, and the angle ψ [0, 2π) between them suffice to characterize
_∈_
the state. Thus, communication states can be compactly represented as (rU _, r, ψ=θ−θU_ )∈Scomm ≜
[0, a][2] [0, 2π). Hence, the policy affects the SMDP state transitions (and its steady-state) only
_×_
through the UAV radial velocity vr in the waiting states, the scheduling decision (direct-BS or
UAV relay) and UAV trajectory’s end radius position ˆrU in communication states. Instead, the
angular velocity θc in the waiting states and the UAV trajectory to reach the target end radius ˆrU
in the communication states only affect the instantaneous Lagrangian ℓν, but not state dynamics.
With this observation, let O(rU )≜vr∈[−Vmax, Vmax] define the radial velocity policy of waiting
states rU _∈Swait, specifying the radial velocity component of waiting action (vr, θc)∈Await(rU_ );
let U (rU _, r, ψ)≜(ξ, ˆrU_ ) define the scheduling and next radius position policy of communication
states (rU _, r, ψ)∈Scomm: either direct-BS with ˆrU = rU (ξ = 0), or any trajectory starting from_
radius rU and ending at radius ˆrU when relaying through the UAV (ξ = 1). Accordingly, O and
_U are the SMDP’s outer decisions and are the only actions affecting the steady-state distribution,_
denoted as ΠO,U under the outer policy (O, U ); thus, (14) can be restated as
1
_g(ν) =_ min
_πcomm_ _O,U_
� [ˆ] ˆ �
ΠO,U (s)ℓ[∗]ν[(][s][;][ O][(][s][))d][s][ +] ΠO,U (s)ℓ[∗]ν[(][s][;][ U] [(][s][))d][s] _,_ (15)
_Swait_ _Scomm_
where ℓ[∗]ν [is the Lagrangian metric optimized with respect to the][ inner decision][ components not]
specified by O and U . In particular, for a waiting state rU, under the radial velocity action
_O(rU_ )=vr, the inner optimization is performed with respect to the angular velocity θc,
-----
�
_ℓ[∗]ν[(][r][U]_ [;][ v][r][) = min]θc _ν (Pmob(V ) −_ _Pavg) ∆0 s.t. V =_
_vr[2]_ [+][ r]U[2] _[θ]c[2]_ (16)
_[≤]_ _[V][max][.]_
Since ν≥0, the optimizer θc[∗] [is the angular velocity minimizing the UAV power consumption: due]
to the quasi-convex structure of Pmob(v) [10], θc[∗][=0][ if][ |][v][r][|≥][v][P]min[≜] [arg min][V] _[P][mob][(][V][ )][ (in fact,]_
�
any angular movement would undesirably increase power consumption), and _vr[2][+][r]U[2]_ [(][θ]c[∗][)][2][=][v][P]min
otherwise (i.e., enough angular movement to yield the power minimizing speed). For communi
cation states, under direct-BS transmission, ℓ[∗]ν[(][s][; 0][, r][U] [) =][ L/R][GB][(][r][)][; on the other hand, when]
relaying through the UAV, ℓ[∗]ν [is obtained by optimizing the trajectory][ p][U] [followed by the UAV,]
starting at radius rU and terminating at radius ˆrU (with final angular position _φ[ˆ] optimized),_
∆
ˆ
_ℓ[∗]ν[(][s][; 1][,][ ˆ][r][U]_ [)= min] (1−νPavg)∆+ν _Pmob(vU_ (η))dη s.t. C.1, C.2. (17)
∆,pU _,tp,φ[ˆ]_ 0
where C.1-C.2 are the data payload, maximum UAV speed and trajectory constraints (see (12)). In
other words, the inner decision on trajectory minimizes the instantaneous delay-energy trade-off,
among all feasible trajectories terminating at the target radius ˆrU . Defining α≜ (1+ν(2νPPmaxmax−Pavg)) _[∈]_
[0, 1] to regulate the trade-off between service delay and UAV energy, (17) can be rewritten as
∆
_ℓ[∗]ν[(][s][; 1][,][ ˆ][r][U]_ [)] ˆ
1+ν(2Pmax−Pavg) [= min]∆,pU _,tp[(1][ −]_ [2][α][)∆+][α] 0
_Pmob(V (η))_
dη s.t. C.1, C.2, (18)
_Pmax_
This reformulation is the focus of our HCSO trajectory design algorithm, detailed in Sec. IV.
Alg. 1 optimizes the outer policy and computes the average cost-per-stage metric g(ν), along
with the average excess energy-per-stage metric for a given ν, by solving problem (15) via value
iteration [25]. Alg. 2 solves the dual maximization maxν≥0g(ν) via projected sub-gradient ascent[1]
[36]. Specifically, in Alg. 1, lines 2 and 3 compute the inner Lagrangian cost metric optimized
with respect to the inner actions—along with the excess energy cost metric—for all states and
outer actions; line 6 computes the value iteration update for waiting states: upon moving to the
new radial position rU +vr∆0, no request is received, w.p. e[−][Λ][′][∆][0], hence moving to a waiting state
(with future value VW,i(rU +vr∆0)); otherwise, the system moves to a communication state, with
future value VC,i(rU +vr∆0) (averaged with respect to the request position); line 12 computes
the value iteration update for communication states, transitioning to a waiting state w.p. 1; the
corresponding optimal outer actions are saved in lines 7 and 13; line 16 averages the value of
communication states with respect to the random request position; lines 8, 14, and 17 similarly
[1The source code for these algorithms is available on GitHub [2].](https://github.com/bharathkeshavamurthy/MAESTRO-X.git)
-----
**Algorithm 1 (O[∗], U** _[∗], g(ν),_ _E[¯], V·,[next]0_ _, E·[next],0_ [) = VITER(][ν, V][·][,][0][,][ E][·][,][0][)]
1: Initialization: i=0; stop criterion δ.
2: Inner optimization in waiting states: ∀rU _∈Swait, ∀vr∈[−Vmax, Vmax], calculate ℓ[∗]ν_ [(][r][U] [;][ v][r][)][ as in (16), with minimizer][ θ]c[∗][; compute]
�
excess energy cost ϵ[∗](rU ; vr)=Pmob( _vr[2] + rU[2]_ [(][θ][c][∗][)][2][)∆][0][ −] _[P][avg][∆][0][.]_
3: Inner **optimization** **in** **communication** **states:** _∀s∈Scomm, ∀rˆU_ _∈[0, a],_ calculate _ℓ[∗]ν_ [(][s][; 1][,][ ˆ][r][U] [)] via Alg. 3 with _α_ =
_νPmax/(1+ν(2Pmax−Pavg)), with minimizer p[∗]U_ [(trajectory); compute excess energy cost][ ϵ][∗][(][s][; ˆ][r][U] [)=][E][(][s][;][ p]U[∗] [)][ −] _[P][avg][T]_ [(][s][;][ p][∗]U [)][.]
4: repeat
5: **for each rU** _∈[0, a] do_ _▷_ Outer optimization in waiting states
6: _VW,i+1(rU_ )←vr _∈[−Vminmax,Vmax]�ℓ[∗]ν_ [(][r][U] [;][ v][r][)+][e][−][Λ][′][∆][0] _[V][W,i][(][r][U]_ [+][v][r][∆][0][)+(1][−][e][−][Λ][′][∆][0] [)][V][C,i][(][r][U] [+][v][r][∆][0][)]�,
7: _Oi+1(rU_ ) ← _vr[∗][, where][ v]r[∗]_ [is the][ arg min][.]
8: _EW,i+1(rU_ )←ϵ[∗](rU ; vr[∗][)+][e][−][Λ][′][∆][0] _[E][W,i][(][r][U]_ [+][v]r[∗][∆][0][)+(1][−][e][−][Λ][′][∆][0] [)][E][C,i][(][r][U] [+][v]r[∗][∆][0][)][.]
9: **end for**
10: **for each rU** _∈[0, a] do_ _▷_ Outer optimization in communication states
11: **for each r∈[0, a], ψ∈[0, 2π) (s = (rU** _, r, ψ)) do_ _▷_ Outer optimization in communication states
12: _Vˆ (s)←_ min � _RGBL_ (r) [+][V][W,i][(][r][U] [)], _rˆUmin ∈[0,a][ℓ]ν[∗]_ [(][s][; ˆ][r][U] [)+][V][W,i][(ˆ][r][U] [)] � _▷_ Value function given GN position
� �� � � �� �
_ξ=0_ _ξ=1_
13: _Ui+1(s) ←_ (ξ[∗], ˆrU[∗] [)][, where][ (][ξ][∗][,][ ˆ][r]U[∗] [)][ is the][ arg min][ (][r][ˆ]U[∗] [=][ r][U][ if][ ξ][∗] [= 0][).]
14: _Eˆ(s)←ξ[∗]_ _· ϵ[∗](s; ˆrU[∗]_ [)+][E][W,i][(ˆ][r]U[∗] [)][.] _▷_ Total excess cost given GN pos., optimized over scheduling/trajectory
15: **end for**
16: _VC,i+1(rU_ )← ´ 20 _π_ 21π ´ a0 _a2r[2][ ˆ][V][ (][r][U]_ _[, r, ψ][)d][r][d][ψ][′]_ _▷_ Value function in comm states, averaged over GN position
17: _EC,i+1(rU_ )← ´ 20 _π_ 21π ´ a0 _a2r[2][ ˆ][E][(][r][U]_ _[, r, ψ][)d][r][d][ψ][′]_ _▷_ Excess energy cost in comm states, averaged over GN position
18: **end for**
19: _∀rU ∈_ [0, a] and X ∈{W, C}, calculate δX[(][V][ )](rU )=VX,i+1(rU )−VX,i(rU ) and δX[(][E][)][(][r][U] [)=][E][X,i][+1][(][r][U] [)][−E][X,i][(][r][U] [)][;][ i][←][i][+1][.]
20: until maxrU,X δX[V] [(][r][U] [)][−] [min][r]U _[,X][ δ]X[V]_ [(][r][U] [)][<δ][ and][ max][r]U _[,X][ δ]X[E]_ [(][r][U] [)][−] [min][r]U _[,X][ δ]X[E]_ [(][r][U] [)][<δ][.] _▷_ Termination condition
21: return g(ν)≈δW[(][V][ )][(0)][/π][comm][,][ ¯][E≈][δ]W[(][E][)][(0)][.] _▷_ dual cost and average excess energy cost
22: _V·[next],0_ (·)=V·,i(·)−VW,i(0), E·[next],0 (·)=E·,i(·)−EW,i(0). _▷_ Relative values (next VITER initialization)
23: _O[∗](·)=Oi(·), U_ _[∗](·)=Ui(·)._ _▷_ Optimal waiting and communication policies
update the total excess energy cost, needed to compute the projected dual sub-gradient ascent
in Alg. 2. In practice, the integrals in lines 16 and 17, and the continuous state/action spaces
are discretized (see MAESTRO-X [2]), leading to an overall complexity of each value iteration
update (lines 5-18) of order O(KR _·(KV +KR[2]_ _[·][K][A][))][, where][ K][R][ is the number of discretized radii]_
levels (rU and r values), KA is the number of angular levels (ψ and ψ[′]), and KV is the number of
discretized radial velocities (vr). Upon convergence (typically, value iteration converges within
(log(1/δ)) iterations to achieve a target accuracy δ [25, Sec. V]), line 21 estimates the values
_O_
of the average cost-per-stage and excess energy-per-stage metrics.
In Alg. 2, line 1 initializes the dual variable and a sequence of step-sizes used for projected
sub-gradient ascent; line 3 calls value iteration (Alg. 1) using the current dual variable ν, and
outputs the optimal outer policy and the average cost-, excess energy- per-stage metrics; line
5 monitors convergence in terms of primal feasibility and complementary slackness conditions;
line 4 updates the value of the dual variable in the direction of its sub-gradient and projects
its value to the non-negative range to ensure dual feasibility; note that Alg. 1 outputs also the
_relative values metrics V and_ : these are used to initialize the total cost and excess energy
_E_
metrics in the next call to Alg. 1, and help speed up convergence. We are left with the trajectory
design (line 3 of Alg. 1), carried out using Hierarchical CSO in the next section.
-----
**Algorithm 2 Projected Sub-gradient Ascent (PSGA)**
1: Initialization: k = 0; dual variable ν≥0; step-size {ρk= _k[ρ]+1[0]_ _[, k][≥][0][}][;][ V][·][,][0][(][·][)=][E][·][,][0][(][·][)][ ≡]_ [0][.]
2: repeat
3: (O[∗], U _[∗], g,_ _E[¯], V·,0, E·,0) ←_ VITER(ν, V·,0, E·,0) via Alg. 1.
4: Update ν ← max �ν+ρkE[¯], 0�; k←k+1. _▷_ Dual variable update
5: until _E[¯]<ϵP F ; ν|E|[¯]_ _<ϵCS_ _▷_ Check KKT optimality conditions
6: return: optimal outer policy (O[∗], U _[∗])._
**Algorithm 3 HCSO Algorithm**
1: Randomly initialize N particles (p, v)1:N : pi is a sequence of way-points, vi a sequence of UAV speeds.
2: while M ≤ _Mmax do_
3: Obtain M -segment trajectory: (p[∗], v[∗])=CSO(p1:N _, v1:N_ _, N, M_ ) (see [26]). _▷_ CSO call
4: Increase M _←2M_ ; interpolate to form reference trajectory: (˜p, ˜v)=interp(p[∗], v[∗], M ). _▷_ Increase resolution via interpolation
5: Reduce swarm size N _←N_ _−Nred._
6: **for n=1, 2, . . ., N do** _▷_ Generate N particles randomly
7: New way-point particle pn with mth way-point xm = ˜xm+(χm, ζm) and xM = ˆrU _∥xxMM−−11∥2_ [.] _▷_ Way-point perturbation
8: New velocity particle vn with mth velocity vm = [˜vm+κm][[][V][low][,V][max][]]. _▷_ Velocity perturbation
9: **end for**
10: end while
IV. TRAJECTORY DESIGN VIA HIERARCHICAL COMPETITIVE SWARM OPTIMIZATION
In this section, we design the UAV trajectory during the D&F protocol. To solve (18), we
propose a CSO scheme [26] defining a meta-heuristic UAV trajectory. First, as done also with
SCA approaches [10], [16], [37], we simplify the continuous UAV trajectory into a finite sequence
of way-points connected by straight lines at constant velocity. However, a direct application of
CSO to high-resolution trajectory design suffers from poor convergence due to exponentially
large solution spaces [38]. We address this weakness by proposing a Hierarchical variant of CSO
(HCSO), wherein a sequence of problems is solved: initially, CSO produces a low-resolution
trajectory; the optimized trajectory is then interpolated to create a higher-resolution one, then
further optimized with CSO. The process repeats until a target resolution is achieved.
Let x0 = (rU _, 0) be the initial UAV position and xG≜(r cos ψ, r sin ψ) be the request position_
(in this section, expressed as Cartesian coordinates), corresponding to the communication state
**s=(rU** _, r, ψ)∈Scomm. Given a target end radius position ˆrU (the outer action), we encode the_
UAV trajectory as a sequence of M way-points xm=(xm, ym), m = 1, . . ., M, ending at xM
at radius ˆrU, and velocities vm∈[Vlow, Vmax] used to traverse each straight trajectory segment
Ψm≜xm−xm−1. The first and second _[M]2_ [segments correspond to the two phases of the D&F]
protocol. Here, the minimum velocity Vlow≪Vmax ensures well-defined segment durations; the
sequences of way-points p≜[x1, . . ., xM ] and velocities v≜[v1, . . ., vM ] are the optimization
variables. Since the number of bits communicated (C.1) during each trajectory segment, coupled
with our throughput-maximizing rate adaptation scheme, cannot be computed in closed-form,
we approximate them numerically. Specifically, between subsequent way-points xm−1 and xm
-----
traversed with velocity vm, we generate a sequence of nres evenly-spaced points with sufficiently
high resolution; letting {Rk[new]}k[n]=1[res] [be the expected throughput at each point, computed via (3)]
and Prop. 1, the number of bits communicated along the mth segment is approximated as
_Fm≜_ _[∥][Ψ]v[m]m[∥][2]_ _nres1_ �nk=1res _[R]k[new], where_ _[∥][Ψ]v[m]m[∥][2]_ is the time taken to traverse it. Thus, (18) becomes
�
1 2α + α _[P][mob][(][v][m][)]_
_−_
_Pmax_
(P.0) min
**p,v∈[Vlow,Vmax][M]**
_M_
�
_m=1_
_∥Ψm∥2_
_vm_
�
(19)
s.t. hi(p, v) ≜ _L −_
_M_
2 [(][i][+1)]
�
_Fm ≤_ 0, i = 0 and 1, ∥xM _∥2 = ˆrU_ _,_ (C)[˜]
_m=_ _[M]2_ _[i][+1]_
where C enforce the data payload and end radius constraints. To solve[˜] (P.0) with CSO, we first
convert it into an unconstrained one, by penalizing constraint violations with a particular solution:
1) if the UAV does not decode (or forward) its data payload by the end of either phase, then it
flies along the circumference of a circle (radius rmin>0, small) around the current position with
its power-minimizing velocity (vPmin=22 m/s [10]) until the transmission/reception is completed;
and 2) we enforce the end radius constraint by projecting the penultimate way-point xM _−1 to the_
circle at radius ˆrU, i.e. xM = ˆrU **xM** _−1/∥xM_ _−1∥2.[2]_ This yields the penalized objective function
ˆ
� _EP,0+ ˆEP,1_
+(1 − 2α)(t[ˆ]P,0+t[ˆ]P,1)+α ;
_Pmax_
�
1 2α + α _[P][mob][(][v][m][)]_
_−_
_Pmax_
ˆ
_f_ (p, v)≜
_M_
�
_m=1_
_∥Ψm∥2_
_vm_
_tˆP,0≜_ [max][{][h][0][(][p][,][ v][)][,][ 0][}] ; E[ˆ]P,i≜Pmint[ˆ]P,i, xM = ˆrU **xM** _−1_ _,_
_R¯GU_ (∥xM/2 − **xG∥2)** [; ˆ][t][P,][1][≜] [max]R¯UB[{][h]([1]∥[(]x[p]M[,][ v]∥[)]2[,])[ 0][}] _∥xM_ _−1∥2_
where _t[ˆ]P,i and_ _E[ˆ]P,i are the time and energy penalties involved in finishing the data communication_
during the decode and forward phases (i=0 and 1). In particular, _t[ˆ]P,i equals the remaining_
payload max{hi(p, v), 0}, divided by the corresponding throughput at the terminal position
(R[¯]GU for the decode phase and _R[¯]UB for the forward phase). Hence, (P.0) becomes minfˆ(p, v)._
**p,v**
To solve this problem, we employ the HCSO algorithm, outlined in Alg. 3 and discussed next.
We initialize N way-point particles p1:N ≜p1, . . ., pN and N UAV velocity particles v1:N ≜
**v1, . . ., vN (line 1). The core of the algorithm is CSO (line 3), detailed in [26]: essentially,**
during the kth iteration within CSO, the N particles are randomly grouped into _N_
2 [pairwise]
competitions. For both members of a pair, _f[ˆ](p, v) is calculated; the winner of the competition_
is passed onto the (k+1)th iteration, while the loser is modified by learning from the winner,
2We let _∥xx∥2_ [=(1][,][ 0)][ for a point in the origin,][ x][=(0][,][ 0)][.]
-----
Fig. 3: An illustration outlining the sequence of operations under MAESTRO-X that occur at each UAV.
as detailed by the update equations in [26]; after repeating these pair-wise competitions, the
CSO algorithm outputs a winning trajectory (p[∗], v[∗]). However, a direct application of CSO
alone suffers from a complexity-accuracy dilemma: high-resolution trajectories are slow to
converge, while low-resolution ones give rise to poor solutions that fail to capture fine-grained
variations in the trajectory way-points and velocities. To overcome this limitation, we embed
CSO within a hierarchical wrapper: starting from a low-resolution trajectory optimized via CSO,
after each CSO iteration (line 3), the resulting trajectory is interpolated to form a reference
higher-resolution trajectory of M 2M way-points (line 4). The new population size is then
_←_
reduced, N _←N_ _−Nred, to lower the computational burden of CSO (line 5), and a new set of N_
particles is generated randomly. To preserve the quality of the previous lower-resolution trajectory
solution, the mth way-point of each new particle is generated by injecting zero-mean Gaussian
noise χm, ζm∼N �0, σm,X[2] � (line 7) around the reference trajectory; similarly, the UAV velocity
is generated by injecting Gaussian noise κm∼N (0, σV[2] [)][ (line 8), followed by projection onto the]
feasible set ([·][[][V][low][,V][max][]]). Here, the way-point variance σm,X[2] [=][ ς][(][∥][x][˜][m][+1][−][x][˜][m][∥][2][+][∥][x][˜][m][−][1][−][x][˜][m][∥][2][)][,]
with scaling factor ς>0, is determined by the spread between neighboring reference trajectory
way-points. This choice accounts for the empirical observation that in areas with clustered UAV
way-points, the objective function _f[ˆ](p, v) is sensitive to large variations. The speed variance_
_σV[2]_ [=][ ε][(][V][max][−][V][low][)][2][, with scaling factor][ ε>][0][, reflects the observation that the UAV velocities]
exhibit faster convergence with CSO than the trajectory way-points and less sensitivity to random
initialization. These steps in Alg. 3 continue until the desired trajectory resolution is reached.
V. MAESTRO-X: AN EXTENSION TO UAV SWARMS
In this section, we extend MAESTRO to swarms of NU UAV-relays. This eXtension, termed
MAESTRO-X, augments the multiscale optimal policy obtained via SMDP value iteration.
Depicting an example scenario of serving data traffic generated by an aggregation of soil
sensors in precision agriculture, Fig. 3 illustrates its control flow. MAESTRO-X is enabled
by replicating the optimal single-agent policy of the SMDP in Sec. III across the swarm and
-----
employing additional enhancements including spread maximization, consensus-driven conflict
_resolution with queuing dynamics, piggybacking, and frequency reuse. These mechanisms[3]_ are
implemented using a fully-connected distributed mesh network overlaid on the BS and UAVs,
that enables periodic exchanges of command-and-control messages, as depicted in Fig. 3.
**Spread Maximization: Note that the inner action of MAESTRO’s optimal waiting policy is**
symmetric in relation to clockwise and counter-clockwise angular UAV movements. For multiple
UAVs, we leverage this symmetry to proactively position idle UAVs for potential new relay
requests. Specifically, each UAV in the waiting state moves either clockwise or counter-clockwise
(with angular velocity given by (16)), so as to maximize its angular distance from the nearest
UAV in the waiting state, in an attempt to spread out and more readily serve future requests.
To this end, UAV i parses the state flag as 0 and GPS event fields in its control frame (see
Fig. 3). By monitoring the control frames received from other UAVs, it constructs a local peer
list of other waiting state UAVs, and determines its closest peer (in the angular dimension)
_L_
_j[∗]= arg minj∈L |θi−θj|, where θj is the current angular coordinate of UAV j. UAV i then executes_
the angular motion away from UAV j[∗], until new control frames (containing updated positions)
are received from its peers (at the end of the synchronized reporting period) or upon receiving
a new GN transmission request, at which time it transitions to the communication state.
**Consensus-driven Conflict Resolution: In our single-UAV formulation (Sec. III), the scheduling**
action was determined by comparing the Lagrangian costs of direct-BS transmission to that
of relayed UAV service. To extend scheduling decisions to UAV swarms—including queueing
dynamics, as well as simultaneous multi-user service via piggybacking at the UAVs and frequency
reuse (both described later in this section)—the augmented scheduling decision must now 1)
resolve conflicts among the BS and UAVs as to whom should serve a new GN request; 2) facilitate
a consensus on the best node to serve the GN; 3) account for queueing delays experienced at
each potential server node while waiting for data channels to become available. Similarly to the
single-UAV setting, this augmentation is driven by a cost-of-service metric computed at the BS
and at each UAV. The new metric consists of several modifications to the original delay-energy
cost trade-off computed in the single-UAV setting. For new requests served directly by the BS,
the new metric equals the original delay metric, plus an estimate of the time needed for a data
channel to become available (and considers the frequency reuse mechanism to be described).
3Due to space constraints, we keep our discussions on these multi-agent mechanisms brief. For more details on their
[implementation, please refer to our source code on GitHub [2].](https://github.com/bharathkeshavamurthy/MAESTRO-X.git)
-----
This time can be estimated based on the time needed to complete the requests currently served
at the BS, and the time needed to complete those already queued. Thus, for a new GN request at
(r, θ), the augmented cost metric associated with direct-BS transmission is _R¯GBL_ (r) [+][t]BS[, where]
the first term accounts for the transmission time, whereas tBS is the additional waiting time.
Meanwhile, for new requests served by UAV i at radius rU _|i, GN request radius r, and angle_
between them ψU _|i, i.e., state si = (rU_ _|i, r, ψU_ _|i), with target end radius ˆrU_ _|i, the augmented cost_
metric is given by _ℓ[˜][∗]ν[(][s][i][; 1][,][ ˆ][r][U]_ _[|][i][)+][t][U][|][i][. The first term,][ ˜][ℓ][∗]ν[(][s][i][; 1][,][ ˆ][r][U]_ _[|][i][)][, is the Lagrangian cost metric,]_
modified to account for the piggybacking mechanism (described later in this section), wherein
the UAV follows a collated trajectory to handle the new request while serving previous requests;
the second term, tU|i, is an estimate of the time needed for a data channel to become available
(and considering the frequency reuse mechanism). Upon calculating these cost-of-service metrics
for the BS and the UAVs, the network arrives at a consensus on the best node to serve the new
request, i.e., if _R¯GBL_ (r) [+][t]BS[≤][ℓ][˜][∗]ν[(][s][i][; 1][,][ ˆ][r][U] _[|][i][)+][t][U][|][i][,][∀][i][∈{][1][,][ 2][, . . ., N][U]_ _[}][, then the BS serves the request;]_
otherwise, the request is relayed through the UAV i[∗]= arg mini∈{1,2,...,NU _}_ _ℓ[˜][∗]ν[(][s][i][; 1][,][ ˆ][r][U]_ _[|][i][) +][ t][U][|][i][.]_
**Frequency Reuse: To improve the spectrum utilization efficiency, we propose a frequency reuse**
mechanism, allowing multiple serving nodes (the BS and UAVs) to share the same data channel
simultaneously when serving their respective GN requests. When direct-BS transmission is used
to serve a new GN request, a single data channel assignment occurs at the start of direct
transmission. When the new request is instead served using a D&F UAV relay, two distinct
data channel assignments occur: one each for the decode and forward phases of the UAV. In
essence, reuse of an occupied data channel is permitted on the condition that the received SNRs
of nodes sharing the data channel degrade no more than an acceptable pre-specified threshold
permits. Moreover, to make operations of the frequency reuse mechanism more amenable to our
problem, which includes UAVs following time-varying trajectories, we equivalently describe this
SNR degradation threshold by instead using a minimum distance threshold dth.
The frequency reuse mechanism proceeds in the same way, regardless of whether the data
channel assignment under consideration is for a GN using direct-BS transmission, a GN sending
its data to a UAV (decode phase), or a UAV relaying its data payload to the BS (forward phase).
To formalize, let k∈{1, 2, . . ., NC} be the data channel under consideration for reuse; let node
_i be the new transmitter (either a GN beginning its uplink transmission or a UAV beginning_
its forward phase) determining whether reuse of data channel k is possible; let node j be the
intended receiver of the transmission originating from node i; let (k) be the set of active
_T_
-----
transmitters already using data channel k to serve their requests, i.e., a GN transmitting to a BS
or UAV, or a UAV transmitting to the BS during its forward phase; let (k) be the set of active
_R_
receivers already using data channel k, i.e., a UAV receiving an uplink transmission from a GN
during the decode phase, the BS receiving an uplink transmission directly from a GN, or the
BS receiving the data payload from a UAV during the forward phase. For data channel k to be
deemed acceptable for reuse, the following two conditions must both be met:
(FR.1) _dℓ,j ≥_ _dth, ∀ℓ_ _∈T (k),_ (20)
(FR.2) _di,ℓ_ _≥_ _dth, ∀ℓ_ _∈R(k),_ (21)
where di′,j′ is the Euclidean distance between any transmitter i[′] and receiver j[′]. From the above
equations, (FR.1) ensures that the distances between the intended receiver and all currently active
transmitters are above the minimum distance threshold dth, at all times during the execution of
the UAVs’ trajectories. Likewise, (FR.2) ensures that distances between the new transmitter
and all currently active receivers are above the minimum distance threshold dth. Effectively,
satisfying conditions (FR.1) and (FR.2) simultaneously ensure that no received SNR experiences
a degradation beyond a pre-specified limit, and hence data channel k is acceptable for reuse. Next,
given its re-usability, the wait time for a channel to become available is estimated by modeling
queuing dynamics, choosing the channel with the smallest wait time for service. Also, note that,
once a channel is chosen with reuse, since the throughput experienced by the UAV during service
degrades due to the added interference from other transmitters using the same channel, the UAV
might not be able to complete its decode or forward phases using the optimal trajectory: the UAV
then flies along the circumference of a circle (rmin>0) around the phase-specific final way-point
with its power-minimizing velocity (22 m/s) to complete the phase; additionally, we evaluate the
service in this case using the same time and energy penalties discussed in Sec. IV.
**Piggybacking: To facilitate simultaneous multi-user service at the UAVs, we incorporate a**
piggybacking mechanism (in the cost-of-service computation of the consensus-driven conflict
resolution process), wherein a UAV follows a collated trajectory to accommodate new GN uplink
requests while serving previous requests. Recalling from the description of conflict resolution,
for a new request served through UAV i, we consider the state si = �rU _|i, r, ψU_ _|i�, with target_
end radius ˆrU _|i, and modified Lagrangian cost metric_ _ℓ[˜][∗]ν[(][s][i][; 1][,][ ˆ][r][U]_ _[|][i][)][. If UAV][ i][ is currently not]_
serving any other request, this modified cost metric simplifies to _ℓ[˜][∗]ν[(][s][i][; 1][,][ ˆ][r][U]_ _[|][i][)=][ℓ][∗]ν[(][s][i][; 1][,][ ˆ][r][U]_ _[|][i][)][,]_
-----
|Notation|Description|Simulation Value|Col4|Notation|Description|Simulation Value|
|---|---|---|---|---|---|---|
|NG|Number of GNs|30||a|Cell radius|1 km|
|L|Data payload|10 Mbits||W|System BW|20 MHz|
|NC|Number of data channels|4||B|Data channel BW|5 MHz|
|κ|NLoS attenuation constant|0.2|||SNR referenced at 1 m|40 dB|
|(α, α˜)|LoS/NLoS pathloss exponents|(2,2.8)|||UAV mobility power consumption|Eq. (4), params. of [10]|
|(k1, k2)|Rician K-factor parameters [14]|(1,0.05)||(z1, z2)|LoS probability parameters [39]|(9.61,0.16)|
|HU / HB|UAV / BS antenna height|200 m / 80 m||Vmax|Max. UAV speed|55 m/s|
||Control frame reporting period|10 ms|||SINR degradation threshold|5 dB|
TABLE II: The system simulation parameters (unless otherwise stated).
i.e., the original Lagrangian cost metric computed for the single UAV. On the other hand, if the
UAV is currently serving other requests, the UAV computes the cost metric to serve the new
request by piggybacking it, i.e., serving it simultaneously with its current requests on a different
data channel. In this case, the modified cost metric becomes _ℓ[˜][∗]ν[(][s][i][; 1][,][ ˆ][r][U]_ _[|][i][)=][ℓ]ν[(pg)](si; 1, ˆrU_ _|i), where_
_ℓ[(pg)]ν_ (si; 1, ˆrU _|i) is defined to encapsulate modifications to the cost-of-service metric corresponding_
to the amount of data payload of the new request that has been either decoded or forwarded (or
both) during the execution of the current trajectory (serving the UAV’s previous requests). Note
that the energy expended by the UAV serving its current trajectory while piggybacking the new
request is not considered in the cost computed for this new request, since the energy cost has
already been accounted for in the execution of the current trajectory; instead, we consider only
the delays experienced by the piggybacked GN during its associated cost computation.
VI. SIMULATION SETUP AND EVALUATIONS
Unless otherwise stated, we use the parameter values in Table II. To solve (15) via Algorithms
1–3, we discretize the SMDP state and action spaces (with 25 equally-spaced radii levels and 25
radial velocity waiting actions) and apply linearly-interpolated value iteration (see implementation
details documented in [2]). Furthermore, we chose ∆0 = 1s.
Validation of surrogate optimization problem (9): First, we justify the efficacy of our alternative
optimization framework that replaces the original metric _D[¯]_ _µ with the lower bound_ _W[¯]_ _µ[(][s][)][. As]_
depicted in Table III, we observe that the optimized value _W[¯]_ _µ[(][s][∗][)]_ [of the alternative formulation]
(9) is practically identical to the expected delay metric _D[¯]_ _µ∗_ of the original formulation (6),
across various data payload sizes (L) and data traffic arrival rates (Λ[′]). Hence, replacing _D[¯]_ _µ_
with its lower bound _W[¯]_ _µ[(][s][)]_ as the optimization metric leads to near-optimal solutions. Notably,
the surrogate optimization problem (9) is amenable to dynamic programming tools such as value
iteration (see Alg. 1) and enables our proposed two-scale policy decomposition that drastically
reduces the size of the action space in our SMDP formulation. These tools would not be directly
applicable to the original formulation (6) that uses _D[¯]_ _µ as the optimization objective._
-----
|Payload: L|Arrival rate: Λ′|Lower bound: W¯ µ(s ∗)|Expected Delay: D¯ µ∗|Direct-to-BS: D¯ BS|
|---|---|---|---|---|
|1 Mbits|1 req/min/UAV|1.15 s|1.15 s|31.64s|
|10 Mbits|0.2 req/min/UAV|16.41 s|16.41 s|316.38 s|
|100 Mbits|0.033 req/min/UAV|82.17 s|82.17 s|3163.81 s|
TABLE III: Pavg=1 kW: A comparison between the lower bound _W[¯]_ _µ[(][s][∗][)]_ [of][ ¯][D][µ][∗] [(Prop. 2) and direct-BS (][D][ ¯] _[BS][).]_
(a) Optimal Wait Policy. (b) Optimal D&F trajectory.
Fig. 4: L=10 Mbits, Pavg=1.2 kW, Λ[′]=0.2 req/min/UAV: Optimal waiting policy (a) and optimized D&F trajectory
during a communication phase (terminating above the BS) (b). The arrows and associated numerical values represent
the direction of motion and the flying speed in m/s.
MAESTRO policy: We now study illustrative examples of the optimal policy (Fig. 4). We note
that, during the waiting phase (Fig. 4a), the UAV moves towards a radius of 94 m; upon
_≈_
reaching it, it flies at power-minimizing speed (22.5 m/s) along a circle: this allows the UAV to
be well-positioned for future requests (not too close to the BS, and not too far away from it), and
at the same time to minimize its power consumption. Next, Fig. 4b depicts the optimal trajectory
obtained via HCSO (Algorithm 3), for a certain configuration of GN request positions, initial and
target final UAV radii (evident from the figure). Intuitively, during the decode phase, the UAV
flies towards the GN to improve the pathloss conditions; for the same reason, it moves towards
the BS during the forward phase. Additionally, Fig 4b depicts two different trajectory choices
for the GNs at [193, 594] m (GN-0 and GN-1, specular to each other), one corresponding
_±_
to minimum service delay and the other corresponding to minimum service energy: here, in
addition to observing the angular symmetry in our formulation (see Sec. III), we notice that,
under the minimum delay trajectory, the UAV flies faster, to improve pathloss quicker and reduce
-----
the transmission delay; in contrast, it flies slower under the minimum energy trajectory, to save
energy. The delay-energy trade-off in trajectory design is regulated via α, as described by (18).
MAESTRO-X delay-power trade-off: We compare the delay-power trade-off of MAESTRO-X
with adaptations of state-of-the-art algorithms to our setup, namely: the CIRCLE heuristic [20];
a CVXPY implementation of the Successive Convex Approximation scheme (SCA) [10]; a
CVXPY implementation of the Constrained SCA scheme with Alternating Direction Method
of Multipliers (CSCA-ADMM) [16], and a TensorFlow implementation of the Double Deep-Q
Networks framework (DDQN) [19]. Note that all these frameworks are optimized under their
original channel and communication models detailed in the corresponding references (see Table
I for a list of their features), while we evaluate their performance under more realistic models
of dynamic traffic arrivals and A2G channels. In addition, we consider the following custom
heuristics: BS-only, in which GNs transmit directly to the BS without using UAVs; HAP-only
in which GNs transmit directly to a High Altitude Platform (HAP, height=2 km); and Static,
in which the UAVs statically hover at fixed locations. We also compute a Lower Bound to the
delay as follows: for a GN at radius level r, it is the minimum between the delay incurred with
direct-BS transmission (with throughput _R[¯]GB(r)), and a D&F scheme in which the UAV is on_
top of the GN during the decode phase (with throughput _R[¯]GU_ (0)), and on top of the BS during
the forward phase (with throughput _R[¯]UB(0)). Note that this lower bound is not attainable, since_
it neglects the mobility of the UAV. We average the results over 1000 requests.
In Fig. 5a, we plot the delay-power trade-off under low congestion (Λ[′]=0.2 req/min/UAV).
Remarkably, MAESTRO-X allows to regulate the delay-power trade-off, whereas the other
schemes do not. Across such trade-off, it outperforms all other schemes. Specifically, exploiting
the mobility and maneuverability of the UAVs via optimized trajectories demonstrate lower
service delays compared to static UAV deployments: for instance, a single UAV optimized via
MAESTRO under 1 kW power constraint delivers the data payload 29% faster than a static UAV,
while using 27% less power. Notably, under the same power consumption as the competitors,
a single UAV optimized with MAESTRO achieves 38% lower delay than 3 UAV relays under
DDQN [19], and 13 faster service times than the CIRCLE heuristic with 3 UAVs [20]. Adding
_×_
UAVs significantly improves the performance of MAESTRO-X: with 3 UAVs MAESTRO-X
delivers the payloads 4.7 faster than SCA [10] and 8.6 faster than CSCA-ADMM [16]. The
_×_ _×_
gains start to saturate with 2-3 UAVs. In fact, MAESTRO-X approaches the theoretical lower
bound to the delay, for large power consumption values: with more power available, UAVs
-----
(a) Delay-Power Trade-off. (b) Delay chart.
Fig. 5: L=10 Mbits, Λ[′]=0.2 req/min/UAV: Delay-power trade-off (a) and delay charts (b) for MAESTRO-X, stateof-the-art algorithms, and custom heuristics. In (b), MAESTRO-X is evaluated under Pavg =1 kW.
leverage their mobility to improve pathloss conditions; thanks to spread maximization, multiple
UAVs are more likely to be in the vicinity of a request and readily serve it.
In Fig. 5b, we show the contributions of the communication and queue wait times to the
overall delay experienced by the GNs, with MAESTRO-X evaluated under a power constraint
of 1 kW (less than any other scheme, see Fig. 5a). We note that the BS-only deployment suffers
severely due to large communication delays of GNs at the cell edge, causing the queue to
become backlogged. The performance is drastically improved by deploying HAPs (HAP-only),
thanks to their higher elevation and improved LoS conditions. Yet, the delay performance offered
by a HAP-only deployment is poorer than a non-terrestrial deployment involving UAVs: 2.7
_×_
slower than a static UAV and 3.8 slower than a UAV optimized with MAESTRO. Across all
_×_
UAV-assisted implementations, increasing the number of UAVs in the swarm not only lowers the
communication delay but also the queue wait times since more GNs can be served simultaneously.
Remarkably, MAESTRO-X demonstrates negligible queue wait times even with a single UAV:
in this low-traffic regime, requests are served quicker than the rate at which they are generated,
thereby bypassing the need for piggybacking and frequency reuse.
To analyze the impact of these mechanisms, in Fig. 6a and Fig. 6b, we study a high congestion
regime (Λ[′]=20 req/min/UAV). The results depicted in Fig. 6a are qualitatively similar to the low
congestion case with some key differences: for all the competitor schemes, we note a performance
-----
(a) Delay-Power Trade-off. (b) Delay chart.
Fig. 6: L=10 Mbits, Λ[′]=20 req/min/UAV: Delay-power trade-off (a) and delay charts (b) for MAESTRO-X, stateof-the-art algorithms, and custom heuristics. In (b), MAESTRO-X is evaluated under Pavg =1 kW.
degradation, due to the large wait times (Fig. 6b); a similar performance degradation is noted for
MAESTRO-X with a single UAV. However, remarkably, MAESTRO-X with 2-3 UAVs appears
to be unaffected by the higher arrival rate, as also demonstrated by the small queue time. This is
attributed to frequency reuse allowing more efficient spectrum use, and to piggybacking allowing
simultaneous service of multiple requests by each UAV.
MAESTRO-X, impact of number of channels for large swarms: In Fig. 7, we study the impact
of the number of channels (each of 5 MHz) on the average service delay offered by a MAESTRO
X deployment of 10 UAV-relays, in the high congestion regime. Note that the competitors
become computationally intractable with more than 5-6 UAVs, whereas the policy replication
mechanism of MAESTRO-X offers scalability to large UAV swarms (see Fig. 8). The delay
quickly improves by increasing the number of channels, and saturates after 5 channels at 2s
delay (consistent with Fig. 6a). This is a remarkable result: for instance, with 4 channels
(service delay of 4 s), if no frequency reuse was allowed, the network could at most service
_≈_
4[data channels] 15[req/min/data channel]=60 req/min. The ability to serve a much larger rate
_×_
of Λ = 200 req/min is attributed to the frequency reuse mechanism.
Policy convergence time: Finally, in Fig. 8, we benchmark MAESTRO-X against SCA from [10]
(single-agent, model-based), CSCA-ADMM from [16] (model-based), and DDQN from [19]
(model-free), in terms of their policy convergence times, when varying the number of UAVs
-----
Fig. 7: 10 UAVs, L=10 Mbits, Pavg=1 kW, Λ=200
req/min: Average service delay (communication time
+ queue wait time) vs the number of channels NC.
Fig. 8: Policy convergence time (in hours) for
MAESTRO-X and the relevant state-of-the-art.
_NU_ . All implementations are in Python, and are executed on a compute node with 2× 64-core
AMD EPYC Milan 7763 CPUs, 16 64 GB DDR4 memory, and 4 NVIDIA A100 GPUs with
_×_ _×_
40 GB VRAM each. Remarkably, the convergence time of MAESTRO-X is irrespective of the
number of UAVs, whereas it grows quickly for CSCA-ADMM and DDQN. This is due to the
policy replication mechanism used by MAESTRO-X: the policy is computed for a single-agent,
and then replicated across the swarm, coupled with the supplementary UAV-swarm mechanisms
developed in Sec. V. On the other hand, the convergence times of CSCA-ADMM and DDQN
grow quickly with the number of UAVs, and become prohibitive when scaled to more than 5
and 6 UAVs, respectively: in fact, it grows linearly for CSCA-ADMM, due to a joint multi
UAV construction involved in its CVXPY-SCS implementation, and exponentially for DDQN,
due to combined multi-agent state and action space construction. Remarkably, MAESTRO
X yields a faster convergence time even for a single UAV, thanks to its ability to leverage
the multiscale structure of the decision process to achieve a more efficient implementation,
in addition to Tensor-ized executions exploiting SIMD processing in CUDA-capable GPUs,
and distributed workers and thread-pool concurrency in Python (TensorFlow). These benefits in
policy convergence coupled with the superior delay-power performance illustrated in Figs. 5 and
6, present MAESTRO-X as an appealing solution for both small and large UAV swarms.
VII. CONCLUSION
In this paper, we propose the MAESTRO-X framework for the decentralized orchestration
of rotary-wing UAV-relay swarms in cellular networks, augmenting the coverage and service
-----
capabilities of a terrestrial BS. First, we specialize our system to single-UAV deployments and
design the optimal scheduling and trajectory optimization policy under an SMDP formulation.
Next, we extend to distributed multi-UAV deployments by employing multi-agent coordination
mechanisms, and then replicate this augmented single-UAV policy across the swarm. Numerical
evaluations demonstrate that MAESTRO-X delivers significant gains over BS- and HAP-only
deployments; furthermore, it exhibits superior performance over static UAV deployments, deep
Q-learning schemes, and successive convex approximation strategies.
APPENDIX A: PROOF OF PROP. 1
Since 2(K+1) _g_ has a non-central χ[2] distribution with 2 degrees of freedom and a non_|_ _|[2]_
_√_ �
centrality parameter 2K, we find that Pout(Υ, β, K)=1−Q1( 2K, 2(K+1)u(Υ, β)), where
_√_ �
_Q1(·, ·) is the standard Marcum Q-function [14]. Hence, R(Υ, β, K) = Υ·Q1(_ 2K, 2(K + 1)u(Υ, β)).
We now maximize it over Υ≥0. Let Z≜2γ[−][1]u (Υ, β) and γ≜ _[N]βP[0][B]_ [, hence][ Υ=][B][ log][2] �1+ _[Z]2_ � ≜f (Z).
It follows that Υ[∗]=f (Z _[∗]), where Z_ _[∗]_ maximizes over Z 0 the function
_≥_
_√_
_h(Z) ≜_ ln R(f (Z), β, K) = ln f (Z) + ln Q1(
�
2K, _γ(K + 1)Z)._ (22)
� _√_ �
Note that Q1 _a,_ _bZ_ is log-concave in Z≥0 for a, b>0 (see [40]), and second derivative of
ln f (Z) satisfies (ln f (Z))[′′]= _[f]_ _[′′][(][Z][)]_
_f_ (Z) (f (Z))[2][ ≤][0][,][ ∀][Z][≥][0][, so that][ h][(][Z][)][ is concave in][ Z][≥][0][. Since]
_[−]_ [(][f] _[′][(][Z][))][2]_
limZ→0+ h(Z)=−∞ and limZ→∞ _h(Z)=−∞, there exists a unique Z_ _[∗]∈(0, ∞) (hence Υ[∗]=f (Z_ _[∗]))_
such that h[′](Z _[∗])=0, solvable with the bisection method, with h[′](Z) given by_
_√_
_∂Q1(_ 2K, b)/∂b��b=[√]γ(K+1)Z
_,_
_√_ �
_Q1(_ 2K, _γ(K + 1)Z)_
_h[′](Z) =_ _[f][ ′][(][Z][)]_
_f_ (Z) [+]
�
_γ(K + 1)_
_√_
2 _Z_
yielding (1) after solving for f _[′]_ and the partial derivative of Q1.
APPENDIX B: PROOF OF PROP. 2
Let _W[¯]_ _µ≜W[¯]_ _µ[(][s][)][+ ¯][W][ (]µ[bs][)]. If ξu=1, then additional requests received during the UAV relay phase_
are served directly by the BS, with delay _R¯GBL_ (r) [for a GN in position][ (][r, θ][)][. Thus, the expected]
average communication delay to serve these additional requests is E[∆[(]u,i[bs][)][]= ¯][D][BS][, yielding]
_W¯_ _µ= ¯Wµ[(][s][)][+ ¯][D]BS[( ¯][N]µ[−][1)][ and][ ¯][D]µ[=]_ _WN¯¯µµ_ [=] _W¯N¯µ[(]µ[s][)]_ [+] �1− _N¯[1]µ_ � _D¯_ _BS. Let µ be any policy (including_
the optimal one) that satisfies _D[¯]_ _µ≤D[¯]_ _BS: under such policy, since_ _N[¯]µ≥1, the expression above_
implies that _W[¯]_ _µ[(][s][)]_ _µ[≤][D][¯]_ _BS[. Moreover, since][ E][[][N]u[|][∆]u[(][s][)][]=∆]u[(][s][)][Λ][′][ and][ ξ]u[≤][1][, it follows that]_
_[≤][D][¯]_
_N¯µ≤1+Λ[′]W¯_ _µ[(][s][)]_ with equality if the UAV always serves requests. This implies (8).
-----
APPENDIX C: PROOF OF PROP. 3
Let πwait=1−πcomm be the SMDP steady-state probability of the UAV being in the waiting
state. Since the probability of remaining in the waiting state (no request is received) in one
SMDP step is pww=e[−][Λ][′][∆][0] and that of moving from a communication state to a waiting state is
_pcw=1, πcomm and πwait are solutions of the stationary equation πwait = πwaitpww + πcommpcw =_
_e[−][Λ][′][∆][0]πwait + πcomm. Solving it with πwait+πcomm=1 yields the expression of πcomm in Prop. 3._
REFERENCES
[1] B. Keshavamurthy and N. Michelusi, “Multiscale Adaptive Scheduling and Path-Planning for Power-Constrained
[UAV-Relays via SMDPs,” 2022. [Online]. Available: https://arxiv.org/abs/2209.07655](https://arxiv.org/abs/2209.07655)
[2] B. Keshavamurthy, “MAESTRO-X: Multiscale Adaptive Energy-conscious Scheduling and TRajectory Optimization
[(eXtended),” April 2022. [Online]. Available: https://github.com/bharathkeshavamurthy/MAESTRO-X.git](https://github.com/bharathkeshavamurthy/MAESTRO-X.git)
[3] A. Fotouhi, H. Qiang et al., “Survey on UAV Cellular Communications: Practical Aspects, Standardization Advancements,
Regulation, and Security Challenges,” IEEE Communications Surveys Tutorials, pp. 1–1, 2019.
[4] M. Mozaffari, W. Saad et al., “A Tutorial on UAVs for Wireless Networks: Applications, Challenges, and Open Problems,”
_IEEE Communications Surveys Tutorials, pp. 1–1, 2019._
[5] B. Keshavamurthy and N. Michelusi, “Learning-Based Spectrum Sensing and Access in Cognitive Radios via Approximate
POMDPs,” IEEE Transactions on Cognitive Communications and Networking, vol. 8, no. 2, pp. 514–528, 2022.
[6] D. Alvear, “Inside Verizon | Up to Speed | Drones, robots and the power of 5G,” July 2021.
[7] S. Guillette, “Verizon News Archives | Robots, drones and sensors are changing the way we farm,” March 2019.
[8] Y. Zeng, R. Zhang et al., “Wireless communications with unmanned aerial vehicles: opportunities and challenges,” IEEE
_Communications Magazine, vol. 54, no. 5, pp. 36–42, May 2016._
[9] Q. Wu, L. Liu et al., “Fundamental Trade-offs in Communication and Trajectory Design for UAV-Enabled Wireless
Network,” IEEE Wireless Communications, vol. 26, pp. 36–44, 02 2019.
[10] Y. Zeng, J. Xu et al., “Energy Minimization for Wireless Communication With Rotary-Wing UAV,” IEEE Transactions
_on Wireless Communications, vol. 18, no. 4, pp. 2329–2345, April 2019._
[11] M. A. Abd-Elmagid and H. S. Dhillon, “Average Peak Age-of-Information Minimization in UAV-Assisted IoT Networks,”
_IEEE Transactions on Vehicular Technology, vol. 68, no. 2, pp. 2003–2008, 2019._
[12] X. Hu, K.-K. Wong et al., “UAV-Assisted Relaying and Edge Computing: Scheduling and Trajectory Optimization,” IEEE
_Transactions on Wireless Communications, vol. 18, no. 10, pp. 4738–4752, 2019._
[13] J. Chen and D. Gesbert, “Optimal positioning of flying relays for wireless networks: A LOS map approach,” in 2017 IEEE
_International Conference on Communications (ICC), May 2017, pp. 1–6._
[14] C. You and R. Zhang, “3D Trajectory Optimization in Rician Fading for UAV-Enabled Data Harvesting,” IEEE Transactions
_on Wireless Communications, vol. 18, no. 6, pp. 3192–3207, 2019._
[15] R. K. Patra and P. Muthuchidambaranathan, “Optimisation of Spectrum and Energy Efficiency in UAV-Enabled Mobile
Relaying Using Bisection and PSO Method,” in 3rd Int. Conference for Convergence in Technology (I2CT), 2018, pp. 1–7.
[16] Q. Hu, Y. Cai et al., “Low-Complexity Joint Resource Allocation and Trajectory Design for UAV-Aided Relay Networks
With the Segmented Ray-Tracing Channel Model,” IEEE Transactions on Wireless Communications, vol. 19, no. 9, 2020.
[17] Q. Wu, Y. Zeng et al., “Joint Trajectory and Communication Design for Multi-UAV Enabled Wireless Networks,” IEEE
_Transactions on Wireless Communications, vol. 17, no. 3, pp. 2109–2121, 2018._
-----
[18] M. Mozaffari, W. Saad et al., “Efficient Deployment of Multiple Unmanned Aerial Vehicles for Optimal Wireless Coverage,”
_IEEE Communications Letters, vol. 20, no. 8, pp. 1647–1650, 2016._
[19] H. Bayerlein, M. Theile et al., “Multi-UAV Path Planning for Wireless Data Harvesting With Deep Reinforcement
Learning,” IEEE Open Journal of the Communications Society, vol. 2, p. 1171–1187, 2021.
[20] L. Wang, K. Wang et al., “Multi-Agent Deep Reinforcement Learning-Based Trajectory Planning for Multi-UAV Assisted
Mobile Edge Computing,” IEEE Transactions on Cognitive Communications and Networking, vol. 7, no. 1, 2021.
[21] A. M. Koushik, F. Hu et al., “Deep Q-Learning-Based Node Positioning for Throughput-Optimal Communications in
Dynamic UAV Swarm Network,” IEEE Transactions on Cognitive Communications and Networking, vol. 5, no. 3, 2019.
[22] Q. Zhang, M. Mozaffari et al., “Machine Learning for Predictive On-Demand Deployment of UAVs for Wireless
Communications,” in 2018 IEEE Global Communications Conference (GLOBECOM), 2018, pp. 1–6.
[23] A. E. A. A. Abdulla, Z. M. Fadlullah et al., “Toward Fair Maximization of Energy Efficiency in Multiple UAS-Aided
Networks: A Game-Theoretic Methodology,” IEEE Transactions on Wireless Communications, vol. 14, no. 1, Jan 2015.
[24] Y. Li and L. Cai, “UAV-Assisted Dynamic Coverage in a Heterogeneous Cellular System,” IEEE Network, vol. 31, no. 4,
pp. 56–61, July 2017.
[25] D. P. Bertsekas, Dynamic Programming and Optimal Control, Vol. II, 4th ed. Athena Scientific, 2007.
[26] R. Cheng and Y. Jin, “A Competitive Swarm Optimizer for Large Scale Optimization,” IEEE Transactions on Cybernetics,
vol. 45, no. 2, pp. 191–204, 2015.
[27] J. Hu, H. Zhang et al., “Reinforcement Learning for Decentralized Trajectory Design in Cellular UAV Networks With
Sense-and-Send Protocol,” IEEE Internet of Things Journal, vol. 6, no. 4, pp. 6177–6189, 2019.
[28] C. Zhou, H. He et al., “Deep RL-based Trajectory Planning for AoI Minimization in UAV-assisted IoT,” in 2019 11th
_International Conference on Wireless Communications and Signal Processing (WCSP), 2019, pp. 1–6._
[29] M. Clerc, Particle Swarm Optimization, ser. ISTE. London ; Newport Beach: ISTE, 2006.
[30] H. Shakhatreh, A. Khreishah et al., “Efficient 3D placement of a UAV using particle swarm optimization,” in 2017 8th
_International Conference on Information and Communication Systems (ICICS), 2017, pp. 258–263._
[31] Z. Yuheng, Z. Liyan et al., “3-D Deployment Optimization of UAVs Based on Particle Swarm Algorithm,” in 2019 IEEE
_19th International Conference on Communication Technology (ICCT), 2019, pp. 954–957._
[32] A. Al-Hourani, S. Kandeepan et al., “Optimal LAP Altitude for Maximum Coverage,” IEEE Wireless Communications
_Letters, vol. 3, no. 6, pp. 569–572, 2014._
[33] D. Tse and P. Viswanath, Fundamentals of Wireless Communication. Cambridge: Cambridge University Press, 2005.
[34] R. Essaadali and A. Kouki, “A new simple Unmanned Aerial Vehicle doppler effect RF reducing technique,” in MILCOM
_2016 - 2016 IEEE Military Communications Conference, 2016, pp. 1179–1183._
[35] J. D. Little and S. C. Graves, “Little’s Law,” in Building Intuition. Springer, 2008, pp. 81–100.
[36] S. Boyd, L. Xiao et al., “Subgradient methods,” Lecture notes, Stanford University, Autumn Quarter, pp. 2004–2005, 2003.
[37] Y. Zeng and R. Zhang, “Energy-Efficient UAV Communication With Trajectory Optimization,” IEEE Transactions on
_Wireless Communications, vol. 16, no. 6, pp. 3747–3760, June 2017._
[38] P. Yang, K. Tang et al., “Turning High-Dimensional Optimization Into Computationally Expensive Optimization,” IEEE
_Transactions on Evolutionary Computation, vol. 22, no. 1, pp. 143–156, 2018._
[39] A. Al-Hourani, S. Kandeepan et al., “Optimal LAP Altitude for Maximum Coverage,” IEEE Wireless Communications
_Letters, vol. 3, no. 6, pp. 569–572, 2014._
[40] Y. Sun and S. Zhou, “Tight bounds of the generalized marcum q-function based on log-concavity,” in IEEE GLOBECOM
_2008 - 2008 IEEE Global Telecommunications Conference, 2008, pp. 1–5._
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/2007.01228, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "GREEN",
"url": "http://arxiv.org/pdf/2007.01228"
}
| 2,020
|
[
"JournalArticle"
] | true
| 2020-07-02T00:00:00
|
[
{
"paperId": "b74155a6cfe3cc8303b6776cd5cff973f9fe3407",
"title": "Multiscale Adaptive Scheduling and Path-Planning for Power-Constrained UAV-Relays via SMDPs"
},
{
"paperId": "b56b3f96468f3ffe3f7142f92e4c01d7ab6c64f7",
"title": "Learning-Based Spectrum Sensing and Access in Cognitive Radios via Approximate POMDPs"
},
{
"paperId": "364ba551cdcc1f9d52ffb462c738fb2ee4152f78",
"title": "Multi-UAV Path Planning for Wireless Data Harvesting With Deep Reinforcement Learning"
},
{
"paperId": "383a2b6362b8fe6c8a922d193383f88fdf87bb44",
"title": "Multi-Agent Deep Reinforcement Learning-Based Trajectory Planning for Multi-UAV Assisted Mobile Edge Computing"
},
{
"paperId": "0bc6770fcb8446ef4532c6f557274b528eaaef0e",
"title": "Low-Complexity Joint Resource Allocation and Trajectory Design for UAV-Aided Relay Networks With the Segmented Ray-Tracing Channel Model"
},
{
"paperId": "074d5b17ab7bab1e94181cd0bfcad29b4f7f7b10",
"title": "Deep RL-based Trajectory Planning for AoI Minimization in UAV-assisted IoT"
},
{
"paperId": "1a9332b0dbcabd7a5535b59a0f3011fd318c2117",
"title": "3-D Deployment Optimization of UAVs Based on Particle Swarm Algorithm"
},
{
"paperId": "da4dc1f8bc2987d47f532b8c7298e4098d0b4f74",
"title": "Deep ${Q}$ -Learning-Based Node Positioning for Throughput-Optimal Communications in Dynamic UAV Swarm Network"
},
{
"paperId": "50bed0baf83b76345a346f4152dad0310127186c",
"title": "3D Trajectory Optimization in Rician Fading for UAV-Enabled Data Harvesting"
},
{
"paperId": "20c41464a232cb13204a26dfacd6e8ff2e87c446",
"title": "UAV-Assisted Relaying and Edge Computing: Scheduling and Trajectory Optimization"
},
{
"paperId": "013a051d2aa72e7595ec63d99f9b294196658430",
"title": "Reinforcement Learning for Decentralized Trajectory Design in Cellular UAV Networks With Sense-and-Send Protocol"
},
{
"paperId": "925f022fbaa8fa7e955beb1c1d825fd9e376d40f",
"title": "Survey on UAV Cellular Communications: Practical Aspects, Standardization Advancements, Regulation, and Security Challenges"
},
{
"paperId": "5fd570b617cd19b121fdf9da1367e5233d0aa29e",
"title": "Fundamental Trade-offs in Communication and Trajectory Design for UAV-Enabled Wireless Network"
},
{
"paperId": "88e31c9264896dcd38fa08a36b870cf6f1752de4",
"title": "Machine Learning for Predictive On-Demand Deployment of Uavs for Wireless Communications"
},
{
"paperId": "31af2e6f860e400711ca3237ff3ff97e2287823d",
"title": "Average Peak Age-of-Information Minimization in UAV-Assisted IoT Networks"
},
{
"paperId": "38019e745cf8d7dc0274882f5d8e485acdf8f001",
"title": "Energy Minimization for Wireless Communication With Rotary-Wing UAV"
},
{
"paperId": "5c351a8a822916e6581747fc17edff162a1a7ab0",
"title": "Optimisation of Spectrum and Energy Efficiency in UAV-Enabled Mobile Relaying Using Bisection and PSO Method"
},
{
"paperId": "0f5a0fbd07155cbf81ae9a5e76a1bc78da10a376",
"title": "A Tutorial on UAVs for Wireless Networks: Applications, Challenges, and Open Problems"
},
{
"paperId": "7660dc6e755e85fbb0b162491dd615b5245d08d5",
"title": "UAV-Assisted Dynamic Coverage in a Heterogeneous Cellular System"
},
{
"paperId": "ee01b802099fbc958a2b572671eff970faf28d57",
"title": "Optimal positioning of flying relays for wireless networks: A LOS map approach"
},
{
"paperId": "3b30b5b5cc861be94387905ad013ac9f4c427a98",
"title": "Joint Trajectory and Communication Design for Multi-UAV Enabled Wireless Networks"
},
{
"paperId": "d349eb54d9aa0b5ba7008334e46c2d526a1f86a6",
"title": "Efficient 3D placement of a UAV using particle swarm optimization"
},
{
"paperId": "a3ad8702915541a170a0f4ec7cd7203f8ad75d27",
"title": "A new simple Unmanned Aerial Vehicle doppler effect RF reducing technique"
},
{
"paperId": "774c755a1c2bda4e312547c64fe1531c1f3ce3b4",
"title": "Energy-Efficient UAV Communication With Trajectory Optimization"
},
{
"paperId": "adf581691aef09a0da06dae8bc87248f33ca348a",
"title": "Efficient Deployment of Multiple Unmanned Aerial Vehicles for Optimal Wireless Coverage"
},
{
"paperId": "3df9de1483091be50d2a686da3d43f39e706db31",
"title": "Turning High-Dimensional Optimization Into Computationally Expensive Optimization"
},
{
"paperId": "e022e57ec1877de91d9f129d31bbae4053d7983e",
"title": "Wireless communications with unmanned aerial vehicles: opportunities and challenges"
},
{
"paperId": "9cfad16564dca92ce99ee0afa061f886d638c698",
"title": "A Competitive Swarm Optimizer for Large Scale Optimization"
},
{
"paperId": "65753ff3e5ab97d0d6fd85c826cc0251c1eefae9",
"title": "Optimal LAP Altitude for Maximum Coverage"
},
{
"paperId": "81a49c1a7421e07afe51fc262a8ea6a6e28307a6",
"title": "Tight Bounds of the Generalized Marcum Q-Function Based on Log-Concavity"
},
{
"paperId": "fda8c1e5080a45613d9c74718ccc2b5ab86d5c00",
"title": "Fundamentals of Wireless Communication"
},
{
"paperId": "b0fa6699b932bf18c404f84bddc4b9fe8ab4a864",
"title": "Particle swarm optimization"
},
{
"paperId": null,
"title": "Inside Verizon | Up to Speed | Drones, robots and the power of 5G"
},
{
"paperId": null,
"title": "Robots, drones and sensors are changing the way we farm."
},
{
"paperId": "0769fdea3345e3c06459a95e39c9ea19ce70845f",
"title": "Toward Fair Maximization of Energy Efficiency in Multiple UAS-Aided Networks: A Game-Theoretic Methodology"
},
{
"paperId": "1dad542838a53b6fed74376db5381e60f95a3b22",
"title": "Little's Law"
},
{
"paperId": "bf184b2650c6261dbacad92d569779f4a840b23a",
"title": "Building intuition insights from basic operations management models and principles"
},
{
"paperId": null,
"title": "Dynamic Programming and Optimal Control, Vol. II, 4th ed"
},
{
"paperId": "25ea0939ecb7504ebb884539e3c4070e00f9b9da",
"title": "Subgradient Methods"
},
{
"paperId": null,
"title": "“ Drones , robots and the power of 5 G . ” Verizon News Center . Jul . 2021 . [ Online ] “ Robots , drones and sensors are changing the way we farm . ” Verizon News Center . Mar . 2019 . [ Online ]"
}
] | 30,200
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Mathematics",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Mathematics",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/02b1964d705661c5a16e0d87b152d14da1372dd8
|
[
"Computer Science",
"Mathematics"
] | 0.83962
|
Optimal Black-Box Secret Sharing over Arbitrary Abelian Groups
|
02b1964d705661c5a16e0d87b152d14da1372dd8
|
Annual International Cryptology Conference
|
[
{
"authorId": "145100790",
"name": "R. Cramer"
},
{
"authorId": "2266310",
"name": "S. Fehr"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"Int Cryptol Conf",
"Annu Int Cryptol Conf",
"CRYPTO",
"International Cryptology Conference"
],
"alternate_urls": null,
"id": "212b6868-c374-4ba2-ad32-19fde8004623",
"issn": null,
"name": "Annual International Cryptology Conference",
"type": "conference",
"url": "http://www.iacr.org/"
}
|
A black-box secret sharing scheme for the threshold access structure Tt,n is one which works over any finite Abelian group G. Briefly, such a scheme differs from an ordinary linear secret sharing scheme (over, say, a given finite field) in that distribution matrix and reconstruction vectors are defined over Z and are designed independently of the group G from which the secret and the shares are sampled. This means that perfect completeness and perfect privacy are guaranteed regardless of which group G is chosen. We define the black-box secret sharing problem as the problem of devising, for an arbitrary given Tt,n, a scheme with minimal expansion factor, i.e., where the length of the full vector of shares divided by the number of players n is minimal.Such schemes are relevant for instance in the context of distributed cryptosystems based on groups with secret or hard to compute group order. A recent example is secure general multi-party computation over black-box rings.In 1994 Desmedt and Frankel have proposed an elegant approach to the black-box secret sharing problem based in part on polynomial interpolation over cyclotomic number fields. For arbitrary given Tt,n with O < t < n - 1, the expansion factor of their scheme is O(n). This is the best previous general approach to the problem.Using certain low degree integral extensions of Z over which there exist pairs of sufficiently large Vandermonde matrices with co-prime determinants, we construct, for arbitrary given Tt,n with O < t < n - 1, a black-box secret sharing scheme with expansion factor O(log n), which we show is minimal.
|
# Optimal Black-Box Secret Sharing
over Arbitrary Abelian Groups
Ronald Cramer and Serge Fehr
BRICS[⋆], Department of Computer Science, Aarhus University, Denmark
_{cramer,fehr}@brics.dk_
**Abstract. A black-box secret sharing scheme for the threshold access**
structure Tt,n is one which works over any finite Abelian group G. Briefly,
such a scheme differs from an ordinary linear secret sharing scheme (over,
say, a given finite field) in that distribution matrix and reconstruction
vectors are defined over Z and are designed independently of the group
_G from which the secret and the shares are sampled. This means that_
perfect completeness and perfect privacy are guaranteed regardless of
which group G is chosen. We define the black-box secret sharing problem
as the problem of devising, for an arbitrary given Tt,n, a scheme with
minimal expansion factor, i.e., where the length of the full vector of
shares divided by the number of players n is minimal.
Such schemes are relevant for instance in the context of distributed cryptosystems based on groups with secret or hard to compute group order. A
recent example is secure general multi-party computation over black-box
rings.
In 1994 Desmedt and Frankel have proposed an elegant approach to
the black-box secret sharing problem based in part on polynomial interpolation over cyclotomic number fields. For arbitrary given Tt,n with
0 < t < n − 1, the expansion factor of their scheme is O(n). This is the
best previous general approach to the problem.
Using certain low degree integral extensions of Z over which there exist
pairs of sufficiently large Vandermonde matrices with co-prime determinants, we construct, for arbitrary given Tt,n with 0 < t < n − 1, a
black-box secret sharing scheme with expansion factor O(log n), which
we show is minimal.
## 1 Introduction
A black-box secret sharing scheme for the threshold access structure Tt,n is one
which works over any finite Abelian group G. Briefly, such a scheme differs from
an ordinary linear secret sharing scheme (over, say, a given finite field; see e.g.
[5,24,6,3,2,20,19,1,16,8]) in that distribution matrix and reconstruction vectors
are defined over Z and are designed independently of the group G from which
the secret and the shares may be sampled. In other words, the dealer computes
the shares for the n players as Z-linear combinations of the secret group element
_⋆_ Basic Research in Computer Science (www.brics.dk), funded by the Danish National
Research Foundation.
-----
of his interest and secret randomizing group elements, and reconstruction of the
secret from the shares held by a large enough set of players is by taking Zlinear combinations over those shares. Note that each player may receive one or
more group elements as his share in the secret. Perfect completeness and perfect
privacy are guaranteed regardless of which group G is chosen. Here, perfect
completeness means that the secret is uniquely determined by the joint shares of
at least t +1 players, and perfect privacy means that the joint shares of at most t
players contain no Shannon information at all about the secret of interest. Note
that these schemes are homomorphic in the sense that the sum of share vectors
is a share vector for the sum of the corresponding secrets.
We define the black-box secret sharing problem as the problem of devising,
for an arbitrary given Tt,n, a scheme with minimal expansion factor, i.e., where
the length of the full vector of shares divided by the number of players n is
minimized[1]. Note the case t = n 1 is easily solved by “additive n-out-of-n
_−_
sharing,” which has expansion factor 1. The cases t = 0, n have no meaning for
secret sharing. For the rest of this discussion we assume 0 < t < n 1.
_−_
The idea of black-box secret sharing was first considered by Desmedt and
Frankel [11] in the context of distributed cryptosystems based on groups with secret order. Shamir’s polynomial based secret sharing scheme over finite fields [24]
cannot immediately be adapted to the setting of black-box secret sharing. Later,
Desmedt and Frankel [12] showed a black-box secret sharing scheme that elegantly circumvents problems with polynomial interpolation over the integers by
passing to an integral extension ring of Z over which a sufficiently large invert_ible Vandermonde matrix exists. Their scheme is then constructed using the fact_
that (sufficiently many copies of) an arbitrary Abelian group can be viewed as
a module over such an extension ring.
For a given commutative ring R with 1, the largest integer l such that there
exists an invertible l _l Vandermonde matrix with entries in R is called the_
_×_
_Lenstra constant l(R) of the ring R. Equivalently, l(R) is the maximal size of a_
subset E of R that is “exceptional” in that for all α, α[′] _∈_ _E, α ̸= α[′], it holds_
that α _α[′]_ is a unit of R.
_−_
Given an integral extension ring R of degree m over Z, they construct a
black-box secret sharing scheme with expansion factor m for a threshold access
structure on at most l(R) 1 players. For any prime p, Lenstra’s constant for
_−_
the ring of integers of the pth cyclotomic number field is p [2]. Given an arbitrary
1 That minimal expansion is at most polynomial in n, even when appropriately gener
alizing the concept to encompass non-Abelian groups as well, is verified by combination of the technique of Benaloh-Leichter [2] with the classical result of complexity
theory that all monotone threshold functions are representable by poly-size monotone Boolean formulas. See also [10].
2 It is not hard to find an exceptional set of size p in this ring. To see that the maximal
size of such a set is p, let K be a number field of degree m, and let ZK denote its
ring of algebraic integers. For an arbitrary non-trivial ideal I of ZK, it is easy to
see that l(ZK ) ≤|ZK _/I| (≤_ 2[m]). In the case where K is the pth cyclotomic number
field, the integer prime p totally ramifies. Hence l(ZK ) ≤|ZK _/P_ _| = p, where P is_
the unique prime ideal of ZK lying above p.
-----
_Tt,n and choosing R as the ring of integers of the pth cyclotomic number field,_
where p is the smallest prime greater than n, they construct a black-box secret
sharing scheme for Tt,n with expansion factor between n and 2n. This is the
best previous general approach to the problem. Further progress on the blackbox secret sharing problem via the approach of [12] depends on the problem of
finding for each n an extension whose degree is substantially smaller than n and
whose Lenstra constant is greater than n. To the best of our knowledge, this is
an open problem of algebraic number theory (see also [12] and the references
therein).
Except for some quite special cases, namely when t is constant or when t
(resp. n _t) is small compared to n [14,4] or the constant factor gain from [15],_
_−_
no substantial improvement on the general black-box secret sharing problem has
been reported since.
The crucial difference with our approach to the black-box secret sharing
problem is that we avoid dependence on Lenstra’s constant altogether. Namely,
first, we observe that a sufficient condition for black-box secret sharing is the
existence (over an extension of Z) of a pair of sufficiently large Vandermonde
matrices with co-prime determinants. And, second, we show how to construct low
_degree integral extensions of Z satisfying this condition. For arbitrary given Tt,n,_
this leads to a black-box secret sharing scheme with expansion factor O(log n).
Using a result of Karchmer and Wigderson [20], we show that this is minimal.
There are several applications of black-box secret sharing. For instance, the
result of [12] is exploited in [13] to obtain an efficient and secure solution for
sharing any function out of a certain abstract class of functions, including RSA.
The interest in application of the result of [12] to practical distributed RSAbased protocols seems to have decreased somewhat due to recent developments,
see for instance [25] and the references therein. However, apart from the fact
that optimal black-box secret sharing is perhaps interesting in its own right,
we note that in [9] our black-box secret sharing scheme is applied in protocols
for secure general multi-party computation over black-box rings. Also, optimal
black-box secret sharing may very well be relevant to new distributed cryptographic schemes for instance based on class groups.
This paper is organized as follows. In Section 2 we give a formalization of the
notion of black-box secret sharing, and show a natural correspondence between
such schemes and our notion of integer span programs (ISPs). This generalizes
the well-known correspondence between monotone span programs over finite
fields [20] and linear secret sharing schemes over finite fields. In Section 3 we
show lower bounds on the size of ISPs computing threshold access structures.
Our main result is presented in Section 4, where we construct an ISP with
minimal size for an arbitrary given threshold access structure. This leads to
an optimal black-box secret sharing scheme for an arbitrary given threshold
access structure. At the end, we point out further combinatorial properties of
our scheme that are useful when exhibiting efficient simulators as required in
the security proofs of threshold crypto-systems such as threshold RSA.
-----
## 2 Black-Box Secret Sharing
**2.1** **Definitions**
**Definition 1. A monotone access structure on** 1, . . ., n _is a non-empty col-_
_{_ _}_
_lection Γ of sets A_ 1, . . ., n _such that_ _Γ and such that for all A_ _Γ_
_⊂{_ _}_ _∅̸∈_ _∈_
_and for all sets B with A_ _B_ 1, . . ., n _it holds that B_ _Γ_ _._
_⊂_ _⊂{_ _}_ _∈_
**Definition 2. Let t and n be integers with 0 < t < n. The threshold access**
structure Tt,n is the collection of sets A ⊂{1, . . ., n} with |A| > t [3].
Let Γ be a monotone access structure on {1, . . ., n}. Let M ∈ Z[d,e] be an
integer matrix, and let ψ : 1, . . ., d 1, . . ., n be a surjective function.
_{_ _} →{_ _}_
We say that the jth row (j = 1 . . . d) of M is labelled by ψ(j) or that “ψ(j)
owns the jth row.” For A ⊂{1, . . ., n}, MA denotes the restriction of M to the
rows jointly owned by A. Write dA for the number of rows in MA. Similarly, for
**x ∈** Z[d], xA ∈ Z[d][A] denotes the restriction of x to the coordinates jointly owned
by A. For each A ∈ _Γ_, let λ(A) ∈ Z[d][A] be an integer (column-) vector. We call
this the reconstruction vector for A. Collect all these vectors in a set .
_R_
**Definition 3. Let Γ be a monotone access structure on** 1, . . ., n _, and let_ =
_{_ _}_ _B_
(M, ψ, ) be as defined above. _is called an integer Γ_ -scheme. Its expansion
_R_ _B_
rate is defined as d/n, where d is the number of rows of M _._
Let G be a finite Abelian group. We use additive notation for its group
operation, and use 0G to denote its neutral element. The group G is of course a
Z-module (see e.g. [23]), by defining the map Z × G → _G, (µ, g) �→_ _µ · g, where_
0 · g = 0G, µ · g = g + . . . + g (µ times) for µ > 0 and µ · g = −((−µ) · g) for
_µ < 0_ [4]. We also write µg or gµ instead of µ _g. Note that it is well-defined how_
_·_
an integer matrix acts on a vector of group elements.
**Definition 4. Let Γ be a monotone access structure on** 1, . . ., n _and let_ =
_{_ _}_ _B_
(M, ψ, ) be an integer Γ _-scheme. Then_ _is a black-box secret sharing scheme_
_R_ _B_
_for Γ if the following holds. Let G be an arbitrary finite Abelian group G, and let_
_A_ 1, . . ., n _be an arbitrary non-empty set. For arbitrarily distributed s_ _G,_
_⊂{_ _}_ _∈_
_let g = (g1, . . ., ge)[T]_ _∈_ _G[e]_ _be drawn uniformly at random, subject to g1 = s._
_Define s = M_ **g. Then:**
**– (Completeness) If A ∈** _Γ_ _, then s[T]A_ _[·][ λ][(][A][) =][ s][ with probability 1, where]_
**_λ(A)_** _is the reconstruction vector for A._
_∈R_
**– (Privacy) If A ̸∈** _Γ_ _, then sA contains no Shannon information on s._
3 Note that some authors define Tt,n as consisting of all sets of size at least t. Our
definition is consistent with a convention in the multi-party computation literature.
4 If the group operation in G is efficient, multiplication by an integer can also be
efficiently implemented using standard “double-and-add.”
-----
Note that these schemes[5] are homomorphic in the sense that the sum s + s[′]
of two share vectors s and s[′], is a share vector for the sum s + s[′] of their
corresponding secrets s and s[′].
**2.2** **Monotone Span Programs over Rings**
In this section we provide quite natural necessary and sufficient conditions under
which an integer Γ -scheme is a black-box secret sharing scheme for Γ . To this
end, we introduce the notion of monotone span programs over rings. This is a
certain variation of monotone span programs over finite fields, introduced by
Karchmer and Wigderson [20]. These are well-known to have a natural one-toone correspondence with linear secret sharing schemes over finite fields (see e.g.
[19,1]). Monotone span programs over Z (ISPs) will turn out to have a similar
correspondence with black-box secret sharing schemes. We also show an efficient
conversion of a monotone span program over an integral extension ring of Z to
an ISP.
As an aside, monotone span programs over rings are the basis for multi-party
computation over black-box rings, as studied in [9]. In particular, the techniques
of [8] for secure multiplication and VSS apply to this flavor of monotone span
program as well.
Throughout this paper, R denotes a (not necessarily finite) commutative ring
with 1. Let Γ be a monotone access structure on 1, . . ., n, and let M _R[d,e]_
_{_ _}_ _∈_
be a matrix whose d rows are labelled by a surjective function ψ : 1, . . ., d
_{_ _} →_
1, . . ., n .
_{_ _}_
**Definition 5. ε = (1, 0, . . ., 0)[T]** _R[e]_ _is called the target vector. Furthermore,_
_∈_
_M = (R, M, ψ, ε) is called a monotone span program (over the ring R). If R = Z,_
_it is called an integer span program, or ISP, for short. We define size(_ ) = d,
_M_
_where d is the number of rows of M_ _._
For N _R[a,b], imN denotes its column space, i.e., the space of all vectors_
_∈_
_N_ **x** _R[a], where x ranges over R[b], and kerN denotes its null-space, i.e., the_
_∈_
space of all vectors x _R[b]_ with N **x = 0** _R[a]._
_∈_ _∈_
**Definition 6. As above, let Γ be a monotone access structure and let** =
_M_
(R, M, ψ, ε) be a monotone span program over R. Then _is a monotone span_
_M_
_program for Γ_ _, if for all A_ 1, . . ., n _the following holds._
_⊂{_ _}_
**– If A ∈** _Γ_ _, then ε ∈_ imMA[T] _[.]_
**– If A ̸∈** _Γ_ _, then there exists κ = (κ1, . . ., κe)[T]_ _∈_ kerMA with κ1 = 1.
_We also say that_ computes Γ _._
_M_
5 See [21] for an equivalent definition. We also note that only requiring reconstruction
to be linear, as some authors do, results in an equivalent definition of black-box
secret sharing. This is an easily proved consequence of Lemma 2, but we omit the
details here.
-----
If R is a field, our definition is equivalent to the computational model of
monotone span programs over fields [20]. Indeed, this model is characterized by
the condition that A ∈ _Γ if and only if ε ∈_ imMA[T] [. The equivalence follows from]
the remark below.
_Remark 1. By basic linear algebra, if R is a field, then ε ̸∈_ imMA[T] [implies that]
there exists κ ∈ kerMA with κ1 = 1. If R is not a field this does not necessarily
hold[6]. The implication in the other direction trivially holds regardless of R.
Using (generally inefficient) representations of monotone access structures as
monotone Boolean formulas and using induction in a similar style as in e.g. [2],
it is straightforward to verify that for all Γ and for all R, there is a monotone
span program over R that computes Γ .
**Definition 7. For any Γ and for any R, mspR(Γ** ) denotes the minimal size of
_a monotone span program over R computing Γ_ _. If R = Z, we write isp(Γ_ ).
Define a non-degenerate monotone span program as one for which the rows
of M span the target-vector. As opposed to the case of fields, a non-degenerate
monotone span program over a ring need not compute any monotone access
structure. This is of no concern here, though.
The following proposition characterizes black-box secret sharing schemes in
terms of ISPs.
**Proposition 1. Let Γ be a monotone access structure on** 1, . . ., n _, and let_
_{_ _}_
= (M, ψ, ) be an integer Γ _-scheme. Then_ _is a black-box secret sharing_
_B_ _R_ _B_
_scheme for Γ if and only if M = (Z, M, ψ, ε) is an ISP for Γ and for all A ∈_ _Γ_ _,_
_its reconstruction vector λ(A) ∈R satisfies MA[T]_ **_[λ][(][A][) =][ ε][.]_**
_Proof. The argument that the stated ISP is sufficient for black-box secret sharing_
is quite similar to the well-known case of linear secret sharing over finite fields.
The other direction of the implication follows in essence from Lemma 1 below.
We include full details for convenience.
Consider the ISP from the statement of the proposition, together with the as
sumption on the reconstruction vectors. Consider an arbitrary set A 1, . . ., n
_⊂{_ _}_
and an arbitrary finite Abelian group G. Define s = M **g for arbitrary g =**
(s, g2, . . ., ge)[T] _∈_ _G[e]. Suppose A ∈_ _Γ_, and let λ(A) ∈R be its reconstruction
vector. It follows that s[T]A[λ][(][A][) = (][M][A][g][)][T][ λ][(][A][) =][ g][T][ (][M][ T]A **_[λ][(][A][)) =][ g][T][ ε][ =][ s][.]_**
Thus the completeness condition from Definition 4 is satisfied. If A _Γ_, then
_̸∈_
there exists κ ∈ Z[e] with MAκ = 0 ∈ Z[d][A] and κ1 = 1, by Definition 6. For
arbitrary s[′] _∈_ _G, define s[′]_ = M (g + (s[′] _−_ _s)κ) ∈_ _G[d][A]_ . The secret defined by s[′]
equals s[′], while on the other hand s[′]A [=][ s][A][. This implies perfect privacy: the]
assignment g[′] = g + (s[′] _s)κ provides a bijection between the set of possible_
_−_
vectors of “random coins” consistent with sA and s, and the set of those consistent with sA and s[′]. Therefore, the privacy condition from Definition 4 is also
satisfied.
6 Consider for example the integer matrix M = (2 0).
-----
In the other direction of the proposition, we start with a black-box se
cret sharing scheme for Γ according to Definition 4. Consider an arbitrary set
_A_ 1, . . ., n . Suppose A _Γ_, and let λ(A) be its reconstruction vec_⊂{_ _}_ _∈_ _∈R_
tor. For an arbitrary prime p, set G = Zp. By the completeness condition from
Definition 4, it follows that (1, 0, . . ., 0)[T] _≡_ (MAIe)[T] **_λ(A) ≡_** _MA[T]_ **_[λ][(][A][) mod][ p][,]_**
where Ie ∈ Z[e,e]p is the identity matrix. This holds for all primes p. Hence,
_MA[T]_ **_[λ][(][A][) = (1][,][ 0][, . . .,][ 0)][T][ =][ ε][. Therefore, the condition on the sets][ A][ ∈]_** _[Γ][ in]_
Definition 6 and the condition on the reconstruction vectors from the state_R_
ment of the proposition are satisfied.
To conclude the proof we show that the privacy condition from Definition 4
implies the condition on the sets A _Γ from Definition 6. The following formu-_
_̸∈_
lation is equivalent. Let y ∈ Z[d][A] denote the left-most column of MA, and let
_NA ∈_ Z[d][A][,e][−][1] denote the remaining e − 1 columns. Then it is to be shown that
the linear system of equations NAx = y is solvable over Z.
By Lemma 1 below, it is sufficient to show that this holds modulo m, for all
_m ∈_ Z, m ̸= 0. With notation as in Definition 4 and considering G = Zm, it
follows from the privacy condition that there exists g[′] _∈_ Z[e]m [such that][ g]1[′] _[≡]_ _[s]_ _[−]_ [1]
and sA ≡ _MAg[′]. Setting κ ≡_ **g −** **g[′]** _∈_ Z[e]m[, we have][ M][A][κ][ ≡] **[0][ with][ κ][1]** _[≡]_ [1. In]
other words, NAx = y is solvable over Zm for all integers m ̸= 0. _⊓⊔_
We note that [21] also gives a characterization. Although there are some
similarities in the technical analysis, the conditions stated there are still in terms
of the black-box secret sharing scheme, rather than by providing simple algebraic
conditions on the matrix M as we do. Therefore, we feel that our approach based
on integer span programs is perhaps more useful and insightful, especially since
monotone span programs over finite fields have since long been known to be
equivalent to linear secret sharing schemes over finite fields.
**Lemma 1. Let N ∈** Z[a,b] _and y ∈_ Z[a]. Then the linear system of equations
_N_ **x = y is solvable over Z if and only if it is solvable over Zm for all integers**
_m_ = 0.
_̸_
_Proof. The forward direction of the proposition is trivial. In the other direction,_
consider the Z-module H generated by the columns of N . By basic theory of
Z-modules (see e.g. [23]), there exists a Z-basis B = (b1, . . ., ba) of Z[a], and
non-zero integers a1, . . ., al such that BH = (a1b1, . . ., albl) is a Z-basis of H.
Let L denote the Z-module with basis BL = (b1, . . ., bl). Note that H ⊂ _L._
Let p be an arbitrary prime, and let ( ) denote reduction modulo p. Since the
_·_
determinant of B is ±1, B (resp. BL) provides a basis for the vector-space F[a]p
(resp. the vector-space L). Note that BL ⊂B.
It follows from the assumptions that y ∈ _H ⊂_ _L. Let (y1, . . ., ya) ∈_ Z[a] denote
the coordinates of y wrt. . Since the latter observation holds for all primes p,
_B_
it follows that yl+1 = . . . = ya = 0. Hence, y ∈ _L. Now set ˆm =_ [�]i[l]=1 _[a][i][. By]_
the assumptions, there exists c ˆm ∈ Z[a] such that y + ˆm · c ˆm ∈ _H. Therefore,_
_mˆ_ _· c ˆm ∈_ _L, and by the definition of L, c ˆm ∈_ _L. By the choice of ˆm, it follows_
that ˆm · c ˆm ∈ _H. We conclude that y ∈_ _H, as desired._ _⊓⊔_
-----
_Remark 2. Let_ = (R, M, ψ, ε) compute Γ . If R is a field or a principal ideal
_M_
domain (such as Z), then we may assume without loss of generality that e ≤ _d,_
i.e., there are at most as many columns in M as there are rows.
This is easily shown using elementary linear algebra, and using the basic
properties of modules over principal ideal domains (see e.g. [23] and the proof of
Lemma 1). Briefly, since is non-degenerate, the last statement in Remark 1
_M_
implies that the space generated by the 2nd up to the eth column of M does
not contain even a non-zero multiple of the first column. Without changing the
access structure that is computed, we can always replace the 2nd up to the eth
column of M by any set of vectors that generates the same space. If R is a field
or a principal ideal domain, this space has a basis of cardinality at most d 1.
_−_
_Remark 3. We may now identify a black-box secret sharing scheme for Γ with_
an ISP M = (Z, M, ψ, ε) for Γ . A reconstruction vector for A ∈ _Γ is simply_
any vector λ(A) ∈ Z[d][A] such that MA[T] **_[λ][(][A][) =][ ε][. Note that the expansion rate]_**
of the corresponding black-box secret sharing scheme is equal to size( )/n. By
_M_
Remark 2 it uses at most size( ) random group elements.
_M_
We now state some lemmas that are useful in the sequel.
**Definition 8. The dual Γ** _[∗]_ _of a monotone access structure Γ on {1, . . ., n} is_
_the collection of sets A_ 1, . . ., n _such that A[c]_ _Γ_ _._
_⊂{_ _}_ _̸∈_
Note that Γ _[∗]_ is a monotone access structure on {1, . . ., n}, that (Γ _[∗])[∗]_ = Γ, and
that (Tt,n)[∗] = Tn−t−1,n. The lemma below generalizes a similar property shown
in [20] for the case of fields.
**Lemma 2. mspR(Γ** ) = mspR(Γ _[∗]), for all R and Γ_ _._
_Proof. Let_ = (R, M, ψ, ε) be a monotone span program for Γ . Select an
_M_
arbitrary generating set of vectors b1, . . ., bl for kerM _[T]_, and choose λ with
_M_ _[T]_ **_λ = ε. Let M_** _[∗]_ be the matrix defined by the l +1 columns (λ, b1, b2, . . ., bl),
and use ψ to label M _[∗]_ as well. Define M[∗] = (R, M _[∗], ψ, ε[∗]), where ε[∗]_ =
(1, 0, . . ., 0)[T] _R[l][+1]. Note that size(_ ) = size( ). We claim that com_∈_ _M[∗]_ _M_ _M[∗]_
putes Γ _[∗]. This is easy to verify._
If A[c] _̸∈_ _Γ_, then by Definition 6, there exists κ ∈ _R[l][+1]_ such that MAc **_κ = 0_**
and κ1 = 1. Define λ[∗] = MAκ. Then (M _[∗])[T]A[λ][∗]_ [= ((][M][ ∗][)][T][ ·][ M] [)][κ][ =][ ε][∗][. On the]
other hand, if A[c] _∈_ _Γ_, then there exists λ[ˆ] ∈ _R[d]_ such that M _[T][ ˆ]λ = ε and λ[ˆ]_ _A = 0._
By definition of M _[∗], there exists κ ∈_ _R[l][+1]_ such that M _[∗]κ = λ[ˆ] and κ1 = 1._
Hence, MA[∗] **_[κ][ = ˆ][λ][A][ =][ 0][ and][ κ][1][ = 1. This concludes the proof.]_** _⊓⊔_
The lemma below holds in a more general setting, but we tailor it to ours.
**Lemma 3. Let f** (X) ∈ Z[X] be a monic, irreducible polynomial. Write m =
deg(f ). Consider the ring R = Z[X]/(f (X)). Suppose M = (R, M, ψ, ε) is a
_monotone span program over R for a monotone access structure Γ_ _. Then there_
_exists an ISP_ _Mˆ_ = (Z, _M,[ˆ]_ _ψ,[ˆ]_ ˆε) for Γ with size( M[ˆ] ) = m · size(M).
-----
_Proof. The proof is based on a standard algebraic technique for representing a_
linear map defined over an extension ring in terms of a linear map defined over
the ground ring. This technique is also used in [20] for monotone span programs
over extension fields. Since our definition of monotone span programs over rings
differs slightly from the definitions in [20], we explain it in detail.
Note that R is a commutative ring with 1 and that it has no zero divisors, but
that it is not a field. Fix w _R such that f_ (w) = 0 (such as w = X, the class of
_∈_
_X modulo f_ (X)). Then for each x _R, there exists a unique coordinate-vector_
_∈_
_→x_ = (x0, . . ., xm−1)T ∈ Zm such that x = x0 · 1 + x1 · w + · · · + xm−1 · wm−1. In
other words, W = {1, w, . . ., w[m][−][1]} is a basis for R when viewed as a Z-module.
For each x ∈ _R there exists a matrix in Z[m,m], denoted as [x], such that, for_
all y _R, [x]→y_ = _xy (the coordinate vector of xy_ _R). The columns of [x] are_
_∈_ _−→_ _∈_
simply the coordinate vectors of x, x · w, . . ., x · w[m][−][1]. If x ∈ Z, then [x] is a
diagonal matrix with x’s on its main diagonal. Furthermore, for all x, y _R, we_
_∈_
have the identities [x + y] = [x] + [y] and [xy] = [x][y].
Consider the monotone span program = (R, M, ψ, ε) from the statement
_M_
of the lemma. As before, write d (resp. e) for the number of rows (resp. columns)
of M . We define the ISP _Mˆ_ = (Z, _M,[ˆ]_ _ψ,[ˆ]_ ˆε) as follows. Construct M[ˆ] ∈ Z[md,me]
from M by replacing each entry x _R in M by the matrix [x]. The labeling ψ_
_∈_
is extended to ψ[ˆ] in the obvious way, i.e., if a player owns a certain row in M,
then that same player owns the m rows that it is substituted with in M[ˆ] . The
target vector ˆε is defined by ˆε = (1, 0 . . ., 0)[T] _∈_ Z[me].
We verify that [ˆ] is an ISP for Γ . First, consider a set A _Γ_ . By definition,
_M_ _∈_
there exists a vector λ = (λ1, . . ., λdA )[T] _∈_ _R[d][A]_ such that MA[T] **_[λ][ =][ ε][. Using the]_**
identities stated above and carrying out matrix multiplication “block-wise,” it
follows that M[ˆ] _A[T]_ [([][λ][1][]][, . . .,][ [][λ][d]A [])][T][ = ([1]][,][ [0]][, . . .,][ [0])][T][ . Define ˆ][λ][ ∈] [Z][md][A][ as the]
first column of the matrix ([λ1], . . ., [λdA ])[T] . Then M[ˆ] _A[T]_ **_[λ][ˆ][ = ˆ][ε][. Now consider a]_**
set A ̸∈ _Γ_ . By definition, there exists κ = (κ1, κ2, . . ., κe)[T] _∈_ _R[e]_ such that
_κ1 = 1 and MAκ = 0 ∈_ _R[d][A]_ . Using similar reasoning as above, it follows that
_Mˆ_ _A([κ1][T]_ _, . . ., [κe][T]_ )[T] = ([0][T] _, . . ., [0][T]_ )[T] . Define ˆκ ∈ Z[me] as the first column
of the matrix derived from κ in the above equation. Then, the first m entries of
**_κˆ are 1, 0, . . ., 0 (since κ1 = 1) and M[ˆ]_** _Aκˆ = 0 ∈_ Z[d][A] .
This proves the lemma. As an aside, it follows directly from the analysis
above that we may delete the 2nd up to mth leftmost colums of _Mˆ and the_
corresponding coordinates of ˆε without changing the access structure computed.
Hence, 1 + m(e 1) columns suffice, rather than me.
_−_ _⊓⊔_
## 3 Lower Bounds for the Threshold Case
**Proposition 2. For all integers t, n with 0 < t < n** _−_ 1, isp(Tt,n) = Ω(n _·_ log n).
_Hence, the expansion factor of a black-box secret sharing scheme for Tt,n with_
0 < t < n 1 is Ω(log n).
_−_
-----
Proposition 2 follows quite directly from the bound shown in Theorem 1 for
binary monotone span programs, as proved in [20][7]. Before we give the details
of the proof of Proposition 2, we include a proof of their bound for convenience.
Note that we have made constants for their asymptotic bound explicit.
Throughout this section, K denotes a field. Let = (K, M, ψ, ε) be a non_M_
degenerate monotone span program. The access structure of, denoted Γ ( ),
_M_ _M_
is the collection of sets A such that ε ∈ imMA[T] [. Note that by Remark 1 this is]
consistent with our Definition 6. We write msp2(Γ ) instead of mspF2 (Γ ).
**Proposition 3. [20] msp2(T1,n) ≥** _n · log n._
_Proof. Consider a monotone span program M = (F2, M, ψ, ε) such that Γ_ (M) =
_T1,n. Define e as the number of columns of M_, d as its number of rows, and di
as the number of rows of Mi for i = 1 . . . n, where we write Mi instead of M{i}
and di instead of d{i}. Without loss of generality, assume that the rows of each
_Mi are linearly independent over F2. Let H1 collect the vectors in F[e]2_ [with first]
coordinate equal to 1. Since {i} ̸∈ _T1,n, Remark 1 implies that |kerMi ∩_ _H1| ̸= ∅._
By assumption on Mi, |kerMi _∩H1| = 2[e][−][1][−][d][i]_ for i = 1 . . . n. On the other hand,
_{i, j} ∈_ _T1,n. Hence, by Remark 1, we have kerMi ∩_ kerMj ∩ _H1 = ∅, for all i, j_
with 1 _i < j_ _n. By counting and normalizing, 2[−][d][1]_ + + 2[−][d][n] 1. By the
_≤_ _≤_ _· · ·_ _≤_
Log Sum Inequality (see e.g. [7]), d = d1 + · · · + dn ≥ _n log n._ _⊓⊔_
**Theorem 1. [20] n · (⌊log n⌋** + 1) ≥ msp2(Tt,n) ≥ _n · log_ _[n][+3]2_ _[, for all][ t, n][ with]_
0 < t < n 1.
_−_
_Proof. The upper bound, which is not needed for our purposes, follows by con-_
sidering an appropriate Vandermonde matrix over the field F2u, where u =
(⌊log n⌋ +1). This is turned into a binary monotone span program for Tt,n using
a similar conversion technique as in Lemma 3.
For the lower bound, note that we may assume that t (n 1)/2, since
_≥_ _−_
msp2(Tt,n) = msp2(Tn−t−1,n) by Lemma 2. We have the following estimates.
_n_ _n_
msp2(Tt,n) ≥
_t + 2_ _t + 2_
_[·][ msp][2][(][T][t,t][+2][) =]_ _[·][ msp][2][(][T][1][,t][+2][)]_
_n_
_._
_≥_
_t + 2_ 2
_[·][ (][t][ + 2)][ ·][ log(][t][ + 2)][ ≥]_ _[n][ ·][ log][ n][ + 3]_
The first inequality is argued as follows. Consider an arbitrary monotone span
program M = (F2, M, ψ, ε) for Tt,n. Assume without loss of generality that the
number of rows in Mi is at most the number of rows in Mi+1, i = 1, . . ., n − 1.
The first t + 2 blocks M1, . . ., Mt+2 clearly form a monotone span program for
_Tt,t+2. Hence, the total number of rows in these blocks is at least msp2(Tt,t+2)._
Each other block Mj with j > t + 2 has at least as many rows as any of the first
_t + 2 blocks. Therefore, Mj has at least msp2(Tt,t+2)/(t + 2) rows. Summing up_
over all i according to the observations above gives the first inequality.
7 Note that isp(Tn−1,n) = n: the case t = n−1 is solved by simple additive “n-out-of-n
secret sharing.”
-----
The equality is implied by Lemma 2, the second to last inequality follows
from Proposition 3, and the last one from t (n 1)/2.
_≥_ _−_ _⊓⊔_
For the proof of Proposition 2, let an ISP for Tt,n be given, and consider the
ISP matrix, but with all entries reduced modulo 2. By our ISP definition and by
arguing the cases A ̸∈ _Tt,n using Remark 1, it follows that a binary monotone_
span program for Tt,n is obtained in this way. The argument is concluded by
applying Theorem 1[8]. The statement about black-box secret sharing follows
from Proposition 1.
Note that our lower bound on black-box secret sharing can also be appre
ciated without reference to Proposition 1, by essentially the same argument as
above. Namely, setting G = Z2 in Definition 4, we clearly obtain a (binary)
linear secret sharing scheme. This is well-known to be equivalent to a binary
monotone span program, as mentioned before. Hence, we can directly apply the
bound from Theorem 1.
## 4 Optimal Black-Box Threshold Secret Sharing
**Theorem 2. For all integers t, n with 0 < t < n −** 1, isp(Tt,n) = Θ(n · log n).
_Hence, there exists a black-box secret sharing scheme for Tt,n with expansion_
_factor O(log n), which is minimal._
_Proof. By Proposition 1 it is sufficient to focus on the claim about the ISPs._
The lower bound follows from Proposition 2. For the upper bound, we consider
rings of the form R = Z[X]/(f (X)), where f (X) ∈ Z[X] is a monic, irreducible
polynomial. Write m = deg(f ), the degree of R over Z.
On account of Lemma 3, it is sufficient to exhibit a ring R together with
a monotone span program M over R for Tt,n such that m = O(log n) and
size( ) = O(n).
_M_
The proof is organized as follows. We first identify a certain technical property
of a ring R that facilitates the construction of a monotone span program over
_R for Tt,n, with size O(n). We finalize the proof by constructing a ring R that_
enjoys this technical property, and that has degree O(log n) over Z.
For x1, . . ., xn ∈ _R, define_
_n_
� �
_∆(x1, . . ., xn) =_ _xi ·_ (xi − _xj)._
_i=1_ 1≤j<i≤n
Assume, for the moment, that there exist α1, . . ., αn ∈ _R and r0, r1 ∈_ _R such_
that
_r0 · ∆(1, . . ., n)[2]_ + r1 · ∆(α1, . . ., αn)[2] = 1.
This assumption implies the existence of a monotone span program over R
for Tt,n with size 2n, as we now show. Define
_∆0 = ∆(1, . . ., n) ∈_ Z, and _∆1 = ∆(α1, . . ., αn) ∈_ _R._
8 See [21,22] for lower bounds on the randomness required in black-box secret sharing
schemes.
-----
Let N0 ∈ _R[n,t][+1]_ (resp. N1 ∈ _R[n,t][+1]) be the matrix in which the i-th row is_
equal to (∆0, i, i[2], . . ., i[t]) (resp. (∆1, αi, αi[2][, . . ., α]i[t][)),][ i][ = 1][ . . . n][. In both cases,]
the ith row is labelled by i. When studied as possible monotone span programs
over R for Tt,n, N0 (resp. N1) satisfies Definition 6 for the sets A ̸∈ _Tt,n. On the_
other hand, in both cases, the rows owned by a set A ∈ _Tt,n do not necessarily_
span the target vector (1, 0, . . ., 0) _R[t][+1]. However, these rows do span[9]_ the
_∈_
vector (∆[2]0[,][ 0][, . . .,][ 0)][ ∈] _[R][t][+1][ (resp. (][∆][2]1[,][ 0][, . . .,][ 0)][ ∈]_ _[R][t][+1][). Both properties stated]_
can be verified immediately, for instance using the well-known expression for a
Vandermonde determinant in combination with Cram´er’s rule (see e.g. [23]);
passing to the fraction field K of R (note that R has no zero-divisors), this rule
implies that a c _c linear system of equations N_ **x = y over the ring R, has a**
_×_
solution at least in case where y det(N ) _R[c]. Another way is by using Lagrange_
_∈_ _·_
Interpolation over K, and clearing denominators.
Define a new monotone span program matrix M _R[2][n,][2][t][+1]_ consisting of all
_∈_
pairs of rows
(∆0, i, i[2], . . ., i[t], 0, . . ., 0), and (∆1, 0, . . ., 0, αi, αi[2][, . . ., α]i[t][)][,]
for i = 1 . . . n. The shown padding consists of t zeroes in both cases, and each
of the rows in a pair is labelled by i. Define ε = (1, 0, . . ., 0)[T] _R[2][t][+1]. The_
_∈_
sets A ̸∈ _Tt,n clearly satisfy Definition 6, and this time the rows owned by sets_
_A ∈_ _Tt,n span the target vector: they span in particular all vectors of the form_
(r · ∆[2]0 [+][ s][ ·][ ∆]1[2][,][ 0][, . . .,][ 0), with][ r, s][ ∈] _[R][. By setting][ r][ =][ r][0]_ [and][ s][ =][ r][1][, these]
include the target vector ε.
To conclude, we exhibit a ring R with degree O(log n) over the integers and
_α1, . . ., αn, r0, r1 ∈_ _R with r0 · ∆[2]0_ [+][ r][1] _[·][ ∆]1[2]_ [= 1, where][ ∆][0] [=][ ∆][(1][, . . ., n][) and]
_∆1 = ∆(α1, . . ., αn)._
These conditions are reformulated as follows. Let Πn denote the set of integer
primes p with 2 _p_ _n and define_
_≤_ _≤_
_Qn =_
�
_p∈Πn_
_p ∈_ Z.
Then we are looking for a ring R with degree O(log n) over the integers and
_α1, . . ., αn ∈_ _R such that_
_∆1 ∈_ (R/(Qn))[∗],
i.e., the residue-class of ∆1 in the ring R/(Qn) is a unit.
Indeed, if ∆1 ∈ (R/(Qn))[∗], then ∆1 ∈ (R/(Q[k]n[))][∗] [as well, for any positive]
integer k. To verify this by induction, suppose that ∆1 · v = 1 + w · Q[i]n [for some]
_v, w ∈_ _R and i ≥_ 1: then ∆1 · (v − _vw · Q[i]n[) = 1][ −]_ _[w][2][ ·][ Q][2]n[i]_ [and 2][i][ ≥] _[i][ + 1. As]_
a consequence, ∆1 ∈ (R/(∆[2]0[))][∗][. Namely, as an integer,][ ∆][2]0 [factors completely]
over the primes p ∈ _Πn. Then choose k∗_ large enough such that ∆[2]0 [divides][ Q][k]n[∗] [,]
2
and apply the previous observation. It follows that ∆1 0[))][∗] [as well, or]
_[∈]_ [(][R/][(][∆][2]
equivalently, there exist r0, r1 ∈ _R such that r0 · ∆[2]0_ [+][ r][1] _[·][ ∆]1[2]_ [= 1.]
9 A similar property was first noticed and exploited in [17,18] and later in [25].
-----
Set m = ⌊log n⌋ + 1. Let f[ˆ](X) ∈ Z[X] be any monic, irreducible polynomial
of degree m such that for all p ∈ _Πn, f[ˆ]p(X) (the polynomial f[ˆ](X) with its_
coefficients reduced modulo p) is irreducible in Fp[X].
One way of constructing such a polynomial is as follows. For all p ∈ _Πn, select_
a monic, irreducible polynomial f[ˆ]p(X) ∈ Fp[X] of degree m. By the theory of
finite fields, this is always possible. Applying the Chinese Remainder Theorem to
each of the coefficients separately, select an arbitrary lift to a monic polynomial
_fˆ(X) ∈_ Z[X] of degree m such that ˆf (X) ≡ _fˆp(X) mod p. Note that the monic_
polynomial f[ˆ](X) is irreducible in Z[X]: if not, reduction modulo p with p ∈ _Πn,_
gives a non-trivial factorization of f[ˆ]p(X) in Fp[X].
Set R = Z[X]/( f[ˆ](X)). By definition of f[ˆ](X), it follows that R/(p) is a finite
field, for all p ∈ _Πn. Indeed, for all p ∈_ _Πn,_
_R/(p) ≃_ Z[X]/(p, _f[ˆ](X)) ≃_ Fp[X]/( f[ˆ]p(X)) ≃ Fpm.
Note that all ideals (p) of R with p ∈ _Πn are distinct and maximal. It follows,_
using the Chinese Remainder Theorem for general rings, that
_R/(Qn) ≃_
�
_p∈Πn_
Fpm.
For all p ∈ _Πn we have |F[∗]p[m][|][ =][ p][m]_ _[−][1][ ≥]_ [2][m] _[−][1][ ≥]_ _[n][. Therefore, for each][ p][ ∈]_ _[Π][n][,]_
distinct non-zero
_β1[(][p][)][, . . ., β]n[(][p][)]_ _∈_ Fpm
can be selected. Finally, select arbitrary α1, . . ., αn ∈ _R such that, for i = 1 . . . n,_
_R/(Qn) ∋_ _αi ←→_ (βi[(][p][)])p∈Πn ∈
�
_p∈Πn_
Fp[m],
where the correspondence is via the (implicit) isomorphism. By construction,
for all i, j with 1 ≤ _i, j ≤_ _n and i ̸= j, it holds that αi ∈_ (R/(Qn))[∗] and
_αi −_ _αj ∈_ (R/(Qn))[∗]. Hence, ∆1 ∈ (R/(Qn))[∗], as desired. _⊓⊔_
**Corollary 1. For all integers t, n with 0 < t < n** 1, there exists an ISP of size
_−_
_n · (⌊log n⌋_ + 2) for Tt,n.
_Proof. Let R, α1, . . ., αn, r0, r1, N0, N1 be as constructed in the proof of The-_
orem 2. Apply the construction from the proof of Lemma 3 to N1, and take
into account the final remark of that proof. This gives an ISP matrix N[ˆ]1 with
_n · (⌊log n⌋_ + 1) rows and 1 + t(⌊log n⌋ + 1) columns. Clearly, the sets A ̸∈ _Tt,n_
satisfy Definition 6. For the sets A ∈ _Tt,n, the rows owned by A span δ1_ _·εˆ, where_
_δ1_ Z is the first coordinate of r1 _∆[2]1[.]_
_∈_ _·_
The ISP matrix N0 has the properties stated in the proof of Theorem 2 also
over Z. Hence, the sets A ̸∈ _Tt,n satisfy Definition 6 over Z. For the sets A ∈_ _Tt,n,_
the rows owned by them clearly span (δ0, 0, . . ., 0) ∈ Z[t][+1], where δ0 ∈ Z is the
first coordinate of r0 · _∆[2]0[. Since][ δ][0]_ [+] _[δ][1]_ [= 1, this leads directly to an ISP for][ T][t,n][,]
where the ISP matrix has n ( log n +2) rows and t( log n +2)+1 columns.
_·_ _⌊_ _⌋_ _⌊_ _⌋_ _⊓⊔_
-----
## 5 Concluding Remarks
**5.1** **A Note on Simulateability**
The ISPs _Mˆ_ = (Z, _M,[ˆ]_ _ψ,[ˆ]_ ˆε) constructed in the proofs of Theorem 2 and Corol
lary 1 satisfy the following additional properties, which are helpful when proving
the security of certain threshold cryptosystems.
Let the share vector s = M[ˆ] **g be computed according to the corresponding**
black-box secret sharing scheme, then the following holds.
1. The entries of sA are independent random group elements for any subset A
of 1, . . ., n with _A_ _t._
_{_ _}_ _|_ _| ≤_
2. Every player i can compute a reconstruction share s[′]i [by taking][ Z][-linear com-]
binations (of course independent of the group) of the entries of his original
share si, such that any t reconstruction shares s[′]i [still allow to reconstruct]
the secret s, and such that any t original shares si together with s allow
to compute the complete reconstruction share vector s[′] (by taking Z-linear
combinations).
The former property is inherited from the two Vandermonde matrices upon
which the construction of _Mˆ_ is based on, and the latter holds for s[′] defined as
**s[′]** = M[ˆ] _[′]g, where the ISP_ _Mˆ_ _[′]_ = (Z, _M[ˆ]_ _[′],_ _ψ,[ˆ]_ ˆε) is constructed from the matrices
_∆0N0 and ∆1N1 in a way similar to which_ _Mˆ_ is constructed from N0 and N1
in the proof of Theorem 2.
Assuming that the group operation is efficiently computatble and that (al
most) random group elements can be sampled efficiently, these properties allow
the players of a set A with |A| ≤ _t to efficiently simulate their joint view sA of_
the distribution phase, by sampling (almost) random elements from the group
and to efficiently simulate their view of the corresponding reconstruction phase
by computing s[′] from sA and the secret s.
When proving the security of a direct application of our black-box secret
sharing scheme to distributed RSA for instance, these properties enable an efficient simulator for the adversary’s view of the distributed decryption or signing
process (see also [12,25]).
**5.2** **Implementation**
We stress that in this paper we are primarily interested in the asymptotically
optimal result from Theorem 2. Several choices in its proof have been made to
simplify the mathematical exposition, while suppressing computational aspects.
There are a number of possible practical implementations of black-box secret
sharing based on our result. We do not optimize its performance here, but merely
indicate below that straightforward implementations run in time polynomial
in n.
Note that the scheme consumes O(n log n) random coins (group elements)
and that the expansion factor is O(log n) in any case, i.e., each player receives
-----
_O(log n) groups elements as his share in a secret group element. For an imple-_
mentation, it is important to limit the necessary computational resources for
dealer and players.
One implementation is based on the well-known fact that for any finite
Abelian group G, G[m] can be viewed as a module over the ring R (see also [12]).
The multiplication of an element of R by an element of G[m] can be performed
having only black-box access to the group operation of G. This way, the monotone span program over R acts directly on vectors of elements of G[m]. This leads
in a straightforward fashion to an attractive implementation of black-box secret
sharing where the actual ISP it is based upon can be left implicit. See for instance [12] for the computational details of this general procedure, taking into
account the remarks below.
By the constructive method from the proof of Theorem 2, we may assume
without loss of generality that the coefficients of the polynomial f (X) have bit
length smaller than log Qn ≤ log(n!) = O(n log n) bits. Recall that its degree
_m is_ log n + 1. For given threshold parameters t, n, it can be fixed once and
_⌊_ _⌋_
for all. One simple possible choice for the αi’s is to identify them with distinct,
non-zero integer polynomials of degree at most log n, such that each of the
_⌊_ _⌋_
coefficients is either 0 or 1. For instance, αi can point to i by basing it on the
bit representation of i. ∆[2]0 [is simply represented by an integer with bit length]
_O(n[2]_ _· log n). The value ∆[2]1_ [is the product of][ O][(][n][2][) elements of][ R][, each of which]
has integer coordinates −1, 0 or 1. The values r0 and r1 can be obtained by
2
computing the inverse u of ∆1 0[), for instance by solving a linear system]
_[∈]_ _[R/][(][∆][2]_
of equations over Z∆20 [, and by computing][ u] _[·]_ _[∆]1[2]_ _[∈]_ _[R][. The reconstruction vectors]_
are computed from r0, r1 and obvious “interpolation coefficients” obtained from
the αi’s.
## Acknowledgments
We thank Ivan Damgaard for many helpful suggestions and discussions. Also
thanks to Yvo Desmedt, Yair Frankel, Anna G´al, Yuval Ishai, Brian King and
the anonymous referees of CRYPTO ’02 for comments.
## References
1. A. Beimel. Secure schemes for secret sharing and key distribution. Ph.D.-thesis,
Technion, Haifa, June 1996.
2. J. Benaloh and J. Leichter. Generalized secret sharing and monotone functions. In:
Proc. CRYPTO ’88, Springer LNCS, vol. 765, pp. 274–285, 1988.
3. M. Bertilsson, I. Ingemarsson. A construction of practical secret sharing schemes
using linear block codes. In Proc. AUSCRYPT ’92, Springer LNCS, vol. 718, pp.
67-79, 1993.
4. S. Blackburn, M. Burmester, Y. Desmedt, and P. Wild. Efficient multiplicative
sharing scheme. In: Proc. EUROCRYPT ’96, Springer LNCS, vol. 1070, pp. 107–
118, 1996.
-----
5. G. R. Blakley. Safeguarding cryptographic keys. In: Proc. National Computer Con
ference ’79, AFIPS Proceedings, vol. 48, pp. 313-317, 1979.
6. E. F. Brickell. Some ideal secret sharing schemes. In: J. Combin. Maths. & Combin.
Comp. vol. 9, pp. 105–113, 1989.
7. T. Cover and J. Thomas. Elements of information theory. Wiley Series in Telecom
munications, 1991.
8. R. Cramer, I. Damgaard, and U. Maurer. Efficient general secure multi-party
computation from any linear secret-sharing scheme. In: Proc. EUROCRYPT ’00,
Springer LNCS, vol. 1807, pp. 316–334, 2000.
9. R. Cramer, S. Fehr, Y. Ishai, and E. Kushilevitz. Efficient multi-party computation
over rings. Manuscript, February 2002.
10. Y. Di Crescenzo, and Y. Frankel. Existence of Multiplicative Secret Sharing
Schemes with Polynomial Share Expansion. In: Proc. SODA ’99, ACM Press, pp.
895–896, 1999.
11. Y. Desmedt and Y. Frankel. Theshold cryptosystem. In: Proc. CRYPTO ’89,
Springer LNCS, vol. 435, pp. 307–315, 1990.
12. Y. Desmedt and Y. Frankel. Homomorphic zero-knowledge threshold schemes over
any finite Abelian group. In: SIAM Journal on Discrete Mathematics, 7(4), pp.
667–679, 1994.
13. Y. Desmedt, A. De Santis, Y. Frankel, and M. Yung. How to share a function
securely. In: Proc. STOC ’94, ACM Press, pp. 22–33, 1994.
14. Y. Desmedt, G. Di Crescenzo, and M. Burmester. Multiplicative non-abelian
sharing schemes and their application to threshold cryptography. In: Proc. ASIACRYPT ’94, Springer LNCS, vol. 917, pp. 21–31, 1995.
15. Y. Desmedt, B. King, W. Kishimoto, and K. Kurosawa. A comment on the effi
ciency of secret sharing scheme over any finite Abelian group. In: Proc. ACISP ’98,
Springer LNCS, vol. 1438, pp. 391–402, 1998.
16. M. van Dijk. Secret key sharing and secret key generation. Ph. D. Thesis, Eind
hoven University of Technology, 1997.
17. Y. Frankel, P. Gemmell, P. MacKenzie, and M. Yung. Optimal resilience proactive
public-key cryptosystems. In: Proc. FOCS ’97, IEEE Press, pp. 384–393, 1997.
18. Y. Frankel, P. Gemmell, P. MacKenzie, and M. Yung. Proactive RSA. In: Proc.
CRYPTO ’97, Springer LNCS, vol. 1294, pp. 440–454, 1997.
19. A. G´al. Combinatorial methods in boolean function complexity. Ph.D.-thesis, Uni
versity of Chicago, 1995.
20. M. Karchmer and A. Wigderson. On span programs. In: Proc. Structures in Com
plexity Theory ’93, IEEE Computer Society Press, pp. 102–111, 1993.
21. B. King. Some results in linear secret sharing. Ph.D.-thesis, University of
Wisconsin-Milwaukee, 2001.
22. B. King. Randomness required for linear threshold sharing schemes defined over
any finite abelian group. In: Proc. ACISP ’01, Springer LNCS, vol. 2119, pp. 376–
391, 2001.
23. S. Lang. Algebra. Addison-Wesley Publishing Co., 2nd edition, 1984.
24. A. Shamir. How to share a secret. In: Communications of the ACM, (22) pp. 612–
613, 1979.
25. V. Shoup. Practical threshold signatures. In: Proc. EUROCRYPT ’00, Springer
LNCS, vol. 1807, pp. 207–220, 2000.
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.1007/3-540-45708-9_18?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1007/3-540-45708-9_18, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBYNCND",
"status": "CLOSED",
"url": "https://tidsskrift.dk/brics/article/download/21726/19160"
}
| 2,002
|
[
"JournalArticle",
"Conference"
] | true
| 2002-02-05T00:00:00
|
[
{
"paperId": "7dbdb4209626fd92d2436a058663206216036e68",
"title": "Elements of Information Theory"
},
{
"paperId": "38647ae73600a9cae911c84d7f2463321b333a89",
"title": "Efficient Multi-party Computation over Rings"
},
{
"paperId": "b7056d0ddc9f4d8e2da0eb2cbbb51f39df883324",
"title": "Equational Axioms for Probabilistic Bisimilarity (Preliminary Report)"
},
{
"paperId": "aac0a5b51b82c5643b78ed95bc7004409a71be90",
"title": "Composing Strand Spaces"
},
{
"paperId": "9da45d0ac15cd4d5877cc70eae4667f7594413fb",
"title": "A Formalization of Linkage Analysis"
},
{
"paperId": "30687a58847d9defc1fdfb478af9ff8163dfe851",
"title": "A Simple Correctness Proof of the Direct-Style Transformation"
},
{
"paperId": "2559ce9a3716cfff210db80265ffc57f0f475f89",
"title": "Thesis"
},
{
"paperId": "cfd0b5fa49cc7cbc8fcb8b58643fe63c400393dc",
"title": "Temporal Logic with Cyclic Counting and the Degree of Aperiodicity of Finite Automata"
},
{
"paperId": "75373008ac7c703a3ce1f815c2c9eb9aa0f9c456",
"title": "A Simple CPS Transformation of Control-Flow Information"
},
{
"paperId": "2a423840a199b5999e623e19551b9c3af11ae822",
"title": "Extracting Witnesses from Proofs of Knowledge in the Random Oracle Model"
},
{
"paperId": "b3e0be5d8981b5e02d07e8c4dde919b835e21e29",
"title": "Randomness Required for Linear Threshold Sharing Schemes Defined over Any Finite Abelian Group"
},
{
"paperId": "970e8206335b0ca91bc25d7e079f14f38c33e3a4",
"title": "Syntactic Theories in Practice"
},
{
"paperId": "d728243147560dd11960bc30f80a14fe6533a508",
"title": "Practical Threshold Signatures"
},
{
"paperId": "64856147da1c26464f183f13be9137d60e23a729",
"title": "General Secure Multi-party Computation from any Linear Secret-Sharing Scheme"
},
{
"paperId": "9232f19cc6419363d0e39752139de518a4bc406e",
"title": "Syntactic accidents in program analysis: on the impact of the CPS transformation"
},
{
"paperId": "29806fa1bb52189c8df2ce98f4e403ebfbe55be9",
"title": "A Comment on the Efficiency of Secret Sharing Scheme over Any Finite Abelian Group"
},
{
"paperId": "1723d3a80ee241c0e04f3870f4373a2de0798ac0",
"title": "Optimal-resilience proactive public-key cryptosystems"
},
{
"paperId": "483f767ed12bc31b97dcf59ed0c019303f195c6d",
"title": "Proactive RSA"
},
{
"paperId": "8fb9ab9929ce1a991b5b2ed8dd29f71b567da3e4",
"title": "Efficient Multiplicative Sharing Schemes"
},
{
"paperId": "da417d59441865b90d0763933e226db09eebc5fe",
"title": "Multiplicative Non-abelian Sharing Schemes and their Application to Threshold Cryptography"
},
{
"paperId": "f5451dd09f0b153ce05a06178a03f97dc1e30562",
"title": "Perfect Homomorphic Zero-Knowledge Threshold Schemes over any Finite Abelian Group"
},
{
"paperId": "9b05ece21561b6d240292df8cc1b7366233ed98e",
"title": "How to share a function securely"
},
{
"paperId": "87b5014904410775f09aa7512376b804541598ed",
"title": "On span programs"
},
{
"paperId": "12d4e2345bb0cb0e6e559d6da9d5458c01adc7c8",
"title": "A Construction of Practical Secret Sharing Schemes using Linear Block Codes"
},
{
"paperId": "6d96340b98082a0da318355361977ab5801694a0",
"title": "Some Ideal Secret Sharing Schemes"
},
{
"paperId": "12113d0c06262e149df06516b77b8257c8b1a492",
"title": "Generalized Secret Sharing and Monotone Functions"
},
{
"paperId": "bdccfcb9e908460052b799d0ccf6a21e92225e35",
"title": "Threshold Cryptosystems"
},
{
"paperId": "88abb2cda4f2a57499a717966ac4fbe9a993027a",
"title": "How to share a secret"
},
{
"paperId": "32d21ccc21a807627fcb21ea829d1acdab23be12",
"title": "Safeguarding cryptographic keys"
},
{
"paperId": null,
"title": "Recent BRICS Report Series Publications RS-02-8 Ronald Cramer and Serge Fehr. Optimal Black-Box Secret Sharing over Arbitrary Abelian Groups"
},
{
"paperId": null,
"title": "RS-02-1 Claus Brabrand, Anders Møller, and Michael I. Schwartzbach. The <bigwig> Project"
},
{
"paperId": "d454a3aedc5b19b4aee6fa68ccca9f240ef4e776",
"title": "Some results in linear secret sharing"
},
{
"paperId": "795633d34a83c571571725610664067c1f144ab8",
"title": "Existence of multiplicative secret sharing schemes with polynomial share expansion"
},
{
"paperId": "8ae4c5492ca8525f0d438c42600d8f8cfd4b80a5",
"title": "Secret key sharing and secret key generation"
},
{
"paperId": null,
"title": "Proc. CRYPTO '97"
},
{
"paperId": "8c77cdf0ae09e931e67037aad106b671d28345a3",
"title": "Secure schemes for secret sharing and key distribution"
},
{
"paperId": null,
"title": "In: Proc"
},
{
"paperId": null,
"title": "Algebra"
},
{
"paperId": "b6418463d62358a421f936c7a709d48f1b857ac4",
"title": "The Faculty of the Division of the Physical Sciences in Candidacy for the Degree of Doctor of Philosophy Department of Computer Science"
}
] | 16,161
|
en
|
[
{
"category": "Physics",
"source": "external"
},
{
"category": "Biology",
"source": "external"
},
{
"category": "Mathematics",
"source": "external"
},
{
"category": "Medicine",
"source": "external"
},
{
"category": "Mathematics",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/02b2d773a7a2249e864f271b4964ef7bbd942499
|
[
"Physics",
"Biology",
"Mathematics",
"Medicine"
] | 0.809525
|
Emergence of Oscillations in a Mixed-Mechanism Phosphorylation System
|
02b2d773a7a2249e864f271b4964ef7bbd942499
|
Bulletin of Mathematical Biology
|
[
{
"authorId": "2285231",
"name": "C. Conradi"
},
{
"authorId": "3291335",
"name": "Maya Mincheva"
},
{
"authorId": "4686418",
"name": "Anne Shiu"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"Bull Math Biology"
],
"alternate_urls": null,
"id": "8a47ce78-e793-4315-8bf8-b09466ebf633",
"issn": "0092-8240",
"name": "Bulletin of Mathematical Biology",
"type": "journal",
"url": "http://link.springer.com/journal/11538"
}
|
This work investigates the emergence of oscillations in one of the simplest cellular signaling networks exhibiting oscillations, namely the dual-site phosphorylation and dephosphorylation network (futile cycle), in which the mechanism for phosphorylation is processive, while the one for dephosphorylation is distributive (or vice versa). The fact that this network yields oscillations was shown recently by Suwanmajo and Krishnan. Our results, which significantly extend their analyses, are as follows. First, in the three-dimensional space of total amounts, the border between systems with a stable versus unstable steady state is a surface defined by the vanishing of a single Hurwitz determinant. Second, this surface consists generically of simple Hopf bifurcations. Next, simulations suggest that when the steady state is unstable, oscillations are the norm. Finally, the emergence of oscillations via a Hopf bifurcation is enabled by the catalytic and association constants of the distributive part of the mechanism; if these rate constants satisfy two inequalities, then the system generically admits a Hopf bifurcation. Our proofs are enabled by the Routh–Hurwitz criterion, a Hopf bifurcation criterion due to Yang, and a monomial parametrization of steady states.
|
# Emergence of oscillations in a mixed-mechanism phosphorylation system
#### Carsten Conradi[∗], Maya Mincheva[†], and Anne Shiu[‡]
#### January 28, 2019
Abstract
This work investigates the emergence of oscillations in one of the simplest cellular
signaling networks exhibiting oscillations, namely, the dual-site phosphorylation and
dephosphorylation network (futile cycle), in which the mechanism for phosphorylation
is processive while the one for dephosphorylation is distributive (or vice-versa). The
fact that this network yields oscillations was shown recently by Suwanmajo and Krishnan. Our results, which significantly extend their analyses, are as follows. First,
in the three-dimensional space of total amounts, the border between systems with a
stable versus unstable steady state is a surface defined by the vanishing of a single
Hurwitz determinant. Second, this surface consists generically of simple Hopf bifurcations. Next, simulations suggest that when the steady state is unstable, oscillations
are the norm. Finally, the emergence of oscillations via a Hopf bifurcation is enabled
by the catalytic and association constants of the distributive part of the mechanism: if
these rate constants satisfy two inequalities, then the system generically admits a Hopf
bifurcation. Our proofs are enabled by the Routh-Hurwitz criterion, a Hopf-bifurcation
criterion due to Yang, and a monomial parametrization of steady states.
Keywords: multisite phosphorylation, monomial parametrization, oscillation, Hopf
bifurcation, Routh-Hurwitz criterion
### 1 Introduction
Oscillations have been observed experimentally in signaling networks formed by phosphorylation and dephosphorylation [20, 21], which suggests that these networks are involved in
timekeeping and synchronization. Indeed, multisite phosphorylation is the main mechanism
for establishing the 24-hour period in eukaryotic circadian clocks [30, 42]. Our motivating
question, therefore, is, How do oscillations arise in phosphorylation networks?
We tackle this question for the network that, according to Suwanmajo and Krishnan,
“could be the simplest enzymatic modification scheme that can intrinsically exhibit oscillation” [39, §3.1]. This network, in (1), is the mixed-mechanism (partially processive, partially
∗HTW Berlin
†Northern Illinois University
‡Texas A&M University
1
-----
distributive) dual-site phosphorylation network (or mixed-mechanism network for short).
Examples of networks that include both processive and distributive elements include the
“processive model” of Aoki et al. [1, Table S2] and a model of ERK regulation via enzymes
MEK and MKP3 [37, Fig. 2].
In the mixed-mechanism network, Si denotes a substrate with i phosphate groups attached, and K and P are, respectively, a kinase and a phosphatase enzyme:
k1
k3 k4
S0 + K ⇄ S0K −→ S1K −→ S2 + K
k2
(1)
k5 k8
k7 k10
S2 + P ⇄ S2P −→ S1 + P ⇄ S1P −→ S0 + P .
k6 k9
When the kinase phosphorylates – that is, adds phosphate groups to – a substrate in the
mixed-mechanism network (via the reactions labeled by k1 to k4), the kinase and substrate
do not dissociate before both phosphate groups are added. Accordingly, the mechanism
for phosphorylation is processive. In contrast, when the phosphatase dephosphorylates –
i.e., removes phosphate groups from – a substrate (via reactions k5 to k10), this mechanism
is distributive: the phosphatase and substrate dissociate each time a phosphate group is
removed. Accordingly, network (1) is said to have a mixed mechanism[1].
The dynamical systems arising from the mixed-mechanism network live in a 9-dimensional
space, but, due to three conservation laws, are essentially 6-dimensional. Specifically, the
total amounts of kinase, phosphatase, and substrate – denoted by Ktot, Ptot, and Stot,
respectively – are conserved. For each choice of three such total amounts and each choice of
positive rate constants ki, there is a unique positive steady state [39]. One focus of our work
is determining when such a steady state undergoes a Hopf bifurcation leading to oscillations
(with any of the ki’s or total amounts as bifurcation parameter).
#### 1.1 Summary of main results
How do oscillations of the mixed-mechanism network emerge, and how robust are they?
These questions are the motivation for our work. Let us describe Suwanmajo and Krishnan’s
progress in this direction. They first found rate constants ki and total amounts, displayed
in Table 1, that yield oscillations [39, Supplementary Information].
k1 k2 k3 k4 k5 k6 k7 k8 k9 k10 Ktot Ptot Stot
1 1 1 1 100 1 0.9 3 1 100 17.5 5 40
Table 1: Rate constants (left) and total amounts (right), from [39, Supplementary Information], which lead to oscillations in the mixed-mechanism network (1).
Next, they examined whether oscillations persist as Ktot varies. What they found,
summarized in Figure 1, is that oscillations persist when Ktot is in the (approximate) interval
(13.03, 29.23), and oscillations arise as the unique steady state undergoes a Hopf bifurcation.
1Network (1) is symmetric to the mixed-mechanism network in which phosphorylation is distributive
(instead of processive) and dephosphorylation is processive (instead of distributive), so our results apply
equally well to that network (cf. [39, networks 21–22]).
2
-----
steady state is steady state is steady state is
Hopf Hopf
locally stable unstable locally stable
K
tot
(oscillations)
0 ≈ 13.03 ≈ 29.23
Figure 1: Stability of the unique steady state of the mixed-mechanism network (1) as a
function of Ktot, as analyzed by Suwanmajo and Krishnan [39, Fig. 4]. (The other total
amounts, Ptot and Stot, and the rate constants ki are those in Table 1.) Oscillations were
found when Ktot is in the “unstable” interval [39].
Subsequently, Conradi and Shiu [7] found that when Ptot also is allowed to vary, oscillations exist for larger values of Ktot (e.g., Ktot = 100). So, how exactly do oscillations
depend on the three total amounts (or, equivalently, the initial conditions)? Concretely, our
goal is to expand Figure 1 to encompass all possible perturbations to the initial conditions
(i.e., the total amounts):
Question 1.1. Consider the mixed-mechanism network (1), with ki’s from Table 1.
1. For which values of (Ktot, Ptot, Stot) ∈ R[3]>0 [is the unique steady state unstable?]
2. Whenever (by perturbing parameters or total amounts) a steady state switches from
being locally stable to unstable, does this always give rise to a Hopf bifurcation?
The direct method for solving Question 1.1(1) is to solve the steady-state equations, and
then apply the six-dimensional Routh-Hurwitz stability criterion. However, this approach is
intractable: the resulting Hurwitz determinants are pages-long.
Accordingly, we take an algebraic shortcut. Namely, we find a parametrization of the set
of steady states, and then use this for the input to Routh-Hurwitz. The result is somewhat
surprising: each Hurwitz determinant except the last two (which are positive multiples of
each other) is always positive. This yields our answer to Question 1.1(1): For every ODE
system arising from the mixed-mechanism network (1), a (two-dimensional) surface in the
three-dimensional space of total amounts defines the border between steady states that are
stable and those that are unstable. (Our result even applies to many systems for which the
ki’s are not those in Table 1; see Proposition 4.1.)
We can now translate Question 1.1(2) as follows: does the surface mentioned above
consist of Hopf bifurcations? We prove, using a Hopf-bifurcation criterion stated in terms of
Hurwitz determinants, due to Yang [43], that the answer, at least generically, is “yes”: When
the unique steady state of the mixed-mechanism network (1) switches from being stable to
unstable, then, generically, it undergoes a Hopf bifurcation.
For general one-parameter ODE systems, there are two types of local bifurcations: saddle
nodes (which require a zero eigenvalue of the Jacobian matrix) and Hopf bifurcations (which
require a pair of pure imaginary eigenvalues of the Jacobian) [16]. We show that a saddle
node bifurcation can not occur for any parameter values (see the proof of Proposition 4.1).
Therefore, only Hopf bifurcations are possible for the mixed-mechanism system.
A second question we aim to answer is the following:
3
|locally stable Ho|unstable opf Ho|locally stable opf|
|---|---|---|
||unstable||
||||
-----
Question 1.2. Consider the mixed-mechanism network (1). What conditions on the ki’s
guarantee a Hopf-bifurcation for some (positive) values of the total concentrations?
As an answer to Question 1.2, we prove that the catalytic constants (k7 and k10) and
association constants (k5 and k8) of the distributive part of the mechanism enable oscillations
to emerge via a Hopf bifurcation. Specifically, under the simplifying assumption that all
dissociation (backward-reaction) constants are equal (k2 = k6 = k9), if the rate constants
satisfy two inequalities – lower bounds on k10 and k5/k8 – then the system generically admits
a Hopf bifurcation (Proposition 4.3 and Theorem 4.5). (As a comparison, for the fully
distributive dual-site network described in Section 1.2 below, the catalytic constants alone
enable bistability [5].) Finally, we encode the relevant inequalities in a procedure to generate
many parameter values for which we expect oscillations (Procedure 5.1).
#### 1.2 Connection to related work
Our work joins a growing number of works that harness steady-state parametrizations. Such
results include criteria for when such parametrizations exist [26, 40] and methods for using
them to determine whether a network is multistationary [25, 29, 32, 34]. Going further,
steady-state parametrizations can also be used to find a witness to multistationarity or even
the precise parameter regions that yield multistationarity [4, 5]. In this work, we use a
steady-state parametrization in a novel way: to study oscillations via Hopf bifurcations.
(Our approach is similar in spirit to using Clarke’s convex parameters together with a Hopfbifurcation criterion [9, 11, 14, 18]).
As mentioned earlier, there has been much interest in the dynamics of phosphorylation
systems [7]. The mixed-mechanism network (1) fits into the related literature as follows.
The mixed network is a dual-site network situated between two extremes: the fully processive dual-site network – in which the phosphorylation and dephosphorylation mechanisms
are both processive – and the fully distributive dual-site network. One might therefore expect the dynamics of the mixed-mechanism network to straddle those of the two networks.
This is indeed the case. As summarized in Table 2, and reviewed in [7], fully processive networks are globally convergent to a unique steady state [6, 10, 35], while mixed-mechanism
networks admit oscillations but not bistability [39], and fully distributive networks admit
bistability [19] (and the question of oscillations is open [7]).
Dual-site network Oscillations? Bistability? Global convergence?
Fully processive No No Yes
Mixed-mechanism Yes No No
Fully distributive (Open) Yes No
Table 2: Dual-site phosphorylation networks and their properties: whether they admit oscillations or bistability, and whether all trajectories converge to a unique steady state.
Finally, we revisit Suwanmajo and Krishnan’s claim mentioned earlier that the mixedmechanism network is among the simplest enzymatic mechanisms with oscillations. In support of this claim, Tung proved that the simpler system obtained from the mixed-mechanism
network by taking its (two-dimensional) Michaelis-Menten approximation, is not oscillatory
4
-----
[41]. Moreover, Rao showed that this approximation is globally convergent to a unique
steady state [36]. The validity of the Michaelis-Menten approximation for phosphorylation
systems has been called into question [38], and what we know about the mixed-mechanism
system concurs: this system is oscillatory, but its Michaelis-Menten approximation is not.
The outline of our work is as follows. Section 2 provides background on multisite phosphorylation, steady states, and Hopf bifurcations. Section 3 gives a monomial parametrization
of the steady states of mixed-mechanism network. In Section 4, we prove our main results
(described above). We use these results in Section 5 to give a procedure for generating rate
constants admitting Hopf bifurcations. In Section 6, we present simulations that suggest
that oscillations are the norm in the unstable-steady-state regime. Finally, we end with a
Discussion in Section 7.
### 2 Background
In this section, we introduce the ODEs arising from the mixed-mechanism network, and recall
two criteria: the Routh-Hurwitz criterion for steady-state stability and Yang’s criterion for
Hopf bifurcations.
#### 2.1 Differential equations of the mixed-mechanism network
For the mixed-mechanism network (1), we let x1, x2, . . ., x9 denote the species concentrations
in the order given in Table 3. The dynamical system (arising from mass-action kinetics)
defined by the mixed-mechanism network (1) is given by the following ODEs:
x˙ 1 = − k1x1x2 + k2x3 + k10x9
x˙ 2 = − k1x1x2 + k2x3 + k4x4
x˙ 3 = k1x1x2 − (k2 + k3)x3
x˙ 4 = k3x3 − k4x4
x˙ 5 = k4x4 − k5x5x6 + k6x7 (2)
x˙ 6 = − k5x5x6 − k8x8x6 + (k6 + k7)x7 + (k9 + k10)x9
x˙ 7 = k5x5x6 − (k6 + k7)x7
x˙ 8 = k7x7 − k8x6x8 + k9x9
x˙ 9 = k8x6x8 − (k9 + k10)x9 .
x1 x2 x3 x4 x5 x6 x7 x8 x9
S0 K S0K S1K S2 P S2P S1 S1P
Table 3: Assignment of variables to species for the mixed-mechanism network (1).
The conservation laws arise from the fact that the total amounts of free and bound
enzyme or substrate remain constant. That is, as the dynamical system (2) progresses, the
5
-----
following three conservation values, denoted by Ktot, Ptot, Stot ∈ R>0, remain constant:
Ktot = x2 + x3 + x4,
Ptot = x6 + x7 + x9, (3)
Stot = x1 + x3 + x4 + x5 + x7 + x8 + x9 .
Also, a trajectory x(t) beginning in R[9]≥0 [remains in][ R]≥[9] 0 [for all positive time][ t][, so it]
remains in a stoichiometric compatibility class, which we denote as follows:
P = {x ∈ R[9]≥0 [|][ the conservation equations (3) hold][}][ .] (4)
#### 2.2 Stability of steady states and the Routh-Hurwitz criterion
The dynamical system (2) arising from the mixed-mechanism network is an example of a
reaction kinetics system. That is, the system of ODEs takes the following form:
dx
= Γ · R(x) =: g(x), (5)
dt
where Γ and R are as follows. Letting s denote the number of species and r the number of
reactions, Γ is an s × r matrix whose k-th column is the reaction vector of the k-th reaction,
i.e., it encodes the net change in each species that results when that reaction takes place.
Also, R : R[s]≥0 [→] [R]≥[r] 0 [encodes the reaction rates of the][ r][ reactions as functions of the][ s]
species concentrations.
A steady state (respectively, positive steady state) of a reaction kinetics system is a nonnegative concentration vector x[∗] ∈ R[s]≥0 [(respectively,][ x][∗] [∈] [R]>[s] 0[) at which the ODEs (5)]
vanish: g(x[∗]) = 0. Letting S := im(Γ) denote the stoichiometric subspace, a steady state x[∗]
is nondegenerate if Im (dg(x[∗])|S) = S, where dg(x[∗]) denotes the Jacobian matrix of g at x[∗].
A nondegenerate steady state is locally asymptotically stable if each of the σ := dim(S)
nonzero eigenvalues of dg(x[∗]) has negative real part. Hence, a steady state is locally stable
if and only if the characteristic polynomial of the Jacobian evaluated at the steady state has
σ roots with negative real part (the remaining roots will be 0).
To check whether a polynomial has only roots with negative real parts, we appeal to the
Routh-Hurwitz criterion below [13].
Definition 2.1. The i-th Hurwitz matrix of a univariate polynomial p(λ) = a0λ[n] + a1λ[n][−][1] +
- · · + an is the following i × i matrix:
a1 a0 0 0 0 - · · 0
a3 a2 a1 a0 0 - · · 0
... ... ... ... ... ...
a2i−1 a2i−2 a2i−3 a2i−4 a2i−5 - · · ai
[,]
Hi =
in which the (k, l)-th entry is a2k−l as long as 0 ≤ 2k − l ≤ n, and 0 otherwise.
Proposition 2.2 (Routh-Hurwitz criterion). A polynomial p(λ) = a0λ[n] + a1λ[n][−][1] + · · · + an
with a0 > 0 has all roots with negative real part if and only if all n of its Hurwitz matrices
have positive determinant (det Hi > 0 for all i = 1, . . ., n).
6
-----
#### 2.3 Hopf bifurcations and a criterion due to Yang
A simple Hopf bifurcation is a bifurcation in which a single complex-conjugate pair of eigenvalues of the Jacobian matrix crosses the imaginary axis, while all other eigenvalues remain
with negative real parts. Such a bifurcation, if it is supercritical, generates nearby oscillations
or periodic orbits [27].
To detect simple Hopf bifurcations, we will use a criterion of Yang that characterizes
Hopf bifurcations in terms of Hurwitz-matrix determinants (Proposition 2.3).
Setup for Yang’s criterion. We consider an ODE system parametrized by µ ∈ R:
x˙ = gµ(x),
where x ∈ R[n], and gµ(x) varies smoothly in µ and x. Assume that x0 ∈ R[n] is a steady
state of the system defined by µ0, that is, gµ0(x0) = 0. Assume, furthermore, that we have
a smooth curve of steady states:
µ �→ x(µ) (6)
(that is, gµ (x(µ)) = 0 for all µ) and that x(µ0) = x0. Denote the characteristic polynomial
of the Jacobian matrix of gµ, evaluated at x(µ), as follows:
pµ(λ) := det (λI − Jac gµ) |x=x(µ) = λ[n] + a1(µ)λ[n][−][1] + · · · + an(µ),
and, for i = 1, . . ., n, let Hi(µ) denote the i-th Hurwitz matrix of pµ(λ).
Proposition 2.3 (Yang’s criterion [43]). Assume the above setup. Then, there is a simple
Hopf bifurcation at x0 with respect to µ if and only if the following hold:
(i) an(µ0) > 0,
(ii) det H1(µ0) > 0, det H2(µ0) > 0, . . ., det Hn−2(µ0) > 0, and
(iii) det Hn−1(µ0) = 0 and [d][(det][ H]dµ[n][−][1][(][µ][))] |µ=µ0 ̸= 0.
Remark 2.4. Liu [27] gave an earlier version of Yang’s Hopf-bifurcation criterion (Proposition 2.3), using a variant of the Hurwitz matrices that differs from ours.
### 3 Steady states of the mixed-mechanism network
In this section, we recall that the mixed-mechanism network admits a unique steady state in
each compatibility class (Proposition 3.1), and prove that the set of steady states admits a
monomial parametrization (Theorem 3.2). We then use this parametrization to analyze the
space of compatibility classes (Proposition 3.6).
7
-----
#### 3.1 Uniqueness of steady states
Suwanmajo and Krishnan proved that, for every choice of positive rate constants and positive total amounts, the mixed-mechanism network does not admit multiple positive steady
states [39, §A.2]. Additionally, there are no boundary steady states in any compatibility
class P, as in (4), and P is compact. Hence, via a standard application of the Brouwer
fixed-point theorem (e.g., [33, Remark 3.9]), there is always a unique steady state:
Proposition 3.1 (Uniqueness of steady states). For any choice of positive rate constants ki
and positive total amounts K, P, and S (2) arising from the
tot tot tot, the dynamical system
mixed-mechanism network has a unique steady state in P, and it is a positive steady state.
Proposition 3.1 precludes the existence of multiple positive steady states, and hence the
existence of a saddle node bifurcation. Thus, a Hopf bifurcation is the only other oneparameter bifurcation which may occur. Indeed, we will show that a Hopf bifurcation exists
for some parameter values in Section 4.
Also, Proposition 3.1 proves part of a conjecture that we posed [6]. The other half
of the conjecture, however, posited that mixed-mechanism systems, like fully processive
systems [6, 10], are globally convergent to the unique steady state. Suwanmajo and Krishnan
demonstrated that this is false: the system can exhibit oscillatory behavior [39]!
This capacity for oscillations is the focus of this work, and our analysis will harness a
monomial parametrization of the steady states. We turn to this topic now.
#### 3.2 A monomial parametrization of the steady states
The steady states of the mixed-mechanism network can be parametrized by monomials (and
thus is said to have “toric steady states” [33]):
Proposition 3.2 (Parametrization of the steady states). For every choice of rate constants ki > 0, the set of positive steady states of the mixed-mechanism system (2) is threedimensional and is the image of the following map χ = χk1,...,k10:
χ : R[3]+ [→] [R]+[9] (7)
(x1, x2, x6) �→ (x1, x2, . . ., x9),
given by
k1 k1k3
x3 := x1x2, x4 := x1x2, x5 := [k][1][k][3][(][k][6][ +][ k][7][)]
k2 + k3 (k2 + k3)k4 (k2 + k3)k5k7
x1x2
,
x6
k1k3
x7 := x1x2, x8 := [k][1][k][3][(][k][9][ +][ k][10][)]
(k2 + k3)k7 (k2 + k3)k8k10
x1x2 k1k3
, x9 := x1x2 .
x6 (k2 + k3)k10
Proof. It is straightforward to check that the image of χ is contained in the set of steady
states: after substituting χ(x1, x2, x3), the right-hand side of the mixed-mechanism network
ODEs (2) vanishes. Conversely, let x[∗] = (x1, x2, . . ., x9) be a positive steady state. The
right-hand side of the ODEs (2) vanish at x[∗], so, in the following order, we use ˙x3 = 0 to
solve for x3 in terms of x1 and x2, use ˙x4 = 0 to solve for x4 via x3 which was already
8
-----
obtained, use ˙x1 = 0 to obtain x9, use ˙x9 = 0 to obtain x8, use ˙x8 = 0 to obtain x7, and
finally use ˙x7 = 0 to obtain x5. This yields precisely the parametrization (7), so x[∗] is in the
image of χ.
Remark 3.3. The parametrization (7) appeared earlier in [7].
Remark 3.4. That we could achieve a steady-state parametrization was expected, due to
Thomson and Gunawardena’s rational parametrization theorem for multisite systems [40].
Remark 3.5. In the parametrization χ in Theorem 3.2, we divide by x6, so χ is technically
not a monomial map. However, χ can be made monomial: we introduce y := x[x]6[1] [, so that the]
parametrization accepts as input (y, x2, x6), and then x1 is replaced by yx6.
#### 3.3 A parametrization of the compatibility classes
Every compatibility class P of the mixed-mechanism network, by definition (4), is uniquely
determined by a choice of total amounts (Ktot, Ptot, Stot) ∈ R[3]>0[. Thus, we identify the]
set of compatibility classes with {(Ktot, Ptot, Stot)} = R[3]>0[. We parametrize this set below]
(Proposition 3.6).
Let φ : R[9]>0 [→] [R]>[3] 0 [denote the map sending a vector of concentrations to the correspond-]
ing total amounts (K, P, S
tot tot tot), as in (3):
φ(x) := (x2 + x3 + x4, x6 + x7 + x9, x1 + x3 + x4 + x5 + x7 + x8 + x9) . (8)
Each compatibility class P contains a unique positive steady state (Proposition 3.1), and the
positive steady states are parametrized by χ from Theorem 3.2, so the space of compatibility
classes is parametrized as follows:
Proposition 3.6 (Parametrization of the compatibility classes). Identify every compatibility class P of the mixed-mechanism network (1), with the corresponding total amounts
(Ktot, Ptot, Stot) ∈ R[3]>0[. Then, for every choice of positive rate constants][ k][i][, the following]
is a bijection that sends a vector (x1, x2, x6) ∈ R[3]>0 [to the compatibility class in which the]
unique steady state is χ(x1, x2, x6):
φ ◦ χ : R[3]>0 [→] [R]>[3] 0 [=][ {][(][K]tot[, P]tot[, S]tot[)][}][,]
where φ is as in (8) and χ is the steady-state parametrization (7). The map φ ◦ χ is given
by
k1
x2 +
k2 + k3
� k1k3
x1x2, x6 +
k2 + k3
� 1
+ [1]
k7 k10
�1 + [k][3]
k4
�
x1x2,
(x1, x2, x6) �→
�
,
�
+ [1]
x6
�k6 + k7 + [k][10][ +][ k][9]
k5k7 k10k8
��
x1x2
�
k1k3
x1 +
k2 + k3
�� 1
+ [1] + [1] + [1]
k3 k4 k7 k10
which becomes, when the rate constants are those in Table 1, the following:
� x1x2
(x1, x2, x6) �→ x1x2 + x2, x6 + [1009]
1800[x][1][x][2][, x][1][ + 2809]1800[x][1][x][2][ + 161]900 x6
�
. (9)
9
-----
Example 3.7. Consider the mixed-mechanism system with rate constants from Table 1. To
compute the unique steady state x[∗] in the compatibility class given by (K, P, S
tot tot tot) =
(17.5, 5, 40), we use Proposition 3.6. Namely, we know that φ ◦ χ(x[∗]1[, x][∗]2[, x][∗]6[) = (17][.][5][,][ 5][,][ 40),]
so we solve (using, e.g., Mathematica [22]) for the unique positive solution:
(x[∗]1[, x][∗]2[, x][∗]6[)][ ≈] [(1][.][0134][,][ 8][.][6916][,][ 0][.][0624)][ .]
We obtain the remaining coordinates of x[∗] using the parametrization χ in (7):
x[∗] = χ(x[∗]1[, x][∗]2[, x][∗]6[)] (10)
≈ (1.0134, 8.6916, 4.4041, 4.4041, 1.4893, 0.0624, 4.8935, 23.7512, 0.0440) .
#### 3.4 Steady states and Hopf bifurcations
Our analysis of oscillations in the mixed-mechanism system is based on Hopf bifurcations.
Hopf-bifurcation diagrams are displayed in Figure 2, where the total amounts are the bifurcation parameters (c.f. Figure 1 which is with respect to Ktot). Figure 2 suggests that, in
the 3-dimensional space of total amounts, there is a surface of Hopf bifurcations. Indeed, we
will see in the next section that this is the case (see Theorem 4.5 and Figure 3).
(a) Bif. parameter Ktot.
(b) Bif. parameter Ptot.
(c) Bif. parameter Stot.
Figure 2: Numerical continuation of the unique positive steady state, in (10), when
(K, P, S .5, 5, 40): (a) For P, 8 and S
tot tot tot) = (17 tot = 5 tot = 40, we observe (supercritical) Hopf bifurcations at Ktot ≈ 13.0296, 29.2251 (Ptot = 5) and Ktot ≈ 18.5758
(Ptot = 8). (b) For Ktot = 5 and Stot = 40, we observe (supercritical) Hopf bifurcations at
Ptot ≈ 4.6310 and Ptot ≈ 7.5479. (c) For Ktot = 17.5 and Ptot = 5, we observe (supercritical) Hopf bifurcations at Stot ≈ 21.8213 and Stot ≈ 43.5944. All figures in this work were
made using Matcont [8].
### 4 Hopf bifurcations in the mixed-mechanism system
We saw in the previous section that the mixed-mechanism network yields a unique positive
steady state in each compatibility class. Now we show that the compatibility classes with
a stable steady state are separated from those with an unstable steady state by a single
surface H (Proposition 4.1 and Theorem 4.2), and, under stronger hypotheses, crossing the
surface H generically corresponds to undergoing a Hopf bifurcation (Theorem 4.5). (Recall
that generically means that the exceptional set has zero measure. So, we will show that the
subset of the surface corresponding to non-Hopf points has dimension at most 1.)
10
-----
To simplify computations, we assume that dissociation (backward-reaction) constants are
equal: k2 = k6 = k9. In chemistry, the forward reaction is usually more thermodynamically
favorable than the backward reaction. Therefore, the rate constant of a forward reaction is
much larger than the rate constant of the backward reaction [2]. We choose small values for
the dissociation rate constants in Section 5, similar to what was done in [12].
Proposition 4.1. Consider the dynamical system (2) arising from the mixed-mechanism
network and any positive rate constants for which k2 = k6 = k9. Then:
1. Every compatibility class P contains a unique (positive) steady state x[∗].
2. Exactly one of the following holds:
(a) The unique steady state x[∗] in each compatibility class P is locally asymptotically
stable.
(b) In the space of total amounts {(Ktot, Ptot, Stot)} = R[3]>0[, which we identify with]
the space of compatibility classes P, a surface H defines the border between those P
whose unique steady state x[∗] is locally asymptotically stable and those P for which
x[∗] is unstable.
Proof. Item 1 follows from Proposition 3.1.
For item 2, let J denote the Jacobian matrix of the mixed-mechanism system (2), with
equal dissociation constants: k2 = k6 = k9 =: kb, evaluated at the parametrized steady state
χ(x1, x2, x6), from (7). The characteristic polynomial of J is:
p(λ) := det(λI − J) = λ[3](λ[6] + b1λ[5] + b2λ[4] + · · · + b6),
where the coefficients bi (displayed below) are rational functions in x1, x2, x6 and the ki’s. To
streamline reading we only give the complete numerator of b6 and b1. The full coefficients
can be found in the Mathematica file mixed coeffs charpoly kb.nb[2].
numerator(b6) = k1[2][k]3[2][k][4][(][k][10][ +][ k][7][)(][k][10][k][5][k][7][ +][ k][5][k][7][k][b][ +][ k][10][k][8][(][k][7][ +][ k][b][))][x][1][x][2]2 (11)
+ k1k10k3k4k7(k3 + kb)(k10k5k7 + k5k7kb + k10k8(k7 + kb))x2x6
+ k10[2] [k][4][k][5][k]7[2][k][8][(][k][3] [+][ k][b][)][2][x]6[2] [+][ k][1][k]10[2] [(][k][3] [+][ k][4][)][k][5][k]7[2][k][8][(][k][3] [+][ k][b][)][x][1][x]6[2]
+ k1k10k5k7(k10k4k7 + k3k4k7 + k10k3(k4 + k7))k8(k3 + kb)x2x[2]6
numerator(b5) = k1[2][k]3[2][k][4][(][k][10][ +][ k][7][)(][k][10][ +][ k][b][)(][k][7][ +][ k][b][)][x][1][x][2]2
+ k1k10k3k4k7(k10 + kb)(k3 + kb)(k7 + kb)x2x6 + . . .
numerator(b4) = k1k3k4(k10 + k7)(k10 + kb)(k3 + kb)(k7 + kb)x1x2 + . . .
numerator(b3) = . . . + k1[2][k][3]�k10[2] [(][k][7] [+][ k][b][) +][ k][7][k][b][(][k][3] [+][ k][4] [+][ k][7] [+][ k][b][)]
+ k10 �(k7 + kb)[2] + k3(2k7 + kb) + k4(2k7 + kb)�[�]x[2]1[x][2][ +][ . . .]
numerator(b2) = . . . + k1[2][k][3][(][k][7][k][b][ +][ k][10][(2][k][7][ +][ k][b][))][x][2]1[x][2][ +][ . . .]
numerator(b1) = k1k3(k7kb + k10(2k7 + kb))x1x2 + k10k7(k3 + kb)(k10 + k3 + k4 + k7 + 3kb)x6
+ k1k10k7(k3 + kb)x1x6 + k1k10k7(k3 + kb)x2x6 + k10k7(k5 + k8)(k3 + kb)x[2]6
2This file and others mentioned below are in the Supporting Information; see Appendix A.
11
-----
And for the denominators:
denominator(b6) = k10(kb + k3)k7
denominator(bi) = k10(kb + k3)k7x6, for i = 2, 3, 4, 5 .
As x1, x2, x6 and the ki are positive, thus b1, b2, . . ., b6 > 0 (in the aforementioned
Mathematica file, we checked the above numerators are sums of only positive monomials).
Recall that, due to the 3 conservation laws (3), the Jacobian matrix has rank 6, not 9.
Accordingly, the relevant Hurwitz matrix, namely, for p(λ)/λ[3], is as follows:
b1 1 0 0 0 0
b3 b2 b1 1 0 0
b5 b4 b3 b2 b1 1
0 b6 b5 b4 b3 b2
0 0 0 b6 b5 b4
0 0 0 0 0 b6
Consider the Hurwitz determinants. First det H1 = b1 > 0. The next 3 Hurwitz determinants are also positive:
numerator(det H2) = k1[3][k]3[2][(][k][7][k][b][ +][ k][10][(2][k][7][ +][ k][b][))][2][x][3]1[x][2]2
+ k1[3][k][10][k][3][k][7][(][k][3][ +][ k][b][)(][k][7][k][b][ +][ k][10][(2][k][7][ +][ k][b][))][x][3]1[x][2][x][6][ +][ . . .]
numerator(det H3) = k1[5][k]3[3][(][k][10][k][5][k][7] [+][ k][5][k][7][k][b] [+][ k][10][k][8][(][k][7] [+][ k][b][))(][k][7][k][b] [+][ k][10][(2][k][7] [+][ k][b][))][2][x]1[5][x][3]2[x][6] [+][ . . .]
numerator(det H4) = k1[7][k]3[4][(][k][10][k][5][k][7] [+][ k][5][k][7][k][b] [+][ k][10][k][8][(][k][7] [+][ k][b][))(][k][7][k][b] [+][ k][10][(2][k][7] [+][ k][b][))][2]
�k5k7(k3 + k4 + k7)kb + k10[2] [k][8][(][k][7][ +][ k][b][)+]
k10(k3 + k4 + k7)(k5k7 + k8(k7 + kb))�x[7]1[x][4]2[x][2]6 [+][ . . .]
where the denominators, which are positive, are, respectively:
denominator(det H2) = k10[2] [k]7[2][(][k][b][ +][ k][3][)][2][x][2]6
denominator(det H3) = k10[3] [k]7[3][(][k][b] [+][ k][3][)][3][x]6[3]
denominator(det H4) = k10[4] [k]7[4][(][k][b] [+][ k][3][)][4][x]6[4]
(We display only the leading terms of the polynomials; the complete polynomials together
with an algorithmic verification of positivity are in mixed Hi.nb.) The final Hurwitz determinant is det H6 = (b6)(det H5), and we saw that b6 > 0. So, by the Routh-Hurwitz criterion
(Proposition 2.2), the steady state χ(x1, x2, x6) is locally stable if and only if det H5 > 0.
Hence, the surface H that delineates the boundary between compatibility classes with
stable steady states vs. those with unstable steady states is defined by det H5 ◦ (φ ◦ χ)[−][1] = 0,
where φ ◦ χ is the parametrization of compatibility classes from Proposition 3.6. If H
intersects the positive orthant R[3]>0[, then case (b) of the proposition holds. Otherwise, if]
H ∩ R[3]>0 [=][ ∅][, then we claim that we are in case (a). To show this, we need to verify that]
det H5(x1, x2, x6) > 0 for some (x1, x2, x6) ∈ R[3]>0[. The denominator of det][ H][5][(][x][1][, x][2][, x][6][) is]
strictly positive:
denominator(det H5) = k10[5] [k]7[5][(][k][3][ +][ k][b][)][5][x][5]6[.]
12
-----
So we need only show that the numerator of det H5(x1, x2, x6) is strictly positive for some
(x1, x2, x6) ∈ R[3]>0[.]
To this end, we view this numerator as a polynomial in x1 (so the coefficients are rational
functions of x2, x6, and the ki’s):
�
numerator(det H5) = x[9]1[x][4]2
� k10k7x6(k3 + kb)
k3(k10(2k7 + kb) + k7kb) [+][ x][2]
+ (12)
�
k8x6
� k5
k8
� � k5
+ k8[2][x][2]6 α02 + α11 + α20
k8
�2[�]
� k5
α01 + α10
k8
� k5
k8
+ lower degree terms in x1,
k5 � k5
α03 + α12 + α21
k8 k8
2
�
+ α30
�3[��]
k8[3][x]6[3]
�
where the coefficients αij are sums of (many) positive monomials and are given in the file
mixed analyis H5N x1 LT.nb. Therefore (for fixed x2 and x6) when x1 is sufficiently large,
the expression (12) is positive, as desired.
The proof of Proposition 4.1 focused on the surface H defined by the equation det H5 ◦
(φ ◦ χ)[−][1] = 0. This surface sometimes meets the positive orthant R[3]>0[, and indeed we show]
that this is the case when certain relationships hold among the rate constants.
Theorem 4.2. Consider the dynamical system (2) arising from the mixed-mechanism network. Assume the positive rate constants satisfy k2 = k6 = k9 and the following inequality:
k10k3k4 − (k3 + k4)(k3 + k7)(k4 + k7) > 0 . (13)
If k5/k8 is sufficiently large, then there is a compatibility class P whose unique steady state
x[∗] is unstable.
Proof. Assume that the rate constants satisfy k2 = k6 = k9 =: kb and (13). By the proof
of Proposition 4.1, a steady state χ(x1, x2, x6) of the mixed-mechanism system (2) is locally
stable if and only if det H5(x1, x2, x6) > 0. We also saw in that proof that the denominator
of det H5(x1, x2, x6) is strictly positive for all (x1, x2, x6) ∈ R[3]>0[. So, by Proposition 2.2, it]
suffices to show that if k5/k8 is sufficiently large, then there exists (x[∗]1[, x][∗]2[, x][∗]6[)][ ∈] [R]>[3] 0 [such]
that the numerator of det H5(x[∗]1[, x][∗]2[, x][∗]6[) is strictly negative: this would show that the steady]
state x[∗] := χ(x[∗]1[, x]2[∗][, x]6[∗][) is unstable.]
To this end, view the numerator of det H5 as a polynomial in x2 with coefficients in
x1, x6, and the ki’s. It is a degree-9 polynomial in x2 of the following form (see the file
mixed analysis H5N x2 LT.nb):
k10k7(k3 + kb)
numerator(det H5) = k1[9] �α0x[3]6 [+][ α][1][x]6[2] [+][ α][2][x][6][ +][ α][3]� [�]x[5]1 [+] 1[x][6]
k3(k10(2k7 + kb) + k7kb)[x][4]
�
x[9]2
+ lower degree terms, (14)
where α0, . . ., α3 are rational functions in kb, k3, k4, k5, k7, k8, k10. These functions αi are
given in mixed analysis H5N x2 LT.nb.
13
-----
We now analyze α0, which has the following form (see mixed analysis H5N x2 LT.nb):
β0
, (15)
�k5
k8
�k5
k8
�
+ β3
3
�
+ β1
2
�
+ β2
�
α0 = k8[3]
�
�k5
k8
where each coefficient βi is a rational function in kb, k3, k4, k7, k10 (and hence does not depend
on k1, k5, or k8). In particular, β0 is the following polynomial:
β0 = − k1[9][k]3[5][k]7[3] [(][k][10][k][3][k][4][ −] [(][k][3][ +][ k][4][)(][k][3][ +][ k][7][)(][k][4][ +][ k][7][)) (][k][10][ +][ k][b][)][3][ (][k][7][k][b][ +][ k][10][(2][k][7][ +][ k][b][))][2][ .]
It follows that β0 < 0 when inequality (13) holds.
Thus, when (13) holds, then, by equation (15), the inequality α0 < 0 holds for k5/k8
sufficiently large. In this case, the cubic polynomial in x6 appearing in (14), and hence also
the coefficient of x[9]2 [in the numerator of det][ H][5][, will be negative for][ x][6] [sufficiently large.]
Hence, if we choose x1 := 1 (or any positive value) and x6 and x2 sufficiently large, then the
numerator of det H5 will be negative.
In the remainder of this section, we focus on the question of whether the surface H
consists of (at least generically) Hopf bifurcations. If so, this would imply that whenever a
steady state of the mixed-mechanism network switches from stable to unstable, we expect
it to undergo a Hopf bifurcation leading to oscillations. We begin our analyses of Hopf
bifurcations by giving a criterion for such bifurcations.
Proposition 4.3. Consider the dynamical system (2) arising from the mixed-mechanism
network and any positive rate constants with k2 = k6 = k9 and k10k3k4 − (k3 + k4)(k3 +
k7)(k4 + k7) > 0. Then there exists (x[∗]1[, x][∗]2[, x][∗]6[)][ ∈] [R]>[3] 0 [such that][ det][ H][5][(][x]1[∗][, x][∗]2[, x][∗]6[) = 0][ (in]
other words, φ ◦ χ(x[∗]1[, x][∗]2[, x][∗]6[)][ is on][ H][). Moreover, for such a vector][ (][x][∗]1[, x][∗]2[, x][∗]6[)][, the system]
undergoes a Hopf bifurcation with respect to x2 at the steady state χ(x[∗]1[, x][∗]2[, x][∗]6[)][ if and only if]
the following inequality holds:
d(numerator(det H5)|x1=x[∗]1[, x][6][=][x][∗]6[)]
|x2=x[∗]2 [̸][= 0][ .] (16)
dx2
Proof. Fix positive rate constants for which k2 = k6 = k9 and k10k3k4 −(k3 +k4)(k3 +k7)(k4 +
k7) > 0. By the proofs of Proposition 4.1 and Theorem 4.2, the function det H5 : R[3]>0 [→] [R]
takes both positive and negative values. So, as det H5 is continuous, det H5(x[∗]1[, x][∗]2[, x][∗]6[) = 0]
for some (x[∗]1[, x][∗]2[, x][∗]6[)][ ∈] [R]>[3] 0 [(by the intermediate-value theorem).]
Assume det H5(x[∗]1[, x][∗]2[, x][∗]6[) = 0. To see whether the steady state][ χ][(][x][∗]1[, x][∗]2[, x][∗]6[) is a Hopf]
bifurcation with respect to the parameter µ = x2, where the curve of steady states is x(µ) =
χ(x[∗]1[, µ, x][∗]6[) and][ µ][0][ =][ x][∗]2[, we use Proposition 2.3 (Yang’s criterion). Parts (i) and (ii) of that]
criterion hold for any steady state χ(x[∗]1[, x][∗]2[, x][∗]6[), because][ b][6] [=][ b][6][(][x][∗]1[, x][∗]2[, x][∗]6[)][ >][ 0, by (11),]
and also det Hi = det Hi(x[∗]1[, x][∗]2[, x][∗]6[)][ >][ 0 for][ i][ = 1][,][ 2][,][ 3][,][ 4 (from the proof of Proposition 4.1).]
Recall from the proof of Proposition 4.1 that the denominator of det H5 is strictly positive and
does not depend on x2; thus, we can focus on the numerator of H5. So, by Proposition 2.3,
χ(x[∗]1[, x][∗]2[, x][∗]6[) is a Hopf bifurcation with respect][ x][2][ if and only if (16) holds.]
14
-----
Remark 4.4. Given rate constants ki as in Proposition 4.3 for which there is a Hopf bifurcation, we can perturb slightly the rate constants involved in (13) (while maintaining the
equality k2 = k6 = k9) and preserve the existence of a Hopf bifurcation. Indeed, this assertion follows from Proposition 4.3 (inequality (16) is maintained under small perturbations
of the xi’s), the fact that simple roots of a polynomial depend continuously – in fact, infinitely differentiably – on the coefficients [28], and the fact that the inequality (13) defines
a (relatively) open set in the parameter space of the ki’s.
Under the hypotheses of Proposition 4.3, we expect that inequality (16) holds generically
on H. We will confirm this when the rate constants are those in Table 1 (Theorem 4.5).
The proof of Theorem 4.5 makes use of discriminants, which we now review. Consider
a degree-n, univariate polynomial f = cnx[n] + cn−1x[n][−][1] + · · · + c0 with coefficients ci ∈ C.
A multiple root of f is some x[∗] ∈ C for which (x − x[∗])[2] divides f or equivalently f (x[∗]) =
f [′](x[∗]) = 0. It is well-known that f has a multiple root in C if and only if a certain multivariate
polynomial in the ci’s, the discriminant, vanishes [15]. For instance, the discriminant of the
quadratic polynomial ax[2] + bx + c is the familiar expression b[2] − 4ac.
Theorem 4.5 (Hopf bifurcations of the mixed-mechanism network). Consider the dynamical
system (2) arising from the mixed-mechanism network and rate constants in Table 1. Let
H denote the surface, from Proposition 4.1, that defines the border between those P whose
unique steady state x[∗] is locally stable and those P for which x[∗] is unstable. Then H consists
generically of compatibility classes P whose unique steady state x[∗] undergoes a simple Hopf
bifurcation (with x2 as bifurcation parameter).
Proof. It is straightforward to check that the rate constants in Table 1 satisfy the inequality (13). Therefore, the surface H as in Proposition 4.1.2(b) exists, and is defined by
det H5 = 0, where H5 is the Hurwitz matrix (specialized to the rate constants in Table 1) as
in the proof of Proposition 4.1.
To prove that H consists generically of Hopf bifurcations, we use Proposition 4.3. That result states that χ(x[∗]1[, x][∗]2[, x][∗]6[) is a Hopf bifurcation with respect to][ x][2] [if and only if (][x][∗]1[, x][∗]2[, x][∗]6[)][ ∈]
H[′] \ S, where
H[′] := V>0(det H5) := �(x1, x2, x6) ∈ R[3]>0 [|][ det][ H][5][(][x][1][, x][2][, x][6][) = 0]�, and
� d(det H5|x1=x[∗]1[, x][6][=][x][∗]6[)] �
S := (x[∗]1[, x][∗]2[, x][∗]6[)][ ∈H][′] |x2=x[∗]2 [= 0] ⊆H[′] .
dx2
����
We have that H = φ ◦ χ(H[′]), and that the following subset of H consists of compatibility
classes whose unique steady state undergoes a simple Hopf bifurcation with x2 as bifurcation
parameter: φ ◦ χ(H[′] \ S). So, it suffices to show that dim(S) < dim(H[′]). Note that
dim(H[′]) ≥ 2, so we will show that dim(S) ≤ 1.
To this end, note that if (x[∗]1[, x][∗]2[, x][∗]6[)][ ∈S][, then][ x][∗]2 [is a multiple root of the univariate]
polynomial numerator(det H5)|x1=x[∗]1[, x][6][=][x][∗]6 [(this also uses the fact the denominator of det][ H][5][,]
which is 188956800000000000000x[5]6[, does not depend on][ x][2][). Thus, any (][x][∗]1[, x][∗]2[, x][∗]6[)][ ∈S]
satisfies D(x[∗]1[, x][∗]6[) = 0, where][ D][ is the discriminant of det][ H][5] [and][ H][5] [is viewed as a univariate]
polynomial in the variable x2. So, we have the map:
S → {(x1, x6) ∈ R[2] | D(x1, x6) = 0} =: D
(x1, x2, x6) �→ (x1, x6) .
15
-----
The preimage of any point of this map has size at most 4 (because numerator(det H5)|x1=x[∗]1[, x][6][=][x][∗]6
has degree 9, so it has at most 4 multiple roots).
Thus, to achieve our desired inequality (namely, dim(S) ≤ 1), we need only prove the
following claim: dim(D) ≤ 1 or, equivalently, the bivariate polynomial D is not the zero
polynomial. It suffices to show that D(1, 1) is nonzero, which in turn would follow if we can
show that the univariate, degree-9 polynomial numerator(det H5)|x1=x[∗]1[, x][6][=][x][∗]6 [=][ H][5][(1][, x][2][,][ 1)]
does not have a multiple root over C. Indeed, using Mathematica, we see that the numerator
of det H5(1, x2, 1) has 9 (distinct) complex roots:
−131.425, − 102.999, − 78.022, − 66.423, − 39.194, − 3.946 ± 0.734i, − 3.677, 268.606 .
Thus, D is a nonzero polynomial, and this completes the proof.
(a) Stot = 40.
(b) Ptot = 5.
(c) Ktot ≈ 13.0296.
Figure 3: Slices of the Hopf-bifurcation surface H, from Theorem 4.5. Specifically, displayed
are the intersections of H with the hyperplanes defined by (a) Stot = 40, (b) Ptot = 5, and
(c) Ktot ≈ 13.0296. Each such curve was obtained numerically, using Matcont [8], by a
two-parameter continuation of the Hopf bifurcation arising from Ktot ≈ 13.0296, Ptot = 5,
and Stot = 40. Each point of the curves in (a) – (c) corresponds to a Hopf bifurcation with
respect to either of the two varying total concentrations. Points “inside” H correspond to
unstable steady states and thus the potential for oscillations.
In Figure 3, we show some slices of the Hopf-bifurcation surface H (where the rate
constants are from Table 1). Accordingly, this figure extends the one-dimensional Figure 1.
The bifurcations analyzed in Proposition 4.3 and Theorem 4.5 are with respect to the
bifurcation parameter x2, the steady-state value of the kinase K. It is natural to ask whether
we also obtain a bifurcation with respect to a more biologically meaningful parameter, such
as a rate constant or a total amount. We now explain how to perform such an analysis.
To use a total amount (here we use Ptot) as a bifurcation parameter (perturbing this
parameter corresponds to perturbing the compatibility class), consider the following maps:
{(Ktot, Ptot, Stot)} = R[3]>0 ←−φ◦χ R[3]>0 h5:=det−→ H5 R>0
Recall that (φ ◦ χ) : R[3]>0 [→] [R]>[3] 0 [is a bijection. Let][ g][ :=][ h][5] [◦] [(][φ][ ◦] [χ][)][−][1][ :][ R]>[3] 0 [→] [R][. Also, let]
p := (φ ◦ χ)2 = x6 + [1009]1800 [x][1][x][2][ denote the second coordinate function of][ φ][ ◦] [χ][ from (9) (here]
we assume the rate constants from Table 1). We are interested in checking whether ∂g
∂Ptot [is]
16
-----
(generically) nonzero whenever g = 0. Accordingly, we use the chain rule:
∂g 1
=
∂Ptot ∂p/∂x1
1800
=
1009x2
∂h5 1800
+
∂x1 1009x1
∂h5 1
+
∂x1 ∂p/∂x2
∂h5 1
+
∂x2 ∂p/∂x6
∂h5
∂x6
∂h5
+ [∂h][5] . (17)
∂x2 ∂x6
For specific values of x1, x2, x6, it is straightforward to check whether the sum (17) is nonzero.
More generally, we expect this sum to be generically nonzero; that is, we expect that the
surface H consists generically of Hopf bifurcations with respect to the total-amount Ptot.
### 5 Generating rate constants admitting oscillations
The proof of Theorem 4.2 yields a recipe for generating rate constants for the mixedmechanism network at which we expect oscillations arising from a Hopf bifurcation. Specifically, we choose rate constants ki for which the equalities k2 = k6 = k9 hold, the inequality (13) holds, and α0 < 0 (as in (15)), and then pick x2 and x6 large enough so that det H5
is negative but close to 0. We summarize these choices in the following procedure.
Procedure 5.1 (Generating rate constants likely to admit oscillations).
Input: The following functions[3]:
(i) α0 as in (15),
(ii) the numerator of det H5,
(iii) q := α0x[3]6 [+][ α][1][x]6[2] [+][ α][2][x][6][ +][ α][3][ as in][ (14)][, and]
(iv) φ ◦ χ given in Proposition 3.6.
Output: Rate constants and total amounts for which det H5 is negative and close to 0.
Steps:
1. Choose positive values for kb := k2 = k6 = k9, x1, k1, k3, k4, k7, and k8.
2. Choose a positive value for k10 for which k10 > [(][k][3][+][k][4][)(][k]k[3]3[+]k[k]4[7][)(][k][4][+][k][7][)] .
3. Choose the remaining rate constant k5 such that α0 < 0.
4. Choose x6 so that q < 0.
5. Choose x2 so that the numerator of det H5 is negative but close to 0.
6. Return the ki’s and (Ktot, Ptot, Stot) := φ ◦ χ(x1, x2, x6), where φ ◦ χ is evaluated at
the ki’s (and x1, x2, x6) chosen in the previous steps.
Remark 5.2. Using the output of Procedure 5.1, one can attempt to exhibit and analyze
oscillations or Hopf bifurcations using software, e.g., Matcont [8]. See Figure 4.
3The functions are provided as a text file in the Supporting Information. See Appendix A.
17
-----
Example 5.3. We follow Procedure 5.1 as follows (to verify our computations see the file
mixed generate rc.nb):
Step 1. We pick kb = 0.143738, k1 = 0.575284, k3 = 3.89096, k4 = 5.05386, k7 = 9.25029,
k8 = 0.621813, and x1 = 5.82148.
Step 2. The inequality for this step evaluates to k10 > 85.5048, so we choose k10 = 90.
Step 3. Evaluating α0 at the chosen ki’s, we obtain the following inequality:
−8.896 × 10[17]k5[3] [+ 1][.][49735][ ×][ 10][20][k]5[2] [+ 4][.][79701][ ×][ 10][20][k][5] [+ 2][.][42695][ ×][ 10][20][ <][ 0][,]
which we find, using Mathematica, is feasible for k5 > 171.471. So, we pick k5 = 172.
Step 4. By evaluating q at the values chosen above, we obtain the following inequality:
−1.41683 × 10[22]x[3]6 [−] [3][.][5508][ ×][ 10][25][x]6[2] [−] [1][.][80374][ ×][ 10][25][x][6] [+ 2][.][15078][ ×][ 10][24][ <][ 0][ .]
This inequality holds when x6 > 0.0996797, so we choose x6 = 0.1.
Step 5. By evaluating the numerator of det H5, we obtain the following inequality:
− 5.42893 × 10[25]x[9]2 [−] [4][.][20944][ ×][ 10][29][x]2[8] [−] [5][.][05393][ ×][ 10][31][x]2[7] [−] [6][.][67609][ ×][ 10][32][x]2[6]
+ 4.66164 × 10[33]x[5]2 [+ 3][.][97617][ ×][ 10][34][x]2[4] [+ 1][.][01289][ ×][ 10][35][x]2[3] [+ 1][.][19894][ ×][ 10][35][x]2[2]
+ 6.7831 × 10[34]x2 + 1.4718 × 10[34] < 0 .
This inequality is feasible, as computed in Mathematica, for x2 > 9.0382; we pick x2 = 10.
Step 6. We have determined the following rate constants:
k1 k2 k3 k4 k5 k6 k7 k8 k9 k10
0.575284 0.143738 3.89096 5.05386 172 0.143738 9.25029 0.621813 0.143738 90
We obtain the following steady state, using (7):
(x1, x2, . . ., x9) = χ(x1, x2, x6) (18)
= (5.82148, 10, 8.30052, 6.39056, 1.90691, 0.1, 3.49146, 520.229, 0.358855) .
Using this steady state, we obtain the total amounts, using (8):
(Ktot, Ptot, Stot) = φ(x1, x2, . . ., x9) = (24.6911, 3.95031, 546.499) . (19)
The resulting bifurcation analysis is shown in Figure 4.
### 6 Dynamics: simulations and conjectures
Are oscillations the norm when the mixed-mechanism system has an unstable steady state?
We conjecture that this is the case.
Conjecture 6.1. Consider the mixed-mechanism network, and any choice of rate constants
and total amounts. If the unique steady state in P is unstable, then P contains a stable
periodic orbit.
18
-----
(a) Bif. parameter Ktot.
0 10 20 30 40
(b) Bif. parameter Ptot.
(c) Bif. parameter Stot.
Figure 4: Numerical continuation of the steady state (18), when total amounts are as
in (19): (a) A (supercritical) Hopf bifurcations are at Ktot ≈ 24.0623 and 107.5635. (b)
(Supercritical) Hopf bifurcations are at Ptot ≈ 4.1022 and Ptot ≈ 2.3275. Matcont reported
a branch point, the leftmost red circle, at Ptot ≈−8.5427 × 10[−][13], i.e., for Ptot ≈ 0 and thus
outside the domain of interest. (c) A (supercritical) Hopf bifurcation is at Stot ≈ 288.4384.
(a) x5 vs. t.
(b) x5 vs. x2.
(c) Increasing Ktot.
Figure 5: Numerical verification of oscillations in the mixed-mechanism system with rate
constants as in Table 1. For (a) and (b), we used (K, P, S, 5, 40) and
tot tot tot) = (14
initial values as in (10). Here the solution converges to a periodic orbit. For (c), we used
(P, S, 40) and three values for K
tot tot) = (8 tot (namely, 100, 1000, and 10000), and again
initial values as in (10), except that x5 = 1.1. Again the solutions seem to converge to a
periodic orbit, and moreover this periodic orbit appears not to depend on the value of Ktot.
See Conjecture 6.2.
Some simulations are shown in Figure 5. In (A) and (B) of that figure, we see solutions
converging to a period orbit; this system arises from total-amounts similar to those that
Suwanmajo and Krishnan found to support oscillations. In contrast, in Figure 5(C), we
see oscillations, when (Ptot, Stot) = (8, 40), for three large values for Ktot: 100, 1000, and
10000. Oscillations persist across these values, which yields a much larger range for Ktot
than Suwanmajo and Krishnan’s results would suggest.
Moreover, the value of Ktot appears not to affect the resulting periodic orbit (when
projected to x5, the concentration of the doubly phosphorylated substrate S2). Could this
be a biological design mechanism for robust timekeeping (for instance, in circadian clocks)?
Mathematically, we conjecture that oscillations indeed persist for arbitrarily large Ktot; and,
that the periodic orbit in x5 indeed does not depend on Ktot.
Conjecture 6.2.
1. Consider the mixed-mechanism network with rate constants as in Table 1. Then there
exist values of Ptot and Stot such that for Ktot arbitrarily large, the unique steady
state in P is unstable.
19
-----
2. For such values of Ptot and Stot and for sufficiently large Ktot, the compatibility class
P contains a periodic orbit such that this orbit in x5 (the concentration of S2) does not
depend on the value of Ktot.
One way to tackle Conjecture 6.2 is to analyze the robustness of the period and the
amplitude with respect to Ktot using the theory developed in [3, 24, 23].
Finally, we consider the dynamics in compatibility classes that contain a locally stable
steady state. Our simulations suggest that such a steady state is in fact globally stable.
Accordingly, we pose the question, Consider the mixed-mechanism network, and any choice
of rate constants and total amounts. If the unique steady state x[∗] in P is locally stable, does
it always follow that x[∗] is globally stable? In the Michaelis-Menten limit, this is true [36].
### 7 Discussion
We return to the question, How do oscillations emerge in phosphorylation networks? Concretely, we would like (1) easy-to-check criteria for exactly which phosphorylation networks
admits oscillations or Hopf bifurcations, and (2) for those networks that admit oscillations,
a better understanding of the “geography of parameter space”, that is, a characterization
of which rate constants and initial conditions yield oscillations. Both of these problems are
still unresolved, and the second problem in particular is very difficult.
Nevertheless, here we made progress on characterizing some of the geography of parameter space for the mixed-mechanism phosphorylation network. Indeed, we found that a single
surface defines the boundary between stable and unstable steady states, and this surface
consists generically of Hopf bifurcations. Hence, when a steady state switches from stable
to unstable, then we expect it to undergo a Hopf bifurcation leading to oscillations. Additionally, we gave a procedure for generating many parameter values leading to oscillations.
We now discuss the significance of our work. At a glance, it might seem that our results
are specific to network (1) and rate constants related to those in Table 1. However, the
approach is general: for other rate constants (e.g., estimated from data) or other networks
(e.g., a version of the ERK network from [37] also has oscillations and a unique steady state),
one could apply the same techniques. Therefore, the potential impact is broad.
Going forward, we hope that the novel techniques we used – specifically, using a steadystate parametrization together with a Hopf-bifurcation criterion – will contribute to solving
other problems. For instance, we expect that such tools could help solve an important open
problem in this area [7], namely, the question of whether oscillations or Hopf bifurcations
arise from the fully distributive phosphorylation network.
#### Acknowledgements
AS was partially supported by the NSF (DMS-1312473/1513364 and DMS-1752672) and the
Simons Foundation (#521874). AS thanks Alan Rendall and Jonathan Tyler for helpful
discussions. CC was partially supported by the Deutsche Forschungsgemeinschaft DFG
(DFG-284057449). The authors two referees for their helpful suggestions.
20
-----
### A Files in the Supporting Information
[The following files can be found at http://www.math.tamu.edu/~annejls/mixed.html:](http://www.math.tamu.edu/~annejls/mixed.html)
Text files:
- mixed H5N kb.txt . . . contains H5N, the numerator of det H5 under the assumption
k2 = k6 = k9 = kb
- mixed W.txt . . . contains a matrix W that defines (3)
- mixed xt.txt . . . contains xt, the parameterization (7)
- mixed Jx.txt . . . contains Jx, the Jacobian evaluated at the parameterization (7)
Mathematica Notebooks:
- mixed analysis H5N x1 LT.nb:
Functionality: This file can be used to obtain numerator(det H5) as in (12), in particular
to examine the coefficients α01, α10, . . .
Input: the file mixed H5N kb.txt
- mixed analysis H5N x2 LT.nb:
Functionality: This file can be used to obtain numerator(det H5) as in (14), in particular
to examine the coefficients α0, . . ., α3 and β0, . . ., β3.
Input: the file mixed H5N kb.txt
- mixed coeffs charpoly.nb:
Functionality: This file can be used to obtain the characteristic polynomial of the
Jacobian of the system (2). It contains the Mathematica commands to establish bi > 0.
Input: the file mixed Jx.txt
- mixed Hi.nb:
Functionality: This file can be used to obtain the determinants of the Hurwitz matrices
H2, . . ., H5. It contains the Mathematica commands to establish det Hi > 0, for i = 2,
3, 4 and that det H5 is of mixed sign.
Input: the file mixed Jx.txt
- mixed generate rc.nb:
Functionality: This file contains a realization of Procedure 5.1.
Input: the files mixed H5N kb.txt, mixed W.txt, mixed xt.txt, mixed Jx.txt.
### References
[1] Kazuhiro Aoki, Masashi Yamada, Katsuyuki Kunida, Shuhei Yasuda, and Michiyuki
Matsuda, Processive phosphorylation of ERK MAP kinase in mammalian cells, P. Natl.
Acad. Sci. USA 108 (2011), no. 31, 12675–12680.
[2] Peter Atkins, Julio De Paula, and James Keeler, Atkins’ physical chemistry, Oxford
University Press, 2018.
21
-----
[3] EG Bure and Ye N Rozenvasser, On investigations of autooscillating system sensitivity,
Avtomat. i Telemekh (1974), no. 7, 9–17.
[4] Carsten Conradi, Elisenda Feliu, Maya Mincheva, and Carsten Wiuf, Identifying parameter regions for multistationarity, PLoS Comput. Biol. 13 (2017), no. 10, e1005751.
[5] Carsten Conradi and Maya Mincheva, Catalytic constants enable the emergence of bistability in dual phosphorylation, J. R. Soc. Interface 11 (2014), no. 95.
[6] Carsten Conradi and Anne Shiu, A global convergence result for processive multisite
phosphorylation systems, B. Math. Biol. 77 (2015), no. 1, 126–155. MR 3303108
[7], Dynamics of post-translational modification systems: recent progress and future
challenges, Biophys. J. 114 (2018), no. 3, 507–515.
[8] Annick Dhooge, Willy Govaerts, and Yuri A. Kuznetsov, MATCONT: A MATLAB
package for numerical bifurcation analysis of ODEs, ACM Trans. Math. Softw. 29
(2003), no. 2, 141–164.
[9] Mirela Domijan and Markus Kirkilionis, Bistability and oscillations in chemical reaction
networks, J. Math. Biol. 59 (2009), no. 4, 467–501.
[10] Mitchell Eithun and Anne Shiu, An all-encompassing global convergence result for processive multisite phosphorylation systems, Math. Biosci. 291 (2017), 1–9.
[11] Hassan Errami, Markus Eiswirth, Dima Grigoriev, Werner M. Seiler, Thomas Sturm,
and Andreas Weber, Detection of Hopf bifurcations in chemical reaction networks using
convex coordinates, J. Comput. Phys. 291 (2015), 279–302.
[12] James E Ferrell and Sang Hoon Ha, Ultrasensitivity part II: multisite phosphorylation,
stoichiometric inhibitors, and positive feedback, Trends Biochem. Sci. 39 (2014), no. 11,
556–569.
[13] Feliks R. Gantmacher, Matrix theory, Chelsea, New York 21 (1959).
[14] Karin Gatermann, Markus Eiswirth, and Anke Sensse, Toric ideals and graph theory
to analyze Hopf bifurcations in mass action systems, J. Symbolic Comput. 40 (2005),
no. 6, 1361–1382.
[15] I.M. Gelfand, M.M. Kapranov, and A.V. Zelevinsky, Discriminants, resultants and multidimensional determinants, Birkh¨auser, 1994.
[16] John Guckenheimer and Philip Holmes, Nonlinear oscillations, dynamical systems, and
bifurcations of vector fields, vol. 42, Springer Science & Business Media, 2013.
[17] Jeremy Gunawardena, Multisite protein phosphorylation makes a good threshold but can
be a poor switch, P. Natl. Acad. Sci. USA 102 (2005), no. 41, 14617–14622.
22
-----
[18] Otto Hadaˇc, Frantiˇsek Muzika, Vladislav Nevoral, Michal Pˇribyl, and Igor Schreiber,
Minimal oscillating subnetwork in the Huang-Ferrell model of the MAPK cascade, PLOS
ONE 12 (2017), no. 6, 1–25.
[19] Juliette Hell and Alan D. Rendall, A proof of bistability for the dual futile cycle, Nonlinear Anal.-Real 24 (2015), 175–189.
[20] Zoe Hilioti, Walid Sabbagh, Saurabh Paliwal, Adriel Bergmann, Marcus D Goncalves,
Lee Bardwell, and Andre Levchenko, Oscillatory phosphorylation of yeast Fus3 MAP kinase controls periodic gene expression and morphogenesis, Curr. Biol. 18 (2008), no. 21,
1700–1706.
[21] Huizhong Hu, Alexey Goltsov, James L Bown, Andrew H Sims, Simon P Langdon,
David J Harrison, and Dana Faratian, Feedforward and feedback regulation of the MAPK
and PI3K oscillatory circuit in breast cancer, Cell. Signal. 25 (2013), no. 1, 26–32.
[22] Wolfram Research, Inc., Mathematica, Version 11.3, Champaign, IL, 2018.
[23] Brian Ingalls, Maya Mincheva, and Marc R. Roussel, Parametric sensitivity analysis
of oscillatory delay systems with an application to gene regulation, B. Math. Biol. 79
(2017), no. 7, 1539–1563.
[24] Brian P Ingalls, Autonomously oscillating biochemical systems: parametric sensitivity
of extrema and period, Systems biol. 1 (2004), no. 1, 62–70.
[25] Matthew D. Johnston, Translated chemical reaction networks, B. Math. Biol. 76 (2014),
no. 6, 1081–1116.
[26] Matthew D. Johnston, Stefan M¨uller, and Casian Pantea, A deficiency-based approach to parametrizing positive equilibria of biochemical reaction systems, Preprint,
arXiv:1805.09295 (2018).
[27] Wei Min Liu, Criterion of Hopf bifurcations without using eigenvalues, J. Math. Anal.
Appl. 182 (1994), no. 1, 250–256. MR 1265895
[28] German Lozada-Cruz, The simple application of the implicit function theorem, Boletin
de la Asociati´on Matem´atica Venezolana XIX (2012), no. 1.
[29] Stefan M¨uller, Elisenda Feliu, Georg Regensburger, Carsten Conradi, Anne Shiu, and
Alicia Dickenstein, Sign conditions for injectivity of generalized polynomial maps with
applications to chemical reaction networks and real algebraic geometry, Found. Comput.
Math. 16 (2016), no. 1, 69–97.
[30] Koji L. Ode and Hiroki R. Ueda, Design principles of phosphorylation-dependent timekeeping in eukaryotic circadian clocks, Cold Spring Harbor Perspectives in Biology
(2017).
[31] Parag Patwardhan and W. Todd Miller, Processive phosphorylation: Mechanism and
biological importance, Cell. Signal. 19 (2007), no. 11, 2218–2226.
23
-----
[32] Mercedes P´erez Mill´an and Alicia Dickenstein, The structure of MESSI biological systems, SIAM J. Appl. Dyn. Syst. 17 (2018), no. 2, 1650–1682.
[33] Mercedes P´erez Mill´an, Alicia Dickenstein, Anne Shiu, and Carsten Conradi, Chemical
reaction systems with toric steady states, B. Math. Biol. 74 (2012), no. 5, 1027–1065.
[34] Mercedes P´erez Mill´an and Adri´an G. Turjanski, MAPK’s networks and their capacity
for multistationarity due to toric steady states, Math. Biosci. 262 (2015), 125–137.
[35] Shodhan Rao, Global stability of a class of futile cycles, J. Math. Biol. 74 (2017), 709–
726.
[36], Stability analysis of the Michaelis–Menten approximation of a mixed mechanism
of a phosphorylation system, Math. Biosci. 301 (2018), 159 –166.
[37] Boris Y. Rubinstein, Henry H. Mattingly, Alexander M. Berezhkovskii, and Stanislav Y.
Shvartsman, Long-term dynamics of multisite phosphorylation, Mol. Biol. Cell 27
(2016), no. 14, 2331–2340.
[38] Carlos Salazar and Thomas H¨ofer, Multisite protein phosphorylation – from molecular
mechanisms to kinetic models, FEBS Journal 276 (2009), no. 12, 3177–3198.
[39] Thapanar Suwanmajo and J. Krishnan, Mixed mechanisms of multi-site phosphorylation, J. R. Soc. Interface 12 (2015), no. 107.
[40] Matthew Thomson and Jeremy Gunawardena, The rational parameterisation theorem
for multisite post-translational modification systems, J. Theoret. Biol. 261 (2009), no. 4,
626–636.
[41] Hwai-Ray Tung, Precluding oscillations in Michaelis-Menten approximations of dualsite phosphorylation systems, Preprint, arXiv:1712.03594 (2017).
[42] David M. Virshup and Daniel B. Forger, Keeping the beat in the rising heat, Cell 137
(2009), no. 4, 602–604.
[43] Xiaojing Yang, Generalized form of Hurwitz-Routh criterion and Hopf bifurcation of
higher order, Appl. Math. Lett. 15 (2002), no. 5, 615–621.
24
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/1809.02886, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "GREEN",
"url": "http://arxiv.org/pdf/1809.02886"
}
| 2,018
|
[
"JournalArticle"
] | true
| 2018-09-08T00:00:00
|
[
{
"paperId": "2df819c0538469793a8389679b4f4bcf6f7145bf",
"title": "Design Principles of Phosphorylation-Dependent Timekeeping in Eukaryotic Circadian Clocks."
},
{
"paperId": "41a39996bcfe17f463f2d887eebbdec691bf3646",
"title": "A Deficiency-Based Approach to Parametrizing Positive Equilibria of Biochemical Reaction Systems"
},
{
"paperId": "2af19e1686fc089448c4d76d70ebaa80c4847e8c",
"title": "Stability analysis of the Michaelis-Menten approximation of a mixed mechanism of a phosphorylation system."
},
{
"paperId": "11019f340361bdc0d84b5e772042182fe9c51caa",
"title": "Precluding oscillations in Michaelis-Menten approximations of dual-site phosphorylation systems."
},
{
"paperId": "9174452c4ee3659b87f47fcd650cf86cae211321",
"title": "BioVis Explorer: A visual guide for biological data visualization techniques"
},
{
"paperId": "65caa10cd354f4e1a15c47b66a4486528305c8e2",
"title": "Minimal oscillating subnetwork in the Huang-Ferrell model of the MAPK cascade"
},
{
"paperId": "213a212e2dcf3eb06c7b306925aaebab89b13d50",
"title": "Parametric Sensitivity Analysis of Oscillatory Delay Systems with an Application to Gene Regulation"
},
{
"paperId": "4a509bbee9455029501b3228ef435485bc1413a6",
"title": "Parametric Sensitivity Analysis of Oscillatory Delay Systems with an Application to Gene Regulation"
},
{
"paperId": "1ce02a8f82b4239ffb1f6b607cc09121ea58b4ad",
"title": "Dynamics of Posttranslational Modification Systems: Recent Progress and Future Directions."
},
{
"paperId": "75555f0a3d2b070de86aae488553f9629e2a1890",
"title": "The Structure of MESSI Biological Systems"
},
{
"paperId": "0245441bee85be067fe74346abfb55e0333bb123",
"title": "An all-encompassing global convergence result for processive multisite phosphorylation systems."
},
{
"paperId": "8e43cb607788a94fd1e18c995f2f27878a883178",
"title": "Identifying parameter regions for multistationarity"
},
{
"paperId": "7f7b8e55714cadfeff8ab27fb425668357428e7e",
"title": "Long-term dynamics of multisite phosphorylation"
},
{
"paperId": "5b64c964040a01b678909ec13840fe2de42199e9",
"title": "Global stability of a class of futile cycles"
},
{
"paperId": "e9d3210e8ab533a6f93ac1315ebb06c229b0a190",
"title": "Detection of Hopf bifurcations in chemical reaction networks using convex coordinates"
},
{
"paperId": "ac7c5862a54c47dc685a6f06487cb0063818ea9d",
"title": "Mixed mechanisms of multi-site phosphorylation"
},
{
"paperId": "0ce9439f2a73ffbc0158e962fbb327307d1848ba",
"title": "Ultrasensitivity part II: multisite phosphorylation, stoichiometric inhibitors, and positive feedback."
},
{
"paperId": "1f9a8259c0d2baa5b5ccb1a8a9c33070927bed37",
"title": "Catalytic constants enable the emergence of bistability in dual phosphorylation"
},
{
"paperId": "8911294959240cadf32a79549a594345be3fdcd9",
"title": "A Global Convergence Result for Processive Multisite Phosphorylation Systems"
},
{
"paperId": "2d77ee9289092ad82423fd5563fd085b21bcbb6b",
"title": "A proof of bistability for the dual futile cycle"
},
{
"paperId": "cc23f1466e05093ffc272cb286b6d5220828c28b",
"title": "MAPK's networks and their capacity for multistationarity due to toric steady states."
},
{
"paperId": "32aace90f177a3e07cf614cf078333bbfd1f80cc",
"title": "Sign Conditions for Injectivity of Generalized Polynomial Maps with Applications to Chemical Reaction Networks and Real Algebraic Geometry"
},
{
"paperId": "68add4e13463db5bf6c05998a6e2a1fccc1c9267",
"title": "Translated Chemical Reaction Networks"
},
{
"paperId": "2860502531c3d2e496c3a7b0e4264d3557a866d3",
"title": "Processive phosphorylation of ERK MAP kinase in mammalian cells"
},
{
"paperId": "92f8857732fe572154eba60a6bcb3d6fd10ed515",
"title": "Chemical Reaction Systems with Toric Steady States"
},
{
"paperId": "2a364c9811520cb77ca9f6512142e751bf8c2fc9",
"title": "The rational parameterization theorem for multisite post-translational modification systems."
},
{
"paperId": "7a85cdb668d5c22d3dad7715597a448bb398f262",
"title": "Bistability and oscillations in chemical reaction networks"
},
{
"paperId": "1da168c6ee12b4f250a658a79407032ba21da779",
"title": "Multisite protein phosphorylation – from molecular mechanisms to kinetic models"
},
{
"paperId": "0ad56a3a5d4955095f84f45ba1e77d01ff540b21",
"title": "Keeping the Beat in the Rising Heat"
},
{
"paperId": "3d4284a2375aa9e9112c481a026dc046fa28d9af",
"title": "Oscillatory Phosphorylation of Yeast Fus3 MAP Kinase Controls Periodic Gene Expression and Morphogenesis"
},
{
"paperId": "76f61eb3f5a3b52a6757bad7a6a2cec27db814eb",
"title": "Processive phosphorylation: mechanism and biological importance."
},
{
"paperId": "235b1fa2aa347949d653680d7753b388de74c21c",
"title": "Toric ideals and graph theory to analyze Hopf bifurcations in mass action systems"
},
{
"paperId": "ca6641651b3653d60280eb88776f7154b06964cb",
"title": "Multisite protein phosphorylation makes a good threshold but can be a poor switch."
},
{
"paperId": "d5066250f3c780d964083719ca7579ffada45556",
"title": "Autonomously oscillating biochemical systems: parametric sensitivity of extrema and period."
},
{
"paperId": "8b6f7d0b305d6cba163ef568cdc361ada56e6b7a",
"title": "MATCONT: A MATLAB package for numerical bifurcation analysis of ODEs"
},
{
"paperId": "b307ed62d5999744414a3b13f2a98b55a4acbcf3",
"title": "Generalized form of Hurwitz-Routh criterion and Hopf bifurcation of higher order"
},
{
"paperId": "9592c5f840a913b2ddb658fedddf591483259807",
"title": "Discriminants, Resultants, and Multidimensional Determinants"
},
{
"paperId": "b1018b0062afae20c467e86cb752d8aacfcc8686",
"title": "Criterion of Hopf Bifurcations without Using Eigenvalues"
},
{
"paperId": "2ce5b9b35b5941749cb614e72a69e7ee0de9184c",
"title": "NONLINEAR OSCILLATIONS, DYNAMICAL SYSTEMS, AND BIFURCATIONS OF VECTOR FIELDS (Applied Mathematical Sciences, 42)"
},
{
"paperId": "70e3da6c426ca384f78f77474cbbf00a436038a2",
"title": "Nonlinear Oscillations, Dynamical Systems, and Bifurcations of Vector Fields"
},
{
"paperId": null,
"title": "Mathematica, Version 11.3"
},
{
"paperId": null,
"title": "A (2016) Sign conditions for injectivity"
},
{
"paperId": "d9bfc83ce4955238d7725f56599b70b3db04e402",
"title": "Feedforward and feedback regulation of the MAPK and PI3K oscillatory circuit in breast cancer."
},
{
"paperId": "2913d69df8537296d587a1831d8debb5258dcaba",
"title": "A simple application of implicit function theorem"
},
{
"paperId": null,
"title": "Boletin de la Asociatión Matemática Venezolana Millán MP"
},
{
"paperId": "985d6e713796a9fd3526d2ed47d764d67ed445f3",
"title": "Matrix Theory"
},
{
"paperId": "bb51541c410466e28955ec9e92768e36d28c61f9",
"title": "Introduction to the Implicit Function Theorem"
},
{
"paperId": null,
"title": "Atkins' Physical Chemistry"
},
{
"paperId": null,
"title": "Rozenvasser, On investigations of autooscillating system sensitivity, Avtomat"
},
{
"paperId": null,
"title": "Gantmacher, Matrix theory, Chelsea"
},
{
"paperId": null,
"title": "Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations"
},
{
"paperId": null,
"title": "Return the k i 's and (K tot , P tot , S tot ) := φ • χ(x 1 , x 2 , x 6 ), where φ • χ is evaluated at the k i 's (and x 1"
},
{
"paperId": null,
"title": "Using the output of Procedure 5.1, one can attempt to exhibit and analyze oscillations or Hopf bifurcations using software"
},
{
"paperId": null,
"title": "< 0 , which we find, using Mathematica, is feasible for k 5 > 171.471. So, we pick k 5 = 172"
}
] | 20,839
|
en
|
[
{
"category": "Medicine",
"source": "external"
},
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
},
{
"category": "Geography",
"source": "s2-fos-model"
},
{
"category": "Environmental Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/02b398677c3144cbb12459c81f06a8c4b350e510
|
[
"Medicine",
"Computer Science"
] | 0.884237
|
Open data products-A framework for creating valuable analysis ready data
|
02b398677c3144cbb12459c81f06a8c4b350e510
|
Journal of Geographical Systems
|
[
{
"authorId": "1403960481",
"name": "Daniel Arribas-Bel"
},
{
"authorId": "2116121340",
"name": "Mark A. Green"
},
{
"authorId": "145192149",
"name": "Francisco Rowe"
},
{
"authorId": "39908645",
"name": "A. Singleton"
}
] |
{
"alternate_issns": [
"1435-5930"
],
"alternate_names": [
"J Geogr Syst"
],
"alternate_urls": [
"https://link.springer.com/journal/10109",
"http://www.springer.com/economics/regional+science/journal/10109"
],
"id": "7fac555d-41ef-410c-9c84-ed584a9cff7f",
"issn": "1026-7050",
"name": "Journal of Geographical Systems",
"type": "journal",
"url": "https://www.springer.com/economics/regional+science/journal/10109"
}
|
This paper develops the notion of “open data product”. We define an open data product as the open result of the processes through which a variety of data (open and not) are turned into accessible information through a service, infrastructure, analytics or a combination of all of them, where each step of development is designed to promote open principles. Open data products are born out of a (data) need and add value beyond simply publishing existing datasets. We argue that the process of adding value should adhere to the principles of open (geographic) data science, ensuring openness, transparency and reproducibility. We also contend that outreach, in the form of active communication and dissemination through dashboards, software and publication are key to engage end-users and ensure societal impact. Open data products have major benefits. First, they enable insights from highly sensitive, controlled and/or secure data which may not be accessible otherwise. Second, they can expand the use of commercial and administrative data for the public good leveraging on their high temporal frequency and geographic granularity. We also contend that there is a compelling need for open data products as we experience the current data revolution. New, emerging data sources are unprecedented in temporal frequency and geographical resolution, but they are large, unstructured, fragmented and often hard to access due to privacy and confidentiality concerns. By transforming raw (open or “closed”) data into ready to use open data products, new dimensions of human geographical processes can be captured and analysed, as we illustrate with existing examples. We conclude by arguing that several parallels exist between the role that open source software played in enabling research on spatial analysis in the 90 s and early 2000s, and the opportunities that open data products offer to unlock the potential of new forms of (geo-)data.
|
p g
**ORIGINAL ARTICLE**
## Open data products‑A framework for creating valuable analysis ready data
**Dani Arribas‑Bel[1] · Mark Green[1] · Francisco Rowe[1] · Alex Singleton[1]**
Received: 17 October 2019 / Accepted: 29 June 2021 / Published online: 20 October 2021
© The Author(s) 2021
**Abstract**
This paper develops the notion of “open data product”. We define an open data
product as the open result of the processes through which a variety of data (open and
not) are turned into accessible information through a service, infrastructure, analytics or a combination of all of them, where each step of development is designed to
promote open principles. Open data products are born out of a (data) need and add
value beyond simply publishing existing datasets. We argue that the process of adding value should adhere to the principles of open (geographic) data science, ensuring openness, transparency and reproducibility. We also contend that outreach, in
the form of active communication and dissemination through dashboards, software
and publication are key to engage end-users and ensure societal impact. Open data
products have major benefits. First, they enable insights from highly sensitive, controlled and/or secure data which may not be accessible otherwise. Second, they can
expand the use of commercial and administrative data for the public good leveraging
on their high temporal frequency and geographic granularity. We also contend that
there is a compelling need for open data products as we experience the current data
revolution. New, emerging data sources are unprecedented in temporal frequency
and geographical resolution, but they are large, unstructured, fragmented and often
hard to access due to privacy and confidentiality concerns. By transforming raw
(open or “closed”) data into ready to use open data products, new dimensions of
human geographical processes can be captured and analysed, as we illustrate with
existing examples. We conclude by arguing that several parallels exist between the
role that open source software played in enabling research on spatial analysis in the
90 s and early 2000s, and the opportunities that open data products offer to unlock
the potential of new forms of (geo-)data.
**Keywords Geographic data science · Open data · Open source**
- Dani Arribas‑Bel
D.Arribas-Bel@liverpool.ac.uk
1 Geographic Data Science Lab, Department of Geography and Planning, University
of Liverpool, Roxby Building, 74, Bedford St S., Liverpool L69 7ZT, UK
-----
**JEL Classification C55 · C63 · C80**
### 1 Introduction
In the current era of digital transformation, data are a central pillar of the global
economy and society. We have passed the point at which more data are being collected than can be physically stored (Lyman and Varian 2003; Gantz et al. 2007;
Hilbert and López 2011).[1] In addition to traditional forms of data, such as social
surveys and censuses, major technological innovations have enabled an explosion in
the generation, collection and use of new forms of data (Timmins et al. 2018). Networked sensors embedded in electronic devices, such as mobile phones, satellites,
vehicles, smart energy meters, computers, GPS trackers and industrial machines can
now sense, create and store data on locations, transactions, operations and people.
Social media, web search engines and online shopping platforms have also spurred
this data revolution by recording and storing users’ activity and personal information. Data are created as a by-product through interaction with these technological
systems. While they are often not designed for research purposes, they can bring
value for answering research questions (Timmins et al. 2018).
The world’s technological capacity to store, communicate and share information has significantly expanded. In 2018, companies worldwide were estimated to
have generated and stored an excess of 33 zettabytes,[2] seven exabytes of new data
(Cisco 2018). Networked sensor technology in the financial services, manufacturing, healthcare and media and entertainment industries was estimated to account for
48 percent of global data generation globally in 2018 (Cisco 2018). In July 2019,
66 percent (over 5 billion people) of the world’s population were estimated to use
mobile phones, 56 percent (over 4.3 billion) to be internet users, and 46 percent
(over 3.4 billion) to comprise active social media users, whose penetration is growing at over 7 percent at year (Hootsuite and We Are Social 2019).
Despite the growing volume and speed of data collection and storage, only a
small share are actually used. In 2019, a global study found that most organisations
analysed less than half of the data they collected (Splunk, 2019). In 2018, a similar
global survey estimated that 96% of all generated data in the engineering and construction industry goes unused (Snyder et al. 2018). In 2011, a small share of scientists from a survey of 1700 leading scientists reported to regularly use and analyse
1 Lyman and Varian (2003), for instance, estimated that 5 exabytes of new data generated through electronic channels, such as telephone, radio, television and the Internet were stored globally in 2002 but that
more than three times that amount (i.e. 18 exabytes) were produced and not stored. Gantz et al. (2007)
estimated that the amount of digital data created and replicated (255 exabytes) exceeded the storage
capacity available (246 exabytes) in 2007. Hilbert and López (2011) estimated that the global generalpurpose computing capacity -as a measure of the ability to generate and process data- grew at an annual
rate of 58 percent, while global storage capacity grew at an annual rate of 23 percent between 1986 and
2007.
2 One zettabyte is equivalent to 1021 bytes. To visualise this, it would take one billion 1 TB external hard
drives to store a zettabyte of data.
-----
large data sets (Science Staff 2011). Only 12 percent reported data sets exceeding
100 gigabytes and use data sets exceeding 1 terabyte (Science Staff 2011).
The low utilisation rate of data may be reflective of barriers to access, as well
as inability to process such vast quantities of information efficiently. Two key challenges involve privacy and confidentiality concerns, as well as the unstructured
nature of data production and storage (Hanson et al. 2011; Manyika et al. 2015).
Privacy and confidentiality concerns restrict access to data collected by companies
and government agencies. The frequency, detail and geographical granularity of
data being generated are unprecedented and therefore ensuring their privacy, confidentiality and integrity is critical. While legislation has been slow in responding
to the changing landscape of digital data, it is now evolving in this direction. Major
changes to ensure data protection and privacy were made to the EU General Data
Protection Regulation (GDPR) which came into effect in 2018. Innovative institutional arrangements, such as data collaboratives (Verhulst, Young and Srinivasan
2017; Klievink et al. 2018) or services (e.g. Consumer Data Research Centre in the
UK), have developed data sharing protocols and secure environments to facilitate
access to commercial and administrative data for research purposes.
New forms of data are often highly unstructured and messy. They are produced
in multiple formats, including videos, images and text; and, are stored in various
organisational structures. Data are often not random samples of populations and
are collected for specific administrative, business or operational purposes, and not
necessarily for research (Hand 2018; Meng 2018; Timmins et al. 2018). In their
original form, new forms of data are thus not readily usable limiting their applications. Significant data engineering is required, involving the use and design of specialised methods, software and expert knowledge, and linkage to other data sources
(Hand 2018). To our knowledge, no formal analytical framework has been developed to chart the critical data engineering processes to develop purposely-built data
products.
In this paper, we propose and develop the idea of Open Data Products (ODPs) as
a framework to transform raw data into Analysis Ready Data (Giuliani et al.2017;
Dwyer et al. 2018) and identify the key features that we contend of this framework.
We define an ODP as the final data outcome resulting from adding value to raw,
highly complex, unstructured and difficult-to-access data to address a well-defined
problem, and making the generated data output openly available. Thus, three fundamental components characterise an ODP: its insightful utility, value added and open
availability. We argue that an open data product has two major benefits. First, it enables developing insights from scattered, and/or highly sensitive, and/or controlled,
and/or secure data which may be difficult to gather and use, or may not be accessible
otherwise. Second, it expands the use of commercial and administrative data for the
public good leveraging on their high temporal frequency and geographic granularity.
We also contend that there is a compelling need for data open products as we experience the current data revolution. New, emerging data sources are unprecedented
in temporal frequency and geographical resolution, but they are large, unstructured,
fragmented and often expensive to assemble and possibly hard to access due to privacy and confidentiality concerns. By transforming raw (open or “closed”) data
into valuable open data products, new dimensions of human geographical processes
-----
can be captured and analysed. Ultimately, ODPs may provide valuable guidance to
develop appropriate policy interventions.
The paper is structured as follows: The next section defines and develops the idea
of ODPs detailing the core elements to developing ODPs. We outline a framework
that covers the initial conception of an ODP and moving through to developing and
disseminating a product. Following, we discuss some of the challenges involved in
the process of developing ODPs. The fourth section introduces some case studies
of ODP exemplars which highlight the potential offered through our framework.
Finally, we conclude the paper discussing the future potential of ODPs.
### 2 Defining open data products
Defining Open Data Products (ODPs) is challenging since their remit is wide and
incorporates several, diverse aspects. In some ways, they share several characteristics with traditional open data, as described in Kitchin (2014) or Janssen et al.
(2012). To the extent ODPs result in open data, they also share most of their main
benefits for society (Molloy 2011). It might be intuitive to assume that making
Open Data [3] available that were previously not accessible would constitute an ODP.
However, building ODPs is a broader project encapsulating frameworks for product development, such as data, delivery channels, transparent processes, etc. Indeed,
ODPs adhere to standard principles of product development (e.g. Bhuiyan 2011)
such as end user feedback or prioritising goals more than almost any other academic
output.
In this context, we define ODPs as:
The open result of transparent processes through which a variety of data (open
and not) are turned into accessible information through a service, infrastructure, analytics or a combination of all of them, where each step of development
follows open principles.
We argue that the key difference between ODP over purely Open Data is the
value added, which widens accessibility and use of data that would otherwise be
expensive or inaccessible. Components of an ODP might include sophisticated data
analysis to transform input data, digital infrastructure to host generated datasets, and
dashboards, interactive web mapping sites or academic papers documenting the process. They almost always merge together data and algorithms but this is not necessarily a requisite.
While we adhere to general open principles, we recognise not all steps of the process can (or even need to) be fully open. We also argue a need for hybrid approaches
that allow for closed data to be incorporated and opened up through the creation of
3 Open Data have numerous definitions but commonly refer to data that are released into the public
domain without restrictive licenses that prevent their reuse or inclusion in derivative products. Such data
are differentiated from free data (e.g. Twitter/Facebook API), that may be restricted in terms of access
limits, but also importantly in the purposes to which the data can be applied or used.
-----
ODPs (Singleton and Longley 2019). Such approaches are necessary for widening
access to information derived from sensitive data. The resulting product should be
released as open data; ideally too, the majority of the process that results in an ODP
should be open, and although it might not be possible to release every component of
an ODP, those related to infrastructure such as computer code, platforms and algorithms required to generate output data should be made available and transparent
(Peng 2011; Singleton et al. 2016). Akin to the argument in open-source software,
this is not only so that third parties re-run every step of the process again before
using the data, but also to build a reproducible environment of trust that contributes
to user adoption of the product’s outputs (Brunsdon and Comber 2020).
Although ODPs can take many forms and shapes, and hence differ greatly from
each other, we think providing a few examples can be useful to land an abstract term
in more practical settings. We will use two case studies that together embody differently but well both the ethos and the building blocks of ODPs: geodemographic
classifications and data generated around the COVID19 pandemic. Below we introduce each, and we will return to different aspects in the next section.
Geodemographic classifications are created with the aim of describing the most
salient characteristics of people and the areas where they live (Webber and Burrows
2018). There are various classifications spanning different countries and substantive uses across both the public and private sector (Singleton and Spielman 2014).
Geodemographic classifications combine diverse sources of publicly and privately
available data to generate insights about the behaviour of existing or prospective
customers, service users and citizens. Technically, a geodemographic classification
collates and combines disparate sources of data through a computational data reduction technique called cluster analysis that groups areas into a set of representative
clusters describing salient patterns based on their similarity across a wide range of
descriptive attributes.
Our second set of illustrations relate to the recent COVID-19 pandemic. The need
to respond rapidly and efficiently to the spread of the virus, to save lives and sustain
the economy, created intense demand for actionable data and information to feed
into responsive decision making. Despite the global scope of the pandemic, many
of the data generation, collection and processing systems originally in place were
national at most, but in many cases regional or local. To bridge the gap between the
available data and insight required, several researchers and organisations launched
efforts to develop open data products. These included, for example, consolidated
databases (e.g. Riffe et al. 2021) as well as ODPs derived from advanced analysis
(e.g. Paez et al. 2020).
### 3 The building blocks of open data products
In this section, we outline the key components of our proposed framework for developing ODPs. First, ODPs are born out of a need or problem that needs insight and
will inform many design choices. Once the need is clearly delineated, the ODP
process adds value to existing data in ways that help meet the original need. Adding value usually takes two forms: potentially complex transformations, fusion and
-----
abstraction of the data in what we call Open (Geographic) Data Science; and outreach activities to ensure the original need is addressed with the maximum impact.
Throughout these explanations, we illustrate key points with the geodemographics
and COVID-19 case studies introduced above.
**3.1 Identifying a problem in need of insight**
Inception of an ODP begins with the identification of a concept or idea to address
some problem that requires insight. Developing meaningful products often requires
thinking less about ‘what’ a product might be, and more about ‘who’ might use it
and what they would want to know. As such, identifying end users, understanding
opportunities for satisfying their needs and mapping such opportunities to what
is possible with the available data, skills and resources available can help to focus
ODPs, and maximise their relevance. We would like to highlight this stage is usually
followed in the research process (i.e. thinking about the “research question”), but
that is not always the case in processes that result in open data. In fact, several open
datasets are explicitly released as a side effect of the data existing for other purposes,
and their release does not always have a clear end goal. While this has sometimes
spurred several innovations (e.g. smartphone apps as a result of transit data made
available as open data), we want to stress that ODPs are most useful when designed
for a purpose and to further a goal.
This process can be independent, however if possible co-designing products can
be an effective approach. Co-design (or co-production) is the involvement of external partners within the research process to help create user-led and user-focused
products (Ostrom 1996). It is not clear what activities might be considered co-design
(Filipe et al. 2017), however this process does not necessarily have to be onerous.
Building trust through collaborations can help to ensure relevant and impactful
products (Klievink et al. 2018). Data or knowledge exchange can facilitate partnerships, as well as opening up new ODPs that often would otherwise have not been
made available. Developing partnerships (termed data collaboratives) is relevant
here which are cross-sector initiatives for sharing or developing new data products
that add value to the work undertaken by each actor in the collaboration (Klievink
et al. 2018).
The principles of co-design are not limited to the identification of the product. It
applies to each part of our framework, and understanding the end user needs is core
to designing a successful product.
Perhaps the clearest example of the importance of a problem needing insight can
be found in the recent pandemic. Understanding the uneven impact of the pandemic
on society requires information about how different demographic groups from a
wide variety of geographic contexts are affected. However, very few readily available datasets exist to understand the dynamics on the pandemic as it unfolds across
different countries and different age groups. To fill this gap, Riffe et al. (2021) introduce a global demographic database of COVID-19 cases and deaths, COVerAGEDB, enabling cross-country comparisons in the experiences of the pandemic.
-----
**3.2 Adding value**
The development of ODPs is not merely about making raw data available, as data
driven innovation is more than opening up availability and use of data (Klievink
et al. 2018). A key tenant of ODPs is to process, analyse and build on the original
data, resulting in analysis ready data[4] (see, for example, the collection introduced by
Zhu 2019). This enhances the value of the information and opportunities for insight.
The added value of ODPs can be achieved through numerous strategies, although
these should ideally be linked to the first step of the framework to maximise their
utility.
Development of new ODPs that extend the uses of existing data create value
through producing new information. Data analysis can extract useful information or
process data to create a new resource that demonstrates clear value added. Sources
that cannot be made available in their raw form (often due to disclosure control
or commercial sensitivity) can be made openly available through processing and
manipulating into new ODPs with data owner permission.
Improving usability of data can help increase access, particularly where data
acquisition is costly, hidden or publicly unavailable. It can be more salient when
data are already available. But utilising or processing the data requires advanced
quantitative skills to derive information, and bridging potential skills and knowledge
gaps can open up existing data to a much wider audience (Klievink et al. 2018). This
is pertinent for lay populations who, if ODPs are combined with interactive visualisations and resources, can engage with complex data in ways that might otherwise
be unavailable to them. In such cases, value is added through focusing on the needs
of the end users.
Matching or linking records can bring added value to existing databases or
resources. Data linkage is the process of merging two or more independent resources
or databases together based upon matching on a set of shared identifiers (Harron
et al. 2017). Given the inherent costs of producing resources or collecting new data
to investigate a research question, linking two or more existing sources together that
could not answer the question by themselves, but possess all of the necessary information between them may provide a more efficient solution (Harron et al. 2017).
Even where data linkage is not the priority, ODPs should be set up to allow future
linkage to other potential resources.
By generating analysis ready data, ODPs bridge the gap between useful but inaccessible data and user needs. In doing so, they unlock potential research findings
that derive from analysis that relies on them, and can feed into decision making that
encourage more evidence-based policy making. Geodemographics and other composite indices are an excellent example of adding value to existing datasets. These
approaches manage to leverage information from multiple data sources, deriving
4 The term “Analysis Ready Data” finds its origin in the remote sensing literature. We use it in this context because we believe the challenges and benefits of processing data before they are made available to
end-users are extend well beyond satellite imagery.
-----
summary measures of the latent information (Green et al. 2018; Vickers and Rees
2007) while preserving the confidentiality of the original data as required.
**3.3 Open (geographic) data science**
Various sources of new data forms are available in a “half-cooked” state (Spielman 2017). They are not available in a form that would be useful or accessible for
interested stakeholders. For instance, data such as open transport data are available
through convoluted processes (e.g. APIs) that non-technical audiences are not able
to easily access. Others, such as satellite imagery or air quality data, can be downloaded easily but their size, complexity and unstructured nature preclude wider use.
Yet, others, such as purchase records from retailers, exist but have restricted access.
Given the accidental nature of many of these data sources (Arribas-Bel 2014), few
undergo thorough quality assurance and assessments for bias, completeness and
statistical representativeness. This is an important feature which differentiates new
forms of data from traditional census and survey-based sources, for which there exist
reliable infrastructure and frameworks for analysis, publication and dissemination.
The “unfinished” nature of new forms of data is a key feature of Data Science
as a discipline. The explosion in the amount, variety and potential uses of new data
has created the need for an interdisciplinary field that combines elements from areas
such as statistics, computer science and information visualisation (Donoho 2017).
Several new forms of data are inherently spatial, so there have been calls to establish closer links between these disciplines and Geography through GISc (Singleton
and Arribas-Bel 2019), computational (Arribas-Bel and Reades 2018) and quantitative Geography (Arribas-Bel 2018).[5]This stage of the analysis has become increasingly sophisticated, increasingly with greater use of advanced algorithms and complex pipelines that transform data in useful ways. As an illustration, Stubbings et al.
(2019) developed a green space index by combining street-level imagery, state-ofthe-art deep learning techniques and hierarchical modelling. Dismissing this component of every data project as merely “data cleaning” involves several risks. It diminishes the credit awarded to a step that can crucially influence the final results, which
compels researchers to relegate this key task to short and vague descriptions that
obscure the steps undertaken, with clear implications for openness, transparency and
reproducibility of their research (Brunsdon 2016).
We consider it vital that the (Geographic) Data Science process embedded
in the generation of ODPs be as open and transparent as possible (Brunsdon
and Comber 2020). Three main reasons underpin this requirement. First, as for
open-source software (Raymond 1999), an open approach fosters collaboration,
pooling of resources and avoids duplicating efforts. Second, an open approach
involves an explicit recognition of the limitations of the datasets generated. Third,
5 We argue that, in this context, the term “Geographic Data Science” is more appropriate to capture the
set of practices that we want to refer to. For more details on the motivation, reasoning and justification,
in particular to how this term relates to more established ones such as GIScience or Geocomputation, we
refer the reader to Arribas-Bel and Reades (2018) and Singleton and Arribas-Bel (2019).
-----
**Fig. 1 The geographic data science stack**
an open approach represents a clear message to users about the commitment to
honesty and transparency by the ODP creator. This is an important element. The
code, packages and platforms used to create an ODP will usually be accessed
only by a small fraction of its users. However, the fact that they can be checked
contributes to build user trust, and ultimately to amplify the use and impact of
ODP by attracting a larger user base.
The open approach that we recommend to maximise the impact of ODPs operates at three layers of the (Geographic) Data Science process: analysis, methods
and infrastructure. Figure 1 shows an overview of what we term the Geographic
Data Science stack. The top layer involves specification of the steps taken to
transform the original input data into a final ODP, which we term ‘analysis’. In
this context, the growing usage of computer code in research allows for the full
documentation and evaluation of how products are developed (Brunsdon 2016).
An open approach requires that the code generating the final dataset from the initial one(s) is available in both machine and human readable form. An increasingly
popular format to meet this requirement within scientific communities is the computational notebook, such as Jupyter notebooks (Rule et al. 2019) or Rmarkdown
notebooks (Casado-Díaz et al. 2017; Koster and Rowe 2019). In cases where
commercial interest and copyright law prevents code sharing, so-called pseudo
code with enough detail to reproduce the steps can be an acceptable compromise.
Code released in the analysis stage should be specifically tailored to the development of the ODP. A good illustration of this approach is the Open SIMD pro[ject to expand on the Scottish Index of Multiple Deprivation (https://github.com/](https://github.com/TheDataLabScotland/openSIMD)
[TheDataLabScotland/openSIMD).](https://github.com/TheDataLabScotland/openSIMD)
The second layer involves _methods. More generalisable code to implement a_
technique that could be applied in different contexts is relegated to this level. In
this case, an open approach requires methods to be packaged as an open-source
software library and released following standard software engineering practices
(e.g. version control and continuous integration; Wolf, Oshan & Rey 2019). This
division between analysis-specific code in notebooks and more modular code
into packages avoids duplication of effort and increases the clarity with which
the analysis is presented. Both R (CRAN) and Python (Conda-forge) are good
-----
examples of community approaches to support packages; similarly, projects such
as scikit-learn (Pedregosa et al. 2011) or the Tidyverse federation of packages
(Wickham et al. 2019) are good illustrations of open source packages.
The third layer comprises _infrastructure. The growing complexity of modern_
software stacks and analysis pipelines requires open access to analysis and methods used, as well as the infrastructure on which the development of ODPs has been
based be transparently detailed. In this context, ODPs can borrow from several
advances in software development to make the data available. A prominent example
is containerisation, the technology underpinning projects like Docker or Singularity,
that allows to isolate the computational environment required to reproduce a set of
commands. The gds_env project (Arribas-Bel 2019) provides an illustration for the
case of GDS.
Full reproducibility may not always be possible or even desirable. For example,
sensitive input data may not be amenable for sharing due to disclosure risks. We
argue that as much of the process from start to finish should be made available,
especially when there are few barriers against it. The purpose of an ODP is to design
products that add value to existing data through opening up opportunities within
data that are messy or unable to be openly shared.
A good example of the value of open geographic data science can be found in
the geodemographics literature. Many of the original classifications were created by
the private sector, where full disclosure of the underlying methods and data input
is not always be possible given associated commercial sensitivity or intellectual
property. Such an approach has drawn criticism as being “black box” (Singleton and
Longley 2009). Arguably, this poses an acute issue for applications in the public
sector, especially where life outcomes are at stake (Longley 2005). Responding to
these concerns, there has been movement towards creating geodemographics that
are more open to scrutiny. Under the umbrella of Open Geodemographics, several
classifications that are fully reproducible have been created in countries such as the
UK (Vickers and Rees 2007; Gale et al. 2016; Martin et al. 2018) and US (Spielman
and Singleton 2015). In these instances, code and data are disseminated openly, and
these academic outputs also have associated journal articles in the peer reviewed
literature. Such an approach was made possible through all of the data integral to
these classifications being disseminated with open licences and enabling reuse and
redistribution.[6]More recent research also discusses alternative reproducible methods
that might also be applicable when data are sourced with wider and more restrictive
licensing arrangements where full reproducibility was not possible (Singleton and
Longley 2019).
6 For example, the 2011 ONS Output Area Classification has a formal page on the Office for National
[Statistics website here: https://www.ons.gov.uk/methodology/geography/geographicalproducts/areaclassi](https://www.ons.gov.uk/methodology/geography/geographicalproducts/areaclassifications/2011areaclassifications)
[fications/2011areaclassifications.](https://www.ons.gov.uk/methodology/geography/geographicalproducts/areaclassifications/2011areaclassifications)
-----
**3.4 Outreach**
The mantra ‘build it and they will come’ should not be the outcome of ODP development. Successful dissemination, circulation and impact should not rely on chance.
Outreach activities and resources are required to encourage end users to engage with
a product. These activities should be designed to guide end users on the use of the
ODP. A full review of various forms of outreach activities is beyond the scope of the
paper, we focus on two main dissemination channels. It is important to recognise
that several of these practices closely relate to and take inspiration from a variety
of literatures, including those of participatory GIS (Dunn 2007) and citizen science
(e.g. Haklay 2013).
A first key channel is user-focused events. These serve the purpose of refining
and promoting a product. They can involve small, focused events such as workshops
with stakeholders or lay community groups, and larger public promotion campaigns.
Online presence and social media can play an important role in accessing wider
coverage if supported with resources and materials. Project-specific social media
accounts and online presence are increasingly more common. For example, the
European Commission devotes an entire website to different aspects of their Global
Human Settlement open data product.[7]Partnerships can also assist in the outreach
process, especially when ODPs are designed to address a particular problem. For
example, the “Access to Healthy Assets and Hazards” project (AHAH, Green et al.
2018) partnered with Public Health England (PHE) to make some of the data available through PHE’s Public Health Profiles resource. Co-designing an ODP requires
engagement and co-development of project ideas with end users at every step so
that the impact of ODP is maximised. Singleton and Longley (2019) co-developed
a bespoke workplace classification in close collaboration with the Greater London
Authority (GLA). The ODP is now available openly through the Consumer Data
Research Centre’s data repository,[8], and the GLA is using it for internal operations.
A second major channel involves the use of open-source platforms, software and
resources. The integration of these assets is key to ensure interaction and engagement of end users with the ODPs, and a key principle is to facilitate end users with
non-technical skills to interact with ODPs. Data stores comprise a useful example to
make available ODPs and associated meta-data. Publishing all technical details, analytical code and documentation is important so that users can evaluate how ODPs
were created and refine the project pipeline (see Paez et al. 2020, for an example
of extensively documented data processes). Open-source platforms can help with
this process, for example, CKAN for publishing open data, or GitHub for sharing
code. Complementing these platforms should be the use of interactive resources
that improve the accessibility and usability of ODPs. Examples of this approach
include AHAH or the classification developed by Rowe et al. (2018) to analyse the
trajectory of socio-economic inequality at the neighbourhood level in UK. These
resources comprise an interactive web mapping tool that has been used by the
7 [https://ghsl.jrc.ec.europa.eu/datasets.php.](https://ghsl.jrc.ec.europa.eu/datasets.php)
8 [http://data.cdrc.ac.uk/dataset/london-workplace-zone-classification.](http://data.cdrc.ac.uk/dataset/london-workplace-zone-classification)
-----
general public and policy makers to point and click to their local areas and engage
with the resource, as well as allow more technical users to download and analyse the
information.
Journals have also emerged as a key mechanism for the explicit dissemination of ODPs. Innovative examples include Data in Brief,[9] Scientific Data[10] or
REGION,[11]which publish papers explicitly focused on ODPs rather than focused on
research from which a side product is an ODP. In doing so, they seek to promote
the creation, sharing and reuse of scientific data. Papers are peer reviewed and published under an open license. This form of publication is useful as it provides essential context, describing how ODPs have been generated as well as assessing their
limitations and identifying potential purposes for the reuse of generated ODPs (e.g.
Rowe et al. 2017), all elements hard to cover on a traditional research paper. Journals, such as REGION, have also started publishing computational notebooks, and a
key aim is their added value in communicating and disseminating ODPs (Koster and
Rowe 2019). Notebooks offer interactivity with the potential to engage policy, discipline-specific or local knowledge experts with data analysis exploration (Rowe et al.
2020). This in turn can enable the identification of new relevant patterns or uses that
may have not been reported or explicitly discussed in the original publication. These
novel ways of publication provide an incentive for researchers to generate ODPs.
Outreach does not mark the end of developing ODPs. It is a continual and circular process that should incorporate constant evaluation and refinements to a product.
Ideally, as data are updated, new relevant sources become available, and feedback
from end users is gathered, they should be incorporated to refine ODPs. Outreach
should therefore be designed to maximise this refinement process, facilitating feedback generation from relevant users.
Examples of outreach into stakeholders and users can be found in geodemographics. Spielman and Singleton (2015) and Patias et al. (2019) produced open classifications for the US and UK, respectively. Through further interaction, engagement
and outreach, the Location intelligence company Carto[12] has integrated them into
their portfolio of data offerings. For the initial release of the US classification, only
a description of the group level (ten clusters) was included, but Carto developed
new labels for the 55 cluster type level, making these available within the public
domain, alongside integration into their mapping platform,[13]used by industry and
government. Thanks to this effort, the original classifications are openly accessible
via their API and can be viewed within an interactive map improving their ease of
access, engagement and dissemination.
ODP development and outreach has also been instrumental in supporting
responses to the COVID-19 pandemic. For example, the Local Data Spaces project
9 [https://www.journals.elsevier.com/data-in-brief/.](https://www.journals.elsevier.com/data-in-brief/)
10 [https://www.nature.com/sdata/.](https://www.nature.com/sdata/)
11 [https://openjournals.wu-wien.ac.at/ojs/index.php/region/.](https://openjournals.wu-wien.ac.at/ojs/index.php/region/)
12 [https://carto.com.](https://carto.com)
[13 The Carto blog describing the work can be found here: https://carto.com/blog/demographic-clusters-](https://carto.com/blog/demographic-clusters-segmentation-data-observatory/)
[segmentation-data-observatory/.](https://carto.com/blog/demographic-clusters-segmentation-data-observatory/)
-----
in the UK saw researchers working with Local Government practitioners to co-produce data insights using data held in secure and centralised researcher data environments (Leech et al. 2021). The aim was to help Local Authorities access these data
directly or undertake research on their behalf, allowing them to gain data insights
from data they did not have access to (including timely COVID-19 data deposited by
the Office for National Statistics (ONS) not available elsewhere). Through continual repeated meetings with the team, researchers were able to co-design how Local
Authorities wanted ODPs shared. Short computational notebooks were one solution, embedding descriptive data analyses as ‘conversation starters’ to show what
data insights could be produced and help Local Authorities see the ‘art of possible’ (rather than sharing analysis ready data initially). For example, through sharing
notebooks mapping asymptomatic COVID-19 test site accessibility in Liverpool,
Liverpool City Council asked where to locate new sites and the team were able then
focus on generating optimised locations to improve access (Green 2021). The added
value of using notebooks meant that any analysis run for a Local Authority could be
replicated for any other the local area resulting in all Local Authorities benefitting
from insights during the co-production process.
### 4 Challenges
Open Data are a good example of a Public Good, being both “non-rivalrous” and
“non-excludable.” Open Data are, however, not free. There are direct costs associated with their collection, extraction, preparation and release; alongside indirect
costs such as the loss of potential income that might be realised through alternative licensing models (Singleton et al. 2016; Johnson et al. 2017). Moreover, their
consumption does not necessarily contribute to their production. For example, we
might use OpenStreetMap data and services, but never commit any new geographic
features or corrections to this open map system. Although some costs might be
argued as being written off over time, others remain in perpetuity such as the cost
of data hosting or download bandwidth. Issues of this nature which are associated
with Open Data are generally enhanced when they are productised, given the additional human resource burden required in their creation, and the generation of necessary meta data or reporting associated with their release, such as extensive technical briefings, or the preparation of linked academic publications. As with Open
Data, the “value” of an ODP is not realised directly (as it is free at the point of
use), and to balance production costs, this would only likely to occur if these enveloped accounting of indirect benefits. For example, within some sectors where funding may be limited, an ODP might replace limited or no insight; potentially returning various economic or social benefits. Where funding is less constrained, ODPs
may add value vis-a-vis commercial offerings if the insights generated are unique
or complementary (Johnson et al. 2017). Capturing such value in both instances is
however complex and lies somewhat outside the scope of this paper. However, given
the costs of Open Data and those additional burdens of ODPs, there does need to be
strategic planning and thought associated with creating ODPs. We would argue that
some strategies that have been adopted by the Open Source Software community
-----
might be applicable within the context of an ODP. This might include the sponsorship of ODPs by organisations who are benefiting from their availability, or the integration of ODPs into commercial software as a service platform (e.g. API). More
specifically, organisations developing ODPs, might also supply these within a ‘freemium’ model where enhanced versions of ODPs might be provided as commercial
offerings.
The creation of ODPs share similarities to those ways in which open source software are produced. It has been argued that major contributions to many open source
software packages are in fact mostly a result of contributions from a more limited set
of developers (Krishnamurthy 2005). In a similar vein, many ODPs are created as a
result of individuals or very focused teams. As with open software where there are a
narrow set of contributors, this creates a challenge for how ODPs can be maintained
and updated over time. Low diversity in teams developing Open Source Software
(OSS) has also been suggested to hinder creativity and productivity (Giuri et al.
2010), which we would also argue is applicable to ODPs. Given these issues with
OSS,one way in which they can be sustained is through code sharing platforms, such
as Github or Bitbucket, where new developers can find out about software, make
contributions or fork developments (Peng 2011). We argue that such platforms are
equally useful for the sharing of code and data associated with the development of
ODPs. However, they are not designed specifically for this purpose, and in essence,
features are repurposed from the software developer community. The size of data
that can be shared within such platforms is often limited, and where more extensive storage is required, this becomes an increasing cost burden. Although explicit
data sharing platforms have emerged (e.g. figshare.com, zenodo.org, datadryad.org,
dataverse.harvard.edu), these tend to focus on dissemination or archiving rather than
development. Such platforms are useful for the promotion of ODPs, but are limited
in functionality to support the process of remixing or update (Singleton et al. 2016).
We would argue that there is space for new platforms with features that are better
tailored to the needs of ODP development, and much like Github or Bitbucket might
reward users through public profiles detailing their contributions to different ODPs.
The extent to which any community of ODP developers might be formalised and
developed akin to those established within OSS will be challenging given the positioning of this emergent area (Harris et al. 2014; Arribas-Bel 2018; Arribas-Bel and
Reades 2018; Singleton and Arribas-Bel 2019). Such issues are accentuated within
our current university curricula. Within the Quantitative Social Sciences and Statistics, focus tends to favour theory and applications of statistical models. Although the
processes of software development are considered within Computer Science, these
focus on applications rather the use of code in development of ODPs. Moreover,
the recent rapid growth of Data Science has so far emphasised visualisation and
new modelling techniques from the cannon of machine learning and artificial intelligence. We argue that there is clearly a role for the better embedding of ODP development both within curricula bearing components of Data Science.
Finally, for those involved in the production of knowledge through research,
historically there would be limited value ascribed to the considerable extra efforts
required to package and document outputs from research as ODPs (Singleton et al.
2016). Within systems where impact is valued or measured, we argue that this might
-----
support engagement for the development of ODPs given their utility as a route to
stakeholder engagement.
### 5 Conclusion
This paper introduces the concept of Open Data Product as a construct that lowers
barriers for a wider audience of stakeholders to access and benefit from the (geo-)
data revolution. The value in framing the challenge of making sense of new forms of
data through ODPs resides in its comprehensive approach. We focus neither exclusively on technical issues, such as the current big data discourse; nor on governance and outreach solely, such as more traditional open data notions. Instead, ODPs
recognise that turning disparate, unstructured and often sensitive data sources into
useful and accessible information for a wider audience of stakeholders requires a
combination of computational, statistical and social efforts. In doing so, we contribute to the Open Data literature by providing a framework that expands the notion of
how Open Data can be generated and what can constitute the basis to generate open
datasets, as well as how to ensure its final usability and reliability.
Although not fully developed in this paper, we see a clear parallel between ODPs
and the role that open-source software played in democratising access to cutting
edge methods and computational power in the 90 s and 2000s. Three decades ago,
a series of technological advances such as the advent of personal computing and
rapid increase of computational power (e.g. Moore’s Law) provided fertile ground
for experimentation in the domain of spatial analysis. Initially, however, this field
of experimentation was hampered by a landscape dominated by proprietary software that was restrictive to access. Besides the obvious monetary cost, commercial
software restricted access to methodological innovations as it used to be oriented
to profitable market areas. In this context, OSS contributed significantly to unlock
much of the potential of new computers and helped spur an era of new research that
would have not been possible otherwise.[14]
We see data, rather than computation, as the defining feature of the present technological context. To make the most of new forms of data, we need more than “just”
OSS; hence the proposal for ODPs in this paper. However, we would also like to
stress the relevance and crucial role that OSS has to play in a world where “raw
data” are so distant from an “analysis ready data”. As highlighted above, ODPs can
only succeed through a transparent process that can build trust among end-users.
Without the ability that currently only OSS provides to access cutting-edge techniques and do so in a transparent way, it is difficult to imagine successful ODPs.
Rather than definitive, our hope for this paper is to be provocative. The current data
landscape is in transition and is very likely that several innovations are still in the notso-distant horizon. Hence, the notion of ODP will necessarily be an evolving one that
14 For a practical illustration of this statement, the reader is advised to examine the number of published
papers that actively cite open-source software projects such as GeoDa (Anselin, Syabri & Kho; 2006);
R’s spdep (Bivand et al. 2011); or PySAL (Rey & Anselin 2010).
-----
adapts to changing conditions to remain useful and valuable. At any rate, we envision
the need for novel approaches and mindsets such as those described in this paper only
to increase in the coming years. There is much that the spatial analysis community
holds to contribute to exploit the data deluge that is rapidly changing every aspect of
society. New ways to communicate and deliver our collective advances in data intelligence and expertise to maximise societal impact are needed. We hope the ideas presented in this paper partially shape the agenda and, more generally, contribute to a
wider conversation about our role in shaping this new world in the making.
**Open Access This article is licensed under a Creative Commons Attribution 4.0 International License,**
which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as
you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article
are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the
material. If material is not included in the article’s Creative Commons licence and your intended use is
not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission
[directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licen](http://creativecommons.org/licenses/by/4.0/)
[ses/by/4.0/.](http://creativecommons.org/licenses/by/4.0/)
### References
Anselin L, Syabri I, Kho Y (2006) GeoDa: an introduction to spatial data analysis. Geogr Anal
38(1):5–22
Arribas-Bel D (2014) Accidental, open and everywhere: emerging data sources for the understanding of
cities. Appl Geogr 49:45–53
Arribas-Bel D (2018) Statistics, modelling, and data science. In: Ash J, Kitchin R, Leszxzynski A (eds)
Digital geographies. Sage, London
Arribas-Bel D, Reades J (2018) Geography and computers: Past, present, and future. Geogr Compass
12(10):751
[Arribas-Bel D (2019) A containerised platform for Geographic Data Science https://github.com/darribas/](https://github.com/darribas/gds_env)
[gds_env](https://github.com/darribas/gds_env)
Bhuiyan N (2011) A framework for successful new product development. J Indus Eng Manag
4(4):746–770
Bivand R, Anselin L, Berke O, Bernat A, Carvalho M, Chun Y, and Lewin-Koh N (2011). spdep: Spatial
dependence: weighting schemes, statistics and models.
Brunsdon C (2016) Quantitative methods I: reproducible research and quantitative geography. Prog Hum
Geogr 40(5):687–696
Brunsdon C, Comber A (2020) Opening practice : supporting reproducibility and critical spatial data sci[ence. J Geogr Syst. https://doi.org/10.1007/s10109-020-00334-2](https://doi.org/10.1007/s10109-020-00334-2)
Casado-Díaz JM, Martínez-Bernabéu L, Rowe F (2017) An evolutionary approach to the delimitation of
labour market areas: an empirical application for Chile. Spat Econ Anal 12(4):379–403
Cisco V (2018) Cisco visual networking index: Forecast and trends, 2017–2022.White Paper 1(1)
Donoho D (2017) 50 Years of Data Science. J Comput Graph Stat 26(4):745–766
Dunn CE (2007) Participatory GIS—a people’s GIS? Prog Hum Geogr 31(5):616–637
Dwyer JL, Roy DP, Sauer B, Jenkerson CB, Zhang HK, Lymburner L (2018) Analysis ready data: enabling analysis of the Landsat archive. Remote Sens 10(9):1363
Filipe A, Renedo A, Marston C (2017) The co-production of what? Knowledge, values, and social relations in health care. PLoS Biol 15(5):2001403
Gale CG, Singleton AD, Bates AG, Longley P (2016) Creating the 2011 area classification for output
[areas (2011 OAC). J Spatial Inf Sci. https://doi.org/10.5311/JOSIS.2016.12.232](https://doi.org/10.5311/JOSIS.2016.12.232)
Gantz JF et al (2007) The expanding digital universe: a forecast of worldwide information growth through
2010. International Data Corporation (IDC)
-----
Giuliani G, Chatenoux B, De Bono A, Rodila D, Richard JP, Allenbach K, Peduzzi P (2017) Building an
earth observations data cube: Lessons learned from the Swiss data cube (SDC) on generating analysis ready data (ARD). Big Earth Data 1(1–2):100–117
Giuri P, Ploner M, Rullani F, Torrisi S (2010) Skills, division of labor and performance in collective
inventions: Evidence from open source software. Int J Ind Organ 28(1):54–68
Green MA (2021) Thinking spatially to communicate and evaluate the roll-out of ‘mass’ testing in Liverpool, 2020. People, Place Policy 15(1):54–56
Green MA, Daras K, Davies A, Barr B, Singleton A (2018) Developing an openly accessible multidimensional small area index of ‘access to healthy assets and hazards’ for great Britain, 2016.
Health Place 54:11–19
Haklay, M. (2013). Citizen science and volunteered geographic information: Overview and typology of
participation. In Crowdsourcing geographic knowledge (pp. 105–122). Springer, Dordrecht.
Hand DJ (2018) ‘Statistical challenges of administrative and transaction data.’ J R Stat Soc Ser a: Stat
[Soc 181(3):555–605. https://doi.org/10.1111/rssa.12315](https://doi.org/10.1111/rssa.12315)
Hanson B, Sugden A, Alberts B (2011) Making data maximally available. Science 331(6018):649.
[https://doi.org/10.1126/science.1203354](https://doi.org/10.1126/science.1203354)
Harris R, Tate N, Souch C, Singleton A, Orford S, Keylock C, Jarvis C, Brunsdon C (2014) Geographers
count: a report on quantitative methods in geography. Enhanc Learn Soc Sci 6(2):43–58
Harron K, Dibben C, Boyd J, Hjern A, Azimaee M, Barreto ML, Goldstein H (2017) Challenges in
[administrative data linkage for research. Big Data. https://doi.org/10.1177/2053951717745678](https://doi.org/10.1177/2053951717745678)
Hilbert M, López P (2011) The World’s technological capacity to store, communicate, and compute
[information. Science 332(6025):60–66. https://doi.org/10.1126/science.1200970](https://doi.org/10.1126/science.1200970)
[Hootsuite & We Are Social (2019) Digital 2019 Global Digital Overview. Available at: https://datareport](https://datareportal.com/reports/digital-2019-global-digital-overview)
[al.com/reports/digital-2019-global-digital-overview.](https://datareportal.com/reports/digital-2019-global-digital-overview)
Janssen M, Charalabidis Y, Zuiderwijk A (2012) Benefits, adoption barriers and myths of open data and
open government. Inf Syst Manag 29(4):258–268
Johnson PA, Sieber R, Scassa T, Stephens M, Robinson P (2017) The cost(s) of geospatial open data.
Transactions in GIS 21(3):434–445
Kitchin R (2014) The data revolution: big data, open data, data infrastructures and their consequences. Sage
Klievink B, van der Voort H, Veeneman W (2018) Creating value through data collaboratives. Informa[tion Polity 23(4):379–397. https://doi.org/10.3233/ip-180070](https://doi.org/10.3233/ip-180070)
Koster S, Rowe F (2019) Fueling Research Transparency: Computational Notebooks and the Discussion
Section. REGION 6(3):1–2
Krishnamurthy S (2005) ‘Cave or community? An empirical examination of 100 mature open source
projects’, First Monday.
Leech S, Green MA, Macdonald J, Gibin M (2021) Using local-level data to investigate Covid-19 inequalities in England. [https://www.adruk.org/news-publications/news-blogs/using-local-level-data-](https://www.adruk.org/news-publications/news-blogs/using-local-level-data-to-investigate-covid-19-inequalities-in-england-404/)
[to-investigate-covid-19-inequalities-in-england-404/](https://www.adruk.org/news-publications/news-blogs/using-local-level-data-to-investigate-covid-19-inequalities-in-england-404/)
Longley P (2005) Geographical Information Systems: a renaissance of geodemographics for public service delivery. Prog Hum Geogr 29(1):57–63
[Lyman P and Hal R. Varian (2003) "How Much Information" 2003. Retrieved from http://groups.ischool.](http://groups.ischool.berkeley.edu/archive/how-much-info-2003/)
[berkeley.edu/archive/how-much-info-2003/ on 03/04/2020.](http://groups.ischool.berkeley.edu/archive/how-much-info-2003/)
Manyika J. et al (2015) Interoperability Integrating multiple IoT systems enables 40 percent of potential
[value. San Francisco, USA: McKinsey Global Institute. Available at: www.mckinsey.com/mgi.](http://www.mckinsey.com/mgi)
Martin D, Gale C, Cockings S, Harfoot A (2018) Origin-destination geodemographics for analysis of
[travel to work flows. Comput Environ Urban Syst 67:68–79. https://doi.org/10.1016/j.compenvurb](https://doi.org/10.1016/j.compenvurbsys.2017.09.002)
[sys.2017.09.002](https://doi.org/10.1016/j.compenvurbsys.2017.09.002)
Meng XL (2018) Statistical paradises and paradoxes in big data (I): Law of large populations, big data
paradox, and the 2016 US presidential election. Ann Appl Stat 12(2):685–726
Molloy JC (2011) The open knowledge foundation: open data means better science. PLoS Biol
9(12):e1001195
Ostrom E (1996) Crossing the great divide: coproduction, synergy, and development. World Dev
24(6):1073–1087
Paez A, Lopez FA, Menezes T, Cavalcanti R, Pitta MGDR (2020) A spatio-temporal analysis of the environmental correlates of COVID-19 incidence in Spain. Geogr Anal 53(3):397–421
Patias N, Rowe F, Cavazzi S (2019) A scalable analytical framework for spatio-temporal analysis of
neighborhood change: a sequence analysis approach. The annual international conference on geographic information science. Springer, Cham, pp 223–241
-----
Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, Vanderplas J (2011) Scikit-learn:
machine learning in Python. J Mach Learn Res 12:2825–2830
Peng RD (2011) Reproducible research in computational science. Science 334(6060):1226–1227
Raymond E (1999) The cathedral and the bazaar. Knowl Technol Policy 12(3):23–49
Rey SJ, Anselin L (2010) PySAL: a Python library of spatial analytical methods In Handbook of applied
spatial analysis. Springer
Riffe T, Acosta E (2021) Data Resource Profile: COVerAGE-DB: a global demographic database of
[COVID-19 cases and deaths. Int J Epidemiol 50(2):390–390f. https://doi.org/10.1093/ije/dyab027](https://doi.org/10.1093/ije/dyab027)
Rowe F, Casado-Díaz JM, Martínez-Bernabéu L (2017) Functional labour market areas for Chile. Region
[4(3):7–9. https://doi.org/10.18335/region.v4i3.199](https://doi.org/10.18335/region.v4i3.199)
Rowe F, Patias N, Arribas-Bel D (2018) Policy brief: neighbourhood change and trajectories of inequality
in Britain, 1971-2011. Policy Brief prepared for UK2070 Commission, pp 1–6
Rowe F, Maier G, Arribas-Bel D, Rey S (2020) The potential of notebooks for scientific publication.
[Reproducib Dissemination Region 7(3):E1–E5. https://doi.org/10.18335/region.v7i3.357](https://doi.org/10.18335/region.v7i3.357)
Rule A, Birmingham A, Zuniga C, Altintas I, Huang S-C, Knight R, Moshiri N, Nguyen MH, Rosenthal
SR, Perez F, Rose PW (2019) Ten simple rules for writing and sharing computational analyses in
Jupyter Notebooks. PLoS Comput Biol 15(7):e1007007
[Science Staff (2011) ‘Challenges and Opportunities’, Science, 331(6018): 692–693. doi: https://doi.org/](https://doi.org/10.1126/science.331.6018.692)
[10.1126/science.331.6018.692.](https://doi.org/10.1126/science.331.6018.692)
[Singleton A, Arribas-Bel D (2019) Geographic data science. Geogr Anal. https://doi.org/10.1111/gean.12194](https://doi.org/10.1111/gean.12194)
Singleton AD, Longley PA (2009) Geodemographics, visualisation, and social networks in applied geog[raphy. Appl Geogr 29(3):289–298. https://doi.org/10.1016/j.apgeog.2008.10.006](https://doi.org/10.1016/j.apgeog.2008.10.006)
Singleton AD, Longley PA (2019) Data infrastructure requirements for new geodemographic classifica[tions: the example of London’s workplace zones. Appl Geogr 109:102038. https://doi.org/10.1016/j.](https://doi.org/10.1016/j.apgeog.2019.102038)
[apgeog.2019.102038](https://doi.org/10.1016/j.apgeog.2019.102038)
Singleton AD, Spielman SE (2014) The past, present, and future of geodemographic research in the
United States and United Kingdom. Prof Geogr 66(4):558–567
Singleton AD, Spielman S, Brunsdon C (2016) Establishing a framework for open geographic informa[tion science. Int J Geogr Inf Sci 30(8):1507–1521. https://doi.org/10.1080/13658816.2015.1137579](https://doi.org/10.1080/13658816.2015.1137579)
Snyder J, Menard A, Spare N (2018) Big Data = Big Questions for the Engineering and Construction
Industry. White Paper. First Myanmar Investment (FMI). Raleigh, US
Spielman S (2017) Keynote address CARTO I spatial data science conference. Sage, Brooklyn
Spielman SE, Singleton A (2015) Studying neighborhoods using uncertain data from the american com[munity survey: a contextual approach. Ann Assoc Am Geogr 105(5):1003–1025. https://doi.org/10.](https://doi.org/10.1080/00045608.2015.1052335)
[1080/00045608.2015.1052335](https://doi.org/10.1080/00045608.2015.1052335)
Splunk (2019) The state of dark data. Report. Splunk Inc. San Francisco, California, U.S.
Stubbings P, Peskett J, Rowe F, Arribas-Bel D (2019) A hierarchical urban forest index using street-level
imagery and deep learning. Remote Sensing 11(12):1395
Timmins K, Green MA, Radley D, Morris M, Pearce J (2018) How has big data contributed to obesity
research? a review of the literature. Int J Obes 42:1951–1962
Verhulst S, Young A and Srinivasan P (2017) An Introduction to Data Collaboratives. New York, USA:
[GovLab. Available at: http://datacollaboratives.org/static/files/data-collaboratives-intro.pdf.](http://datacollaboratives.org/static/files/data-collaboratives-intro.pdf)
Vickers D, Rees P (2007) ‘Creating the UK National Statistics 2001 output area classification.’ J R Stat
[Soc Ser a: Stat Soc 170(2):379–403. https://doi.org/10.1111/j.1467-985X.2007.00466.x](https://doi.org/10.1111/j.1467-985X.2007.00466.x)
Webber R, Burrows R (2018) The predictive postcode: the geodemographic classification of british society. SAGE, London
Wickham H, Averick M, Bryan J, Chang W, McGowan L, François R, Kuhn M (2019) Welcome to the
Tidyverse. J Open Sour Softw 4(43):1686
Wolf LJ, Rey SJ, Oshan TM. (2019) Open code is not enough: towards a replicable future for geographic
[data science http://ljwolf.org/post/opencode/](http://ljwolf.org/post/opencode/)
Zhu Z (2019) Science of landsat analysis ready data. Remote Sens 11:2166
**Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published**
maps and institutional affiliations.
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://pmc.ncbi.nlm.nih.gov/articles/PMC8528182, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBY",
"status": "HYBRID",
"url": "https://link.springer.com/content/pdf/10.1007/s10109-021-00363-5.pdf"
}
| 2,021
|
[
"JournalArticle"
] | true
| 2021-10-01T00:00:00
|
[
{
"paperId": "8409cf2149de062d22c6c79c2680752e347a8031",
"title": "Thinking spatially to communicate and evaluate the roll-out of ‘mass’ testing in Liverpool, 2020"
},
{
"paperId": "4d17e9ce4d485d3dd862cb2337847836c1f13157",
"title": "Data Resource Profile: COVerAGE-DB: a global demographic database of COVID-19 cases and deaths"
},
{
"paperId": "3c02e4244178c3bd56434d7c5cd5d238405c64c1",
"title": "The Potential of Notebooks for Scientific Publication, Reproducibility and Dissemination"
},
{
"paperId": "82620503cacf8ff6f8f3490e7bdf7508f1ab2021",
"title": "Opening practice: supporting reproducibility and critical spatial data science"
},
{
"paperId": "03bb6ed47a7a639ce54ba96d83d0b877a348270f",
"title": "A Spatio‐Temporal Analysis of the Environmental Correlates of COVID‐19 Incidence in Spain"
},
{
"paperId": "3f133afba45eecff1318de4fe6da5efdbb2114e9",
"title": "Fueling Research Transparency: Computational Notebooks and the Discussion Section"
},
{
"paperId": "6492197d6460d54b5358ea1be9de60ed7ec2f800",
"title": "Welcome to the Tidyverse"
},
{
"paperId": "123550fbc9f5420a2a06522e7dc3aac7d54d1b4e",
"title": "Open Code is not enough: Towards a replicable future for geographic data science"
},
{
"paperId": "a6b13c667df9a58c1776b64b9e029aa90eb1459b",
"title": "Science of Landsat Analysis Ready Data"
},
{
"paperId": "0abc1fcb955825792c717045b45ccd4ba9f80670",
"title": "Data infrastructure requirements for new geodemographic classifications: The example of London's workplace zones"
},
{
"paperId": "55f80d39d1903438b2a72767ff755fe8cee98b97",
"title": "Ten simple rules for writing and sharing computational analyses in Jupyter Notebooks"
},
{
"paperId": "6810ce73c832620d125a9ffa3af332e0accc91b8",
"title": "A Hierarchical Urban Forest Index Using Street-Level Imagery and Deep Learning"
},
{
"paperId": "6d71c3bf52123c2fd6b11551050865311c31853c",
"title": "A Scalable Analytical Framework for Spatio-Temporal Analysis of Neighborhood Change: A Sequence Analysis Approach"
},
{
"paperId": "160cd84c09e1f5658349a847a3c36ec039c85ca4",
"title": "Policy Brief: Neighbourhood Change and Trajectories of Inequality in Britain, 1971-2011"
},
{
"paperId": "b29e9d93c58b0d61a111fa2af1023d842623249c",
"title": "Creating value through data collaboratives"
},
{
"paperId": "e4a5d87a508710aabf82988878dbaef852259d98",
"title": "Developing an openly accessible multi‐dimensional small area index of ‘Access to Healthy Assets and Hazards’ for Great Britain, 2016"
},
{
"paperId": "0cd67d7f1f66162dc48c40372c1fce9fb3d7f0ec",
"title": "Geography and computers: Past, present, and future"
},
{
"paperId": "fdb5a2bf7f5f565cf98df71f00c02d4076c64a15",
"title": "Analysis Ready Data: Enabling Analysis of the Landsat Archive"
},
{
"paperId": "2c0d246fb2a28eb0e54582215a0bb3a0606d3057",
"title": "How has big data contributed to obesity research? A review of the literature"
},
{
"paperId": "2aa8348a499fc3d6adceb90ac54d6e13f4c757b6",
"title": "Statistical challenges of administrative and transaction data"
},
{
"paperId": "8b58f608261132a950a5e83ec00aa3b3836ccab7",
"title": "Statistical paradises and paradoxes in big data (I): Law of large populations, big data paradox, and the 2016 US presidential election"
},
{
"paperId": "ac9284630b27668ad6d592a6d928ab48f90139d2",
"title": "The Predictive Postcode: The Geodemographic Classification of British Society"
},
{
"paperId": "258709301bff9401bf448ce83c47fd439ad4961f",
"title": "Building an Earth Observations Data Cube: lessons learned from the Swiss Data Cube (SDC) on generating Analysis Ready Data (ARD)"
},
{
"paperId": "f56425ec56586dcfd2694ab83643e9e76f314e91",
"title": "50 Years of Data Science"
},
{
"paperId": "4a6d46962d3f58d278cfb46d3ddebbb30bf275f5",
"title": "Geographic Data Science"
},
{
"paperId": "367a9ee47615907d6787489eeb615147a7261779",
"title": "Functional Labour Market Areas for Chile"
},
{
"paperId": "e80f918d9b6f219c84045b4a81d573aec857760e",
"title": "Challenges in administrative data linkage for research"
},
{
"paperId": "75789f541bd5693889b5ba06166404fabfff80ea",
"title": "The Cost(s) of Geospatial Open Data"
},
{
"paperId": "b968e8dac7355748d80ccea547ac5a0dea9a8f69",
"title": "The co-production of what? Knowledge, values, and social relations in health care"
},
{
"paperId": "dd14b3c6a3cf0873c37c6692aa8e9f7d9a4676f2",
"title": "An evolutionary approach to the delimitation of labour market areas: an empirical application for Chile"
},
{
"paperId": "75acbc5aeb26906953fd43fd450743c40e8dd4fc",
"title": "Establishing a framework for Open Geographic Information science"
},
{
"paperId": "0d14e684edbf2a876d8d3ef24861ac59106ebd38",
"title": "Creating the 2011 area classification for output areas (2011 OAC)"
},
{
"paperId": "b16c6550de331b694d1d5030a2318b4ff17e57e0",
"title": "Spatial Dependence: Weighting Schemes, Statistics and Models"
},
{
"paperId": "69f4cf3191ba82a27d272a79f04fdc8f718c3078",
"title": "Studying Neighborhoods Using Uncertain Data from the American Community Survey: A Contextual Approach"
},
{
"paperId": "92b8a8124872dfa576fbe9ea44d2a8ab723fe477",
"title": "The Data Revolution: Big Data, Open Data, Data Infrastructures and Their Consequences"
},
{
"paperId": "3f558e813321e48930f0b211c4c81c37f1ee75b0",
"title": "Geographers Count: A Report on Quantitative Methods in Geography"
},
{
"paperId": "171d0bc1598858bb728dd9373f97ca5b363b5edc",
"title": "The Past, Present and Future of Geodemographic Research in the United States and United Kingdom"
},
{
"paperId": "27e3eb01ec182959ec1bd579b90c5a8eec3075ad",
"title": "Challenges and Opportunities"
},
{
"paperId": "4b606805da01c61e4422fd90fe33877a6d71951c",
"title": "Benefits, Adoption Barriers and Myths of Open Data and Open Government"
},
{
"paperId": "823b50be0da9b84c4945f947ecbacc83f2c2cc25",
"title": "A framework for successful new product development"
},
{
"paperId": "e4625641d1072542315c14083befb43cd7096aba",
"title": "Reproducible Research in Computational Science"
},
{
"paperId": "f41f3ad8c286a8dc509b2fa42b5febe81bcec9d8",
"title": "The Open Knowledge Foundation: Open Data Means Better Science"
},
{
"paperId": "64d2fe89ba7f105cb96796742f4dcbf4e56bd2ca",
"title": "The World’s Technological Capacity to Store, Communicate, and Compute Information"
},
{
"paperId": "9e7a7c605264f886bee1e0b53909f825ed6e20e6",
"title": "Making Data Maximally Available"
},
{
"paperId": "ad4fd2c149f220a62441576af92a8a669fe81246",
"title": "Scikit-learn: Machine Learning in Python"
},
{
"paperId": "1b49b613a4a3518b44803a417de2c85dfec2900f",
"title": "Geodemographics, visualisation, and social networks in applied geography"
},
{
"paperId": "ba612664f5ae2a8f6c02fd369f259f216c55be23",
"title": "Participatory GIS — a people's GIS?"
},
{
"paperId": "8e93610ebb1774904db423dd2513042cfb0bf290",
"title": "Creating the UK National Statistics 2001 output area classification"
},
{
"paperId": "df48f0c4131672b95187bc2a0089682263a5937c",
"title": "Keynote address"
},
{
"paperId": "49f4803a0dbdc0bf4199593c3e40e2ab8db38f9e",
"title": "Geographical Information Systems: a renaissance of geodemographics for public service delivery"
},
{
"paperId": "2fc0516f700b490b7e13db0f0d73d05afa5e346c",
"title": "Cave or Community? An Empirical Examination of 100 Mature Open Source Projects"
},
{
"paperId": "5eaf9f92564b1ba8060c5d1021f8cfd70ae6b36b",
"title": "The cathedral and the bazaar"
},
{
"paperId": "8bd9fc0acc518db3d1efefe7fef37c3817bbbda5",
"title": "Crossing the great divide: Coproduction, synergy, and development"
},
{
"paperId": null,
"title": "Using local-level data to investigate Covid-19 inequalities in England"
},
{
"paperId": null,
"title": "Digital 2019 Global Digital Overview"
},
{
"paperId": "6276081391d85136b2e9140ba03c82ae0f48e135",
"title": "Statistics, Modelling, and Data Science"
},
{
"paperId": null,
"title": "Open Sour Softw"
},
{
"paperId": null,
"title": "The state of dark data"
},
{
"paperId": "2f1f46b45be18b117660682e84e93b31654c4c2a",
"title": "Origin-destination geodemographics for analysis of travel to work flows"
},
{
"paperId": null,
"title": "Harfoot A (2018) Origin-destination geodemographics for analysis"
},
{
"paperId": null,
"title": "Big Data = Big Questions for the Engineering and Construction Industry"
},
{
"paperId": null,
"title": "Quantitative methods I: reproducible research and quantitative geography"
},
{
"paperId": "390715d7f73579d0fe508e494bea2ef5929c00b4",
"title": "University of Birmingham Accidental, open and everywhere: Emerging data sources for the understanding of cities"
},
{
"paperId": null,
"title": "Interoperability Integrating multiple IoT systems enables 40 percent of potential value"
},
{
"paperId": "43789305e5d2212da05f9c16b148e84aae5614b2",
"title": "Citizen Science and Volunteered Geographic Information: Overview and Typology of Participation"
},
{
"paperId": "6ac3d1790c7be421f47be53f337c7eda0f57224a",
"title": "PySAL: A Python Library of Spatial Analytical Methods"
},
{
"paperId": null,
"title": "The expanding digital universe: a forecast of worldwide information growth through"
},
{
"paperId": "d8f16b07147e2cbedd8227ea5a8ef2cd7f8ae413",
"title": "GeoDa: An Introduction to Spatial Data Analysis"
},
{
"paperId": "0f7cb51fef6ff49ce293534e58935acf1cf5df73",
"title": "Skills, Division of Labor and Performance in Collective Inventions. Evidence from the Open Source Software"
},
{
"paperId": null,
"title": "How Much Information"
},
{
"paperId": null,
"title": "Policy Brief prepared for UK2070 Commission, pp"
},
{
"paperId": null,
"title": "Open data products‑A framework for creating valuable analysis…"
}
] | 15,175
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Engineering",
"source": "s2-fos-model"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/02b64b590630a0ebcbf2f75e045c1d933e3f2667
|
[
"Computer Science"
] | 0.865984
|
A Data-Driven Distributed Adaptive Control Approach for Nonlinear Multi-Agent Systems
|
02b64b590630a0ebcbf2f75e045c1d933e3f2667
|
IEEE Access
|
[
{
"authorId": "102908639",
"name": "Xian Yu"
},
{
"authorId": "1682560",
"name": "S. Jin"
},
{
"authorId": "121866036",
"name": "Genfeng Liu"
},
{
"authorId": "2029346528",
"name": "Ting Lei"
},
{
"authorId": "2109922523",
"name": "Ye Ren"
}
] |
{
"alternate_issns": null,
"alternate_names": null,
"alternate_urls": [
"http://ieeexplore.ieee.org/servlet/opac?punumber=6287639"
],
"id": "2633f5b2-c15c-49fe-80f5-07523e770c26",
"issn": "2169-3536",
"name": "IEEE Access",
"type": "journal",
"url": "http://www.ieee.org/publications_standards/publications/ieee_access.html"
}
|
In this paper the distributed leader-follower consensus tracking problem is investigated for unknown nonlinear non-affine discrete-time multi-agent systems. Via a dynamic linearization method both for the agent system and the local ideal distributed controller, a distributed adaptive control scheme is proposed in this paper using the Newton-type optimization method. The proposed approach is data-driven since only the local measurement information among neighboring agents is utilized in the control system design. The consensus tracking stabilities of the proposed approach are rigorously guaranteed in the cases of fixed and switching communication topologies. The simulations are conducted to verify the effectiveness of the proposed approach.
|
Received October 31, 2020, accepted November 11, 2020, date of publication November 17, 2020,
date of current version November 30, 2020.
_Digital Object Identifier 10.1109/ACCESS.2020.3038629_
# A Data-Driven Distributed Adaptive Control Approach for Nonlinear Multi-Agent Systems
XIAN YU 1, SHANGTAI JIN 1, GENFENG LIU 1, TING LEI 2, AND YE REN3
1School of Electronic and Information Engineering, Beijing Jiaotong University, Beijing 100044, China
2College of Electrical and Information Engineering, Zhengzhou University of Light Industry, Zhengzhou 450002, China
3School of Electrical and Control Engineering, North China University of Technology, Beijing 100144, China
Corresponding author: Xian Yu (yuxian@bjtu.edu.cn)
This work was supported in part by the National Natural Science Foundation of China (NSFC) under Grant 61433002 and Grant 61833001.
**ABSTRACT In this paper the distributed leader-follower consensus tracking problem is investigated for**
unknown nonlinear non-affine discrete-time multi-agent systems. Via a dynamic linearization method both
for the agent system and the local ideal distributed controller, a distributed adaptive control scheme is
proposed in this paper using the Newton-type optimization method. The proposed approach is data-driven
since only the local measurement information among neighboring agents is utilized in the control system
design. The consensus tracking stabilities of the proposed approach are rigorously guaranteed in the cases
of fixed and switching communication topologies. The simulations are conducted to verify the effectiveness
of the proposed approach.
**INDEX TERMS Dynamic linearization, data-driven control, adaptive control, multi-agent systems, consen-**
sus tracking.
**I. INTRODUCTION**
The recent two decades have witnessed a burgeoning research
direction in the automatic control of interconnected systems [1], [2]. Such systems are the well-known multi-agent
systems (MASs). Cooperative control of MASs is aimed to
exploit the local interactive control protocols among networked agents for achieving a global objective that is difficult
to be accomplished by a single agent. Due to its powerful
potential applications [3]–[5], a considerable attention has
been attracted for different cooperating tasks, such as consensus, formation, coverage control, flocking and containment
control. Among these research topics, consensus control is an
important and fundamental problem. Remarkable results on
consensus have been investigated from different perspectives,
and readers are referred to [6]–[8] and references therein.
Because of the pioneering works [9], [10] on consensus, many scholars have extensively investigated different
consensus problems. For instance, in [11]–[13], the distributed consensus of linear continuous-time and discretetime homogeneous systems was discussed. Moreover, some
works, such as [14], [15], were extended to heterogeneous
network systems. Since almost all the physical dynamics
The associate editor coordinating the review of this manuscript and
approving it for publication was Fei Chen.
of controlled systems in practice are inherently nonlinear,
the aforementioned control schemes cannot be directly
applied to nonlinear systems. Recently, the adaptive control
schemes were developed for nonlinear MASs [6], [7]. However, the aforementioned works are usually based on availability of the dynamic models or structural information of
the controlled MAS. This means that the first principle or an
identification method is required for these distributed control
schemes, which have the problems of unmodeled dynamics
and model/controller reduction [16].
For the control problems of unknown systems, datadriven control methodologies are a concerned research topic,
in which the model free adaptive control (MFAC), proposed
by Hou in [17] and further developed in [18], is valuable.
The MFAC has been extended from the original single-input
single-output systems [19], [20] and multi-input multi-output
systems [21], [22] to nonlinear MASs [23]–[25]. Besides,
the dynamic linearization based control methods have been
successfully applied to many practical applications, such
as servo motor systems [26] and exoskeleton robotic systems [27]. The detail of the dynamic linearization based
control methods can be found in [16].
However, some challenging issues still have not been
developed for unknown nonlinear discrete-time MASs on
data-driven consensus control. One issue is to how to
-----
design a distributed consensus controller structure through a
systematic approach. The local controller structures, such as
the distributed proportional and proportional-integral control
laws in [7], [28], are usually determined a priori by experience, which leads to the difficulty in determining their appropriation or effectiveness in applications. Another issue is to
how to design a distributed control gain updating algorithm
on condition that the local measurements are only applicable.
The control gains in existing distributed control schemes are
usually calibrated heuristically and chosen as fixed values if
the physical dynamics of the controlled MAS are unknown.
The dynamic linearization based control methods motivate us
to explore a novel data-driven distributed consensus tracking
approach for addressing these issues.
Comparing to existing distributed control methods,
the main contributions of this paper are as follows.
- Provide a systematic way of directly designing the distributed controller structure for unknown MASs on the
consensus tracking, and the designed distributed controller is independent of the controlled MAS.
- Propose a data-driven distributed adaptive control
approach, where the local control law and control gain
updating algorithm are designed only using the local
information among neighboring agents.
- Establish the consensus tracking stability properties
of the proposed approach under fixed and switching
topologies.
The rest of this paper is organized as follows. Section II
formulates the consensus tracking problem. Section III concludes the main results, including the designed control law,
distributed control gain updating algorithm, adaptive control
approach and its convergence properties. Section IV conducts some simulations. Section V provides some concluding remarks. In this paper, denotes any generic vector
∥· ∥
or matrix norm.
**II. PROBLEM FORMULATION**
The communication topology of a leader-follower system
including the leader is represented by the graph, where the
_G_
topology among agents is fixed and directional. The leader’s
command is only accessible to a subset of the follower agents
with unidirectional paths from the leader to the follower
agents. Each follower agent exchanges local measurement
information only with its neighboring follower agents under
a directional graph. It is assumed that the topology among
the follower agents is a fixed strongly connected graph and at
least one follower agent is communicated to the leader.
We consider a set of N heterogeneous nonlinear non-affine
discrete-time follower agents, where the physical model of
follower agent q, q = 1, 2, · · ·, N, is described by
_yq(k + 1) = fq�yq(k), · · ·, yq(k −_ _ny),_
_uq(k), · · ·, uq(k −_ _nu)�,_ (1)
where yq(k) ∈ R[1] and uq(k) ∈ R[1] are the system output and
control input of agent q at the time instant k, respectively,
and k = 1, 2, · · · ; ny ∈ Z+ and nu ∈ Z+ are unknown orders
for the system outputs and control inputs of agent q; fq(·) :
R[n][y][+][n][u][+][2] �→ R[1] is an unknown nonlinear function.
Following [19], the agent system (1) can be transformed
into the following equivalent dynamic linearization data
model:
_δyq(k + 1) = ψq(k)δuq(k),_ (2)
where δyq(k + 1) = yq(k + 1) − _yq(k); δuq(k) = uq(k) −_
_uq(k −_ 1); the unknown time-varying parameter ψq(k) is
called the pseudo partial derivative (PPD) of the agent system
(1), satisfying |ψq(k)| ≤ _bp, bp > 0 is a known constant._
_Assumption 1: The sign of ψq(k), remains invariant, satis-_
fying ψq(k) > 0 or ψq(k) < 0 for all k = 1, 2, · · · . Without
loss of generality, we consider 0 < ψq(k) ≤ _bp in this paper._
_Remark 1: The considered condition 0 < ψq(k) ≤_ _bp in_
Assumption 1 implies that the control direction is known and
positive. This condition is reasonable since many practical
systems, such as autonomous underwater vehicles, unmanned
aerial vehicles and mobile robots, feature this property.
Note that the time-varying parameter ψq(k) is only a
concept in the sense of mathematics, and its existence is
rigorously guaranteed by the theorem in [19]. While the timeinvariant parameters usually introduced for traditional adaptive control methods indicate the variables of dynamics of a
controlled plant. It can be seen from the theorem in [19] that
_ψq(k) is obviously time-varying even if the controlled plant_
(1) is linear time-invariant since ψq(k) is only related to the
system outputs and control inputs by the time k. Moreover, all
of the possible properties of the controlled system (1), such
as the nonlinearity and time-varying parameters or structures,
are involved into ψq(k), which may lead to its complicated
characteristic, but the simple numerical behavior that can
be easily estimated. This implies that ψq(k) is capable of
managing these properties that are difficult to be handled
in the framework of traditional adaptive control due to its
possible insensitivity to these properties. More detail on the
parameter can be referred to [29].
Equation (2) is only a data model that is equivalent to (1)
in the sense of mathematics and it has no physical meanings. This equivalent transformation is achieved by using the
compact form dynamic linearization method, and the detailed
derivations can be referred to [19] and [29]. The data model
(2) is only related to the system outputs and control inputs
of the controlled plant, and it not explicitly or implicitly
includes the parametric and structural information of physical
dynamics of a controlled plant. In addition, the data model
(2) is only purposed to the control system design, and it
is not suitable for other objectives, such as monitoring and
diagnosis.
The consensus tracking problem of the leader-follower
MAS described by (1) under the graph is summarized as
_G_
follows.
The leader’s command at the time instant k is denoted as
_yd_ (k). The global control objective is to develop a data-driven
distributed adaptive control approach that drives the system
-----
output yq(k) to yd (k) when the time instant k tends to infinity;
that is, the tracking error lim
_k→∞_ _[e][q][(][k][)][ =][ y][d]_ [(][k][)][ −] _[y][q][(][k][)][ =][ 0,]_
although the local information among neighboring agents is
only used and the leader’s command is only accessible to a
subset of the follower agents. To describe the local measurement information of agent q with its neighbors, we define the
following distributed tracking error ξq(k) under G as:
_ξq(k) =_ � _aq,p�yp(k) −_ _yq(k)�_ + dq�yd (k) − _yq(k)�, (3)_
_p∈Nq_
where Nq denotes the set of neighbors of agent q; aq,p 1
=
if agent q can receive information from its neighboring agent
_p, otherwise aq,p = 0, specially aq,q = 0; dq = 1 if agent q_
receives the leader’s command yd (k), otherwise dq = 0.
The first issue for developing the distributed control
approach is the structure design of a distributed control law.
Since the physical model of the MAS described by (1) is completely unknown, so far there is not a systematic way to determine the distributed controller structure. The second issue is
related to the design of the distributed control gain updating
algorithm using only the local measurement information.
The last important issue is the stability properties of the
developed data-driven distributed control approach. In next
section, these issues are discussed in detail.
**III. MAIN RESULTS**
_A. DISTRIBUTED CONTROL LAW_
This subsection considers the design of the distributed
controller structure through only known local information.
Assume there exists an ideal distributed consensus controller
in theory that can guarantee the system output of agent q equal
to yd (k +1) in one step-ahead. The ideal distributed controller
can be written in the following mathematical form:
_uq(k) = Cq�ξq(k + 1), · · ·, ξq(k −_ _ne + 2),_
_uq(k −_ 1), · · ·, uq(k − _nc)�,_ (4)
where Cq(·) : R[n][e][+][n][c] �→ R[1] is an unknown nonlinear
function; ne ∈ Z+ and nc ∈ Z+ are the unknown orders
of ideal distributed controller (4) on the distributed tracking
errors and control inputs, respectively.
The assumption on (4) implies that it is reachable and
the detailed discussions are referred to [30]. In practice,
the controller (4) is difficult to derive. Thus the key task is
to transform it into a practical distributed controller, while
keeping it equivalent to (4) in the input-output data sense.
In achieving this, the following two assumptions are required.
_Assumption 2: The partial derivative of Cq(·) with respect_
to the distributed tracking error ξq(k + 1) is continuous.
_Assumption 3: Cq(·) satisfies the generalized Lipschitz_
condition, that is, if |δξq(k + 1)| ̸= 0, then there exists an
unknown constant β > 0 such that
|δuq(k)| ≤ _β|δξq(k + 1)|,_ (5)
where δξq(k + 1) = ξq(k + 1) − _ξq(k)._
Assumption 2 is common since many controllers, such as
the distributed proportional controller [31] and the distributed
adaptive controller [32], generally satisfy this condition.
Assumption 3 implies that the ideal distributed controller (4)
is required to be stable [16].
_Lemma 1: The controller (4) satisfies Assumptions 2_
and 3. If |δξq(k + 1)| ̸= 0, then there exists an unknown
controller parameter θq(k), such that (4) can be transformed
into the following equivalent distributed control law using the
compact form dynamic linearization (CFDL) method:
_δuq(k) = θq(k)δξq(k + 1),_ (6)
where |θq(k)| ≤ _bt_, and bt > 0 is an unknown constant.
_Proof:Lemma 1 can be proved by utilizing the differen-_
tial mean value theorem through Assumptions 2 and 3. The
derivative detail is similar to the results in [30], [33] and thus
the derivations are omitted.
For simplicity, we label the obtained distributed control law
(6) as CFDL controller (CFDLc).
_Remark 2: The CFDLc (6), with a time-varying lineariza-_
tion structure, is equivalent to the ideal distributed controller
(4), which implies two points. One point is that the structure
complexity of (6) does not increase even though the agent system (1) is highly nonlinear. Another point is that the CFDLc
(6) can be considered as a candidate consensus controller for
unknown nonlinear MASs as described by (1) since (6) can
drive eq(k + 1) = 0 in one-step. In other words, the issue
of designing a distributed controller structure is addressed
through the CFDL method, while the existing distributed
controller structures are usually given in an ad hoc way.
_Remark 3: Lemma 1 indicates that the CFDLc (6) is inde-_
pendent of the agent system (1), and θq(k) can be obtained
through only the local information using some data analytical
approaches when the dynamic model of agent q is unavailable. It can also be obtained via the model based optimization
method through submitting (6) into the agent system (1) when
its dynamic model is known. This paper just considers the
issue of obtaining θq(k) using only the local information
communicated with agent q.
The CFDLc (6) cannot be implemented in practice due to
the presence of the noncausal term ξq(k + 1) in (6). Similar
to [33], the following practical CFDLc is obtained:
_δuq(k) = −θq(k)ξq(k),_ (7)
which means that uq(k) can be computed directly according
to the measured ξq(k) at current time instant k. Note that (7)
is not an approximation to (6), but a direct derivation from the
observation that (6) can drive eq(k + 1) = 0.
_B. DISTRIBUTED CONTROL GAIN UPDATING ALGORITHM_
This subsection considers the second issue on tuning θq(k)
in the CFDLc (7) using only the local information communicated with agent q via the data model (2).
-----
We first consider the following control objective function:
_Jq =_ 2[1]
� _aq,p�yp(k + 1) −_ _yq(k + 1)�2_
_p∈Nq_
+ [1] �yd (k + 1) − _yq(k + 1)�2 + 1_ _q[(][k][)][,]_ (8)
2 _[d][q]_ 2 _[λδ][u][2]_
_µ > 0 is a weight factor, η ∈_ (0, 1] is the step size of ψq(k),
and σ is a tiny positive constant.
Based on the estimation given in (13) and (14), the estimation of the noncausal term ξq(k + 1) is given by
_ξˆq(k + 1) =_ � _aq,p�yˆp(k + 1) −ˆyq(k + 1)�_
_p∈Nq_
+ dq�yd (k + 1) −ˆyq(k + 1)�, (15)
where
_yˆp(k + 1) = yp(k) −_ _ψ[ˆ]_ _p(k)δup(k),_ (16)
_yˆq(k + 1) = yq(k) −_ _ψ[ˆ]_ _q(k)δuq(k)._ (17)
_C. SUMMARIZED DISTRIBUTED ADAPTIVE_
_CONTROL APPROACH_
The CFDLc (7), the two updating algorithms (11) and (13)
with the resetting mechanisms (12) and (14), and the distributed tracking error estimation (15), formulate the distributed consensus tracking approach. The detailed steps are
as follows.
_Step 1: Set k_ 1, initialize the local measurement data
=
and the _ψ[ˆ]_ _q(1) satisfying σ ≤_ _ψ[ˆ]_ _q(1) ≤_ _bp, and randomly set_
_θq(1) satisfying −bt ≤_ _θq(1) < 0._
_Step 2: Compute the control input_
_uq(k) = uq(k −_ 1) − _θq(k)ξq(k),_ (18)
rewritten from the CFDLc (7), apply it to agent q, and collect
_yq(k + 1) and uq(k)._
_Step 3: Update the PPD estimation value_ _ψ[ˆ]_ _q(k) using (13)_
with the resetting mechanism (14).
_Step 4: Compute_ _ξ[ˆ]q(k + 1) by (15) with (16) and (17)._
_Step 5: Update the control gain_
_ψˆ_ _q(k)ξˆq(k + 1) + λθq(k)ξq(k)_
_θq(k + 1) = θq(k) −_ _γ_ _, (19)_
�λ + ψ[ˆ] _q[2](k)�ξq(k)_
with the resetting mechanism (12).
_Step 6: Set k_ _k_ 1, and go back to Step 2.
= +
For convenient descriptions, we label the proposed
approach as CFDL based distributed adaptive control
(CFDL-DAC).
_Remark 4: The proposed CFDL-DAC illustrates that no_
physical dynamics of the controlled MAS are involved into
the distributed controller design. The parameter updating
algorithm (13) is based on only the input-output data of
each agent. The distributed tracking error estimation (15),
the distributed control law (18) and the distributed control
gain updating algorithm (19) are designed using only the local
information communicated to agent q. Further, the design for
the distributed control law (18) is an independent process of
the dynamics of agent q. Hence, the proposed CFDL-DAC is
a pure data-driven distributed control approach.
Note that the proposed CFDL-DAC can be extended to
deal with the leaderless control problems [34], [35] since it
is applicable as long as the local measurement information
can be described, which is demonstrated by (3). However,
where λ > 0 is a weight factor used as penalty for δuq(k).
In order to obtain the optimal control gain θq(k) under the
control objective function (8), the relationship between yq(k+
1) and uq(k) for agent q is required. In achieving this, the data
model (2) is applied and we rewrite it as
_yq(k + 1) = yq(k) + ψq(k)δuq(k)._ (9)
Taking the CFDLc (7) and data model (9) into the control
objective function (8), we obtain
_Jq =_ [1]2 � _aq,p�yp(k + 1) −_ _yq(k) + ψq(k)θq(k)ξq(k)�2_
_p∈Nq_
+ [1] �yd (k + 1) − _yq(k) + ψq(k)θq(k)ξq(k)�2_
2 _[d][q]_
+ [1] �θq(k)ξq(k)�2. (10)
2 _[λ]_
Equation (10) indicates that the control objective function (8)
is transformed into an identification function of θq(k). Then
the tuning of θq(k) is achieved by applying the following
Newton-type optimization method:
�−1
_∂Jq_
_∂θq(k)_
_θq(k + 1) = θq(k) −_ _γ_
�
_∂_ [2]Jq
_∂θq[2](k)_
= θq(k) − _γ [ψ][q][(][k][)][ξ]�[q]λ[(] +[k][ +] ψ[ 1)]q[2]([ +]k)�[ λθ]ξq[q](k[(][k])[)][ξ][q][(][k][)]_ _, (11)_
with the given resetting mechanism
_θq(k + 1) = −bt_ if _θq(k + 1) < −bt_,
or _θq(k + 1) = 0_ if _θq(k + 1) > 0,_ (12)
where γ ∈ (0, 1] is the step size of θq(k).
Note that the control gain θq(k) is not required to be
updated if the distributed tracking error ξq(k) = 0 since
_yd_ (k) = yq(k) in this case; that is, a perfect tracking for agent
_q is achieved._
However the control gain updating algorithm (11) is not
realizable since ψq(k) is unknown and ξq(k + 1) is noncausal.
For simplicity, we consider the updating algorithm given
in [23] to estimate ψq(k):
_ψˆ_ _q(k) = ˆψq(k −_ 1) + _[ηϵ][q][(][k][)][δ][u][q][(][k][ −]_ [1)] (13)
_µ + δu[2](k −_ 1) _[,]_
with the resetting mechanism
_ψˆ_ _q(k) = ˆψq(k −_ 1) if _ψ[ˆ]_ _q(k) < σ or_ _ψ[ˆ]_ _q(k) > bp, (14)_
where
_ϵq(k) = δyq(k) −_ _ψ[ˆ]_ _q(k −_ 1)δuq(k − 1),
-----
the results in [34], [35] are obtained for continues-time multiagent systems with unknown control directions, and this
paper considers discrete-time multi-agent systems where the
control directions are known. Therefore, the two obstacles are
required to be tackled before utilizing the proposed approach.
_D. CONVERGENCE ANALYSES_
The lemma following [36] is applied to facilitate the convergence analyses.
_Lemma 2: HHH_ (k) ∈ R[N] [×][N] is an irreducible stochastic
matrix with positive diagonal entries and is the set of all
_H_
possible HHH (k). The multiplication of Q matrixes satisfies
∥HHH (Q)HHH (Q − 1) · · · _HHH_ (1)∥≤ _ι,_ (20)
where {HHH (r)|r = 1, 2, · · ·, Q, Q ∈ Z+} is the subset of the
Q matrixes arbitrarily selected from, and 0 < ι < 1.
_H_
_Theorem 1: Let the MAS described by (1) under the com-_
munication graph satisfying Assumptions 1–3, be con_G_
trolled by the proposed CFDL-DAC. The leader’s command
_yd_ (k) is assumed to be time-invariant, namely, yd (k) ≡ _c, c_
is a constant. Then eq(k) converges to zero as k →∞ for all
_q = 1, 2, · · ·, N_, if the following condition is satisfied:
1
_bt <_ _bp�_ _q=max1,···,N_ �Np=1 _[a][q][,][p][ +][ d][q]�_ _._ (21)
_Proof:We let_
Substituting (26) into (24) yields the following closed-loop
error dynamics:
_eee(k + 1) = eee(k) + ���(k)���(k)ξξξ_ (k). (27)
Then based on equation (25), it has
_eee(k + 1) = eee(k) + ���(k)���(k)(LLL + DDD)eee(k)_
= �III + ���(k)���(k)(LLL + DDD)�eee(k). (28)
From Assumption 1, we have that 0 < ψq(k) ≤ _bp._
Besides, III + _���(k)���(k)(LLL +_ _DDD) must be an irreducible matrix_
since the communication graph is assume to be strongly
_G_
connected. Based on the resetting mechanism (12), for the
matrix III +���(k)���(k)(LLL +DDD), if the condition (21) is satisfied,
then there is at least one row sum of the matrix strictly less
than one, which means that it is an irreducible stochastic
matrix with positive diagonal entries.
With equation (28), we can conclude that
_eee(k + 1) = GGG(k, 1)eee(1),_ (29)
where
_GGG(k, 1) =_
_k_
�
_j=1_
� �
_III + ���(k −_ _j + 1)���(k −_ _j + 1)(LLL + DDD)_ _._
_yyy(k) = [y1(k), y2(k), · · ·, yN_ (k)][T] ∈ R[N] _,_
_uuu(k) = [u1(k), u2(k), · · ·, uN_ (k)][T] ∈ R[N] _,_
_eee(k) = [e1(k), e2(k), · · ·, eN_ (k)][T] ∈ R[N] _,_
_ξξξ_ (k) = [ξ1(k), ξ2(k), · · ·, ξN (k)][T] ∈ R[N] _,_
and rewrite equations (2) and (3) respectively as the following
form based on yd (k) ≡ _c:_
_eq(k + 1) = eq(k) −_ _ψq(k)δuq(k),_ (22)
_ξq(k) =_ � _aq,p�eq(k) −_ _ep(k)�_ + dqeq(k), (23)
_p∈Nq_
Then equations (22) and (23) can be respectively expressed
by the following vector forms:
_eee(k + 1) = eee(k) −_ _���(k)δuuu(k),_ (24)
_ξξξ_ (k) = (LLL + DDD)eee(k), (25)
where
_���(k) = diag(ψ1(k), ψ2(k), · · ·, ψN_ (k)) ∈ R[N] [×][N] _,_
_δuuu(k) = uuu(k) −_ _uuu(k −_ 1),
_DDD = diag(d1, d2, · · ·, dN_ ) ∈ R[N] [×][N] _,_
and LLL ∈ R[N] [×][N] is the Laplacian matrix of the follower agents
under the communication graph .
_G_
Similarly, we rewrite (7) as the following vector form:
_δuuu(k) = −���(k)ξξξ_ (k), (26)
where ���(k) = diag(θ1(k), θ2(k), · · ·, θN (k)) ∈ R[N] [×][N] .
Taking norms on both sides of equation (29) yields
∥eee(k + 1)∥≤∥GGG(k, 1)∥∥eee(1)∥, (30)
By grouping all Q matrixes together for _GGG(k, 1) in (30), and_
applying Lemma 2, we have
∥eee(k + 1)∥≤ _ι[⌊]_ _Q[k]_ [⌋]∥eee(1)∥, (31)
where indicates the smaller but nearest integer to the real
⌊·⌋
number k/Q. Then it is obtained that lim
_k→∞_ [∥][eee][(][k][ +][ 1)][∥=][ 0;]
that is, the tracking error eq(k) converges to zero as k →∞
for all q = 1, 2, · · ·, N . This proof is completed.
Next the communication graph for the MAS described by
(1) is extended to switching topologies, where each communication graph is strongly connected and at least one agent is
communicated to the leader’s command at each time instant k
for each graph. To facilitate the description of the switching
topologies, we denote (k) as a time-varying graph at time
_G_
instant k, then we can have matrixes LLL(k) and DDD(k) with the
same denotation as aforementioned. Furthermore, we denote
_Gt = {G1, G2, · · ·, GP} as the set of all possible directed_
graphs, describing the switching communication topologies,
where P ∈ Z+ is the total number of possible communication
topologies. In this case, the stability of the CFDL-DAC is
summarized as follows.
_Corollary 1: Let the MAS described by (1) satisfying_
Assumptions 1–3, be controlled by the proposed CFDL-DAC,
where the communication topology is the switching graphs
_Gt = {G1, G2, · · ·, GP}, each graph is strongly connected, and_
the leader’s command yd (k) is assumed to be time-invariant,
namely, yd (k) ≡ _c. Then the tracking error eq(k) converges_
-----
to zero as k →∞ for all q = 1, 2, · · ·, N, if we select bt
satisfying the following condition:
1
_bt <_ _bp�_ _q=max1,···,N_ �Np=1 _[a][q][,][p][(][m][)][ +][ d][q][(][m][)]�_ _,_ (32)
_m=1,···,P_
where (aq,p(m)) is the weighted adjacent matrix of Gm, dq(m)
is the entries of D(m) = diag(d1(m), · · ·, dN (m)) under Gm,
_Gm is the element of set Gt_, and m = 1, 2, · · ·, P.
_Proof:In this case, equation (25) is rewritten as_
_ξξξ_ (k) = (LLL(k) + DDD(k))eee(k), (33)
then based on equations (24) and (26), we have
_eee(k + 1) =_ �III + ���(k)���(k)(LLL(k) + DDD(k))�eee(k). (34)
Since all the possible communication topologies are
strongly connected, III + ���(k)���(k)(LLL(k) + DDD(k)) is still an
irreducible matrix. It is noted that the set {LLL1 + DDD1, _LLL2 +_
_DDD2, · · ·,_ _LLLP + DDDP} includes all the possible matrices of_
_LLL(k)_ _DDD(k). If the condition (32) is satisfied, then the greatest_
+
diagonal entry of III + ���(k)���(k)(LLL(k) + DDD(k)) is less than 1;
that is, the matrix III + ���(k)���(k)(LLL(k) + DDD(k)) is irreducibly
stochastic with positive diagonal entries. Similar to the proof
of Theorem 1, it then can be obtained that the tracking error
_eq(k) converges to zero as k →∞_ for all q = 1, 2, · · ·, N .
This completes the proof.
_Remark 5: The results of Theorem 1 and Corollary 1 are_
based on the time-invariant leader’s command, and the convergence conditions (21) and (32) require a global communication topology to determine bt . The limitation in
determining bt probably can be avoided by introducing the
stability analysis methods given in [32]. However, the results
in [32] are based on the availability of physical model of a
controlled plant. The agent system considered in this paper is
unknown. Therefore, it may need to integrate other analysis
methods in addressing this problem. In future work the case
of time-varying leader’s command will be investigated for
generalizing the proposed approach given in this paper.
The conditions (21) and (32) given in Theorem 1 and
Corollary 1 seem from their mathematical forms that they
can be ensured by simply choosing −bt ≥ _θq(k) < 0._
However, it should be noted that θq(k) is designed to achieve
its automatic tuning using only the local information among
neighboring agents. This is different from the most existing distributed control schemes where the control gains are
usually calibrated heuristically and chosen as fixed values
if the physical dynamics of a controlled plant are unknown.
As presented by the control gain updating algorithm (11), this
automation helps search and approximate the optimal value
of the control gain in the sense of minimizing the control
objective function (8). Moreover, bt is purposed to be chosen
a value as large as possible under the conditions of (21) and
(32), so that θq(k) can be searched in a larger space in order
to better approximate the optimal value.
and DDD = diag(1, 0, 0, 1).
It is seen that the communication topology is strongly
connected. The bound of θq(k) is set as bt = 0.2, and the
**FIGURE 1. Fixed communication topology.**
**IV. SIMULATION RESULTS**
To illustrate the effectiveness of the proposed CFDL-DAC,
three examples are simulated in this paper. The three examples consider the same nonlinear heterogeneous discrete-time
MAS, where the first two examples are conducted under the
fixed and switching communication topologies, respectively,
with time-invariant leader’s command, and the third example
is proceeded with time-varying leader’s command.
The nonlinear heterogeneous discrete-time MAS consists
of four follower agents, where the follower agent models are
governed by
_y1(k −_ 1)y1(k)
_y1(k + 1) =_
1 + y[2]1[(][k][ −] [1)][ +][ y][2]1[(][k][)][ +][ 3][u][1][(][k][)][,]
_y2(k)_
_y2(k + 1) =_ 2[(][k][)][,]
1 + y[4]2[(][k][)][ +][ u][3]
(35)
_y3(k + 1) =_ _[y][3][(][k][ −]_ [1)][y][3][(][k][)][u][3][(][k][ −] [1)][ +][ u][3][(][k][)]
1 + y[2]3[(][k][ −] [1)][ +][ y][2]3[(][k][)]
+u[3]3[(][k][)][,]
_y4(k + 1) =_ _[y][4][(][k][)][u][4][(][k][)]_ + 2u4(k),
1 + y[6]4[(][k][)]
and the initial system outputs of the four follower agents are
set as y1(1) = y2(1) = y3(1) = y4(1) = 0. Furthermore,
we would like to point out that the dynamic models of the
simulated MAS are only for generating the input-output data,
and are not involved in the control system design.
_A. EXAMPLE 1: FIXED COMMUNICATION TOPOLOGY_
Fig. 1 shows the communication topology, where the leader
is described by vertex 0, and only agents 1 and 4 receive the
leader’s command. The communication among neighboring
agents is depicted by solid arrows. We use 0 and 1 as weights
for the information communicated between two adjacent
follower agents, therefore the Laplacian matrix among the
follower agents is given by
2 1 1 0
− −
0 1 0 1
−
1 0 1 0
−
0 1 1 2
− −
_,_
_LLL_
=
-----
**FIGURE 2. Tracking performance (Example 1).**
**FIGURE 3. Tracking error (Example 1).**
**FIGURE 4. Control gains (Example 1).**
bound of ψq(k) is given as bp = 1. Thus it can be obtained:
1
{0.2, 0.15} < 1 × �q=max1,···,4 �4p=1 _[a][q][,][p][ +][ d][q]�_
1
(36)
=
1 3
× [≈] [0][.][3][,]
which indicates that the convergence condition (21) of Theorem 1 is satisfied.
We set the leader’s command as yd (k) = 6, and the
simulation is executed with 120 time instants. The simulation
results are shown in Figs. 2–4. Fig. 2 and Fig. 3 are the
tracking performances and tracking errors of the four follower
agents, respectively. Fig. 4 shows the updating values of the
control gains for the four follower agents.
It is obvious that the system outputs of the four follower
agents have large deviations from the leader’s command at
**FIGURE 5. Switching communication topologies.**
**FIGURE 6. Tracking performance (Example 2).**
**FIGURE 7. Tracking error (Example 2).**
the primary time instants, but all the tracking errors of the
four agents gradually decrease and the consensus tracking is
basically achieved after 60 time instants. Furthermore, we can
conclude from Fig. 4 that the proposed CFDL-DAC keeps
automatically tuning and updating the control gains for the
four follower agents to search the optimal values before the
consensus tracking is achieved.
_B. EXAMPLE 2: SWITCHING COMMUNICATION_
_TOPOLOGIES_
In this subsection, we represent that the proposed
CFDL-DAC also works well under switching communication
topologies. The communication topologies switch randomly
among three graphs, which are described by the set Gt =
{G1, G2, G3}, as shown in Fig. 5. Fig. 5 shows that each graph
-----
**FIGURE 8. Control gain (Example 2).**
**FIGURE 9. Tracking performance for fixed topology (Example 3).**
**FIGURE 10. Tracking error for fixed topology (Example 3).**
of the three communication topologies is strongly connected.
The bound of θq(k) and ψq(k) are respectively set as bt =
0.12 and bp = 2, therefore it can be obtained that
1 1
0.12 < 2 × �q=max1,···,4 �4p=1 _[a][q][,][p][ +][ d][q]�_ = 2 × 3 [≈] [0][.][17][,]
(37)
which indicates that the convergence condition (32) for
Corollary 1 is satisfied.
We set the leader’s command as yd (k) = 4, and the simulation results are shown in Figs. 6–8. It is observed that the consensus tracking is achieved, and the tracking errors of all the
follower agents converge to zero after 120 time instants which
verifies the result of Corollary 1. The automatic tuning of the
control gain, as shown in Fig. 8, contributes to the ability of
tracking the leader’s command for the proposed CFDL-DAC
even under the switching communication topologies.
**FIGURE 11. Tracking performance for switching topologies (Example 3).**
**FIGURE 12. Tracking error for switching topologies (Example 3).**
_C. EXAMPLE 3: TIME-VARYING LEADER’s COMMAND_
To further demonstrate the effectiveness of the proposed
approach, in this subsection we consider the time-varying
leader’s command described by
_yd_ (k) = 3 + 2 sin(0.9kπ/260) + cos(kπ/240), (38)
and the simulation is executed with 700 time instants. In this
simulation, the fixed graph as shown in Fig. 1 and the switching topologies as depicted in Fig. 5 are considered.
The simulation results are presented in Figs. 9-12. These
results show that the system outputs of the four follower
agents rapidly approximate to the neighborhood of the
leader’s command from a large deviation at the initial time
instant. Although the tracking errors do not converge to zero,
they reduce to a small bound.
**V. CONCLUSION**
This paper investigated a distributed leader-follower consensus tracking approach for a class of unknown nonlinear nonaffine discrete-time MASs. A data-driven distributed adaptive
control scheme was designed using only the local measurements exchanged among neighboring agents via the dynamic
linearization method applied to the controlled MAS and the
ideal distributed controller. The stabilities of the proposed
distributed adaptive control approach were rigorously guaranteed under both the fixed and switching communication
topologies. In future, investigating a more general distributed
adaptive control scheme and analyzing its stability properties
for a time-varying leader’s command are interesting topics.
-----
**REFERENCES**
[1] F. L. Lewis, H. Zhang, K. Hengster-Movric, and A. Das, Cooperative Con_trol of Multi-Agent Systems: Optimal and Adaptive Design Approaches._
Berlin, Germany: Springer-Verlag, 2013.
[2] W. Ren and R. Beard, Distributed Consensus in Multi-vehicle Cooperative
_Control: Theory and Applications. London, U.K.: Springer-Verlag, 2008._
[3] Y.-J. Zhou, G.-P. Jiang, F.-Y. Xu, and Q.-Y. Chen, ‘‘Distributed finite
time consensus of second-order multi-agent systems via pinning control
(August 2018),’’ IEEE Access, vol. 6, pp. 45617–45624, 2018.
[4] J.-Q. Liang, X.-H. Bu, Q.-F. Wang, and H. He, ‘‘Iterative learning consensus tracking control for nonlinear multi-agent systems with randomly
varying iteration lengths,’’ IEEE Access, vol. 7, pp. 158612–158622, 2019.
[5] Z. Gao, L. Wang, and S. Qiao, ‘‘Cucker-smale flocking control of leaderfollower multiagent systems with intermittent communication,’’ IEEE
_Access, vol. 7, pp. 172676–172682, 2019._
[6] M. Khalili, X. Zhang, Y. Cao, M. M. Polycarpou, and T. Parisini, ‘‘Distributed fault-tolerant control of multiagent systems: An adaptive learning approach,’’ IEEE Trans. Neural Netw. Learn. Syst., vol. 31, no. 2,
[pp. 420–432, Feb. 2020, doi: 10.1109/TNNLS.2019.2904277.](http://dx.doi.org/10.1109/TNNLS.2019.2904277)
[7] F. Chen and W. Ren, ‘‘On the control of multi-agent systems: A survey,’’
_Found. Trends Syst. Control, vol. 6, no. 4, pp. 339–499, 2019._
[8] A. Dorri, S. S. Kanhere, and R. Jurdak, ‘‘Multi-agent systems: A survey,’’
_IEEE Access, vol. 6, pp. 28573–28593, 2018._
[9] A. Jadbabaie, J. Lin, and A. S. Morse, ‘‘Coordination of groups of mobile
autonomous agents using nearest neighbor rules,’’ IEEE Trans. Autom.
_Control, vol. 48, no. 6, pp. 988–1001, Jun. 2003._
[10] R. Olfati-Saber and R. M. Murray, ‘‘Consensus problems in networks of
agents with switching topology and time-delays,’’ IEEE Trans. Autom.
_Control, vol. 49, no. 9, pp. 1520–1533, Sep. 2004._
[11] L. Zhao, J. Yu, C. Lin, and H. Yu, ‘‘Distributed adaptive fixed-time consensus tracking for second-order multi-agent systems using modified terminal
sliding mode,’’ Appl. Math. Comput., vol. 312, pp. 23–35, Nov. 2017.
[12] K. You and L. Xie, ‘‘Network topology and communication data rate
for consensusability of discrete-time multi-agent systems,’’ IEEE Trans.
_Autom. Control, vol. 56, no. 10, pp. 2262–2275, Oct. 2011._
[13] W. Ren and R. W. Beard, ‘‘Consensus seeking in multiagent systems
under dynamically changing interaction topologies,’’ IEEE Trans. Autom.
_Control, vol. 50, no. 5, pp. 655–661, May 2005._
[14] M. di Bernardo, A. Salvi, and S. Santini, ‘‘Distributed consensus strategy
for platooning of vehicles in the presence of time-varying heterogeneous
communication delays,’’ IEEE Trans. Intell. Transp. Syst., vol. 16, no. 1,
pp. 102–112, Feb. 2015.
[15] Y. Zheng and L. Wang, ‘‘Consensus of heterogeneous multi-agent systems without velocity measurements,’’ Int. J. Control, vol. 85, no. 7,
pp. 906–914, Jul. 2012.
[16] Z. Hou, R. Chi, and H. Gao, ‘‘An overview of dynamic-linearization-based
data-driven control and applications,’’ IEEE Trans. Ind. Electron., vol. 64,
no. 5, pp. 4076–4090, May 2017.
[17] Z. Hou, ‘‘Parameter identication, adaptive control and model-free learning adaptive control for nonlinear systems,’’ Ph.D. dissertation, School
Automat., Northeastern Univ., Shenyang, China, 1994.
[18] Z. Hou and S. Xiong, ‘‘On model-free adaptive control and its stability
analysis,’’ IEEE Trans. Autom. Control, vol. 64, no. 11, pp. 4555–4569,
[Nov. 2019, doi: 10.1109/TAC.2019.2894586.](http://dx.doi.org/10.1109/TAC.2019.2894586)
[19] Z. Hou and S. Jin, ‘‘A novel data-driven control approach for a class
of discrete-time nonlinear systems,’’ IEEE Trans. Control Syst. Technol.,
vol. 19, no. 6, pp. 1549–1558, Nov. 2011.
[20] D. Xu, B. Jiang, and P. Shi, ‘‘Adaptive observer based data-driven control
for nonlinear discrete-time processes,’’ IEEE Trans. Autom. Sci. Eng.,
vol. 11, no. 4, pp. 1037–1045, Oct. 2014.
[21] Z. Hou and S. Jin, ‘‘Data-driven model-free adaptive control for a class
of MIMO nonlinear discrete-time systems,’’ IEEE Trans. Neural Netw.,
vol. 22, no. 12, pp. 2173–2188, Dec. 2011.
[22] D. Xu, B. Jiang, and P. Shi, ‘‘A novel model-free adaptive control design
for multivariable industrial processes,’’ IEEE Trans. Ind. Electron., vol. 61,
no. 11, pp. 6391–6398, Nov. 2014.
[23] X. Bu, Z. Hou, and H. Zhang, ‘‘Data-driven multiagent systems consensus
tracking using model free adaptive control,’’ IEEE Trans. Neural Netw.
_Learn. Syst., vol. 29, no. 5, pp. 1514–1524, May 2018._
[24] X. Bu, Q. Yu, Z. Hou, and W. Qian, ‘‘Model free adaptive iterative
learning consensus tracking control for a class of nonlinear multiagent systems,’’ IEEE Trans. Syst., Man, Cybern. Syst., vol. 49, no. 4, pp. 677–686,
Apr. 2019.
[25] R. Chi, Y. Hui, B. Huang, and Z. Hou, ‘‘Adjacent-agent dynamic
linearization-based iterative learning formation control,’’ _IEEE_
_Trans. Cybern., vol. 50, no. 10, pp. 4358–4369, Oct. 2020, doi:_
[10.1109/tcyb.2019.2899654.](http://dx.doi.org/10.1109/tcyb.2019.2899654)
[26] Z.-H. Pang, G.-P. Liu, D. Zhou, and D. Sun, ‘‘Data-based predictive
control for networked nonlinear systems with network-induced delay and
packet dropout,’’ IEEE Trans. Ind. Electron., vol. 63, no. 2, pp. 1249–1257,
Feb. 2016.
[27] X. Wang, X. Li, J. Wang, X. Fang, and X. Zhu, ‘‘Data-driven modelfree adaptive sliding mode control for the multi degree-of-freedom robotic
exoskeleton,’’ Inf. Sci., vol. 327, pp. 246–257, Jan. 2016.
[28] D. Meng and Y. Jia, ‘‘Iterative learning approaches to design finite-time
consensus protocols for multi-agent systems,’’ Syst. Control Lett., vol. 61,
no. 1, pp. 187–194, Jan. 2012.
[29] Z. Hou and S. Jin, Model Free Adaptive Control Theory and Applications.
New York, NY, USA: CRC Press, 2013.
[30] Y. Zhu and Z. Hou, ‘‘Data-driven MFAC for a class of discrete-time
nonlinear systems with RBFNN,’’ IEEE Trans. Neural Netw. Learn. Syst.,
vol. 25, no. 5, pp. 1013–1020, May 2014.
[31] T. Ma, F. L. Lewis, and Y. Song, ‘‘Exponential synchronization of nonlinear
multi-agent systems with time delays and impulsive disturbances,’’ Int. J.
_Robust Nonlinear Control, vol. 26, no. 8, pp. 1615–1631, May 2016._
[32] Z. Li, W. Ren, X. Liu, and M. Fu, ‘‘Consensus of multi-agent systems with
general linear and lipschitz nonlinear dynamics using distributed adaptive
protocols,’’ IEEE Trans. Autom. Control, vol. 58, no. 7, pp. 1786–1791,
Jul. 2013.
[33] Z. Hou and Y. Zhu, ‘‘Controller-dynamic-linearization-based model free
adaptive control for discrete-time nonlinear systems,’’ IEEE Trans. Ind.
_Informat., vol. 9, no. 4, pp. 2301–2309, Nov. 2013._
[34] G. Wang, ‘‘Distributed control of higher-order nonlinear multi-agent systems with unknown non-identical control directions under general directed
graphs,’’ Automatica, vol. 110, Dec. 2019, Art. no. 108559.
[35] G. Wang, C. Wang, and X. Cai, ‘‘Consensus control of output-constrained
multiagent systems with unknown control directions under a directed
graph,’’ Int. J. Robust Nonlinear Control, vol. 30, no. 5, pp. 1802–1818,
Mar. 2020.
[36] S. Yang, J.-X. Xu, and X. Li, ‘‘Iterative learning control with input
sharing for multi-agent consensus tracking,’’ Syst. Control Lett., vol. 94,
pp. 97–106, Aug. 2016.
XIAN YU received the B.S. degree in mechanical
engineering and automation and the M.S. degree
in mechatronic engineering from Guangxi University, Nanning, China, in 2012 and 2015, respectively. He is currently pursuing the Ph.D. degree
with the Advanced Control Systems Laboratory,
Beijing Jiaotong University, Beijing, China.
From 2019 to 2020, he was a Visiting
Researcher of the KIOS Research and Innovation Center of Excellence, University of Cyprus,
Nicosia, Cyprus. His current research interests include data-driven control,
multi-agent systems, adaptive control, model free adaptive control, and
iterative learning control.
SHANGTAI JIN received the bachelor’s, master’s,
and Ph.D. degrees from Beijing Jiaotong University, Beijing, China, in 1999, 2004, and 2009,
respectively.
He is currently an Associate Professor with
Beijing Jiaotong University. His current research
interests include model-free adaptive control, data
driven control, learning control, and intelligent
transportation systems.
-----
GENFENG LIU received the bachelor’s degree
in electric engineering and automation and the
master’s degree in control science and engineering from Henan Polytechnic University, Jiaozuo,
China, in 2012 and 2015, respectively. He is currently pursuing the Ph.D. degree in control science and engineering with the Advanced Control
Systems Laboratory, Beijing Jiaotong University,
Beijing, China.
His research interests include data-driven control, iterative learning control, model free adaptive control, adaptive control,
fault-tolerant control, multiagent systems, and train control systems.
TING LEI received the bachelor’s degree
from Zhengzhou University, Zhengzhou, China,
in 2012. He is currently pursuing the Ph.D. degree
with the Advanced Control Systems Laboratory,
Beijing Jiaotong University, Beijing, China.
He is also working with the College of Electrical and Information Engineering, Zhengzhou
University of Light Industry, Zhengzhou. His current research interests include urban transportation
systems, data-driven control, and optimization and
control of large scale networks.
YE REN received the bachelor’s degree from
Beijing Jiaotong University, Beijing, China,
in 2013. He is currently pursuing the Ph.D. degree
with the Advanced Control Systems Laboratory.
He is also working with the School of Electrical
and Control Engineering, North China University of Technology, Beijing. His current research
interests include optimization and control of large
scale networks, data-driven control, and multiagent systems control.
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://api.unpaywall.org/v2/10.1109/ACCESS.2020.3038629?email=<INSERT_YOUR_EMAIL> or https://doi.org/10.1109/ACCESS.2020.3038629, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": "CCBYNCND",
"status": "GOLD",
"url": "https://ieeexplore.ieee.org/ielx7/6287639/8948470/09261580.pdf"
}
| 2,020
|
[
"JournalArticle"
] | true
| null |
[
{
"paperId": "dcd12a6d27fc79cad08ed283c38dfa03384fea2e",
"title": "Adjacent-Agent Dynamic Linearization-Based Iterative Learning Formation Control"
},
{
"paperId": "695d0a78140696a410b0f9ecd6655e35e12e4e54",
"title": "Distributed Fault-Tolerant Control of Multiagent Systems: An Adaptive Learning Approach"
},
{
"paperId": "285e8df98886ebfa3d13bba9e22207d5808ba32d",
"title": "Consensus control of output‐constrained multiagent systems with unknown control directions under a directed graph"
},
{
"paperId": "6b675f32e60383ee98177384f0666b5e1217b30c",
"title": "Distributed control of higher-order nonlinear multi-agent systems with unknown non-identical control directions under general directed graphs"
},
{
"paperId": "32dffdf55ac7d340aff4eccd491d5a89d4f58e62",
"title": "Iterative Learning Consensus Tracking Control for Nonlinear Multi-Agent Systems With Randomly Varying Iteration Lengths"
},
{
"paperId": "e2a7b6f5e817a70a1a1645ba1a964626d3c6002e",
"title": "On the Control of Multi-Agent Systems: A Survey"
},
{
"paperId": "1de86a998a07bdd339dabd51e6703605cb6df466",
"title": "Model Free Adaptive Iterative Learning Consensus Tracking Control for a Class of Nonlinear Multiagent Systems"
},
{
"paperId": "b9073df6b915c0ebf92277b71cd2e3c805c29316",
"title": "On Model-Free Adaptive Control and Its Stability Analysis"
},
{
"paperId": "ca872af455a646e22bf3d016d0c90f751140c56e",
"title": "Data-Driven Multiagent Systems Consensus Tracking Using Model Free Adaptive Control"
},
{
"paperId": "04e348d70aab71b6b346e45e632136380ac5a88f",
"title": "Distributed adaptive fixed-time consensus tracking for second-order multi-agent systems using modified terminal sliding mode"
},
{
"paperId": "0fe543ca1755ed50e2e110e8827ba68b6f18d68e",
"title": "An Overview of Dynamic-Linearization-Based Data-Driven Control and Applications"
},
{
"paperId": "3d9a14f21b8969a5bebf90394b11fb5043625e4b",
"title": "Iterative learning control with input sharing for multi-agent consensus tracking"
},
{
"paperId": "f650e7d22aacd1c08e9bb61d4d457be31c2859ce",
"title": "Exponential synchronization of nonlinear multi‐agent systems with time delays and impulsive disturbances"
},
{
"paperId": "51c3b8d8d42b85bb3bd3c2707d21223f81758903",
"title": "Data-Based Predictive Control for Networked Nonlinear Systems With Network-Induced Delay and Packet Dropout"
},
{
"paperId": "40a02aaacc1efe0b9a8bc937fa07fae3856cae72",
"title": "Data-driven model-free adaptive sliding mode control for the multi degree-of-freedom robotic exoskeleton"
},
{
"paperId": "e058a1cf36df18c2e3fbe82224b6fc2c0d9df6ba",
"title": "Distributed Consensus Strategy for Platooning of Vehicles in the Presence of Time-Varying Heterogeneous Communication Delays"
},
{
"paperId": "8d5d83d8b56c18d1290e908cb61b57425d2530fa",
"title": "Data-Driven MFAC for a Class of Discrete-Time Nonlinear Systems With RBFNN"
},
{
"paperId": "df00836a72b09da059055713691d44267570c3ba",
"title": "A Novel Model-Free Adaptive Control Design for Multivariable Industrial Processes"
},
{
"paperId": "0851aa9a4173c9bf11c87e865a084e43c9f2d3f2",
"title": "Cooperative Control of Multi-Agent Systems: Optimal and Adaptive Design Approaches"
},
{
"paperId": "67c0b6026e506ade0aa999b752cdff83cd7b6a22",
"title": "Model Free Adaptive Control: Theory and Applications"
},
{
"paperId": "cbb4898de5545a509442128f118fc067e7982283",
"title": "Controller-Dynamic-Linearization-Based Model Free Adaptive Control for Discrete-Time Nonlinear Systems"
},
{
"paperId": "d3426bc252fa5aa29f7d098754e8fd337c686f1f",
"title": "Consensus of heterogeneous multi-agent systems without velocity measurements"
},
{
"paperId": "f693c9a254c841cd7eb0edf243e9f8d3af450660",
"title": "Data-Driven Model-Free Adaptive Control for a Class of MIMO Nonlinear Discrete-Time Systems"
},
{
"paperId": "2590f90b0245059e828b9c2598625720e559bed9",
"title": "Consensus of Multi-Agent Systems With General Linear and Lipschitz Nonlinear Dynamics Using Distributed Adaptive Protocols"
},
{
"paperId": "0122e7e784aae390dcc230f23bc79e8d19073092",
"title": "Network Topology and Communication Data Rate for Consensusability of Discrete-Time Multi-Agent Systems"
},
{
"paperId": "bbadd4370cb353cd31a260668824f27a41c9af56",
"title": "Distributed Consensus in Multi-vehicle Cooperative Control - Theory and Applications"
},
{
"paperId": "ee6ff99457245822545a32c436f359cac84ccb68",
"title": "Consensus seeking in multiagent systems under dynamically changing interaction topologies"
},
{
"paperId": "9839ed2281ba4b589bf88c7e4acc48c9fa6fb933",
"title": "Consensus problems in networks of agents with switching topology and time-delays"
},
{
"paperId": "20be6a4e06295792977d2d6e4c9eb9a8405226e9",
"title": "Coordination of groups of mobile autonomous agents using nearest neighbor rules"
},
{
"paperId": "a6ad73fa48ec6ef753a1572cc076479fdca06b01",
"title": "Cucker-Smale Flocking Control of Leader-Follower Multiagent Systems With Intermittent Communication"
},
{
"paperId": "993fc07af995c9900a2682103842c58053290ae7",
"title": "Distributed Finite Time Consensus of Second-Order Multi-Agent Systems via Pinning Control (August 2018)"
},
{
"paperId": "737f439d1b90b018a36130f48c084e87b8bb413b",
"title": "Adaptive Observer Based Data-Driven Control for Nonlinear Discrete-Time Processes"
},
{
"paperId": "8cfdf6b9d95e8a4a418da0cb9afc9c6739c27a53",
"title": "Iterative learning approaches to design finite-time consensus protocols for multi-agent systems"
},
{
"paperId": "f0fb49cebb954213b767452757d75c748887fe74",
"title": "A Novel Data-Driven Control Approach for a Class of Discrete-Time Nonlinear Systems"
},
{
"paperId": "f16a0dcbf96f5a6b41ce1855c68e9f67a1514a10",
"title": "Multi-Agent Systems: A Survey"
},
{
"paperId": null,
"title": "‘‘Parameter identication, adaptive control and model-free learning adaptive control for nonlinear systems,’’"
}
] | 13,851
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/02b768bce3ec87af9443f2bb1d890ce09f1ca916
|
[
"Computer Science"
] | 0.855051
|
Semi-Decentralized Federated Learning with Collaborative Relaying
|
02b768bce3ec87af9443f2bb1d890ce09f1ca916
|
International Symposium on Information Theory
|
[
{
"authorId": "31525269",
"name": "M. Yemini"
},
{
"authorId": "8352011",
"name": "R. Saha"
},
{
"authorId": "120677219",
"name": "Emre Ozfatura"
},
{
"authorId": "1727814",
"name": "Deniz Gündüz"
},
{
"authorId": "1746299",
"name": "A. Goldsmith"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"International Symposium on Information Technology",
"Int Symp Inf Theory",
"Int Symp Inf Technol",
"ISIT"
],
"alternate_urls": null,
"id": "234ccdc0-f58f-4f94-b86a-428d11a0c5ad",
"issn": null,
"name": "International Symposium on Information Theory",
"type": "conference",
"url": "http://www.wikicfp.com/cfp/program?id=1719"
}
|
We present a semi-decentralized federated learning algorithm wherein clients collaborate by relaying their neighbors’ local updates to a central parameter server (PS). At every communication round to the PS, each client computes a local consensus of the updates from its neighboring clients and eventually transmits a weighted average of its own update and those of its neighbors to the PS. We appropriately optimize these averaging weights to ensure that the global update at the PS is unbiased and to reduce the variance of the global update at the PS, consequently improving the rate of convergence. Numerical simulations substantiate our theoretical claims and demonstrate settings with intermittent connectivity between the clients and the PS, where our proposed algorithm shows an improved convergence rate and accuracy in comparison with the federated averaging algorithm.
|
# Semi-Decentralized Federated Learning with Collaborative Relaying
### Michal Yemini Rajarshi Saha Emre Ozfatura Deniz G¨und¨uz Andrea J. Goldsmith
Princeton University Stanford University Imperial College London Imperial College London Princeton University
**_Abstract—We present a semi-decentralized federated learning_**
**algorithm wherein clients collaborate by relaying their neighbors’**
**local updates to a central parameter server (PS). At every**
**communication round to the PS, each client computes a local**
**consensus of the updates from its neighboring clients and**
**eventually transmits a weighted average of its own update and**
**those of its neighbors to the PS. We appropriately optimize**
**these averaging weights to ensure that the global update at**
**the PS is unbiased and to reduce the variance of the global**
**update at the PS, consequently improving the rate of convergence.**
**Numerical simulations substantiate our theoretical claims and**
**demonstrate settings with intermittent connectivity between the**
**clients and the PS, where our proposed algorithm shows an**
**improved convergence rate and accuracy in comparison with the**
**federated averaging algorithm.**
I. INTRODUCTION
Federated learning (FL) algorithms iteratively optimize a
common objective function to learn a shared model over data
samples that are localized over multiple distributed clients
[1]. FL approaches aim to reduce the required communication overhead and improve clients’ privacy by training
local models of private dataset at the clients and forwarding
them periodically to a centralized parameter server (PS). In
practical FL setups, some clients are stragglers and cannot
send their updates regularly, either because: (i) they cannot
finish their computation within a prescribed deadline, or (ii)
due to communication limitations [2], where they suffer from
intermittent connectivity to the PS since their wireless channel
is temporarily blocked [3]–[8]. Stragglers deteriorate the convergence of FL as the computed local updates become stale.
This can even result in bias in the final model in the case of
persistent stragglers. On the other hand, Communication strag_glers (type (ii)) are inherently different from computation-_
_limited stragglers (type (i)), since it can be solved by relaying_
the updates to the PS via neighboring clients.
Communication quality at the wireless edge as a key design
principle is considered in the federated edge learning (FEEL)
framework [9], which takes into account the wireless channel
characteristics from the clients to the PS to optimize the
convergence and final model performance at the PS. So far the
FEEL paradigm has mainly focused on direct communication
from the clients to the PS, and aimed at improving the
M. Yemini, R. Saha, and A.J. Goldsmith are partially supported by the
AFOSR award #002484665 and a Huawei Intelligent Spectrum grant.
E. Ozfatura and D. G¨und¨uz received funding from the European Research
Council (ERC) through Starting Grant BEACON (no. 677854) and the UK
EPSRC (grant no. EP/T023600/1) under the CHIST-ERA program.
performance by resource allocation across clients [9]–[18];
these approaches have ignored possible cooperation between
clients in the case of intermittent communication blockages.
Motivated by our prior works [19]–[21], where client cooperation is used to improve the connectivity to the cloud
and to reduce the latency and scheduling overhead, this work
proposes a new FEEL paradigm, where the clients cooperate to
mitigate the detrimental effects of communication stragglers.
In our proposed method, clients share their local updates with
neighbors so that each client sends to the PS a weighted
average of its current update and those of its neighbors. Using
this approach, the PS receives new updates from disconnected
clients, which would otherwise become stale and be discarded.
We demonstrate the success of our relaying scheme through
both theoretical analysis and numerical simulations.
_Related Works:_ Decentralized collaborative learning
frameworks have been introduced as an alternative to centralized FL, in which the PS is removed to mitigate a potential
communication bottleneck and a single point of failure [22]–
[33]. In decentralized learning, each client shares its local
model with its neighbors through device-to-device (D2D)
communications, and model aggregation is executed at each
client in parallel. This aggregation strategy is determined
at each client according to the network topology, i.e., the
connection pattern between the clients.
An alternative approach to both centralized and decentralized schemes is hierarchical FL (HFL) [21], [34]–[36], where
multiple PSs are employed for the aggregation to prevent a
communication bottleneck. In HFL, clients are divided into
clusters and a PS is assigned to each cluster to perform
local aggregation. The aggregated models at the clusters are
later aggregated at the main PS in a subsequent step to
obtain the global model. HFL has significant advantages over
centralized and decentralized schemes, particularly when the
communication takes place over wireless channels since it
allows spatial reuse of available resources [21]. Nonetheless,
HFL requires employing multiple PSs that may not be practical
in certain scenarios. Instead, the idea of hierarchical collaborative learning can be redesigned to combine hierarchical and
decentralized learning, which is referred as semi-decentralized
_FL, where the local consensus follows decentralized learning_
with D2D communications, whereas the global consensus is
orchestrated by the PS [37], [38]. One of the major challenges
in FL that is not considered in [37], [38] is the partial
client connectivity [39], [40]. Unequal client participation due
-----
to intermittent connectivity exacerbates the impact of data
heterogeneity [41]–[44], and increases the generalization gap.
Most existing works on FL assume error-free rate-limited
orthogonal communication links, with an underlying communication protocol that takes care of wireless imperfections.
However, this separation between the communication medium
and the learning protocol can be suboptimal [9]. An alternative
approach treats the communication of the model updates to the
PS as an uplink communication problem and jointly optimizes
the learning algorithm and the communication scheme [9].
Within this framework is an original and promising approach
known as over-the-air computation (OAC) [14]–[16], which
exploits the superposition property of wireless signals to
convey the sum of the model updates that are transmitted by
each client in an uncoded fashion. In addition to bandwidth
efficiency, the OAC framework provides a certain level of
anonymity to clients due to its superposition nature; hence,
it can enhance the privacy of the participating clients [17],
[18]. We emphasize that in OAC, PS receives the aggregate
model, and it is not possible to disentangle the individual
model updates. Therefore, any strategy that utilizes a PS
side aggregation mechanism with individual model updates to
address unequal client participation is not compatible with the
OAC framework. One of the major advantages of our proposed
scheme is that it mitigates the drawbacks of unequal client
participation without requiring the identity of transmitting
clients or their individual updates at the PS. Therefore, our
solution is compatible with OAC.
Client connectivity is a particularly significant challenge
in FEEL, where the clients and the PS communicate over
unreliable wireless channels. Due to their different physical
environments and distances to the PS, clients can have different connectivity to each other and the PS. This problem
has been recently addressed in [10]–[13], [45]–[48] by considering customized client selection mechanisms to balance
the participation of the clients and the latency for the model
aggregation in order to speed up the learning process. We
adopt a different approach to this problem, where instead of
designing a client selection mechanism, or optimizing resource
allocation to balance client participation, we introduce a relay_ing mechanism that takes into account the nature of individual_
clients’ connectivity to the PS and ensures that, in case of poor
connectivity, their local updates are conveyed to the PS with
the help of their neighboring clients.
_Paper Organization: Sec. II presents the FL system_
model and the proposed FL collaborative relaying scheme.
Sec. III presents conditions for the unbiasedness of our proposed scheme and an analysis of the convergence rate. Sec. IV
optimizes the convergence rate of our proposed scheme, while
Sec. V presents numerical results that validate our theoretical
analysis and highlight the performance improvement in terms
of training accuracy. Finally, Sec. VI concludes this paper.
**Remark. Due to space limitations, all proofs are omitted, and**
_can be found in an online extended version of this paper [49]._
PS
_c1_
_c3_
_c8_
_c9_
_c5_ _c9_
Fig. 1: System model with intermittent uplink communication between clients and PS (dotted lines) and reliable communication
between neighboring clients (solid lines).
II. SYSTEM MODEL FOR COLLABORATIVE RELAYING
Consider n clients communicating periodically with a PS
that trains a global model x ∈ R[d]. Let L(x, ζ) be the loss
evaluated for a model x at data point ζ. Denote the local
loss at client1 � _i by fi : R[d]_ _× Zi →_ R, where f (x; Zi) =
_|Zi|_ **_ζ∈Zi_** _[L][(][x][;][ ζ][)][. Here,][ Z][i][ is the local dataset of client]_
_i. The goal of PS is to solve the following empirical risk_
minimization (ERM) problem:[1]
1
**x[∗]** = arg minx∈R[d] _·_ _f_ (x) ≜ arg minx∈R[d] _·_ _n_
_A. FL with Local SGD at Clients_
_n_
�
_fi(x; Zi)._
_i=1_
Denote the local gradient as ∇fi(x) ≜ _∇xf_ (x; Zi), and let
_gi(x) be an unbiased estimate of it. In the r[th]_ _round of FL,_
the PS broadcasts the global model x[(][r][)] to the clients. For a
local averaging period of, each client performs iterations
_T_ _T_
of local training, after which the local models are sent to the
PS for aggregation. For local iteration k [0 : ] of the r[th]
_∈_ _T_
round, client i applies the local update rule:
� �
**x[(]i[r,k][+1)]** = x[(]i[r,k][)] _−_ _ηrgi_ **x[(]i[r,k][)]** _,_ (1)
where ηr is the learning rate for round r and x[(]i[r,][0)] = x[(][r][)].
_B. Communication Model_
**Communication between clients and PS. We consider a**
setting where the uplink connections between the clients and
the PS are intermittent. As shown in Fig. 1, we model the
connectivity of client i to the PS at round r by the Bernoulli
random variable τi(r) ∼ Bern(pi), where τi = 1 indicates
the presence of an uplink communication opportunity, whereas
_τi(r) = 0 indicates a blocked uplink. For simplicity of expo-_
sition, we consider the uplink channel to be either completely
blocked or perfectly available without any noise, and the
downlink from PS to clients does not suffer from intermittent
dropouts.
**Remark 1. The connectivity probabilities {pi}i∈[n] can be**
_easily estimated using pilot signals. Moreover, clients can_
_share their pi with each other using local links in a pre-_
_training phase. On the other hand, we do not assume that_
1For simplicity, we assume |Zi| = |Zj _| for all i, j ∈_ [n]. Our method can
be extended to the setting of imbalanced local dataset sizes as well.
-----
**Algorithm 1 COLREL-CLIENT: Collaborative Relaying**
**Input: Round index r, Step-size ηr, Local avg. period T,**
Neighborhood of client i Ni, αij for every j ∈Ni ∪{i}.
**Output: ∆x�[r]i** [+1].
1: Receive x[(][r][)] from PS.
2: Set x[(]i[r,][0)] = x[(][r][)].
3: for k ← 0 to T − 1 do
Compute (stochastic) gradient gi(x[(]i[r,k][)]t).
� �
**xi[(][r,k][+1)]** = x[(]i[r,k][)] _−_ _ηrgi_ **x[(]i[r,k][)]** .
4: end for
5: Set ∆x[r]i [+1] = x[(]i[r,][T][ )] _−_ **x[(][r][)].**
6: Send ∆xi to every j ∈Ni.
7: Receive ∆xj from every j ∈Ni.
8: Compute ∆x�[r]i [+1] = [�]j∈Ni∪{i} _[α][ij][ ·][ ∆][x]j[r][+1]._
9: Transmit (relay) ∆x�[r]i [+1] to the PS.
_the instantaneous connectivity information, i.e., τi(r), r ∈_ [n]
_is available to any of the clients._
**Communication between clients. The connectivity be-**
tween clients is modeled by an undirected graph G = (V, E)
where V = [n] and (i, j) _E_ client i can communicate
_∈_ _⇐⇒_
with client j. Let Ni ≜ _{j ∈_ _V_ : {i, j} ∈ _E}. We do_
not assume that the graph G is connected. Instead, it can be
composed of multiple connected subgraphs.
_C. Collaborative Relaying of Local Updates_
Let ∆x[(]i[r][+1)] denote client i’s update at the end of T _[th]_
local iteration of round r, i.e., ∆x[(]j[r][+1)] = x[(]j[r,][T][ )] _−_ **x[(][r][)]. We**
assume that client i’s update ∆x[(]j[r][+1)] is readily available to
its neighbors. Then client i computes a weighted average of
its own update and those of its neighbors in Ni, i.e.,
**Algorithm 2 COLREL-PS: PS Aggregation**
**Input: Number of rounds R, a set of clients [n].**
**Output: Global model x[(][R][)].**
1: Set x[(0)] = 0
2: for k ← 0 to T − 1 do
Broadcast x[(][r][)] to all clients.
Set τi(r + 1) = 1 or 0 depending on connectivity.
Update x[(][r][+1)] = x[(][r][)] + _n[1]_ �i∈[n] _[τ][i][(][r][ + 1)∆][x][�]i[r][+1]_
3: end for
III. CONVERGENCE ANALYSIS
_A. Unbiasedness of COLREL FL_
In our collaborative relaying scheme, the local update of a
particular client i can be transmitted to the PS by itself, or
by one or more of its neighbors j ∈Ni. Since the PS may
be blind to the identities of the clients, the clients collaborate
among themselves to ensure that this redundancy is mitigated.
This is done by appropriately choosing the weights αij. In
particular, Lemma 1 gives a sufficient condition on the values
of {αij}i,j∈[n] that ensures that the aggregated global update
at the PS is an unbiased estimate of _n[1]_ �i∈[n] [∆][x]i[(][r][)][, the true]
aggregate in the case of perfect channel connectivity.
**Lemma 1. Let w = 1/n and {αij}i,j∈[n] be such that**
�
= piαii + _pjαji = 1._ (3)
_j∈Ni_
E
�
_τj(r + 1)αji_
_j∈Ni∪{i}_
_Then, for every i_ [n],
_∈_
� _r+1_
_τj(r + 1)αji∆x[r]i_ [+1]���∆xi
_j∈Ni∪{i}_
= [1] _i_ _._
_n_ [∆][x][r][+1]
� � �
∆x�[(]i[r][+1)] = _αij∆x[(]j[r][+1)]_ = _αij_ **x[(]j[r,][T][ )]** _−_ **x[(][r][)][�]** _,_
_j∈Ni∪{i}_ _j∈Ni∪{i}_
where αij is a non-negative importance weight assigned by
client i while relaying the client j’s update. Note that weighted
averaging entails a complexity of O �maxi∈[n] |Ni| + 1�.
_D. PS Aggregation_
In our setting, the PS does not explicitly select the subset
of clients from which it wants to receive information, rather
it receives updates from all communicating clients at the
beginning of every round. The PS uses the following re-scaled
sum of received updates:
�
**x[(][r][+1)]** = x[(][r][)] + w _τi(r + 1)∆x�[(]i[r][+1)]._ (2)
_i∈[n]_
This update can be computed over-the-air and does not require
the PS to know the identities of the communicating clients.
We set w=1/n to preserve the unbiasedness of the objective
function at the PS, as discussed in Sec. III. Our Collaborative
Relaying (ColRel) algorithm is presented in Algs. 1 and 2.
Note that the standard FL setting under intermittent client
connectivity to the PS but with no collaboration between the
clients is captured by the choice w = 1/n, Ni = ∅, pi =
_p, αii = 1, αij = 0 for all i, j ∈_ [n] and j ̸= i.
_B. Expected Suboptimality Gap_
Next, Thm. 1 presents the convergence rate of COLREL as
a function of {αij}, under the following assumptions.
**Assumption 1. For any i, the local loss fi is L-smooth w.r.t.**
**x, i.e., for any x, y ∈** R[d], ∥∇fi(x) _−∇fi(y)∥2 ≤_ _L∥x_ _−_ **y∥2.**
**Assumption 2. The stochastic gradients gi(x) are unbiased**
_and have bounded variance, i.e.,_ _i_ [n]:
_∀_ _∈_
_1) E[gi(x)] = ∇fi(x), and_
_2) E∥gi(x) −∇fi(x)∥2[2]_ _[≤]_ _[σ][2][ for some finite][ σ][2][.]_
**Assumption 3. For any i, the loss fi is µ-strongly convex, i.e.,**
_for any x, y ∈_ R[d], (∇fi(x) _−∇fi(y))[⊤](x_ _−_ **y) ≥** _µ∥x_ _−_ **y∥2[2][.]**
_w · E_
-----
**Algorithm 3 OPT-α: Optimization of relay weight matrix A**
**Input: Connectivity graph G, Transmission probability vector**
**p, Maximum number of iterations L.**
**Output: Relay weight matrix A[(][L][)]** that approximately solves
(6).
1
1: Set A[(0)]ji [=] (|Ni|+1)·pj _[·][ 1][{][j][∈N][i][∪{][i][}][:][p][j]_ _[>][0][}][.]_
2: for ℓ _←_ 0 to L − 1 do
Set ℓ _ℓ_ + 1.
_←_
Set i = ℓ mod n + n · 1{ℓ mod n=0}.
(ℓ)
Compute **A[�]** _i_ according to (9).
Set A[(]k[ℓ][)] according to (7) for every k ∈ [n].
3: end for
Let A = (αij)i,j∈[n] denote the n × n matrix of relay
weights, and let Nil = (Ni ∪{i}) ∩ (Nl ∪{l}) denote the
common neighborhood of nodes i and j. Suppose,
Fig. 2: Homogeneous connectivity with pi = 0.2, ∀i ∈ [n] and FCT.
**A[(]i[ℓ][)]** =
� (ℓ)
**A�** _i_ if ℓ mod n + n · 1{ℓ mod n=0} = i, _. (7)_
**A[(]i[ℓ][−][1)]** otherwise
�
_S(p, A) =_
_i,l∈[n]_
�
_pj(1 −_ _pj)αjiαjl._ (4)
_j:j∈Nil_
**Theorem 1. Under Asms. 1-3 and condition (3), COLREL, as**
_specified by Algs. 1 and 2, with ηr =_ _r[4]T[µ] +1[−][1]_ _[, satisfies for every]_
_r ≥_ _r0(p, A),_
E∥x[(][r][+1)] _−_ _x[⋆]∥[2]_ _≤_ [(][r][0][T][ + 1)]
(r + 1)[2][ ∥][x][(0)][ −] _[x][⋆][∥][2][ +][ C][1]k[(][p] + 1[,][ A][)][T]_
_T_ _T_
( 1)[2]
_T −_ _T_
+C2
_k_ + 1 [+][ C][3][(][p][,][ A][)] (k + 1)[2][,]
_T_ _T_
_where B(p, A) =_ [2]n[L][2][2][ S][(][p][,][ A][)][,][ C][1][(][p][,][ A][) =][ 4]µ[2][2][ ·][ 2]n[σ][2][2][ S][(][p][,][ A][)][,]
� �
_C2 =_ _µ[4][2][2][ ·][L][2][ σ]n[2]_ _[e][,][ C][3][(][p][,][ A][) =][ 4]µ[4][4][ ·]_ _L[2]σ[2]e +_ [2][L]n[2][σ][2] [2][e] _S(p, A)_ _,_
_and r0(p, A) = max_ � _Lµ_ _[,][ 4]_ � _B(µp[2],A)_ + 1� _,_ _T[1]_ _[,]_ _µ4[2]nT_ �.
As a consequence of Thm. 1, it follows that,
2
E ���x(r+1) − _x⋆���_ = O
���x(0) − _x⋆��2_
+ _[S][(][p][,][ A][)]_
_r[2]_ _r_
�
_._ (5)
(ℓ)
Here, **A[�]** _i_ is given by
(ℓ) �
**A�** _i_ = arg min _pj(1 −_ _pj)αji[2]_
_j∈Ni∪{i}_
� �
+ 2 _pj(1 −_ _pj)αjiαjl[(][ℓ][−][1)],_
_l∈[n],l≠_ _i_ _j∈Nil_
�
s.t.: _pjαji = 1,_ _αji ≥_ 0 _∀j ∈_ [n]. (8)
_j∈Ni∪{i}_
Let Lji = {l : l ∈ [n], l ̸= i, j ∈Nil}, that is, the set of all
clients that have j as a mutual neighbor with i, and let βji =
�l∈Lji _[α]jl[(][ℓ][−][1)]. Let p(i) = maxk∈Ni∪{i} pk. Using Lagrange_
multipliers we solve (8) for j ∈Ni ∪{i} as follows:
�−βji + 2(1λ−ipj ) �+ if pj ∈ (0, 1), p(i) < 1,
_α�ji[(][ℓ][)]_ [=] 0�k∈[n] [1][{][p]k1[=1][,k][∈N]i _[∪{][i][}}]_ otherwise[if][ p][j][ = 1][,]. (9)
Here, λi satisfies [�]j∈Ni∪{i} _[p][j]_ �−βji + 2(1λ−ipj ) �+ = 1, and
(·)[+] ≜ max{·, 0}. We can find λi using the bisection method.
The complete algorithm is detailed in Alg. 3. Its overall
_complexity is O(L_ (n[2] + K)), where K is the number of
_·_
iterations used in the bisection method for optimizing λi.
**Remark 2. The optimization (6) only requires client i to know**
_the weight values for its neighbors of distance 2. Thus, we can_
_exploit the communication links between clients, and optimize_
(6) distributively. We present the distributed algorithm in [49].
V. NUMERICAL SIMULATIONS
We consider training a ResNet-20 model for image classification on CIFAR-10 dataset over 10 clients; each executes 8
local training steps of local-SGD. All plots have been averaged
over 5 different realizations. We used a learning rate of 0.1 for
SGD, a coefficient of 1e − 4 for ℓ2-regularization to prevent
overfitting, and a batch-size of 64.
In Figs. 2 and 3, the dataset is distributed across the clients
in an IID fashion. As benchmarks, we consider Federated
Therefore, the convergence rate can be improved by minimizing the term S(p, A) subject to the unbiasedness condition in
Lemma 1. Minimizing S(p, A) can also reduce r0(p, A).
IV. OPTIMIZING THE RELAYING WEIGHTS
We choose the relay weight matrix A to minimize the upper
bound on the expected distance to optimality as given by
Thm. 1. Thus, we solve the following optimization problem:
arg min
**A** _[S][(][p][,][ A][)][,]_
�
s.t.: _pjαji = 1,_ _αji ≥_ 0 _∀i, j ∈_ [n]. (6)
_j:j∈Ni_
The function S(p, A) is convex with respect to (w.r.t.) A for
**p** [0, 1][n]. It can be shown that the domain of (6) is separable
_∈_
w.r.t. Ai, the i[th] column of A, and we can use the GaussSeidel method [50, Prop. 2.7.1] to iteratively solve (6). At
every iteration ℓ, we compute the estimate A[ℓ] as
-----
Fig. 3: Different connectivity across clients with a ring topology.
Fig. 4: Non-IID data + global momentum.
_Averaging (FedAvg) - No Dropout, in which all clients are able_
to successfully transmit their local updates to the PS at every
communication round. We also consider FedAvg - Dropout,
in which the PS is unaware of the identity of the clients,
and simply assumes that the update for any client unable
to successfully transmit is zero. These benchmarks serve as
natural upper and lower bounds to the performance of the
proposed algorithm.
In Fig. 2, we have a homogeneous connectivity setup with
equal probability pi = 0.2 that client i ∈ [n] successfully
transmits its local updates to the PS. Furthermore, we assume a fully-connected topology (FCT) where each clients
is connected to all the other clients in the system. COLREL
achieves a performance on par with FedAvg - No Dropout.
We also consider a non-blind strategy, FedAvg - Dropout (Non_Blind) where the PS is aware of the identity of the clients, and_
knows exactly which clients have been successful in sending
their local updates to the PS. This is common in point-to-point
learning settings. In this case the PS simply ignores the clients
that have been unable to send their updates, and averages the
successful updates by dividing the global aggregate at the PS
by the number of successful transmissions.
In Fig. 3 (and also in Fig. 4), we consider every client has
a different probability of successful transmission to the PS
according to p = [0.1, 0.2, 0.3, 0.1, 0.1, 0.5, 0.8, 0.1, 0.2, 0.9].
We have deliberately chosen some clients to have a very low
connectivity, some others moderate, and others very high. We
consider a ring topology where client i is connected to clients
(i 1) mod n and (i + 1) mod n. For this setting, we
_−_
distinguish the cases with and without optimized weights. The
weights are optimized in order to minimize the term S(p, A),
which consequently minimizes the variance of the iterates,
subject to ensuring that the updates are unbiased according to
Alg. 3. Note that explicitly optimizing the consensus weights
that the clients use for their neighbors was not essential in Fig.
2 because the initial weights of Alg. 3 are optimal for a FCT
with homogeneous connectivity to the PS, i.e., pi = p∀i ∈ [n].
Finally, in Fig. 4, we consider the setting in which the
training data is distributed across the clients in a non-IID
fashion. To emulate non-IID-ness, we consider the sort-and_partition approach in which the training data is initially sorted_
based on labels, and then divided into blocks and distributed
among clients in a skewed fashion so that each client has data
from only a few classes. For the ring topology in this plot, we
have considered each client to be connected to 4 of its nearest
neighbors. We also use global momentum at the PS to update
the global model. Remarkably, FedAvg (even with non-blind
averaging) fails to converge in this setting. This is because
in the absence of collaboration, clients that have important
training samples that are critical for training a good model
with high accuracy, may have a low probability of successful
transmission and thus are rarely able to convey their updates
to the PS. Therefore, when these clients are unable to convey
their updates to the PS, the resulting test accuracy of the global
model is 10%, as good as a random classifier for 10 classes.
_∼_
Collaborative relaying ensures that the information from these
critical datapoints are also conveyed to the PS even when the
data owner does not have connectivity to the PS.
VI. CONCLUSIONS
Our goal in this paper is to mitigate the detrimental effect
of clients’ intermittent connectivity on the training accuracy
of FL systems. For this purpose, we proposed a collaborative
relaying strategy, which exploits the connections between
clients to relay potentially missing model updates to the PS
due to blocked clients. Our algorithm allows the PS to receive
an unbiased estimate of the model update, which would not be
possible without relaying. We optimized the consensus weights
at each client to improve the rate of convergence. Our proposed
approach can be implemented even when the PS is blind to
the identities of clients which successfully communicate with
it at each round. Numerical results showed the improvement
in training accuracy and convergence time that our approach
provides under various settings, including IID and non-IID
data distributions, different communication graph topologies,
as well as blind and non-blind PSs.
REFERENCES
[1] B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas,
“Communication-Efficient Learning of Deep Networks from Decentralized Data,” in Proceedings of the 20th International Conference on
_Artificial Intelligence and Statistics, vol. 54, Apr 2017, pp. 1273–1282._
[2] M. Chen, D. G¨und¨uz, K. Huang, W. Saad, M. Bennis, A. V. Feljan, and
H. V. Poor, “Distributed learning in wireless networks: Recent progress
and future challenges,” arxiv:2104.02151, 2021.
-----
[3] M. R. Akdeniz, Y. Liu, M. K. Samimi, S. Sun, S. Rangan, T. S.
Rappaport, and E. Erkip, “Millimeter wave channel modeling and
cellular capacity evaluation,” IEEE J. Sel. Areas Commun., vol. 32, no. 6,
pp. 1164–1179, June 2014.
[4] M. Gapeyenko, A. Samuylov, M. Gerasimenko, D. Moltchanov,
S. Singh, M. R. Akdeniz, E. Aryafar, N. Himayat, S. Andreev, and
Y. Koucheryavy, “On the temporal effects of mobile blockers in urban
millimeter-wave cellular scenarios,” IEEE Trans. Veh. Technol., vol. 66,
no. 11, pp. 10 124–10 138, Nov 2017.
[5] Y. Yan and Y. Mostofi, “Co-optimization of communication and motion
planning of a robotic operation under resource constraints and in fading
environments,” IEEE Trans. Wireless Commun., vol. 12, no. 4, pp. 1562–
1572, April 2013.
[6] M. M. Zavlanos, M. B. Egerstedt, and G. J. Pappas, “Graph-theoretic
connectivity control of mobile robot networks,” Proc. IEEE, vol. 99,
no. 9, pp. 1525–1540, Sep. 2011.
[7] N. Michael, M. M. Zavlanos, V. Kumar, and G. J. Pappas, “Maintaining
connectivity in mobile robot networks,” in Experimental Robotics, 2009.
[8] S. Gil, S. Kumar, D. Katabi, and D. Rus, “Adaptive communication in
multi-robot systems using directionality of signal strength,” Int. J. Rob.
_Res., vol. 34, no. 7, pp. 946–968, 2015._
[9] D. G¨und¨uz, D. B. Kurka, M. Jankowski, M. M. Amiri, E. Ozfatura,
and S. Sreekumar, “Communicate to learn at the edge,” IEEE Comm.
_Magazine, vol. 58, no. 12, pp. 14–19, 2020._
[10] M. E. Ozfatura, J. Zhao, and D. G¨und¨uz, “Fast federated edge learning
with overlapped communication and computation and channel-aware fair
client scheduling,” in IEEE International Workshop on Signal Processing
_Advances in Wireless Communications (SPAWC), 2021, pp. 311–315._
[11] D. Liu, G. Zhu, J. Zhang, and K. Huang, “Data-importance aware user
scheduling for communication-efficient edge machine learning,” IEEE
_Trans. Cogn. Commun. Netw., vol. 7, no. 1, pp. 265–278, 2021._
[12] W. Xia, T. Q. S. Quek, K. Guo, W. Wen, H. H. Yang, and H. Zhu,
“Multi-armed bandit based client scheduling for federated learning,”
_IEEE Trans. Wireless Commun., pp. 1–1, 2020._
[13] H. H. Yang, A. Arafa, T. Q. S. Quek, and H. Vincent Poor, “Agebased scheduling policy for federated learning in mobile edge networks,”
in Proc. - ICASSP IEEE Int. Conf. Acoust. Speech Signal Process.
_(ICASSP), 2020, pp. 8743–8747._
[14] M. M. Amiri and D. G¨und¨uz, “Federated learning over wireless fading
channels,” IEEE Trans. Wireless Comms., vol. 19, no. 5, pp. 3546–3557,
2020.
[15] E. Ozfatura, S. Rini, and D. G¨und¨uz, “Decentralized sgd with over-theair computation,” in GLOBECOM 2020 - 2020 IEEE Global Communi_cations Conference, 2020, pp. 1–6._
[16] G. Zhu, Y. Wang, and K. Huang, “Broadband analog aggregation for low-latency federated edge learning (extended version),”
arxiv:1812.11494, 2019.
[17] B. Hasircioglu and D. Gunduz, “Private wireless federated learning with
anonymous over-the-air computation,” arxiv:2011.08579, 2021.
[18] M. Seif, W.-T. Chang, and R. Tandon, “Privacy amplification for
federated learning via user sampling and wireless aggregation,”
arxiv:2103.01953, 2021.
[19] M. Yemini, S. Gil, and A. J. Goldsmith, “Exploiting local and cloud
sensor fusion in intermittently connected sensor networks,” in 2020 IEEE
_Global Communications Conference (Globecom), December 2020._
[20] ——, “Cloud-cluster architecture for detection in intermittently connected sensor networks,” arXiv:2110.01119, 2021.
[21] M. S. H. Abad, E. Ozfatura, D. G¨und¨uz, and O. Ercetin, “Hierarchical
federated learning across heterogeneous cellular networks,” in Proc. _ICASSP IEEE Int. Conf. Acoust. Speech Signal Process. (ICASSP), 2020,_
pp. 8866–8870.
[22] X. Lian, C. Zhang, H. Zhang, C.-J. Hsieh, W. Zhang, and J. Liu, “Can
decentralized algorithms outperform centralized algorithms? a case study
for decentralized parallel stochastic gradient descent,” in NIPS, Dec.
2017.
[23] H. Tang, X. Lian, M. Yan, C. Zhang, and J. Liu, “d[2]: Decentralized
training over decentralized data,” in 35th Int. Conf. Mach. Learn.,
vol. 80. PMLR, Jul 2018, pp. 4848–4856.
[24] Z. Jiang, A. Balu, C. Hegde, and S. Sarkar, “Collaborative deep learning
in fixed topology networks,” in NIPS, Dec. 2017.
[25] K. Yuan, Q. Ling, and W. Yin, “On the convergence of decentralized
gradient descent,” SIAM Journal on Optimization, 2016.
[26] M. Kamp, L. Adilova, J. Sicking, F. H¨uger, P. Schlicht, T. Wirtz, and
S. Wrobel, “Efficient decentralized deep learning by dynamic model
averaging,” in Conf. Mach. Learn. Knowl. Discovery in Databases, 2019,
pp. 393–409.
[27] J. Zeng and W. Yin, “On nonconvex decentralized gradient descent,”
_IEEE Trans. Signal Process., vol. 66, no. 11, pp. 2834–2848, June 2018._
[28] L. Kong, T. Lin, A. Koloskova, M. Jaggi, and S. U. Stich, “Consensus
control for decentralized deep learning,” arXiv:2102.04828, 2021.
[29] T. Vogels, L. He, A. Koloskova, S. P. Karimireddy, T. Lin, S. U.
Stich, and M. Jaggi, “Relaysum for decentralized deep learning on
heterogeneous data,” in Thirty-Fifth Conference on Neural Information
_Processing Systems, 2021._
[30] R. Saha, S. Rini, M. Rao, and A. J. Goldsmith, “Decentralized optimization over noisy, rate-constrained networks: Achieving consensus
by communicating differences,” IEEE Journal on Selected Areas in
_Communications, vol. 40, no. 2, pp. 449–467, 2022._
[31] A. Koloskova, N. Loizou, S. Boreiri, M. Jaggi, and S. Stich, “A unified
theory of decentralized SGD with changing topology and local updates,”
in 37th Int. Conf. Mach. Learn., vol. 119. PMLR, Jul 2020, pp. 5381–
5393.
[32] J. Wang, A. K. Sahu, Z. Yang, G. Joshi, and S. Kar, “MATCHA:
speeding up decentralized SGD via matching decomposition sampling,”
arxiv:1905.09435, 2019.
[33] M. Assran, N. Loizou, N. Ballas, and M. Rabbat, “Stochastic gradient
push for distributed deep learning,” in 36th Int. Conf. Mach. Learn.
PMLR, Jun 2019, pp. 344–353.
[34] L. Liu, J. Zhang, S. Song, and K. B. Letaief, “Client-edge-cloud
hierarchical federated learning,” in IEEE Int. Conf. Commun. (ICC),
2020, pp. 1–6.
[35] W. Y. B. Lim, J. S. Ng, Z. Xiong, J. Jin, Y. Zhang, D. Niyato, C. Leung,
and C. Miao, “Decentralized edge intelligence: A dynamic resource
allocation framework for hierarchical federated learning,” IEEE Trans.
_Parallel Distrib. Syst., vol. 33, no. 3, pp. 536–550, 2022._
[36] T. Castiglia, A. Das, and S. Patterson, “Multi-level local {sgd}: Distributed {sgd} for heterogeneous hierarchical networks,” in International
_Conference on Learning Representations, 2021._
[37] F. P.-C. Lin, S. Hosseinalipour, S. S. Azam, C. G. Brinton, and
N. Michelusi, “Semi-decentralized federated learning with cooperative
d2d local model aggregations,” IEEE Journal on Selected Areas in
_Communications, vol. 39, no. 12, pp. 3851–3869, 2021._
[38] Anonymous, “Hybrid local SGD for federated learning with
heterogeneous communications,” in Submitted to The Tenth International
_Conference on Learning Representations, 2022, under review. [Online]._
[Available: https://openreview.net/forum?id=H0oaWl6THa](https://openreview.net/forum?id=H0oaWl6THa)
[39] H. Yang, M. Fang, and J. Liu, “Achieving linear speedup with partial
worker participation in non-IID federated learning,” in International
_Conference on Learning Representations, 2021._
[40] X. Gu, K. Huang, J. Zhang, and L. Huang, “Fast federated learning in
the presence of arbitrary device unavailability,” arxiv:2106.04159, 2021.
[41] T. H. Hsu, H. Qi, and M. Brown, “Measuring the effects of nonidentical data distribution for federated visual classification,” CoRR, vol.
abs/1909.06335, 2019.
[42] T.-M. H. Hsu, H. Qi, and M. Brown, “Federated visual classification
with real-world data distribution,” CoRR, vol. abs/2003.08082, 2020.
[43] K. Hsieh, A. Phanishayee, O. Mutlu, and P. Gibbons, “The non-IID data
quagmire of decentralized machine learning,” in 37th Int. Conf. Mach.
_Learn., vol. 119._ PMLR, Jul 2020, pp. 4387–4398.
[44] Y. Zhao, M. Li, L. Lai, N. Suda, D. Civin, and V. Chandra, “Federated
learning with non-iid data,” CoRR, vol. abs/1806.00582, 2018.
[45] E. Ozfatura, D. Gunduz, and H. V. Poor, “Collaborative learning over
wireless networks: An introductory overview,” arxiv:2112.05559, 2021.
[46] H. H. Yang, Z. Liu, T. Q. S. Quek, and H. V. Poor, “Scheduling policies
for federated learning in wireless networks,” IEEE Trans. Commun.,
vol. 68, no. 1, pp. 317–333, 2020.
[47] W. Shi, S. Zhou, and Z. Niu, “Device scheduling with fast convergence
for wireless federated learning,” in IEEE Int. Conf. Commun. (ICC),
2020, pp. 1–6.
[48] M. M. Amiri, D. G¨und¨uz, S. R. Kulkarni, and H. V. Poor, “Convergence
of update aware device scheduling for federated learning at the wireless
edge,” IEEE Trans. Wireless Comm., vol. 20, no. 6, pp. 3643–3658,
2021.
[49] M. Yemini, R. Saha, E. Ozfatura, D. G¨und¨uz, and A. J. Goldsmith, “Robust federated learning with connectivity failures: A semi-decentralized
framework with collaborative relaying,” arXiv:2202.11850, 2022.
[50] D. Bertsekas, Nonlinear Programming. Athena Scientific, 1999.
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/2205.10998, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "GREEN",
"url": "https://arxiv.org/pdf/2205.10998"
}
| 2,022
|
[
"JournalArticle"
] | true
| 2022-05-23T00:00:00
|
[
{
"paperId": "f221c036aa17270b57ea254b1e7ceb3a21b36479",
"title": "Decentralized Edge Intelligence: A Dynamic Resource Allocation Framework for Hierarchical Federated Learning"
},
{
"paperId": "7609583b23d37f38b675271b099dbe91340df286",
"title": "Robust Federated Learning with Connectivity Failures: A Semi-Decentralized Framework with Collaborative Relaying"
},
{
"paperId": "4e55c40ddf5584aa8bdfb44732d335d63da70d87",
"title": "Collaborative Learning over Wireless Networks: An Introductory Overview"
},
{
"paperId": "aa6c2814ff94ca098d90f188f95126b5b06ebb69",
"title": "Nonlinear Programming"
},
{
"paperId": "b9b9f79b9644404c38c280b611812db39b2df9a4",
"title": "RelaySum for Decentralized Deep Learning on Heterogeneous Data"
},
{
"paperId": "7fb49222bc676580c03e2e9175449818c7006ef7",
"title": "Cloud-Cluster Architecture for Detection in Intermittently Connected Sensor Networks"
},
{
"paperId": "afe686ee1a7521bc6ef16f48f38b9420f65c9aa1",
"title": "Fast Federated Edge Learning with Overlapped Communication and Computation and Channel-Aware Fair Client Scheduling"
},
{
"paperId": "263762edd4e8d5866e4bfb71015d1e8d03941f0d",
"title": "Fast Federated Learning in the Presence of Arbitrary Device Unavailability"
},
{
"paperId": "654eaff6b4d8890b7b54dabcfa63822c90da42d5",
"title": "Distributed Learning in Wireless Networks: Recent Progress and Future Challenges"
},
{
"paperId": "16a7dcb0332e02c502174a6210e8cfcde9f8762a",
"title": "Semi-Decentralized Federated Learning With Cooperative D2D Local Model Aggregations"
},
{
"paperId": "19cc2c685f9683743750615df93d4cb1cf3eee15",
"title": "Privacy Amplification for Federated Learning via User Sampling and Wireless Aggregation"
},
{
"paperId": "842a4e524f97a644e9995b9966da33b119f8bf52",
"title": "Consensus Control for Decentralized Deep Learning"
},
{
"paperId": "433000baf18bb4403681fde5740bccd1fa2034a9",
"title": "Achieving Linear Speedup with Partial Worker Participation in Non-IID Federated Learning"
},
{
"paperId": "7e4a83495c15fc3c4ff98f74ed3b6b03db393aac",
"title": "Private Wireless Federated Learning with Anonymous Over-the-Air Computation"
},
{
"paperId": "eafd4efd5a5f90a22d714f13d63ec36120059533",
"title": "Decentralized Optimization Over Noisy, Rate-Constrained Networks: Achieving Consensus by Communicating Differences"
},
{
"paperId": "6b8f1c25e47efd019210086d20318488e00c8897",
"title": "Communicate to Learn at the Edge"
},
{
"paperId": "b97d07c6dc6132dbcbddf203d4a3f2d2a803f414",
"title": "Multi-Level Local SGD for Heterogeneous Hierarchical Networks"
},
{
"paperId": "7843a2e12c0aa7f642806652358fa2a8fb058daf",
"title": "Multi-Armed Bandit-Based Client Scheduling for Federated Learning"
},
{
"paperId": "552ec63759f9eccd7780edb0e755b6bcc3e54a3e",
"title": "Exploiting Local and Cloud Sensor Fusion in Intermittently Connected Sensor Networks"
},
{
"paperId": "c933fed82e7b5cbf7230f0f970b69590b40f86a1",
"title": "A Unified Theory of Decentralized SGD with Changing Topology and Local Updates"
},
{
"paperId": "e740a2b706fcae34850fd0e56619a2df7ee4dce7",
"title": "Federated Visual Classification with Real-World Data Distribution"
},
{
"paperId": "45776d8fc6e9fccf87d923b9dcfede962910ee8c",
"title": "Decentralized SGD with Over-the-Air Computation"
},
{
"paperId": "b93f5a3f8194baa3f3333775a9553694c58e0256",
"title": "Convergence of Update Aware Device Scheduling for Federated Learning at the Wireless Edge"
},
{
"paperId": "fba1082dac7c90bfe97e7c3e85a55499ba982c14",
"title": "Device Scheduling with Fast Convergence for Wireless Federated Learning"
},
{
"paperId": "2f3b15e68b793d7e5d834ce3356d2cf033ea7660",
"title": "Age-Based Scheduling Policy for Federated Learning in Mobile Edge Networks"
},
{
"paperId": "cb4f814bc755ee7d4083a529cc1c84c2965547b4",
"title": "Data-Importance Aware User Scheduling for Communication-Efficient Edge Machine Learning"
},
{
"paperId": "206261db1196e4e391ca42077f6fca6b3ece34d0",
"title": "The Non-IID Data Quagmire of Decentralized Machine Learning"
},
{
"paperId": "46d8c9e2dc9c12615eb5f6813d18f967d61c7e0d",
"title": "Measuring the Effects of Non-Identical Data Distribution for Federated Visual Classification"
},
{
"paperId": "bcb2d1c9cdc321d192925cc97c563470b30b8251",
"title": "Hierarchical Federated Learning ACROSS Heterogeneous Cellular Networks"
},
{
"paperId": "efc088a6df785ffc96fb7ff3a65c9b549ac54832",
"title": "Scheduling Policies for Federated Learning in Wireless Networks"
},
{
"paperId": "da2bd3d3d82ba6828dd90cb6777432099d1a1e02",
"title": "Federated Learning Over Wireless Fading Channels"
},
{
"paperId": "72055aff17a462bcccb3250becaf62c9911cdd8b",
"title": "MATCHA: Speeding Up Decentralized SGD via Matching Decomposition Sampling"
},
{
"paperId": "afb1acd9cb0caa50b9b9170e3cd63fa4a6f65478",
"title": "Client-Edge-Cloud Hierarchical Federated Learning"
},
{
"paperId": "43572a7cc087e388f7f312a0f2e17915682ff27c",
"title": "Broadband Analog Aggregation for Low-Latency Federated Edge Learning"
},
{
"paperId": "b42c38975f3d3f8bfe4b0c1a6e576c3e297cec38",
"title": "Stochastic Gradient Push for Distributed Deep Learning"
},
{
"paperId": "2231677181ab63a5ab6f98aa417eed91d831f896",
"title": "Efficient Decentralized Deep Learning by Dynamic Model Averaging"
},
{
"paperId": "5cfc112c932e38df95a0ba35009688735d1a386b",
"title": "Federated Learning with Non-IID Data"
},
{
"paperId": "42f7bd35df5a280ccd47115a30901afab9f0776b",
"title": "D2: Decentralized Training over Decentralized Data"
},
{
"paperId": "4d886c571c0849fda73a2d24e944d59fd37bcf9c",
"title": "Collaborative Deep Learning in Fixed Topology Networks"
},
{
"paperId": "3f1ab8b484f7881a68c8562ff908390742e4ba90",
"title": "Can Decentralized Algorithms Outperform Centralized Algorithms? A Case Study for Decentralized Parallel Stochastic Gradient Descent"
},
{
"paperId": "f7415d74aa51f0b665c2fc5da9ba76305dd90e19",
"title": "On the Temporal Effects of Mobile Blockers in Urban Millimeter-Wave Cellular Scenarios"
},
{
"paperId": "de43968e11cb1c5f3a09e5d9b3cc3a271c1b0379",
"title": "On Nonconvex Decentralized Gradient Descent"
},
{
"paperId": "d1dbf643447405984eeef098b1b320dee0b3b8a7",
"title": "Communication-Efficient Learning of Deep Networks from Decentralized Data"
},
{
"paperId": "fce4412e5be888229b939fc80766092549395c0f",
"title": "Adaptive communication in multi-robot systems using directionality of signal strength"
},
{
"paperId": "69a6cab0dd64f91548ce9b7feb24cdf92a0d0fe2",
"title": "Millimeter Wave Channel Modeling and Cellular Capacity Evaluation"
},
{
"paperId": "9267430a1a7be8040f1afb09a2f2929a2e7d5489",
"title": "On the Convergence of Decentralized Gradient Descent"
},
{
"paperId": "7586392a28c345a923a209b2a6fcfcc251f1c27a",
"title": "Co-Optimization of Communication and Motion Planning of a Robotic Operation under Resource Constraints and in Fading Environments"
},
{
"paperId": "ffb03ba337560ccf1204e57e1d1bb831c55e21bf",
"title": "Graph-theoretic connectivity control of mobile robot networks"
},
{
"paperId": "0aa7cb075978ed8c24f3e2a8ddc3ccb14df3e9a5",
"title": "Hybrid Local SGD for Federated Learning with Heterogeneous Communications"
},
{
"paperId": "524d14d172db722a02e772f83e5f81023bf032bb",
"title": "Maintaining Connectivity in Mobile Robot Networks"
}
] | 11,187
|
en
|
[
{
"category": "Computer Science",
"source": "external"
},
{
"category": "Computer Science",
"source": "s2-fos-model"
}
] |
https://www.semanticscholar.org/paper/02b7726c2e069342adbf3dd51abbeeb68dd32bda
|
[
"Computer Science"
] | 0.868341
|
Anomaly Detection Through Unsupervised Federated Learning
|
02b7726c2e069342adbf3dd51abbeeb68dd32bda
|
International Conference on Mobile Ad-hoc and Sensor Networks
|
[
{
"authorId": "1726077791",
"name": "Mirko Nardi"
},
{
"authorId": "2619829",
"name": "L. Valerio"
},
{
"authorId": "2174176466",
"name": "A. Passarella"
}
] |
{
"alternate_issns": null,
"alternate_names": [
"MSN",
"Mobile Ad-hoc and Sensor Networks",
"Int Conf Mob Ad-hoc Sens Netw",
"Mob Ad-hoc Sens Netw"
],
"alternate_urls": null,
"id": "72a6d50c-86ae-47c7-9a0e-54e5746aacee",
"issn": null,
"name": "International Conference on Mobile Ad-hoc and Sensor Networks",
"type": "conference",
"url": null
}
|
Federated learning (FL) is proving to be one of the most promising paradigms for leveraging distributed resources, enabling a set of clients to collaboratively train a machine learning model while keeping the data decentralized. The explosive growth of interest in the topic has led to rapid advancements in several core aspects like communication efficiency, handling non-IID data, privacy, and security capabilities. However, the majority of FL works only deal with supervised tasks, assuming that clients' training sets are labeled. To leverage the enormous unlabeled data on distributed edge devices, in this paper, we aim to extend the FL paradigm to unsupervised tasks by addressing the problem of anomaly detection (AD) in decentralized settings. In particular, we propose a novel method in which, through a preprocessing phase, clients are grouped into communities, each having similar majority (i.e., inlier) patterns. Subsequently, each community of clients trains the same anomaly detection model (i.e., autoencoders) in a federated fashion. The resulting model is then shared and used to detect anomalies within the clients of the same community that joined the corresponding federated process. Experiments show that our method is robust, and it can detect communities consistent with the ideal partitioning in which groups of clients having the same inlier patterns are known. Furthermore, the performance is significantly better than those in which clients train models exclusively on local data and comparable with federated models of ideal communities' partition.
|
# Anomaly Detection through Unsupervised Federated Learning
### Mirko Nardi
_Scuola Normale Superiore_
Pisa, Italy
mirko.nardi@sns.it
### Lorenzo Valerio
_IIT-CNR_
Pisa, Italy
lorenzo.valerio@iit.cnr.it
### Andrea Passarella
_IIT-CNR_
Pisa, Italy
andrea.passarella@iit.cnr.it
**_Abstract—Federated learning (FL) is proving to be one of the_**
**most promising paradigms for leveraging distributed resources,**
**enabling a set of clients to collaboratively train a machine learn-**
**ing model while keeping the data decentralized. The explosive**
**growth of interest in the topic has led to rapid advancements**
**in several core aspects like communication efficiency, handling**
**non-IID data, privacy, and security capabilities. However, the**
**majority of FL works only deal with supervised tasks, assuming**
**that clients’ training sets are labeled. To leverage the enormous**
**unlabeled data on distributed edge devices, in this paper, we aim**
**to extend the FL paradigm to unsupervised tasks by addressing**
**the problem of anomaly detection in decentralized settings. In**
**particular, we propose a novel method in which, through a**
**preprocessing phase, clients are grouped into communities, each**
**having similar majority (i.e., inlier) patterns. Subsequently, each**
**community of clients trains the same anomaly detection model**
**(i.e., autoencoders) in a federated fashion. The resulting model**
**is then shared and used to detect anomalies within the clients**
**of the same community that joined the corresponding federated**
**process. Experiments show that our method is robust, and it**
**can detect communities consistent with the ideal partitioning**
**in which groups of clients having the same inlier patterns are**
**known. Furthermore, the performance is significantly better than**
**those in which clients train models exclusively on local data**
**and comparable with federated models of ideal communities’**
**partition.**
**_Index Terms—federated learning, unsupervised, anomaly de-_**
**tection**
I. INTRODUCTION
Distributed/decentralized ML executed at the edge represents
one of the most promising approaches capable of addressing
the issues that afflict centralized solutions.
In this regard, the Federated Learning (FL) [1] paradigm
has proved to be an effective and promising approach to face
the hard challenges triggered by these distributed settings. It
essentially aims to collaboratively train an ML model while
keeping the data decentralized through the exchange of models’
parameters updates (instead of raw data) that, in its vanilla
version, are iteratively aggregated and shared by a central
coordinating node.
Given its effectiveness, in the last years plenty of subsequent
research works have been released focusing on different core
This work has been partly funded under the H2020 MARVEL (grant 957337),
HumaneAI-Net (grant 952026), SoBigData++ (grant 871042) and CHIST-ERA
SAI (grant CHIST-ERA-19-XAI-010, by MUR, FWF, EPSRC, NCN, ETAg,
BNSF).
aspects: improving communication efficiency, increasing model
performance in combination with non-IID data, extending
privacy and security capabilities and addressing client hardware
variability.
Nevertheless, FL applications and implementations for
mobile edge devices are still largely designed for supervised
learning tasks as a spontaneous consequence of its original
development purpose [2]. Thus, one of the least treated aspects
is the extension of FL to other ML paradigms like unsupervised
learning, reinforcement learning, active learning, and online
learning [3].
This paper specifically aims to apply FL on unsupervised
tasks for mobile edge devices. Unsupervised learning (as well as
semi-supervised and self-supervised learning) has recently been
considered one of the next great frontiers for AI [4]. Unlabeled
data far surpasses labeled data in real-world applications. Hence
its integration with federated contexts is mandatory to fully
unleash the potential of this approach.
In this paper, we consider nodes that have to learn a common
ML model (e.g., a classifier). We assume that sets of these
nodes “see” similar data patterns. However, as we assume
that data are not labeled, nodes need to automatically group
themselves into those sets, to perform FL across members
of the same set. As a specific application case, we consider
anomaly detection. Specifically, our methodology consists of a
preprocessing phase in which each node of the system detects
a membership group (cluster or community) such that each
member shares similar majority (i.e., inlier) patterns. In fact,
to ensure the effectiveness of an anomaly detection task, a
federated model must be trained on data coming from the same
distribution. Once the nodes are grouped in communities, a
federated learning process is spawned for each of them: nodes
of the same group use their local data to collaboratively train an
autoencoder to recognize their majority pattern (i.e., the inlier
class). Autoencoders are particularly suitable for this purpose
since typical FL protocols involve using a neural network-based
model. However, the methodology is orthogonal to the specific
model trained via FL. Once the federated process is finished,
each client gets a much more accurate global model than it
would have obtained using only its local data, as long as it has
joined the proper community.
The proposed methodology is particularly suited for mobile
environments for several reasons. First, it allows nodes not
-----
to exchange local data, thus addressing privacy and network
resource limitations. Second, it supports heterogeneous settings
when the federation is not under the control of a single
entity (like in a datacenter), but where nodes join “freely”
the federation. Third, it is tailored to using tiny ML models
on individual nodes, which is mandatory for realistically
implementing decentralized model training on mobile devices.
This work can subsequently be framed in a more general
context of anomaly detection in which normal data belong to
multiple classes (in contrast to the typical AD task involving
only a single inlier class). For instance, the methodology
proposed, whose output is a set of models each specialized in
identifying a single normal pattern, can be further extended with
ensemble-based methods to efficiently tackle the multi-class
anomaly detection problem, as shown in [5].
The remainder of the paper is organised as follows. In
Section II an overview of the problem and the related works are
discussed. In Section IV we list the preliminaries and describe
our method in detail. In Section V we discuss the results of
the experiments, and in Section VI we draw the conclusions.
II. RELATED WORKS
Federated Learning is a distributed learning framework
particularly amenable to optimize the computing power and the
data management on edge devices. It is now widely considered
modern and more effective evolution of the more traditional
distributed paradigms [6]–[10], in which models are trained on
large but ‘flat’ datasets within a fully controlled environment
in terms of resource availability and data management.
FL enables to relax many of the traditional constraints and,
since its introduction [1], several lines of research contribute to
fast advances [3]; additionally, from the application perspective,
many specific use-case solutions have already been deployed
by major service providers [2], [11], [12].
Due to space reasons, in the rest of the section, we provide
an overview of unsupervised approaches to FL, which are the
closest area with respect to the focus of this paper.
_A. Unsupervised Federated Learning_
Very few works combining federated learning and unsupervised approaches have been released, each of them dealing
with limited scenarios and settings. Reference [13] is the
first to introduce unsupervised representation learning in a
federated setting, but it simply combines the two concepts
without assuming the typical issues of distributed settings,
particularly for mobile environments (e.g., dealing with nonIID data, scaling the number of devices, different application
domains).
Reference [14] make progress on the same problem by
adding and facing two relevant challenges: (i) inconsistency
of representation spaces, due to non-IID data assumption, i.e.,
clients generate local models focused on different categories;
(ii) misalignment of representations, given by the absence of
unified information among clients.
Reference [15] introduced an unsupervised federated learning
(FL) approach for speech enhancement and separation with
non-IID data across multiple clients. An interesting aspect of
this work is that a small portion of supervised data is exploited
to boost the main unsupervised task through a combination of
updates from clients, with supervised and unsupervised data.
In [16] authors present a first effort for introducing a
collaborative system of autoencoders for distributed anomaly
detection. However, the data collected by the edge devices
are used to train the models in the cloud, which violates an
essential FL feature. Locally, the models are used for inference
only.
A more recent work [17] in a similar direction proposes a
federated learning (FL)-based anomaly detection approach for
identification and classification intrusion in IoT networks using
decentralized on-device data. Here the authors use federated
training rounds on Gated Recurrent Units (GRUs) models and
keep the data intact on local IoT devices by sharing only the
learned weights with the central server of the FL. However,
dealing with a classification task still assumes the availability
of labeled data.
III. PROBLEM FORMULATION AND PRELIMINARIES
We consider a distributed learning system with a set of clients
_M and a set of data distributions C, such that_ _C_ _M_ . With
_|_ _| ≤|_ _|_
data distribution, we refer to a set of identically distributed data
representing a specific pattern (e.g., observations of phenomena
belonging to the same class of events, in case of a classification
task). We assume that every client receives a portion d
_∈_
(0%, 50%) of its samples from a single distribution Cout ∈
_C, and the remaining (100 −_ _d)% from Cin ∈_ _C, such that_
_Cin ̸= Cout. Thereby, the two samples partitions within each_
client form the outlier and inlier classes, respectively. This
split represents a basic assumption when dealing with AD
tasks [18]. d [5%, 15%] is generally a realistic value [19],
_∈_
thus adopted in the majority of related works. Note that this
scenario corresponds to assuming local skewed data, i.e., that
each node “sees” a prevalence of data of a single class (its
inlier class) and a minority of data from (one of the) other
classes. This is also quite realistic in practice in AD tasks.
The challenge addressed in the paper is the following. In case
of supervised learning, data belonging to each class are labelled,
so each node knows which other nodes “see” the same majority
class, and therefore forming FL groups is straightforward. In
unsupervised cases, each node can detect its majority class from
local data, but has no direct information to know which other
nodes see the same majority class. Therefore, the main objective
of our methodology is to identify an effective algorithm for
nodes to form consistent groups (i.e., groups that see the same
majority class), to then run a standard FL process across nodes
of the same group.
Note that, as will be clear from the detailed description in
Section IV, at the end of the first step of our methodology
clients become partitioned into k disjoint groups S1, . . ., Sk.
In the ideal case, each group corresponds to the (unknown to
the clients) set of nodes seeing the same inlier class Cin, and
therefore in the ideal case k = _C_ .
_|_ _|_
-----
IV. PROPOSED METHODOLOGY
As anticipated in Section I our methodology consists in two
logical steps. In the first step we group clients that “see” the
same inlier class, via a fully autonomous and unsupervised
process. In the second step, we run a standard FL process
among clients belonging the same group. We present the two
steps in the following sections.
_A. Step I: group identification_
The aim of this phase is to make the clients join a group
(i.e. cluster) having the same (or similar) majority class Cin.
To achieve this, we firstly train a “classical” AD model (e.g.,
OCSVM) on every client, using only its local data, such that
each of them is able to compute a preliminary split of its data
into inliers and outliers. Thereafter, every couple of clients
perform the following steps: (i) they exchange their respective
models, and (ii) they use the partner’s model to split its local
data into “normal” and “anomalous” data through an inference
step. In other words, for every pair of nodes (mi, mj), node
_mi uses node’s mj local model to classify its own local data,_
and vice versa. If the classification accuracy is high enough,
it means that node’s mj model has been trained on the same
inlier class of node a, and therefore mi and mj should be in
the same group.
Note that, it is not necessary to use a very complex local
model at this step. Although the local model of a client only
enables an approximate preliminary inliers/outliers split, it
suffices to detect clients sharing the same majority class of
data, as long as those patterns of data in those classes are
sufficiently different (as it is the case in typical AD tasks).
Given a client mi, from its perspective this phase is detailed
in Algorithm 1. Specifically, on the local dataset of the i-th
client, i.e., xi, an inference step is computed using its own
locally trained model (line 4) and all the models of other clients
(line 9). yj,i is the output binary vector given by the AD model
of the j-th client on the data of the i-th client. Thus, inj,i is the
portion of inliers in the vector yj,i. The boolean bj,i indicates
whether the i-th client flags the j-th client as a candidate for the
association. The output the process corresponds to the group
of candidate clients Gi with inlier classes similar to mi.
At the end of algorithm 1, each client has a local view
of which other clients should belong to its group. However,
different clients in the same group may have different local
views (i.e., even if mj is in Gi, Gj may not be identical to
_Gi). In order to obtain an overall view of the groups, shared_
by all nodes, we adopt the following method.
Since the association of two clients is reciprocal (line 14),
a undirected graph can be built from all the resulting groups
of candidates of each client. A link between two nodes means
that those two nodes mutually “think” to be in the same group.
Finally, a community detection algorithm is run on this graph
to detect which groups of nodes should be considered part of
the same set and thus undergo a standard FL step. In other
words, we assume that communities found at the end of this
step are the groups of clients with the same inlier class.
**Algorithm 1 Client mi local training and association**
**Input: AD Model Modi, contamination d, association thresh-**
old q, set of other clients M
**Output: Group Gi of candidate clients similar to mi**
1: procedure LOCALAD(Modi, d, q, M )
2: _Gi ←∅_
3: _Modi = Modi.fit(xi, d)_
4: _yi,i = Modi.predict(xi)_
5: _ini,i = inlierPercCount(yi,i)_
6: _send(Modi, M_ )
7: **for all mj in M do**
8: _Modj = receive(mj)_
9: _yj,i = Modj.predict(xi)_
10: _inj,i = inlierPercCount(yj,i)_
11: _bj,i = ini,i −_ _q ≤_ _inj,i ≤_ _ini,i + q_
12: _send(bj,i, mj)_
13: _bi,j = receive(bi,j, mj)_
14: **if bj,i AND bi,j then**
15: _Gi ←_ _mi_
16: **end if**
17: **end for**
18: **return Gi**
19: end procedure
_B. Step II: federated outlier detection_
The result of the first phase is a set of k groups (or
communities) G0, . . ., Gk; for each of them a FL instance
is started using autoencoders as models. Autoencoders are
suitable for the purpose for two main reasons: (i) they naturally
fit into the FL framework, being NN-based; (ii) they can be
effectively used in AD task. In fact, they essentially learn
a compressed representation of the unlabeled data used for
the training, performing a nonlinear dimensionality reduction.
Once trained, the reconstruction error of a given sample can
be used to classify it using a threshold.
We use the vanilla version of the Federated Averaging
(FedAvg) [1], a FL protocol based on averaging the local
stochastic gradient descent updates to compute the global model.
At the end of each federation process, the trained autoencoder
is shared among the clients of the same group.
Note that, the community detection step requires either a
central entity that runs the algorithm once and for all nodes,
or that the graph is shared among all nodes and each runs
the same community detection algorithm individually. Even
in the former case, our methodology does not require that
nodes share local data with any central controller, and thus
can address situations where centralized learning is unfeasible
or impractical (e.g., due to data ownership reasons).
V. EXPERIMENTS
In this section, we describe the numerical simulations to
assess the performance of the proposed methodology. The
baseline is given by the local model scheme, in which every
client trains its model using only local data. We show a further
-----
comparison with an ideal partitioning scheme in which the
groups of clients having the same inlier patterns are known.
This corresponds to a supervised FL algorithm, where all data
are labeled by a central entity. Our code is based on wellaccessed and standard frameworks: Tensorflow, Scikit-Learn,
PyOD and Flower. For the sake of reproducibility, the code is
available at https://github.com/mirqr/FedAD
_A. Datasets and setup_
We test our methodology on the MNIST [20] and the fashionMNIST [21] datasets, using the original 60000-10000 train-test
splits. Since both have ten classes, we have _C_ = 10 data
_|_ _|_
distributions.
Locally, given a portion of outlier d, the train set of every
client has d percent of its samples from a single distribution
_Cout ∈_ _C, and the remaining (100_ _−_ _d) percent from Cin ∈_ _C,_
such that Cin ̸= Cout.
With a view to a collaborative anomaly detection task, we
ensure that all the datasets owned by the clients are numerically
balanced and disjoint. The set of clients M that compose
an experimental setup is configured as follows: we define a
parameter p as the number of clients within the same data
distribution (i.e., class), meaning that the train samples of a
class Cin of the original dataset (e.g., MNIST) are evenly and
randomly spread to form the inliers of p clients. Accordingly,
the portion of outliers for each client within the same group
and characterized by the same Cin, is given by the samples of
a class different from Cin. We ensure that the outlier classes
_C \ Cin are equally represented within the group, meaning_
that for each client of the group the minority class is “circular”
through the set C \ Cin.
As an example, using all the available data distributions
of the dataset (i.e., 10 classes), and by setting p = 9, then
the training data distribution among the clients of the group
_Cin = 0 is shown in Fig. 1. The same applies to every group,_
i.e., an experimental system configuration ends up with _M_ =
_|_ _|_
_C_ _p clients. Consequently, the ideal partitioning we aim to_
_|_ _|_
find through the community detection phase is composed by
_k =_ _C_ = 10 groups with p clients each.
_|_ _|_
Note that, without loss of generality, to obtain an balanced
distribution of the outliers classes among the clients of a group,
it is convenient to set p = (|C| − 1)n, n ∈ N. Additionally,
since for each configuration run, we exploit all the samples of
the dataset involved, as a higher value of p leads to smaller
local datasets for the clients.
_B. Models_
In the first phase, every client detects the partners having
the same inlier class. As explained in Section IV, a client
tests the others’ trained models on its local data and selects as
partners those whose model produces an inliers/outliers ratio
similar to its own. We select the “association” threshold q in
the interval [0.01, 0.10], i.e., q represents the maximum of the
percentage difference between the data classified as normal
by the local model, and those considered normal by using
the partner’s model. In other words, the local client considers
another client as partner if the model of the latter produces
a fraction of normal data on the local dataset equal to the
percentage produced by the local node, _q. In particular, we_
_±_
found that the value q = 0.08 turns out to work well on every
experiment.
We choose the model of the first phase with the following
requirements: (i) it must be easy to set up and fast to train;
(ii) it must be light to store and to be transmitted; (iii) it must
provide a preliminary sufficiently good outlier detection to
allow the clients to correctly group for the next phase.
There is not a model generally suitable for this purpose;
it strongly depends on the type of data used, especially for
AD tasks [22]. Moreover, the abovementioned requirements
force us to discard any NN-based AD model. Thus, we have
identified OC-SVM [23] to be a good choice for our cases.
It requires essentially two parameters to be set: the kernel
and the parameter ν (0, 1], which is an upper bound on the
_∈_
fraction of training errors and a lower bound on the fraction of
support vectors. The fine-tuning of ν in contaminated data can
be challenging without any assumptions on the distribution of
the outliers. However, since in our tests we assume to know
(only) the contamination value d = 10% for every dataset, we
can set ν = 0.1. Moreover, we use the RBF kernel.
For the second phase, we use a fully connected autoencoder,
a NN-based model that naturally fits into a federated learning
framework, with a three-layers topology (64-32-64), ReLU
activations on the hidden layers, and Sigmoid activation on
the output layer. Thirty-two neurons for the middle layer is
a reasonable value to avoid an information bottleneck. We
empirically observed that using more layers/neurons does not
significantly improve the effectiveness due to the tendency of
the neural network to overfit on this specific dataset.
_C. Group detection and anomaly detection performance_
For both the MNIST and the fashion-MNIST datasets, we
run four tests varying the value of p : 9, 18, 27, 36 . In all
_{_ _}_
the tests we use the contamination parameter d = 10% and we
take into account all the available classes, i.e., _C_ = 10. Let
_|_ _|_
_mCi,j be the j-th client with majority class Ci; we define ICi_
as the ideal set of clients having the same majority class Ci,
e.g., I0 = {m0,0, . . . m0,p−1}.
In Table I, we show the results of the community detection
phase for the MNIST dataset: we find nine communities, and
in most cases, they match with the ideal group of clients. The
major exception is given by G4, that in all the four cases
is given by the union of I4 and I9, meaning that the clients
having 4 and 9 as inlier class join the same community. This is
a consequence of the OC-SVM model’s inability to distinguish
the two digits, and it represents a typical behaviour when
dealing with image classification using MNIST. A similar
result occurs for G5 when p = 36 (Table Id), in which the
union of I5 and I8 is detected as single community. In this
case, recalling that a higher value of p leads to smaller local
datasets for the clients, it is reasonable that for p = 36 the local
models do not have enough samples and are no longer able
to distinguish the two digits. We can observe the anticipation
-----
Fig. 1: Histograms of training data distribution for the group Cin = 0 (i.e., 0 is the common inlier class) with p = 9.
TABLE I: Community detection for MNIST
(a) p = 9
**Community ID** **Members**
_G0_ _I0_
_G1_ _I1_
_G2_ _I2_
_G3_ _I3_
_G4_ _I4 ∪_ _I9_
_G5_ _I5_
_G6_ _I6_
_G7_ _I7_
_G8_ _I8_
(c) p = 27
**Community ID** **Members**
_G0_ _I0_
_G1_ _I1_
_G2_ _I2_
_G3_ _I3_
_G4_ _I4 ∪_ _I9_
_G5_ _I5 ∪_ _m8,18_
_G6_ _I6_
_G7_ _I7_
_G8_ _I8 \ m8,18_
(b) p = 18
(d) p = 36
|Community ID|Members|
|---|---|
|G0 G1 G2 G3 G4 G5 G6|I0 I1 ∪I3 I2 ∪I4 ∪I6 I5 I6 I7 I8|
|Community ID|Members|
|---|---|
|G0 G1 G2 G3 G4 G5 G6 G7 G8|I0 I1 I2 I3 I4 ∪I9 I5 I6 I7 I8|
|Community ID|Members|
|---|---|
|G0 G1 G2 G3 G4 G5 G6 G7 G8|I0 I1 I2 I3 I4 ∪I9 I5 I6 I7 I8|
|Community ID|Members|
|---|---|
|G0 G1 G2 G3 G4 G5 G6|I0 ∪I2 ∪I4 ∪I6 I1 ∪I3 I5 \ m5,6 I6 I7 I8 m5,6|
|Community ID|Members|
|---|---|
|G0 G1 G2 G3 G4 G5 G6 G7 G8|I0 I1 I2 I3 I4 ∪I9 I5 ∪m8,18 I6 I7 I8 \ m8,18|
|Community ID|Members|
|---|---|
|G0 G1 G2 G3 G4 G5 G6 G7 G8|I0 I1 I2 I3 I4 ∪I9 I5 ∪I8 I6 I7 I8|
of this behaviour when p = 27 in Table Ic in which the client
_m8,18 mistakenly joins I5._
Similar considerations can be done for the fashion-MNIST
case (Table II). Here the ideal groups of clients I1 and I3 are
detected as a single community in the four testes. The same
applies to the groups I0, I2, I4, I6, excluding the case p = 9
(Table IIa), in which I0 is correctly isolated. This result is
expected as fashion-MNIST is notably harder than MNIST.
_D. Experimental result: federated outlier detection_
We compare our methodology with two baselines: (i) local,
where clients only train on local data; (ii) ideal, in which a
client mCi,j uses the model trained through federated learning
on the set of clients ICi, i.e., the set of the clients sharing
the same majority class. The test samples for each client are
randomly sampled from the MNIST/fashion-MNIST test set,
following the same inlier/outlier classes and the ratio of the
corresponding client.
In Tables III and IV we show the test AUC score on MNIST
and fashion-MNIST by varying the value of p, meaning that for
each row we compute the average AUC score of p _C_ clients.
_|_ _|_
Our methodology performs almost as the upper bound baseline,
represented by the ideal federations of clients. Nevertheless,
the results are consistent with the partitioning we obtain in
TABLE II: Community detection for fashion-MNIST
(a) p = 9
**Community ID** **Members**
_G0_ _I0_
_G1_ _I1 ∪_ _I3_
_G2_ _I2 ∪_ _I4 ∪_ _I6_
_G3_ _I5_
_G4_ _I6_
_G5_ _I7_
_G6_ _I8_
(b) p = 18
**Community ID** **Members**
_G0_ _I0 ∪_ _I2 ∪_ _I4 ∪_ _I6_
_G1_ _I1 ∪_ _I3_
_G2_ _I5 \ m5,6_
_G3_ _I6_
_G4_ _I7_
_G5_ _I8_
_G6_ _m5,6_
(c) p = 27
**Community ID** **Members**
_G0_ _I0 ∪_ _I2 ∪_ _I4 ∪_ _I6_
_G1_ _I1 ∪_ _I3_
_G2_ _I5_
_G3_ _I6_
_G4_ _I7_
_G5_ _I8_
(d) p = 36
**Community ID** **Members**
_G0_ _I0 ∪_ _I2 ∪_ _I4 ∪_ _I6_
_G1_ _I1 ∪_ _I3_
_G2_ _I5_
_G3_ _I6_
_G4_ _I7_
_G5_ _I8_
the first step with the community detection that, especially for
MNIST, identifies the right groups of clients in most of the
cases. In the fashion-MNIST case, there are more exceptions to
this behaviour. For instance, clients with different inlier classes
all join a common group, as shown in Tabel IV (e.g., G1).
This affects the average AUC scores, which appear slightly
less than the ideal upper bound (as opposed to nearly identical
MNIST scores), but are still satisfactory.
More detailed results are shown in Tables V and VI, in which
we only consider the detected communities that do not match
the ideal cases. In these tables, each row corresponds to the
average test AUC score for a fixed p and all the clients having
|Community ID|Members|
|---|---|
|G0 G1 G2 G3 G4 G5|I0 ∪I2 ∪I4 ∪I6 I1 ∪I3 I5 I6 I7 I8|
|Community ID|Members|
|---|---|
|G0 G1 G2 G3 G4 G5|I0 ∪I2 ∪I4 ∪I6 I1 ∪I3 I5 I6 I7 I8|
-----
TABLE III: Test AUC on MNIST. For each p, mean std are
_±_
computed on p _C_ clients
_|_ _|_
Local Community (ours) Ideal
_p_
9 0.773 ± 0.205 0.836 ± 0.18 0.839 ± 0.185
18 0.769 ± 0.207 0.835 ± 0.18 0.836 ± 0.181
27 0.77 ± 0.208 0.836 ± 0.18 0.84 ± 0.181
36 0.766 ± 0.207 0.819 ± 0.191 0.838 ± 0.182
TABLE IV: Test AUC on fashion-MNIST. For each p, mean
std are computed on p _C_ clients
_±_ _|_ _|_
Local Community (ours) Ideal
_p_
9 0.714 ± 0.166 0.761 ± 0.161 0.772 ± 0.155
18 0.71 ± 0.173 0.747 ± 0.166 0.769 ± 0.155
27 0.706 ± 0.165 0.75 ± 0.162 0.765 ± 0.154
36 0.707 ± 0.166 0.749 ± 0.161 0.765 ± 0.151
majority class CIN . The difference between the community
(ours) and the ideal case is that in the former the clients of
_CIN are trained through the corresponding federation G such_
that CIN ∈ _G (Tables I and II), while in the latter they are_
trained through the perfect federation CIN = IIN .
As regards the MNIST case, we always obtain a community
_G4 = I4∪I9 and, for p = 36, we have an additional community_
_G5 = I5_ _I8. We ignore the one client mismatch in the_
_∪_
_p = 27 (Table Ic) as we verified that its influence is negligible._
In Table V we observe that the clients with majority class
_CIN = 4 still perform well with our methodology, with an_
average increase of 6% in the AUC score from the local case
and an average decrease of 2% from the ideal case. CIN = 9
scores end up approximately in the middle of the two bounds,
highlighting, however, that the local case already reaches a
good score of 0.83 for any p. CIN = 5 is the only case that
performs noticeably worse than the ideal case, with a decrease
of 9% in the AUC score. However, also in this case there is
a noticeable improvement over using the local models only.
For the fashion-MNIST case (Table VI), the scores are
predictably lower than in the previous case: the gaps between
the two bounds are generally tighter, but in any test, the scores
of our methodology still fall in the middle. Clients of CIN = 1
almost reach the ideal result, although the difference with the
local one is minimal, while clients with CIN = 3 have on
average a 4% increase/decrease on both the lower/upper
_∼_
baseline. Clients of CIN = 2, CIN = 4 have an average AUC
score very close (+1%) to the lower baseline for p > 8; this is
precisely the value beyond which their federation is the union
of four sets, i.e., I0 ∪ _I2 ∪_ _I4 ∪_ _I6, thus totalling four different_
majority classes. On the other hand, the remaining clients of
this big federation, CIN = 0 and CIN = 6, are still able to
reach a 7% increase on the local case and be very close to
_∼_
the ideal case.
VI. CONCLUSIONS AND FUTURE WORK
In this paper we propose a new methodology for federated
learning in unsupervised settings, particularly amenable for
dynamic mobile environments without central coordination.
|mputed|± d on p|C| clients|
|---|---|
|p|Local Community (ours) Ideal|
|||
|9 18 27 36|0.773 ± 0.205 0.836 ± 0.18 0.839 ± 0.185 0.769 ± 0.207 0.835 ± 0.18 0.836 ± 0.181 0.77 ± 0.208 0.836 ± 0.18 0.84 ± 0.181 0.766 ± 0.207 0.819 ± 0.191 0.838 ± 0.182|
|td are|e computed on p|C| clients|
|---|---|
|p|Local Community (ours) Ideal|
|||
|9 18 27 36|0.714 ± 0.166 0.761 ± 0.161 0.772 ± 0.155 0.71 ± 0.173 0.747 ± 0.166 0.769 ± 0.155 0.706 ± 0.165 0.75 ± 0.162 0.765 ± 0.154 0.707 ± 0.166 0.749 ± 0.161 0.765 ± 0.151|
We specifically focus on Anomaly Detection tasks to define
the details and test the methodology. The methodology is
composed by two sequential steps: in the first step we detect
the communities of clients having similar majority patterns
(i.e., inlier class); this is achieved by having the clients perform
a preliminary inlier/outlier split of their local data through the
training of an AD model. Two clients join the same community when both agree in the inliers/outliers proportion after
exchanging their respective models and computing an inference
step on their local data. Then, each of the resulting community
collaboratively trains a NN-based anomaly detection model
through the federated learning framework.
We tested our methodology on the MNIST and fashionMNIST datasets; in most cases, the communities found match
with the ideal groups of clients, which are used as an upper
bound baseline in experimental part. When the ideal groups
TABLE V: Test AUC std on MNIST
_±_
Local Community (ours) Ideal
_p_ _CIN_
9 4 0.749 ± 0.245 0.833 ± 0.197 0.833 ± 0.232
9 0.823 ± 0.184 0.86 ± 0.159 0.881 ± 0.138
18 4 0.774 ± 0.2 0.819 ± 0.204 0.855 ± 0.19
9 0.828 ± 0.176 0.872 ± 0.149 0.881 ± 0.139
27 4 0.762 ± 0.214 0.823 ± 0.208 0.84 ± 0.205
9 0.836 ± 0.158 0.862 ± 0.161 0.882 ± 0.132
36
4 0.76 ± 0.215 0.799 ± 0.213 0.84 ± 0.201
9 0.838 ± 0.156 0.862 ± 0.157 0.881 ± 0.13
5 0.708 ± 0.194 0.718 ± 0.188 0.807 ± 0.177
8 0.677 ± 0.196 0.696 ± 0.195 0.719 ± 0.219
TABLE VI: Test AUC std on MNIST
_±_
Local Community (ours) Ideal
_p_ _CIN_
9
18
27
36
1 0.911 ± 0.051 0.94 ± 0.028 0.946 ± 0.025
3 0.741 ± 0.127 0.788 ± 0.139 0.83 ± 0.094
2 0.663 ± 0.146 0.686 ± 0.154 0.719 ± 0.125
4 0.714 ± 0.128 0.762 ± 0.13 0.782 ± 0.117
6 0.642 ± 0.142 0.675 ± 0.144 0.698 ± 0.137
1 0.913 ± 0.04 0.935 ± 0.036 0.944 ± 0.026
3 0.751 ± 0.107 0.792 ± 0.14 0.831 ± 0.082
0 0.683 ± 0.125 0.742 ± 0.124 0.775 ± 0.089
2 0.665 ± 0.153 0.667 ± 0.16 0.711 ± 0.126
4 0.713 ± 0.142 0.724 ± 0.134 0.775 ± 0.115
6 0.626 ± 0.142 0.68 ± 0.137 0.704 ± 0.133
1 0.907 ± 0.04 0.937 ± 0.033 0.944 ± 0.024
3 0.74 ± 0.099 0.77 ± 0.164 0.813 ± 0.088
0 0.688 ± 0.109 0.743 ± 0.101 0.773 ± 0.08
2 0.674 ± 0.136 0.692 ± 0.145 0.763 ± 0.109
4 0.71 ± 0.13 0.725 ± 0.117 0.777 ± 0.107
6 0.63 ± 0.125 0.705 ± 0.126 0.714 ± 0.133
1 0.907 ± 0.041 0.936 ± 0.035 0.943 ± 0.024
3 0.73 ± 0.118 0.762 ± 0.16 0.803 ± 0.093
0 0.68 ± 0.113 0.754 ± 0.095 0.772 ± 0.078
2 0.675 ± 0.127 0.694 ± 0.144 0.733 ± 0.123
4 0.714 ± 0.132 0.743 ± 0.119 0.783 ± 0.107
6 0.639 ± 0.131 0.698 ± 0.125 0.717 ± 0.127
-----
are not found, our methodology merges 2-4 ideal groups into
one community; it occurs in two MNIST classes, obtaining 9
groups, and in 6 fashion-MINIST classes, obtaining 6 groups
in the worst case. The aggregation usually occurs for clients
having similar majority classes (e.g., 4 and 9 in the case of
MNIST).
We finally test the resulting AD federated models trained
by the detected communities in term of AUC score, with
local test sets on each client. In both cases, the results show
clear advantage over the models locally trained (i.e., the
lower baseline), while the performance is comparable with
the federated models of ideal communities’ partition, even for
detected communities in which different majority classes are
merged. This indicates that, even though we may not always
be able to group clients as in the ideal (supervised) case, still
the accuracy of the resulting model is close to optimal, and
significantly better than using local models trained only on
local data.
Future directions can involve several aspects of the proposed
solution. Firstly, the optimization of the community detection
phase, i.e., the all-to-all exchange of the local models may
be suboptimal for high numbers of clients. Moreover, another
possible improvement is the selection of the specific algorithms
used to train local and federated models. For example, the “flat”
fully connected autoencoder we use for the federated training
may be too simple; as an example, when dealing with images,
convolutional autoencoders may be introduced.
Finally, we aim to frame this solution in a more general
context of anomaly detection in which normal data belong to
multiple classes, in contrast to the typical AD task that only
involves a single inlier class.
ACKNOWLEDGMENT
This work has been partly funded under the H2020 MARVEL
(grant 957337), HumaneAI-Net (grant 952026), SoBigData++
(grant 871042) and CHIST-ERA SAI (grant CHIST-ERA-19XAI-010, by MUR, FWF, EPSRC, NCN, ETAg, BNSF).
REFERENCES
[1] B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y.
Arcas, “Communication-Efficient Learning of Deep Networks from
Decentralized Data,” in _Proceedings_ _of_ _the_ _20th_ _International_
_Conference on Artificial Intelligence and Statistics, ser. Proceedings_
of Machine Learning Research, A. Singh and J. Zhu, Eds., vol. 54.
PMLR, 20–22 Apr 2017, pp. 1273–1282. [Online]. Available:
[https://proceedings.mlr.press/v54/mcmahan17a.html](https://proceedings.mlr.press/v54/mcmahan17a.html)
[2] T. Yang, G. Andrew, H. Eichner, H. Sun, W. Li, N. Kong, D. Ramage,
and F. Beaufays, “Applied federated learning: Improving google keyboard
query suggestions,” arXiv preprint arXiv:1812.02903, 2018.
[3] P. Kairouz, H. B. McMahan, B. Avent, A. Bellet, M. Bennis, A. N.
Bhagoji, K. Bonawitz, Z. Charles, G. Cormode, R. Cummings, R. G. L.
D’Oliveira, H. Eichner, S. E. Rouayheb, D. Evans, J. Gardner, Z. Garrett,
A. Gascon, B. Ghazi, P. B. Gibbons, M. Gruteser, Z. Harchaoui, C. He,´
L. He, Z. Huo, B. Hutchinson, J. Hsu, M. Jaggi, T. Javidi, G. Joshi,
M. Khodak, J. Konecny, A. Korolova, F. Koushanfar, S. Koyejo,´
T. Lepoint, Y. Liu, P. Mittal, M. Mohri, R. Nock, A. Ozg[¨] ur, R. Pagh,¨
H. Qi, D. Ramage, R. Raskar, M. Raykova, D. Song, W. Song, S. U.
Stich, Z. Sun, A. T. Suresh, F. Tramer, P. Vepakomma, J. Wang,`
L. Xiong, Z. Xu, Q. Yang, F. X. Yu, H. Yu, and S. Zhao, “Advances
and open problems in federated learning,” Foundations and Trends®
_in Machine Learning, vol. 14, no. 1–2, pp. 1–210, 2021. [Online]._
[Available: http://dx.doi.org/10.1561/2200000083](http://dx.doi.org/10.1561/2200000083)
[4] Y. LeCun, “The next ai revolution will not be supervised,” 2018.
[5] M. Nardi, L. Valerio, and A. Passarella, “Centralised vs decentralised
anomaly detection: when local and imbalanced data are beneficial,” in
_Proceedings of the Third International Workshop on Learning with_
_Imbalanced Domains: Theory and Applications, ser. Proceedings of_
Machine Learning Research, N. Moniz, P. Branco, L. Torgo, N. Japkowicz,
M. Wozniak, and S. Wang, Eds., vol. 154. PMLR, 17 Sep 2021, pp. 7–20.´
[[Online]. Available: https://proceedings.mlr.press/v154/nardi21a.html](https://proceedings.mlr.press/v154/nardi21a.html)
[6] K. Ota, M. S. Dao, V. Mezaris, and F. G. B. D. Natale, “Deep learning
for mobile multimedia: A survey,” vol. 13, no. 3, pp. 1–22. [Online].
[Available: https://dl.acm.org/doi/10.1145/3092831](https://dl.acm.org/doi/10.1145/3092831)
[7] K. Chahal, M. S. Grover, and K. Dey, “A hitchhiker’s guide on
distributed training of deep neural networks.” [Online]. Available:
[http://arxiv.org/abs/1810.11787](http://arxiv.org/abs/1810.11787)
[8] T. Ben-Nun and T. Hoefler, “Demystifying parallel and distributed
deep learning: An in-depth concurrency analysis.” [Online]. Available:
[http://arxiv.org/abs/1802.09941](http://arxiv.org/abs/1802.09941)
[9] J. Verbraeken, M. Wolting, J. Katzy, J. Kloppenburg, T. Verbelen, and J. S.
Rellermeyer, “A survey on distributed machine learning,” vol. 53, no. 2,
[pp. 30:1–30:33. [Online]. Available: https://doi.org/10.1145/3377454](https://doi.org/10.1145/3377454)
[10] T. Tuor, S. Wang, K. K. Leung, and K. Chan, “Distributed machine
learning in coalition environments: Overview of techniques,” in 2018
_21st International Conference on Information Fusion (FUSION), pp._
814–821.
[11] K. Bonawitz, H. Eichner, W. Grieskamp, D. Huba, A. Ingerman,
V. Ivanov, C. Kiddon, J. Konecnˇ y, S. Mazzocchi, H. B. McMahan` _et al.,_
“Towards federated learning at scale: System design,” arXiv preprint
_arXiv:1902.01046, 2019._
[12] W. A. Group, “Federated learning white paper v1.” [Online].
Available: [https://aisp-1251170195.cos.ap-hongkong.myqcloud.com/](https://aisp-1251170195.cos.ap-hongkong.myqcloud.com/fedweb/1552917186945.pdf)
[fedweb/1552917186945.pdf](https://aisp-1251170195.cos.ap-hongkong.myqcloud.com/fedweb/1552917186945.pdf)
[13] B. van Berlo, A. Saeed, and T. Ozcelebi, “Towards federated unsupervised
representation learning,” in Proceedings of the Third ACM International
_Workshop on Edge Systems, Analytics and Networking, 2020, pp. 31–36._
[14] F. Zhang, K. Kuang, Z. You, T. Shen, J. Xiao, Y. Zhang, C. Wu,
Y. Zhuang, and X. Li, “Federated unsupervised representation learning,”
_arXiv preprint arXiv:2010.08982, 2020._
[15] E. Tzinis, J. Casebeer, Z. Wang, and P. Smaragdis, “Separate but together:
Unsupervised federated learning for speech enhancement from non-iid
data,” arXiv preprint arXiv:2105.04727, 2021.
[16] T. Luo and S. G. Nagarajany, “Distributed anomaly detection using
autoencoder neural networks in WSN for IoT,” Tech. Rep., 2018.
[17] V. Mothukuri, P. Khare, R. M. Parizi, S. Pouriyeh, A. Dehghantanha,
and G. Srivastava, “Federated-learning-based anomaly detection for iot
security attacks,” IEEE Internet of Things Journal, vol. 9, no. 4, pp.
2545–2554, 2021.
[18] V. Chandola, A. Banerjee, and V. Kumar, “Anomaly detection,” ACM
_Computing Surveys, vol. 41, pp. 1–58, 7 2009. [Online]. Available:_
[https://dl.acm.org/doi/10.1145/1541880.1541882](https://dl.acm.org/doi/10.1145/1541880.1541882)
[19] C. C. Aggarwal, Outlier Analysis. Springer International Publishing,
2017.
[20] Y. LeCun, C. Cortes, and C. Burges, “Mnist handwritten digit database,”
2010.
[21] H. Xiao, K. Rasul, and R. Vollgraf. (2017) Fashion-mnist: a novel image
dataset for benchmarking machine learning algorithms.
[22] S. Han, X. Hu, H. Huang, M. Jiang, and Y. Zhao, “Adbench: Anomaly
detection benchmark,” arXiv preprint arXiv:2206.09426, 2022.
[23] B. Scholkopf, R. C. Williamson, A. Smola, J. Shawe-Taylor, and J. Platt,¨
“Support vector method for novelty detection,” Advances in neural
_information processing systems, vol. 12, 1999._
-----
|
{
"disclaimer": "Notice: Paper or abstract available at https://arxiv.org/abs/2209.04184, which is subject to the license by the author or copyright owner provided with this content. Please go to the source to verify the license and copyright information for your use.",
"license": null,
"status": "GREEN",
"url": "https://arxiv.org/pdf/2209.04184"
}
| 2,022
|
[
"JournalArticle",
"Conference"
] | true
| 2022-09-09T00:00:00
|
[
{
"paperId": "0bff4af924788d9779041513b6894385eac51ffd",
"title": "ADBench: Anomaly Detection Benchmark"
},
{
"paperId": "048cb4d1b7712b7499a9d7db6d24caaab5ddd9ce",
"title": "Separate But Together: Unsupervised Federated Learning for Speech Enhancement from Non-IID Data"
},
{
"paperId": "795308ca0a281865b42b612045e5074076a82a75",
"title": "Federated-Learning-Based Anomaly Detection for IoT Security Attacks"
},
{
"paperId": "bf2ca8386bfc6a4c65a91f4628da7c49f931e9f2",
"title": "Federated unsupervised representation learning"
},
{
"paperId": "2baca0b49c71388abe13e7c59bda9414bab24497",
"title": "Towards federated unsupervised representation learning"
},
{
"paperId": "f9a855ae59579d16dca6a5133cd8daddd3305582",
"title": "A Survey on Distributed Machine Learning"
},
{
"paperId": "07912741c6c96e6ad5b2c2d6c6c3b2de5c8a271b",
"title": "Advances and Open Problems in Federated Learning"
},
{
"paperId": "2a3d09bbdfe21418ce75d6973f71028fa9192b89",
"title": "Federated Learning for Edge Networks: Resource Optimization and Incentive Mechanism"
},
{
"paperId": "afa778ba0ba6333e25671cfb691a4bdda13b2868",
"title": "Federated Learning With Differential Privacy: Algorithms and Performance Analysis"
},
{
"paperId": "79cf9462a583e1889781868cbf8c31e43b36dd2f",
"title": "Towards Federated Learning at Scale: System Design"
},
{
"paperId": "b97047c4dc75cbe8d6fc5cb3dd5a81d36458892d",
"title": "APPLIED FEDERATED LEARNING: IMPROVING GOOGLE KEYBOARD QUERY SUGGESTIONS"
},
{
"paperId": "7143230a68aecbce640e53b6cde171699a1e4270",
"title": "A Hitchhiker's Guide On Distributed Training of Deep Neural Networks"
},
{
"paperId": "b7ba8f3fa0c587695ed2f87b92e2b5410284413e",
"title": "Distributed Machine Learning in Coalition Environments: Overview of Techniques"
},
{
"paperId": "5cfc112c932e38df95a0ba35009688735d1a386b",
"title": "Federated Learning with Non-IID Data"
},
{
"paperId": "7389d201825f3581649adb15e0246367d1ed9d97",
"title": "Distributed Anomaly Detection Using Autoencoder Neural Networks in WSN for IoT"
},
{
"paperId": "d8c09661b1bebfb690f0566167c87d64c5628d73",
"title": "Demystifying Parallel and Distributed Deep Learning"
},
{
"paperId": "f9c602cc436a9ea2f9e7db48c77d924e09ce3c32",
"title": "Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms"
},
{
"paperId": "80ff2c726bdb4efdaac712bfc8712cfd4bb939ad",
"title": "Deep Learning for Mobile Multimedia"
},
{
"paperId": "7fcb90f68529cbfab49f471b54719ded7528d0ef",
"title": "Federated Learning: Strategies for Improving Communication Efficiency"
},
{
"paperId": "d1dbf643447405984eeef098b1b320dee0b3b8a7",
"title": "Communication-Efficient Learning of Deep Networks from Decentralized Data"
},
{
"paperId": "1bc042ec7a58ca8040ee08178433752f2c16f25e",
"title": "Outlier Analysis"
},
{
"paperId": "71d1ac92ad36b62a04f32ed75a10ad3259a7218d",
"title": "Anomaly detection: A survey"
},
{
"paperId": "bf206bad6a74d27b40c8ea77ee54e98e492fb7f9",
"title": "Support Vector Method for Novelty Detection"
},
{
"paperId": "0575cd39742118cb04c9df4e262fb5d22af48af8",
"title": "Centralised vs decentralised anomaly detection: when local and imbalanced data are beneficial"
},
{
"paperId": null,
"title": "The next ai revolution will not be supervised"
},
{
"paperId": null,
"title": "query suggestions,” arXiv preprint arXiv:1812.02903"
},
{
"paperId": null,
"title": "Mnist handwritten digit database"
}
] | 12,470
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.